Вы находитесь на странице: 1из 12

Using Collective Decision System Support to

Manage Error in Wireless Sensor Fusion

Arnold B. Urken
Professor of Political Science
Stevens Institute of Technology
Hoboken, NJ 07030
aurken@stevens.edu

Presented at Fusion 05, the International Conference on Information Fusion, Philadelphia, PA., July, 2005.

1
Using Collective Decision System Support to
Manage Error in Wireless Sensor Fusion*

Arnold B. Urken
Professor of Political Science
Stevens Institute of Technology
Hoboken, NJ 07030
aurken@stevens.edu

Abstract – When sensor fusion uses voting methods to network communications and faulty decision making.
produce collective decisions on the basis of incomplete This paper outlines a new approach to wireless sensor
and imperfect information that would be produced if fusion that uses voting systems to manage these errors.
voting information were perfect and complete, the The paper is organized to explain how voting system
collective outcomes will be error-resilient. These can be designed to provide error-resilient sensor fusion.
outcomes will not be changed by breakdowns in wireless Section 2 provides a framework that explains the
network communications or decision making errors. motivation for developing a new approach and
Error-resilient collective outcome (ERCO) analysis summarizes the state of the art in building error
makes it possible to predict how long to wait or how management into voting processes. Section 3 presents
many votes to reach an optimal collective decision. the concept of an error-resilient collective outcome
ERCO analysis also provides a new framework for (ERCO) and explains how voting systems can be
gaining strategic and tactical advantages from network- designed to measure ERCO efficiency for complex
centric information sharing. This framework raises new decision tasks in risky network environments. In these
theoretical and empirical research opportunities for systems, voting methods are used to answer different and
integrating voting theory and fusion research. complementary questions about the same data. Section 4
applies this theoretical approach to a complex decision
Keywords: wireless sensor networks, distributed
detection, error, decision fusion, voting systems, error- task in which voter ratings are processed through
resilient. plurality, approval, and Copeland scoring methods in a
Monte-Carlo simulation to compare their ERCO
Patent pending. Portions of this work were supported by
contract DAAE30-00-D-1011 to the Stevens Wireless efficiency. And Section 5 discusses the simulation
Network Security Center, 2004. Approved for General results and outlines key questions for future research.
Public Release.
2 Sensor Design and Deployment
1 Introduction New sensor designs and deployment plans are driving
forces in the evolution of sensor fusion techniques. For
Current developments in the design and deployment of example, sensors that use light-scattering technology to
sensors are challenging existing methodologies for identify and detect more than one agent have led to
collecting data and producing useful information in proposals [1] for deploying large sensor arrays of
wireless networks. In commercial and security multipurpose sensors in cities to provide protection
applications of sensor technology, producing precise and against NBC (Nuclear, Biological, and Chemical)
accurate intelligence is being constrained by new attacks. These deployments could provide early
standards for reliability, cost, processing speed and warnings that enable targeted populations to take evasive
energy conservation. Although voting methods have action and permit first responders to mitigate damage.
been used to address problems of sensor communications Innovative use of materials is extending such
in networks, sensor fusion techniques have not been capabilities by creating smaller, mobile, and inexpensive
developed to overcome errors caused by breakdowns in sensor systems that can increase the scope and accuracy

