Petroleum Exploration With Bayesian Networks

Petroleum exploration with Bayesian Networks: from
prospect risk assessment to optimal exploration.

Gabriele Martinelli, Department of Mathematical Sciences, NTNU
PhD thesis
Preface
This thesis is submitted in partial fulfilment of the requirements for the

degree of Philosophiae Doctor at the Department of Mathematical Sciences
of the Norwegian University of Science and Technology (NTNU). The thesis
contains the result of my research work under the supervision of my principal
advisor prof. Jo Eidsvik and my second and third advisors, dr. Ragnar
Hauge and prof. H
akon Tjelmeland.
The research was financed by the project FindOil of the research center
Statistics for Innovation (sfi)2 in Oslo.
I would like to take some time here to thank all those who helped me during
the time I spent at NTNU and in the other academic institutions that have
hosted me.
First and foremost, I would like to thank my principal advisor prof. Jo
Eidsvik. He has provided me with an invaluable guidance throughout these
three and a half years, from both a professional and a personal point of view.
Then, I would like to thank my second advisor dr. Ragnar Hauge for
his precious suggestions and inputs and for the opportunity of sitting with
the SAND group at the Norwegian Computing Center in Oslo for the fall
semester of 2010, and my third advisor prof. H
akon Tjelmeland for the
fruitful discussions.
I also would like to thank prof. Jim Berger and prof. James Smith from
Duke University for the opportunity of spending the spring semester of 2010
in the US and for the collaborations and the work that I have developed
during those months.
A personal thank goes to prof. Henning Omre, for his generous welcome in the spatial and computational statistics group within the Math
Department, and for the possibility of taking part and contributing to the
development of the MASTMO project in Awassa, Ethiopia.
I acknowledge all my coauthors, and in particular prof. Tapan Mukerji
from Stanford University, dr. Debarun Bhattacharija from IBM, dr. Ketil
Hokstad from Statoil and prof. Richard Sinding-Larsen from NTNU for the
opportunity of working with them and developing competences in different
domains.
I want here to express all my gratitude to Valeria, who has accepted

and sustained from the very beginning my decision of moving to Norway for
my PhD, and with whom I have shared an incredible, wonderful, sometimes
difficult, but at the end rewarding, long-term relationship for almost four
years. Skype and low-cost airlines have been good partners in this journey,
but the same journey would not have been possible without the love and
the commitment that I feel for you and from you every day, and for this I
want to thank you.
A warm and grateful thank goes to my wonderful family, that has been
on my side during this four years, especially in the cold and dark evenings at
the beginning of my experience in Norway. The passion, the dedication and
the sharp but sincere advices that I have received from my parents especially
during these last months have been invaluable and for this I want to thank
them.
In Trondheim and in the US, I have always found colleagues with whom
I have shared more than just seminars and work discussions and that have
become important friends for me: I want here to mention Antonio, Valeria,
Valentina, Luce and Maria from my experience in the US, and Thiago,
Rikke, Kjartan, H
akon, Erlend, David, Petter, Simone, Sara, Javad, GeirArne, Xiangping and Rupali from the department of mathematics at NTNU.
If my experience in Norway has been so positive it is also because of
the many special friends that I have met here during these years. With
many of them I have shared dinners, trips in Norway and abroad, many
sessions of biking and cross-country skiing, amusing and intense moments,
and especially that common and unique felling of being able to feel yourself
at home even when you are 2000 km far from it. I would like to spend a
sentence or even a page for each of them, but I barely have the room here
for naming them here: Fabio, Roberto, Alessandro O, Claudia, Stefano N,
Stefano C, Ivan, Alessandro N, Giovanni, Leonardo, Ezequiel, Gursu, Elena,
Francesco C, Cristina, Tania, Vangelis, Nicla, Magnus, Simone, Luca O,
Francesco S, Tanu, Valentin, Celine, Atsede, Luca F, Max, Daniele, Nicola,
Alfredo, Tayo, Francesca, Marta, . . .
I would like finally to express my greatest gratitude to two friends that
have been particularly close to me during my years in Trondheim. The
first is Aldo, who has taught me that in order to be a good scientist you
need also a broad mind, and the willingness of knowing and understanding
different fields and worlds. The second is Steffi, whose culture, empathy
and kindness have been a continuous and fascinating discovery for me, and
whose friendship will always remain one of the greatest gift I will carry away
from Trondheim.
Gabriele Martinelli
September 4, 2012, Oslo
Introduction
Introduction
Contents
1 Motivation
2 Geological background
2.1
The geology of a petroleum system . . . . . . . . . . . . . . .
2.2
Modeling dependencies in a petroleum system . . . . . . . . .
3 Bayesian Networks
3.1
3.2
Inference engines for BN . . . . . . . . . . . . . . . . . . . . .
10
3.1.1
Variable elimination algorithm . . . . . . . . . . . . .
11
3.1.2
Junction tree algorithm . . . . . . . . . . . . . . . . .
16
Specific network structures for HC problems . . . . . . . . . .
22
4 Static and Sequential decision problems

4.1
Basic notation
. . . . . . . . . . . . . . . . . . . . . . . . . .
30
4.2
Static decision problem . . . . . . . . . . . . . . . . . . . . .
31
4.3
Dynamic decision problem . . . . . . . . . . . . . . . . . . . .
33
4.4
A small example . . . . . . . . . . . . . . . . . . . . . . . . .
34
5 Outline of the thesis
25
38
Motivation
The primary goal of the thesis is to investigate the possibility of using

Bayesian Networks (BN) as a tool for modeling geological dependencies and
to improve decision making in a Hydrocarbons (HC) basin.
The setting of the problem is the following: we consider a HC basin
where geologists and geophysicists have gathered a series of preliminary
information and are now looking forward to establish an exploration plan.
1
GEOLOGICAL BACKGROUND
They can choose among a set of fixed predetermined drilling locations that
we will call segments throughout this Introduction. The geology of the field
and the multiple interactions between different geological elements, that will
be discussed in detail in Section 2 of this Introduction, makes it difficult to
correctly assess the correlation between different segments. It is natural
to imagine that if we find oil in a certain segment this will increase the
probability of finding oil in neighboring segments, but classical geostatistical
methods like kriging cannot be applied for solving this problem, mainly for
two reasons. The first one is that the process is still very much expert-driven,
and it is not easy to incorporate the geological knowledge in completely
data-driven geostatistical frameworks; the second one is that we are trying
not just to model the outcome of the exploration, but also to explain the
geological process that lies behind the presence or absence of HC, and BN are
a more natural tool to encode causal relationships than other geostatistical
tools. For all these reason the importance of quantitative geology has risen
in recent years, providing more accurate models to address this issue, and
this thesis has the ambition of showing the potential of a new tool, namely
BN.
The initial goal of this thesis has been therefore to study how statistical
tools like BN and in general graphical models could help in rephrasing the
problem of modeling dependencies in HC fields: the main geological concepts and some preliminary modeling assumption are discussed in Section
2. The reason why we have chosen BN and how inference in BN works
will be discussed in detail in Section 3. The idea of being able to model
not just the outcome (probability of oil or dry in the whole field), but the
single geological processes that drive the generation and accumulation of
HC is crucial to introduce the second main topic of this thesis, namely the
possibility of exploiting the graphical model structure to design optimal exploration strategies: the main concepts and references that we have used for
this study will be discussed in Section 4. An outline of the thesis is finally
presented in Section 5.
Geological background
As we have mentioned in the previous section, the evaluation of prospect dependencies, and how these can affect the analysis of a geologically appealing
area, are becoming an important factor to take into account when planning
an exploration program. It has been pointed out, see e.g. Rose (2001) and
VanWees et al. (2008), that the introduction of a prospect-interdependency
perspective may result in a substantial difference when discussing a sequential drilling program or an expected portfolio evaluation.
In current practice, the estimation of undiscovered resources is usually
based on a MonteCarlo simulation where each segment is characterized by
a certain size and probability of discovery. The prospect database can be
based on extrapolation of discoveries made in the past through creaming
methods or discovery models, see Meisner and Demirmen (1981) and Xu
and Sinding-Larsen (2005). The effect of prospect dependencies within this
framework has been studied by Carter and Morales (1991) and Kaufman and
Lee (1992), where sampling of common geological elements is considered
in order to improve the estimation. As VanWees et al. (2008) point out,
though, a strong disadvantage of all these probabilistic approaches is that
there is not a systematic method to update the probability of discovery in
the segments as the exploration progresses. Other approaches have been
characterized by dividing the risk associated to certain geological elements
in common risks and conditional segment risks. This approach, see Murtha
(1996) and Stabell (2000), has the advantage of allowing an update of the
probability of discovery, but is still quite simple and could not capture the
complex interactions existing among the segments.
Recently, Smalley et al. (2008) and Cunningham and Begg (2008) proposed approaches where modeling of prospect dependencies is coupled with
designing an optimal, possibly sequential, drilling program. The idea of using BN for updating the probability of discovery has been first presented in
VanWees et al. (2008). In their work a distance-weighted function is used
to determine the correlation coefficient between the prospects. The analysis shows that introducing a way to update the probability of discovery in
an interdependent network has substantial effects on portfolio simulations.
What is lacking in their analysis is a consistent way of integrating expert
opinions in the BN and in designing an exploration procedure that could
keep into account the intrinsic sequentiality of the process. Our work takes
its origin from these premises.
2.1
The geology of a petroleum system
Almost all of the worlds petroleum occur in sedimentary basins, i.e in basins
characterized by a first phase of subsidence, followed by sediment accumulation (Gluyas and Swarbrick, 2003). There are several important factors
that need to be kept into account when studying the formation of a sedimentary basin: among the others, we mention the type of basin, the sediment
rate supply, the burial and the thermal history and the possible uplifts and
erosions. All these components concur to determine whether a sedimentary
basin has the ability of becoming a petroleum province.
In the petroleum geology literature it is common to consider four elements as fundamental for this process. Three of them are geological features
that need to be present in every petroleum province, and are namely the
presence a source rock, the presence of a reservoir and the existence of a
structure called trap that prevents the HC from leaking out. The fourth key
factor is the timing between the three aforementioned geological elements,
especially the timing between the burial of the source rock deposition, the
migration and accumulation in the reservoir rock and the deposition of the
trapping seal rock. The importance of the timing effect is shown in Figure
1, that represents a typical petroleum system elements chart; we recognize
the deposition and burial time for the different factors and we can estimate
the critical moment, i.e. the moment when all the components are correctly
in place for the HC to accumulate.
The source rock is usually a sedimentary rock that contains sufficient
organic matter such that when it is buried and heated it will produce HC.
The main HC are liquid petroleum (oil) and hydrocarbon gas, and they
differ for density, gravity, and chemical composition. The term petroleum is
usually referred to a mixture of HC molecules and other molecules in smaller
quantities that are mixed with the HC. Different sorts of organic matters
(algae, woody tissues, . . . ) yield different sorts of petroleum. This concept
will be important for our analysis (Martinelli et al., 2011b), since we will
consider just oil-prone or gas-prone prospects or prospects that can trap
both kinds of HC.
Oil and gas are less dense than water and therefore they tend to migrate
upwards according to buoyancy until they reach a seal. Seals are usually
low-permeability rocks, typically shales, cemented limestones or salt rocks.
The term trap identifies the geometrical structure of the sealed petroleum-
Figure 1: Petroleum system elements chart: we recognize the deposition

time for the different factors and we estimate the critical moment (indicated
with an arrow), i.e. the moment when all the components are correctly in
place for the HC to accumulate.
bearing container. The simplest trapping configurations are domes (fourway anticlines), but fault blocks or other more complex structures can act
as potential traps and they will be taken into account in this work, especially
in Martinelli et al. (2012b). An example of different trap structures is shown
in Figure 2.
Anticline trap 1
Anticline trap 2
Fault trap
Figure 2: Anticlinal and fault traps. The picture comes from the Bezurk
case study, used in Martinelli et al. (2012a).
By the term reservoir we mean a rock plus void space contained in a trap.
Since oil-filled caves are uncommon, we prefer to talk about reservoir rocks,
usually porous and permeable rocks such as coarse-grained sandstones or
carbonate rocks. The petroleum together with some water migrate into the
pore spaces between the grains or the crystals of the reservoir rock. Porosity,
permeability, grain shape and size and other factors contribute to the success
of a potential reservoir rock. These factors will be crucial in our analysis,
and we will consider them in Martinelli et al. (2012b) and in Martinelli et al.
(2012a). The porosity is the void space in the rock and is usually reported
as a percentage (often the effective porosity, i.e. the void space that can be
actually filled with HC, is reported). The permeability is a measure of the
degree to which fluid can be transmitted, and it is reported in milliDarcies
(mD). It is important to recognize that usually gas may be produced from
reservoir with very low permeability, while a higher permeability is required
to produce oil reservoirs.
An example with the different source, reservoir, and seal layers, taken
from a synthetic case used in Martinelli et al. (2012a) is shown in Figure 3.
Seal
Reservoir
Source
Reservoir
Source
Figure 3: Elements of the petroleum system, recognizable in different layers

deposited through geological time. The picture comes from the Bezurk case
study, used in Martinelli et al. (2012b).
Pressure and temperature differ from the trap (subsurface) to the Earths
surface. This is why when we report volumes of expected HC accumulations
in reservoirs we have always to specify whether they are intended at surface
or subsurface conditions. The ratio between these volumes is called formation volume factor. Further, not all of the HC present in the reservoir can
be brought to surface. For this reason we need to take into account the
so-called recovery factor, that specifies the ratio between the HC in place
and the recoverable quantity of oil and gas that can be effectively produces.
We will consider this factor in both Martinelli et al. (2012b), and Martinelli
et al. (2012a). In the latter we will distinguish between geological discovery
and commercial discovery to remark the fact that, because of this problem, it is not convenient to produce fields whose volume is under a certain
threshold.
2.2
Modeling dependencies in a petroleum system
We are particularly interested in this thesis in modeling the dependency

structure of these geological elements in a HC field. A first assumption
that we will use throughout the whole work (see especially Martinelli et al.
(2011b), Martinelli et al. (2012a) and Martinelli et al. (2012b)) is the independence between the three main components of the petroleum system,
namely source, reservoir and trap. This is a common assumption, see for
example Bickel et al. (2008) where three separate models, one for each component of the petroleum system, are used for identifying pairwise correlations
between prospects. The assumption is crucial because it allows us to model
the dependency separately for each of these three factors.
For the dependency modeling, we use BN. These allow us to actually
build the geological understanding of the phenomenon into the model. Due
to the nature of the problem, with no repeated data, at the beginning of
our work, see Martinelli et al. (2011b), neither structure nor parameters
in the BN could be estimated. Everything must be set by experts, and so
our model is an expert system, where we try to represent the geological
knowledge already present.
The graphical nature of the BN give a nice representation of the elements
present, and all components in the model have a direct physical interpretation. We take the prior probabilities for success for each factor at each
prospect as given input, and build the correlation model without perturbing
the initial values of these. Setting these priors is a very different modeling
problem.
In the most recent works, see Martinelli et al. (2012a) and especially Martinelli et al. (2012b), parameters and structure of the network are learned
directly from data, in an attempt of solving the problem of the elicitation
that is a common well-known problem in the BN community. In recent years

mixed expert-based and data-trained approaches have been developed with
success in other fields, such as environmental management, see Pollino et al.
(2007) and Borsuk et al. (2004). In VanWees et al. (2008) another application of BN for oil and gas exploration can be found. A similar approach has
been used in another very recent paper by Rasheva and Bratvold (2011).
The study of petroleum geology in recent years has been greatly improved
by the foundation of Basin and Petroleum System Modeling (BPSM) (Magoon and Dow, 1994) (Allen and Allen, 2005). By this term we mean the
development of modeling systems able to allow the geoscientists to examine
the dynamics of sedimentary basins and their associated fluids to determine
if past conditions were suitable for hydrocarbons to fill potential reservoirs
and be preserved there. BPSM software has become popular in the oil & gas
exploration industry and has improved significantly the understanding of the
main processes driving the generation, the migration and the accumulation
of HC.
In essence, BPSM tracks the evolution of a basin through time as it fills
with fluids and sediments that may eventually generate or contain hydrocarbons. The main partial differential equations that describe the petroleum
generation and migration are discussed in Hantschel and Kauerauf (2009).
In contrast with the more common Reservoir Modeling, the BPSM model
geometry is dynamic and not static, and often changes significantly during simulation, as a consequence of the deposition of new layers or of the
presence of faulting episodes. The different processes of sediment deposition, burial, maturation kinetics, fluid flow and faulting may be examined
at several levels, and complexity typically increases with spatial dimensionality. Two-dimensional modeling, either in map or cross section, can be
used to reconstruct HC generation, migration and accumulation along a
cross section. Three-dimensional modeling reconstructs petroleum systems
at reservoir and basin scales and has the ability to display the output in 1D,
2D or 3D, and through time.
One of the main weakness of this technology is the lack of efficient riskanalysis tools and especially the lack of a decisional framework that could
organize the different information that we get from the modeling. The interest of uncertainty analysis in basin modeling increases as the computer
power makes it possible to assess multiple models in a reasonable time. Much
BAYESIAN NETWORKS
work has been done on the uncertainty ( Zwach and Carruthers (1998) and
Corradi et al. (2003)), but the focus is mainly on traditional Monte Carlo
techniques which randomly draw values from statistical distributions of the
input parameter and compare the differences in the result for each drawn
input parameter. More recently even spatial and structural uncertainty has
been considered (Jia, 2010).
In Martinelli et al. (2012b) we try to study how the uncertainty of some
critical parameters in BPSM could affect decision making, and we show how
it is possible to integrate classical VoI based decision techniques with BPSM.
We believe that there is actually need of more integration between quantitative geology methodologies such as BPSM and decision making analysis, in
order not only to describe the uncertainties present in the model, but also
to show how this uncertainty can affect our decision.
Bayesian Networks
BN or Bayesian belief networks are probabilistic models where a graph structure is used to represent a set of random variables and their conditional
dependencies. BN belong to a larger class of models, generally referred to
as expert systems.
An expert system is defined as a knowledge base plus an inference engine
(Cowell et al., 2007): the knowledge base contains the information of a
problem, encoded in some manner, while the inference engine consists in
algorithms for processing the encoded knowledge together with any further
specific information. For BN, the knowledge base is represented by the
causal relationships between the variables, encoded in a Directed Acyclic
Graph (DAG), while the inference engine is based on Bayes theorem. We
will discuss in the next sections the main concepts and the main inference
engine used nowadays for performing inference in large networks.
When designing a BN model or in general a graphical model, we can distinguish a qualitative phase and a quantitative one. In a BN the qualitative
part is represented by a graph, that should be directed and acyclic. Vertices
in the graph represent variables. Directed edges represent probabilistic influence or causal mechanisms Pearl (1984). A graph G can be described by
a pair G = (V, E), where V is a finite set of vertices or nodes, and E is a
subset of the ordered pairs of vertices V V called edges of G. If (, ) E,
BAYESIAN NETWORKS
10
but (, )
/ E, then the edge is directed and is a parent of . The set of
parents of a vertex is denoted by pa(). The set of children is denoted by
ch(). The union between these two sets (parents and children) is denoted
by ne(), i.e. neighbors of . If there is a path from to (i.e. a sequence
of vertices n such that (i1 , i ) E i, and 1 = , N = ), we write
7 . The acyclicity condition is ensured by the absence of cycles, i.e.
directed paths whose starting point coincides with the ending point. The
set of vertices such that 7 is defined ancestors of or an(), and
the set of vertices such that 7 is called descendants of or de().

A graph is called complete if every pair of vertices is joined. A complete
subgraph which is maximal with respect to is called a clique.
A BN is completed by its quantitative part: the most important in-
gredient is a collection of random variables (Xv ), v V , and a probability
distribution P on Xv which factorizes according to G. For details about
the exact probabilistic definition of factorization we suggest Cowell et al.

(2007). For our purposes it is sufficient that the probability distribution P
has density p with respect to the product measure = vV v given by
Q
p(x) = vV p(xv |xpa(v) ). We will usually deal with discrete BN, i.e. net-
works where all the variables have a discrete distribution over a finite number
of states k. In our applications k ranges from 2 or 3 in (Martinelli et al.,
2011b) and Martinelli et al. (2011a) to 6 in Martinelli et al. (2012a). The

conditional probability distributions p(xv |xpa(v) ) are therefore expressed by
Conditional Probability Tables (CPT) of size k #[pa(v)]1 , where #[pa(v)] is

the number of parents of node v. If a node has no parents, his CPT is simply
given by a vector of size k. In Martinelli et al. (2012b) we will deal with a
mixed discrete-continuous BN, where the continuous nodes are distributed
according a Gaussian distribution whose parameters depend on the state of
the discrete parents. The inference for this kind of networks remains outside the scope of this section. An important property among the Markovian
properties on DAGs is the d-separation property Geiger and Pearl (1993):
given A, B and S disjoint subsets of G, we say that S d-separates A from B
A
B | S.
3.1
Inference engines for BN
By inference in BN we refer to the task of computing a-posteriori probabilities. This task can be found under different names such as probability prop-
BAYESIAN NETWORKS
11
agation, belief updating. Most queries involve observations or evidence;

evidence on a variable is a statement of the certainties of its state.
We can be interested in knowing the likelihood of a particular evidence
P
xi P (x1 , . . . , xn , e), or to compute the a-posteriori probability
e: p(e) =
for a given target or interest variable: P (xi |e) =
P (xi ,e)
P (e) .
In general, we are
not only interested in a single variable but in a set of them, usually all the
unobserved ones.
Computing the a-posteriori probability of a given variable is useful in
different situations:
Predictive or deductive reasoning: What is the probability of observ-
ing a symptom knowing the presence of a given disease? Or, what is

the probability of observing oil in a certain prospect, given that we
know that a producing source rock has been found in the area? In this
case the target variable usually is a descendant of the evidence.
Diagnostic or abductive reasoning: What is the probability of disease

D being the correct diagnosis given some symptoms? Or, what is the
probability that a prospect where we have found the presence HC has

been charged by a certain type of source rock? In this case the target
variable usually is an ancestor of the evidence.
In the BN framework the direction of the links between the variables does
not constraint the type of query to be posed.
We will now discuss briefly the simplest algorithms (brute force, variable elimination), and we will move later to the Junction Tree Algorithm
(Lauritzen and Spiegelhalter, 1988), that is the algorithm implemented in
Murphy (1999) and widely used throughout the whole thesis.
3.1.1
Variable elimination algorithm
Given a BN with N variables X1 , . . . , XN and their discrete probability

distributions, then the computation p(xi ) (or p(xi |e)) can be considered
conceptually relatively easy, but usually computationally complex.
From here on it is important to point out that we will use the short
notation p(xi ) in stead of p(Xi = xi ) and we will use the functions fi (xi ) or
simply fi to indicate p(xi |xpa(i) ). We will study the more common inference
engines for BN considering a simple case study with N = 5 nodes presented
BAYESIAN NETWORKS
12
in Figure 4. For this case study therefore f1 = p(x1 ), f2 = p(x2 |x1 ), f3 =

p(x3 |x2 , x4 ), f4 = p(x4 ), f5 = p(x5 |x1 ).
X2
X1
X3
X5
X4
Figure 4: Simple BN used for inference purposes in Section 3.1

The brute-force approach consists in computing:
X
p(xi ) =
x1 ,...,xi1 ,xi+1 ,...,xn
n
Y
j=1
fj (xj )
As we can see, it presents the same problems of the computation of the

joint probability density, i.e it is computationally very inefficient and even
intractable in most of the cases.
If our goal is to compute p(x2 ), according the brute-force approach, we
proceed in this way:
X
p(x2 ) =
x1 ,x3 ,x4 ,x5
x1 ,x3 ,x4 ,x5
5
Y
j=1
fj
p(x1 )p(x2 |x1 )p(x3 |x2 , x4 )p(x4 )p(x5 |x1 )
If all the variables are binary, this implies that we have to construct a a
probability table with 32 entries (i.e. the joint probability distribution).
This means that we need a huge number of multiplications and additions
for the marginalization of x1 ,x3 ,x4 and x5 .
We can simplify the process using the distributive law and moving in
some additions: in this way, we have:
p(x2 ) =
X
x1
X
x5
p(x5 |x1 ) p(x1 )p(x2 |x1 )
"
X X
x4
x3
p(x3 |x2 , x4 )
!#)
BAYESIAN NETWORKS
13
This means that now we are dealing with just one table of size 8 and three
tables of size 4.
This technique works (or allow an easy inference) if the DAG which we
are dealing with has no cycles (i.e it is a poly-tree). In this case, indeed,
we can can move the additions in such a way that we never create a table
larger than those included in the BN representation. In the general case
(the underlying undirected graph has cycles) inference is NP-Complete, see
Cooper (1990). The complexity of the previous method is exponential in
the width (number of variables minus one) of the largest factor set involved
in the process. The key to efficient inference with this method lies in finding
a good summation order (or elimination order).
What happens if we have some observations, i.e.e? Before running our
algorithm, for each ei e, we identify the potentials or probability functions
fi (xi ) in which it is included, then:

fi (xi ) =
fi (xi ) if xi is consistent with e

0
otherwise
We can summarize this approach in the Variable Elimination (VE) Algorithm:

1. Let F be a list containing all the probability functions {f1 , . . . , fn }
2. Enter the evidence e
3. Select an elimination order containing all the variables but the target
one Xi .
4. for k = 1 to n 1:
Let Xk be (k).
Let F be the set of probability functions in F that contains Xk

P Q
Let f = xk
f F f
F = (F \ F ) f
5. Combine in a single function f all the functions in F. Normalize f to

obtain p(xi )
BAYESIAN NETWORKS
14
The elimination order chosen is very important, since it allows to decrease in a sensible way the complexity of the computations involved, as we
can see in an example.
Let us consider again the example if Figure 4; ci () are functions that
copy fi , but keep tracks of all the variables involved, i.e. c2 (x1 , x2 ) = f2 =
p(x2 |x1 ).
In order to calculate p(x2 ) with the VE algorithm, we proceed in this
way:
P
1. Delete X5 : g(x1 ) =
x5 c5 (x5 , x1 )
2. Delete X4 : g(x2 , x3 ) =
3. Delete X3 : g(x2 ) =
x3
4. Delete X1 : g 0 (x2 ) =
x4 c4 (x4 )
c3 (x3 , x2 , x4 )
g(x2 , x3 )
x1 c1 (x1 )
g(x1 ) c2 (x2 , x1 )
5. Calculate p(x2 ) normalizing g 0 (x2 ) g(x2 )

VE algorithm can be implemented also as a propagation algorithm, i.e.
as an algorithm that base their operation mode in a process of message
passing among the nodes in the network. Thus, nodes act as processors that
receive, calculate and send information.
In a propagation algorithm there are two basic operations that a node
can perform:
Ask-information(i, j). Node i asks information to node j. If node
j has already received information (messages) from all its neighbours

but i, then Send-message(j, i); otherwise, node j has to propagate the
request: Ask-information(j, k): for all k ne(j), k 6= i.

Send-message(i, j). Once node i has received messages from all its
neighbours but j, it sends a message to node j; a message is a potential
defined over Fi Fj , computed in the following way:
M
ij
X F
/ i Fj
fi
k6=j
ki
BAYESIAN NETWORKS
15
Finally, we can calculate p(xi |e) locally in each node i as a normalization of

the following quantity:
p(xi |e) =
xj 6=xi
fi
kne(xi )
ki
We can compute the same probability p(x2 ) as before through the message passing algorithm: for the sake of clarity we show in Figure 5 the
family of each node. The directions of the arrows represent the flow of the
messages.
X2 ,X1
M1-->2
M3-->2
X1
X3 ,X2, X4
M5-->1
X5, X1
X4
M4-->3
Figure 5: Propagation Algorithm (one-way) applied to the example of Figure

4.
The goal is to compute p(x2 ), so we proceed in this way:
1. Ask(2,1) and Ask(2,3)
2. Ask(3,4) and Ask(1,5)
3. M 51 =
x5 c5 (x5 , x1 )
4. M 43 = c(x4 )
5. M 12 = c1 (x1 ) M 51
6. M 32 =
x3 ,x4
c(x3 , x2 , x4 ) M 43
7. p(x2 ) = normalize (
x1 ,x3 [c(x2 , x1 , x3 )
M 12 M 32 ])
In general, computing the a-posteriori probability for all the variables in

the network can be done through the Message Passing algorithm by using
exactly two waves of messages. Let us consider figure 6; the process is the
following:
BAYESIAN NETWORKS
X2 ,X1
M1-->2
X1
16
M2-->1
M5-->1
M3-->2
M2-->3
M1-->5
X3, X2, X4
M3-->4
X5 ,X1
X4
M4-->3
Figure 6: Propagation Algorithm (two-ways) applied to the example of Figure 4.

1. Select a node as root (example, node 2)
2. Ask evidence from the root to its neighbors. This step provokes messages in upward direction in all the links of the network.
3. Distribute evidence from the root to its neighbors (and go on recursively). This operation provokes to have message in the opposite
(downward) direction.
4. Calculate marginal distributions in each node by local computation,
i.e, using its incoming messages.
3.1.2
Junction tree algorithm
Let us consider now the DAG shown in Figure 7. As we can see, this DAG
differ from the one presented in Figure 4 for the presence of one more edge
from node X4 to node X5 .
X2
X1
X3
X5
X4
Figure 7: Example used in Section 3.1.2. Note that the BN is no longer a

poly-tree, but it presents a cycle.
BAYESIAN NETWORKS
17
In this case the message propagation algorithm and the variable elimination algorithm as they have been described before do not work. Here, if
we choose, as an example, X2 as root, we have the following sequence of
operations:
Ask(2,1) and Ask(2,3)
...
The algorithm does not work since there are multiple paths going from a
node to another node, i.e. there are cycles in the network and we are not
dealing any longer with poly-trees.
The solution is to eliminate the cycles, through the creation of a tree
of cliques or Junction Tree or Join tree. Briefly, we can reformulate the
inference problem as a two-stage process:
Preprocess the network in order to get a secondary structure: a tree;
the idea is to group in an appropriate way the variables in the network
and to structure the clusters in a tree-shaped structure.
Then, apply propagation algorithms (or variable elimination) over the
secondary structure.
The entire process takes in the name of Junction Tree Algorithm (JTA)
(Lauritzen and Spiegelhalter, 1988).
The first step is therefore to build the junction tree itself. A junction
tree verifies the running intersection property: given two nodes (or clusters)
G1 and G2 , if X = G1 G2 6= , then X has to contain all the nodes in the
path connecting G1 with G2 . The probability tables (potentials) associated
to the join tree nodes must represent a factorization of the joint probability
distribution defined by the original network.
In order to achieve this goal, standard algorithms exists. Cowell et al.
(2007) present the following recipe:
BAYESIAN NETWORKS
18
1. Moralize the DAG.

2. Triangulate the moral graph and obtain the cliques its contain.
3. Organize the cliques in a junction-tree structure.
4. Build a potential representation of the original probabilistic model.
The first step is therefore represented by the moralization technique, that
consists in the following:
i) For each node Xi , connect pairwise all parents of Xi with undirected
edges.
ii) Drop the direction of all original (directed) edges.
The moralized version of the DAG shown in Figure 7 is presented in
Figure 8. In this simple case the moralized graph contains just a an edge
more than the original graph.
X2
X2
X1
X3
X5
X4
X1
X3
X5
X4
Figure 8: Moralization technique applied to the example of Figure 7: an

extra edge is added in order to marry the parents of node X3 , then the
original DAG is made undirected.
An undirected graph is triangular or chordal if there exists no cycle of
length 4 without a chord. Thus, we can obtain a triangular graph from
a non-triangular one by adding new links (so called fill-ins). Note that the
possible triangulation is not unique and optimal triangulation (optimal for
guiding the JTA) exist, but are often not easy to find.
The triangulation is guided by a deletion sequence . Given an undirected graph G = (V, E) the fill-in process adds new edges (fill-ins) until the
graph is triangular. In details, the steps are presented in Algorithm 1. If we
choose a simple deletion sequence = {1, 2, 3, 4, 5} and we apply the algo-
rithm to the example shown in Figure 7, we get the undirected triangulated

graph shown in Figure 9.
BAYESIAN NETWORKS
19
X2
X2
X1
X3
X5
X4
X1
X3
X5
X4
Figure 9: Triangulation technique applied to the example of Figure 7: an

extra edge is added to the undirected graph in order to make the graph
triangular.
Algorithm 1 Triangulation algorithm
E = ;
for i = 1 to dim(V) do
Xi = (i)
Let E be the set of fill-ins to be added in order to make ne(Xi )
completely connected.
E = E E
Delete from the original graph node Xi and all its incident edges
end for
return GT = E E .
If G is a triangular graph we can always find a perfect deletion sequence,
that is, a sequence that when used to guide the triangulation sequence does
not add any fill-in.
The third step of the JTA is to organize the cliques in a junction-tree
structure. A clique is a maximal complete subgraph of GT , where complete
means that every pair of vertices is joined, and maximal means that we
cannot find a bigger set with the same properties, that includes all the
elements of the original set . We can identify the cliques during the deletion
process or directly from the graph, using algorithms such as the maximum
cardinality search.
In the graph of Figure 9, for example, we can identify the following
cliques: {X1 , X2 , X5 }, {X2 , X4 , X5 }, {X2 , X3 , X4 }. The goal is to arrange
the cliques in a junction tree structure. The pseudo-algorithm is the following:
1. Create an undirected graph with the cliques (C1 , . . . , Ck ) as nodes and
the set of edges corresponding to the cliques that share at least one
BAYESIAN NETWORKS
20
node: {(Ci , Cj ) such thatCi Cj 6= }.

2. Assign a weight w(Ci , Cj ) = |Ci Cj | to each edge of the graph.
3. Obtain a maximum spanning tree, i.e. dispose the cliques in a tree of
maximum length.
In the example shown during this presentation of the algorithm we have
just two possibilities of building the junction tree. We show the undirected
clique graph in Figure 10. As we can see, two out of the three connecting
edges have a weight equal to 2, while the remaining one has a weight equal
to 1. When building the maximum spanning tree, therefore, we select the
order that dispose the cliques in a tree of maximum length and we obtain
the tree shown in Figure 11.
X1,X2,X5
X2
X1
X3
X2,X3,X4
X5
X4
X2,X4,X5
Figure 10: The cliques of Figure 9 are arranged in a Junction Tree structure.
X1,X2,X5
X1,X2,X5
X2,X4,X5
X2,X3,X4
X2,X4,X5
X2,X3,X4
Figure 11: The final Junction Tree is built from the cliques of Figure 9 , in
order to respect the Maximum spanning tree rule.
The fourth and last step is building the potential representation. Now
we have a set of new nodes, that we will call clusters; each cluster Ci has a
potential Ci obtained by the following process:
BAYESIAN NETWORKS
21
Initialize all the potentials as unitary.

For each variable Xi in the network, identify a cluster Cj containing
all the variables involved in fi (if they appear in more than one cluster,
break ties randomly).

Update Cj by multiplying it by fi .
After this process we get a potential representation that is a factorization
of the j.p.d.:
p(x) =
t
Y
Ci (xCi )
i=1
In the previous example a possible potential representation is the following:

C1 (xC1 ) = f1 (x1 )f5 (x5 |x1 )f2 (x2 |x1 )
C2 (xC2 ) = f4 (x4 |x5 )
C3 (xC3 ) = f3 (x3 |x2 , x4 )
If we apply the propagation algorithm described in Section 3.1.1 over
a join tree, we get the so-called Shenoy-Shafer (Shenoy and Shafer, 1990)
architecture. In this architecture we need space to store two messages in each
separator (one mailbox for each direction), see Figure 12. Initial potentials
associated to cluster are not modified during the propagation.
The messages through each separator are computed in the following way:
M ij (e) =
Ci (e)
M ki (e)
k6=j
Ci \Cj
Procedurally, cluster i computes the product of its local potential Ci and

the messages from all clusters except j, marginalizes out all variables that
are not in j and then send the result to j.
At the end, after a full-propagation (upward + downward) all the separators are full and the following is true:
pCi (e) = Ci (e)
M ki
This is the so-called cluster belief related to cluster Ci after observing an

evidence e, and is the product of the clusters local potential and the message
received from all his neighbors.
Propagation over
over join
Propagation
join trees:
trees: Shenoy-Sha
Shenoy-Sha
44 NETWORKS
NETWORKS WITH
WITH CYCLES
CYCLES
3
16
16
we apply
apply the
the propagation
propagation
alg
If we
algo
some slides
slides ago
ago over
overaajoin
jointree
tre
some
so-called
Shenoy-Shafer
Archi
so-called22Shenoy-Shafer Archit
BAYESIAN NETWORKS
Ci
Ci
Ci
!i
!
Computing messages:
Mji Mij
Cj
Initial potentials associated to
Initial
potentials
to c
modified
during associated
the propagatio
modified during the propagation
Computing messages:
Mji Mji
Mij Mij
Cj
In this
this architecture
architecturewe
weneed
needspa
sp
In
messages
in
each
separator
messages in each separator (o(o
direction).
direction).
Mi
Mi
!j
Cj
!j
"
!(Ci )
Ci \Si j!(Ci )
"
Ci \Si j
After a full-propagation (upwar

the
are full and
the
Afterseparators
a full-propagation
(upward
Figure 12: Shanoy-Shefer architecture for the Junction tree algorithm:

in
the separators
are full and the
14:from the cliques potentials.
each separator we store two messages,Figure
computed
P(Ci , e) = !(Ci )
Figure 14:
P(Ci , e) = !(Ci ) #
Last part:
givenetwork
explanations
about why
it actually
works and discussion on
3.2
Specific
structures
for HC
problems
the complexity of the algorithm
c Jose A. Gamez - EPSA/UCLM
Different
networks
structures haveabout
been used
the
thesis
in order
c Jose throughout
A. Gactually
amez - EPSA/UCLM
Last part:
give explanations
why
it
works
and
discussion
tothe
incorporate
the of
desired
correlation structures among the variables. The
complexity
the algorithm
main idea is that the BN used for modeling HC problems are not learned
directly from the data; their structure is in fact imposed by the modeler in
order to obtain a correlation structure consistent with the geological understanding of the problem.
As pointed out in Section 2.2 the first basic assumption that we have always used when building BN for modeling HC problems is the independence
between the different elements that compose a petroleum system. This assumption translates in the possibility of building separate BN for each of the
elements of the petroleum system: the different branches are then linked to
a common node at the bottom that represents the effective presence of HC
in the considered segment. An example is shown in Figure 13.
In Martinelli et al. (2011b) we propose a model restricted to the source
component of the petroleum system. In the proposed network we distinguish
three kinds of nodes, kitchens, prospects and segments. The latter represent
physical locations that the exploration management and the geologists could
choose for an exploration campaign. The structure of the network is guided
by the geological map of the likely HC migration paths indicated by experts
in local geology. The edges are directed, and the direction is imposed by the
on
Advance
Advanced
BAYESIAN NETWORKS
23
TrapA
ResA
A
SouA
..A
Figure 13: Independence among geological elements modeled in a BN

migration direction. Each node has three possible outcomes, oil, gas or dry.
We have excluded here the simultaneous presence of oil and gas, but in later
works we have relaxed this assumption. The CPT between the variables of
the network are built in conformity with two principles: a physical principle
and a simplicity principle. The physical principle states dry states only
propagates dry and that gas can only appear if at least one parent is gas
as well, while oil can be generated even if the parents are all gas, since the
presence of oil can be seen as a failure in the propagation of gas (for details
see Martinelli et al. (2011b)). It is always considered the probability of a
local failure: this means that even if all the components (in this case the
source presence) are present at a prospect level, the same component can
fail at a segment level. This is due to the possibility of a local migration
failure that in conventional HC basins is always present.
The weakness of this approach is the complete subjectivity of the network
building process. For this reason in subsequent studies we have chosen to
focus our attention on more general network structures that could still model
correlation among geological elements, and at the same time be more robust
to sensitivity analysis or changes in the prior knowledge.
For this reason, in Martinelli et al. (2012a), we have proposed different
network structures for the components of the petroleum system, in order
to describe different kinds of dependence for various building blocks (subnetworks) of the large BN. We will here resume four possible structures,
used both in Martinelli et al. (2012a) and in Martinelli and Eidsvik (2012).
Independent: Some geological features can be essentially independent. This means that a partial or complete knowledge about a feature
in a particular site does not affect the distribution of the same feature
BAYESIAN NETWORKS
24
in the neighbor sites. See Figure 14 (left).

Common parent network: Some other features show a common
cause structure, i.e. they are affected by the presence of a common
element, as we can see in Figure 14, center. In this case a success in the
source observation will strongly increase the likelihood of the presence
of the charge in all the other sites (it is a confirmation), while a failure
will just slightly decrease this likelihood. Plus, once we get a positive
evidence, other positive or negative evidences would not change the
local probabilities, since we already know that the common parent is
in place.
E
CP
L2
B
A
B
A
E
D
L1
B
A
Figure 14: Independent Network (left), Common Parent or Counting Network (center) and Multi-Level Network (right): four possible ways to model
mutual interaction among prospects.
Common parent counting network: The structure of the network
is identical to the structure chosen for the common parent network, see
Figure 14 (center). The difference is that here we consider a common
parent with multiple states. In this way, the more positive answers
we get from the children, the more likely to get a success in the other
nodes. The physical reasoning in this case is that we are trying to
model a phenomenon whose success rate is uncertain, and the more
evidence we collect, the more certain this rate becomes. This structure is especially useful when we are trying to model a feature with
continuous levels, such the producibility of a reservoir (a feature linked
to its porosity and permeability), and a confirmation of its quality in
certain locations makes it more likely to obtain positive answers from
other sites that share the same reservoir rock.
Multi-Level network: We use this structure to model features that

may depend on covariates, often other geological mechanisms. We
model this phenomenon by grouping children node to a few parent
STATIC AND SEQUENTIAL DECISION PROBLEMS
25
nodes. The resulting graph is shown in Figure 14 (right). This network

introduces asymmetry in the correlation structure. It works in the
following way: an observation in a node belonging the second level (D
or E) affects the upper level (A, B or C) in the very same way as an
observation coming from the first level itself. This effect does not work
in the other direction, since the propagation from L1 to L2 can fail.
The more complex is the correlation network, the more parameters we
have to fix in order to reproduce the required marginal probabilities. The
idea is that the marginal probability of presence/absence of each geological
feature can be estimated quite accurately a priori, using software such as
R These marginals are kept in the bottom nodes, and they help to
GeoX .
set some of the parameters; the others are optimized as a function of the
degree of correlation that we want to introduce. In this way we can play

with the network and study the behavior in different conditions.
Ultimately, we believe that the process can be at least partially automatized by the mixed use of data and expert knowledge. We have anticipated
this argument, that has been explored and developed in Martinelli et al.
(2012b).
The proposed approach and the implementation of different correlation
structures for different HC models can find practical application on the
oil & gas industry. Currently, segments and prospects reserves of HC are
estimated by imputing a range or a distribution for each of the possible
volume/fluid parameters. Correlation elements are then introduced between
segments/prospects, both for what concern the volumetric parameters and
for what concerns the risking parameters (presence/absence of key geological
elements). Usually these correlation elements are simply given by common
parent structures (if reservoir fails in A, it is likely that it will fail also in
B) or maximum correlation structures (if reservoir fails in A, it will fail also
in B). We believe that our structures and our formulation of the problem
through BN could substantially help in capturing more complex and realistic
geological situations.
Static and Sequential decision problems
The original motivation for studying optimal exploration decision problems

comes from Martinelli et al. (2011b), where a BN has been developed in order
26
to include spatial covariates when organizing the exploration of a petroleum

field. As we have seen in the previous sections, the covariates are defined
with a deep understanding of the critical geological elements of the area,
and they are then translated in a BN, that proves as an efficient instrument
for evaluating a high number of possible scenarios.
In HC exploration we can distinguish four main actions that need to be
performed in order to guarantee the success of the exploration: acquisition,
exploration, appraisal, and production. The costs usually increase in this
sequence, while the risk decreases in this sequence, since we gather more
information about the presence, the volume and the quality of the resources
in place. In the acquisition phase the company performs an initial evaluation
and acquires the working interest. In the exploration phase the geological
evaluation continues and some preliminary exploration wells are drilled. In
the appraisal phase, engineers perform reservoir simulations, a production
plan is established, and the platform and other infrastructures are designed.
Finally, in the development or production phase, the platform is installed,
the production wells are drilled and the production starts.
In the following discussion we will slightly simplify this model by considering just the exploration and the production phases: in the exploration
phases we gather perfect or imperfect information through exploration wells
or seismic collection (or other kinds of geological surveys). In the production phase we define which prospects we are going to drill and possibly their
order, based on the information that we have collected in the previous phase.
Let us now first consider an exploration problem over a graphical model.
The distribution assigned to the model encodes the spatial dependency
among the prospects. This prior distribution p(x) refers to a certain distinction of interest, that in our case can be the lithology expected at a certain
depth. We work with a discrete model, with two or three possible outcomes,
(gas, oil and dry, or oil saturated sand, brine saturated sand and shale, . . . ),
that usually coincide with the lithology-fluid outcomes of an inverse seismic
problem. The model can be complicated if we furthermore allow the decision maker (DM) to have the possibility whether to acquire data directly
from a certain geological/seismic survey y or to acquire perfect information
directly through an exploration well.
As presented above, the ultimate goal is to design an optimal drilling
sequence that can be used as a road map by the decision maker in the
27
production phase. The sequence has to take into account the possibility
that any of the many multiple scenarios can arise, and therefore has to be
flexible enough to incorporate the outcome of the previously drilled wells
in the sequence itself. Since neither performing the survey nor drilling an
exploration well come for free, several different questions can arise. Among
them, for example:
Is it worth acquiring the survey under consideration?
Where is the best location for drilling an exploration well?
Is it really worth to drill an exploration well or is it better to start
immediately with production wells?
We need to schedule in advance all the exploration campaign: where

is it better to go (possibly first)?
In order to answer some of these questions, we have first to set on the

table the different options that exist for the DM. The first main option is
whether we are considering a static or a dynamic decision problem:
Static decision problem: once we have collected the information, we
have to decide immediately the full drilling plan. We may be allowed
to drill a single well or many wells. In this case we know in advance
the number, and we have to be able to identify a priori where it is
optimal to drill.
Dynamic decision problem: once we have collected the information, we
can sequentially decide which prospects we can drill and their order,
depending on the outcome of the previously drilled prospects. In this
case the idea is to solve a decision tree similar to that one proposed in
Figure 15, by working backwards all the possible scenarios.
The first approach may look simplistic, but it is often used in real situations where the size of the problem and other factors, related to decisions,
resources allocation or results requirements, do not allow a full evaluation
of the decision tree. We have chosen this approach in two works, Martinelli et al. (2011b) and Martinelli et al. (2012b), following similar works
like Eidsvik et al. (2008) and Bhattacharjya et al. (2010). It is important
to remark that even in this case the combinatoric dimension of the problem
drill1
drill 2
drill 3
....
drill 1
....
drill 2
o
well 3
well 3
28
well 2
well 2
oil
dry
il
dry
quit
quit
....
....
....
Figure 15: Decision tree for a simple 3-nodes discrete example with two
possible outcomes (oil or dry) per node.
makes it impossible to solve exactly when the number of prospects that we
are allowed to drill is high. A short discussion of the problem is presented
in Martinelli et al. (2011b).
The second approach is more complete and adequate for studying optimal exploration problems, but the drawback is that the classical formulation
in terms of Dynamic Programming (DP) equations usually breaks up when
the number of potential drilling locations is larger than 10. A good presentation of this approach can be found in Bickel and Smith (2006). The main
problem is given by the combinatorial growth of the number of scenarios:
a field with N possible locations and two possible
! outcomes (oil or dry)
PN
N
i!2i (n i + 1) scenarios.
means the necessity of considering i=0
i
This means a treatable 105 when N = 6, but an already unmanageable
1010 scenarios when N = 10. The problem can be simplified if we rec-
ognize the recombining structure of the tree: if we observe 1 to be dry and

3 wet, the future (conditional) probabilities and cash flows are the same
regardless if we drilled well 1 or 3 first. This a consequence of Bayesian
updating: posterior distributions do not depend on the order in which information is received. Even in this case, though, we get 103 scenarios when
N = 6 and 106 scenarios when N = 10.
This is the main reason that has pushed us to explore how we can solve
sub-optimally the problem when N > 10. The literature in optimal decision modeling applied to oil &gas investments includes a series of papers
considering the problem of option pricing, see Paddock et al. (1988) Smith
and McCardle (1999). Kokolis et al. (1999) and Wang et al. (2000), on the
other side, describe a problem of decision making under uncertainty for oil
29
& gas prospects with focus on the technical risks connected to a project.
They do not consider how to design an optimal sequential drilling strategy,
but discuss the combinatorial increase of the number of scenarios that has
to be considered. Some special cases of a generic optimal sequential problem are solved and presented in Smith and Thompson (2008). Bickel and
Smith (2006) and later Bickel et al. (2008) aim to solve the same problem
in a more general setting, where not all the prospects have positive intrinsic
values (i.e. not all prospects would be independently economically viable).
Our work presented in Martinelli et al. (2011a) ideally continues the same
line of research, showing a possible way to approximate the DP procedure
based on heuristics. In Martinelli et al. (2012a), on the other side, we show
a case study where it is possible to solve exactly the DP, given the small
of dimension of the problem. In a recent paper Brown and Smith (2012)
show a different way to approximate the same problems, and in particular
they focus on how to build lower and upper bounds to the DP value when
it is not possible to solve the equation. In Martinelli and Eidsvik (2012)
we use similar ideas, with more effort in defining ways to derive optimal sequences rather than bounds. Interestingly, both in Brown and Smith (2012)
and in Martinelli and Eidsvik (2012), the best cluster is selected using Gittins Indeces (GI) Gittins (1979), that were originally introduced for solving
bandit-problems Weber (1992). The same indeces were then used in the
context of oil exploration by Benkherouf and Bather (1988) and Benkerhouf
et al. (1992): GI provide the optimal sequential solution when considering
a set of independent prospects.
So far we have focused on the production phase, but it is important
to keep into account that the problem of establishing an optimal drilling
sequence is strictly related with the problem of how to optimize the information gathering (exploration phase). In a specular way to what we have
described above, we can distinguish two possible ways of gather the information: static gathering and dynamic gathering, where we can decide whether
to collect more information depending on the outcome of the nodes already
tested. In the decision analysis community the problem is known as Value of
Information (VoI) problem, and dates back to Miller (1975), Howard (1966)
and Raiffa (1969). In the oil & gas community the VoI is becoming a very
important tool in the early phases of the exploration, and a number of recent
papers suggested and propose different applications, see Cunningham and
30
Begg (2008) and Bratvold and Begg (2008).

We will now present the basic concepts and notation of the statistical
formulation of static and sequential decision problems with static and dynamic (perfect) information gathering, and we will discuss their implications
on a small example with two possible outcomes per prospect (oil or dry).
We will use the word test or explore in order to indicate the exploration or
information gathering action, and the word drill to indicate the action that
we will perform as a result of our decision problem.
4.1
Basic notation
We consider again a set of N prospects with a discrete set of possible outcomes. The random vector of all variables is x = (x1 , . . . , xN ), and its
distribution on a DAG p(x) is specified from the product of conditional distributions p(xi |xpa(i) ) , as shown in the previous section. We let i be the
observable in node i = 1, . . . , N . If node i is not yet observed, we set i = .
If we choose to observe node i, i is the actual outcome of the random vari-
able xi at this node. For instance, i = 1 can mean that well i has been explored and found dry, i = 2 if found oil in a simple two-outcomes example.
Initially, before acquiring any observables, we have = 0 = (, . . . , ).
For the likelihood of this scenario we need the marginal p(x2 = 2). This is
computed by summing out all scenarios that share the second component
equal to 2. In order to compute the conditional probabilities of a node i,
given evidence, we need p(xi = j|), j = 1, . . . , k, where the empty elements
() of are unobserved and marginalized out. We can associate to each
prospect i a revenue ri and a drilling cost ci that must be payed if we decide
to produce that prospect, no matter the outcome of the prospect itself. We
further associate to each prospect a certain exploration cost Ei , that must be
paid if we decide to explore the prospect, in order to get perfect information
about its state.
In order to decide which well is better to drill first, we must compare
the prior value (PV), i.e. the value of the field given the prior information
p(x), and the value of free clairvoyance (VFC) associated to a certain set
of wells j, i.e. the value of the field had we the possibility of knowing with
certainty (like clairvoyants!) the outcome of j. Note that j can be a whatever
subset of the original set of prospects, and in case of static collection we
have to be able to compare the value provided by all these sets. In case
31
of dynamic collection, since we are planning a sequential collection, we can

simply compute the value of the first best site, and than, given its outcome,
update the whole procedure.
4.2
Static decision problem
When dealing with static decision problems, we can not get any benefit from
the intrinsic sequentiality of the drilling procedure, and this is reflected both
in the prior value and in the value of free clairvoyance. Since we have to
decide at once all the drilling strategy, the best information that we can
get a priori is simply the sum of the positive Intrinsic Values (IV) of the
prospects, where IV (i) = p(xi = oil)ri ci . Therefore:
P Vst ( 0 ) =
N
X
max{IV (i), 0} =
i=1
N X
2
X
i=1 k=1
max{p(xi = k)rik ci , 0}
(1)
where rik = ri if k = 1 (oil) and rik = 0 if k = 2 (dry). In general, we can

compute the prior value given whatever prior configuration:
Vst () =
N X
2
X
i=1 k=1
max{p(xi = k|x )rik ci , 0},
where x = {xi = i i 6= {}}.
In case of static collection, the value of free clairvoyance needs to be
computed for every possible decision about the field. In a case with N
prospects and two possible decisions per prospect (test or no test), we have
2N possible combinations. For each of these combinations we have 2M possible evidences, where M is the number of sites for which the decision is test
or explore. The value of free clairvoyance given a binary decision vector j is
therefore:
V F Cst.st (j) =
2
X
p(xkj )Vst ( kj )
2
X
p(xkj )
kj =1
M
kj =1
N X
2
X
i=1 s=1
max{p(xi = s|xkj )ris ci , 0},
(2)
32
where p(xkj ) is the probability of the configuration kj computed on all

the sites in j for which we have chosen to test. As mentioned in the previous
section, this approach has been used in Martinelli et al. (2011b) for sets j
of dimension 1, 2 or 3, meaning that we were allowed to choose maximum 3
sites where to perform the exploration.
In Table 1 we show all the possible decision vectors j for a simple 3prospects case, and all the possible configurations kj for j = (1, 0, 1) (test A,
No test B and test C).
Decision
vector j
(0,0,0)
(1,0,0)
(0,1,0)
(1,1,0)
(0,0,1)
(1,0,1)
(0,1,1)
(1,1,1)
Prosp. A
Prosp. B
Prosp. C
No test
Test
No test
Test
No test
Test
No test
Test
No test
No test
Test
Test
No test
No test
No test
Test
No test
No test
No test
No test
Test
Test
Test
Test
Outcome
k(1,0,1)
1
2
3
4
Prosp. A
Prosp. C
Dry
Oil
Dry
Oil
Dry
Dry
Oil
Oil
Table 1: Decision vectors j for a simple 3-prospects case, and test outcomes
kj for the decision vector j = (1, 0, 1). In this case M = 2, therefore the
number of possible outcomes is 22 = 4.
In case of dynamic collection we are interested in indicating which out
of the original N prospects is the best to start the exploration campaign. In
this case we have the possibility of quitting the exploration campaign after i,
or to continue depending on the outcome of the exploration in i (see Miller
(1975) for details). The dynamic value of free clairvoyance associated to the
site i is therefore:
V F Cst.dyn (i, 0 ) =
2
X
ki =1

p(xi = ki ) max Vst ( ki i ), max{Vst.dyn (j, ki i ) Ej } ,
j6=i
(3)
where ki i = { i , i = ki } and:
Vst.dyn (j, ki i ) =
2
X
kj =1

k ,k
k ,k
p(j = kj |i = ki ) max Vst ( i,ji j ), max{Vst.dyn (l, i,ji j ) El } .
l6=i,j
This means that after a generic test in prospect i we can either decide to
stop our exploration, and therefore collect just the value Vst ( ki i ) depending
on the outcome ki of prospect i, or decide to perform other explorations
and iterate the procedure. In the latter case, the successive test will be
33
performed in the site j that maximizes the new information, keeping into
account its testing cost Ej .
4.3
Dynamic decision problem
In a dynamic decision problem we exploit the sequentiality of any real

drilling campaign, by letting our strategy depend on the outcome of the
previously drilled prospects. The prior value, i.e. the value before any exploration is performed, shall also reflect this possibility. This means that
even without any information we are able to rank the prospects such that the
most lucrative/informative come first in our sequence. The balance between
information and monetary value is governed by a discounting parameter
chosen between 0 and 1.
The dynamic prior value assumes now the following expression:
P Vdyn ( 0 ) =
where:
Vdyn ( ki i )
max
2
X
i{1,...,N }
ki =1
= max
j6=i
2
X
kj =1
p(xi = ki )
riki
p(xj = kj |xi = ki )
k
rj j
Vdyn ( ki i )
ci , 0 , (4)
k ,k
Vdyn ( i,ji j )
cj , 0
We can see that in the dynamic case we need a DP formulation also for
the prior value. This happens because we need to keep into account that
even without any previous information, we would be able to conduct our
drilling campaign in best possible way, i.e. by updating our future sequence
with the results coming from the already drilled prospects.
As we have done with the static case, we have two possibilities of collecting the information before we start producing the wells, a static collection
or a dynamic collection. If we chosen the static collection it means that
we have to decide in advance the full exploration schedule, while we can always change and update our strategy in the production phase. The value of
free clairvoyance can be now written, symmetrically to what we have done
before, as:
34
V F Cdyn.st (j) =
2
X
p(xkj )Vdyn ( kj )
kj =1
M
l=
2
X
p(xkj )
kj =1
max
(5)
2
X
i{1,...,N },
ki =1
p(xi = ki |xkj )
riki
Vdyn ( kjk,ii )
ci , 0
The summation is out of all the possible 2M combinations of places where
we can explore, since once a combination has been chosen, it is not possible
to change it any longer.
In a rather similar way we can derive the expression for the VFC in the
case where both exploration and production are allowed in a dynamic way:
V F Cdyn.dyn (i, 0 ) =
2
X
ki =1

p(xi = ki ) max Vdyn ( ki i ), max{V F Cdyn.dyn (j, ki i ) Ej } ,
j6=i
(6)
where:
V F Cdyn.dyn (j, ki i ) =
2
X
kj =1

k ,k
k k
p(xj = kj |xi = ki ) max Vdyn ( i,ji j ), max{V F Cdyn.dyn (l, i,ji j ) El } .
l6=i,j
As we can see, in case we decide to stop the exploration, we have to keep into
account that we can always perform a sequential production, and therefore
we have to use the dynamic prior value. In case we decide to continue, we
have the usual maximization within the integration of the possible evidences,
to represent the idea that we can perform the best choice after collecting
every possible evidence.
It is worth noticing that a criterion very similar to this last one was
developed in Bickel and Smith (2006), and later used in many of our works,
especially in Martinelli et al. (2011a) and in Martinelli and Eidsvik (2012)
for optimal sequential exploration purposes. In this case we do not consider
the exploration phase as separated from the drilling phase, since we are
receiving perfect information and we assume that if we explore and find a
positive outcome we can immediately produce that well. Details on the DP
can be found in the cited papers.
4.4
A small example
We consider a small example with 2 prospects and two possible states per
prospect, oil or dry. We are interested in optimizing the exploration strategy,
35
in the two cases of static and dynamic decision problems. The two prospects
A and B are connected through a common parent, as shown in Figure 16.
If either A or B is observed oil, then the top common node has to be oil. If
either of the two is observed dry, it can be a local failure or a dry common
parent.
CP
Figure 16: Small example with two prospects A and B correlated through a
common parent CP. The example is used in Section 4.4
Let us consider first the static decision problem. We can intuitively
understand that, if the marginal and conditional probability are fixed, and
if the expected revenues and drilling costs are given, our optimal exploration
strategy will depend strongly on the exploration costs assigned to the two
prospects. If we are in a static collection setting we have to compare four
different quantities:
V oIst.
V F Cst.st (no test A, no test B) P Vst = 0
V oIst.A (EA ) EA
V F Cst.st (test A, no test B) P Vst EA
V oIst.B (EB ) EB
V F Cst.st (no test A, test B) P Vst EB
V oIst.AB (EA , EB ) EA EB
V F Cst.st (test A, test B) P Vst EA EB
In this case we use equations 1 and 2 for computing respectively P Vst

and V F Cst.st (). If we allow the possibility of sequential collection we have
to consider two more quantities:
V oIdyn.A (EA , EB ) EA
V F Cst.dyn (test A f irst) P Vst
V oIdyn.B (EA , EB ) EB
V F Cst.dyn (test B f irst) P Vst ,
36
computed using equation 3.

The results can be visualized on a plane with EA on the x-axis and
EB on the y-axis (see Figure 17 on the left). In dark colors we find the
decision regions where it is indifferent whether to perform static or dynamic
collection, in light colors regions where it is optimal to perform dynamic
collection, since our second decision will depend on the outcome of the first
exploration well. Finally, in black we find a region where the exploration
costs for both A and B are too high, and therefore the solution is not to drill
any exploratory well before the production campaign; and in cyan/white we
find a region where both the exploratory wells are cheap, and therefore the
optimal solution is to collect perfect information from both prospects, A and
B.
In Figure 17, on the right, we compare the same quantities for the dynamic decision problem, namely:
V oIst.
V F Cdyn.st (no test A, no test B) P Vdyn = 0
V oIst.A (EA ) EA
V F Cdyn.st (test A, no testB) P Vdyn
V oIst.B (EB ) EB
V F Cdyn.st (no testA, test B) P Vdyn
V oIst.AB (EA , EB ) EA EB
V F Cdyn.st (test A, test B) P Vdyn
V oIdyn.A (EA , EB ) EA
V F Cdyn.dyn (test A f irst) P Vdyn
V oIdyn.B (EA , EB ) EB
V F Cdyn.dyn (test B f irst) P Vdyn
In this case we use equations 4 and 5 for computing respectively P Vdy

and V F Cdyn.st (), and equation 6 for computing V F Cdyn.dyn ().
The color codes used for the figure on the right are the same as for the
figure on the left. It is important to notice that in this case the regions of
EA and EB for which the exploration is suggested are much smaller than
in the Static Decision Problem. This reflects the different prior value that
we have used for the Dynamic decision problem, that is greater than the
static prior value. Ultimately, this means that it is less convenient to drill
preliminary exploratory wells when we already know that we will have the
possibility to update the drilling strategy on the way, while if our decision
is static we need to have the best possible information before delivering the
drilling plan.

Decision regions, Dynamic decision problem
20
20
18
18
EB (Exploration cost, prospect B)
EB (Exploration cost, prospect B)
Decision regions, Static decision problem
16
14
12
10
8
6
4
2
0
37
16
14
12
10
8
6
4
2
5
10
15
EA (Exploration cost, prospect A)
20
5
10
15
EA (Exploration cost, prospect A)
20
Figure 17: Decision regions for the problem presented in Section 4.4. On
the left, decision regions for the static decision problem, on the right for the
dynamic one. In red, region where it is optimal to test just prospect A, in
blue just prospect B, in black where it is optimal to quit the exploration from
the beginning, in cyan where it is optimal to test both (static exploration).
Finally, in light red and right blue, regions where it is optimal to start testing
(dynamic exploration) from prospect A and from prospect B, respectively.
While in Figure 17, we plot just the V oI of the prospect whose value
is maximum among the considered ones, it is interesting also to compare
all the six quantities taken into account in a single graph (Figure 18). We
can notice that the purple and the cyan planes, that correspond to the VoI
associated to dynamic exploration, coincide with the green plane (static
exploration in both A and B) for small values of EA and EB , and with the
red and blue planes for high values of respectively EA and EB . For medium
values it exist a region where lie the real benefits of having the possibility
of performing sequential exploration in stead of static. We can furthermore
notice that, while the blue and the red planes depend just by the values
assumed by EA and EB respectively, the purple and the cyan planes depend
both on EA and on EB .
In this Section we have presented the main methods for computing the
Value of Information in a static and in a dynamic setting, using standard
DP procedures. This topic has been a central focus throughout the whole
thesis: VoI computations can be found both in Martinelli et al. (2011b) and
in Martinelli et al. (2012b), while optimal sequential exploration strategies
(with approximations) are discussed broadly in Martinelli et al. (2011a),
Martinelli et al. (2012a) and Martinelli and Eidsvik (2012).
OUTLINE OF THE THESIS
38
Figure 18: Decision planes for the problem presented in Section 4.4. On
the left, decision planes for the static decision problem, on the right for the
dynamic one. In red, plane corresponding to V oIst.A EA (test just A), in
blue V oIst.B EB (test just B), in green V oIst.A+B EA EB (test both A
and B) and in green V oIst.none (test none). Finally, in purple plane for the
dynamic decision of starting in A V oIdy.A EA , and in cyan plane for the
dynamic decision of starting in B: V oIdy.B EB .
Outline of the thesis
The thesis develops in five parts, corresponding to five pieces of works that
have been published or submitted for publication by myself and my coauthors in the last three years, in relation to this thesis. I will now go briefly
through these works, in order to introduce them and to show the logical and
chronological ratio that is behind the entire work.
Part I. This first paper, referred to as Martinelli et al. (2011b), aims
to illustrate the preliminary ideas and results that we have developed
when we first approached the problem of the thesis. The work is

intended to introduce BN as a modelization tool for a geologists and
geophycists audience. Much effort is therefore placed in explaining
and discussing the genesis of the network, and the mathematical and
logical meanings of nodes and edges in a graphical model. In the final
part of the work we propose possible uses of the model that we have
developed, borrowing ideas from the VoI community and introducing a
new criterion, the Abandoned Revenue (AR) criterion, for determining
when it is optimal to quit the exploration, given that we have explored
a sequence of dry wells. The motivation for this work and for this
specific criterion comes from a real problem of a major norwegian
39
oil & gas company, Statoil. Given that the majority of the big oil
fields has already been discovered, this company argues that it would
have be more beneficial to have a criterion that prevents from big
losses in frontier areas rather than a generic criterion like the VoI
that weights in the same way revenues and losses. In this case our
contribution has been both in the development of the BN model, and
in the definition of the optimization criteria for the exploration. The
work was first presented (oral) at the AAPG (American Association of
Petroleum Geologists) Conference and Exhibition held in New Orleans
in April 2010, and then accepted and published in the AAPG Bulletin
in August 2011.
Part II. The second paper, referred to as Martinelli et al. (2011a),
takes the lead from a more detailed analysis about the possibility of
building an optimal sequential drilling sequence on the network introduced in the previous work. In particular, we have recognized the
impossibility of running a full DP evaluation of the decision tree, and
we have therefore tried to develop approximated strategies that could
lead to sub-optimal solutions, and to assess the quality of these approximations. The work is intended to show a methodology that could have
a wide range of applications for decision makers dealing with graphical
models. For this reason we have tested and compared the strategies
both to BN and to Markov Random Fields (MRF). The results show
that the suggested strategies clearly improve the simpler intuitive constructions (naive and myopic approaches), and they prove to be useful
when when selecting exploration policies. The work has been presented (oral) part at the First Spatial Statistics Conference held in
Enschede, the Netherlands, in March 2011, and part at the IAMG
(International Association of Mathematical Geosciences) Conference,
held in Salzburg in September 2011. The work has been submitted
for publication in the European Journal of Operational Research in
February 2012 and it is currently in the review process.
Part III. This third paper, referred to as Martinelli et al. (2012a),
shows the first application to a real case study of the methodology

proposed in the first paper. When dealing with a real case study,
we have had to face a number of problems that were not considered
40
in the original work. First, we had to introduce a modeling scheme

for every component of the petroleum system, and not just for the
source rock: this has lead us to the necessity of translating into the
model different kinds of geological arguments, discussed in Section 3.2.
Second, we needed to consider the possibility of dealing with a larger
sample space, i.e. with outcomes more complex than the traditional
binary model (oil or dry) presented for the sake of simplicity in the
first two works. In the work we propose a complete model and we
discuss the implications of our choices when we develop a sequential
exploration strategy. The work has been inspired by a real case study
from Statoil, who is currently in the process of defining the exploration
strategy in the field. The initial decisions confirm the results presented
in the paper. The work has been accepted for presentation (oral) at
the SPE (Society of Petroleum Engineers) Conference and Exhibition,
to be held in San Antonio, USA, in October 2012, and submitted for
publication in the SPE Journal in June 2012.
Part IV. This fourth paper, referred to as Martinelli et al. (2012b),
aims to give a more sound foundation to the modeling BN based techniques proposed in the previous works. The idea is to use BPSM as
a driving tool for building the BN model: the choice has a geological
foundation, since in BPSM the different components of the petroleum
system are clearly identifiable, and therefore they are easier to be
translated into a decisional framework like ours, that wants to maintain
and to exploit this diversification. The training process is guided by
the construction of multiple scenarios that keeps into account the uncertainty in four key parameters of the basin. The outcome is a series
of accumulation distribution for the different prospects, whose correlation is learned directly from the different scenarios. We conclude the
work by demonstrating important decision making applications such
as evidence propagation and again the VoI. In the current industrial
framework the prospect risk assessment is almost completely separated from basin modeling, and the decision makers, usually engineers
or economists, need to estimate the exploration risk based on subjective judgements of geologists and geophysicists. We believe therefore
that this work, though being little more than a proof of concept, could
have an impact in the way the decisional process is performed in in-
41
dustry. The work has been presented (oral) at the 9th Geostatistics
Conference, held in Oslo in June 2012, and submitted for publication
in Petroleum Geoscience in July 2012.
Part V. The fifth paper, referred to as Martinelli and Eidsvik (2012),
aims to solve the same problem developed in Part II, but with a dif-
ferent technique, that could possibly exploit the peculiar structure of

the graphical models used in the problems under consideration. The
idea, borrowed from a recent work of Brown and Smith (2012), is to
split the original network or model in small sub-networks, to solve the
DP exactly within these subnetworks, and to combine the results. The
goal is to develop sequential drilling strategies that could guide the decision makers; a byproduct of the work is to provide lower and upper
bounds that could prove the goodness of the chosen approximation.
The work is intended to be useful not just for the oil & gas practitioners, but for a larger community of decision makers dealing with
graphical models. For this reason we have sided our original utility
function based on revenues and costs with more generic utility functions based on entropy, that appear commonly in environmental and
biological problems. The work has been partially presented in a seminar at NTNU in March 2012, and has been submitted for publication
in the Journal of the Royal Statistical Society, Series C, in September
2012.
REFERENCES
42
References
Allen, P. and Allen, J. (2005). Basin Analysis, Principles and Applications.
2th ed. Blackwell Publishings.
Benkerhouf, L., Glazebrook, K. and Owen, R. (1992). Gittins indexes and
oil exploration. Journal of the Royal Statistical Society. Series B 54,
229241.
Benkherouf, L. and Bather, J. (1988). Oil Exploration: Sequential Decisions
in the Face of Uncertainty. Journal of Applied Probability 25, 529543.
Bhattacharjya, D., Eidsvik, J. and Mukerji, T. (2010). The Value of Information in Spatial Decision Making. Mathematical Geosciences 42,
141163.
Bickel, J. and Smith, J. (2006). Optimal Sequential Exploration: A Binary
Learning Model. Decision Analysis 3, 1632.
Bickel, J., Smith, J. and Meyer, J. (2008). Modeling Dependence Among
Geologic Risks in Sequential Exploration Decisions. SPE Reservoir Evaluation & Engineering 11, 352361.
Borsuk, M. E., Stow, C. A. and Reckhow, K. H. (2004). A Bayesian network of eutrophication models for synthesis, prediction, and uncertainty
analysis. Ecological Modelling 173, 219239.
Bratvold, R. B. and Begg, S. H. (2008). I would rather be vaguely right than
precisely wrong: A new approach to decision making in the petroleum
exploration and production industry. AAPG Bulletin 92, 13731392.
Brown, D. and Smith, J. (2012). Optimal Sequential Exploration: Bandits,
Clairvoyants, and Wildcats. submitted .
Carter, P. and Morales, E. (1991). Probabilistic addition of gas reserves
within a major gas project, SPE paper 50113. 2007 SPE Asia Pacific Oil
and Gas Conference and Exhibition 1, 367374.
Cooper, G. F. (1990). The computational complexity of probabilistic inference using Bayesian belief networks. Artificial Intelligence 42, 393405.
REFERENCES
43
Corradi, A., Ponti, D. and Ruffo, P. (2003). A methodology for prospect

evaluation by probabilistic basin modeling. Multidimensional basin modeling, AAPG/Datapages Discovery Series, S. Duppenbecker and R. Marzi
eds. 7, 283293.
Cowell, R., Dawid, P., Lauritzen, S. and Spiegelhalter, D. (2007). Probabilistic Networks and Expert Systems. Springer series in Information
Science and Statistics.
Cunningham, P. and Begg, S. H. (2008). Using the value of information
to determine optimal well order in a sequential drilling program. AAPG
Bullettin 92, 13931402.
Eidsvik, J., Bhattacharjya, D. and Mukerji, T. (2008). Value of information
of seismic amplitude and CSEM resistivity. Geophysics 73, R59R69.
Geiger, D. and Pearl, J. (1993). Logical and algorithmic properties of conditional independence and graphical models. Annals of statistics 21,
20012021.
Gittins, J. (1979). Bandit processes and dynamic allocation indices. Journal
of the Royal Statistical Society, Series B 41, 148177.
Gluyas, J. and Swarbrick, R. (2003).
Petroleum Geoscience.
Wiley-
Blackwell.
Hantschel, T. and Kauerauf, A. I. (2009). Fundamentals of Basin and
Petroleum Systems Modeling. Springer.
Howard, R. (1966). Information Value Theory. IEEE Transactions on Systems Science and Cybernetics SSC-2.
Jia, B. (2010). Linking geostatistics with basin and petroleum system modeling - Assessment of spatial uncertainties. M.Sc thesis, Stanford University.
Kaufman, G. M. and Lee, P. J. (1992). Are wildcat well outcomes dependent or independent? Working papers 3373-92 Massachusetts Institute of
Technology (MIT), Sloan School of Management.
Kokolis, G., Litvak, B., Rapp, W. and Wang, B. (1999). Scenario selection
for valuation of multiple prospect opportunities: a MonteCarlo simula-
REFERENCES
44
tion approach. SPE paper 52977, presented at the SPE Hydrocarbon

Economics and Evaluation Symposium, Dallas, TX, 20-23 March 1999 .
Lauritzen, S. L. and Spiegelhalter, D. J. (1988). Local Computations with
Probabilities on Graphical Structures and Their Application to Expert
Systems. Journal of the Royal Statistical Society, Series B 50, 157224.
Magoon, L. B. and Dow, W. G. (1994). Petroleum System: From Source to
Trap. AAPG memoir 60.
Martinelli, G. and Eidsvik, J. (2012). Dynamic exploration designs for
graphical models using clustering. submitted for publication in the Journal of the Royal Statistical Society, Series C .
Martinelli, G., Eidsvik, J. and Hauge, R. (2011a). Dynamic Decision Making
for Graphical Models Applied to Oil Exploration. Mathematics Department, NTNU, Technical Report in Statistics 12.
Martinelli, G., Eidsvik, J., Hauge, R. and Drange-Forland, M. (2011b).
Bayesian Networks for Prospect Analysis in the North Sea. AAPG Bulletin 95, 14231442.
Martinelli, G., Eidsvik, J., Hauge, R. and Hokstad, K. (2012a). Strategies
for petroleum exploration based on Bayesian Networks: a case study, SPE
Paper 159722. Accepted for presentation at the SPE Annual Technical
Conference and Exhibition to be held in San Antonio, TX, USA, 8-10
October 2012, submitted for publication to the SPE Journal .
Martinelli, G., Eidsvik, J., Rekstad, S., Sinding-Larsen, R. and Mukerji,
T. (2012b). Building Bayesian networks from basin modeling scenarios
for improved geological decision making. submitted for publication in
Petroleum Geoscience .
Meisner, J. and Demirmen, X. (1981). The creaming method: A Bayesian
procedure to forecast future oil and gas discoveries in mature exploration
provinces. Journal of Royal Astronomical Society A 144, 131.
Miller, A. (1975). The Value of Sequential Information. Management Science
22, 111.
REFERENCES
45
Murphy, K. P. (1999). A Variational Approximation for Bayesian Networks

with Discrete and Continuous Latent Variables. Proceedings of the Fifteenth conference on Uncertainty in artificial intelligence, UAI 99 .
Murtha, J. (1996). Estimating Reserves and Success for a Prospect With
Geologically Dependent Layers. SPE Reservoir Engineering 11, 3742.
Paddock, J. L., Siegel, D. R. and Smith., J. L. (1988). Option valuation
of claims on real assets: The case of offshore petroleum leases. Quart. J.
Econom. 103, 479508.
Pearl, J. (1984). Heuristics: Intelligent Search Strategies for Computer
Problem Solving. Addison-Wesley.
Pollino, C. A., Woodberry, O., Nicholson, A., Korb, K. and Hart, B. T.
(2007). Parameterisation and evaluation of a Bayesian network for use in
an ecological risk assessment. Environmental Modelling and Software 22,
11401152.
Raiffa, H. (1969). Decision Analysis. Addison-Wesley.
Rasheva, S. and Bratvold, R. B. (2011). A new and improved approach for
geological dependency evaluation for multiple-prospect exploration, SPE
paper 147062. Accepted for presentation at the SPE Annual Technical
Conference and Exhibition in Denver, Colorado, USA, 30 October - 2
November 2011 .
Rose, P. (2001). Risk analysis and management of petroleum exploration
ventures. AAPG methods in Exploration Series 12.
Shenoy, P. P. and Shafer, G. (1990).
Axioms for Probability and
Belief-Function Propagation. Uncertainty in Artificial Intelligence, R.D.

Shachter, T. Levitt, J.F. Lemmer and L.N. Kanal (eds.) 4, 169198.
Smalley, P. C., Begg, S. H., Naylor, M., Johnsen, S. and Godi, A. (2008).
Handling risk and uncertainty in petroleum exploration and asset management: An overview. AAPG Bulletin 92, 12511261.
Smith, J. . and Thompson, R. (2008). Managing a Portfolio of Real Options:
Sequential Exploration of Dependent Prospects. The Energy Journal International Association for Energy Economics 29, 4362.
REFERENCES
46
Smith, J. E. and McCardle, K. F. (1999). Options in the real world: Lessons

learned in evaluating oil and gas investments. Operational Research 47,
115.
Stabell, C. (2000). Alternative Approaches to Modeling Risks in Prospects
with Dependent Layers, SPE paper 63204-MS. SPE Annual Technical
Conference and Exhibition, 1-4 October 2000, Dallas, Texas .
VanWees, J., Mijnlieff, H., Lutgert, J., Breunese, J., Bos, C., Rosenkranz,
P. and Neele, F. (2008). A Bayesian belief network approach for assessing
the impact of exploration prospect interdependency: An application to
predict gas discoveries in the Netherlands. AAPG Bulletin 92, 1315
1336.
Wang, B., Kokolis, G. P., Rapp, W. J. and Litvak, B. L. (2000). Dependent
Risk Calculations in Multiple-Prospect Exploration Evaluations. SPE
paper 63198, presented at the 2000 SPE Annual Technical Conference
and Exhibition held in Dallas, Texas, 1-4 October 2000 .
Weber, R. (1992). On the Gittins Index for Multiarmed Bandits. The Annals
of Applied Probability 2, 10241033.
Xu, J. and Sinding-Larsen, R. (2005). How to choose priors for Bayesian
estimation of the discovery process model. Natural Resources Research
14, 211233.
Zwach, C. and Carruthers, D. (1998). Honoring uncertainties in the modelling of migration volumes and trajectories. IFE Symposium on Advances
in Understanding and Modelling Hydrocarbon Migration, Oslo, Norway,
December 7-8, 1998 .
Paper I
Bayesian networks for prospect analysis in the North Sea
G. Martinelli, J. Eidsvik, R. Hauge and M. Drange Frland
AAPG Bulletin 95(8):1423 -1442, 2011.
Bayesian networks for prospect

analysis in the North Sea
Gabriele Martinelli, Jo Eidsvik, Ragnar Hauge, and
Maren Drange Frland
ABSTRACT
We propose a flexible framework for evaluating prospect dependencies in oil and gas exploration and for solving decisionmaking problems in this context. The model uses a Bayesian
network (BN) for encoding the dependencies in a geologic system at source, reservoir, and trap levels. We discuss different
evaluation criteria that allow us to formulate specific decision
problems and solve these within the BN framework. The BN
model offers a realistic graphic model for capturing the underlying causal geologic process and allows fast statistical computations of marginal and conditional probabilities.
We illustrate the use of our BN model by considering two
situations. In the first situation, we wish to gain information
about an area where hydrocarbons have been discovered, and
use the value of perfect information to determine which locations are the best to drill. In the second situation, we consider the problem of abandoning an area when only dry wells
are drilled. For this latter, we use an abandoned revenue criterion to determine the drilling locations.
The application is from the North Sea. Our main focus is
the description, visualization, and interpretation of the results
for relating the statistical modeling to the local understanding
of the geology.
AUTHORS
Gabriele Martinelli Norwegian University of
Science and Technology, Norway;
gabriele.martinelli@math.ntnu.no
Gabriele Martinelli holds a B.Sc. degree (2006)

and an M.Sc. degree (2008) in mathematical
engineering from Politecnico di Milano, Milan,
Italy. He is currently a graduate student at the
Norwegian University of Science and Technology.
Jo Eidsvik Norwegian University of Science and
Technology, Norway; joeid@math.ntnu.no
Jo Eidsvik got his M.Sc. degree (1997) from the

University of Oslo and his Ph.D. (2003) from
the Norwegian University of Science and Technology (NTNU), Norway. He worked for the
Norwegian Defense Research Establishment and
for Statoil. His current position is associate
professor of statistics at NTNU.
Ragnar Hauge Norwegian Computing Center,
Norway; ragnar.hauge@nr.no
Ragnar Hauge has a masters degree in statistics

from the Norwegian University of Science and
Technology and a Ph.D. in statistics from the University of Oslo. He is currently an assistant research director at the Norwegian Computing Center, where he has been working since 1995.
The main focus of his work has been stochastic
modeling of facies and stochastic seismic inversion.
Maren Drange Frland Norwegian Computing
Center, Norway; maren.drange.forland@nr.no
Maren Drange Frland received an M.Sc.

degree in statistics from the Norwegian
University of Science and Technology in 2008.
She is currently a research scientist at the
Norwegian Computing Center.
INTRODUCTION
ACKNOWLEDGEMENTS
In recent years, there has been great interest in the evaluation

of prospect dependencies and how these dependencies can
affect the analysis of a geologically appealing area. It has been
pointed out (e.g., Van Wees et al., 2008 and Cunningham and
Copyright 2011. The American Association of Petroleum Geologists. All rights reserved.
Manuscript received June 28, 2010; provisional acceptance August 27, 2010; revised manuscript received
October 12, 2010; revised provisional acceptance October 19, 2010; 2nd revised manuscript received
November 15, 2010; final acceptance January 3, 2011.
DOI:10.1306/01031110110
AAPG Bulletin, v. 95, no. 8 (August 2011), pp. 14231442
1423
We thank two anonymous referees and the associate editor for useful comments and suggestions.
We thank Statoil and, in particular, Knut Birger Hjelle
and Erling Siring for providing us the expertise
and the data necessary to perform this study. We
thank the Statistics for Innovation (SFI2) Research
Center in Oslo that partially financed Gabriele
Martinellis scholarship through the FindOil project.
This work has been performed using the Bayes
Net Toolbox (Murphy, 2007) for Matlab and the
network package of R.
The AAPG Editor thanks the two anonymous reviewers for their work on this paper.
Begg, 2008), that the introduction of a prospectinterdependency perspective may result in a substantial difference when discussing a sequential drilling program or an expected portfolio evaluation.
The prospect dependencies entail information
about one prospect that leads to information about
other prospects. This is useful in decision making, if
one plans to drill the best two prospects. The best
two prospects might not be the top two ranked
prospects a priori because the dependence can cause
two a priori quite independent prospects to carry
more information in total.
All dependencies between prospects come
from geologic processes. A general model of dependency should consider all the geologic elements
that are needed for oil and gas to accumulate in
sufficient quantities to be worth producing. These
elements include an organic-rich source rock to
generate the oil or gas, a porous reservoir rock to
store the hydrocarbons (HCs), and a trap to prevent
the oil and gas from leaking. The source, reservoir,
and trap at one prospect might share some features
with those of other prospects. One discriminant
factor could be the distance (prospects nearby are
more likely to be interdependent), but commonly,
the geologic constraints and conditions that have
developed through time are more representative. A
complex model of interdependencies can therefore
be developed, as assessed in Bickel et al. (2008). In
this work, we focus our attention on the source
dependencies, which is the most interesting element for our case study. The procedure is valid for
all three factors, and they could be joined at the
prospect level.
Within each prospect, there may be several
segments, which are the potential reservoirs that
can be chosen for drilling. Segments belonging to
the same prospect can be distinguished by the depth
(i.e., different formations), by the relative position
with respect to some geologic element (i.e., on
either side of a major fault), or by some other local
geographic distinction.
A priori geologic information can be encoded
in a qualitative part and a quantitative part. Here,
the qualitative part contains a list and map of prospects, segments, and sources (kitchens), that is, areas
where the HC formation developed. At sources1424
segments-prospects, we also know the possible outcomes (oil, gas, dry). It further includes a road map
of the area, indicating the migration paths of HC
between sources-segments-prospects. The quantitative part is described by probabilities, where the
expert belief is translated to specific numbers. This
is assigned by conditional probabilities, stating the
likelihood of an outcome at one node, depending
on the outcomes of its parents.
Bayesian networks (BNs) (Pearl, 1986; Jensen,
1996) are attractive models for encoding qualitative and quantitative information. Van Wees et al.
(2008) presented a BN model for prospect dependencies, where the dependence is essentially modeled as a function of the distance. In our process, the
involvement of local geology experts in the networkbuilding phase allowed us to overcome this simplification and to generate a more reliable model.
The BN model provides a flexible dependency
structure between the prospects that is not only
based on the spatial distance, but instead includes a
complex geologic modeling. A good understanding
of the geologic mechanisms that govern HC migration is a prerequisite to building an effective BN
model. It is well known that the integration between simulation results or observed data with experts opinions can sensibly improve the quality and
the predictive ability of the network. For instance,
biological applications in Imoto et al. (2003) successfully integrate a priori expert biological knowledge with the microarray information. Similarly, an
ecological case study presented in Hamilton et al.
(2005) uses expert knowledge to show how the
development of a bacterium and its interaction
with the environment can be analyzed. The expert knowledge can be used both in the definition
of the possible dependencies and in the quantification of the conditional probability tables (CPTs).
The Bayesian (Belief) network becomes an effective framework for integrating various contributions.
We present a BN model based on a set of
prospects in the Norwegian part of the North Sea.
The general purpose is to build a reliable and coherent network for source, reservoir, and trap. Our
main focus is on the source network, which is the
most complex model for the geologic situation. The
idea is to use BNs to address decision questions in
Bayesian Networks for Prospect Analysis in the North Sea
Figure 1. Bayesian network. The K nodes

represent kitchens, the P nodes represent
the prospects, whereas the bottom nodes of
the network, marked with numbers and
letters, are the segments.
the context of complex geologic interdependencies. In particular, we demonstrate the applicability

of a BN model for what-if scenarios and static exploration strategies. By what-if scenarios, we imagine knowing the outcome at one or more prospects
and study how this information propagates to the
other prospects. For the exploration strategy, we
compute which fixed number of segments are best to
drill to maximize some utility criterion, see for example, Smalley et al. (2008), Cunningham and Begg
(2008), Eidsvik et al. (2008), and Bhattacharjya
et al. (2010). Note that we do not attempt to find
the optimal exploration sequence, instead, we use
various criteria to study the impact of different kinds
of evidence. Finding the optimal exploration sequence (Bickel and Smith, 2006) is a huge combinatorial challenge in our model with 25 prospects.
The first proposed criterion is the value of
perfect information (VoPI), which assumes that
information only has value if it can change our decisions. In our case, the decision is which segments
to drill. The second criterion we propose is the
abandoned revenue (AR), which is a modification
of the VoPI. The AR is intended for areas with no
proven successes, and the challenge is to minimize
the possible loss by leaving the area under dry
evidence.
We explain the BN model obtained by geologic

constraints in Building the Model. In Exploratory
Analysis, we present what-if examples to provide
insight to the model. In Exploration Strategies, we
propose and discuss VoPI and AR for selection of
segments in exploration drilling. A brief discussion
of the benefits and limitations of our approach
and of possible improvements is given in Closing
Remarks.
BUILDING THE MODEL

In this section, we build a model with prospect
dependencies. We first describe BNs, then build a
network model for the source dependencies.
A BN (Jensen, 1996) consists of a set of nodes
and directed edges and associated conditional probability distributions. A BN model is a flexible tool
for establishing causal relationships between the
nodes, here representing geologically defined prospects. The BNs are graphs G fV ; Eg, where V is
the set of nodes and E is a subset of the set V V
of ordered pairs of vertices, called the edges or
links of G. The BN we use for the source variable
in our North Sea case study is presented in Figure 1.
The circles are nodes, whereas the arrows are edges.
Martinelli et al.
1425
The physical meaning of these will be discussed in