2
of multifaceted data that can be collected to produce objectives [11] have drawn from voting theory as well as
knowledge [2]. By understanding the underlying theoretical insights about processing data in computer
complex patterns in such artificial environments, controls networks [4, 5]. Each of these analytical perspectives
can be designed to generate precise and accurate sensor addresses the problem of error in different ways, though
information. neither perspective incorporates the concept of error-
However innovations in the ability of sensors to resilient fusion into the voting process itself.
produce complex knowledge have not taken full account The computer science and computer engineering
of the problems of producing information in wireless literatures have used voting methods in sensor fusion
networks. Sensor capabilities are limited by decision- because processing of votes does not require much
making and transmission errors. Physical interactions bandwidth or computational overhead. With notable
with sensed environments can degrade sensor reliability exceptions [6], applications of voting systems are
and speed in detecting phenomena. Even if sensor confined to the use of simple methods of weighting votes
performance is not degraded by environmental with majority rule to control sensor fusion processes. In
conditions, technology costs and energy constraints may these studies, there is usually a focus on how to weight
limit the feasibility of deploying enough sensors to the votes.
monitor a situation. Voting systems contain subsystems that enable
Moreover, when sensors are not attacked by physical or individual voters to communicate information about
cyber attacks, the wireless networks that are needed for preferences and judgments to form a collective outcome.
transmitting data and producing knowledge pose risks. Each voting system contains subsystems based on rules
Data collection can be thwarted by malicious actions that for the endowment of votes that can be used to express
divert messages to the wrong destination or overwhelm individual information, rules for the allocation of votes,
the processing speed and energy constraints provided by and rules for aggregating votes to create a collective
network architecture. Although malicious attackers can outcome. For instance, “plurality voting,” commonly
use commercial jamming devices to thwart wireless used in elections in many Western democratic polities,
communications, the same effect can be caused includes an endowment of one vote, an allocation
inadvertently by environmental distortions from constraint that restricts assigning the vote to a single
background radiation from buildings. choice—normally without splitting or saving the vote—
For these reasons, decision fusion should not be and a plurality aggregation rule that recognizes the
considered an afterthought in the development and choice with the most votes as the winner.
deployment of new techniques for sensor knowledge. Computer scientists counterbalance the application of
Error should be integrated into the design of wireless voting methods with techniques for managing problems
sensor architecture. Wireless systems based on such caused by breakdowns in network communication that
designs will facilitate the development and deployment prevent the production of collective outcomes. For
of innovative sensors in two ways. First, they can remove example, statistical techniques are used to remove clutter,
obstacles that limit deployment of emerging sensor cleanse data, and weight votes. These techniques depend
techniques for producing more complex intelligence. on the assumption that data collection satisfies
And second, wireless sensor systems that are resilient to quantitative and qualitative requirements necessary for
error will enable designers of sensors to increase the the application of statistical methodology. These
complexity of inputs that can enhance the scope and requirements entail the creation of networks that are
accuracy of knowledge. computationally intensive, energy inefficient, and
dependent on sequential sharing of information [7],
2.1 Voting, Error, and Decision Fusion In the theoretical voting literature, where elections are
the focus of analysis, the most common assumption is
Sensor fusion models have been developed to plan that votes are collected successfully to produce a
decision tasks and the collection of sensor data, address collective outcome that reveals the group preference.
the ontological basis of fusion processes [10], use Problems in voting theory, social choice, and collective
Bayesian techniques to manage the integration of sensor decision making analyses focus on how to process the
data [13], and design local sensor decision thresholds to voting data once it received so that collective outcomes
maximize detection performance [12]. Fusion models with particular attributes can be created. And although
that have used voting systems to achieve similar analytic

3
voting theorists often disagree about the properties of outstanding 4 votes cannot change the collective
these attributes, the question of what happens if votes are outcome. The score in favor of A may increase or stay
missing does not normally arise in arguments for or the same, but the outcome cannot be changed by lack of
against the use of a voting method [9]. information caused by network communications and/or
In pre and post-election surveys sampling and data sensor errors.
cleansing techniques are used to deal with missing data In our hypothetical convoy assessment example with
by making a priori assumptions that permit the creation only two choices, six votes satisfied the aggregation
of stratified samples. But voting analysts focus on requirement to make choice A an ERCO. However
problems of preference aggregation such as paradoxical ERCO-based distributed inference provides a generic
and manipulated collective outcomes. Outside of the form of automated decision support that enables
preference aggregation mainstream, voting theorists who commanders to manage risk and uncertainty when the
address the problem of voter error by weighting collective outcome is not as clear-cut as it is in our
individual votes based on the cognitive ability or hypothetical example. For instance, suppose that the
competence of each decision maker. Neither of these commander has received 5 votes in favor of A and 2
analytical traditions of voting analysis considers what to votes in favor of B. In this case, the commander must
do if and when error caused by network communications either wait to receive more voting information or make
breakdown occurs. If all of the votes cannot be an inference that might be very risky and
collected, group preferences cannot be inferred and counterproductive. In such complex situations, it is not
individual voter preferences cannot be weighted to clear if the outstanding information has been delayed by
optimize group performance. network traffic or if breakdowns in network
Neither the voting literature nor the computer science communications or sensor failures have caused the
literature builds error-management into the vote- problem. Under such conditions, without collective
collection process itself [2]. Integrating error decision system support, commanders may unwittingly
management into voting process design makes it possible avoid risky choices that are reasonable and make risky
leverage communication in the network applications choices that are unreasonable.
layer to improve wireless sensor fusion. By interpreting For more complex, non-binary decision tasks, ERCOs
network communications from the viewpoint of the can be defined for collective decisions with two or more
recipient of votes, incomplete and imperfect voting data choices for a fixed number of human or machine sensor
can be transformed into instantaneous and accurate voters and an aggregation rule that determines that a
information. Understanding the complex patterns decisive collective outcome has been produced. At each
underlying such voting transactions can be used manage stage of data collection, the number of outstanding votes
sensor decisions in risky network environments. in the network can be represented in terms of the
percentage of the total number of voters or time required
3. Error-Resilient Voting Analysis to collect a particular segment of the outstanding votes.
If at any stage of the process of collecting votes, the
This section provides a formal definition of an error-
number of collected votes satisfies the aggregation rule
resilient collective outcome (ERCO) and explains how
to analyze voting systems to compute the probability of and the collective outcome cannot be changed by receipt
producing ERCOs. of any combination of outstanding votes, then the
collective outcome is an error-resilient collective
3.1 What is an ERCO? outcome (ERCO).
ERCOs can be produced for choices on single or
An ERCO, error-resilient collective outcome, is a multiple dimensions for collective decisions that take
voting outcome based on incomplete and imperfect place in client-server or peer-to-peer computer
information that would be produced if information about networking environments [8]. However this paper
the voting situation were complete and perfect. Humans describes ERCO production for a complex decision task
make intuitive use of ERCOs in simple situations. For along a single dimension: the number of vehicles in a
example, suppose a central commander relies on ten convoy.
sensors to report whether an object is A or B and bases Regardless of the cause(s), the costs of waiting can be
an inference on the majority outcome. If the commander significant. Lives and property will be lost and
has already received 6 votes in favor of A, then the