Bayesian Network: Qualitative Part.
We divide the model presentation into two
parts: qualitative and quantitative. This duality is
fictitious, but it reflects the way we built the network itself, which led us to design the network first
and then to design a method to define the CPTs.
In particular, the qualitative part describes which
geologic elements enter the model and how they are
connected. The quantitative part states the probability model for all nodes, which denote the strength
of geologic expert belief attributed to the different
nodes and edges of the network.
Bayesian Network: Qualitative Part
We focus on the source part for two reasons. First,
the reservoir quality is excellent in a large part of
the field, as studies have shown, see for example,
Ramm and Ryseth (1996). Second, the source network is rich from a BN point of view. It contains
some interesting prospect dependencies, whereas
the currently used network for the trap applies to
the local segment level.
We use a discrete model with three possible
states for each node (dry, gas, or oil), representing
the actual state of the source in that node. Our
network includes three possible kinds of nodes:
Kitchens. We define the kitchens as areas where
source rock has reached appropriate conditions
of pressure and temperature to generate HCs. In
our model, the kitchens are nodes that with
probability 1 assume either the state oil or the
state gas. Kitchens are denoted by K in Figure 1.
Prospects. We define prospects as structures
that may contain HCs that have been fully evaluated and are ready to drill; the prospects are the
key node of the network because, through them,
we can define the spatial relationships that we
need. Prospects are denoted by P in Figure 1.
Segments. The segments are the bottom nodes
of our network, and they correspond to potential
drilling sites associated with a certain prospect.
For each prospect, we can have one or more
segments that share part of the infrastructure
needed for drilling and production. In Figure 1,
1426
segments inherit the number of their prospect,

with alphabetical listing.
The basic point of the analysis is a list of prospects, where great interest still exists in HC discoveries. For each segment, and therefore for each
prospect, the following information is available
given by a team of experts in local geology: marginal probability of HC source; conditional probability of gas given the presence of HC; list of
possible paths (edges) for the gas and the oil migration between different kitchens, prospects, and
segments; expected volume, both for gas and for
oil; expected monetary value of oil and gas; and
expected operational costs.
The three first points are needed for building
the network, whereas the three last points concern
the decision-making process. Other information is
also available, such as age and formation. These were
helpful when constructing edges and probabilities
for the BN, but are not discussed further here.
The edges represent expert geologists assumptions of the prospect, segment, and kitchen
connections. Within the BN formulations, the edges
define a dependence structure for the nodes. For
instance, in Figure 1, the prospect P4 has just one
directed edge from the kitchen, K2, because a
strong prior belief indicates a unique kitchen for
this prospect. However, prospect P1 has directed
edges from two kitchens, K1 and K3. The experts
in local geology assume that the source can flow
from any of these kitchens to provide HCs in
prospect P1.
A crucial feature of our BN is that segments
never are parents, all connections are on higher
levels. This allows for a local failure, where HC
may have reached the prospect area, but not the
actual segment we look at.
We end up with a network of 42 connected
nodes, where 4 are kitchens, 13 are prospects, and
25 are segments (Figure 1). Each node, vi, has three
possible states: dry, gas, or oil, that is,
vi s;
i 2 f1; . . . ; 42g;
s 2 fdry; gas; oilg
We are particularly interested in the segment

nodes, where we can drill an exploratory well to
collect an evidence of dry, gas, or oil. The evidence at one segment will affect the probability of
outcomes at other segments, via the BN model.
We will investigate this further in the computation of exploration drilling strategies.
A characteristic of our model is that the variables have an immediate physical and easy-tounderstand meaning, directed by the geologic description of the area. In general, only the observable
nodes in a BN must have a physical interpretation.
However, when BNs are used for modeling, it is
common that all nodes have a physical interpretation (Imoto et al., 2003; Hamilton et al., 2005).
Edges generally only represent probabilistic relations, and not causality (Howard and Matheson,
2005), although they may encode a specific causal
relation as in Hamilton et al. (2005). In our specific case study, we are not trying to infer the directions of the edges from a set of data. Instead, we
assign physical interpretation to the edges. An edge
represents possible flow from one node (area) to
another. We can do this because the network is
specified by experts, and thus the possible flow
directions are known. When we use the BN structure to answer specific decision problems, we of
course invert edges to propagate evidence.
Bayesian Network: Quantitative Part

The key feature as we build the BN is that we can
define the conditional dependence locally. After
defining the edges and nodes of the network, the
next step is to assign probabilities for all the required conditional distributions (from the kitchens
to the prospects, within the prospects, and from
the prospects to the segments). We require a conditional probability for the gas, oil, and dry outcomes
for all nodes, given the outcome of the parent nodes.
When all the conditionals are defined, a valid joint
distribution is also available.
We now focus on the directed edges in Figure 1.
The direction of the edge has both a physical and
probabilistic interpretation: physical in terms of
a more likely direction for gas or oil flow from
one prospect to another and probabilistic in the
sense that the distribution for the child node is
defined via the conditional distribution given all

parent nodes.
If vi ; vj 2 E; but vj ; vi 2
=E )
vi ! vj and vi is parent of vj
Each arrow corresponds to a possible flow of
gas or oil, assigned with a certain probability. In
particular, for each arrow, we have two values:
eHC
P vj gas _ vj oil j vi gas _ vi oil
ij
and
eG
ij P vj gas j vj gas _ vj oil
^ vi gas _ vi oil
that is, eHC

ij represents the probability of HC flow
from i to j, and eG
ij represents the conditional probability of node j being gas, given that there has
been an HC flow between node i and node j.
This information alone is not enough to build
a unique BN. Given the marginal probabilities for
the top nodes and the net of possible flows, we
must be able to build a table of conditional probabilities given the parents.
To handle the ambiguity challenge present
when two possible types of flow (oil and gas) mix
in the same node, we use physical considerations.
A well-known theory on HC mixture, see for example, Gluyas and Swarbrick (2003), tell us that
if both oil and gas enter the same trap, the result
will be gas. This is because gas is lighter than oil,
and it forces itself into the pockets and squeezes
the oil out of the trap. For this mechanism to be in
place, we need an abundance of gas and enough
time to complete the process. We consider that
these requirements are met in the area under consideration, and we disregard the possibility of reservoirs with both oil and gas.
Another physical assumption that we make is
that gas can only appear if at least one parent is
gas as well, whereas oil can be generated even if
the parents are all gases. This occurs because the
presence of oil can be seen as a failure in the propagation of gas. Obviously, the dry state only propagates dry.
Martinelli et al.
1427
Table 1. Structure of the Conditional Probabilities for Node i with Two Parents, j and k
vj
Dry
Gas
Oil
Dry
Gas
Oil
Dry
Gas
Oil
vk
P (vi = Dryvj, vk)
P (vi = Gasvj, vk)
P (vi = Oilvj, vk)
Dry
Dry
Dry
Gas
Gas
Gas
Oil
Oil
Oil
1
1 eHC
ji
1 eHC
ji
1 eHC
ki
HC
1 eHC
ji 1 eki
HC
1

e
1 eHC
ji
ki
1 eHC
ki
HC
1 eHC
ji 1 eki
HC
1 eji 1 eHC
ki
0
G
eHC
ji eji
0
G
eHC
ki eki
HC
G
G
1 1 eji eji 1 eHC
ki eki
G
eHC

e
ki
ki
0
G
eHC
ji eji
0
0
eHC

1
eGji
ji
HC
eji
eHC

1
eGki
ki
HC
G
HC
G
HC
1 eji eji 1 eki eki 1 eHC
ji 1 eki
HC
HC
G
1 1 eHC
1

e

e

e
ji
ki
ki
ki
eHC
ki
HC
HC
G
1 1 eHC
ji 1 eki eji eji
HC
HC
1 1 eji 1 eki eki
These two behaviors reflect in the entries of the

CPT. In the simple case with a node j with a single
G
parent i, where eHC
ij > 0 and eij > 0, we have
P vj dry j vi dry 1 and

P vj gas j vi dry
P vj oil j vi dry 0
P vj dry j vi gas 1 eHC
ij ;
G
P vj gas j vi gas eHC
ij eij ; and
HC
P vj oil j vi gas eij 1 eG
ij
P vj dry j vi oil 1 eHC
ij ;
P vj gas j vi oil 0; and
P vj oil j vi oil eHC
ij
When we have more than one parent, we assume that flow from each parent is independent of
the flow from the others. This is a key assumption
that allows us to use only these two parameters
per edge, independent of how many parents a node
has. It is also a physically reasonable assumption,
as no obvious correlation exists between separate
migration paths. Combined with our previous assumption of abundance of gas, and with the basic
propagation rules just described, we can easily build
Table 1 for a prospect having two parents using just
four parameters. In the very same way, the assessment
of the conditional probability for the node P13,
with five parent nodes, requires 10 parameters.
We set the parameters using expert geologic
knowledge of the area. This includes geologic simulation, information about the diagenetic processes
occurring, accounting for formation, age, and so on.
Note that some prospects have a marginal probability of gas equal to 0, no matter which kind of
1428
information they receive from other parts of the

network. For instance, this is the case for segments
10B and 10C, whose conditional probability of gas
flow from the parent is 0. We recall that we started
our construction of the network from the local
marginal probabilities, and we built the conditional
probabilities so that they could match the information that we had from the geologic experts. This
behavior is therefore consistent with the a priori
belief.
Another direct consequence of the way in
which the network is built is that the marginal HC
probabilities for the segments are always lower than
the corresponding prospect probabilities. Even if
HC is effectively present, a positive probability (local failure probability) of having a local bypass of
the HC exists, and this makes the HC less likely at a
segment level than at a prospect one.
EXPLORATORY ANALYSIS
With the assumptions listed in the previous section,
a network with 42 nodes is obtained. Through the
Junction Tree algorithm (Lauritzen and Spiegelhalter,
1988; Cowell et al., 2007), it is possible to compute
the joint distribution and the marginal distributions
P vi ; i 2 f1; . . . ; 42g; and vi 2 fdry; gas; oilg.
Figure 2 shows these a priori marginal probabilities of gas (Table 2). We can immediately see
that all the kitchens are of type gas, and this fact
is reflected, at least at a first glance, in the marginal distribution. Nonetheless, the mechanism of
oil squeezing explained in the previous section is
Figure 2. Marginal probabilities of the state gas for the

whole network.
working quite well, and allows a non-zero probability of oil, especially in the bottom segments.
Single-Segment Evidence
A couple of what-if examples will help increase
our understanding of the BN model. We specify a
particular evidence in one segment and compute
the difference between the prior and conditional
probabilities given the evidence. In practice, this
evidence would be the outcome of an exploratory
well at the chosen segment.
1. Segment 9A dry: Figure 3A shows the difference
in the marginal probabilities of dry wells before
and after observation of a dry well in node 9A.
The difference appears to be quite high, not
only in the well belonging to the same prospect
(9B), but also in the backward and forward paths.
Such evidence has much influence on prospects
P5 and P8, and even P4. As a result, the whole
area becomes less interesting to explore.
2. Segment 10A dry: When dry evidence is observed in node 10A, the difference in prior and
conditional probabilities is not nearly as large as
for 9A (Figure 3B). The difference is 0.390 for

the prospect itself, 0.179 and 0.163 for segments
belonging to the same prospect (10B and 10C),
and much smaller for segments far away (0.054
and 0.048 for 6A and 6B). This small influence
is caused by the fact that the P (HC flow from
P10 to 10A) (0.54) is much smaller than P (HC
flow from P9 to 9A) (0.99). In this latter case,
the lack of HC is therefore more probably a
local failure, whereas in the former case, local
failure was unlikely, and thus we got a large regional impact.
Multiple Sites Evidence
We next study what happens if the evidence
involves more than a single segment. As in the
previous subsection, we will present some explicative examples.
1. Segment 9A dry: Figure 3A shows what happens with a dry evidence in segment 9A. In that
case, we had a large number of segments in the
central sector with a high dry probability given
the single-segment evidence at 9A. Now, we
Martinelli et al.
1429
Table 2.
Node
Type
Gas Volume
(billion m3)
Oil Volume
(million m3)
Marginal Dry
Marginal Gas
Marginal Oil
K1
K2
K3
K4
P1
P2
P3
P4
P5
P6
P7
P8
P9
P10
P11
P12
P13
1A
2A
3A
4A
4B
5A
5B
5C
6A
6B
6C
7A
8A
9A
9B
9C
10A
10B
10C
11A
12A
12B
13A
13B
13C
Kitchen
Kitchen
Kitchen
Kitchen
Prospect
Prospect
Prospect
Prospect
Prospect
Prospect
Prospect
Prospect
Prospect
Prospect
Prospect
Prospect
Prospect
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
Segment
9.7
0.4
3.5
0.6
1.9
0.5
1.3
2.4
8.8
8.8
1.6
0.4
0.1
0.6
1.8
0.1
0.8
2.2
0.8
3.1
4.4
10.3
0.5
8.9
7.2
5.9
0.5
2.3
0.8
2.8
2.4
6.9
14.7
3.0
3.4
9.7
1.7
0.1
0.2
1.3
1.3
6.7
19.0
7.1
1.4
7.3
4.8
2.7
3.4
2.7
0.00
0.00
0.00
0.00
0.19
0.39
0.59
0.19
0.09
0.09
0.17
0.22
0.10
0.29
0.18
0.41
0.09
0.20
0.40
0.60
0.28
0.20
0.21
0.34
0.52
0.10
0.20
0.80
0.19
0.36
0.10
0.10
0.10
0.61
0.30
0.37
0.18
0.50
0.41
0.10
0.10
0.30
1.00
1.00
1.00
1.00
0.70
0.49
0.33
0.65
0.77
0.78
0.75
0.69
0.68
0.41
0.66
0.47
0.73
0.52
0.21
0.26
0.57
0.29
0.04
0.03
0.00
0.72
0.64
0.00
0.00
0.32
0.45
0.45
0.00
0.22
0.00
0.00
0.41
0.25
0.47
0.00
0.72
0.56
0.00
0.00
0.00
0.00
0.11
0.12
0.08
0.16
0.14
0.13
0.08
0.09
0.22
0.30
0.16
0.12
0.18
0.28
0.39
0.14
0.15
0.51
0.75
0.63
0.48
0.18
0.16
0.20
0.81
0.32
0.45
0.45
0.90
0.17
0.70
0.63
0.41
0.25
0.12
0.90
0.18
0.14
1430
Figure 3. Difference between

the a priori probability of dry
and the conditional probability
of dry after observing a dry
evidence in sites (A) 9A and
(B) 10A.
could be interested in excluding a larger area

with just another single exploration well. We
study what happens if we have another dry evidence in segment 6A. As we see in Figure 4,
thanks to this second observation, we can exclude a larger area and we increase almost all
dry probabilities considerably. However, if we
want to enforce our previous opinion near prospect P9, we might decide to drill segment 9B.
Figure 5A shows the updated probabilities in the
case of dry evidence at 9B, whereas Figure 5B
shows the more probable result of gas.
2. Segment P12 oil: One interesting effect of using
a BN model in this context is that we can construct behaviors that may be considered counterintuitive. By this term we mean that the network
may weaken a prior assumption in the presence
of positive observations, instead of enforcing the
assumption. Consider as an example prospect
P12 in the western part of Figure 1. This pros-
pect does not receive any direct contribution

from a kitchen but is fed by three other prospects. If we observe oil in a segment belonging
to the prospect P12, this evidence will strengthen
our confidence in all the surrounding prospects.
If we furthermore observe oil in segment 8A,
the local oil probability in segment 11A decreases with respect to the single evidence. We
would expect to find gas in this area but find oil
and a feeder that provides only oil. This decreases the probability of HC in all other prospects leading to P12 because these would most
likely carry gas, and we see oil there. A similar
phenomenon occurs at segment 7A because the
new observation reinforces the confidence in a
particular flow direction (from prospect P8 to
P12). In general, confirming a flow direction
does not influence our belief in other flow directions, as it should be, unless other conditions
hold at the same time.
Martinelli et al.
1431
Figure 4. Difference between

the a priori probability of dry
and the conditional probability
of dry after observing a dry evidence in segments 9A and 6A.
EXPLORATION STRATEGIES
The preliminary analysis that we described so far
is useful for understanding the way evidence propagates through the network. We now use the BN
model for decision making. The problem setting
is similar to the one discussed in Van Wees et al.,
2008, and it occurs in the field of managing risks
and uncertainty in a portfolio of projects (Smalley

et al., 2008). Note that our approach is different
from the goal of optimal sequential exploration
discussed in Bickel and Smith (2006) and in Bickel
et al. (2008). We are not looking for an optimal
dynamic exploration strategy here, but aim for
an exhaustive static search of an optimal set of
wells.
Figure 5. Difference between the

a priori probability of dry and the conditional probability of dry after observing a dry evidence in segment 9A
and a dry (A) or gas (B) evidence in
segment 9B.
1432
The exploration strategy depends on the choice

of the optimization criterion. We have chosen two
different criteria to study this segment selection
problem. These criteria are similar to the ones proposed in Van Wees et al. (2008) based on the
maximization of the net present value, but they
are slightly different and they answer to precise
questions coming from the oil industry:
Which is the best segment combination to use
to drill a certain (fixed) number of exploration
wells to maximize the total revenues?
Which are the best segments to drill to maximize our confidence that we can leave the area
without losing a great potential? In other words,
which are the best segments that we should drill
to be (almost) sure that in the worst scenario
(segments dry), the whole area is actually dry.
The first question is more traditional, in the
sense that it has been discussed in other case studies
for the same purposes. In our analysis, this question
is not trivial. We use the VoPI as a criterion. This is
widely used in the petroleum industry, see for example, Bratvold et al. (2009) and Bhattacharjya
et al. (2010). In some situations with uncertainty
connected with the evidence, one could compare
the VoPI with the value of imperfect information
(Bickel, 2008).
The second question relates to the decision of
leaving an area after drilling dry wells. We suggest
maximizing the probability that we are leaving a
dry area, and this entails minimizing the AR criterion. We fix the maximum number of segment p
that we can afford to drill.
Both for the VoPI and for the AR, the computational effort is increasing with the dimension p.
Indeed, when p = 1, we have to consider just 25
different prospects, whereas
when p = 2, the number

of combinations is 25
300.
It becomes 2300
2
when p = 3 and so on. We acknowledge this combinatorial problem for a large p. Nonetheless, it is
reasonable to fix a static number of sites and then
consider all the possible solutions. A standard evaluation of the network takes about 0.16 s. A strategy
up to p = 4 includes at most 1 million combinations
(12650 34), which is feasible in tens of hours.
Background for Case Study

Table 2 shows the expected volumes of oil and gas
for each segment. These values are specified by
experts of the area. We assign no uncertainty on
these values, as discussed in Cunningham and Begg
(2008). Given these expected volumes and the expected price for oil and gas (respectively, $75/bbl
and $0.35/m3), we can compute the expected revenues for each segment.
We next consider the costs related to the
exploration and development. We use four kinds
of costs: a fixed cost that represents the cost of an
exploratory well (well fixed cost [WFC]), a second
exploration cost per prospect (exploration fixed
cost [EFC]), a third major cost for the infrastructure and the well development costs (drilling
fixed cost [DFC]), and a fourth cost proportional
to the expected volume of gas and oil representing
the operation and maintenance costs. We assume
that the second and the third fixed costs have to be
paid just once per prospect because the segments
belonging to one prospect can share infrastructure.
Thus, if we are going to drill different segments
belonging to the same prospect, we save considerably. This assumption is realistic because different segments commonly correspond to formations
at different levels of depth. We note that WFC is
valid at the segment level because all the wells are
considered independent at the exploration stage.
The expert reference cost values for the area
under consideration are on the order of WFC =
$20 million, EFC = $390 million, whereas the proportional part is on the order of $20 million per
billion m3 of gas and per million m3 of oil. The
DFC = $500 million in our first analysis, but we
later check sensitivity to this cost. We do not include any uncertainty in these values nor in the
volumes. The cost parameters and the prices of oil
and gas are subject to change with time, they depend on the location and depth, and they could
therefore vary between prospects and segments.
They should be viewed more as the attempt to
mimic a likely situation, than as a set of invariant
parameters. The quantities allow us to compute for
each segment the partial revenues for oil (PRO),
that is, the expected oil revenues minus the flexible
Martinelli et al.
1433
Figure 6. Intrinsic value (in

million USD) of the 25 segments
in our network, that is, the expected value of each well in different configurations (oil or gas)
times the marginal probability
of the configuration itself. In this
graph, the prospect costs are
not considered because they can
be shared by multiple segments.
The total sum of intrinsic revenues is approximately $23.7 billion.
costs proportional to the estimated oil volume of

the segment, and the partial revenues for gas (PRG),
that is, the expected gas revenues minus the flexible costs proportional to the estimated gas volume
of the segment.
Given the above quantities, we compute the
intrinsic value (Bickel and Smith, 2006) that is the
unconditional expected value of the segment, that is,
the expected value given success times the probability of success (in our case, oil and gas) minus the
expected value given the failure times the probability of dry. The result is shown in Figure 6. Segment 10B is the a priori most lucrative segment
with an intrinsic value of about $6 billion.
The Value of Perfect Information
The goal is to build a single index related to a
segment or a set of segments that prospectively (i.e.,
on expectation before drilling any segments) is able
to help us formulate a drilling strategy. We are interested in building an a priori exploration timetable no matter the result of possible exploration
wells. This would be a useful tool for oil and gas
companies because the planned drilling strategy has
to be submitted in advance, along with the number
of trial segments. The definition of VoPI in this
context is strongly affected by the choice of the
strategy that the company wants to pursue. By the
word strategy, we mean both the exploration part
and the subsequent development strategy. As dis1434
cussed in the previous section, we introduce some

simplifications to handle this problem, but many
possible solutions still exist, depending both on the
personal choice of the decision maker and on the
actual sequence of events. The intrinsic difficulty
of integration of risk analysis with production strategy definition has been effectively underscored in
Suslick and Schiozer (2004, p. 4), where the authors point out that several alternatives are possible, and restrictions have to be considered.
The basic ideas of our strategy can be resumed
as follows:
The first decision concerns the best exploration
strategy, that is, the best segments to explore
first. Each segment has an exploration cost equal
to WFC + EFC.
Given the result of the exploration, the probabilities of the other segments are updated.
For each prospect, the first decision is whether
to develop, considering the potential revenues
associated with the segments within that prospect and the sum of all the costs of prospect
development and segment development.
Given a positive decision at prospect level, we can
decide which segments to develop, depending
on the local expected marginal revenues.
A consistent procedure for this value computation, segment by segment, can be outlined using
the specified decision strategy. This is summarized
mathematically by the computation of the total

expected revenue (ER) for segment i:
ERi
evidence j
maxfRevkjvi j DFCg; 0A
prospect k
Pvi j
where the revenues for prospect k, given vi = j, are

given by
Revkjvi j
PROl P vl oil j vi j

+ PRGl P vl gas j vi j
2
segment l2k
with notation introduced in the previous section.

Equation 1 is a weighted sum of nonnegative prospect revenues with perfect evidence at site i alone.
Note that although the information collected at
site i is perfect, we want to study how this transfers
to the other segments. Thus, we have only partial
perfect information (site i in this setting). We are
not valuing just the information related to a decision on site i, but rather the impact of this perfect
information on all other decisions.
For every prospect, k, and every type of evidence, j, we decide to produce HC only if the
expected revenues are larger than the drilling costs.
Equation 2 collects the expected revenues for all
segments, l, belonging to prospect, k, given evidence, j, according to the previously defined rule.
No revenues exist when the segment is dry. The
VoPI is defined by subtracting the prior value, that
is, the expected revenue without evidence:
VoPIi ERi
maxRevk DFC; 0
prospect k
3
where
Revk

PROl P vl oil
segment l2k

+ PRGl P vl gas
X
The VoPI is always nonnegative because new information at worst leads to the same decisions as
before and may lead to improved decisions. The
definition of VoPI is consistent with the one proposed in Bhattacharjya et al. (2010), but this standard formulation is complicated by our segment or
prospect model, that is, we do not have a standard
costk for a selected prospect k, but a set of costs
considered at different segment or prospect levels.
We will use the VoPI to choose the best segments for exploration. As suggested in the literature (Eidsvik et al., 2008), the VoPI should be
compared with the actual cost of acquiring evidence. These costs are represented by the exploration costs, that is, by the EFC and WFC. Our
formulation of the model says that the former are
costs shared by all the segments within a prospect,
whereas the latter are costs for the single segment
(well). Therefore, we have to consider all the valuable segments within prospect k and compare their
VoPI sum with the following sum of costs:
Costk EFC +
(
where Il
Il WFC;
segment l2k
1 if VoPIl > 0
0 otherwise
Equations 1 and 2 can be extended to include

evidence at multiple segments of dimension p > 1.
The probabilities of segment outcomes are then
computed conditional on the joint evidence, and
the sum is over all kinds of joint evidence at the
selected exploration segments. In practical terms,
we propose a strategy as follows:
Select all the possible combinations of the segments of size p.
For each combination, compute all the evidence
configurations (3p because we have three possible states for each segment), their prior probability of occurrence, and the conditionals for
the other segments given that evidence.
Compute the expected revenue associated with
each evidence, considering that some fixed costs
are shared for segments belonging to the same
prospect.
Martinelli et al.
1435
Figure 7. Value of perfect information (in million USD),

with a single segment evidence
(p = 1). The parameter settings
are presented in Background
for Case Study, with DFC =
$500 million (o) or $2000 million
(x). The dashed line represents
the EFC cost ($390 million). DFC =
drilling fixed cost; EFC = exploration fixed cost.
Find the corresponding VoPI by subtracting the

expected revenue without evidence.
Results are shown in Figures 7 ( p = 1) and 8
( p = 2). Figure 7 presents the VoPI as a function
of the segments on the first axis, with two possible
values of DFC, respectively, $500 million (o) and
$2000 million (x). The largest VoPI when DFC =
$500 million is about $215 million and occurs at
segment 3A. This segment has a small intrinsic
value (Figure 6) but has a large VoPI because the
evidence would resolve ambiguity about oil and
gas at this segment, which have relatively small
prior probabilities. The VoPI is also a result of evidence propagation giving increased expected revenues at neighboring segments. Very little corre-
lation exists between the intrinsic values and the

VoPI results. For instance, segment 10B with the
highest intrinsic value has only a moderate VoPI.
To understand this, we have to reflect about the
mechanisms that influence the VoPI in this context. A particular segment has a high VoPI if its
impact is able to change the decision that we would
make at this segment and at other segments. Note
then, in this case, when comparing these values
with the exploration costs in equation 5, we notice
that none of the segments has a VoPI that is able to
justify a drilling, that is, the benefits gained by the
surplus of information is not worthy of the exploration costs needed to collect the information
itself, as shown by the dashed line in Figure 7,
that corresponds to the major exploration cost
Figure 8. Value of perfect information (in million USD)

for evidence at two segments
( p = 2). The parameter settings
are presented in Background for
Case Study.
1436
Figure 9. Value of perfect information (in million USD) for single-segment evidence (p = 1). On the x axis is the drilling fixed cost (DFC)
from $0 to $15 billion; on the y axis is the list of the segments.
EFC = $390 million. When DFC = $2000 million,

the maximum value of VoPI is above $800 million,
and it is reached at segment 12B. In this case, the
VoPI widely exceeds the exploration costs; therefore, an exploration well on segment 12B, 12A, or
10B can improve our knowledge of the network
and, more importantly, can sensibly change our
decision in the area. Note further that, as we will
point out in this section, VoPI(i) can be higher for
either of the two considered values of DFC; that is,
no monotonicity exists in the value because the
development costs strongly affect our decision.
Figure 8 shows the VoPI computed for all pairs
of segments (p = 2), for DFC = $500 million. The
highest VoPI occurs in the pair 3A and 12B. Both
of these had high marginal VoPI in Figure 7. The
VoPI seems to be close to the sum of the marginal
VoPIs but, in general, the joint value can be higher
or lower than the sum of marginals. The nonadditivity of the value of (perfect) information is
well known in the literature, and it was first pointed
out in Samson et al. (1989). If two segments carry
similar information, the joint value is smaller. However, multiple evidence (p 2) could trigger segments that would not be worth drilling based on
single-segment evidence.
The reason for the VoPI of different segments
is not immediately identifiable. Because the dril-
ling decision is a direct function of the chosen

costs, we next return to the case of single-segment
evidence (p = 1) and try to understand how different segment VoPIs depend on these costs. In
Figure 9, we plot the VoPI as a function of the
DFC. We will next point out several interesting
elements in this display.
First, for all segments, we see a common behavior of VoPI increase up to a certain level and
then a decrease down to 0. For a very low DFC, a
chance that an evidence collected somewhere could
make us change our decision does not exist because
we would drill anyway. However, if the DFC is
too high, a chance to overcome that threshold does
not exist, not even with a high increase of the marginal probability because of the evidence collected.
For some intermediate range of DFC, the VoPI is
highest because evidence will change the decisions.
Second, the VoPI plot at any segment is not
unimodal, but has multiple modes for different
costs. The VoPI for a given prospect is maximized
when the decision for this prospect is at its most
uncertain, that is, when the expected revenue from
drilling is at the same level as the costs. This feature
is typical of the two-action problems (Bickel, 2008),
like the one under consideration. As the cost varies,
different prospects approach their break-even point,
and thus give local maxima in the VoPI.
Martinelli et al.
1437
Figure 10. Value of perfect information (in million USD) for single-segment evidence, not including the self-evidence. On the x axis is
the drilling fixed cost (DFC) from $0 to $15 billion; on the y axis is the list of the segments.
Third, a sort of continuity in the segments belonging to the same prospect exists, even if their
volume could be different. This is also consistent
with the intuition because we expect that neighboring segments give rise to similar information
and decisions, although their volumes and their
marginals could be very different (see e.g., segments 13A, 13B, and 13C).
From this point of view, it is quite important to
distinguish between two different sources that constitute the VoPI. In equation 1, we see that the sum
is over all the prospects and all the segments. The
larger component of that sum could be the so-called
self-evidence, that is, the component of the sum
related to the segment under consideration. This
contribution is most significant when the a priori
product of revenues and marginal probabilities is
not above the fixed threshold, whereas the revenues
themselves (i.e., multiplied by one) are above the
threshold. In the last analysis, this behavior is caused
by the fact that we are dealing with perfect information under the hypothesis to receive a real evidence from the segment, and not just a measurement or some data.
We next consider the part of the VoPI that
does not include the self-evidence to measure the
1438
impact of an evidence on the other segments. Results that do not include the self-evidence are shown
in Figure 10. To analyze these results, we have to
compare them with the ones in Figure 9. We focus
attention on segment 1A. From the map of the area
(Figure 1) and from the initial exploratory analysis,
we know that segment 1A has mostly impact on
prospects P6 and P10. We can recognize signs of
this situation from the comparison of the two figures. In Figure 9, three peaks along the 1A line
exist, whereas in Figure 10, just two peaks remain.
Therefore, we can easily infer that the first peak in
Figure 9 is caused by self-evidence. The second
and the third peaks are mainly caused by propagation of evidence to neighboring prospects. In
particular, the second peak corresponds to a change
in the drilling decision in segments belonging to
prospect P6. The cost at this peak is exactly the
same as that of the self-evidence peaks for segments
belonging to prospect P6. In the same way, and to
a greater extent, the third peak corresponds to a
change in drilling decisions in the segments of
prospect P10, whose huge peak of self-evidence
is easily seen in the comparison between the two
figures. Recall that results in Figure 7 use DFC =
$500 million, which actually is in the very lower
end of the DFC range used in Figures 9 and 10.

Thus, the VoPI of about $220 million in Figure 7
is mainly caused by self-evidence.
A similar analysis can be done for all the other
segments, gaining similar information that guide
decision making. The relationship between VoPI
and intrinsic value exists only through self-evidence.
The main difference between Figures 9 and 10 is
in segment 10B, whose huge volume (Table 2) leads
to a big intrinsic value (Figure 6). We have seen
that self-evidence is a consistent component of the
VoPI, but it is not the unique one, especially if the
volumes associated with the segments are small.
We believe therefore that in this context, the VoPI
is able to blend the different components of the
geologic modeling in an interesting way. The volumes, the costs, and the structure of the network
are combined to deliver the VoPI as an index for
decision making.
Abandoned Revenue
We define AR as the remaining expected value of
the area given a set of dry observations at particular
locations. The criterion is similar to the VoPI, but
here, we include all the possible costs that affect a
segment, both the exploration (EFC and WFC)
and the production (DFC and variable costs) ones.
The AR of a particular segment, i (single observation) is defined as
ARi
max Revkjvi dry DFC EFC; 0
prospect k
6
where Revkjvi dry is defined as follows:
Revkjvi dry

max PROl P vl oil j
segment l2k
vi dry + PRGl P vl gas j

vi dry WFC; 0
Equation 6 is valid in case of single-segment

evidence. In particular, equation 6 is the sum of
nonnegative revenues given the dry evidence observed at segment i. The maximum implies that
prospect k is exploited only if its revenues exceed

the costs. If we can choose p 2 segments, the
conditioning would be over several dry segments,
and the procedure can be easily generalized as
seen for the VoPI. In this case, we are not considering the possibility of getting evidence different from dry. Therefore, the conclusions and the
suggestions provided by this index are valid just in
the aforementioned case.
Figure 11 shows the AR as a function of the
segments on the first axis. The dashed line ($23.7
billion) is the expected revenue without evidence
(sum of positive intrinsic values). Segments 6A and
10B give the smallest abandoned revenue and thus
the largest decrease as compared with the a priori
value. Both of these sites give a small remaining
value in the area, and one might ask if, given the
possibility of drilling two segments, it is best to
drill both of these. However, the two segments are
close in the network, and one might provide more
information by exploring segments that are further apart. In Figure 12, we compute AR for all p =
2 pairs. The hinted potential of the BN is really
exploited for p = 2, and the best choice is not to
drill the best 2 segments of the case p = 1 (6A,
10B), but to choose two segments far apart ([6A,
9A] or [6A, 9B]). It is, therefore, preferable to
choose those pairs that carry information from
different parts of the network, instead of the two
segments that appear most promising using a
marginal analysis, having overlapping information.
The AR for p = 2 is commonly far from the sum of
two p = 1 cases. A quite strong joint interaction
reduces the pair effect. Recall that this difference
was not very pronounced for the VoPI, where we
marginalized over the evidence. The interaction
effect depends on the segment locations in the BN.
The results from p = 3 confirm our results; the
best choice includes sites 6A, 9A, and 13A, that is,
three segments far from each other. Although the
improvement brought by adding one more segment is quite consistent with respect to the minimum for p = 2 (45% reduction), the percentage of
triple segments whose AR is lower than the minimum double segment is very low (4%).
The AR criterion can be useful in situations
with dry evidence. Our strategy is always based on
Martinelli et al.
1439
Figure 11. Abandoned revenue (in million USD) with single segment (p = 1) dry evidence. The parameter setting is described in
Background for Case Study.
static criteria, therefore, we have to fix a priori the

number of wells that we want to drill and act according to this defined plan. In this case, sites 6A
and 9A are indicated as the best choice for p = 2.
They return an AR so small that (on expectation)
the area is no longer worthwhile.
CLOSING REMARKS
The modeling of prospect dependencies with a BN
appears to be a good technique to model the geologic connections between different prospects and
the impact on the evaluation of an exploration
Figure 12. Abandoned revenue (in million USD) with dry evidence at two segments (p = 2). The parameter setting is described in
Background for Case Study.
1440
strategy while maintaining statistical computational

benefits. The key elements in building a BN here
are to separate the geologic components (kitchen,
prospect, segment), keep the connections simple
in the BN, and allow for a local failure probability
(oil or gas ! dry) between prospect and segment.
For using the model to select exploration strategies, we proposed two criteria: the VoPI and the
AR. We indicated guidelines for the interpretation
of the results. The VoPI is a monetary amount based
on the expected revenues with and without evidence. The AR is the expected remaining revenues,
given dry evidence. Both have their merits and
seem very useful in exploratory decision making.
One challenge with the presented approach
is the construction of the BN model. We benefited
from the consultancy of expert knowledge of the
local geology. We believe, nonetheless, that this
procedure can be automized to a certain level
through the run of an ensemble of geologic models.
The result of geologic modeling under different
diagenetic and depositional conditions could help
define the required case-specific conditional probabilities of the network. One could also imagine
combining several BN models with different edges
and probabilities.
Another choice that has to be made for each
specific application is the evaluation criterion. It is
possible to use more complex criteria based on
specific utility functions. This seems to be particularly important to avoid delayed drilling of the
most promising segments. The presented approach
can run a large number of statistical computations in
a short time. Nonetheless, for increasing dimensions
of evidence p, a combinatorial problem will arise.
For instance, to solve the optimal sequential exploration problem for our problem, we would have
to turn to some approximations. The use of decision trees, combined with some kind of pruning,
could provide interesting solutions.
The proposed criteria share the feature of relying on perfect information. Instead of observing
perfect evidence in a given segment, we could just
have the possibility of imperfect information from
that segment. In this situation, the value of information could be used to evaluate the attractiveness of buying such data (Eidsvik et al., 2008).
In the current article, we ignored the uncertainties in the volumes, revenues, and costs. We
simply conducted a sensitivity analysis on one
cost parameter. Possibly, hierarchical modeling
would be a flexible way to include more sources of
uncertainty.
We believe that this work summarizes how an
extensive use of statistical modeling tools can supply instruments to improve the understanding of
prospect dependencies and can help formalize a
decision problem.
REFERENCES CITED
Bhattacharjya, D., J. Eidsvik, and T. Mukerji, 2010, The value of information in spatial decision making: Mathematical Geosciences, v. 42, no. 2, p. 141163, doi:10.1007
/s11004-009-9256-y.
Bickel, J. E., 2008, Relationship between perfect and imperfect information: Decision Analysis, v. 5, no. 3, p. 116
128, doi:10.1287/deca.1080.0118.
Bickel, J. E., and J. E. Smith, 2006, Optimal sequential exploration: A binary learning model: Decision Analysis,
v. 3, no. 1, p. 1632.
Bickel, J. E., J. E. Smith, J. L. Meyer, 2008, Modeling dependence among geologic risks in sequential exploration
decisions: Society of Petroleum Engineers Reservoir
Evaluation & Engineering, v. 11, no. 2, p. 233251.
Bratvold, R. B., J. E. Bickel, and H. L. Lohne, 2009, Value
of information in the oil and gas industry: Past, present,
and future: Society of Petroleum Engineers Reservoir
Evaluation & Engineering, v. 12, no. 4, p. 630638.
Cowell, R. G., P. Dawid, S. L. Lauritzen, and D. J. Spiegelhalter, 2007, Probabilistic networks and expert systems,
Springer series in information science and statistics, New
York, Springer, 324 p.
Cunningham, P., and S. H. Begg, 2008, Using the value of information to determine optimal well order in a sequential
drilling program: AAPG Bullettin, v. 92, no. 10, p. 1393
1402.
Eidsvik, J., D. Bhattacharjya, and T. Mukerji, 2008, Value of
information of seismic amplitude and CSEM resistivity:
Geophysics, v. 73, no. 4, p. R59R69, doi:10.1190/1
.2938084.
Gluyas, J., and R. E. Swarbrick, 2003, Petroleum geoscience:
Oxford: Blackwell Publishing, v. XV, 359 p.
Hamilton, G. S., C. Alston, T. Chiffings, E. Abal, B. Hart,
and K. Mengersen, 2005, Integrating science through
Bayesian belief networks: Case study of Lyngbya in Moreton Bay: International congress on modeling and simulation (MODSIM05), Modelling and Simulation Society
of Australia and New Zealand, Melbourne, Australia,
p. 20892095.
Howard, R. A., and J. E. Matheson, 2005, Influence diagrams: Decision Analysis, v. 2, no. 3, p. 127143.
Martinelli et al.
1441
Imoto, S., T. Higuchi, T. Goto, K. Tashiro, S. Kuhara, and

S. Miyano, 2003, Combining microarrays and biological knowledge for estimating gene networks via
Bayesian networks: Proceedings of the IEEE Computer Society Bioinformatics Conference (CSB03),
p. 104113.
Jensen, F. V., 1996, An introduction to Bayesian networks:
New York, Springer Verlag.
Lauritzen, S. L., and D. J. Spiegelhalter, 1988, Local computations with probabilities on graphical structures and
their application to expert systems: Journal of the Royal
Statistical Society, Series B, v. 50, p. 157224.
Murphy, K., 2007, Bayes Net Toolbox for Matlab: http://www
.cs.ubc.ca/murphyk/Software/BNT/bnt.html (accessed
January 28, 2010).
Pearl, J., 1986, Fusion, propagation and structuring in belief
networks: Artificial Intelligence, v. 29, p. 241288,
doi:10.1016/0004-3702(86)90072-X.
Ramm, M., and A. E. Ryseth, 1996, Reservoir quality and
burial diagenesis in the Statfjord Formation, North Sea:
1442
Petroleum Geoscience, v. 2, p. 313324, doi:10.1144

/petgeo.2.4.313.
Samson, D., A. Wirth, and J. Rickard, 1989, The value of information from multiple sources of uncertainty in decision
analysis: European Journal of Operational Research, v. 39,
p. 254260, doi:10.1016/0377-2217(89)90163-X.
Smalley, P. C., S. H. Begg, M. Naylor, S. Johnsen, and A.
Godi, 2008, Handling risk and uncertainty in petroleum
exploration and asset management: An overview: AAPG
Bulletin, v. 92, no. 10, p. 12511261.
Suslick, S. B., and D. J. Schiozer, 2004, Risk analysis applied
to petroleum exploration and production: An overview:
Journal of Petroleum Science and Engineering, v. 44,
no. 12, p. 19, doi:10.1016/j.petrol.2004.02.001.
Van Wees, J. D., H. Mijnlieff, J. Lutgert, J. Breunese C. Bos,
P. Rosenkranz, and F. Neele, 2008, A Bayesian belief
network approach for assessing the impact of exploration prospect interdependency: An application to predict gas discoveries in the Netherlands: AAPG Bulletin,
v. 92, no. 10, p. 13151336.
Paper II
Dynamic Decision Making for Graphical Models Applied to Oil
Exploration
G. Martinelli, J. Eidsvik, and R. Hauge
Technical Report in Statistics, NTNU, 12, 2011 Submitted for publication.
Dynamic Decision Making for Graphical Models Applied to Oil Exploration

Gabriele Martinelli, Jo Eidsvik
Dept. of Mathematical Sciences, Alfred Getz vei 1,
Norwegian University of Science and Technology, Trondheim, Norway
Ragnar Hauge
Norwegian Computing Center, Gaustadalleen 23, Oslo, Norway
Abstract
We present a framework for sequential decision making in problems described by graphical models. The
setting is given by dependent discrete random variables with associated costs or revenues. In our examples,
the dependent variables are the potential outcomes (oil, gas or dry) when drilling a petroleum well. The goal
is to develop an optimal selection strategy that incorporates a chosen utility function within an approximated
dynamic programming scheme. We propose and compare different approximations, from simple heuristics
to more complex iterative schemes, and we discuss their computational properties. We apply our strategies
to oil exploration over multiple prospects modeled by a directed acyclic graph, and to a reservoir drilling
decision problem modeled by a Markov random field. The results show that the suggested strategies clearly
improve the simpler intuitive constructions, and this is useful when selecting exploration policies.
Keywords: Bayesian Networks, Dynamic Programming, Graphical model, Heuristics, Petroleum
Exploration
1. Introduction
This paper considers the problem of sequential decision making, where the outcome of one decision
will influence the others, and the decisions are based on the expected utility. Our motivation and main
applications are from oil and gas exploration, where a petroleum company has a set of potential drilling
sites, called prospects. For each prospect, we may either drill or not. There is a cost of drilling, but revenues
if the well discovers oil or gas. The prospects are statistically dependent, and drilling at one prospect gives
information that is used to update the probability of success at other prospects. The goal is to find an
optimal drilling sequence, including when to stop drilling and abandon the remaining prospects. Thus, we
are interested in designing a strategy or a policy for selecting the sequence of prospects, or at least the first
few best prospects in such a sequence.
The optimization of the expected utility function is a trade-off between two factors: the direct gain
from the exploitation, and the indirect gain of learning, or exploration, that helps us make informed future
decisions. The balance between these is controlled by a discounting factor. With no discounting, the problem
becomes a maximization of the value of information (VOI), whereas a high discounting factor leads to a
greedy approach where only immediate gain counts.
We have no theoretical restrictions on the underlying statistical model for dependence between outcomes.
In practice, there is a requirement that conditional distributions can be computed and updated fast, since
Corresponding
author
Email addresses: gabriele.martinelli@math.ntnu.no (Gabriele Martinelli ), joeid@math.ntnu.no (Jo Eidsvik),
ragnar.hauge@nr.no (Ragnar Hauge)
Preprint submitted to European Journal of Operational Research
February 21, 2012
many of these will be computed when designing a strategy. For comparing strategies, it is also advantageous
if we can easily simulate from the models. In our examples, we use Bayesian networks (BN) and Markov
random fields (MRF), which both have these properties.
This sequence selection challenge is a discrete optimization problem and the optimal strategy can be
found by Dynamic Programming (DP), see Bellman (1957) and Nemhauser (1966). However, DP becomes
computationally infeasible when the number of possible actions increases. A remedy for this is to apply a
heuristic approach. These strategies have been studied in many contexts due to the curse of dimensionality,
which affects most DP methods (Powell, 2008). The simplest heuristic is to run an independent strategy,
disregarding the information gain caused by dependent variables. A more sophisticated alternative is to use
a myopic strategy. This strategy conditions on past outcomes, but does not account for future scenarios in
a proper way.
A possible solution to large DP problems is also offered by approximate DP methods, see Bertsekas
and Tsitsiklis (1996) and Powell (2008). The main idea of approximate DP is to replace the optimization
function with a statistical model that captures the impact of decisions now on the future. Approximate
DP techniques for solving a multivariate knap-sack problem (Bertsimas and Demir, 2001) resembles the
situation of drilling wells, but in our graphical representation of dependent prospects it is not obvious how
to find a statistical model that approximates the future value function. Further, our main goal is to find an
optimal sequence, and most approximate methods do not give this as a byproduct when approximating the
utility function.
When considering a set of independent prospects, the optimal sequential decisions are offered by the
Gittins indeces (Gittins, 1979), introduced for solving bandit-problems (Weber, 1992). These methods
were used for a petroleum example by Benkherouf and Bather (1988). Here, the discovery probabilities in
different prospects are apriori independent, and later dependent just through the total number of discoveries.
In our context the correlation is much more complex, and the actions influence the model probabilities in a
complicated manner.
Branch and bound methods are non-heuristic in the sense that they produce lower and upper bounds of
the values (Goel et al., 1979). In practice the gap between bounds can be wide. Moreover, it is not obvious
how to generalize these methods for graphical models with dependence between prospects. In our context
we will typically lack monotonicity when computing the best (discounted) sequence. Branch-and-bound
methods seem more suited for the actual maximum value of the utility function, instead of an approximate
sequential decision strategy.
The challenge of constructing drilling strategies is of course well known in the oil and gas industry,
but no one seems to have looked at it from a modern statistical modeling viewpoint applying graphs to
couple many dependent prospects. Kokolis et al. (1999) describe a similar problem with a focus towards
decision making under uncertainty and the technical risks connected to a project. They do not consider
how to design an optimal sequential drilling strategy, but discuss the combinatorial increase of the number
of scenarios that has to be considered. Smith and Thompson (2008) analyze the consequences of dependent
versus independent prospects, and give drilling guidelines that are optimal in special situations. In Bickel
and Smith (2006) and Bickel et al. (2008), DP is used to compute the optimal sequences and profits from
six dependent prospects, but they do not indicate solutions for the large scale challenge.
Our approach is a classical DP procedure with the use of heuristics for approximating the continuation
value (CV). The CV is defined as the value of the prospects that have not yet been revealed in the sequential
exploration. This value of course depends on the outcome of the current sequence. The simplest form of
this is the naive strategy sketched above, where the CV is computed under independence. We use this for
benchmarking. In addition, we apply pruning of the decision tree, where we ignore unlikely branches to
reduce the combinatorial problem.
We use profit as utility function, which is quite reasonable for a large oil company. Alternatives would
be profit given that loss at no time exceeds a given value, or, in the case of entering new exploration areas,
minimum loss before concluding that there is no oil present. The profit criterion we use is not dissimilar
to the VOI. For instance, Eidsvik et al. (2008), Bhattacharjya et al. (2010) and Martinelli et al. (2011)
study the effects of more data acquisition, the ability to make improved decisions, and the associated VOI
for spatially dependent variables. However, they do not compute the VOI in a sequential manner (Miller,
2
1975), neither are they focusing on the best sequential exploration program.
Operational research practitioners could find immediate links between this work and other ranking and
selection problems in a sequential setting, see e.g. Frazier et al. (2008) and Frazier and Powell (2010). What
makes our case peculiar is i) the ability of the action of influencing the probability of success in the other
nodes, since all the nodes are connected though a graphical model, and ii) the combinatorial flavor of the
problem.
The paper develops as follows: In Section 2 we introduce the notation, statistical framework, and the
assumptions required for applying our methods. In Section 3 we present the DP algorithm for our problem.
In Sections 4 and 5 we propose the various heuristic strategies, and the algorithms used to evaluate the
properties of the sequential exploration strategies. Finally, in Section 6 we provide results for a small BN
model and a BN case study of 25 prospects in the North Sea, and a MRF for a oil reservoir represented on
a 5 20 lattice.
2. Assumptions and notation
We consider a set of N prospects with a discrete set of possible outcomes. These N prospect nodes
are a subset of the total M nodes in a graph. The remaining M N auxiliary nodes impose the specified
dependency structure in the model, but are not observable. For every node i = 1, . . . , M we have a discrete
random variable xi {1, . . . , ki }. In the examples below we use ki = k, and k = 3. The random vector of
all variables is x = (x1 , ..., xM ), where the N first components correspond to the prospect variables.
The directed acyclic graph (DAG) in one of our case studies is built from the causal large scale processes
required to make sufficient amounts of oil and gas, see VanWees et al. (2008) and Martinelli et al. (2011).
pa
A DAG defines the joint probability model p(x) from the product of conditional distributions p(xi |xi ), for
pa
all nodes i = 1, . . . , M , where xi denotes the set of outcomes at parent nodes of i. In the MRF example
for a lattice of cells in a specific reservoir unit, the model is defined over neighborhoods on the lattice, where
p(xi |xi ) = p(xi |xj ; j Ni ), and xi is the vector of all variables except xi , while Ni is the neighborhood
of node i. The particular type of model is not critical, but for our purposes fast updating of the conditional
probabilties is important. This updating is required when we get sequential evidence. BNs are fast to update
using for instance the junction tree algorithm, see e.g. Lauritzen and Spiegelhalter (1988) and Cowell et al.
(2007). Moderate size MRFs can be computed recursively by forward-backward algorithms (Reeves and
Pettitt, 2004). Moreover, we will use Monte Carlo samples to generate realistic future scenarios. It is easy
to draw samples x = (x1 , . . . , xM ) p(x) from the BNs and MRFs we consider.
Given a probabilistic model with a certain dependence structure, we want to develop a drilling strategy,
i.e. a dynamic road map that leads us through the exploration phase of the prospects. Since the prospects
are dependent, the outcome of one changes the probability of success in the others. The strategy of continued
drilling thus entails a sequential updating of the probability model.
We let i be the observable in node i = 1, . . . , N . If node i is not yet observed, we set i = . If we choose
to observe node i, i is the actual outcome of the random variable xi at this node. For instance, i = 1
can mean that well i has been drilled and found dry, i = 2 if found gas, and i = 3 if oil. Initially, before
acquiring any observables, we have = (, . . . , ). If we start to explore nodes, we put the outcomes at
the corresponding indices of the vector . Say, if node 2 is selected first, and observed in state 2 = x2 = 2,
we set = (, 2, , . . . , ). For the likelihood of this scenario we need the marginal p(x2 = 2). This is
computed by summing out all scenarios that share the second component equal to 2. In order to compute
the conditional probabilities of a node i, given evidence, we need p(xi = j|), j = 1, . . . , k, where the empty
elements () of are unobserved and marginalized out.
The CV associated with the state vector is denoted v(). This is the expected value of all currently
unobserved states given the observed states, the objective function, and the chosen strategy. One objective
is to find the initial value before any sites have been explored, i.e. v( 0 ) where 0 = {, , . . . , }. This
initial value is in theory given by DP. As an integral part of the DP algorithm one must evaluate the values
v() of all possible combinations of evidence. This becomes impossible when we have many nodes in the
graph.
3
The DP algorithm also gives the optimal sequential decisions, but since this is not feasible for large N ,
we instead construct forward selection strategies, approximating v() to different accuracies. When building
such strategies we make assumptions about the way decisions are made. First, we assume that the decision
maker selects one node at a time. Without this assumption, the problem would grow to allow all orders of
two-tuples, three-tuples, etc. Second, we assume that there are fixed revenues and costs associated with each
node. If we choose to explore a node, we have to pay a cost. For certain outcomes of the node variable, we
receive a revenue. For instance, if the outcome is oil, we get the fixed revenues associated with this outcome.
The revenues and costs change from node to node, but introducing random distributions on the costs and
revenues for each type of outcome would make our optimization problem harder. Finally, we assume the
utility function contains separate parts for every node, without any coupling of the nodes. This utility
function expresses the decision makers inclination to collect the revenues or cost at any site. In principle,
there could be shared costs or revenues for nodes, say if certain HC prospects have common infrastructure
(Martinelli et al., 2011). We could include this into our framework, but it gives extra computation time,
and obscures the presentation of the sequential strategies, that is the focus of our work.
Given these assumptions, we will next show how DP presents a recipe for computing the optimal strategy.
We will discuss why this is not possible for a model with many nodes, and we will instead propose strategies
to overcome the problem.
3. Dynamic programming
In our context DP recursively explores backwards all the possible sequences that may occur, and it
uses these evaluations to select the best dynamic decisions. See e.g. Bickel and Smith (2006) for a similar
application of DP.
By the word sequence we mean each of the possible situations that may arise. Sequences are indexed
by adding one element i {1, . . . , k} at a time to the evidence vector = (1 , . . . , N ). With N = 4
prospects, the state = {, 1, , 2} means that the node 1 has not yet been explored, node 2 has been
observed to be in state 1, node 3 has not yet been explored, and node 4 has been observed to be in state
2. Two different scenarios may correspond to this sequence, one when node 2 is explored before node 4,
and another when node 4 is explored before 2. This order is of course relevant when we have only explored
node 2, and consider observing node 4, or vice versa, but once both node 2 and 4 have been explored, we
no longer distinguish between these two scenarios (except for discounting purposes). Thus, we tend to use
the terms sequence and scenario as synonyms.
The decision tree (Figure 1) visualizes the chosen strategy. It works in the following way:
1. First, decide which site, if any, to observe first.
2. Then, depending on the outcome xi {1, . . . , k}, which node to observe next, if any, and so on.
DP solves the tree by working backwards:
1. First, decide whether to drill the last prospect, conditional on the first N 1 observables.
2. Then, decide which prospect to drill if there are two nodes left, and so on, to the initial empty set.
In order to pursue this strategy, we have to maximize a certain utility function. We use maximum profit,
and v() then represents the expected revenues of future cash flows given that we are in state . Initially,
the vector of observables is empty: 0 = {, , . . . , }. The maximization is among all possible free states:
"
( k
)#
k
X
X
j
l
v() = max
p(xi = j) ri + max
p(xs = l|xi = j)(rs + . . .), 0
,0 ,
(1)
iN
sN 1
j=1
l=1
where the second and the subsequent maximizations are over all nodes not yet considered. Here, is a
discounting factor that depends on the specific case and on the inclination of the decision maker. The rij are
4
2
1
6g
3
gas
4
oil
2
6o
3g
dry
gas
5
6d
3
4
oil
3o
dry
3d
Figure 1: Illustration of a decision tree. At the first branch we can select any of the 6 nodes, or quit (Q). Node 6 is explored
first here. If node 6 is dry, we select node 3 at the next branch. The outcome of node 3 can influence which branch to enter
next, and so on.
revenues or costs of node i with outcome j. When all the sites have been drilled, the CV is v(, , . . . , ) = 0,
and we can proceed backwards, one step at a time, to extract the DP solution.
Equation (1) can be rewritten (Bickel and Smith, 2006), and it can be seen as a maximization over all
free nodes and 0 (not exploring any further). This means that v() = maxi {0, vi ()}, where:
vi () =
k n
o
X
p(xi = j|)(rij + v( ji )) ,
(2)
j=1
where ji = { -i , i = j} and v( ji ) is the CV of the state ji , i.e. v( ji ) = maxl6=i {0, vl ( ji )}.

The main problem with this optimal DP solution is the exponential growth of the number of scenarios
that have to be considered. Bickel and Smith (2006) derives the computational cost for a non-recombining
tree, i.e. a tree ignoring the order of the observed nodes. Then,
Number of possible scenarios in a non-recombining tree: :

N
X
N
k i (N i + 1).
i
i=0
This entails an order of 104 scenarios for six nodes (Bickel and Smith, 2006), and 1015 when N = 25nodes.
The computational cost (proportional to the number of scenarios) is therefore in the order of N2 !k N/2 .
Bickel and Smith (2006) suggest to save the local results of the computations in order to reduce the number
on configurations to consider. Say, for the purposes of the CV, it does not matter whether we first drilled
first one well or another, given that we observe their outcomes. Nonetheless, the exact procedure remains
unfeasible when the N increases. Furthermore, we need to mention that the introduction of the discounting
5
factor makes impossible the use of classical non-recombining algorithms, and gives us few chances other
than following the described approach.
4. Heuristic strategies
Because of the rapid growth in scenarios, one must look for approximate solutions. The problem shares
some features with that of a chess game. The player has to choose among all the possible moves she can carry
out, and at the same time he has to consider all the possible replies of his opponent, and the consequential
replies of herself, and so on. What is done in practical chess algorithms is to limit the search to a reasonable
amount of moves forward, and to evaluate the best move in that restricted match, see Shannon (1950)
and Feldmann et al. (1994).
Similarly, we push the search through a certain number of steps, figuring out some rules to approximate
the remaining value of the scenarios. The idea is to introduce different and simpler rules, in order to
approximate the CV in equation (2) without going all the way down through the branches of the decision
tree. We will call these rules heuristics, following the literature described in Pearl (1984).
4.1. Naive strategy
The naive strategy ignores the dependence among nodes. Therefore, the decision is just based on a priori
knowledge. There is no learning. The CV is then estimated as a simple sum of a priori intrinsic values:
N
k
X
X
vN () =
max
rij p(xi = j), 0 .
(3)
i=1
j=1
The best sequence is therefore computed just once, at the beginning of the algorithm, and the nodes are
chosen according to:
k
k
X
X
j
k
ri p(xi = j), 0 . . . .
(4)
ri p(xi = j), 0 , i(2) = arg max
i(1) = arg max
i
i\i(1)
j=1
j=1
As we can see, the outcome of the first best prospect is irrelevant when choosing the second best site. This
approach, though being very simple (the computational cost is linear in N ), still captures a large part of
the value if the correlation between nodes is small. The main problem is the individuation of the correct
best sequence, since disregarding any correlation effect can lead to focused attention on nodes that might
not be appealing given the evidence of the previous steps.
4.2. Myopic approach
A second natural approach is represented by the myopic strategy (Bollapragada and Morton, 1999).
According to this strategy, the best sequence is computed step-by-step in a forward selection scheme. The
conditional probabilities in the different nodes are now updated, given the previous outcomes. This represents
an intuitive sequential strategy, but it only exploits the dependence in the graph through the past, and does
not consider what the future might bring.
The strategy for finding the first best prospect coincides with the naive approach:
X
i(1) = arg max
rij p(xi = j), 0 .
(5)
i
j=1
Given an outcome x(1) at this first selected node i(1) , the second myopic best site is then chosen as a function
of the observable in the first node:
k
X
j
ri p(xi = j|xi(1) = j1 ), 0
(6)
i(2j1 )|xi(1) =j1 = arg max
i\i(1)
j=1
k
X
rij p(xi = j|xi(1) = j2 ), 0 , . . .

i(2j2 )|xi(1) =j2 = arg max
i\i(1)
j=1
Now, the second best choice, therefore, involves k different maximizations,

depending on the outcome of
PN
xi(1) . Thus, using a myopic strategy leads to a decision tree with i=0 k i scenarios.
The myopic approach approximates the CV in equation (2) by
k
X
v1 = max
rij p(xi(1) = j), 0
j=1
(
)!
k
k
X
X
l
v2 =
max
rxi p(xi(2j) = l|xi(1) = j), 0
p(xi(1) = j)
(2j)
j=1
vM ()
N
X
l=1
i1 vi .
i=1
The complexity of designing an entire strategy with this myopic approach is of order k N . This remains
considerably high, keeping in mind that we are just using a small part of the information.
One way of evaluating the myopic strategy is by Monte Carlo sampling x1 , . . . , xB p(x). For each
of the B samples the decision is given by the past outcomes, say xbi(1) = j, xbi(2j) = l, . . ., and different
samples would follow different branches of the decision tree. One could also imagine truncating the myopic
evaluation and using the (conditional) naive approach from a certain branch on. We will discuss such
approaches in more depth in the next section, when we study more refined forward selection strategies
applying the heuristics for the CV at every stage.
5. Look-Ahead and Rolling Horizon strategies
The methods considered in the previous section have the common goal of providing an approximation to
the CV. It is therefore natural to use them at different stages of the forward selection procedure. We next
propose look-ahead strategies that apply a depth n forward search combining DP with approximations of
the CV. The depth n can be chosen by the user. It will depend on the desired accuracy and on the available
time and computation power.
In our oil and gas prospect application, there is typically no need to push the forward-backward selection
procedure until the very last node. The oil and gas company is most interested in deciding the first few
prospects to drill. On the other hand, the approximations we consider apply heuristics for the CV, and in
the presence of a large and non-homogeneous number of sites, the associated sequences are not necessarily
optimal.
5.1. Look-ahead strategies
Assume that n decisions have been made and that the CV of the field is estimated by a naive or myopic
strategy. We propose to assign a large contribution to the first n < N decisions, and a smaller contribution to
the remaining N n. We approximate all CVs, and use them to run a restricted n-steps DP. The complexity
of the algorithm depends on the size n chosen in the approximation, and it is order of ( n2 )!k n/2 (N n),
when approximating the CV with the naive approach. The strategy is the following:
7
Starting point: no nodes have been observed yet: = {, , . . . }.

n steps are evaluated with DP, i.e. v() = max{v1 (), v2 (), . . .}. At each step vi () is computed
according to equation (2).
After n steps the decision vector has n observed components and N n still empty (not observed).
We define the decision vector at this stage . For instance, if N = 6 and n = 2, with observations
x2 = 2 and x6 = 1, then = {, 2, , , , 1}.
The CV v( ) is always approximated according to one of the methods introduced in Section 4:
Naive:
vN ( ) =
N
n
X
max
i=1
k
X
j=1
rij p(xi
= j| ), 0 ,
We can also fix an order for the N n prospects, based of their intrinsic values, in order to
discount the values in a particular way.
Myopic:
Similar to what was has been done in Section 4, we now approximate the CVs with a stepwise
procedure, computed in the following way:
( 3
)
X
k
v1 = max
ri p(xi(1) = k| ), 0
k=1
v2
3
X
j=1
vM ( )
N
n
X
max
3
X
rxki
(2j)
)!
p(xi(2j) = k|xi(1) = j, ), 0
k=1
p(xi(1) = j| ), . . .
i1 vi ,
i=1
5.2. Rolling horizon look-ahead strategies

We next combine different look-ahead searches and forward selection strategies. We suggest the idea
depicted in Figure 2, where one first runs a look-ahead search of depth n. Next, the best node is selected.
Given the outcome of this node, a second search of depth n is performed, and so on.
We call these strategies Depth n (in the following Dpt n) rolling horizon look-ahead (RHLA) strategies
(see Le and Day (1982) and Alden and Smith (1992)). It is interesting to note that a Dpt 0 strategy coincides
with a full naive or myopic approach (depending on the approximation chosen for the CV), while a Dpt
N 1 strategy coincides with a full evaluation of the decision tree, and therefore with the DP presented in
equation (1).
This RHLA strategy is a forward selection, but it partially accounts for future scenarios in its look-ahead
length-n DP procedure. In the RHLA strategy we explore the tree up to a certain fixed depth n, but we
draw conclusions just about the first best site. Since at every step we rerun the strategy, we can incorporate
at this step the outcome of the sample, instead of exploring all the possible combinations of evidence.
The resulting algorithm has the same computational complexity as the myopic strategy, with an additional factor due to the complexity of the look-ahead strategy in itself. In total we have a complexity
of ( n2 )!k n/2 (N n) k N . Note that this strategy can always be computed in a forward selection manner.
It is however much harder to evaluate the strategy, for instance to compute the associated value, or the
variability in the computed sequences over different outcomes.
For a small number of nodes N , one can compute the values probabilistically for different depths n
RHLA strategies. For larger dimensions we suggest to use Monte Carlo sampling to evaluate the different
strategies.
8
Figure 2: Rolling horizon look-ahead strategy, Dpt n; at every step we run a DP strategy using n sites for finding the best
node, and then we update the strategy with the outcome of that node.
We then draw samples from the graphical model with joint distribution p(x). We run the RHLA depth
n procedure to select nodes, and for each step in the forward selection we plug in the outcomes according
to the relevant sample at that node. This approach mimics what would happen in hypothetical scenarios,
and we can say that we are playing the game.
Given one realization from the graphical model, the pseudo algorithm is presented in Algorithm 1. The
algorithm presents two parts: a first one, that constitutes the core of the algorithm from where we call the
recursion, and a second one that presents the recursive function itself. In the core we find a while loop
that is necessary to terminate the algorithm when all the nodes have been explored and an if condition
that breaks the process if none of the nodes presents a positive CV. In the recursive function we have an
if condition that ensures that the correct depth is achieved, and a for loop that goes through all the
not-yet-explored nodes. When running a RHLA strategy on small examples (cfr. Section 6.1) there is the
possibility to run a RHLA for every possible evidence, spanning the whole sample space. By averaging the
revenues and costs collected through the strategy, we get a value that coincides (exact and myopic case) or
approximates (RHLA case) the estimated final value. In large examples (cfr. Section 6.2) this is not possible
and we estimate the final values through a Monte Carlo sampling procedure.
5.3. Pruning strategies
The look-ahead strategies share the idea of choosing a priori the depth n of the search tree. This choice
must be done before running an approximation. In practice, we choose n based on the available computation
time.
The problem is that we often explore branches of the decision tree that are useless for designing an
optimal strategy, and we do not privilege enough branches that can give a stronger contribution to the
value. We next design adaptive strategies based on tree-pruning, accounting for the value of the different
branches. These idea is inspired by similar ideas applied in contiguous fields, like the chess computer-based
algorithms.
We prune the branches of the tree that are very unlikely. In this way we do not have to explore all the
combinations, and we reduce the complexity of the algorithm. We define threshold parameter such that
if P ( ji ) < then v( ji ) v( i ),
and we use one of the methods described in Section 4 in order to approximate the CV.
9
Algorithm 1 Evaluating a Rolling Horizon Look-Ahead strategy of Depth n

= [, , . . . , ]
# Dynamic programming outcome vector
y=0
# Rolling horizon counter
val = 0
# Value counter
seq = [ ]
# Best sequence vector
Sample s p(x)
# Current sample
while #[i = {}] > 0 do
[v, j] = f (, 1)
if v > 0 then
# CV positivity condition
# Set sampled outcome sj at selected node j
j = sj
s
val + = y rj j
# Discounting of revenues
seqy+1 = j
# Selected node is j
else
break
end if
y++
end while
return val
return seq
function [v, j] = f (, d)
if #[i = {}] == 0 then
j=0
v=0
else if d n then
for i : i = {} do
for l = 1 : k do
[v, j] = f ( li , d + 1)
vil = ril + v
end for
Pk
vi = l=1 {p(xi = l|) vil }
end for
v = maxi {vi , 0}
j = arg max{vi }
else
j=0
nP
o
P
k
l
v() = i:i ={} max
l=1 ri p(xi = l|), 0
end if
end function
# Input: Current state, current depth

# Last iteration condition, stop
# Depth n condition, continue DP
# DP iteration at next depth level
# Reached depth n, compute naive CV
A more refined approach is to decide which branches to explore based on the value of the nodes. This
reduces the number of nodes to explore. The method can either be based on the intrinsic value of the
individual node under consideration or a look-ahead evaluation of depth 1.
The pseudo-algorithm is the following:
0 = {, , . . . , }
for i=1:N we order the segments on the basis on an approximate CV, that can be either of the
following:
Pk
Intrinsic value: v( i ) = j=1 rik p(xi = j)
10
Look-ahead Dpt 1 value:h

nP
oi
Pk
PN
k
s l
v( i ) = j=1 p(xi = j) rij + s=2 max
l=1 r(s) p(x(s) = l|xi = j), 0
Keep only the N Nprun maximum nodes with the highest values and move to the second level of
depth in a RHLA framework. For the Nprun nodes with minimum values, use the approximated values
already computed (Intrinsic or Look-ahead Dpt 1).
In practice, Nprun cannot be too small (too many paths to explore), nor too large (we risk to abandon
paths that may result being interesting). We will use the pruning strategies to speed up the computations
on large graphs.
6. Results
We first study a small BN model, where the exact DP solution is available. This allows us to compare
the suggested strategies with the exact solution. This synthetic study also anticipates the behavior of the
approximations on the BN case study from the North Sea, with 25 prospects. Finally, we analyze a MRF
model for an oil reservoir. We construct sequential exploration schemes and interpret the results of different
strategies.
6.1. A small Bayesian Network example
We are first interested in exploring the accuracy and the results of our methods on a small BN example
(Figure 3). We use a small DAG with M = 12 nodes. The nodes denoted K1, K2, P 1, P 2, P 3 and P 4
are auxiliary nodes that cannot be drilled. They are motivated by geological mechanisms that are needed
to introduce a realistic correlation structure in the network. The two K-nodes represent kitchens, i.e. areas
where the hydrocarbon (HC) generation has been or still is in place, and where the migration of HC started.
The P-nodes represent geological macro-regions able to store HC. Finally, the bottom numbered nodes,
1, . . . , 6 = N in Figure 3, are prospect nodes where the oil and gas company considers drilling wells. The
cost and revenues and marginal probabilities are summarized in Table 1. We designed the DAG to have
large variabilities both in the likelihood of finding HC and in the related volumes (revenues). The intrinsic
values, i.e. the marginal a priori values of the prospect, are all very close to 0: this makes the case harder
to solve. The conditional probabilities defined by the edges are based on geological reasoning and explained
in details in (Martinelli et al., 2011). They impose some learning in the model, once we collect evidence.
Table 1: Parameters rki for the 1st case study, and relative Intrinsic Values (marginal probability of success/failure times
revenues/costs)
k\i
r1k (dry)
r2k (gas)
r3k (oil)
p(xk = 1)
p(xk = 2)
p(xk = 3)
Intrinsic Value
1
-20
6
3
0.20
0.52
0.28
-0.04
2
-25
3
1
0.10
0.72
0.18
-0.16
3
-1
9
6
0.80
0.01
0.19
0.43
4
-15
0
7
0.30
0.02
0.68
0.15
5
- 22
4
2
0.15
0.68
0.17
-0.25
6
-8
5
1
0.34
0.53
0.13
0.05
In this small case we can compare the result of approximate strategies with the exact DP solution. The
discounting parameter is fixed, here and in the next simulations, to a realistic value of 0.99, as suggested
in (Bickel and Smith, 2006). The first comparison is presented in Table 2. Here, the result of the strategies
up to the third best choice are presented, for the naive and myopic strategies, for exact DP and for Dpt n
strategies, up to n = 4. According to the exact strategy, if oil or gas is found in the first segment chosen
11
K1
1
6
P1
P4
5
K2
P2
P3
Figure 3: BN used in the 1st case study. We indicate with the letter K the nodes denominated kitchens, i.e. zones where HC
have been generated, with letter P auxiliary nodes that are functional to establish the desired correlation structure, and with
numbers the six nodes where we can drill.
(in this case, number 6), the suggestion is to keep drilling in the same area (under P4 node) with segment
number 5. If the well reports a negative result, it makes sense to immediately explore another part of the
field. The naive approach does not take this dichotomy into account because the sequence is fixed a priori.
The myopic approach uses a different strategy for the oil/gas and the dry case, but since the depth of the
search is in this case short-sighted, the conclusion is to stop drilling immediately after a dry well.
Table 2: Results of the sequential exploration program for the 1st case study, for naive, myopic, exact and Dpt1 to Dpt 4
strategies. i(1) , i(2) and i(3) are respectively the first, the second and the third best site selected. Q means quit (the strategy).
Final value is v( 0 ).
i(1)
i(2) |xi(1) = dry
i(2) |xi(1) = gas
i(2) |xi(1) = oil
i(3) |xi(1) = dry, xi(2) = dry
i(3) |xi(1) = dry, xi(2) = gas
i(3) |xi(1) = dry, xi(2) = oil
i(3) |xi(1) = gas, xi(2) = dry
i(3) |xi(1) = gas, xi(2) = gas
i(3) |xi(1) = gas, xi(2) = oil
i(3) |xi(1) = oil, xi(2) = dry
i(3) |xi(1) = oil, xi(2) = gas
i(3) |xi(1) = oil, xi(2) = oil
Final Value
Time
Naive
3
4
4
4
6
6
6
6
6
6
6
6
6
Myopic
3
Q
2
2
Q
Q
Q
4
4
4
4
4
4
Exact
6
3
5
5
Q
2
2
4
4
4
4
4
4
Dpt1
6
3
2
2
Q
2
2
5
5
5
5
4
4
Dpt2
6
3
5
4
Q
2
2
4
4
4
3
2
2
Dpt3
6
3
5
4
Q
2
2
4
4
4
5
2
2
Dpt4
6
3
5
5
Q
2
2
4
4
4
4
4
4
0.63
0.24 sec
1.67
0.24 sec
4.960
85.6 sec
3.85
0.43 sec
4.84
3.52 sec
4.93
16.11 sec
4.957
48.22 sec
In addition to comparing strategies, we study the computational time and the final value, v( 0 ). We
notice that, despite slightly different strategies, the final values are quite close to the exact for Dpt 2 or
even Dpt 1, with a much smaller computational time. The final value reported in the table is only the
approximate value found when optimizing the strategy for the Dpt 1-4 algorithms. In practice, their value
12
will be higher, since the approximation is based on using a naive strategy at the end, whereas the algorithm
always looks ahead running new Dpt n searches. We therefore believe that the best comparison is not much
about comparing values, but more about comparing the proposed strategies on real scenarios.
Since the dimension of the problem is relatively small, we can directly span the whole sample space and
compute all RHLA strategies exactly, as anticipated in Section 5.2. This is the approach adopted in Table
3. Here we compare the evaluation of the different strategies (naive, myopic and different depths of look
ahead strategies) on the whole sample space generated by the BN of reference. We therefore test 36 = 729
combinations of evidence on the nodes of interest, and we compute the likelihood of these scenarios by
summing out the outcome at the top nodes. In this way, we can compute the average performance of the
strategies, and the related variance.
Table 3: Sequential exploration program, methods comparison following a complete RHLA procedure (Section 5.2)
Revenues Distribution
Average value
Standard deviation
Naive
0.63
12.664
Myopic
1.68
8.815
Dpt1
4.89
15.268
Dpt2
4.95
14.878
Dpt3
4.959
14.877
Dpt4
4.960
14.869
The result tells us that, when applied in practice on this simple test case, the two simple strategies
perform extremely poorly, while the look ahead strategies perform significantly better. In particular, Dpt 2
and Dpt 3 perform almost as good as Dpt 4 (which in this case corresponds exactly to the Exact Strategy),
with a significant reduction in the computational time. An interesting argument in favor of the look-ahead
strategies can also be made considering the variance. If we consider the second row of Table 3, we observe
an increasing variance between the simpler strategies and the look-ahead strategies. We first notice that the
variance of the revenues distribution under the naive strategy just reflects the variance of the marginal a
priori distribution for prospects 3, 4 and 6:
2
N
= 12.6642 =
3 X
3 X
3
X

i=1 j=1 k=1

(ri3 + rj4 + 2 rj6 r)2 p(x3 = i, x4 = j, x6 = k)
Furthermore, we can relate the low variance of the myopic strategy to a spike on the value 1, that corresponds (see Table 1) to the loss for a likely (p = 0.8) dry observation in segment 3. Since a dry outcome
at the first site in the myopic strategy would imply quitting the search, we are ultimately left with a high
number of scenarios whose revenues outcome is simply 1. If we remove these scenarios, the variance
shrinks from myopic to Dpt 1 to Dpt 4, providing another argument in favor of these strategies. A lower
variance in this case coincides with a more stable estimate and a lower risk when starting an exploration
campaign, and this can be almost as important as a high final value.
6.2. A Case Study from the North Sea

We next study a BN model developed for 25 HC prospects in the North Sea. The network (Figure 4) is
taken from Martinelli et al. (2011), and represents a model of HC fields in the Norwegian part of the North
Sea. The network includes the same characteristics as the small test study, but there are now 25 possible
drilling locations (numbered 1 through 25 in Figure 4). We use the same probability model as in Martinelli
et al. (2011). This gives the marginal probabilities in Table 4. The joint model is constructed from the DAG.
Many geological assumptions are used when building the model. In particular, gas will tend to replace oil in
the HC migration. Thus, with a single edge between two nodes in the graph we have p(xk = 1|xpa
k = 2) = 0,
p(xk = 2|xpa
k = 1) > 0, where 1 is gas and 2 is oil. Dry outcomes result from migration failures. Similar
to the previous model, the DAG has a three-level structure representing the geological mechanisms. For
decision making we are interested in the bottom nodes of the network, that represent identified prospects
whose volumes and costs are assumed known. The corresponding revenues and costs (in Million USD) are
listed in Table 4. Here, we avoid shared prospect costs that would make the computational task harder,
13
and the interpretation more difficult. In this real case, there are still some nodes where the probability of
success (and consequently the intrinsic value) may change substantially given the outcome in other nodes.
However, some nodes would be drilled or not drilled in any event, no matter the strategy.
Figure 4: Network used in the 2nd case study. In this case we have 25 drilling prospects, identified with the nodes from 1 to
25, where we can possibly drill. The BN was first presented in (Martinelli et al., 2011).
Given the BN model we are interested in identifying a drilling sequence that gives maximum profit
under some criterion. Table 5 shows the results of comparing the naive, myopic and three depth (Dpt)
level heuristic strategies. Note that final values are now quite close to each other for all the approximations
considered. The dynamic decisions depend less on the strategy than in the synthetic case in the previous
section. Still, there is a clear increase of about 3000 Million USD when using the Dpt 3 strategy rather than
the naive one. We have again run the different strategies on a number of simulated scenarios (Table 6).
Since the computational time required by the RHLA strategy is order of hours per step, we have considered
a sample size of 200 and followed the algorithm described in Section 5.2. For the same reason we will focus
from now on in a comparison between simple strategies, such as naive or myopic, and two RHLA strategies,
namely Dpt 1 and Dpt 2.
The difference is not very large, but the Dpt 1 and Dpt 2 strategies perform better than the myopic
one. In particular, Dpt 2 strategies give on average around 400 Million USD more than the myopic strategy.
It is particularly important to investigate the reason of this improvement. A first hint is given by the last
three lines of Table 6. Here we can notice that more complex strategies suggest in general to drill more
than simpler strategies. The typical case is that whenever an area is found dry, the intrinsic values for
all the segments around drop, and just long-sight strategies can look for the potential remaining values.
Nonetheless, a higher number of drilled sites translates into an effective improvement of the result just if the
newly drilled sites have a positive outcome. This is the case that we are considering, since among the 1.49
sites more drilled with Dpt 2 strategy, just 0.13 are on average dry, while an outstanding 1.36 are found gas
or oil.
Figure 5 shows what happens to all the 25 prospects when treated with different strategies. In many cases
14
Table 4: Costs, revenues, marginal probabilities and intrinsic values for the 25 sites taken into account in the 2nd case study.
Prospect k
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
r1k
-3000
-900
-2400
-1800
-600
-1500
-3600
-2100
-2700
-1200
-2400
-2700
-4500
-1800
-2100
-3600
-3300
-1200
-2100
-5400
-1800
-2400
-3000
-2400
-2700
r2k
3032
125
1094
188
594
156
406
750
2751
2751
500
125
0
188
563
31
250
688
250
969
1375
3220
156
2782
2251
r3k
2783
236
1085
377
1321
1132
3255
6934
1415
1415
4576
802
0
94
613
613
3161
8963
3349
660
3444
2264
1274
1604
1274
p(xk = 1)
0.20
0.40
0.60
0.28
0.20
0.21
0.34
0.52
0.10
0.20
0.80
0.19
0.36
0.10
0.10
0.10
0.61
0.30
0.37
0.18
0.49
0.41
0.10
0.10
0.30
p(xk = 2)
0.52
0.21
0.26
0.57
0.29
0.04
0.03
0.02
0.72
0.64
0.01
0.04
0.32
0.45
0.45
0.03
0.22
0.02
0.02
0.41
0.26
0.47
0.04
0.72
0.56
p(xk = 3)
0.28
0.39
0.14
0.15
0.51
0.75
0.63
0.46
0.18
0.16
0.19
0.77
0.32
0.45
0.45
0.87
0.17
0.68
0.61
0.41
0.25
0.12
0.86
0.18
0.14
Int. Value
1756
-242
-1004
-337
729
534
844
2107
1965
1747
-1040
123
-1620
-53
319
172
-1410
5697
1285
-312
336
783
806
2052
629
Table 5: Results of the sequential exploration program for the 2nd case study, for naive, myopic, and Dpt1-3 strategies. i(1)
and i(2) are respectively the first and second best sites selected. Final value is v( 0 ).
i(1)
i(2) |xi(1) = dry
i(2) |xi(1) = gas
i(2) |xi(1) = oil
Final Value
Time
Naive
18
8
8
8
20213
< 1 sec
Myopic
18
8
19
19
21321
< 1sec
Dpt1
15
21
22
22
21841
4.72 sec
Dpt2
22
18
18
18
22535
175 sec
Dpt3
18
24
22
22
23197
4h
Table 6: Sequential exploration program, methods comparison following RHLA procedure (Section 5.2) with a sample of 200
scenarios.
Average value
Standard deviation
Average # sites drilled
Average # sites found dry
Average # sites found gas or oil
15
Myopic
24256
13632
16.62
2.89
13.73
Dpt1
24500
12474
18.01
3.02
14.99
Dpt2
24668
12586
18.11
3.02
15.09
(segments 2, 6, 7, 8, 9, . . .) the marginal probability of a positive discovery is higher for the Dpt 1 approach
wrt to the myopic approach. It is interesting to note that, considering for example prospect 8, both the
marginal probability of a positive discovery is increased and of a negative one is decreased. This is explained
by Table 7, that tells us that we are drilling prospect 8 a smaller number of times with the Dpt 1 strategy,
but with higher efficiency. Conversely, in the cases of prospect 14, we have the same marginal accuracy
for myopic and Dpt 1 strategy, but we still have a benefit in economical terms, since we are drilling the
site a higher number of times: technically, in this case, with Dpt 1 strategy we drill prospect 14 only and
all the times that this segment is valuable. Finally, for prospect 20, we increase both the accuracy and
(substantially!) the percentage of drilled times, resulting in a strong economical return. The results are
difficult to interpret in some extreme cases, like prospect 2. Here we note how the accuracy of Dpt 1 strategy
is 100%, while the accuracy of myopic strategy is not known (both the bars are 0). This is due to the fact
that with myopic strategy we never drill prospect 2, thus we can not say anything about the accuracy of
such strategy here; on the other side, with Dpt 1 we drill it just 2% of the times, but in these cases we
always find oil or gas, therefore the accuracy boosts at 100%. This is the reason for listing P (drilled) in
Table 7 as an important diagnostic factor.
Figure 5: Probabilities of positive and negative discoveries for the 25 sites analyzed in the 2nd case study. We compare the
marginal probabilities a priori with the frequency of successes following a myopic or Dpt 1 strategy.
We finally consider (Table 8) what happens in single scenarios, i.e. what are the results when playing
the game on a few samples with different strategies (myopic, Dpt 1 and Dpt 2). We immediately see that
the myopic approach performs either brilliantly (sample 2) or extremely poorly (samples 1 and 3), while
the revenues guaranteed by the other two approaches are, in a way, more stable: this is consistent with the
type of approach, since we understand that being more long-sighted correspond to being more cautious in
our decision. The difference in the revenue variances recorded in the two samples confirms this statement,
with a strong decrease recorded when comparing myopic strategy with RHLA strategies.
If we look closer, we discover other signs that agree with this statement. The first 5 sites picked by a
myopic approach are all on the left part of the network. In simple words, we start our search from the left
side (prospect 18), and keep exploring the same side for a long period as long as the results are positive.
The Dpt 1 approach suggests to jump 3 times between the left and the right side of the network just in the
first five picks (15 and 22, then 18, then 12, then 24), even if the results are very good: this means that
while we consolidate the strength of a part of the network, we also explore if other parts of the networks
16
Table 7: Marginal probabilities of positive and negative discoveries and probability of drill for three prospects, namely prospect
8, 14 and 20. P (drilled) reports the frequency of exploration provided by myopic (Myo) or depth 1 RHLA (Dpt 1) strategy.
Prospect
P(oil/gas)
P(dry)
P(oil/gas | Myo)
P(dry | Myo)
P(oil/gas | Dpt1)
P(dry | Dpt1)
P(drilled, Myo)
P(drilled, Dpt1)
8
0.55
0.45
0.55
0.45
0.59
0.41
1
0.93
14
0.93
0.07
1
0
1
0
0.8
0.93
20
0.88
0.12
0.98
0.02
0.99
0.01
0.5
0.86
are likewise strong. This way of exploring has the further benefit, in this particular case, to allow a longer
series of straight good results (7 versus 5). The myopic strategy looks to perform better in very lucrative
scenarios: this is consistent with the theoretical definition of myopic strategy, that goes for the best first.
In an hypothetical scenario of all prospects containing oil, the myopic strategy would be difficult to beat,
and this situation is very similar to the one drawn in the second sample. In such situation an even simpler
naive strategy could beat both myopic and RHLA strategies, provided that there is not enough correlation
to confirm the nodes characterized by low probabilities and high volumes.
Table 8: Ordered list of sites chosen with myopic, Dpt 1 and Dpt 2 strategy for 3 different samples taken from the RHLA
evaluation.
1-Myo
18
19
9
24
10
8
1
22
21
5
12
20
0
0
0
0
0
0
0
0
0
0
0
0
0
16081
3
3
2
1
3
1
2
3
1
1
3
3
0
0
0
0
0
0
0
0
0
0
0
0
0
1-Dpt1
15
22
18
12
24
21
19
8
9
10
1
7
5
6
20
16
14
0
0
0
0
0
0
0
0
18126
2
3
3
3
1
1
3
1
2
3
2
2
1
3
3
3
3
0
0
0
0
0
0
0
0
1-Dpt2
22
18
15
24
21
19
8
9
10
1
7
5
6
16
12
20
14
0
0
0
0
0
0
0
0
18196
3
3
2
1
1
3
1
2
3
2
2
1
3
3
3
3
3
0
0
0
0
0
0
0
0
2-Myo
18
19
9
24
10
8
1
23
25
22
21
5
7
6
15
16
12
20
14
4
0
0
0
0
0
37293
3
3
2
2
2
1
2
3
2
3
3
3
3
3
3
3
3
3
2
2
0
0
0
0
0
2-Dpt1
15
22
18
12
24
21
19
8
9
10
1
5
7
23
25
6
20
16
14
4
0
0
0
0
0
36859
3
3
3
3
2
3
3
1
2
2
2
3
3
3
2
3
3
3
2
2
0
0
0
0
0
2-Dpt2
22
18
15
24
21
19
8
9
10
1
5
7
23
12
25
6
20
16
14
4
0
0
0
0
0
37087
3
3
3
2
3
3
1
2
2
2
3
3
3
3
2
3
3
3
2
2
0
0
0
0
0
0
3-Myo
18
8
24
1
23
22
5
25
10
9
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
-2455
1
1
3
1
3
1
1
3
3
3
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
3-Dpt1
15
22
18
24
10
12
8
1
5
9
7
23
6
20
16
25
14
0
0
0
0
0
0
0
0
-1208
2
1
1
3
3
3
1
1
1
3
3
3
3
1
3
3
2
0
0
0
0
0
0
0
0
3-Dpt2
22
18
24
10
15
8
1
9
5
7
23
6
16
25
12
20
14
0
0
0
0
0
0
0
0
-1146
1
1
3
3
2
1
1
3
1
3
3
3
3
3
3
1
2
0
0
0
0
0
0
0
0
In summary we learnt that there are clear differences in the suggested drilling strategies for the naive,
myopic and Dpt n computations. A myopic strategy gives a large improvement over the naive strategy in our
network, and this will always be the case as long as the prospects are dependent and not obviously profitable
or unprofitable. The extra gain from running Dpt n strategies is in this 25 prospect case seen as a larger
payoff in money for the computing time spent. The Dpt n strategies also suggest other drilling locations. In
a practical setting, our recommendation is to run a Dpt n search with as large n as computationally feasible.
Note that this can be done stepwise. In many situations we only need to identify the first prospect, and can
wait for the result there before computing the next. This is the practical exploration scenario a petroleum
company faces.
17
6.3. MRF case study