4
opportunities for redeploying resources to counterattack
or take evasive action will be missed. And emergency
responders will not receive early warnings to prepare to
care for victims.

3.2 Designing ERCO-Efficient Processes

The probability of producing an ERCO depends on the


voting system used to process voting data and produce
collective outcomes.

3.21 How Voting Systems Work

Ratings, inputs for a voting system, can be based on


ordinal or cardinal scales. These inputs are processed
according to rules for communicating the rating
information, converting this information into votes and
aggregating the results into a collective outcome. Rating
communication depends on vote endowment and
allocation rules that govern the expression of rating
information.
The vote endowment fixes the number of votes that can
be used to express ratings while the vote allocation rule
sets constraints on the allocation of the endowment. The
aggregation rule determines how many votes are
required to form a winning coalition.

5
When the original inputs are processed in AV and
The following chart illustrates the definition of these voters cast one approval vote for each choice that equals
rules for three systems. or exceeds their average rating, the vote allocations are
Table 1—Subsystems of Three Voting Systems

The OPOV system, a fully-specified description of


“plurality” voting, reveals which choice is most
frequently top-ranked in voter ratings. Approval voting
(AV) shows which choices are approved (or disapproved)
by a plurality or majority of voters. And Copeland
voting reveals the relative collective intensity of
preference that voters express in their ratings.
Although voting theorists debate which voting system
is best, each system answers different questions about the
collective outcomes produced by the same voting or
rating inputs. All voting systems may have paradoxical
attributes and do not necessarily generate consistent
collective outcomes.
Consider the processing of the following voter cardinal
ratings for three choices, A, B, and C.
Table 2—Hypothetical Voter Rating Scenario

When these inputs are processed in a OPOV system with


plurality rule, the allocations are

Table 3—Conversion of Ratings into Single Votes

and the collective outcome is a three-way tie, a


phenomenon associated with the “paradox” between
individual and collective transitivity.

6
determine if an attack is reasonable. The sensors report
Table 4—Conversion of Ratings into Approval Votes the correct number of vehicles by rating an overlapping
set of choices (from 0 to 4 vehicles) on a 0-10 scale, as
shown below.

and B is the plurality winner (based on the definition of


the aggregation rule determined by the number of voters
who expressed approval (3 out of 3), which would not
satisfy the requirement of a majority of the total number
of allocated approval votes).
Under Copeland voting with plurality rule, the ratings
are first processed with Condorcet scoring, which
computes the number of times that each choice is ranked
higher than every other choice in voter preference
ratings. These Condorcet scores are
Table 5—Conversion of Ratings
into Condorcet Scores

Copeland scores are then computed by subtracting the


Condorcet scores to produce

Table 6—Derivation of Copeland Scores

This illustration shows that voting systems can produce


inconsistent collective outcomes, but also generate
collective outcomes with different scores with consistent
relationships. For example, B wins under AV and
Copeland voting.