In the third application we apply our sequential exploration technique on a larger dataset, where the
current knowledge consists of geological knowledge combined with seismic data. The data and the case study
are explained in Bhattacharjya et al. (2010). The MRF model has 3-colors, where the three distinctions of
interest represent respectively oil saturated sand (xi = 1), brine saturated sand (xi = 2) and shale (xi = 3).
We use a lattice representation of the field, with 20 5 cells, i.e. M = N = 100.
The prior model is a categorical first-order MRF (Besag, 1974):
N
X
X
p(x) exp
I(xi = xj ) +
i (xi ) ,
ij
i=1
where i j denotes the sum over all neighboring lattice nodes (north, east, south, and west). The parameter
imposes spatial interaction. The i terms are set from a priori geological knowledge (Bhattacharjya et al.,
2010). We work with a highly correlated MRF (=0.8).
The seismic data y are incorporated in the MRF model x through a Gaussian likelihood model (Eidsvik
et al., 2004). At each cell bivariate seismic data, shown in Figure 6 are modeled by:

1 (xj )
0.062 0.007
p(yj |xj ) N
,
2
2 (xj )
0.007 0.17
where: 1 = (0.03, 0.08, 0.02) and 2 = (0.21, 0.15, 0).
The posterior model is defined by:
p(x|y) p(x)
100
Y
j=1
p(yj |xj ).
This posterior is a MRF with new i terms which now also depend on the data values.
As was done in Bhattacharjya et al. (2010), we assign a fixed cost of 2 Million USD for drilling a dry
well (state 2 or 3), while we have a potential revenue of 5 million USD when finding an oil saturated sand
(state 1). Before drilling we have the situation represented in Figure 6. In the top row we see the bivariate
seismic data, in the bottom row we see the the prior geological knowledge and the posterior oil saturated
sand probability.
The combinatorial complexity prevents us from running a full search, therefore we try different levels
of approximations, from the myopic strategy to more complex depth searches. We present in Figure 7 the
results of myopic, Dpt 1 and Dpt 2 strategies. While the first myopic strategy reproduces the same pattern
that we observe in the posterior probability of oil (bottom right, Figure 6), the second Dpt 1 strategy shows
a different pattern. The sites on the eastern part of the basin, those that get the higher expected revenues
(due to a strong prior probability of oil sand), are not anymore selected in the first step, because they are
surrounded by sites with low profitability. On the other hand, the central sites, whose profitability was not
that high, but overall good over a large area, are privileged by a Dpt 1 strategy. The same behaviors appear
in the bottom part of Figure 7, that report the best first and second choice for Dpt 2 strategy. We can
further note that the expected final values increase with more complex strategies.
For a petroleum company that wants to explore a reservoir zone, we expect the drilling strategy to
depend heavily on the amount of data available (seismic data and well data in the neighborhood of the
reservoir), and the cost of establishing new infrastructure. In this example we built the first element into
the MRF model and the second as part of the case-specific utility function. In our situation, the Dpt n
strategies clearly select different drilling locations than the myopic approach. This kind of information is
useful in an appraisal stage of a reservoir unit
7. Closing remarks
The paper proposes a new approximate solution to sequential decision making. The approximations
apply heuristic procedures to estimate the optimization function at different stages of the algorithm. Pruning
18
Figure 6: Initial conditions of the MRF described in Section 6.3. Top left: reflectivity seismic data. Top right: amplitude seismic
data. Bottom left: prior geological knowledge. Bottom right: Probability of oil saturated sand with interaction parameter
= 0.8.
strategies are also proposed in order to speed up the computation by cutting the less valuable branches of
the decision tree.
The methodology is applied to case studies from the petroleum industry. First, a BN model for 25
prospects in the North Sea (Martinelli et al., 2011) is solved. Second, a MRF with 100 lattice cells for
a local reservoir is studied. In both cases, we construct approximate drilling sequences. We show how
sequential decision making, coupled with a statistical model for the dependence of the field, can yield
strategies very different from those based on independent or myopic searches.
We recommend running a strategy of depth n, where n is as large as computationally feasible. In practice a
petroleum company would often wait for the outcome of the first well(s) to continue its exploration strategy.
It is also possible to run different depth searches and see if results are very dissimilar. In practice the
petroleum company can test the depth n strategies over different utility functions, various kinds of risk
behavior, and a range of cost and revenue inputs. This means only minor edits to inputs parameters in our
implemented algorithms, and provides helpful guidelines when selecting the final exploration policy.
The applications do not limit the scope and the merit of the developed algorithms. One can use the
methodology to other selection problems on graphical models. Nodes could for example correspond to
clinical tests, in a problem where the practitioners make sequential decisions. Also, generic variable selection
problems or design of experiments for graphs could be envisioned utilizing the same instruments.
We believe that there is large potential for interplay between operational research and recent development
for computing multivariate statistical models. The current paper is just one example. Here, the search is
built on heuristic strategies, and we have made no attempts to justify the approximation as the optimal
solution. It would be interesting to study these problems from a more theoretical perspective, merging
knowledge from both operations research, decision theory and statistics.
8. Acknowledgments
We thank the Statistics for Innovation (SFI2 ) research center in Oslo, that partially financed GMs
scholarship through the FindOil project. We acknowledge David Brown and Jim Smith, Duke University,
19
Figure 7: Best 1st and 2nd sites using myopic (top), Dpt1 (center) and Dpt2 (bottom) strategies. The colors correspond to
vi () under the three strategies, where vi () (cfr. equation 2) represents the CV of the chosen strategy, given that we start
drilling at prospect i.
for interesting and useful discussions on this topic.
20
Alden, J. M., Smith, R. L., 1992. Rolling horizon procedures in nonhomogeneous markov decision processes. Operations
Research 40, S183S194.
Bellman, R., 1957. Dynamic Programming. Princeton University Press.
Benkherouf, L., Bather, J., 1988. Oil exploration: Sequential decisions in the face of uncertainty. Journal of Applied Probability
25 (3), 529543.
Bertsekas, D. P., Tsitsiklis, J., 1996. Dynamic Programming. Athena Scientific.
Bertsimas, D., Demir, R., 2001. An approximate dynamic programming approach to multidimensional knapsack problems.
Management Science 48 (4), 550565.
Besag, J., 1974. Spatial interaction and the statistical analysis of lattice systems. Journal of the Royal Statistical Society, Series
B 36, 192236.
Bhattacharjya, D., Eidsvik, J., Mukerji, T., 2010. The value of information in spatial decision making. Mathematical Geosciences
42 (2), 141163.
Bickel, J., Smith, J., 2006. Optimal sequential exploration: A binary learning model. Decision Analysis 3 (1), 1632.
Bickel, J., Smith, J., Meyer, J., 2008. Modeling dependence among geologic risks in sequential exploration decisions. SPE
Reservoir Evaluation & Engineering 11 (2), 352361.
Bollapragada, S., Morton, T. E., 1999. Myopic heuristics for the random yield problem. Operations Research 47 (5), 713722.
Cowell, R., Dawid, P., Lauritzen, S., Spiegelhalter, D., 2007. Probabilistic Networks and Expert Systems. Springer series in
Information Science and Statistics.
Eidsvik, J., Avseth, P., Omre, H., Mukerji, T., Mavko, G., 2004. Stochastic reservoir characterization using pre-stack seismic
data. Geophysics 69, 978993.
Eidsvik, J., Bhattacharjya, D., Mukerji, T., 2008. Value of information of seismic amplitude and csem resistivity. Geophysics
73 (4), R59R69.
Feldmann, R., Mysliwiete, P., Monien, B., 1994. Studying overheads in massively parallel min/max-tree evaluation. Proceedings
of the sixth annual ACM symposium on Parallel algorithms and architectures, 94103.
Frazier, P., Powell, W. B., Dayanik, S., 2008. A knowledge gradient policy for sequential information collection. SIAM J.
Control Optim. 47 (5), 24102439.
Frazier, P. I., Powell, W., 2010. Paradoxes in learning and the marginal value of information. Decision Analysis 7 (4), 378403.
Gittins, J., 1979. Bandit processes and dynamic allocation indices. Journal of the Royal Statistical Society, Series B 41, 148177.
Goel, V., Grossmann, I., El-Bakry, A., Mulkay, E., 1979. A novel branch and bound algorithm for optimal development of gas
fields under uncertainty of reserves. Computers and Chemical engineering 30, 10761092.
Kokolis, G., Litvak, B., Rapp, W., Wang, B., 1999. Scenario selection for valuation of multiple prospect opportunities: a montecarlo simulation approach. SPE paper 52977, presented at the SPE Hydrocarbon Economics and Evaluation Symposium,
Dallas, TX, 20-23 March 1999.
Lauritzen, S. L., Spiegelhalter, D. J., 1988. Local computations with probabilities on graphical structures and their application
to expert systems. Journal of the Royal Statistical Society, Series B 50, 157224.
Le, K., Day, J., 1982. Rolling horizon method: A new optimization technique for generation expansion studies. IEEE Transactions on Power Apparatus and Systems 101 (9), 31123116.
Martinelli, G., Eidsvik, J., Hauge, R., Drange-Forland, M., 2011. Bayesian networks for prospect analysis in the north sea.
AAPG Bulletin 95 (8), 14231442.
Miller, A., 1975. The value of sequential information. Management Science 22 (1), 111.
Nemhauser, G., 1966. Introduction to dynamic programming. Wiley.
Pearl, J., 1984. Heuristics: Intelligent Search Strategies for Computer Problem Solving. Addison-Wesley.
Powell, W. B., 2008. Approximate dynamic programming: Lessons from the field. Proceedings of the 2008 Winter Simulation
Conference, eds. S. J. Mason and R. R. Hill and L. Moench and O. Rose.
Reeves, R., Pettitt, A., 2004. Efficient recursions for general factorisable models. Biometrika 91 (3), 751757.
Shannon, C., 1950. Programming a computer for playing chess. Philosophical Magazine, Series 7 41 (314), 256275.
Smith, J. ., Thompson, R., 2008. Managing a portfolio of real options: Sequential exploration of dependent prospects. The
Energy Journal, International Association for Energy Economics 29, 4362.
VanWees, J., Mijnlieff, H., Lutgert, J., Breunese, J., Bos, C., Rosenkranz, P., Neele, F., 2008. A bayesian belief network
approach for assessing the impact of exploration prospect interdependency: An application to predict gas discoveries in the
netherlands. AAPG Bulletin 92 (10), 13151336.
Weber, R., 1992. On the gittins index for multiarmed bandits. The Annals of Applied Probability 2 (4), 10241033.
21
Paper III
Strategies for petroleum exploration based on Bayesian Networks:
a case study
G. Martinelli, J. Eidsvik, K. Hokstad and R. Hauge
SPE paper 159722, Submitted for publication, 2012.
SPE 159722
Strategies for petroleum exploration based on Bayesian Networks: a case study
Gabriele Martinelli, Jo Eidsvik, Ketil Hokstad, NTNU, Ragnar Hauge, NR
Copyright 2012, Society of Petroleum Engineers
This paper was prepared for presentation at the SPE Annual Technical Conference and Exhibition (ATCE) to be held 8 - 10 October 2012 in San Antonio, Texas.
This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents of the paper have not been
reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect any position of the Society of Petroleum Engineers, its
officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written consent of the Society of Petroleum Engineers is prohibited. Permission to
reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.
Abstract
The paper presents a new approach for modeling important geological elements, such as reservoir, trap and source, in a unified
statistical model. This joint modeling of these geological variables is useful for reliable prospect evaluation, and provides a framework
for consistent decision making under uncertainty. A Bayesian Network, involving different kinds of dependency structures, is used to
model the correlation within the various geological elements, and to couple the elements. Based on the constructed network, an optimal
sequential exploration strategy is established via dynamic programming. This strategy is useful for selecting the first prospect to explore,
and which decisions to make next, depending on the outcome of the first well. A risk neutral decision maker will continue exploring
new wells as long as the expected profit is positive.
The model and choice of exploration strategy is tailored to a case study represented by five prospects in a salt basin, but it will also
be useful for other contexts. For the particular case study we show how the strategy clearly depends on the exploration and development
cost, and the expected volumes and recovery factors. The most lucrative prospect tends to be selected first, but the sequential decisions
depend on the outcome of the exploration well in this first prospect.
1. Introduction
When deciding whether to explore and later to produce hydrocarbons (HC) at a prospect, it is important to consider any kind of
information that could help us in make better decisions. When doing this, it is crucial to recognize how prospects are correlated within
the same field, and to model this dependence.
Today, more fields are characterized by poor quality of the traditional sources of information, such as seismic surveys. In the basin
that we will discuss in this paper, the main problem is the presence of huge salt formations, called diapirs. One believes that these salt
domes may hide potential HC traps, but the real presence is difficult to localize and discover, because such salt structures impose seismic
shadow zones, where interpretation of reflectors is not possible.
A second distinction this field under our consideration shares with other HC fields nowadays is the presence of several smaller
prospects, whose risk evaluation is uncertain. For this reason it gets important to provide alternative risk assessments, that account for
all the available geological and geophysical data, and introduce realistic correlation mechanisms among the prospects. Some main ideas
in prospect risk evaluation were proposed in Rose (2001) and Suslick and Schiozern (2004).
The current work borrows the original idea from Martinelli et al. (2011b), where a Bayesian Network (BN) model was proposed
to model geological correlation between prospects. Similar ideas were developed in VanWees et al. (2008). In Martinelli et al. (2011b)
the modeling was inspired by an underlying spatial model and was restricted to a single geological element, the source presence. In
the current work we model all the elements contributing to the risk evaluation: source, trap and reservoir. They must all be present to
have a producible HC prospect. In the modeling we use geological and geophysical data as covariates, and we rely on expert geological
understanding of the basin. Since the case study is anonymized, we do not give references about the geological details. Our BN model
is built for this case study, but it could be edited without much work to apply in other geological settings.
The ultimate goal here is to show how a dependent prospect model provides a framework for aggregate decision making. We
propose a sequential discrete utility function that incorporates the main costs and revenues that characterize an exploration campaign,
and we show how the possible choices evolve with different input scenarios and are strongly dependent. The ideas rely on a Dynamic
Programming (DP) formulation of the problem and they have been previously developed in Bickel and Smith (2006), Bickel et al. (2008)
and Martinelli et al. (2011a). Similar concepts have also been proposed in Cunningham and Begg (2008) and Smith and Thompson
(2008), for determining the best sequential exploration strategy, but not for the current situation with BN models for various geological
elements and multiple prospects.
The paper evolves as follows: in Section 2 we describe the main geological features of the basin, in Section 3 we present the main
ideas behind the BN model that we use for modeling prospect dependencies, in Section 4 we describe the DP procedure for optimal
SPE 159722
sequential exploration, and finally in Section 5 we show and discuss the main results.
2. The geology of the basin

The basin that we use throughout this case study is a late Paleozoic intra-continental syn-rift salt basin, characterized by a number
of salt diapirs and pillow-like salt structures. Evaporites could represent the lower part of these salt structures; they have been deposited
in different periods of time on an area larger than the present basin.
Several episodes of salt reactivation causing a pronounced subsidence of the sediments, have been postulated from Jurassic to recent
time, with a major episode occurring in the Tertiary. Salt movements have been explained as a consequence of buoyancy and density
contrast, or of tectonic forces. The basin was subsequently uplifted and deeply eroded in Pliocene-Pleistocene era.
Two possible source rocks have been identified in the basin. The first one is partially immature, but it is believed that at least the
most organic-rich shales are able to generate and expel liquid HC. The second formation is generally mud-rich and mature to overmature.
The most prolific reservoir intervals come from late Triassic to upper Jurassic, with porosities up to 35 % and good permeabilities.
For the basin we consider, exploration started in the 1980s, but only few exploration wells have been drilled so far. The petroleum
system has already proven to be in place, after drilling two preliminary exploratory wells nearby. The first well found just salt, entering
the bulk of one of the domes, while the second one reached a sandstone reservoir and proved the presence of oil and gas. It is important to
notice that for the particular feature of this field and for its remote location, just the presence of oil justifies a possible future development
of the field, while the presence of gas alone would not be sufficient to recommend the development. Gas is known to be present in the
entire area, and the estimated amount in each of the 5 potential prospect will be given in Section 4.2 .
From SW to NE, the five potential prospects are identified by the first five letters of the latin alphabet, A, B, C, D and E. There are
other minor areas of interest outside the main field, where the presence of gas has been proven, but we will not consider these areas here.
Several stratigraphic analyses, basin models and seismic interpretations have helped in understanding the geological history of the
area. They further provide useful inputs for quantitative evaluation of the probabilities for the key geological elements to be in place.
These key elements can be local factors or share interesting patterns worth accounting for.
We list here the main elements that emerge from the analysis. For each element, we consider how it correlates between prospects.
Trap geometry (TG): the correct identification of the trap geometry is a key element for understanding the history of the
petroleum accumulation through geological time. It is important to consider that after periods of deposition and uplifting, in
late Pliocene/Pleistocene the area underwent a number of periods of glaciations with consequent extensive erosion. This removed
a number of previous strata, leaving the area around the diapirs heavily tilted and truncated, and therefore difficult to reconstruct
precisely. In spite of a common depositional environment and formation mechanism for all the diapirs, this factor is extremely
local.
Trap seal (TS): the sealing mechanism may be represented by the diapir itself that acts as a stratigraphic trap, or by another shaly
cap rock. A minor correlation effect is considered for this element.
Reservoir presence (RP): the presence of an effective reservoir can be seen through seismic surveys. The depositional environment is believed to be common for the whole area, and it is a shallow marine / estuary environment, characterized by good sorting
and high porosities. The actual depth (and therefore the main properties) of the reservoir varies in the basin, but the common
elements should be taken into account. Most reservoirs are sandstone deposited in the Carnian age (upper Triassic).
Producibility (PR): the effective potential of the field, in terms of HC quantity that can be extracted, is related to some properties
of the reservoir rock like porosity and permeability. The porosity of the sandstone is enhanced and preserved by the presence of
chlorite, that guarantees levels of porosity higher than 20% in the whole field. The AVO anomalies show a spatial trend with a
maximum between prospects C and D.
Source presence (SP): the presence of the source rock and its timing relative to the other petroleum system elements is important.
It is well known that the main source rock in the basin is the Triassic to Late Jurassic Formation, but it is undecided whether the
trap formation happened before or after the complete maturation of this source rock. Geologists assume that the ratio between the
drainage area and diapir area is a good predictor for this element.
Source migration (SM): the local migration from the source rock to the reservoir rock is mainly vertical. Therefore it is a local
element, with scarce correlation.
Source abundance (SA): this element discriminates between a geological discovery and a commercial discovery. As for the
producibility, this SA element is connected with the AVO anomalies, that show a clear spatial dependency pattern.
We are interested in distinguishing between situations where the trap is filled to spill, and situations where the trap mechanism is
in place, while the source have generated just a part of the possible HC. It is useful here to define Geological and Commercial aspect:
both the situations described above show that the geological elements needed for HC generation are in place, but only the first one has
SPE 159722
immediate commercial impact. For this reason we have introduced a non-physical element in the list above, namely the SA feature, that
controls the transition between Geological and Commercial probability of discovery.
As we can see from this analysis, a very important element that is undecided in the evaluation of this field is the timing effect: if the
deposition of all the characteristic elements has not occurred in the right order, a premature leaking of the HC could happen.
We show the marginal probability of the different elements in Table 1. These numbers come from a previously done prospect
R
evaluation of the risk connected to different geological elements, conducted through the commercial software GeoX .
In the current work we integrate specific prospect dependencies to provide a better joint model for the basin.
Geological element
Trap geometry
Trap seal
Reservoir presence
Producibility
Source presence
Source Migration
Geological prospect prob.
Source Abundance
Commercial prospect prob.
A
0.7
0.5
0.8
0.9
0.6
0.8
0.121
0.8
0.097
B
0.6
0.5
0.8
0.9
0.7
0.7
0.106
0.8
0.085
Prospects
C
0.6
0.4
0.8
1
0.6
0.5
0.058
0.9
0.051
D
0.8
0.9
0.85
1
0.8
0.8
0.392
0.9
0.349
E
0.7
0.6
0.8
1
0.6
0.9
0.181
0.9
0.162
Table 1: Marginal probabilities of success per geological element, in the five prospects under consideration.
A brief explanation of the risk evaluation connected to the different prospects in the basin is given here. Prospect D consists of three
smaller diapirs within one closure. Two of the diapirs are connected through a faulted ridge. The Carnian reservoir sandstone is not at
any risk of reaching seafloor in these diapirs, and the top of the structures seem quite undisturbed (limited faulting). Prospect D has a
fairly sized drainage area for HC and a high volume estimated from seismic data. In prospects B and C, seismic data have revealed heavy
faulting caused by the diapir(s), resulting in a high risk of failure due to trap geometry and seal. Prospect E is believed to be potentially
the second best reservoir in the area, in terms of probability of discovery.
The Drainage Areas of the five prospects under consideration, the areas of the prospect itself, and the AVO class 4 anomalies are
shown in Table 2.As we can see, there are maximum areas between prospects C and D.
Drainage Area (km2 )

Prospect Area (km2 )
Ratio
AVO class 4 anomalies
A
66
11.8
small
poor
B
99
8.7
medium
poor
Prospects
C
218
18.4
medium
rich
D
224
46
small
rich
E
171
10
large
rich
Table 2: Drainage areas, prospect areas and AVO anomalies in the five prospects under consideration.
3. Modeling with Bayesian networks

The conditional modeling framework used in a BN assigns relationships between variables at the local level. It is helpful to view the
dependence structure by graphical representations. The BN model consists of nodes representing random variables and edges indicating
the interaction between node variables. In our case the node variable have a geological interpretation, related to prospects. In the models
we consider here we use directed edges between nodes, and the resulting graph is known as a directed acyclic graph (DAG).The conditional independence structure imposed by edges simplifies the model specification and introduces sparseness or parsimony Lauritzen
and Spiegelhalter (1988).
The joint model for a DAG is defined by the marginal distributions for the parent nodes and the conditional probabilities for the
children. The joint model for variables (x1 , . . . , xn ) is then
n
pa
p(x1 , . . . , xn ) = p(xi |xi ),
(1)
i=1
pa
where xi denotes the variables at the parent nodes of node i, and this parent set is empty for the top nodes of the DAG. The full
conditional distribution of xi , given the outcomes at all other nodes, xi = (x1 , . . . , xi1 , xi+1 , . . . , xn ), only depends on the children, the
SPE 159722
parents, and the other parents of its children. This is formalized through the notion of cliques, see Cowell et al. (2007), and utilized
in the computation of DAGs. The updating of node probabilities relies on the concepts of marginalization and the use of conditional
independence structure. For large scale networks it is useful to incorporate a number of computational tricks, and this can be effectively
done through existing software. In our implementation we use the Bayes Net Toolbox (BNT) package, see Murphy (2001).
3.1 A motivating example with two prospects We will start by an illustrative BN example (Figure 1) with three nodes. Nodes 2 and 3
are prospects where we consider oil exploration. Node 1 represents a geological feature, which we cannot observe directly, but the edges
going from 1 to 2 and 3 mimic a causal geological mechanism. Note that the two prospects will be dependent, but given the outcome of
node 1, they are independent since there is no direct edge between node 2 and 3.
2
1
3
Figure 1: Motivating example, two prospects (nodes 2 and 3) correlated through a common parent (node 1).
With binary variables xi {0, 1}, i = 1, 2, 3, where xi = 0 corresponds to failure (no HC), while xi = 1 is a success (HC presence),
the model is completely specified by assigning p(x1 = 1), p(x2 = 1|x1 = 0), p(x2 = 1|x1 = 1), p(x3 = 1|x1 = 0) and p(x3 = 1|x1 = 1).
This gives a unique joint distribution p(x1 , x2 , x3 ) = p(x1 )p(x2 |x1 )p(x3 |x1 ) for all outcomes at the three nodes.
The marginal probabilities of success at prospects 2 and 3 are obtained by summing over all states in node 1:
p(x2 = 1) =
p(x2 = 1|x1 = j)p(x1 = j),
j=0
p(x3 = 1) =
p(x3 = 1|x1 = j)p(x1 = j).
j=0
We will next discuss a special case with p(x2 = 1|x1 = 0) = p(x3 = 1|x1 = 0) = 0, i.e. a success cannot happen at leaf nodes when the
parent node is a failure. This gives a simple interpretation of the BN probabilities. The geological mechanism prevents HC to enter node
2 and 3, when the parent node is dry. The opposite can occur, and this is referred to as a local failure probability: p(x2 = 0|x1 = 1) =
1 p2 and p(x3 = 0|x1 = 1) = 1 p3 . We set p(x1 = 1) = p1 . The marginal probabilities at prospects become p(x2 = 1) = p2 p1 and
p(x3 = 1) = p3 p1 .
We will briefly show the concept of evidence propagation for BNs by evaluating the effect of drilling at node 2. This is useful
in prospect evaluation, and illustrates an important feature of BN models. When drilling at node 2 we of course hope to discover
HC, and start exploiting this resource. At the same time the evidence we obtain by drilling is valuable for learning what chances we
have of finding HC at the other prospect, since the two are dependent through the common parent node 1. Assume we drill at node 2
and discover hydrocarbons (x2 = 1). This entails that x1 = 1, because otherwise the prospect could not have contained HC. Then, the
conditional probability at node 3 is p(x3 = 1|x2 = 1) = p(x3 = 1|x1 = 1) = p3 . This is a 1/p1 - times increase in probability compared
with the marginal above.
The situation with a failure when exploring node 2 is a little harder, because this can happen for node 1 HC or not. We have to
marginalize over both these events for the top node:
p(x3 = 1|x2 = 0) =
=
j=0
j=0
p(x3 = 1, x1 = j|x2 = 0) = p(x3 = 1|x1 = j)p(x1 = j|x2 = 0)
p(x3 = 1|x1 = 1)p(x1 = 1|x2 = 0) = p3 p(x1 = 1|x2 = 0),
where the latter term must be calculated by Bayes rule:

p(x1 = 1|x2 = 0) =
In summary we get:
p(x1 = 1, x2 = 0)
p(x2 = 0|x1 = 1)p(x1 = 1) (1 p2 )p1
=
=
.
p(x2 = 0)
p(x2 = 0)
1 p2 p1
p(x3 = 1|x2 = 0) =
p3 (1 p2 )p1
,
1 p2 p1
SPE 159722
where the numerator can be interpreted as HC at node 1, a failure of node 2, and HC at node 3. The denominator is the marginal
probability of no HC at node 2.
Assuming p1 = 0.2, p2 = 0.5, p3 = 0.9, we get a marginal of p(x3 = 1) = 0.18, and the conditionals depending on the evidence
at node 2 are p(x3 = 1|x2 = 1) = 0.9 and p(x3 = 1|x2 = 0) = 0.1. Consider instead drilling in node 3, and study how this influences
the probability of HC in the other prospect. The marginal p(x2 = 1) = 0.1, while the conditionals are p(x2 = 1|x3 = 1) = 0.5 and
p(x2 = 1|x3 = 0) = 0.012. Since p3 > p2 , a failure in node 3 indicates with more certainty that node 1 is a failure.
We return to this illustrative 3-node DAG when we discuss exploration strategies below.
3.2 The Bayesian Network representation of the basin We propose different network structures for the components of the petroleum
system. With these models we describe different kinds of dependence for various building blocks (sub-networks) of the large BN:
Independent: Some features are essentially independent, therefore there is no point in linking different nodes. See Figure 2 (left).
We propose this structure for the TG and SM.
E
CP
L2
L1
C
B
A
B
A
E
D
B
A
Figure 2: Independent Network (left), Common Parent or Counting Network (center) and Multi-Level Network (right): three possible ways to
model mutual interaction among prospects.
Common parent counting network: Some features show a common cause structure, i.e. they are affected by the presence of
a common element, as we can see in Figure 2 (center). We consider a common parent with multiple states. In this way, the more
positive answers we get from the children, the more likely to get a success in the other nodes. The physical reasoning in this case
is that we are trying to model a phenomenon whose success rate is uncertain, and the more evidence we collect, the more certain
this rate becomes. We propose this structure for the TS, RP and SA.
Multi-Level network: Some features may depend on other causal geological mechanisms. We model this phenomenon by
grouping children node to a few parent nodes. The resulting graph is shown in Figure 2 (right). This network introduces asymmetry
in the correlation structure. It works in the following way: an observation in a node belonging the second level (D or E) affects
the upper level (A, B or C) in the very same way as an observation coming from the first level itself. This effect does not work in
the other direction, since the propagation from L1 to L2 can fail. We propose this structure for the PR and the SP.
3.3 Parameter setting The probabilities associated with the edges of a BN can be summarized by a conditional probability table (CPT)
having the parent node outcomes as rows and the children node outcome in the columns. We next discuss our CPT modeling assumptions
about the different networks.
Independent network: Here there are no edges, and we just fix the marginals equal to the values provided by GeoX.
Counting network: When the parent node has 4 states, we have potentially to fix 4*5 + (4-1) = 23 parameters. In order to reduce
the number of free parameters, we propose a simple parametrization with fixed prior parameters p(PN = 0) = 0 , . . . , p(PN =
3) = 3 for the parent node, and conditional parameters 0 , . . . , 3 related to the success at the child nodes. To ensure i i = 1,
and fulfill the required marginals pi in each of the i = 1, . . . , 5 children nodes, we need that T > maxi pi . In this way we can
compute the marginal success parameters mi = pi /(T ), and finally write the CPT as in Table 3.
Our geological understanding of the process leads us to the conclusion that our confidence in the presence/absence of this structure
is increased/decreased when new evidence appear, and that all the prospects have an equal importance in contributing to this
process.
Multilevel network: Here we relate the geological feature of interest with a covariate that is correlated with our feature. We group
the levels of this covariate in two or more Level Nodes (L1 , L2 , . . . ), whose CPT is of the kind specified in Table 4:
Likewise, for the Children Nodes Ci , we have the CPT in Table 5.
SPE 159722
PN \ Ci
0
1
2
3
0
1 0 mi
1 1 mi
1 2 mi
1 3 mi
1
0 mi
1 mi
2 mi
3 mi
Table 3: CPT for Counting Network; parameters mi and i are chosen in order to match the marginal probabilities, as described in Section 3.3 .
L1 \ L2
0
1
0
1
1 1
1
0
1
Table 4: CPT for Multi Level Network, from first to second level.
LN1 \ Ci
0
1
0
1
1 i
1
0
i
Table 5: CPT for Multi Level Network, from level nodes to children nodes.
We have now to set 5+2=7 parameters. We fix first (P(L1 ) = 1) and 1 (P(L2 = 1|L1 = 1)), and then we set the five i values to
match the marginals in the following way: i = pi / if node i belongs to the first level, or i = pi /(1 ) if node i belongs to the
second level. We choose multi level networks for modeling variables that have a clear trend, such as SP and PR. In the first case
we select 3 classes, in accordance with the drainage area ratio discussed above. In the second case we select two classes, based on
the AVO anomalies described in Table 2.
The resulting network is presented in Figure 3. Here, we can see on the top part the two subnetworks for RP and PR, on the left the
Trap (TR) subnetwork, made by TS and TG, on the bottom the Source (SO) subnetwork, made by SP and SM, and on the right the SA
subnetwork.
The different subnetworks are then linked together through the bottom nodes A, B, C, D and E. The bottom nodes are not binary as
all the other nodes of the network, but they have three states, namely dry, partial oil and commercial oil. The positive/negative outcome
of the SA nodes controls this partition.
Since the basic assumption of the model is the independence of the subnetworks, the CPT of one of the bottom nodes, say A, given
its parents, is a simple table structured in the following way: the probability of the states partial oil and commercial oil are non zero only
when all the parents SO, TR, RP and PR are 1 and the final parent SA is 0 or 1 respectively. This means that unless all the major factors
are 1, there is no possibility of a geological or commercial discovery, but it is of course possible that the results of an exploration are not
limited to a dry/oil observation. In the next section we will discuss the different observables, i.e the possible outcomes of an exploration
well, and how the network that we have built is updated as a consequence of such observables.
3.4 Illustration of evidence propagation The sample space refers to the possible observables or evidences that we can collect from
drilling one (or more) exploration wells. The outcome is informative of HC or not at the bottom level (A, B, C, D and E), but it also
provides information about the source, trap or reservoir, since the presence or absence of some parameters may be hidden by the absence
of other parameters. We classify six types of possible evidence, spanning the whole sample space:
1. We get a positive (oil) evidence in the bottom node; in this case all the variables are in place, and we can consequently update all
its parents. An example is given in Figure 4 (left).
2. We get a partially positive (oil) evidence; in this case all the variables are in place, but a study of the well log reveals that the
formation is not filled to spill, i.e. lacking confidence of a commercial exploitation. In this case we update positively all the parent
nodes, except the SA node. An example is given in Figure 4 (right).
3. We get a negative (dry) evidence and we observe no reservoir. I.e. from the well log there is no evidence of a layer that works as
reservoir. In this case we assign a 0 evidence to the corresponding RP node, while we cannot say anything about the source and
trap. An example is given in Figure 5 (left).
4. We get a negative (dry) evidence, and we observe a reservoir, but no traces of HC in the log. Here, we are in the situation where
we can make the hypothesis that there is a failure in the source: this failure can be due to SP or SM. In this case we assign a 1
evidence to the RP node, a 0 to the SO node, while we cannot say anything about the trap. An example is given in figure 5 (right).
SPE 159722
RP
PRl2
prA
rpA
PRl1
prB
prD
prC
rpB
rpC
tgD
saD
D
trD
tgC
tsC
saC
SA
trC
tgB
tsB
saB
B
trB
tgA
tsA
saE
E
trE
tsD
TS
rpE
prE
tgE
tsE
rpD
saA
A
trA
soA
smA
soB
smB
spA
soC
smC
spB
SPl3
soD
smD
spC
SPl2
soE
smE
spD
spE
SPl1
Figure 3: BN representation of the basin, result of different modeling choices for different geological elements. On the right the trap network,
on the top the producibility and the reservoir presence network, on the left the source abundance network, on the bottom the source network.
5. We get a negative (dry) evidence, and we observe a reservoir and traces of HC in the log. Here, we conclude that the weak link is
the trap. Possibly, there was HC that remained trapped for some time and then leaked out completely. Therefore we assign 1 to
both RP and SO nodes and 0 to TR node.
6. We get a negative evidence, where all the critical factors seem to be in place. However,there is failure in the PR factor, which
means a bad reservoir quality and the discovery is very hard to exploit.
The sample space is summarized in Table 6, and the associated probabilities of the six outcomes, for each of the five prospects, are
shown in Figure 6. The marginal probabilities for oil are very low for most prospects. It is largest for prospect D. The most common
evidence a priori is a reservoir in place, but problems with the HC generation or migration from the source rock. The producibility is a
potential problem just in 3 out of the 5 prospects (the others have probability 1).
It is important to remark here that this classification is not meant to replace extensive analysis of the real exploration outcome. The
purpose of our work is to allow an evaluation of the possible consequences before getting the real outcome. We believe that our network
and the proposed structure for the evidence can be an important and valuable instrument in this respect.
4. Sequential exploration and dynamic programming

Solving a sequential exploration problem means going through all possible sequential outcomes, and for each type evaluating all
possible decisions. When doing this, one must account for the likelihood of each scenario, the information provided by every scenario
SPE 159722
Difference Oil, Oil A

Difference Oil, Partial Oil A
RP
PRl2
rpA
PRl1
rpB
rpC
rpD
RP
1.0
rpE
PRl2
prA
prB
prD
prC
rpA
PRl1
rpB
rpD
rpC
rpE
prA
prB
prD
prC
prE
0.5
tgE
tsE
tgD
tgE
tsC
saC
0.0
tsD
tsB
tsC
saB
tsA
0.5
soA
tsA
0.5
soA
soB soC soD soE
smA smB smC smD smE
spB spC spD spE

SPl2
saA
trA
smA smB smC smD smE
SPl3
0.0
saB
B
A
soB soC soD soE
spA
SA
trB
tgA
trA
saC
trC
tsB
saA
tgB
trB
tgA
saD
D
trD
tgC
TS
trC
tgB
trE
tgD
SA
saE
tsE
saD
D
trD
tgC
TS
0.5
saE
E
trE
tsD
1.0
prE
spA
1.0
spB spC spD spE
SPl3
SPl1
SPl2
1.0
SPl1
Figure 4: Difference between conditional probabilities given evidence and prior probabilities. Evidence is observed in prospect A, and follows
the explanations in section 3.4 . Figure shows the effect of Evidence 1 (left) and 2 (right).
Difference Oil, Yes Res, No Sou A
Difference Oil, No Res A
RP
RP
PRl2
prA
prB
rpA
PRl1
prD
prC
rpB
rpC
rpD
rpE
1.0
PRl2
prA
prE
prB
rpA
PRl1
prD
prC
rpB
rpC
rpD
rpE
prE
0.5
0.5
tgE
tsE
tsE
tsC
SA
0.0
trC
tgA
saA
soA
soB soC soD soE
spB spC spD spE
SPl3
SPl2
SPl1
0.0
saB
saA
0.5
trA
soA
smA smB smC smD smE

spA
tsA
SA
trB
tgA
0.5
trA
saC
trC
tsB
A
tgB
saB
B
trB
saD
D
trD
tgC
TS
saE
trE
tsD
tsC
tgB
tsB
tsA
saC
tgD
saD
D
trD
tgC
TS
tgE
saE
trE
tgD
tsD
1.0
soB soC soD soE
smA smB smC smD smE
1.0
spA
spB spC spD spE
SPl3
SPl2
1.0
SPl1
Figure 5: Difference between conditional probabilities given evidence and prior probabilities. Evidence is observed in prospect A, and follows
the explanations in section 3.4 . Figure shows the effect of Evidence 3 (left) and 4 (right).
(in terms of probability updating), and the potential costs/revenues associated with different outcomes. The solution to this sequential
exploration problem is given by DP. Appropriate probability weighting enables the extraction of the optimal sequential exploration
strategy. The procedure is described in the next sections.
SPE 159722
Prospect
1
1
0
0
0
0
Evidence 1
Evidence 2
Evidence 3
Evidence 4
Evidence 5
Evidence 6
RP
1
1
0
1
1
1
SO
1
1
0
1
1
TR
1
1
0
1
PR
1
1
0
SA
1
0
-
Table 6: Classification of the possible observables (evidence). The marginal probabilities for the six observables sum to one for every prospect
(see Figure 6), therefore they define the whole sample space.
Marginal
0.8
0.6
Oil
0.4
Partial Oil
Dry, no Res
Dry, ok Res, no Sou
0.2
Dry, ok Res&Sou, no Trap

Dry, ok Res&Sou&Trap, no Prod
0
E
Prospect
Figure 6: Marginal probabilities of the six observables for the five prospects under consideration.
4.1 A motivating two-prospect example Consider again the 3-node BN example in Section 3.1 , where we can collect evidence at
node 2 or 3. We set a fixed cost of c = 100 of exploration drilling, and revenues r2 = 2200 c and r3 = 1100 c when the prospects have
HC. The question is now which prospect to explore first, if any. Marginally, the expected values are: r2 p(x2 = 1) cp(x2 = 0) = 120
and r3 p(x3 = 1) cp(x3 = 0) = 98 for the two prospects. They are both positive, so we should decide to drill a well. At first sight it
appears as if node 2 is more attractive (120 > 98). However, node 3 may provide more valuable information about the other prospect.
Assume that node 3 is explored first and is dry (x3 = 0). Then we must pay a cost c here, and next evaluate if the other prospect is
profitable, given the information at node 3. We continue drilling if the expected value is positive. The expected revenue from node 2 is
now r2 p(x2 = 1|x3 = 0) cp(x2 = 0|x3 = 0) = 73, and we decide to stop. We are better off by not drilling any further. On the other
hand, if node 3 is drilled and contains HC, we get conditional expectation r2 p(x2 = 1|x3 = 1) cp(x2 = 0|x3 = 1) = 1000. In this event,
we decide to drill prospect 2, because the continuation value is positive.
Similarly, when node 2 is explored first and is dry (x2 = 0), we get expected revenue from the other prospect: r3 p(x3 = 1|x2 =
0) cp(x3 = 0|x2 = 0) = 10. If node 2 contains HC, the expected revenue is r3 p(x3 = 1|x2 = 1) cp(x3 = 0|x2 = 1) = 890. In both
events the continuation value of node 3 is positive (10 or 890). Thus, no matter the outcome of node 2, we decide to continue to drilling
the other prospect.
Since information at node 3 influences the decision about node 2, the optimal drilling strategy is to start in node 3. Following the
argument above, the expected value of starting in node 3 becomes:
v3
=
+
p(x3 = 0){max(0, r2 p(x2 = 1|x3 = 0) cp(x2 = 0|x3 = 0)) c}
p(x3 = 1){max(0, r2 p(x2 = 1|x3 = 1) cp(x2 = 0|x3 = 1)) + r3 }
= 0.82{max(0, 73) 100} + 0.18{max(0, 1000) + 1000)} = 278
(2)
In comparison, the expected value is only 218 when starting in node 2.

In the case study situation, the network is much bigger than in this illustration. Moreover, the prospects will share costs related to
infrastructure and production. Nevertheless, the optimal strategy is derived in a similar manner as in this 3-node example, considering
10
SPE 159722
immediate profits and expected future profits from unexplored prospects, given the various kinds of evidence.
4.2 Costs and revenues for case study Potential oil and gas volumes have been suggested for the field under consideration. In
particular, P90, P50 and P10 values of the prospect volume distributions are shown in Table 7. As we have discussed earlier, we are
mostly interested in oil volumes here, but the geologists have provided us with estimates of gas volumes in place, and have proposed
plan for their development, therefore we will keep both oil and gas in the analysis.
Volume
Oil P90
Oil P50
Oil P10
Gas P90
Gas P50
Gas P10
A
4.2
5.4
7.5
7.6
10.2
12.9
B
2.7
3.8
4.8
3.0
4.1
5.1
Prospects
C
D
7.9
5.3
12.5 39.9
17.9 87.9
12.4 3.6
17.2 8.9
22.6 17.3
E
1.1
7.5
16
0.3
1.4
2.8
Table 7: Oil (million m3 ) and gas (billion m3 ) volumes for the five prospects under consideration.
Three main sources of costs have been identified:

DFC (Development fixed cost): Main infrastructure cost, common for the entire field, set to 2.5 billions USD.
CPP (Cost per Prospect): variable per prospect, proportional to the (expected) volume of HC in place. It is quantified in 1.5
billions USD for prospect D and respectively 0.55, 0.27, 1.02 and 0.27 billions USD for prospects A, B, C and E.
EFC (Exploration Fixed Cost): The average cost for an exploratory well ranges between 15 and 30 millions USD.
4.3 Utility function We implement a DP procedure to find the optimal sequential exploration strategy. The idea is to develop a decision
tree that can act as a road map for the exploration procedure, including all the possible outcomes that we can get. As a byproduct of this
procedure we can evaluate the value of the whole field, and and at every stage of information gathering. The computational ideas are
borrowed from Bickel and Smith (2006) and Martinelli et al. (2011a). An example is shown in Figure 7.
The idea of DP is to solve the decision tree by working backwards: first, we decide whether to drill the fifth prospect, conditional
on the first four observables. Then, for every scenario, we decide which prospect to drill if there are two prospects left, and so on, until
the initial empty set where no prospects have been explored.
C.1
A
D.1
D.2
evd 1
evd 1
evd 2
C.3
evd 2
C
evd 3
D.3
evd 3
evd 4
evd 4
evd 5
D
D
D.4
evd 6
evd 5
E
evd 6
C.2
D.5
C.4
C.5
Q
C.6
D.6
Figure 7: Decision tree for the basin. The first choice takes into consideration the five possible prospects plus the decision of quitting (Q). In
this example first prospect selected is D (S). The second choice considers the four remained prospects plus again the possibility of quitting
(Q).
SPE 159722
11
We define as the state of the system of evidences : is a vector of length 5, with 7 possible states per cell (-=not yet drilled,
plus the six evidence states shown in Figure 6), i.e. i = j means that we have observed outcome j in prospect i.
The DP we use is governed by the following expression:
(
"
(
)# )
v() = max
jNc
P(x j = k)
k=1
rkj CPPj + max
or
v() = max
jNc
sNn
P(xs = l|x j = k)(rsk CPPs + . . .), 0
)
h
i
k
k
P(x j = k|) r j CPPj + v( j ) , 0 ,
6
,0 ,
(3)
l=1
(4)
k=1
where rkj (revenues for prospect j in state k) is different from 0 just when k = 1 (full oil), and v(kj ) corresponds to the continuation value
from the state kj = {i i 6= j, i = k for i = j}. The notation j Nc means a prospect j selected among the possible prospects at the
current stage, while s Nn means a prospect s selected at the next generation, after j was selected at the current one. The parameter
is a discounting factor. When no prospects have been explored, i.e. in the outermost sum, we have to subtract also the DFC. This
intialization part is the largest cost. In this case we have define v0 (), that differs from v() just for the presence of the additional DFC
cost:
(
)
h
i
6
k
k
v0 (0 ) = max P(x j = k) r j DFC CPPj + v( j ) , 0 ,
(5)
jN5
k=1
where 0 = {, . . . , } is the state when no evidence has been observed at neither of the 5 prospects, and v(kj ) corresponds to the
continuation value from the state kj = {0 0 i 6= j, 0 k0 for i = j}
In the following part we will denote with v j () the expression within the first max sign. Equation (3) then becomes v() =
max j {v j (), 0}, and equation (5) becomes v(0 ) = max j {v j (0 ), 0}. Note that solving equation (4) requires first solving v(kj ), again
and again until the decision is restricted to a single node, following all the paths sketched in Figure 7. The original problem of identifying
where to drill first is now equivalent to the problem of solving v(0 ).
Thhe main problem with this optimal DP solution is the exponential growth of the number of scenarios that have to be considered.
In this particular case with 5 prospects, we can solve it without specific approximate strategies Martinelli et al. (2011a).
When we solve the DP, we compare the final result with the exploration cost EFCi in the best selected prospect. If the result is below
0, it means that the expected revenues that we can get from an exploration campaign starting with site i do not cover the exploration cost
in the same prospect, see for example Martinelli et al. (2011b) for a similar approach.
5. Results
Just a positive final value v0 (0 ) implies starting a drilling procedure. When this happens, it means that the expected value of the
first prospect and its continuation value exceeds the huge DFC. The first prospect selected is, depending on the discounting rate, the
biggest one in terms of intrinsic revenues ( = 0) or the a greater impact on the other prospects, in terms of the value of information
( = 1). For values of the parameter slightly smaller than 1, a balance between the two trends emerges, and it is possible to develop an
optimal decision strategy where the best site is selected on the basis of both trends. In the following studies we will use = 0.95 and an
average EFC of 20 million USD.
We select four cases with different revenue profiles and recovery factors. We will consider both cases with the possible contemporary
presence of oil and gas, and cases when just the oil component is considered. The decision maker is interested in testing the quality of
his decision even when the development of gas resources is not economically viable. We will not consider pure gas cases, because the
company has already excluded a gas-alone development of the field. For each case study, we run the sequential exploration procedure
described in Section 4 . Here are the results:
1. 1st case study:
P50 scenario, oil & gas
recovery factor 34% for oil and 65% for gas

Oil value: 75 USD/barrel
The strategy suggests as first choice a no drill, since the value v(0 ) is negative. Of course, there could be other reasons (new data,
risk seeking behavior, or other) that push the decision maker to explore, but such arguments would not be based on the statistical
model or the utility function with these costs and revenues. Among all the initial values, the highest occurs for prospect D. In
case we drill D anyway, the choice for the second well depends on the outcome of prospect D. This means that v(, , , k, )
is a function of evidence k. If k = 1 or k = 2 (commercial or geologically important discovery) we get the suggestion to drill E.
If k = 3 or k = 4 (reservoir not in place or source not in place), the DP suggests we terminate the exploration campaign. If k = 5
12
SPE 159722
(reservoir and source in place, but trap not in place), we get the indication of continuing the exploration campaign and drill E. The
case k = 6 can not happen within our framework, since the factor PR is assumed to be present with probability 1 in Prospect D (see
Figure 6). Note that, even though there is a strong negative advice initially, the secondary advice, given D, can be positive. This is
reasonable and it is explained by the form of the utility function that we have chosen: once we have drilled the first site, the biggest
cost (DFC) has already been paid. Therefore, if the evidence is positive, potential future discoveries may cover the CPPj s, and it
is optimal to continue the exploration campaign. If, on the other side, the outcome from the first well drilled is unsatisfactory, the
policy chosen (correctly) suggests to stop the campaign.
2. 2nd case study:
P50 scenario, oil & gas
recovery factor 50% for oil and 65% for gas

With an increased recovery factor we observe positive values for three prospects (D, E and A). The highest is reached in prospect
D. As second choice we get the advice to keep on exploring prospect E, no matter the outcome of the exploration in prospect D.
Here, the negative observation in D only has a limited impact, and now we have paid the DFC. It seems easier to boost a prospect
than to kill completely its chances. This is due both to the structure of the nodes (especially the counting nodes), and the quite low
a priori probability of success. A negative evidence in for instance prospect A lowers the probability of positive RP in prospect D
from 0.85 to 0.74. Therefore, in this extremely lucrative case, the future values are still high enough to suggest further exploration.
3. 3rd case study:
P50 scenario, just oil
recovery factor 34% for oil

In a pure oil scenario, even with a higher perspective price (100 USD per barrel), the DP suggests not to start an exploration
campaign. The initial v(0 ) is below 0. The highest initial value is again in prospect D. If we nevertheless decide to drill prospect
D, getting = {, , , k, }, the advice is to keep drilling just for evidence k = 1 (commercial oil) or k = 5 (reservoir and source
in place, not trap). Here, a non-commercial discovery is too weak for a future prosecution of the campaign. It is interesting that
the outcome k = 5 is better than the outcome k = 2. The lack in source abundance means a non commercial discovery, and this
has a strong impact on the likelihood of the other prospects. On the other hand, k = 5 means that just the trap has failed.
4. 4th case study:
P10 scenario, just oil
recovery factor 34% for oil

This is the best scenario among the ones taken into consideration so far. It is a just oil scenario with P10 volumes. In this situation,
four out the five initial values are positive, with the highest value reached in prospect D. The best choice for the second well is
to drill prospect E, no matter the outcome of the first prospect. Considerations are similar to those developed for the 2nd case
study. Again, the high profitability makes it convenient to keep drilling even with a very bad outcome in D, since single negative
evidence is not enough to kill the entire field.
As a partial temporary conclusion we state that the high DFC represents a threshold, that is difficult to overcome in most situations
even for the very profitable prospect D. It is true likewise that, once we have passed this threshold, the strategy suggests in many cases
to keep on drilling no matter the outcome. The higher start cost has already been paid, and the new costs are balanced by the possible
future revenues.
To further understand which mechanisms govern the sequential exploration strategy, we use a simulation study. We run 100 simulated scenarios, with volumes sampled from truncated Normal distribution interpolated between the P10, P50 and P90 values. The
histograms with Volumes related to Prospect D and E are shown in Figure 8 on the left.
For each of the samples we run a sequential exploration strategy. The samples marked with squares in Figure 8 (right), are representative of volumes where the advice is to start the exploration campaign. As we can see, there is a strong correlation with the volume
of the prospect D. As shown in Table 7, the oil volumes expected for this prospect are an order of magnitude higher than the volumes for
the other prospects, and this causes the strong correlation between prospect D dominant volume and the first outcome of the strategy.
SPE 159722
13
30
50
D
P10
P50
P90
E
P10
P50
P90
20
Empirical distribution
Samples
Positive 1st best choice
45
40
35
Oil Volume E (mmboe)
25
15
10
30
25
20
15
10
5
0
50
100
150
200
Oil Volume (mmboe)
250
300
50
100
150
200
Oil Volume D (mmboe)
250
300
Figure 8: Oil volumes, D in red and E in blue (left) and samples when first best choice is positive (right).
Next, the best choice for the second well depends from the outcome of prospect D. Even though the largest price has already been
paid, the exploration and possible exploitation of E depends on the expected volumes in E itself. In Figure 9 (left) we can see marked
with crosses the samples where a positive outcome for the first prospect (k = 1 in prospect D) implies a continuation of the drilling
campaign. If the outcome of the prospect D is k = 3 (no Reservoir), the number of samples advising a continuation decreases in a
sensible way (Figure 9, right).
50
50
Samples
Positive 2nd best choice given evidence 1 in D
45
40
35
35
40
Samples
Positive 2nd best choice given evidence 3 in D
45
30
25
20
15
30
25
20
15
10
10
5
0
0
0
50
100
150
200
250
300
50
100
150
200
250
300
Figure 9: Oil volumes D and E, in red positive first best choice and in cyan positive best choice for the second well, figuring out a positive
outcome (k = 1) for D (left) or a partially negative outcome (k = 3) for prospect D (right).
14
SPE 159722
6. Conclusion
The work proposes a framework for evaluating exploration risk when prospects are strongly correlated. We propose several correlation mechanisms based on BNs that are new compared with the ones currently used in commercial software. The main advantages of
such mechanisms are added flexibility and the ability to incorporate geological inputs when designing the correlation between two or
more prospects. For example, with the introduction of counting nodes or multi-level nodes, we introduce asymmetrical mechanisms that
can model more realistic geological situations.
We further couple this correlation model with an efficient and analytically consistent framework for deciding the best exploration
strategy. We use DP to construct the optimal exploration sequence.
The main results underline that the bigger prospect D stands out as the most profitable one. If the oil price and recovery factor are
sufficiently high, it is selected for the first exploration well. Our methods also analyze what happens next, after drilling prospect D. The
second and third best choices are non trivial, and depend strongly on the evidence collected by a possible exploration well in prospect
D. Sometimes the difference between a geological and a commercial discovery leads to different decisions at the second stage of the
exploration campaign. We conduct a sensitivity analysis to study the effect of volumes on the first and second best decisions. This is
very useful for what if? analysis, and helps decision makers in the risk analysis. In practice, one could run a much larger spectrum of
analysis within the same framework, testing for example different production schemes, or assess the impact of the probabilities in the
BN model.
We believe that this work adds useful elements for understanding risk evaluation methodologies. By applying the model and methods
on the real case study, we demonstrate its potential in a real-world setting. But, it is important to acknowledge that the real setting offers
data of many aspects, and perhaps one would allow re-tuning of the input variables, during the sequential decision making. Nonetheless,
the possibility of an a priori evaluation and quantification of every possible scenario makes the decisions much more informed.
References
Bickel, J. and Smith, J. (2006). Optimal Sequential Exploration: A Binary Learning Model. Decision Analysis 3, 1632.
Bickel, J., Smith, J. and Meyer, J. (2008). Modeling Dependence Among Geologic Risks in Sequential Exploration Decisions. SPE
Reservoir Evaluation & Engineering 11, 352361.
Cowell, R., Dawid, P., Lauritzen, S. and Spiegelhalter, D. (2007). Probabilistic Networks and Expert Systems. Springer series in
Information Science and Statistics.
Cunningham, P. and Begg, S. H. (2008). Using the value of information to determine optimal well order in a sequential drilling program.
AAPG Bullettin 92, 13931402.
Lauritzen, S. L. and Spiegelhalter, D. J. (1988). Local Computations with Probabilities on Graphical Structures and Their Application
to Expert Systems. Journal of the Royal Statistical Society, Series B 50, 157224.
Martinelli, G., Eidsvik, J. and Hauge, R. (2011a). Dynamic Decision Making for Graphical Models Applied to Oil Exploration. Mathematics Department, NTNU, Technical Report in Statistics 12.
Martinelli, G., Eidsvik, J., Hauge, R. and Drange-Forland, M. (2011b). Bayesian Networks for Prospect Analysis in the North Sea.
AAPG Bulletin 95, 14231442.
Murphy, K. (2001). The Bayes Net Toolbox for Matlab. Computing Science and Statistics 33.
Rose, P. (2001). Risk analysis and management of petroleum exploration ventures. AAPG Methods in Exploration Series 12.
Smith, J. . and Thompson, R. (2008). Managing a Portfolio of Real Options: Sequential Exploration of Dependent Prospects. The
Energy Journal, International Association for Energy Economics 29, 4362.
Suslick, S. and Schiozern, D. (2004). Risk analysis applied to petroleum exploration and production: an overview. Journal of Petroleum
Science and Engineering 11, 19.
VanWees, J., Mijnlieff, H., Lutgert, J., Breunese, J., Bos, C., Rosenkranz, P. and Neele, F. (2008). A Bayesian belief network approach
for assessing the impact of exploration prospect interdependency: An application to predict gas discoveries in the Netherlands. AAPG
Bulletin 92, 13151336.
Paper IV
Building Bayesian networks from basin modeling scenarios for
improved geological decision making
G. Martinelli, J. Eidsvik, R. Sinding-Larsen, S. Rekstad and T. Mukerji
Submitted for publication, 2012.
Building Bayesian networks from basin

modeling scenarios for improved geological
decision making
Gabriele Martinelli, Jo Eidsvik
Dept. of Mathematical Sciences, Alfred Getz vei 1,
Richard Sinding-Larsen, Sara Rekstad

Dept. of Geology and Mineral Resources Engineering, Sem Slands veg 1,
Tapan Mukerji
Department of Energy Resources Engineering, School of Earth Sciences,
Stanford University, USA
Abstract
Basin and Petroleum Systems Modeling is important for understanding the geological mechanisms that characterize reservoir units. Bayesian
Networks are useful for decision making in geological prospect analysis and
exploration. In this paper we propose a framework for merging these two
methodologies: by doing so, we take into account in a more consistent and
explicit way the dependencies between the geological elements. The probabilistic description of the Bayesian Network is trained by using multiple
scenarios of Basin and Petroleum Systems Modeling. A range of different
input parameters are used for total organic content, heat flow, porosity, and
faulting, to span a full categorical design for the Basin and Petroleum Systems Modeling scenarios. Given the consistent Bayesian Network for trap,
reservoir and source attributes, we demonstrate important decision making
applications such as evidence propagation and the value of information.
Keywords: Bayesian Networks, Scenario Evaluation, Basin Modeling,
Uncertainty Quantification, Petroleum Exploration
Corresponding author
Email address: gabriele.martinelli@math.ntnu.no (Gabriele Martinelli )
Preprint submitted to Petroleum Geoscience
1. Introduction
The correct integration of geological and geophysical information within
a decisional framework for the purpose of oil and gas exploration is a challenge that will increase in importance with increasing cost and exploration
difficulties of new targets. Currently it is common practice among scientists
to quantify information about risk through detailed exploration analysis,
and then forward these results to management. From the geologists side
the evaluation include for example basin modeling, seal capacity and sedimentology analysis. From the geophysicists side we can consider electromagnetic, magnetic and gravimetric data, and of course, 2D and 3D seismic
surveys. In the transition towards the decision makers the information is
processed and quantified through expert opinions and commercial software
R for risk assessment, multiple-scenario evaluation and es(such as GeoX )
timation of the amount and value of hydrocarbons (HC) resources under
study. In this work we propose a supplement to the existing framework by
integrating directly basin modeling scenarios and decision strategies.
We can identify the problem by analyzing how currently we move from
the Earth model to the decision space: the geological and geophysical knowhow is first translated into basin and petroleum system modeling (BPSM).
Outputs from multiple runs of basin modeling under different geologic scenarios are then used to establish a Bayesian network (BN) that models play
element dependencies. The BN is used to test decisions. In this work we
R
have used a common commercial software for BPSM, namely PetroMod .
Petromod is based on a finite-element simulator (Hantschel and Kauerauf,
2009) that numerically solves the coupled system of equations for sediment
compaction, heat flow, petroleum generation and migration, accounting for
both chemical and physical processes.
In this framework a sensitivity analysis is then carried out, and a database
with multiple runs (corresponding to different geologic scenarios) is built.
The database is the starting point for the value assessment part that provides the basis for efficient decisions.
The idea of modeling play element prospect dependencies with BN was
proposed in VanWees et al. (2008) and Martinelli et al. (2011). Martinelli
et al. (2011) constructed a BN model for assessing the likelihood of source
presence in a part of the North Sea. The network describes the prior distribution of the source system in terms of kitchen, prospects and segments. We
will use the word segment for identifying a volume possibly filled with HC
resulting from a source-reservoir-trap system, while we will use the word
prospect for describing a collection of segments that share some common
features.
When the BN is established, one can use standard techniques to propagate the evidence at certain nodes to all other nodes. This allows us to
study the value of information (VOI) at one or more segments (Bhattachar2
jya et al., 2010). Similar ideas were developed in VanWees et al. (2008).
One of the main critical points of Martinelli et al. (2011) was the substantial belief in expert opinion when designing the BN. In the present paper we
propose an alternative idea for building the BN, integrating expert opinions
with quantitative geological data. The main idea is to train the probabilistic structure of the BN from the multiple basin modeling outputs. This is
done by statistical parameter estimation, together with discretization and
clustering guided by geological intuition. This BN model couples the geological processes and their responses with risk assessment. Assigning expected
revenues to segments, the production strategy and other required economic
variables can now easily be communicated. The BN model provides explicit
probability statements, at single-segments and for prospects.
Using statistical design of experiment (DOE) with oil and gas forecasting problems is not new: Damsleth et al. (1992) and Dejean and Blanc
(1999) propose a DOE based approach for reservoir modeling simulations;
Corre et al. (2000) extends DOE and MonteCarlo methods in order to study
uncertainties in geophysics, geology and reservoir engineering. Dependency
among wildcat wells has been discuss in Kaufman and Lee (1992), where
a binary logit model for the number of successes is proposed. citekaufman
mention, though, that the forecasting capacity of the model was poor in
absence of a correct geological model of the basin.
The paper is organized as follows: In Section 2 we introduce BPSM
and the synthetic case study; Section 3 discusses the DOE simulation setup
with interpretations. In Section 4 we show the procedure for developing the
BN model. Finally, in Section 5, we apply the model for decision making;
Section 6 is the conclusions.
2. A Case study for basin and petroleum systems modeling
2.1. Basin and Petroleum Systems Modeling
BPSM is a useful component in exploration risk assessment and is applicable with increasing reliability during all stages of exploration, from frontier
basins with no well control to more mature areas. The idea is to simulate the
geological and chemical reactions that have occurred in the basin through geological time, in order to identify the critical aspects of the HC generation,
migration and accumulation. Important geological risk factors in oil and
gas exploration are the trapping (consisting of trap geometry, reservoir and
seal), the oil and gas charge (reservoir and source factors), and the timing
relationship between the charge and the formation of potential traps. These
risk factors apply equally to basin, play and prospect scale assessments.
BPSM software combine seismic, well, geological and petrophysical inforR
mation to model the evolution of a sedimentary basin. As output Petromod
will predict if, and how, a reservoir has been charged with HC, including
the source and timing of HC generation, migration routes and amount of

HC both at subsurface or at surface conditions.
In this paper we will use the 3D version of the software, that allows for
full visualization of the migration paths that lead to the accumulation of
HC in the basin.
2.2. The Bezurk case study
We have decided to use as training model a synthetic basin developed
in the Petroleum Geology class at NTNU, Trondheim, Norway (Tviberg,
2011). The controlled basin environment is called Bezurk Basin (Figure 1),
and it includes three potential kinds of prospects, namely anticlinal type,
fault type and a shoestring type. The latter is located within impermeable
shale and consequently the chances for HC to migrate into this reservoir are
low. The Bezurk basin mimics the behavior of a possible real basin with a
main anticlinal trap on the NE sector of the basin, and a series of faults in
the NS direction. A major uplift followed by a strong erosion has occurred
in the western part of the basin, and this activity has caused the major
faulting shown by Faults 1 and 2.
Mmd
Mlf
Ou
Fault 1
Eek
Fault 2
Figure 1: Bezurk basin; we see the 100 km2 area and the different thicknesses of the layers;
in the west part of the basin we identify the two faults that characterize the system.
The history of the basin has been characterized by the deposition of

organic-rich shale and good-porosity sandstone layers. In particular, we
recognize two main possible HC producing layers, the deepest being the
coal bed layer denominated Eek, and the shallowest being a shale rich in
organic content denoted Mlf. Environmental and lithological descriptions
are provided in the supplementary material. Other assumptions are that
4
the Bezurk basin is an onshore basin, with sediment surface at zero meters
above the sea level, and that the basin is located somewhere in the Middle
East. The depositional history started 55 Ma ago and has continued until
today, with a number of erosional episodes. Figure 2 shows two cross-sections
of the basin. Marked in black (Eek) and pink (Mlf) are the two main source
rocks, and in yellow (Mmd) and red (Ou) the two main reservoirs. The third
reservoir layer, a shoestring reservoir, is visible in the second cross-section.
This is between the two Mua seal layers just in the synclinal part of the
basin. The main anticlinal reservoirs are clearly visible in the first cross
section, in the eastern part of the basin.
We have identified 2 main plays, corresponding to the two main potential
reservoir rocks:
The reservoir of the Mmd play in the Bezurk Basin is made up of
sandstone, deposited in a regressive shallow marine environment during the time interval 20Ma to 15Ma. The sandstone reservoir has
porosity ranging from 12% to 30%, which is considered to be a good
porosity. The reservoir covers the whole area on the east side of the
faults, and has a thickness ranging from about 300-900m.
The reservoir of Ou play is deposited from 34Ma to 23Ma in a transgressive shallow marine environment with the overlying Mlf shale acting as a seal. The underlying Eek-coal is deposited on a coastal plain
in the same transgressive system as the reservoir and it is expected
to generate HC due to its depth of burial and the corresponding Heat
Flow. Potential traps are the western faults and the northeastern anticlinal, which are similar to the traps of the younger Mmd-play. The
porosity of the Ou reservoir ranges from 7% to about 20%, which overall is lower than the porosity in the Mmd reservoir. Both reservoirs
have the same kind of sandstone, but due to compaction the lower
reservoir (Ou) has a smaller porosity than the upper reservoir (Mmd).
Generated HC are expected to migrate to the overlaying reservoir. The
critical factor is the geological timing, both for the Ou-play and for the
Mmd-play. In both plays the seal is deposited on top of the reservoir rock.
The sealing efficiency may be inadequate to keep the HC inside the trap in
scenarios with early generation and migration. This can cause large amounts
of HC to be lost.
The basin is exposed to normal faulting at a young age (11 Ma). Two
faults are observed in the profiles (western part). The faults are considered
to be closed faults. HC accumulated in these traps constitute the faultprospect. The critical factors of the prospects are the uplift and erosion
related to the faulting.
Basic Building Process
3.3 Building the Model in PetroMod

The maps are imported into PetroMod as additional maps, in the top to bottom order (Qal to
Basic Building Process
Basement) and the layers are created based on the imported maps.
3.3 Building the Model in PetroMod

The maps are imported into PetroMod as additional maps, in the top to bottom order (Qal to
Basement) and the layers are created based on the imported maps.
Figure 5. Cross section GPY 80
The gap in the fault zone are explained in 3.3.7 Faults.
The gap in the fault zone are explained in 3.3.7 Faults.
Figure 2: Cross sections. In the first one we can easily recognize the four way anticlinal
trap located in the eastern part of the basin; in the second we can identify the Mlq
Figure 6. Cross
section pinches
GPY 38
shoestring reservoir
that
through the Mua seal layer.
12
2.3. Expected Results

HC generation: Both source12rocks are buried deep enough to generate HC. Eek is deposited in a coastal plain environment in the time
interval 34.8 Ma to 34 Ma, and is today located at a depth of about
3000m to 5000m. The lithology of the deepest source rock (Eek) is
coal, which mainly generates gas, but can be also oil prone. The
source rock which today is at the depth of 2000m to about 4500m is
the Mlf black shale. Mlf is deposited in a deep marine environment in
the time interval of 20.60Ma to 20.00Ma, and is expected to generate
both oil and gas. The generated HC are expected to migrate into the
overlying Anticlinal-prospect and the Fault-prospect.
Anticlinal prospect: The Anticline prospect is expected to contain
HC in both the Mmd reservoir and the Ou reservoir. The trap mechanism is as the name implies an anticline. It has a four-way closure and
no large risks are related to trap mechanism. The sealing rocks for
both reservoirs are shale, which over time are expected to obtain adequate sealing capacity and thus prevent the HC from leaking during
the rest of the migration process. The lower accumulation is expected
to contain more gas than the upper accumulation, due to Eek source
rock being more gas prone.
Fault prospect: The Fault prospect contains some more uncertainties regarding HC preservation. The trap mechanism is a normal fault,
which has remained closed from 11Ma to today. However, the effect
of the uplift and the subsequent erosion in the western part of the
basin needs to be modeled: will the timing of the fault and its sealing
capacity be adequate to hold accumulations in place throughout the
basin development? Other crucial questions that need to be evaluated
relate to the change in the geometry of the basin with time and how it
affects the flow paths, and the size of the drainage area of the anticline
that influences the accumulation of the Fault prospect.
The Shoestring prospect: No HC are expected to migrate into the
Mlq reservoir that constitutes the Shoestring prospect. The sand is
deposited as a single river channel in a fluvial system and is isolated
inside the continental Mua impermeable shale. However, if the Mua
shale had been modeled to contain permeable pathways, or the presence of fracture zones, then HC could potentially have migrated into
the reservoir.
2.4. The master model
We have designed a master model by establishing a plausible petroleum
system scenario and a series of boundary conditions. In particular we have
chosen a constant heat flow (HF) of 60 mW per m2 , that corresponds
to a moderately active basin (Allen and Allen, 2005). We have estimated
the paleo water depth (PWD) according to the depositional environment
through time (see Table 1). Finally, since Bezurk is conceived as an onshore
basin there is no water present and the sediment-water-interface temperature
(SWIT) is in reality the sediment-air-interface temperature.
Layer
Qal
Plj
Muh
Mua
Mlq
Mmd
Mlf
Ou
Eek
Depositional Environment
Continental deposits
Oxic environment
Shallow marine
River channel
Continental
Shallow marine
Deep marine
Shallow marine
Coastal plain
PWD
0m
50 m
50 m
5m
5m
100 m
500 m
100 m
0m
Table 1: Paleo water depth (PWD) input values for each of the layers in the model.
An illustrative run (Figure 3) shows that the sole prospect that today is
filled with HC are the two trap segments of the anticlinal formations on the
eastern part of the basin. We see traces of HC against the wall of the closed
faults, but no significant accumulation. Figure 4 shows paths and drainage
areas, illustrating how the migration at the present time converges on the
anticlines, while a minor part of HC migrates westwards toward Fault 2.
Figure 3: Illustrative run: we see the oil (green) and gas (red) accumulations in the
anticlinal segments, with traces of HC in the fault segment.
The HC that migrate into the fault prospect are mainly lost during the
time step of 1.77Ma - 1.55Ma (Figure 5), which is the critical time when
the Muh seal is eroded. This particular uplift creates erosion, and losses
can consequently be explained by the change in the geometry of the basin.
The reservoir layer creates a small anticlinal trap structure against the fault
where the HC accumulate. After the uplift the trap structure flattens out
and the HC migrate out of the trap.
3. Basin modeling scenarios
During the analysis of the basin we have been able to identify four critical elements that constitute possible sources of uncertainty in our model.
In real life all basin parameters are more or less unsure. To accomodate this
uncertainty in our synthetic basin several scenarios for Total Organic Carbon (TOC) content, HF and porosity levels are considered. The TOC is a
measure of the concentration of organic material in source rocks (Allen and
Allen, 2005). We have also noticed that there is a zone in the western part
of the basin characterized by a prominent faulting activity; for this reason
8
Simulation of the Bezurk Basin

Accumulations and flow path
Simulation of the Bezurk Basin

None or very few hydrocarbons are trapped in the fault-prospect. The hydrocarbons that migrate into
the prospect are mainly lost during the time step of 1.77Ma 1.55Ma (Figure 36), which is when the
Muh seal erodes (critical moment). This particular erosion creates uplift; as such the losses can be
explained by the change in the geometry of the basin. Figure 37a illustrates a close-up view of the
accumulation at 1.77Ma. The Mmd reservoir is shown as a transparent layer and the oil (green) and
gas (red)
are also displayed.
The reservoir
layera) creates
a smallb)anticlinal
trap structure against the
Figure
35. Hydrocarbon
accumulation
and flow path
Mmd reservoir
Ou reservoir
Figure
4: HC
and flow paths
for uplift
Mmdthe
play
and flattens
Ou playout
(right).
fault where
theaccumulations
hydrocarbons accumulate.
After the
trap(left)
structure
(FigureThe
37b)
drainage area of the anticlinal traps is much larger than the drainage area for the fault
and
the
hydrocarbons
migrate
out
of
the
trap.
The results from the simulation show two accumulations in the Anticline prospect (Figure 35). The
traps.
information in Table 13 is extracted from PetroMod and shows that The Anticlinal prospect constitutes
almost 100% of the total resources in the basin. As expected the lower accumulation contains more
gas.
Oil (1e6 STB)
866.5
Assoc. Gas (1e9 scf)

236.49
Non. Assoc. Gas (1e9 scf)

0.12
Condensate (1e6 STB)

0.01
Total Mmd reservoir

Anticlinal bot. segment
(Ou), acc. Nr.38
866.5
149.97
236.49
151.26
0.16
92.03
0.01
2.18
Total Ou reservoir
149.98
151.44
95.64
2.19
Anticlinal top segment

(Mmd), acc. Nr. 25
Table
in The
Bezurk Basin
Figure13.
36.Accumulations
HC acc., a) 1.77Ma
b) 1.55Ma
Figure 5: Accumulation in the fault segments; screenshot of the process at 1.77 Ma and
1.55 Ma. Most of the HC leak out during and after the uplift of the basin.
we can hypothetize a possible structural uncertainty, by adding or removing

one of these fault elements from our model.
We next run multiple-scenarios of BPSM changing the key factors in a
controlled design of experiment (DOE).
3.1. A full factorial design
In order to study the interactions among the different factors, we have
designed a full factorial study (Fisher, 1971), where each factor is represented
by
two to three levels. We have chosen three levels for the HF (HF): cool
Figure 37. HC acc. Close up, a) 1.77Ma, b)1.55Ma
(50 mW/m2 ), normal (60 mW/m2 ) or hot (70 mW/m2 ); it is expected that
a cool basin mainly will stay in the oil window, consequently generating
mostly oil, while a warm basin will reach the gas window at an earlier stage,
and therefore generate more gas. We have further chosen two levels for the
porosity of the reservoir rock, good or bad (see profiles in Figure 6). We
use two levels for the TOC content of both source rocks, with TOC ranging
from 8% to 4% for the Mlf black shale and from 20% to 10% for the Eek
coal. Finally, we select two levels, open or close, for the presence of a new
fault (Fault 3) located east of Fault 2. Table
2 summarizes the results from
36
37
the scenarios. From the master model (see Section 2.4), we observe that the
HC which accumulated in The Fault Trap were lost during the time period
of 1.77Ma to 1.55Ma. The reason for adding the Fault 3 is to see if this
could trap HC and potentially create a prospect.
Figure 6: Porosity profiles; on the left the high case, with initial porosity around 40 %;
on the right the low case, with initial porosity around 30 % and a rapid decrease.
3.2. Simulation outcomes

In each of the 24 different BPSM runs, we measure the size and type
of HC accumulations. We further measure which source rock has generated
them and we observe the migration path. We gain insight into the HC
production, the expulsion from the source rock and the accumulation in the
reservoirs. As a result, the amounts of HC that have leaked is available, and
we can try to explain this leakage phenomenon through the observation of
the complete evolution of the basin.
In this section we discuss the main effects of the different scenarios. A
more complete analysis is provided by analysis-of-variance printouts and
diagrams in the supplementary material.
In the supplementary material is reported a Table that contains the main
data, concerning the generation, the expulsion, the accumulation and the
leakage for each of the 24 scenarios. Data are in MMBOE (Million barrels
oil equivalent) throughout the whole analysis.
The generation phase is divided in oil and gas generation and further
subdivided in the two source rocks that are responsible for the HC generation, respectively the Eek source rock and the Mlf source rock. The main
factor driving the HC generation is the level of maturation of the source rock
itself, that ultimately depends on the burial depth, the HF and the TOC.
The analysis shows that higher HF allows a earlier and faster maturation
and therefore a more abundant generation of gas in both the source rocks.
For the oil generation there are no significant differences in the impact of
HF for the Eek source rock. This means that the oil generation has reached
the maximum potential already when HF is on the medium level, and this is
10
Model
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
Porosity
high
low
high
low
high
low
high
low
high
low
high
low
high
low
high
low
high
low
high
low
high
low
high
low
HF
cool
cool
normal
normal
hot
hot
cool
cool
normal
normal
hot
hot
cool
cool
normal
normal
hot
hot
cool
cool
normal
normal
hot
hot
Fault 3
closed
closed
closed
closed
closed
closed
open
open
open
open
open
open
closed
closed
closed
closed
closed
closed
open
open
open
open
open
open
TOC
high
high
high
high
high
high
high
high
high
high
high
high
low
low
low
low
low
low
low
low
low
low
low
low
Table 2: Experimental table, full factorial design with 4 factors (Porosity, Heat Flow,
Fault 3 and TOC) and 2 2 2 3 = 24 total levels.
consistent with our hypothesis. It turns out that when HF is high, most of
the oil generated by the deeper source rock leaks out before being trapped.
Therefore, the overall effect is a smaller oil accumulation in the Ou reservoir
when HF is high. HF and TOC are also the main parameters responsible for
controlling the quantity of expelled HC, i.e. the amount of HC that leave
the source rock after the generation.
Regarding the size of accumulations, we can see that the main factor
is the porosity, followed by the HF again, especially for what concerns the
Ou accumulations. It is quite natural that the porosity is relevant, since a
sandstone with good porosity can trap much more HC than a bad reservoir.
It is interesting to notice how the effects of HF and TOC tend to disappear,
showing that the surplus of HC generated has almost totally been lost before
the seal rock had reached its sealing capacity. Actually, we notice that the
oil accumulations in the Ou reservoir decrease sensibly with increased levels
11
of HF, and are quite stable with respect to TOC.

Finally, Fault 3 in the western part has a strong effect only when it is
leaking. When the fault is not present, there is no leaking through the faults
wall. On the contrary, when the fault is present, there is some leaking along
the fault, especially when there is an early maturation (high HF). The fault
has clearest effect when measuring the outflow from the side. In contrast,
the outflow from the top and the total outflow is governed by the HF and
TOC, since the scenarios with early maturation leak most of the HC before
the seal is adequately sealing.
We have run a similar analysis on the second major kinds of data that
we get from a BPSM analysis, i.e. the oil and gas accumulations. In the
supplementary materials we attach a Table with the expected accumulation
values at surface conditions for each of the 24 scenarios.
We have distinguished four main accumulations, two in the eastern part
of the basin, under anticlinal traps, and two in the western part of the basin,
against the Fault 3. We name the first two accumulations as TE (Top East)
and BE (Bottom East), and the latter as TW (Top West) and BW (Bottom
West). TE and TW refer to the Mmd play, while BE and BW refer to the
Ou play. In the following sections these 4 accumulations will represent our
4 segments; TE and BE belong to the anticlinal prospect, while TW and
BW belong to the fault prospect. The data confirm what is already observed
through the previous analysis: oil accumulations in the Ou reservoir decrease
with increasing HF, while gas accumulations increase from cool to normal
HF, and then remain almost stationary. The main effect is the porosity of
the reservoir rock, that accounts for the largest part of the total variability.
4. Building the Bayesian network
The experimental design setup gives better insight of the key factors
responsible for the main geological processes in the basin. We will next
use the multiple-scenario information to build a dependency structure for
the segments. This takes the form of a BN that will be useful for decision
making.
A BN is characterized by a set of nodes and edges. The nodes are random variables, that may be discrete or continuous. As an example, we will
define nodes for trap presence (on/off), which is a binary random variable.
Edges define the conditional probability structure of the variables, connecting parents to children. For instance, we will define a parent node for Trap
Anticline that can be on/off. This node has two children: TrapTopEast
and TrapBottomEast, which are also on/off, and they have conditional
probability distributions depending on the outcome of the parent node. Let
V be the set of all nodes, xv the variable at node v V , and x the vector
12
of all node variables, the joint probability model can then be defined by
Y
p(x|) =
p(xv |xpa(v) , v ).
vV
Here, pa(v) denotes the parents of node v. Further, denotes the set of
model parameters required for the conditional probabilities tables (CPT),
where v is the local parameter for node xv . We show below how we can train
or learn these parameter values from the multiple-scenario BPSM outputs.
We have chosen to use a BN structure similar to that of Martinelli et al.
(2011). The CPT are then parameterized by incorporating basic geological
mechanisms and allowing for local failure in the propagation of HC elements,
and we train parameters within this context. The formulation in Martinelli
et al. (2011) appears to be a flexible way of modeling dependencies steming
from different geological elements (trap, source, reservoir). The separate assignment of these elements gives a generic model specification that is easy to
interpret and communicate. Finally, a BN formulation allows explicit evaluation of the changes in the probabilities when single elements are observed,
which leads to what-if studies or VOI calculations.
By using trap, source and reservoir elements in the BN, we thus avoid
direct use of the factors involved in the DOE. We have seen that the HF
for example interact both at source and at trap level, and that the porosity
affects both the accumulation and the leaking phase.
4.1. Learning the network
The complete set of 24 scenarios, and associated observations, are shown
in Figure 7 (generation) and in Figure 8 (accumulation). We have used a
standard k-means algorithm with k = 2 (accumulation) or k = 3 (generation) for assessing the threshold for categorizing the data. Note that the
data in this way become proxy for the knowledge of geological elements,
that could potentially be observed at segment level.
We next consider the three main geological elements (trap, reservoir and
source) separately. The BN model we have established is shown in Figure
9.
Trap: We have developed a network with 6 nodes: two parents,
TrapAnticlinal and TrapFault, and four children, TrapTE, TrapBE,
TrapTW and TrapBW. The marginals probabilities for the top nodes
are {0, 1} for the anticlinal trap and {0.5, 0.5} for the fault trap. This
is set by direct learning from the DOE output. The local CPTs for the
children nodes include the possibility of a local failure, quantified in
the success probability T (0.9). This allows a strong and effective
learning when the fault trap presence is confirmed or ignored.
13
Generation Total
15000
10000
5000
0
0
10
15
20
25
Generation Mlf
Generation Eek
10000
3000
8000
2500
6000
2000
4000
1500
2000
1000
500
0
10
15
20
Generation Mlf Gas
25
Generation Mlf Oil
1000
10000
800
8000
600
6000
400
4000
200
2000
10
15
20
Generation Eek Gas
25
Generation Eek Oil
800
2000
600
1500
400
1000
10
15
20
25
200
0
0
10
15
20
25
500
0
10
15
20
25
10
15
20
25
Figure 7: Data for learning the source network. Top: values for the HC generation.
Middle: values for Eek and Mlf generation. Bottom: values for oil and gas generation in
each of the Eek and Mlf source rock. Values in MMBOE.
Accumulation Total
1500
1000
500
0
0
10
15
20
Accumulation Mmd
25
Accumulation Ou
1000
600
800
400
600
200
400
200
0
0
10
15
Accumulation Mmd Gas

60
20
25
Accumulation Mmd Oil
10
15
Accumulation Ou Gas
1000
60
20
25
Accumulation Ou Oil
400
800
300
40
40
600
200
20
20
400
100
200
0
10
20
0
0
10
20
0
0
10
20
10
20
Figure 8: Data for learning reservoir network. Top: values for the HC accumulation.
Middle: values for Mmd and Ou generation. Bottom: values for oil and gas generation in
each of the Mmd and Ou source rock. Values in MMBOE.
14
Reservoir: The reservoir network is another small network, with

7 nodes: one common parent, that represents the total accumulation, two mid-level parents that represent respectively Mmd and Ou
reservoirs and four children representing oil and gas accumulations in
reservoirs (ResMmdGas, ResMmdOil, ResOuGas and ResOuOil). The
nodes can be efficiently learned through a simple BN algorithm, given
the data provided by the BPSM simulations.
The learning process follows the classical maximum likelihood procedure with complete data using the joint model (Cowell et al., 2007).
Because of conditional independence, and the database output from
the DOE, we can maximize each term separately. This means that we
can restrict our attention to the term we are interested in, locally, and
then we can find the maximum likelihood estimate v locally, based
on our database:
#(xv xpa(v) )
xv v =
,
#(xpa(v) )
i.e. the CPTs are estimated by the ratio of the corresponding counts
in the database. For the top nodes, we just count the fractions directly. In our case we update on the basis of a limited number of
experiments. Different and possibly more complex prior distributions
can be assigned to the top nodes, such as Dirichlet priors. When there
are missing or incomplete data more refined techniques are suggested,
such as Expectation-Maximization (EM) or penalized EM algorithms,
see e.g. Jordan (1998) and Cowell et al. (2007).
In our example all reservoir nodes are binary (two states, high and
low ), and we impose a threshold for the accumulations being larger
than a certain value. These values are the separating planes indicated
by the k-mean algorithm. (Figure 8).
With this procedure we can derive explicitly the correlation between
the different nodes that we have created. Note that we have not imposed these correlation, but derived them from the data. They are
nonetheless possible to tune, if the values are in contradiction with
expert belief or other sources of data. Just to give an example, for
the reservoir subnetwork just described, we can check the correlation
between some of the bottom nodes: the correlation between the node
ResMmdGas and ResMmdOil is 0.816, while the correlation between
ResOuOil and ResMmdOil is 0.326. The first result is natural since
a good porosity of the reservoir rock increases its ability to hold both
oil and gas. The second result tells us that the porosity in the Ou
reservoir and in the Mmd reservoir are weakly but positively correlated, and this comes from the fact that we have imposed the same
sandstone to both reservoir rocks. We report in Table 3 the condi-
15
tional probabilities of the given variables (note that they are not in
a parent-child relationship!), derived from our BN; the marginals for
the state high for the variables ResMmdGas ResOuOil are respectively
0.427 and 0.289.
ResMmdGas / ResMmdOil low high
low
1
0
high
0.2 0.8
ResOuOil / ResMmdOil low
high
low
0.867 0.133
high
0.575 0.425
Table 3: Conditional Probability Tables for the variables ResMmdGas vs ResMmdOil
(left) and ResOuOil vs ResMmdOil (right), within the Reservoir subnetwork.
Source: The source network is more complicated, since we have to

take into account two phenomena that interact with a difficult correlation structure, namely one for the gas generation and one for the
oil generation. As we have previously discussed, for the shallower top
rock an increase in HF has the duplex effect of a higher oil and gas
generation. On the other hand, for the deeper source rock, it affects
just the gas generation. TOC affects both generation in similar ways.
We learn the statistical effect of this behavior from the DOE outputs
concerning the generation phase. We include a correlation structure
with 3 levels: a top node for the total generation, intermediate nodes
for the Mmd and Eek generation and bottom nodes for the gas and
oil generation in each of the source rocks.
In this case we have assigned three levels to all the nodes, respectively
high generation, medium and low. Most of the CPTs are learned
directly from the data, using GenTot, GenMlf, GenEek, GenMlfOil,
GenMlfGas, GenEekGas and GenEekOil along with the thresholds
discretizing the data. All the nodes are discrete, with three possible
stases, i.e. k = 3 in Figure 7.
Finally, we gather the information that we get from source, reservoir
and trap in a single node, using our geological understanding of the process.
We know that the source rock is essential for the presence of HC in the
prospect, while a poor reservoir quality or a poor trap makes it less likely
to have a commercial discovery in the prospect. We will next discuss other
considerations for joining the last part of the network.
4.2. Gaussian nodes
So far in the BN building, we have not used the accumulation volumes
extracted from the multiple-scenario BPSM (for learning the reservoir net16
work we have used joint layer accumulation values and not prospect/segment
values). We will now incorporate this information in the bottom nodes of
the network. It seems reasonable to have discrete nodes in the top parts of
the network, since attributes such as source, reservoir and trap are on/off
or multi-level features. In the bottom part of the network it may be more
realistic to have continuous nodes that mimic the actual behavior of the
simulated scenarios. We therefore split each of the bottom nodes TE, BE,
TW and BW in two nodes, one for gas volume and the other for oil volume,
and state that they represent accumulation distributions whose mean and
(possibly) variance depend on the states of their parents. The simultaneous
use of discrete and continuous variables in BN has been explored in Chang
and Fung (1995) and Friedman and Goldszmidt (1996). A good inference
algorithm is presented in Murphy (1999). The related CPTs have to be
assessed, for example the conditional probability density of BEg (BE gas)
is:
2
pBEg (x|T raBE , ResOuGas , SouEekGas ) N (BEg , BEg
),
where BEg is the conditional mean value and BEg is the conditional standard deviation of this Gaussian distribution. This means assessing 12 mean
and variance parameters (2 states for Trap and Reservoir and 3 for Source)
for each of the 8 nodes. We use the simulated accumulation values as references for the mean values of our Gaussian distributions.
TraAnti
TraTE
TraFault
TraBE
TraTW
TraBW
BWo
BEo
BWg
ResOuOil
BEg
ResOu
TWo
ResOuGas
SouEekOil
ResTop
TEo
TWg
ResMmdOil
SouEekGas
ResMmd
TEg
ResMmdGas
SouEek
SouMlfOil
SouMlfGas
SouMlf
SouTop
Figure 9: BN with trap(top), reservoir(left) and source(lower right) branches. Top nodes
are all discrete, while bottom nodes are Gaussian, as explained in Section 4.2.
Further, we include the possibility of local failure of one element, with a