ERCO Production

The following example shows how voting systems


produce ERCOs under OPOV system; the results for AV
and Copeland voting follow the same logic and are
presented below.
In this example, ten sensors (including acoustic (AC)
and infrared (IR) sensors) provide feedback to a
commander to collectively identify the number of
vehicles in a convoy so that the commander can

7
Table 7—Sensor Ratings for
Convoy Assessment Task

When these ratings are converted into OPOV allocations,


as shown below in Table 8:

Table 8—OPOV Allocations based on Table 7

there is no majority winner (since “4 vehicles” receives


only 4 out of 10 votes), leaving the commander, C2,
without advice about whether to attack. And if IR1’s
ratings are not received, the commander would be faced
with a tied collective outcome.
In this type of decision scenario, if all of the sensors
were equally competent, ERCO analysis would focus on
the probability of satisfying the aggregation rule at any
point during the voting process. However sensors have
diverse competencies in detecting objects depending on
the manufacturer’s specifications and sensor limitations
caused by different operating conditions. So we will
assume that the ERCO objective is to produce a
collective outcome that optimizes the group probability
of making a correct collective choice with complex
preferences and competencies. In this scenario, sensor
votes can be weighted using the Shapley-Grofman
theorem [10], which assigns weights to votes based on
sensor competence or reliability using the following
formula:

8
ln(p/1-p) (10), Table 11—Example of an ERCO

where

p = probability of a correct choice (11)

and

(1-p) = the probability of an incorrect choice (12)

If p=.2 for sensors AC1-AC3, .5 for sensors AC4-


AC6, and .8 for sensors IR1-IR4 have a .8 competence,
the following chart shows the Shapley-Grofman (SG)
[10] weights that would be used to adjust the value of the
In this scenario, if the votes of AC2 and AC6 are not
votes in Table 8:
received, “4 vehicles” would be an ERCO.
Table 9—Shapley-Grofman Weights
4. Monte Carlo Analysis
To investigate the probability of producing ERCOs under
OPOV, AV, and Copeland voting systems, Monte Carlo
experiments were conducted in Mat Lab. Random
variables include voter ratings (homogeneous or
And the sensor votes in Table 8 would be transformed as heterogeneous), decision competencies, and the time
shown below: required for each set of voting information that
successfully makes it from a sensor to the commander to
Table 10—Sensor OPOV Allocations form a collective outcome. Shapley-Grofman weights
Based on Tables 8 & 9 are used to adjust individual votes and time is
represented by a Rayleigh distribution with a mean of 5
seconds.
The following results are based on 20,000 runs for 100
sensors in a bimodal culture. In this culture, 75% of the
sensors have homogeneous ratings and a high (.9)
competence, and 25% of the sensors have heterogeneous
ratings with a competence of .48. This scenario is typical
of situations in which sensor disagreement can lead
decision makers to take unreasonable risks. In such
cases, human consumers of sensor fusion may be forced
to rely on experience or intuition to resolve sensor
disagreement. When only 1 or 2, sensors disagree,
with the “4 vehicles” choice identified as the winner educated guesses may be reasonable, but, as in our
under plurality or majority rule. scenario, when 25 out of 100 sensors disagree, additional
decision support is needed to augment human
capabilities.
In each simulation run, a group of voters are randomly
selected from a randomly chosen set of 100 voters that
are drawn from a population with the bimodal cultural
preference, competence, and vote transmission time