17
reduced volume, according to Table 4. Let the parameters R and S be the

local importance factors for the elements source and trap. The effect is that
the failure of a single element can still produce minor accumulations. This
occurs for instance if we believe that low/high states for factors like porosity
do not totally preclude the formation of HC, but simply produces a sensible
reduction in the quantity (as seen in the simulations), due to unpredictable
local variations as a consequence of porosity reducing or enhancing effects.
The choice of the parametrization for Table 4 has some immediate and
evident effects that the reader must be aware of: we are implicitly assuming,
for example, that if we find a volume equal to 0 in a segment and we know
that a trap is in place, we have to blame the source for this (rows 3 and
4 in the table), but the same situation can also occur when both trap and
reservoir fail (rows 5 and 9), no matter which is the outcome of the source,
as it is natural to assume. When just one of these two elements fail, on
the other side, we still allow a marginal possibility of finding HC, and this
is resumed in the parameters R and S . We have fixed R and S to be
equal to 0.2. The effects are multiplicative, therefore a factor 0.2 reduces
the expected accumulations to 20% of the expected accumulation with all
the elements in place.
The second important point to discuss is how to assign the variances
to the Gaussian distributions. We acknowledge that this is a crucial point,
with large and important implications when analyzing the effect of nodes
behaviour, as shown in the previous paragraph. We have decided to assign
the variances in order to have a constant coefficient of variation in all the
possible scenarios described in Table 4; a constant coefficient of variation
will be our a standard hypothesis for the variability of HC volumes. The
variances built in this way happen to be in quite good accordance with
the variances from a parallel GeoX analysis with a much larger number
of samples. We need to stress, though, that we do not have a definitive
answer or suggestion to this point, since we will never be able to consider
and describe all the possible scenarios that could possibly happen in the
basin that we are considering.
The complete BN is shown in Figure 9. Since the accumulations cannot
be negative, we will concentrate in 0 the probability mass corresponding to
negative values. The resulting distributions will therefore be a mixture of
truncated gaussians distributions.
The effects of this parametrization on the HC distributions can be seen in
Figure 10. The distributions are truncated at 0, resulting in mixed discretecontinuous distributions. The probabilities of discovery are 0.919 for Top
Anticlinal and 0.797 for Bottom anticlinal.
The distributions are multimodal, and the different modes reflect the
likelihood of being in each of the 24 configurations taken into account. The
comparison between the empirical distribution (24 configurations, shown
with blue stars) and the BN distribution can be found in Figure 11. Here
18
Reservoir
1
2
1
2
1
2
1
2
1
2
1
2
Trap
1
1
2
2
1
1
2
2
1
1
2
2
Source
1
1
1
1
2
2
2
2
3
3
3
3
0
0
0
0
0
R
S
R + S
0
2 R
2 S
1
Table 4: Conditional Probability Table for the oil and gas accumulations in the four
prospects; the column represents the multiplicative factor assigned to the mean of the
gaussian conditional distribution.
0.06
Distribution Gas Ou (BE)
Distribution Gas Mmd (TE)
0.05
0.04
0.03
0.02
0.01
0
30
20
10
10
20
30
Volume (MMBOE)
40
50
60
70
0.01
Distribution Oil Ou (BE)
Distribution Oil Mmd (TE)
0.008
0.006
0.004
0.002
0
200
200
400
Volume (MMBOE)
600
800
1000
Figure 10: Oil and gas volume distributions in prospects BE and TE. The multimodality of
the distribution is due to failure of local geological elements that do not totally jeopardize
the likelihood of finding HC
19
the bivariate distributions for the states oil and gas for the main TE (left)
and BE (right) accumulations are shown. As we can see, there is a positive
correlation between the oil and gas accumulations, due to the positive effect
of TOC and HF in the maturation of the source rock. Second, the BN
distribution covers quite well the empirical distribution, though there are
discrepancies due to the thresholds introduced in Section 4.1 and to the
prior values (again learned from the data) imposed to the upper nodes of
the network. Recall that the main goal of this work is not to reproduce
exactly the BPSM behavior, but to integrate the results in a probabilistic
framework where it is easier to evaluate the effect of particular observables.
Nonetheless, we have good reasons to believe that our distributions would
constitute an ideal contour line (envelope) of a much larger range of scenarios
than our original 24, and therefore would capture most of the uncertainties
that characterize this case study.
4
x 10
1000
x 10
300
2.2
7
250
200
600
5
400
4
3
200
Volume Oil (MMBOE)
Volume Oil (MMBOE)
800
1.8
1.6
150
1.4
100
1.2
1
50
0.8
0
0.6
2
0
0.4
50
1
200
20
0
20
40
Volume Gas (MMBOE)
100
60
0.2
20
0
20
40
Volume Gas (MMBOE)
60
Figure 11: Oil and gas volumes joint bivariate distributions. Values are given for the
accumulations TE (left) and BE (right).
For economical purposes it is interesting to analyze the inverse cumulative distributions of recoverable HC. In order to compute such distributions
we need to take into account the recovery factor, that is estimated to be
0.45 for oil accumulations and 0.75 for gas accumulations. In Figure 12
we show the inverse cumulative distributions for segments TE and BE of
the anticlinal prospect. The black line represents the contribution of the
oil part, while the red line represents the added value brought by the gas
accumulation. As we can see the gas accumulation is more important for
prospect BE since this has a source rock maturity level sufficient to produce
commercial quantities of gas.
These distributions are immediately updated when more information
gets available. Let us focus our attention on the gas accumulation relative
to segment BE of the anticlinal prospect. In this case we may receive information that confirms our likelihood about the presence/absence of the
reservoir or the trap in that prospect. The network is updated, and the
20
Inverse cumulative distribution of recoverable resources, Anticlinal Mmd (TE)

1
Oil resources
Oil+gas resources
0.8
0.6
0.4
0.2
50
100
150
200
250
300
Volume, MMBOE
350
400
450
500
Inverse cumulative distribution of recoverable resources, Anticlinal Mmd (BE)

1
Oil resources
Oil+gas resources
0.8
0.6
0.4
0.2
50
100
150
Volume, MMBOE
200
250
300
Figure 12: Inverse Cumulative Distribution of recoverable resources for segments TE (Top)
and BE (Bottom) of the anticlinal prospect. In black volumes related to the oil accumulations, in red volumes composed of the joint contribution of oil and gas accumulations.
conditional accumulation distributions can be seen in Figure 13. The effect

of confirming an adequate reservoir layer is much stronger than a positive
trap, since the prior likelihood for the anticlinal trapping to be adequate is
already equal to 0.9, while the uncertainty about the quality of the reservoir
layer (porosity) is much larger.
5. Applications for decision making
In this section we demonstrate a couple of different applications of the
network.
5.1. What-if scenarios
We are interested in the behavior of the network in case of observing
a HC column in another prospect. In order to mimic a real situation, we
consider drilling a well on the anticlinal prospects, and observe the impact
of various evidence in TE, the top segment, on BE, the bottom segment
(Figure 14). We then compare with a similar observations made on the
fault prospect BW (Figure 15).
21
x 10
marginal BE oil
marginal BE oil | Res BE OK
marginal BE oil | Tra BE OK
50
100
150
200
250
300
Volume (MMBOE)
350
400
450
500
Figure 13: Distribution of the oil accumulation in segment BE before and after observing
positive Reservoir and Trap evidence in the same segment.
In the first figure we see that even a rich observation in BE is not sufficient to solve the bi-modality of the marginal distribution, since the possible
uncertainty about the quality of the reservoir remains (TE and BE belong
to 2 different reservoirs). In the second figure we see that both an extremely
poor and a rich observation in the fault prospect BW can substantially
change the shape of the posterior oil BE distribution. As we have already
pointed out, a positive HC column observation in a high risk prospect such
as BW confirms both the quality of the reservoir and the existence of a
charge, having a higher impact on BE than an observation in TE.
3
x 10
marginal BE oil
marginal BE oil | TE oil Acc = 0
7
6
5
4
3
2
1
0
50
100
150
200
250
300
Volume (MMBOE)
350
400
450
500
an oil column of different height in segment TE
22
0.01
marginal BE oil
marginal BE oil | BW oil Acc = 0
0.009
0.008
0.007
0.006
0.005
0.004
0.003
0.002
0.001
0
50
100
150
200
250
300
350
400
450
500
an oil column of different height in segment BW
5.2. Value of Information

Geologists and decision makers need to establish the probability of having recoverable HC larger than an economic threshold. We can use a similar
criterion for assessing the VoI, saying that a value of expected resources
falling below the economic threshold is equivalent to having no resources at
all. Furthermore, when we compute the VoI we have always to specify the
cost of collecting that information. In this case since our reference unit are
the volumes expressed in MMBOE, we will convert the costs in such units.
Given these premises, the prior value for segment i, threshold t and cost
C would be:
Z

P V (i) = max
P (xi = x) vx dx C, 0
x>t
The value of having free clairvoyance in segment i would then be:

Z

Z
X
V F C(i) =
max
P (xj = x|xi = e) vx dx C, 0 P (xi = e) de.
j{BE,T E,BW,T W }
x>t
and finally:
V oP I(i) = V F C(i)
P V (i).
i{BE,T E,BW,T W }
In these expressions the quantity vx is intended to be proportional to the

recoverable resources x. It is worth noticing that when i = j the integral
collapses in a single point, and we observe what is called self-evidence, i.e.
the effect of observing a prospect itself.
Since the distribution are numerically approximated, the integrals are
computed through a discretization and this makes the process computation23
ally intensive.
When computing the VoPI, we state that a certain prospect will be
drilled if its expected recoverable resources exceeds a certain threshold. We
have considered two possible scenarios for t, t = 0 and t = 80. The value t
may also represent risk averse behavior for the decision maker: the higher
is t, the more conservative is the decision maker. For each possible scenario
we have computed the VoPI for the four prospects for different costs C,
C representing the operational cost connected to developing the prospect.
We have decided not to introduce monetary units, but to refer everything
in MMBOE, that is the reference unit for the prospects volumes; for this
reason C is expressed in the same terms. We have repeated the procedure
with and without the self evidence.
VOPI, t=0
VOPI, t=80
50
50
VOPI TE
VOPI BE
VOPI TW
VOPI BW
40
40
30
30
20
20
10
10
50
VOPI TE
VOPI BE
VOPI TW
VOPI BW
100
Cost C
150
200
50
100
Cost C
150
200
Figure 16: Value of Perfect Information for the four prospects BE, TE, BW and TW, as
a function of the threshold t and of the operation costs C.
VOPI without SelfEvidence, t=0
VOPI without SelfEvidence, t=80
50
50
VOPI TE
VOPI BE
VOPI TW
VOPI BW
40
40
30
30
20
20
10
10
50
100
Cost C
150
VOPI TE
VOPI BE
VOPI TW
VOPI BW
200
50
100
Cost C
150
200
Figure 17: Value of Perfect Information without self evidence for the four prospects BE,
TE, BW and TW, as a function of the threshold t and of the operation costs C.
Results are in Figures 16 and 17. We can immediately see two major spikes, corresponding to the range of costs that affects decisions in the
biggest prospects, namely BE and TE. This means that for operation costs
in the regions close to the spikes, having the possibility of observing the
state of one of the prospects would sensibly change our decision about other
24
prospects. We can immediately recognize that the first spike corresponds

to a decision change in prospect BE, and the second spike to a change in
prospect TE. This is confirmed both by the VOPI computed without self evidence (the spikes corresponding to self evidence disappear, see for example
the dashed line of TE that goes to 0 in Figure 17 for high values of C), and
by observing which prospects have the highest impacts: BW in case of the
first spike and TW in case of the second spike. The geological reasons have
been discussed when commenting Figures 14 and 15, and they are confirmed
by this VOI analysis, which compresses the information into outputs useful
for decision making. Similar discussions and considerations can be found in
Martinelli et al. (2011).
The values of VOPI that we get from such analysis must be compared
with the exploration cost necessary to get that information, again expressed
in MMBOE. As we can see, VOPI values are much smaller than the operation costs C, and this is consistent since they need to be compared to the
exploration costs and not to the operation costs. If the exploration cost is,
say, equal to 10 MMBOE, the threshold is fixed to t = 0, and the development costs C are equal to 50, it is optimal (more informative) to focus on
segment BE; in this case TE is not very informative since its high volume
makes it profitable anyway. If the costs C are equal to 150, on the other
side, it is more informative to explore TE; in this situation even TW could
be a good candidate (its VOPI lies above 10 MMBOE for C = 150), while
TE becomes irrelevant since its volume does not cover the operation costs
and it is less informative than TW for estimating the outcome of BE. The
last comment is about the threshold t: we can see that higher values of t
lead to a shift towards smaller costs C for the VOPI peaks. This is reasonable since higher values of t reduce the chance of the prospects to be
commercially viable, and therefore make them interesting just if the operational costs are smaller. This effect is bigger for prospects with low volumes
(first and foremost TW and BW, but also BE), while it is almost impossible
to detect when the volume is large (TE), since the imposed threshold makes
this prospect very appealing in any situation.
6. Discussion and Conclusions
We have shown how BPSM can help in assessing the probability structure of the Bayesian network that models prospect and play element dependencies. The workflow moves from the Earth model to the decision space.
The geological and geophysical know-how is translated into BPSM. Outputs
from multiple runs of basin modeling under different geologic scenarios are
then used to establish the Bayesian network which is used to test decision
scenarios, and analyze value of information.
The work underlines the importance of assessing uncertainty in petroleum
systems. The emphasis is less on knowing the right answer, that may never
25
be known before drilling, but rather on determining the range of outcomes

given the available data and state of understanding of the petroleum system.
Problems are caused by the complex and often non linear interactions among
the different parameters, that make the prediction problem extremely difficult. Currently these problems are solved running a bunch of simulations
with different parameters, and studying the uncertainty in the resulting accumulation distribution as main or sole output. We believe that this process
is not sufficient any longer, since there are too many parameters that remain
hidden (implicit parameters) when the effect of many parameters is tested
at the same time. With our framework we provide an alternative solution by
making explicit all the correlation parameters, though not chosen arbitrarily
a priori, but derived from a multiple scenario evaluation.
Acknowledgments
We thank the Statistics for Innovation (SFI2 ) research center in Oslo,
that partially financed GMs scholarship through the FindOil project. We
acknowledge the Stanford BPSM group for the opportunity given to GM of
learning and practicing the software used in this work.
References
Allen, P., Allen, J., 2005. Basin Analysis, Principles and Applications. 2th
ed. Blackwell Publishings.
Bhattacharjya, D., Eidsvik, J., Mukerji, T., 2010. The value of information
in spatial decision making. Mathematical Geosciences 42 (2), 141163.
Chang, K., Fung, R., 1995. Symbolic probabilistic inference with both discrete and continuous variables. IEEE Transactions on Systems, Man and
Cybernetics 25 (6), 910917.
Corre, B., Thore, P., deFeraudy, V., Vincent, G., 2000. Integrated uncertainty assessment for project evaluation and risk analysis. SPE European
Petroleum Conference.
Cowell, R., Dawid, P., Lauritzen, S., Spiegelhalter, D., 2007. Probabilistic
Networks and Expert Systems. Springer series in Information Science and
Statistics.
Damsleth, E., Hage, A., Volden, R., 1992. Maximum information at minimum cost: A north sea field development study with an experimental
design. Journal of Petroleum Technology 44 (12), 13501356.
Dejean, J.-P., Blanc, G., 1999. Managing uncertainties on production predictions using integrated statistical methods. SPE Annual Technical Conference and Exhibition.
26
Fisher, R., 1971. The Design of Experiments, 9th Edition. Macmillan.

Friedman, N., Goldszmidt, M., 1996. Discretizing continuous attributes
while learning bn. Machine Learning: Proceedings of the International
Conference.
Hantschel, T., Kauerauf, A. I., 2009. Fundamentals of Basin and Petroleum
Systems Modeling. Springer.
Jordan, M., 1998. Learning in graphical models. Kluwer Academic Publishers.
Kaufman, G. M., Lee, P. J., 1992. Are wildcat well outcomes dependent
or independent? Working papers 3373-92., Massachusetts Institute of
Technology (MIT), Sloan School of Management.
Martinelli, G., Eidsvik, J., Hauge, R., Drange-Forland, M., 2011. Bayesian
networks for prospect analysis in the north sea. AAPG Bulletin 95 (8),
14231442.
Murphy, K. P., 1999. A variational approximation for bayesian networks
with discrete and continuous latent variables. Proceedings of the Fifteenth
conference on Uncertainty in artificial intelligence, UAI 99.
Tviberg, S., 2011. To assess the petroleum net present value and accumulation process in a controlled petromod environment. Master Thesis at the
Department of Geology and Mineral resources engineering, NTNU.
VanWees, J., Mijnlieff, H., Lutgert, J., Breunese, J., Bos, C., Rosenkranz, P.,
Neele, F., 2008. A bayesian belief network approach for assessing the impact of exploration prospect interdependency: An application to predict
gas discoveries in the netherlands. AAPG Bulletin 92 (10), 13151336.
27
Paper V
Dynamic exploration designs for graphical models using clustering
G. Martinelli and J. Eidsvik
Submitted for publication, 2012
Dynamic exploration designs for graphical models using

clustering: applied to petroleum exploration
Gabriele Martinelli and Jo Eidsvik
Department of Mathematical Sciences, NTNU, Norway
Abstract
The paper considers the problem of optimal sequential design for graphical models. The
joint probability model for all node variables is considered known. As data is collected, this
probability model is updated. The sequential design problem entails a dynamic selection of
nodes for data collection, where the goal is to maximize utility, here defined via entropy or
total expected profit. With a large number of nodes, the optimal solution to this selection
problem is not tractable. An approximation based on a subdivision of the graph is considered.
Within the small clusters the design problem can be solved exactly. The results on clusters are
combined in a dynamic manner, to create sequential designs for the entire graph. The merging
of clusters also gives upper bounds for the actual utility. Several synthetic models are studied,
along with two real cases from the oil and gas industry. In these examples Bayesian networks
or Markov random fields are used. The sequential model updating and data collection provide
useful guidelines to policy makers.
Introduction
Our interest is a sequential selection problem over dependent variables. The main motivation is
to construct policies for oil and gas exploration, where the outcomes at prospects are dependent
by spatial proximity or by common geological mechanisms. The probability of success for any
exploration well is then highly influenced by the outcomes at other prospects.
More generally the challenge is to construct an optimal dynamic design of nodes in a graph.
For instance, in the situation with a Bayesian Network (BN) or a Markov Random Field (MRF)
we evaluate which variables are most useful to observe. We assume a fixed probability model a
priori. As we acquire data at nodes in the BN or the MRF, the original probability distribution
is updated, according to Bayes rule. Relevant design questions are then: Which nodes are more
informative? Which sequence of nodes gives the best policy? In the petroleum industry drilling
wells is extremely costly, and getting the right information is critical.
At each stage of the dynamic strategy, we choose to observe one additional variable, or quit the
search. If we acquire data at a node, we incorporate the observation in the current (a priori) model
to compute the updated (a posteriori) model. For the next stage, the updated model serves as a
prior model, and so on. The sequential decisions account for two aspects: i) the immediate profit in
terms of monetary units or information gain by knowing the current variable, and ii) the expected
future benefits induced by the predictive capacity, conditional on the current variable. These two
aspects are combined in a utility function. If the expected utility of choosing one more node is too
small, we stop collecting data. The trade off between i) and ii) is related to more general explore
or exploit problems in decision making. An oil and gas company may want to target the most
1
2 SEQUENTIAL DESIGN
lucrative prospects, but it is also important to know the key variables, which give us the chance to
make better, informed, decisions at the later stages. The future values in ii) then play an important
role in the utility function.
With our focus on oil and gas exploration we note some similarities and differences with common
spatial design problems, e.g. Shewry and Wynn (), Zimmerman(), Le and Zidek (). The most
common problem treated in the literature is to allocate a fixed amount of monitoring stations to
improve overall predictive performance in some sense. The selection is thus done in the static sense,
not allowing the decision maker to modify her choices after observing the outcome at the previously
selected spatial sites. In this paper we consider the dynamic decision problem, with one observation
at a time and the ability to make sequential decisions. Moreover, in spatial design problems the
model is typically Gaussian. Our paper is new in the sense that is studies design for graphical
models with discrete outcomes at all nodes.
Our sequential design problem is a discrete optimization problem which is in theory solved via
dynamic programming (DP). This method defines a forward-backward algorithm that constructs
the optimal sequences and the expected utility. Bickel and Smith (2006) present a DP algorithm
tailored for our sequential design problem with dependent oil and gas prospects. However, their
approach is not applicable when the number of variables gets too large. For more than, say, ten
variables, we must instead look for approximate strategies. The appropriate solution seems to be
very case-specific. See e.g. Powell (2007) for more background. Various heuristic approaches are
important for special applications, but it is very difficult to assess the properties of these solutions.
For graphical models it seems natural to utilize the structure. One approach is to split the original
graph in several disjoint clusters. This idea was originally presented in Brown and Smith (2012).
They next solved the DP exactly for the clusters, and combined the results to get approximations
for the expected utilities on the full-size graph. The approach also allows an upper bound on the
utility, indicating the quality of the approximation.
Our main contribution in this paper is to use the clustering strategies for graphs to construct sequential designs for BNs and MRFs. A critical element in the method is to compute the cluster-wise
Gittins index. This extends the original index pioneered in statistics by Gittins (1979) and Whittle
(1980) for so-called bandit problems, studied by Benkerhouf et al. (1992) and Glazebrook and Boys
(1995) for oil and gas exploration problems. We consider the sensitivity of cluster orientation and
size, and various levels of approximation in the Bayes updating scheme. We use utility functions
based on entropy and more traditional cost/revenue aspects. For the situation with dependent
oil and gas prospects, the resulting designs can work as a road map for the exploration company.
In this way we combine statistical models and Bayesian updating with decision making to create
policies. Our focus is on oil and gas resources applications, but similar methods are relevant for
e.g. machine scheduling (Abdul-Razaq and Potts, 1988), medical treatments selection (Claxton
and Thompson, 2001), subset selection problems and more generic search problems.
The paper develops in the following way: in Section 2 we give the main ideas about sequential
design, in Section 3 we discuss how splitting the problem in clusters can help in building approximate
strategies, in Section 4 we provide results on synthetic examples, in Section 5 we show results on
real case studies.
Sequential design
A sequential strategy is illustrated in Figure 1 for the context of petroleum exploration. Here, we
initially choose to drill one of three prospects, or nothing. If we start by drilling prospect 3, the
design criterion for the next stage depends on the outcome of prospect 3. The decision is then to
2 SEQUENTIAL DESIGN
drill1
drill 2
we
drill 3
....
drill 1
....
drill 2
well 2
ll 3 oil
well 3
well 2
oil
dry
quit
....
....
....
dry
quit
Figure 1: Decision tree for a simple 3-nodes discrete example with two possible outcomes (oil or
dry) per node.
choose among prospect 1 and 2, or quit.
Similarly, the sequential design problem we consider here entails a selection of nodes, one at a
time, to maximize a utility function. We first introduce the statistical notation and assumptions
required to frame this sequential design problem. We next outline the theoretical solution given
by DP. A small example is then used to illustrate the sequential strategies resulting from different
utility functions.
2.1
Notation and modeling assumptions
Consider N nodes, and let xi {1, . . . , ki }, i = 1, . . . , N denote the discrete random variables.
Without loss of generality, we assume ki = k possible states for all nodes i. In Figure 1 k = 2 with
oil or dry outcomes. We represent the probabilistic structure for x = (x1 , . . . , xN ) via a graph. For
a BN defined by a directed acyclic graph the joint distribution is
p(x) =
N
Y
i=1
p(xi |xpa(i) ),
(1)
where pa(i) denotes the parent set of node i, which is empty for the top nodes. Undirected graphs
are defined via the full conditionals over a neighborhood, or, by the Hammersley-Clifford theorem,
via a joint distribution over clique potentials. For a first-order MRF (Besag, 1974) we use:
N
X
X
p(x) exp
I(xi = xj ) +
i (xi ) ,
(2)
ij
i=1
where i j denotes neighboring lattice nodes (north, east, south, and west). The parameter
imposes spatial interaction, while the i (xi ) terms include prior preferences about states at node i.
We assume known, fixed, statistical model parameters in p(x), such as and i (xi ) in equation
(2) and the conditional probabilities in equation (1). Associated with the probabilistic model
we can of course compute several attributes that are important for design purposes. Assuming
that we know the revenues or cost, denoted rij , for outcomes xi = j, the decision value (DV)
P
is DV (i) = max(0, kj=1 rij p(xi = j)), i = 1, . . . , N . This DV is useful for decision making.
It is non-zero
only when the expected profit is positive. The entropy (disorder) is defined by
P
H = log(p(x))p(x) = E(log p(x)), and the reduction in entropy is often used for design
purposes, see Wang and Suen (1984).
2 SEQUENTIAL DESIGN
In our sequential design situation, we rely on the ability to extract the marginal probabilities at
all nodes, and to update the probability distributions when evidence is collected. Since we are going
to update the model at each stage of the sequential strategy, for many different kinds of evidence,
we require these computations to be reasonably fast. For BNs the updating of probabilities can
be done effectively by the junction tree algorithm (Lauritzen and Spiegelhalter (1988)). MRFs
can similarly be updated by forward-backward algorithms, see e.g. Reeves and Pettitt (2004) and
Tjelmeland and Austad (2012).
Assume we can acquire data at one node in the graph, and incorporate the outcome to get a
posterior distribution. For the next stage, this updated distribution serves as a prior distribution.
We can then select another node, acquire information, update the probabilities, and so on. The
sequential design of nodes is constructed by optimizing the expected utility, which means that we
integrate over all possible data when finding the optimal sequence. In our case, the utility is based
on monetary profits or entropy reduction. One could of course imagine other selection criteria here.
Minimum entropy entails a dynamic design that attempts to stabilize or minimize the uncertainty
in the graph.
Let i be the observable or evidence in node i = 1, . . . , N . If node i is not yet observed, we
set i = . If we choose to observe node i, i is the actual outcome of the random variable xi at
this node. For instance, in a petroleum example, i = 1 can mean that prospect i has been drilled
and found dry, i = 2 if found gas, and i = 3 if oil. A priori, before acquiring any observables,
we have = 0 = (, . . . , ). When we observe nodes, we put the outcomes at the corresponding
indices of the vector . Say, if node 2 is selected first, and observed in state 2 = x2 = 1, we
set = (, 1, , . . . , ). At each stage, one more entry of is assigned. The posterior that is
updated at every stage of the sequential design is generically denoted by p(x|), with marginals
p(xi = j|), i = 1, . . . , N , j = 1, . . . , k. Since we get perfect information about the selected node
variables, we get p(xi = j|) = 0 or 1 if node i is already observed.
In our setting it is important to monitor the design criterion or utility at all stages of sequential
conditioning.
P When we get evidence , the entropy is reduced, so that H() H 0. For the DV
we have DV (i|)p() DV (i), where the probabilities for the DVs in the sum are conditional
on the evidence . This entails that the pre-posterior DV is always larger than the prior value, and
the value of information is always non-negative, see Bhattacharjya et al. (2010). The sequential
design will be guided by immediate entropy reduction or gain in monetary value, as well as the
expected future impact an observable can have.
2.2
Dynamic Programming for Sequential Design
The sequential design procedure forms a decision tree, where a fork represents a decision to choose
a node (or quit), and each branch points to the future decisions, and the conditional utilities,
depending on the outcome of the chosen node (Figure 1). See also Cowell et al. (2007), Chapter
8. We next present the method of DP to solve the sequential design problem. This algorithm
computes the utilities of all possible sequential designs, and then picks the most lucrative sequence.
We first consider expected profit as utility function. This criterion is relevant for the petroleum
examples with N prospects to explore and hopefully produce. Let v() represent the expected
revenues, i.e. future cash flows, given that we are in observation state . Initially, the vector of
observables is empty: 0 = {, , . . . , }, and the value is v( 0 ). DP computes v( 0 ) and finds
the associated optimal sequential design.
At the first stage we select the optimal node i among all nodes N , or quit. The expected initial
2 SEQUENTIAL DESIGN
value becomes
v( 0 ) = max
iN
k
X
p(xi = j)
j=1
"
rij
+ max
sN /i
( k
X
p(xs = l|xi =
j)(rsl
+ . . .), 0
l=1
)#
,0
where the second and the subsequent maximizations are over nodes not yet considered in the
sequential strategy. Here, is a discounting factor. In practice, a near 1 encourages learning
the dependent model, while a smaller means that we choose the bigger DV s at the early stages.
Note that the expected value contains immediate profit (rij ) and a continuation value (CV) with
conditioning on the outcome of variable xi = j in the selected node. For short, we can write the
expected revenues by starting at node i by
vi () =
k n
o
X
p(xi = j|)(rij + v( ji )) ,
(3)
j=1
where ji = { -i , i = j} and v( ji ) is the CV of the state ji , i.e. v( ji ) = maxl6=i {0, vl ( ji )}. If

we know the outcomes at all nodes, the CV is v(, , . . . , ) = 0. This forms the starting point of
DP, which proceed backwards, one step at a time, extracting the solutions for all sequences. We
show an example in the next section.
As suggested in e.g. Weber et al. (2000) and Krause and Guestrin (2009), the reduction in
entropy is useful for design. Let H() be the (conditional) entropy with current evidence . We
construct a sequential design based on
H( ji ) = H() H( ji ),
i.e. the reduction in entropy caused by additionally observing xi = j at one stage in the strategy.
This entropy reduction can be computed efficiently utilizing fast updates of BNs and MRFs, and
the conditional properties of entropy. We again introduce a price Pi of observing node i. This is
now fixed, no matter the outcome of node i. Similar to what we did for the profit-based utility,
we set the CV as the possible future reductions in entropy brought by the new observation. The
expected utilities when including node i in the design becomes
vi () =
k h
X
j=1
i
H( ji ) Pi + v( ji ) p(xi = j|), i = 1, . . . , N.
(4)
The decision maker selects the node with the highest vi (), i.e. the most informative part of the
graph. If no nodes contribute with positive values, we quit the search. This means that the price
Pi exceeds the expected immediate gain and future information reduction. When there are no
more nodes to observe, the CV is 0. Similar to the situation above, the DP constructs the optimal
sequential designs of nodes, and computes the associated reductions in entropy. An example is
presented in the next section.
Note that the current situation with sequential decisions can also be phrased as a Markov
Decision Process, where the generic state of the system develops as a function of the actions at
each stage, see e.g. Puterman (2005). No matter what method we use for the sequential design
problem, the computational cost grows exponentially with the number of nodes N . In the example
below we construct optimal selection strategies among eight nodes. For graphs much larger than
this, exact DP is not possible, and we outline the new approximate strategies in Section 3.
2 SEQUENTIAL DESIGN
2.3
A simple motivating example
We now present the DP strategies driven by the cost/revenue utility function and entropy on
a small example. The BN case study is shown in Figure 2. Here, the eight leaf nodes can be
observed, {1A, 2A, . . . , 5C}, while the remaining six auxiliary nodes, {K, P 1, . . . , P 5}, impose the
desired (causal) dependency structure in the BN (See Section 5). The goal is to determine where
to observe first, and which would be the consequent choice, given data at the first node, and so
on. We assume the initial probability structure of the BN is fixed. Each node has binary outcomes
(k = 2). Inspired by the petroleum exploration, we refer to these two by oil and dry.
K
5C
5B
1A
P1
P5
5A
2A
P2
4B
4A
P4
P3
3A
Figure 2: Simple example used in Section 2.3. We can collect data in the leaf nodes. By Dynamic
Programming we construct optimal sequential designs that maximize expected utility.
The main input parameters for this example are given in Table 1. There is much dependence
in the BN. Based on cost/revenues only two nodes, 3A and 4A, have positive marginal expectation
E(xi ). The DV is 0 when this value is non-positive. Thus, a naive decision maker, looking for
profit, and ignoring the dependence between nodes, would forget about six of the prospects. The
naive value of the field is 661 + 514 = 1170 for a specified discounting of = 0.99.
Prospect
p(xi = 0)
p(xi = 1)
Entropy reduction
Costs
Revenues
E(xi )
1A
0.44
0.56
0.6859
3000
1368
-554
2A
0.46
0.54
0.6899
900
707
-32
3A
0.48
0.52
0.6920
2400
3443
661
4A
0.61
0.39
0.6682
1800
4151
514
4B
0.70
0.30
0.6129
600
1321
-19
5A
0.40
0.60
0.6743
1500
943
-41
5B
0.48
0.52
0.6922
3600
3254
-20
5C
0.48
0.52
0.6922
2100
1887
-18
Table 1: Input parameters for the example in section 2.3: Marginal probabilites, marginal entropy
reductions and monetary parameters.
An optimal decision maker, using (3) and (4), would account for the ultimate consequences of
the actions. Results are shown in Table 2. We here compare the outcomes of the naive and myopic
strategies with the optimal using DP.
The myopic (nearsighted) strategy relies on forward selection, as opposed to the forwardbackward approach of DP. Using cost/revenue utility, the myopic strategy starts from the most
lucrative prospect 3A. If this variable is dry, we update the network and find out that all the DV
are negative. In particular P (4A = oil|3A = dry) = 0.975, and this ensures that prospect 4A is
3 CLUSTERING STRATEGIES FOR LARGE GRAPHS
no longer attractive. If 3A is oil, the success probabilities in most nodes increases significantly. In
fact, six of the seven remaining DVs are positive. The myopic approach goes for the greatest of all
DVs, and selects 4A as the next candidate. If 3A is oil and 4A is dry, we still have one positive DV.
Not surprisingly, this is the prospect above 3A in the graph, and we go for prospect 2A. If both 3A
and 4A are oil, we go again for the most lucrative prospect which is 5B.
The optimal DP solution defines values vi ( 0 ) in 3) are the following: [3352, 3952, 3595, 3427, 3852,
3926, 3443, 3738] for [1A, 2A, 3A, 4A, 4B, 5A, 5B, 5C]. Note that all these values are much bigger
than the naive value of the field, which is natural since the correlation in the graph is high. The
first selected prospect is then 2A, which has an intrinsic value close to 0, but a large influence on
the neighboring nodes. If 2A is dry, we focus on another area (prospect 5A). If 2A is oil, we remain
relatively close (3A). For the second stage, in the event of 2A dry: If 5A is dry, the network has
been entirely killed, and we stop observing. If 5A is oil, we remain in the same area (5B). The
third stage is shown in 2.
For the entropy-based design, we again compare myopic with a full DP based strategy. For
myopic the first node selected is either 5B or 5C, since their contribution to the reduction of the
entropy is highest (see Table 1). No matter if segment 5B is found dry or oil, we move away from
the 5-nodes, since most of the uncertainty in that part of the graph has been resolved.
Using DP, the first selected node is 1A. This node has a balanced prior probability and a high
impact on the probability structure in the network. In fact, the entropy values vi ( 0 ) are now as
follows: [0.8534, 0.8487, 0.8066, 0.7850, 0.7713,0.8520, 0.8353, 0.8353]. Nodes 4A and 4B, which are
characterized by prior probabilities far from 0.5, get the lowest initial entropy reduction. Node
4A is nevertheless selected when 1A is dry and 5A is oil (see Table 1). In this situation, when
the left and right part of the network has been explored, 4A is the one with the highest marginal
uncertainty, p(4A = oil|1A = dry, 5A = oil) = 0.445. The price P = Pi is set relatively low, and
under the entropy criterion we keep observing no matter the outcomes of the first two nodes.
Strategy
i(1)
i(2) |xi(1) = dry
i(2) |xi(1) = oil
i(3) |xi(1) = dry, xi(2) = dry
i(3) |xi(1) = dry, xi(2) = oil
i(3) |xi(1) = oil, xi(2) = dry
i(3) |xi(1) = oil, xi(2) = oil
Naive M
3A
4A
4A
Q
Q
Q
Q
Myopic M
3A
Q
4A
Q
Q
2A
5A
Sequential M
2A
5A
3A
Q
5B
5A
4A
Myopic E
5B
2A
4A
1A
1A
2A
1A
Sequential E
1A
5A
2A
2A
4A
5C
5B
Table 2: Results of sequential design for the motivating example. Utility is monetary based (M)
and entropy based (E). Here, i(1) , i(2) and i(3) are the first, second and third nodes selected. Q
means to quit the strategy.
Clustering strategies for large graphs
For large graphs the number of possible scenarios to evaluate exceeds what is computationally
tractable. The objective function in (3) must then be approximated in some way. Brown and Smith
(2012) use clusters to overcome the computational limitations and to get an upper bound for the
expected utility. We now apply this method to build sequences. We study different complexity
levels in the sequential Bayes update of the probability structure.
3.1
Sequential strategies based on clustering
The idea is to partition a large graph in smaller subgraphs, which can be computed efficiently. Let
d
C d , d = 1, . . . , L be disjoint nodes of the entire node set N , i.e. C d C e = , and L
d=1 C = N .
d
We denote by xC d the random variables in cluster C , and C d the cluster specific evidence. The
number of nodes in cluster C d will be in the order of one to around ten. The approximations we
present here improve as the cluster sizes grow, with a large increase in computational cost. As an
example of the increase in computing time, consider a situation with binary outcomes k = 2. The
computing time for evaluating a size 2 cluster is about 0.007 seconds, for 5 nodes we have 0.37 sec,
and for nine nodes it is 50 seconds.
To construct an approximate sequential design, we suggest to rank the clusters and select the
optimal node within the best cluster. The ranking is based on DP within clusters, given the current
information. Once we collect data in a cluster, we update the probabilities, use DP again, and get
a new ranking. This provides the basis for the selection at the next stage of the sequential design.
It is important here to introduce the Gittins index (GI), see Gittins (1979) and Whittle (1980).
We consider the cluster-wise GIs in the spirit of Brown and Smith (2012). First, consider a variation
of equation (3), with a generic retirement value M instead of 0 in the decision rule. Moreover,
assume that this DP equation is set up for each cluster, given the current evidence. We have
expected value for cluster d given by:
)#
( k
"
k
X
X
, M . (5)
v d (, M ) = max
p(xs = l|xi = j)(rsl + . . .), M
p(xi = j) rij + max
iC d
sC d /{i}
j=1
l=1
Now, when the computation is restricted to cluster C d , the GI is MC d (), defined as the smallest
retirement value M such that v d (, M ) = M . This is the value which makes the decision maker
indifferent between retiring and continuing the sequential strategy. Below, we will discuss various
levels of conditioning on the generic evidence in (5).
Brown and Smith (2012) derived some important properties for the value function v d (, M ),
for any evidence . Figure 3 illustrates the value functions for one of the examples below. Here,
we plot v d ( C d , M ) M for some clusters related to an example below, for fixed evidence. The
GI corresponding to each cluster is the crossing point with the first axis. Note that the ordering of
the clusters is not monotone in M , indicating the changes in decision paths within the cluster. The
cluster-wise GIs determine the cluster to be selected at the current stage of the sequential design.
We find the cluster with the largest GI by gradually reducing the M in conjunction with DP for
the clusters. Since the value function is piecewise linear, solving v d ( C d , M ) M = 0 for fixed
M, is relatively fast. Alternatively, one can use theory from Markov decision processes to tranform
the DP into a linear programming problem, see Chen and Katehakis (1986) and Brown and Smith
(2012).
To study the policies induced by this cluster strategy, we suggest to generate realizations of
the selected nodes. This entails running hypothetical scenarios where we sequentially observe (i.e.
sample the outcomes) of the chosen nodes, update the probabilities, and then proceed to the next
stage. At subsequent stages we may move between clusters or stay in the same cluster. The chosen
cluster often depends on the outcomes at the previously selected nodes.
Note that the updating step can be quite time-consuming. In principle, the evidence vector in
(5) is the full observation sequence until this stage, in all clusters, not only in cluster d. This means
that all cluster probability models must be updated when we acquire new data at a node. A faster,
but more approximate strategy is to update only the cluster where the current data is collected.
This means that just one of the GIs changes at each stage. We implement both of these methods
4000
3500
v(Cd,M)M
3000
2500
2000
1500
Highest GI
1000
500
0
0.5
1.5
M
2.5
3
5
x 10
Figure 3: Illustration of GIs: Values of v( C d , M ) M for six clusters, as a function of retirement

value M . The GI is the crossing point with the first axis. The cluster with highest GI is selected
in the sequential design.
for updating the probability distribution given the sequential observations. More specifically, we
have:
Multiple clusters update (MCU): We rank the cluster according to their GIs. DP in the
best cluster gives the first node. We update the probability model for all the clusters, given
the observation (sample outcome) in the selected node. All cluster GIs are also modified
based on the updated probabilities. Then, we choose the best cluster at the second stage
using these GIs. We proceed until all the nodes have been observed or there are no more GIs
greater than 0.
Single cluster update (SCU): We rank the clusters on the basis of their GIs. We start
from the cluster with the biggest index, and select a node according to a DP strategy within
the cluster. We update the joint distribution just for that cluster, given the observation in the
first node. The probability model for the other clusters are not updated, and the GIs for other
clusters then stays the same. Then, we choose the cluster with the highest GI among the
updated cluster and the other initial cluster values. The cluster with highest GI is selected,
and the best node based on DP in the cluster is chosen. We update this cluster only, given
all data acquired so far. We continue until all the nodes have been observed, or until there
are no more GIs greater than 0.
The MCU method is of course more consistent since we update the entire probability structure, and all GIs, whenever new information is available. The drawback is the computational
cost required to recompute the joint probability distribution and apply the DP strategy in every
cluster after each observation. The SCU method is faster since only one cluster GI needs to be
10
updated at each stage. However, the sequential design from SCU could suffer lack of accuracy.
Pseudo-algorithms constructing sequences over Monte Carlo samples (observations) are described
in Algorithm 1 and 2.
Algorithm 1 Evaluating a Single cluster Update strategy
= [, , . . . , ]
seq = [ ]
Sample t p(x)
# Current sample
for Clusters d = 1 : L do
[vC d , sC d ] = v( C d )
# Initial cluster-based DP values
# Initial GIs
GIC d = M : v( C d ) M = 0
end for
while d : vC d > 0 do
C = arg maxd {GIC d }
# Best cluster
seq = [seq sC ]
# Best node in cluster C
# Set sampled outcome tsC at selected node sC
s C = t s C
t
[vC , sC ] = v( C ,sCC )
# Updated cluster-based DP value for cluster C
GIC = M : v( C ) M = 0
# Updated GI for cluster C
end while
Algorithm 2 Evaluating a Multiple clusters update strategy