9
attributes associated with our scenario. Then voters are much less ERCO efficient than OPOV even when the
randomly selected to simulate the process of votes proportion of outstanding voters approaches zero.
arriving at C2 (Command & Control) from distributed
sensors. After a vote is received, the simulation finds the Figure 1—Homogeneous Results
collective outcome under OPOV, AV, and Copeland based on Votes Collected
voting methods and counts the number of times that a
given collective outcome would be produced if all of the
votes were collected. These counts are correlated with
the proportion of outstanding votes and the cumulative
vote transmission time associated with each ERCO.
Then the counts of successes and failures in ERCO
production are used to compute the probability of
producing an ERCO. This probability can be used to In this bimodal culture, Figure 2 shows that making the
compare the ERCO efficiency of voting systems under preferences of 75% of the sensors heterogeneous makes
different scenarios. Copeland voting, not OPOV, most ERCO efficient.
Ties can occur under all voting systems, though the However Copeland voting’s ERCO efficiency increases
probability of generating a tie is greater under some more slowly and achieves a lower maximum than the
voting systems. For example, when preferences are OPOV system does (in Figure 1) when preferences are
heterogeneous, the probability of a tie is greater under homogeneous. In Figure 2, with 50% of the voters
AV than it is under OPOV or Copeland voting systems. outstanding, Copeland voting is 80% ERCO efficient and
The simulation results reported here are based on the only reaches a maximum of .9 ERCO efficiency.
assumption that ties are randomly broken. This
assumption increases the performance of the AV system Figure 2—Heterogeneous Results
more than it improves the ERCO-efficiency of OPOV based on Votes Collected
and Copeland systems. Ties could also be resolved
optimally by, for example, by selecting the Copeland
winner in a tied set if one existed.
Other control variables such as false positive and false
negatives are kept low and constant in the results
presented in Figures 1-4 (below). All of these scenarios
are based on the same initial conditions for a simulation
run, but Figures 1 and 3 include homogeneous (similar)
ratings or preferences, while Figures 2 and 4 show what Under these conditions, the OPOV and AV systems are
happens when ratings or preferences are heterogeneous much less ERCO efficient and display a relatively flat,
(diverse). overlapping pattern of ERCO production once 10% of
Although all of the simulation results show that ERCO the outstanding votes have been collected.
efficiency increases monotonically as a function of the
proportion of outstanding voters or cumulative time, the Figures 3 and 4 present similar contrasts when
relative efficiency of voting systems varies. preferences in this bimodal culture become more
For example, Figure 1 shows that the OPOV system is heterogeneous.
most ERCO-efficient when 75% of the voters have
homogeneous preferences. As the proportion of Figure 3—Homogenous Results based on Time
outstanding voters declines, the probability of producing
an ERCO increase rapidly so that ERCO efficiency
exceeds .95 even when only half of the votes have been
received. Under the same conditions, the AV and
Copeland display an overlapping, less efficient ERCO
production pattern. Both of these voting systems are

10
In Figure 3, when preferences are homogeneous, In addition to revealing information about how long to
wait and how much information to collect before
OPOV’s ERCO efficiency closely approaches a reaching a distributed inference, ERCO analysis draws
maximum efficiency of more than .95 when only 250 out our attention to tradeoffs between voting systems and
of 500 seconds have elapsed. In contrast, AV and waiting and information collection.
Copeland voting display an overlapping pattern of ERCO Since voting systems answer different questions about
efficiency that increases slowly and produces a the same data set, choosing a voting system depends on
what C2 wants to learn from the fusion process. In our
maximum ERCO efficiency that is approximately 30%
scenario, for instance, if C2’s goal is only to learn which
less ERCO efficient than the OPOV maximum. of the five choices in the convoy assessment situation the
most frequently top-rated choice is in the shortest
Figure 4—Heterogeneous Results based on Time possible time, then, for example, the ERCO results in
Figures 3 can be used with the OPOV results to reach an
ERCO inference that meets the time constraint. For
instance, C2 might set a time to decide of 200 seconds
into the decision making process to achieve a high (.95)
level of confidence.
However, if C2 wants to know more about the
collective assessment of the number of convoy vehicles,
Figure 3 might be used to derive additional insights. For
instance, if C2 wants to know how much more each of
the five choices was preferred to every other choice, then
In Figure 4, preferences are heterogeneous and the Figure 3 would be used to take account of the Copeland
Copeland method is most ERCO efficient. However the results to verify the results derived from the OPOV
rate of change in ERCO efficiency is slower and the pattern. In our scenario, the Copeland system is
maximum ERCO efficiency is lower for Copeland voting approximately 30% less ERCO-efficient than the OPOV
system, does not become significantly more ERCO-
than they are for OPOV when preferences are
efficient as more time passes or more votes are collected,
homogeneous. AV and OPOV display the same and displays volatility.. For these reasons, C2 might
overlapping, less-efficient, and flat ERCO production discount the value of gaining a second opinion based on
pattern. Copeland voting and consider the AV results. So C2
might wait another 50 or 60 seconds to check the results
for AV that produces higher ERCO efficiency than can be
4. Discussion derived from Copeland voting.

ERCO efficiency can augment decision support for


sensors by making it possible to predict how much 5.2 Further Research
(voting) information to collect or how long to wait to
infer that the collective inference at any point in a The scenario investigated in this paper illustrates the
decision making process is actionable. potential usefulness of ERCO analysis in providing
At the beginning of a sensor collective decision, C2 can collective decision support when the network
determine how long to wait or how many votes to collect
environment includes imperfections in sensor decision
before taking action. ERCO efficiency analysis gives C2
an information advantage that can be used to take making and network communication channels. The
evasive action, launch a neutralizing counterattack, or results of this scenario will be compared to those ERCO
dispatch first responders to mitigate the effects of an efficiencies produced when a) sensors are assumed to be
attack. perfect and communications channels are imperfect and
During an attack, C2 can obtain on-demand b) sensors are imperfect and communications channels
intelligence about the implications of waiting longer to
make a decision or collecting more information before are perfect.
making an inference. ERCO-based analysis may reveal ERCO production is also being studied for peer-to-peer
if to wait for more votes to be received or time to pass or scenarios and for situations in which decision tasks are
how much longer to wait or how many more votes must multi-dimensional. In convoy assessment task, for
be collected to reach a distributed inference about the example, additional dimensionality would be added by
number of vehicles in the convoy.
asking about other attributes of the convoy (e.g., vehicle
shape, color, etc.) in addition to number.
.1 Tradeoffs among Voting Systems

11
As probability results are developed for a variety of International Conference on Telecommunications
ERCO scenarios, analytic models will be developed to Systems, Monterey, October, 2003.
describe the effects of increasing or decreasing the
[9] Lloyd Shapley and Bernard. Grofman, Optimizing
number of sensors and introducing periodic signal Group Judgmental Accuracy in the Presence of
interruptions into the fusion process. In addition, Interdependencies, Public Choice, 1984.
situations in which the number of sensors is not known
will be investigated. [10] Erik P. Blasch and Susan Plano. Ontological Issues
Concurrently, empirical tests will be conducted in in Higher Levels of Information Fusion: User
Refinement of the Fusion Process. Fusion 03, July,
wireless sensor test beds.
2003.

* I would like to thank Russ Ovans and anonymous [11] David L. Hall and Sonya A.H. McMullen.
reviewers for their comments on an earlier draft of this Mathematical Techniques in Multisensor Data Fusion,
paper. Artech, 2004.

[12] Ruixin Niu, Pramrod K. Varshney, Michael Moore,


References and Dale Klamer. Decision Fusion in a Wireless Sensor
Network with a Large Number of Sensors, Fusion 04,
July, 2004.
[1] Philip J. Wyatt, Early Warning and Remediation:
Minimizing the Threat of Bioterrorism, Journal of
[13] L. Ronnie M. Johansson and Robert Suzić.
Homeland Security, April, 2002.
Bridging the Gap between Information Need and
Information Acquisition, Fusion 04, July, 2004.
[2] Omid S. Jahromi, Bruce A. Francis, and Raymond
H. Kwong, Relative information of multi-rate
sensors, Information Fusion, Vol. 5, pp. 119-129
(2004)

[3] Michele Zorzi and Ramesh R. Rao, Energy-


Constrained Error Control for Wireless Channels,
IEEE Personal Communications, Vol. 46, No. 3,
pp. 279-289, March, 1997

[4] Behrooz Parhami, A Taxonomy of Voting Schemes


For Data Fusion and Dependable Computation,
Reliability Engineering and System Safety, vol. 52,
139-151 (1996).

[5] Carl G. Looney, Exploring fusion architecture for a


common operational picture, Information Fusion,
Vol 2, (2001) pp. 251-260.

[6] Ofer Melnik, Yehuda Vardi, and Cun-Hui Zhang,


Mixed Group Ranks: Preference and Confidence in
Classifier Combination, IEEE Transactions on
Pattern Analysis and Machine Intelligence, Vol. 28,
No.8, pp. 973-981, Aug. 2004.

[7] Mark A. Paskin and Carlos E. Guestrin, A Robust


Architecture for Distributed Inference in Sensor
Networks, Proceedings of the Fourth International
Conference on Information Processing in Sensor
Networks, 2005.

[8] Arnold B. Urken, Time, Error, and Collective


Decision Support Systems, Proceedings of the

12

Вам также может понравиться