= [, , . . . , ]
seq = [ ]
Sample t p(x)
# Current sample
[vC d , sC d ] = v( C d )
# Initial cluster-based DP values
GIC d = M : v( C d ) M = 0
# Initial GIs
end for
while d : vC d > 0 do
C = arg maxd {GIC d }
# Best cluster
seq = [seq sC ]
# Best node in cluster C
# Set sampled outcome tsC at selected node sC
s C = t s C
t
[vC d , sC d ] = v( C d ,s C )
# Updated cluster-based DP value for cluster C d
C
GIC d = M : v( C d ) M = 0
# Updated GI for cluster C d
end for
end while
3.2
Computing independent and sequential lower bounds and an upper bound
Associated with a cluster-based sequential design we can approximate the expected utility value
v( 0 ). Of course, the clustering strategy gives a sub-optimal value compared to the full DP solution,
but the optimal one is not tractable for large graphs. A useful aspect of the clustering approach is
that we can get upper bounds for the value v( 0 ) by using clairvoyant information.
Let us first discuss various ways of approximating v( 0 ). The Monte Carlo strategies in Algorithm 1 and 2 provide a sampling-based approach for estimating this value. Here, each Monte
11
Carlo sample constructs a design sequence which depends on the outcome at the selected nodes.
We sum the tsC selected at every step of the while cycle in Algorithm 1 and 2, possibly with
discounting. Finally, these output values are averaged over B Monte Carlo runs. The estimates
will differ between MCU and SCU, since the full updating scheme gives less better sequences on
average. A challenge with this Monte Carlo sampling approach is a large associated Monte Carlo
error for moderate B.
Simpler approximations exist if we disregard the discounting. For instance, we get a lower
bound on the intial value v( 0 ) through an independent evaluation on each of the clusters: Let
v( C d ) be the DP value computed based on the evidence vector restricted to cluster d as follows:
( k
"
)#
k
X
X
v( 0,C d ) = max
p(xi = j) rij + max
p(xs = l|xi = j)(rsl + . . .), 0
,0 .
(6)
iC d
sC d /i
j=1
l=1
A lower (independent) bound for the expected utility is defined by the sum of the marginal values
for all the clusters:
L
X
v( 0,C d ).
vLB(1) ( 0 ) =
d=1
Clearly, vLB(1) ( 0 ) v( 0 ), since this cluster-by-cluster approach ignores the dependence between
clusters. However, this procedure requires no simulations, and if the clusters are chosen well, the
bound can be reasonable.
This lower bound defined via (6) can be improved by sequential cluster selection. Assume we
start by evaluating the cluster with the highest GI. Its value is v( 0,C d ). We next generate an
outcome for this cluster td , and use DP restricted to this cluster, plugging in the sampled data at
the selected nodes. Based on this we update the probability model at the remaining L 1 clusters,
and choose the next cluster with highest GI, say C e , and so on. This sequential cluster average
value defines an improved lower bound. Over B Monte Carlo samples we have
" L
#
B
1 X X
C e<d
vLB(2) () =
v( C d |tb
) ,
B
b=1
d=1
e<d
where the conditioning is the empty set for d = 1, and tC

is the bth sample restricted to all clusters
b
considered previously. Because of the imposed learning, we have vLB(1) () vLB(2) () v().
The quality of this sequential strategy depends on the choice of clusters and on the Monte Carlo
sample size B.
We next consider the construction of an upper bound. This is based on clairvoyant information
in the sequential strategy. This means that we know the outcome of all other nodes, and use this
when making decisions at the current stage. Since we are using information that is not really
available in practice, we get an upper bound for the initial value v( 0 ). This works as a benchmark
for sequential strategies. Together with the various lower bounds, we can squeeze the initial value.
The Monte Carlo strategies in Algorithm 1 and 2 can be extended to provide a sampling-based
approach for the upper bound. Now, at each stage, the GIs are computed by DP within-cluster,
using the updated probability model, given the cluster evidence available at the current stage (if
any), and all sample values outside the cluster. If we again disregard discounting, we can solve the
clairvoyant bound separately for each cluster. In this case we compute the value of a cluster, given
all observations (samples) outside the cluster, i.e using clairvoyant information. This calculation
requires the full conditional for all clusters, given the sampled outcomes in other clusters, but there
4 SYNTHETIC EXAMPLES
12
is no computation of GIs. The upper bound of the initial value is in this case:
" L
#
B
1X X
C e6=d
vU B () =
v( C d |tb
)
T
b=1
C e6=d
where tb
d=1
is the bth sample restricted to clusters different from C d . In summary, we get

vLB(1) () vLB(2) () v() vU B ().
Synthetic examples
We first study small BNs and MRFs to compare various cluster configurations. The number of
nodes is at most 12, and we manage to compare the clustering sequences and values with the
optimal solution obtained by full DP.
4.1
Small BN, Entropy utility
The entropy reduction is relevant in many applications, see e.g. Marcot et al. (2001) and Aalders
et al. (2011). When the BNs get large, and sequential strategies are requested, the current approach
should be interesting.
We refer in this section to the DP defined in equation 4: we run it on two small BN, shown
in Figure 4. The two BNs are small clusters of a bigger network, connected through a Common
Parent (CP) node. Both BNs have 5 nodes that represent sites or variables that can be selected.
In the first network the structure is made by a common node and 4 children, while in the second
network the structure is linear with two chains departing from a common top node.
CP
1R
1L
2R
2L
3L
4L
5L
4R
3R
5R
Figure 4: Simple BNs used for testing the entropy criterion, connected through a
Let us start with the network on the left. Each node is binary, with two states that we will
indicate as A and B. The top node has a symmetrical prior probability distribution, with 50% chance
of being in state A and likewise of being in state B. Nodes 2 and 3 have a CPT with propagation
of information just through state A, while nodes 4 and 5 have perfectly balanced CPT, as shown
in Table 3
The original entropy of the network in configuration {, , , , } is 2.3615. We intuitively
may expect that the reduction in entropy is higher if we observe node 1; actually, as we can see
from Table 3, since the probability distribution in node 4/5 is symmetrical, we observe the same
x2 , x3 \ x1
A
B
13
A
0.9
0.5
x4 , x5 \ x1
A
B
B
0.1
0.5
A
0.9
0.1
B
0.1
0.9
Table 3: CPT for Multi Level Network, from level nodes to children nodes
reduction for both node number 1 and nodes 4/5. We mean that:
(
(
)
)
X
X
X
X
p(x1 )
p(x4 )
p(x|x1 )log(p(x|x1 )) =
p(x|x4 )log(p(x|x4 )) = 1.6684
x1
x4
In particular, the entropy is substantially reduced if we observe the state A in either of these
nodes (in such case a single configuration, {A, A, A, A, A} collects the 65% of conditional probability). If we observe B in 1 or 4/5, on the other side, the result is having four configurations
equally likely, {B, A, A, B, B}, {B, A, B, B, B}, {B, B, A, B, B} or {B, B, B, B, B}, each of them
with a little more than 20% conditional probability of occurrence. The situation is different is we
observe A in nodes 2/3: here, the configuration {A, A, A, A, A} is still the most likely, but with a
mere 46% of occurrence since we are more uncertain about the occurrence in prospect 1 than with
a corresponding observation in nodes 4/5. Finally, observing B in node 2 (or 3), there are two
equally most likely configurations, namely {B, B, A, B, B} and {B, B, B, B, B}, since the marginal
likelihood of a B observation in node 1 in increased, but this has no effect on the conditional
distribution for node 3.
The overall effect is that an observation in 1 or 4/5 produces an average decrease in entropy of
0.6931, higher than the reduction brought by an observation in nodes 2 or 3, equal to 0.6109. This
is consistent with the intuition.
The question now is, are the results consistent with the intuition even when a full DP strategy
is taken into account, i.e. when we have the possibility of keep on drilling until the reduction is
higher than a certain cost P ? The results (final values for all the nodes) are reported in Table 4.
DP, P=0.2
DP, P=0.5
DP, P=0.65
Myo, P=0.2
Myo, P=0.5
Myo, P=0.65
1
1.3615
0.3863
0.0863
0.4931
0.1931
0.0431
2/3
1.4828
0.3803
<0
0.4109
0.1109
<0
4/5
1.4828
0.4234
0.0823
0.4931
0.1931
0.0431
Table 4: Final values of the DP and Myopic strategies applied to the network on the left in Figure
4, for different prices P of experiment, and for = 1.
We can immediately see some surprising results: in the myopic case, the first best choice is
determined just by the reduction in entropy brought by the first node, and therefore it is not
surprising that nodes 1 and 4/5 emerge as winners no matter what price Pi = P of data collection.
If the price is higher than 0.6931, no node is profitable, because this is the maximum reduction in
entropy that we can possibly achieve with a single observation.
The situation when taking into account DP strategies is more complex: if the cost is very high
(0.65), than we might have to stop after a single observation, at least if the reduction after the
14
first observation is already consistent. Thats why node 1 is selected as best choice. We have to
remember here, that though the average reduction for node 1 and nodes 4/5 is equal, the marginal
entropy of {A, , , , } is different (namely smaller) than {, , , A}, therefore where an
observation A is 1 could be sufficient, an observation A in node 4/5 could not be sufficient: this
reflects in the different final values. When the cost is medium (0.5), node 4/5 are selected first.
Here we choose to observe usually 2 or 3 nodes (depending on their outcome), and in such case
starting in 4 or 5 is optimal. When the cost is very small (0.2) it is almost always convenient to keep
observing up to end, and this makes the values for nodes 2/3 and 4/5 optimal and identical (since
the discounting is 0). The only exception is when starting in node 1. In such case we might stop
one step before, and the effect is an overall smaller reduction in entropy. Is this necessarily bad? It
is true that we have achieved a smaller reduction, but with a smaller number of observation! It is
likewise true, though, that if the criterion is keep collecting evidence until the marginal reduction
is smaller than the cost C, the strategy of starting in 2/3 or 4/5 is optimal, as identified by the
algorithm.
In order to understand if the role of node 1 is really central, we have ran the same strategies
on the network on the right in Figure 4, and we have imposed for all the nodes symmetrical CPT,
such that the marginal reduction in entropy is now equal to 0.6931 for all the 5 nodes. Results are
in Tables 5.
DP, P=0.2
DP, P=0.5
DP, P=0.65
Myo, P=0.2
Myo, P=0.5
Myo, P=0.65
1
1.2135
0.2055
0.0431
0.4931
0.1931
0.0431
2/3
1.1607
0.3139
0.0431
0.4931
0.1931
0.0431
4/5
1.2135
0.4139
0.0431
0.4931
0.1931
0.0431
Table 5: Final values of the DP and Myopic strategies applied to the network on the right in Figure
4, for different prices P , and for = 1.
The results for the three nodes are equal in the myopic strategy (as expected, since the marginal
reduction in entropy is the same) and in DP with high costs (0.65), since in this case just an
observation is allowed; with higher discounting even when costs are medium (0.5), the three results
are identical and just one observation is allowed. In every other case it is optimal to start far from
the center, since when 2 or more observations are allowed the reduction in entropy achieved by
observing, say, nodes 4 and 5 is higher than the reduction achieved by observing 1 and any other
node.
When considering the full network, we can consider the two subnetworks as two different clusters.
The first cluster (left subnetwork) results in an average case (C = 0.5) the one with the highest
GI, therefore it is selected first. This should not be surprising, since we have already seen that this
is a more complex case with a total entropy of 2.36, higher than the tool entropy of the second
cluster, equal to 1.99. Within this cluster, the nodes 4 or 5 are the most valuable, as confirmed
also in Table 4. After observing either of these two nodes, cluster 2 becomes more valuable, and
the decision maker therefore should move there to find a prospect that reduces the entropy. The
suggested nodes are again 4 or 5, as indicated also in Table 5, on the left. The choice is in this case
independent on the outcome of the first selected node, but this is in general not true: in particular,
as we have seen before, if the outcome is A, the entropy in the first cluster is drastically reduced
and the indication of moving to the second cluster is strong; if the outcome is B, the entropy is just
15
slightly reduced, just enough to make cluster 2 more valuable in terms of GI. The strategy suggests
to keep alternating nodes from cluster 1 and from cluster 2, following the same criteria.
This small example is synthetic, but similar situations may arise in practical real problems in
fields such as biology, fishing, and natural resources collection in general. We can think that the
two sub clusters are different areas that share little information, and we have to place monitor
stations in order to maximize the entropy reduction of our data. In this situation, the proposed
approach would suggest where to place the first stations, following a sequential approach.
4.2
Small BN: revenues/ cost
We focus our attention on two small BNs (Figure 5) with 12 correlated nodes; in this setting
(revenues/costs based utility function) we will often use the word prospect in stead of node, since
the most natural application is in petroleum explorations. The small dimension allows an exact
solution of the problem. The idea is to show how a different network structure influences the quality
of the bounds, and to compare these bounds with the ones produced by approximated sequential
strategies presented in Martinelli et al. (2011a).
The first case study (on the left in Figure 5) shows 12 prospects mutually correlated through
a single common parent. For a similar use of common parent networks in oil and gas exploration
contexts, see also Martinelli et al. (2012).
CP
10
11
12
10
12
11
CP
10
11
12
1
10
12
11
CP
10
11
10
12
12
1
11
CP
2
10
11
12
10
12
11
CP
2
10
11
12
10
12
11
Figure 5: Small BN case studies used in Section 4.2.

We are first interested in studying the outcome of sequential strategies under different clustering
configurations. It is important to keep into account Table 6, where the marginal probabilities and
the intrinsic values for all the 12 prospects are stored.
Let us assume we are still in an oil exploration setting, with two possible outcomes, oil and dry,
and for simplicity let us consider here strategies SCU. We have few prospects with positive IV, the
Prospect i
p(xi = dry)
p(xi = oil)
Intrinsic Value
1
0.5
0.5
-108
2
0.5
0.5
375
16
3
0.5
0.5
-657
4
0.5
0.5
-711
5
0.5
0.5
360
6
0.5
0.5
-184
7
0.5
0.5
-172
8
0.5
0.5
2417
9
0.5
0.5
-642
10
0.5
0.5
815
11
0.5
0.5
1088
12
0.5
0.5
-949
Table 6: Marginal probabilities and Intrinsic Values for the 12 prospects of the case studies in
Section 4.2
most prominent being number 8, 10 and 11.

# Clusters
Cluster size
i(1)
i(2) |xi(1) = dry
i(2) |xi(1) = oil
i(3) |xi(1) = dry, xi(2) = dry
i(3) |xi(1) = dry, xi(2) = oil
i(3) |xi(1) = oil, xi(2) = dry
i(3) |xi(1) = oil, xi(2) = oil
Average Value
1
12
8
10
2
5
1
5
11
7295
2
6
8
10
11
11
7
10
5
6992
3
4
8
10
10
11
7
11
11
6779
4
3
8
10
10
11
7
11
11
6638
6
2
8
11
11
7
7
10
10
6398
12
1
8
11
11
10
10
10
10
5783
Table 7: Results of the sequential exploration program for the simple BN example described in
Section 4.2, for strategies with different cluster size. i(1) , i(2) and i(3) are respectively the first, the
second and the third best node selected. Q means quit (the strategy).
The network is designed in order to allow for both positive and negative correlations between
the prospects. This means that, for example, finding oil in prospect 8 boosts the probability of a
discovery in prospect 11, but lowers the probability of finding oil in prospects 10 and 7, as shown
in the CPTs in Table 8.
P1, P4, P7, P10 / CP
0
1
P2, P5, P8, P11 / CP
dry
0.2 0.8
dry
oil
0.8 0.2
oil
P3, P6, P9, P12 / CP
0
1
dry
0.6 0.4
oil
0.4 0.6
0
0.8
0.2
1
0.2
0.8
Table 8: Conditional Probability Tables for the children nodes of the case study presented in Section
4.2, shown in Figure 5, on the left
Let us now consider and discuss the results of Table 7, where strategies run under different
clustering configurations are compared. We begin from the 2-clusters approach (second column):
first prospect to be selected is prospect 8, characterized by the highest IV. If prospect 8 is found
initially dry, we choose prospect 10, that is negatively correlated with prospect 8, while if prospect 8
is found oil we move to prospect 11, that is positively correlated with it, and so on for the third best
choice. In general we observe a much less flexibility in strategies characterized by many clusters,
17
and therefore a resulting smaller final value. It is interesting to notice that the strategy with 12
clusters, played with the approach SCU, coincides in fact with the naive strategy, and the first 3
selected nodes are 8, 11 and 10, no matter their outcome.
The most important finding is that, when strategies are actually played on real samples drawn
from this network, the final results, shown in the last line of Table 7, support and confirm the
theoretical results of Table 9. In this case the results are computed by averaging and discounting
the final revenues on 100 samples drawn from the network and played with each of the considered
strategies. These results are bounded by the estimated Lower and Upper Bounds computed in
Table 9. The value for the 2-clusters approach, that looks higher than the exact DP strategy with
a single cluster, is simply explained by MC variability. The 2-clusters approach reveals a very high
efficiency, with a computational time that is several order of magnitudes less than the single cluster
approach.
# Clusters
Cluster size
Independent value (LB1)
Sequential value (LB2)
Exact DP value
Clairvoyant value (UB)
DP Time per cluster (sec)
RHLA Dpt
1
12
7168
6521
6114
2
6
6660
6953
8263
0.962
6534
3
4
6177
6980
8289
0.073
6880
4
3
6115
6974
8295
0.020
7024
6
2
6043
6892
8295
0.006
7146
12
1
5056
6179
8295
0.001
7158
Table 9: Lower and upper bounds with clusters of different size, for the case study presented in
Section 4.2
As we have remarked in Section 3.2, one of the main problems is that in many cases we dont
know how far these approximations are from the real value, especially when it is not possible to
compute the exact DP strategy because of the size of the problem. For this reason we have studied
the possibility of bounding the value of the strategy, for different clusters configurations.
Let us begin from the example that we have used before, shown on the left in Figure 5.
As we can see, in this case there is nothing that suggests a good hint for placing the clusters,
therefore we just follow a naive order when defining different clusters. The results show that without
a natural clustering, it is difficult to assign correctly the nodes to each cluster. In a way, choosing
{1, 2, 3, 4, 5, 6} and {7, 8, 9, 10, 11, 12} as the members of the first partition might be suboptimal.
Further, it might be even lessoptimal
to build the sequential bound in a random order. In this
12
small case, though, there are
/2 = 462 possible ordering of the first two clusters, and we
6
have potentially to test all of the them before deciding which is the optimal partition. We report
all the results in Table 9 and in Figure 6 on the left.
The results show that the sequential value (LB2) remains quite close to the optimal value,
even for relatively small clusters. The independent value (LB1) increases with increasing clusters
dimension, but it is never as good as the sequential bound. Correspondingly, the clairvoyant bounds
(UB) slightly decreases as the clusters become larger. In general, an increase in the cluster size does
not seem to have a strong effect, and it is proved by the fact that the gap between the sequential
and the upper bound is 1421 for clusters of size 3, and 1321 for clusters of size 2. This is related to
what previously said about the absence of a natural clustering criterion.
We have also tried to compare these results with similar results obtained via Rolling-Horizon
Look-Ahead (RHLA) strategies, presented in Martinelli et al. (2011a). These strategies approximate
18
the DP value after a certain number n of steps with heuristics approximations, resulting in what
we have defined as Depth n (Dpt n) RHLA strategies. The values are shown in the bottom rows
in Table 9. The columns indicate the depth of the look-ahead procedure in this method. For
increasing depths, the value goes towards the correct value, because we are using the heuristics
after a higher number of exact steps. We observe that in this case it is sufficient to compute a Dpt
6 strategy for reaching a value that is very close to the exact one. From a computational point of
view, a Dpt 6 strategy costs as much as a clustering strategy with cluster size equal to 6, with a
substantial improvement in the quality of the approximation (the gap between a Dpt 6 strategy
and the exact one is less than 10 units).
9000
9000
8000
8000
7000
5000
4000
6000
5000
4000
3000
3000
2000
2000
1000
0
Independent EVs
Sequential EVs
Upper bound EVs
Exact value
7000
Independent EVs
Sequential EVs
Upper bound EVs
Exact value
RHLA strategies
6000
1000
6
Cluster size / Depth
10
12
6
Cluster size
10
12
Figure 6: Independent LB, sequential LB, clairvoyant UB and exact values for the two examples shown in Section 4.2; for the first example we show also the final values provided by RHLA
strategies.
We now consider the second case study shown in the right in Figure 5. The transition matrices
(CPTs) between the nodes are the same used for the previous example, but in this case the covariance structure is imposed through a Markov chain. Thus, there is a predetermined order that
makes more natural the choice of the clusters. As we might expect, the bounds are in this case much
tighter: in the previous case (common parent structure) we had conditional independence between
children only through the common parent, which was not possible to observe, and this made the
learning very hard. With the Markov chain structure, on the other hand, we have conditional independence between clusters, given a separating cluster: this suggests a clear hint about the location
of the clusters, and it furthermore simplifies the learning process. The results are shown in Figure
6, on the right. In this case the bounds shrink much faster than in the common parent network,
and a cluster size of 3 prospects is sufficient to capture virtually all the learning throughout the
network. Here, the gap between the sequential and the upper bound is 149 for clusters of size 3,
and 371 for clusters of size 2, and this shows that the clustering strategies are much more effective
in this kind of scenario.
4.3
MRF, small cases
We have adapted the same ideas originally developed for BN for a Markov Random Field structure.
Again, we can imagine that we have a lattice where each node represents a possible prospect, and
we are interested in finding the best drilling sequence and approximating the expected value of
the whole field. Since solving DP on a medium/large lattice (more than 10 nodes) is not feasible
exactly, we can split the lattice in a number of sub lattices, and we can solve the problems in each
of these small sublattices (clusters).
We first test our method on a small 3 4 lattice with 12 nodes corresponding to 12 potential
prospects. The MRF is an Ising field with 2 colors, representing oil and dry states. Revenues and
costs are equal for every prospect and symmetrical (+3 and 3 units). The field is non-symmetrical,
i.e. marginal probabilities a priori for oil and dry states follow a parabolic trend with a maximum
19
in cell 6 (2nd row, 2nd column). As a direct consequence, the Intrinsic Revenues are oscillating
around 0, with positive values just on the left part of the lattice. Marginal prior probabilities and
Intrinsic values are shown in Figure 7. The nodes are numbered from left to right and from top to
bottom, see any of the examples shown in Figure 8.
Figure 7: Marginal prior probabilities of oil and Intrinsic Values for the case study presented in
Section 4.3
10
11
12
10
11
12
10
11
12
10
11
12
10
11
12
10
11
12
Figure 8: Six possible clusters configurations for the MRF presented in Section 4.3
20
As done in the previous study, we are here interested in the effect that different clusters size,
shape and different values of the parameter have on the best exploration strategies and bounds.
For this reason we propose in Figure 8 six possible cluster combinations for the lattice under
consideration. We first analyze the best sequences computed with SCU for a fixed value = 1 and
100 simulations: results are shown in Table 10.
We notice immediately that, given the high correlation present in the field, if the first node is
found dry there are few hopes to find other good spots: this is the reason behind the suggestion
of quitting (Q) the strategy for the 1-cluster scenario, after the first node (prospect 6) is found
dry. Whenever more than one cluster is present, the strategy moves out from the cluster where the
dry node has been found, but suggests to keep the exploration campaign alive. When oil is found
in the first place, the suggestion is to keep drilling in the same cluster where the discovery has
happened, and again the best choice is provided by the 1-cluster configuration, since the suggestion
of drilling prospect 7 is crucial for exploring the right part of the field. It is interesting to observe
that configurations 3 and 4 are more rigid, in the sense that the third best choice does not depend
on the outcome of the second, but just on the outcome of the first best prospect, namely prospect 6;
this reflects the rigid vertical and horizontal clusters present in configurations 3 and 4. As usually,
the best way to compare the strategies is to study their effects on a number of fields sampled from
the model. The results are in the last row of Table 10 and show average effective values for 100
simulations; for the sake of comparison, we have to remember that the naive value of the field (sum
of positive IVs) is 1.305. There is an evident reduction of value when the clusters number increase,
as expected. There is also a less immediate, yet interesting and comforting, increase in value for
more compact clusters: results for configuration 1 are better than results for configuration 2, and
results for configuration 6 are better than those for configuration 4. The results of the exact DP
are in this case much better than any of the clustering configurations; this is most likely due to the
decision of quitting right after the first dry node, without further losses.
Configuration
# Clusters
Cluster size
i(1)
i(2) |xi(1) = dry
i(2) |xi(1) = oil
i(3) |xi(1) = dry, xi(2) = dry
i(3) |xi(1) = dry, xi(2) = oil
i(3) |xi(1) = oil, xi(2) = dry
i(3) |xi(1) = oil, xi(2) = oil
Average Value
0
1
12
6
Q
7
Q
Q
5
2
8.02
1
2
6
6
7
2
Q
8
10
5
6.98
2
2
6
6
10
5
Q
9
2
2
6.89
3
4
3
6
5
10
2
2
7
7
5.10
4
3
4
6
10
5
2
2
7
7
5.95
5
4
3
6
5
10
7
7
5
9
5.01
6
3
4
6
5
2
11
11
5
7
6.30
Table 10: Results of the sequential exploration program for the simple MRF example described in
Section 4.3, for strategies with different cluster size and shape. i(1) , i(2) and i(3) are respectively
the first, the second and the third best node selected. Q means quit (the strategy).
5 REAL EXAMPLES
21
Real examples
5.1
Large BN, sequential strategies
The original motivation for this work comes from a large BN describing a geological feature feature:
the migration paths of the Hydrocarbons (HC) expelled by the source rock in a field located in
the North Sea. The network and its parameters have been originally presented in Martinelli et al.
(2011b), and extensively discussed for similar purposes of optimal exploration in Martinelli et al.
(2011a) and Brown and Smith (2012). The idea of dividing the network in sub clusters and
studying the corresponding bounds has been presented in Brown and Smith (2012): here, the
authors propose, among the others, two possible ways to divide in clusters the original network.
K2
K2
1
2
K3
12
P1
13
P7
P6
13
P3
K4
24
P7
P6
24
P11
25
23
P11
14
25
20
P13
P9
23
P5
15
16
20
P9
22
P5
21
15
16
22
21
8
P6
19
9
11
P12
P6
P10
14
P13
P12
K3
10
12
P2
P3
K4
K3
P1
P2
18
P4
11
4
K3
10
17
5
19
P10
9
18
P4
17
4
Figure 9: Bayesian Network describing the migration paths of the HC expelled from the source
rock. The letter K marks the kitchens, i.e. places where the formation of HC has started, the
letter P marks the prospects, large areas of possible accumulations, while the numbers mark the
segments, corresponding to potential drilling locations.
The first one, with clusters of small dimension, is shown in Figure 9 on the left, while the
second one, with clusters of bigger dimension, is presented in Figure 9 on the right. Brown and
Smith (2012) describe the effect of different clusters size on the bounds, and show that the gap is
sufficiently small even for small clusters. This is not surprising, given that the learning is limited to
few parts of the network and the overall learning is relatively poor; similar comments can be found
also in Martinelli et al. (2011a). Now we are interested in studying how the strategies described in
described in Section 3.1, for both SCU and MCU, perform on this large network. The results are
reported in Table 11.
The first comment is that the expected revenues increase with increasing cluster size, as expected, and increase when we update every cluster with the new information and not just the
cluster where we have collected the piece of information. The difference between the two methods
SCU and MCU is stronger when the clusters are small and many, while is less relevant when the
clusters and big and few. This should not surprise since in the first case with SCU we may disregard an impact on a large part of the network, while in the second case, since the cluster size is
considerable, we are loosing just a peripheral information when using SCU in stead of MCU.
The second consideration is that the average number of nodes drilled increases when going from
small to big clusters, but the number of nodes drilled and found dry (that is a good measure of
our accuracy) decreases, meaning that we are more and more accurate. What is more surprising is
5 REAL EXAMPLES
Estimated value
Average # nodes drilled
Average # nodes dry
Time per sample
22
Small clusters
SCU MCU
22637 23117
17.07 16.75
3.23
3.05
15sec 25sec
Big clusters
SCU
MCU
23981 24001
18.02
17.70
3.10
3.00
30min 50min
Table 11: Sequential clustering strategies applied to the case presented in Section 5.1. Small clusters
refer to the partition of Figure 9 on the left, while big clusters refer to the partition on the right
that when we move from SCU to MCU both the average number of nodes and the number of dry
nodes decrease, but with an increase in the revenues: this means that we are avoiding to drill just
the dry nodes, while we are keeping the good nodes. In this whole analysis we can not disregard
the fact that applying MCU with big clusters is computationally much more expensive than any
other option, and that this is often a constraint that the decision maker has to take into account
when planning an optimal decision.
When analyzing the sequences (Table 12) we find out that the difference between the SCU and
MCU approaches for the small-clusters partition does not appear in the first (obvious) nor in the
second choice, but just at the third choice. The equal second choice is due to the fact that cluster 6,
that includes prospect 18, the best prospect, remains the one with highest GI even after removing
prospect 18, if the update is positive (oil or gas). Therefore it is selected as second best choice by
both approaches. For what concerns the third choice, we can, on the other side, see the difference
between the two approaches: in the MCU approach, after leaving cluster 6 with a good outcome
(at least one prospect oil or gas), we move to a neighboring cluster and we drill prospect 9. In the
SCU approach, where the cluster containing prospect 9 has not received the positive information,
we pick the second cluster with the a priori highest GI, moving far away towards prospect 8. In
the big-clusters partition the same hold, but being now prospect 9 and 18 in the same cluster, we
can no longer notice differences between the two approaches, at least in the first 3 choices.
i(1)
i(2) |xi(1) =dry
i(2) |xi(1) =oil or gas
i(3) |xi(1) =dry, xi(2) =dry
i(3) |xi(1) =dry, xi(2) =oil or gas
i(3) |xi(1) =oil or gas , xi(2) =dry
i(3) |xi(1) =oil or gas, xi(2) =oil or gas
Small clusters
SCU MCU
18
18
8
8
19
19
10
24
10
24
8
9
8
9
Big clusters
SCU MCU
18
18
8
8
19
19
24
24
24
24
9
9
9
9
Table 12: Results of the sequential exploration program for the large BN case study shown in
Section 5.1, for single cluster update and multiple clusters update strategies with different cluster
size. i(1) , i(2) and i(3) are respectively the first, the second and the third best node selected.
5 REAL EXAMPLES
5.2
23
MRF, big case
In this application we use sequential design on a MRF. The case study is from an oil reservoir in
the North Sea. Bhattacharjya et al. (2010) use this example to evaluate static acquisition strategies
for imperfect data. Here, we consider the sequential drilling problem over the dependent reservoir
units. We use a lattice representation of the field with 10 4 cells, i.e. 40 nodes. The model is
a categorical first-order MRF as in equation (2). The MRF model has 3-colors, where the three
distinctions of interest represent respectively oil saturated sand (xi = 1), brine saturated sand
(xi = 2) and shale (xi = 3). The external field parameter i (xi ) is set from geological information
and from existing seismic data, see Bhattacharjya et al. (2010).
As was done in Bhattacharjya et al. (2010), we assign a fixed cost of 2 Million USD for drilling
a dry well (state 2 or 3), while we have a potential revenue of 5 million USD when finding an oil
saturated sand (state 1). Before drilling we have the situation represented in Figure 10; here we
can see in the top left the marginal probability for the state oil in a part of the initial field.
Figure 10: Initial conditions of the MRF described in Section 5.2. Top left: marginal probability
for state oil in the 10 4 possible prospects. Top right: amplitude seismic data. Bottom left: prior
geological knowledge. Bottom right: Probability of oil saturated sand with interaction parameter
= 0.8.
Even if we consider a small subset of the initial dataset (a 4 10 square, with 40 potential
prospects), the combinatorial complexity prevents us from running a full search. In Martinelli et al.
(2011a) we have considered solutions based on an approximation with myopic/naive heuristics to
the original DP procedure. Here we use an approach based on clustering in order to show how it
is possible to design an optimal strategy and to bound the value of the field. We compare three
possible clustering strategies, the first based on 20 very small 2-cells clusters the second based on
10 small clusters of size 2 2, and the other based on larger 2 4 clusters. The main problem is
that since the field is not homogeneous, when we compute the joint cluster probabilities p(xC(i) )
we have still to condition on the entire field, and therefore the computational time required for
computing p(xC(i) ) is proportional to the size of the clusters sample space. We show the clusters
in Figure 10.
Because of the complexity and the size of the field, it is possible to compute with accuracy
just the independent lower bound for all the different configurations. Better and more complete
results could be obtained by approximating the forward-backward algorithm used for computing
5 REAL EXAMPLES
24
p(x), using the arguments presented in Tjelmeland and Austad (2012) We are also interested here
of comparing these results with results obtained with different approximations, namely the Rolling
Horizon Look-Ahead strategies of different depth presented in Martinelli et al. (2011a) and already
used for comparison in Section 4.2. Results are shown in Table 13.
Cluster size
Independent LB
Sequential LB
RHLA LB
Clairvoyant UB
2-cells
8.04
9.12
28.23
4-cells
10.44
13.71
17.00
8-cells
12.17
-
Naive
4.21
-
Dpt 1
8.36
-
Dpt 2
10.74
-
Table 13: Lower and upper bounds with clusters of different size and RHLA depth 1 and depth 2
final values, for the case study presented in Section 5.2. Parameters: = 1, = 0.99.
We immediately notice how in this case a clustering strategy with large clusters produce better
results than the the RHLA strategies until Dpt 2 (Dpt 3 is too expensive to compute). Even simple
2-cells clusters give a much better result than the classical naive approach (sum of positive intrinsic
values), and the result is further improved when using 4-cells and 8-cells clusters. In this case the
computation of the sequential and clairvoyant bounds has been possible just the smaller clusters
and for a number of samples much smaller than the previously considered one. It is worth to
notice that the gap between sequential LB and Clairvoyant LB is already quite narrow with 4-cells
clusters, and that the improvement is consistent.
i(1)
i(2) |xi(1) =brine or shale
i(2) |xi(1) =oil
i(3) |xi(1) =brine or shale, xi(2) =brine or shale
i(3) |xi(1) =brine or shale, xi(2) =oil
i(3) |xi(1) =oil, xi(2) =brine or shale
i(3) |xi(1) =oil, xi(2) =oil
2-cells clusters
14
24
13
10
23
24
24
4-cells clusters
14
10
13
24
20
10
4
Dpt 1
19
14
14
40
4
18
18
Dpt 2
14
19
19
40
18
4
18
Table 14: Results of the sequential exploration program for the large MRF case study shown in
Section 5.2, for single cluster update strategies with different cluster size. i(1) , i(2) and i(3) are
respectively the first, the second and the third best node selected.
When we move to sequences (Table 14), we notice that the cluster size and shape have a greater
influence than in the BN case of section 5.1, since the best nodes are now more spread out in
different clusters. The first best pick is a typical myopic first best pick and corresponds to prospect
14. If this is oil, we remain in the same cluster (we are following a SCU strategy), and we go
for prospect 13. If it is dry, the algorithm suggests to move to prospect 10 in the 4-cells cluster
configuration and to prospect 24 in the 2-cells cluster configuration. This happens because the GI
of cluster 4 in the 4-nodes clustering is influenced by the presence of two almost-sure dry nodes
at the bottom, while cluster 7 in the 2-cells clustering has a good GI, since it is made just by two
nodes whose presence of oil is quite likely. In general, the 4-nodes clustering strategy shows a better
ability to test new areas and to come back to the more certain places in case of dry discoveries. Iif
prospect 10 is oil we remain close and drill prospect 20, while if it is dry we move back to prospect
6 CONCLUSIONS
25
24. The 2-cells clustering shows, on the other side, some apparently less rational behavior like the
suggestion of drilling prospect 24 no matter the outcome of prospect 13, due to the small size of
its clusters and to the absence of updating given by the SCU strategy. RHLA Dpt 1 and Dpt 2
strategies do not have the constraint of the clusters and therefore their behavior is more flexible:
the first two nodes belong to different zones of the field, no matter the outcome of the first choice.
At the third step, if both prospect 14 and 19 are found dry we move to prospect 40, exploring a
third new area.
Conclusions
What we have observed so far is that when there is a clear structure in the graphical model (parts
of the network that are almost uncorrelated with other branches) it is extremely convenient to
proceed with clustering algorithms such those proposed here. When on the other side, there is
not a clear structure, even big clusters do not solve the problem of crossed-learning, and RHLA
strategies perform better in terms of expected future revenues. It is worth noticing, though, that
the main drawback of this method lies in the assumption that = 1, thus removing that sequential
effect that is so important in planning strategies. For this reason we believe that this method can
help in approximating the continuation value, but is not really a viable alternative to RHLA if we
are interested in sequential strategies.
The results show that without a natural clustering, it is difficult to assign correctly the nodes
to each cluster. In a way, choosing {1, 2, 3, 4, 5, 6} and {7, 8, 9, 10, 11, 12} as the members of the
first partition might be suboptimal. Further, it might be even less
to build the sequential
optimal

12
/2 = 462 possible ordering
bound in a random order. In this small case, though, there are
6
of the first two clusters, and we have potentially to test all of the them before deciding which is
the optimal partition.
The MRF results show that the both the clusters size and shape are important when we are
trying to build a sequential strategy, and prove that the learning process can be captured quite
well even when we split the main problem in small sub-problems. In all the examples that we
have tested, the proposed strategies and the proposed bounds perform much better than classical
myopic/naive strategies.
In this paper we looked at the dynamic decision problem. Many design problems are different
in the sense that a fixed number of nodes can be selected at once, without being able to modify
the node selection after observing the first node(s). This is another discrete optimization problem,
but some of the presented cluster ideas might be re-used here. We have not considered budgetary
constraints in the current paper. The design can be over as many nodes as is profitable in terms of
the utility. It would be interesting to study constraints in the sense that only N ? < N nodes can
be selected. Another related topic we did not study is the situation with imperfect information,
i.e. when the decision nodes are only observed indirectly. This amounts to another layer in the
graphical models considered here. The design sequences may show more flexibility when both
imperfect and perfect data collection is possible. There are several interesting problems at the
interface of statistical modeling and inference and operations research / decison making. New
insights in such problems will be useful for policy making.
7 ACKNOWLEDGMENTS
26
Acknowledgments
We thank the Statistics for Innovation (SFI2 ) research center in Oslo, that partially financed GMs
scholarship through the FindOil project. We acknowledge Arnoldo Frigessi and Ragnar Hauge
(Norwegian Computing Center) and David Brown and Jim Smith (Duke University) for very interesting discussions on this topic.
References
Aalders, I., Hough, R. L. and Towers, W. (2011). Risk of erosion in peat soils an investigation
using Bayesian belief networks. Soil Use and Management 27, 538549.
Abdul-Razaq, T. S. and Potts, C. N. (1988). Dynamic Programming State-Space Relaxation for
Single-Machine Scheduling. The Journal of the Operational Research Society 29, 141152.
Benkerhouf, L., Glazebrook, K. and Owen, R. (1992). Gittins indexes and oil exploration. Journal
of the Royal Statistical Society. Series B 54, 229241.
Besag, J. (1974). Spatial interaction and the statistical analysis of lattice systems. Journal of the
Royal Statistical Society, Series B 36, 192236.
Bhattacharjya, D., Eidsvik, J. and Mukerji, T. (2010). The Value of Information in Spatial Decision
Making. Mathematical Geosciences 42, 141163.
Bickel, J. and Smith, J. (2006). Optimal Sequential Exploration: A Binary Learning Model.
Decision Analysis 3, 1632.
Brown, D. and Smith, J. (2012). Optimal Sequential Exploration: Bandits, Clairvoyants, and
Wildcats. submitted .
Chen, Y. R. and Katehakis, M. N. (1986). Linear Programming for Finite State Multi-Armed
Bandit Problems. Mathematics of Operations Research 11.
Claxton, K. and Thompson, K. (2001). A dynamic programming approach to the efficient design
of clinical trials. Journal of Health Economics 20, 797 822.
Cowell, R., Dawid, P., Lauritzen, S. and Spiegelhalter, D. (2007). Probabilistic Networks and
Expert Systems. Springer series in Information Science and Statistics.
Gittins, J. (1979). Bandit processes and dynamic allocation indices. Journal of the Royal Statistical
Society, Series B 41, 148177.
Glazebrook, K. and Boys, R. (1995). A class of Bayesian models for optimal exploration. Journal
of the Royal Statistical Society. Series B 57, 705720.
Krause, A. and Guestrin, C. (2009). Optimal Value of Information in Graphical Models. Journal
of Artificial Intelligence Reseach 35, 557591.
Lauritzen, S. L. and Spiegelhalter, D. J. (1988). Local Computations with Probabilities on Graphical Structures and Their Application to Expert Systems. Journal of the Royal Statistical Society,
Series B 50, 157224.
REFERENCES
27
Marcot, B., Holthausen, R., Raphael, M., Rowland, M. and Wisdom, M. (2001). Using Bayesian
belief networks to evaluate fish and wildlife population viability under land management alternatives from an environmental impact statement. Forest Ecology and Management 153, 2942.
Martinelli, G., Eidsvik, J. and Hauge, R. (2011a). Dynamic Decision Making for Graphical Models
Applied to Oil Exploration. Mathematics Department, NTNU, Technical Report in Statistics
12.
Martinelli, G., Eidsvik, J., Hauge, R. and Drange-Forland, M. (2011b). Bayesian Networks for
Prospect Analysis in the North Sea. AAPG Bulletin 95, 14231442.
Martinelli, G., Eidsvik, J., Hauge, R. and Hokstad, K. (2012). Strategies for petroleum exploration
based on Bayesian Networks: a case study. SPE Paper 159722, submitted .
Powell, W. (2007). Approximate dynamic prgramming: solving the curses of dimensionality. Wiley.
Puterman, M. (2005). Markov Decision Processes: Discrete Stochastic Dynamic Programming.
Wileys Series in Probability and Statistics.
Reeves, R. and Pettitt, A. (2004). Efficient recursions for general factorisable models. Biometrika
91, 751757.
Tjelmeland, H. and Austad, H. (2012). Exact and approximate recursive calculations for binary
Markov random fields defined on graphs. Journal of Computational and Graphical Statistics
doi: 10.1080/10618600.2012.632236.
Wang, Q. R. and Suen, C. Y. (1984). Analysis and Design of a Decision Tree Based on Entropy
Reduction and Its Application to Large Character Set Recognition. Pattern Analysis and Machine
Intelligence, IEEE Transactions on PAMI-6, 406 417.
Weber, J., Sun, W. and Le, N. (2000). Designing and integrating composite networks for monitoring
multivariate Gaussian pollution fields. JRSS Series C 49, 6379.
Whittle, P. (1980). Multi-armed bandits and the Gittins index. Journal of the Royal Statistical
Society. Series B 42, 143149.

Petroleum Exploration With Bayesian Networks

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Petroleum Exploration With Bayesian Networks

Загружено:

Авторское право:

Доступные форматы

Petroleum exploration with Bayesian Networks: from

prospect risk assessment to optimal exploration.

This thesis is submitted in partial fulfilment of the requirements for the

I want here to express all my gratitude to Valeria, who has accepted

The geology of a petroleum system . . . . . . . . . . . . . . .

Modeling dependencies in a petroleum system . . . . . . . . .

Inference engines for BN . . . . . . . . . . . . . . . . . . . . .

Variable elimination algorithm . . . . . . . . . . . . .

Junction tree algorithm . . . . . . . . . . . . . . . . .

Specific network structures for HC problems . . . . . . . . . .

4 Static and Sequential decision problems

Static decision problem . . . . . . . . . . . . . . . . . . . . .

Dynamic decision problem . . . . . . . . . . . . . . . . . . . .

5 Outline of the thesis

The primary goal of the thesis is to investigate the possibility of using

The geology of a petroleum system

Figure 1: Petroleum system elements chart: we recognize the deposition

Figure 3: Elements of the petroleum system, recognizable in different layers

Modeling dependencies in a petroleum system

We are particularly interested in this thesis in modeling the dependency

that is a common well-known problem in the BN community. In recent years

7 . The acyclicity condition is ensured by the absence of cycles, i.e.

the set of vertices such that 7 is called descendants of or de().

A BN is completed by its quantitative part: the most important in-

gredient is a collection of random variables (Xv ), v V , and a probability

distribution P on Xv which factorizes according to G. For details about

the exact probabilistic definition of factorization we suggest Cowell et al.

2011b) and Martinelli et al. (2011a) to 6 in Martinelli et al. (2012a). The

Conditional Probability Tables (CPT) of size k #[pa(v)]1 , where #[pa(v)] is

Inference engines for BN

agation, belief updating. Most queries involve observations or evidence;

for a given target or interest variable: P (xi |e) =

ing a symptom knowing the presence of a given disease? Or, what is

Diagnostic or abductive reasoning: What is the probability of disease

probability that a prospect where we have found the presence HC has

Variable elimination algorithm

Given a BN with N variables X1 , . . . , XN and their discrete probability

conceptually relatively easy, but usually computationally complex.

engines for BN considering a simple case study with N = 5 nodes presented

in Figure 4. For this case study therefore f1 = p(x1 ), f2 = p(x2 |x1 ), f3 =

Figure 4: Simple BN used for inference purposes in Section 3.1

x1 ,...,xi1 ,xi+1 ,...,xn

As we can see, it presents the same problems of the computation of the

x1 ,x3 ,x4 ,x5

x1 ,x3 ,x4 ,x5

p(x1 )p(x2 |x1 )p(x3 |x2 , x4 )p(x4 )p(x5 |x1 )

p(x5 |x1 ) p(x1 )p(x2 |x1 )

fi (xi ) in which it is included, then:

fi (xi ) if xi is consistent with e

We can summarize this approach in the Variable Elimination (VE) Algorithm:

Let F be the set of probability functions in F that contains Xk

5. Combine in a single function f all the functions in F. Normalize f to

In order to calculate p(x2 ) with the VE algorithm, we proceed in this

5. Calculate p(x2 ) normalizing g 0 (x2 ) g(x2 )

j has already received information (messages) from all its neighbours

request: Ask-information(j, k): for all k ne(j), k 6= i.

Finally, we can calculate p(xi |e) locally in each node i as a normalization of

Figure 5: Propagation Algorithm (one-way) applied to the example of Figure

In general, computing the a-posteriori probability for all the variables in

Figure 6: Propagation Algorithm (two-ways) applied to the example of Figure 4.

Junction tree algorithm

Figure 7: Example used in Section 3.1.2. Note that the BN is no longer a

1. Moralize the DAG.