0 оценок0% нашли этот документ полезным (0 голосов)

2 просмотров59 страницAug 17, 2013

10.1.1.151.4931.pdf

© Attribution Non-Commercial (BY-NC)

PDF, TXT или читайте онлайн в Scribd

Attribution Non-Commercial (BY-NC)

0 оценок0% нашли этот документ полезным (0 голосов)

2 просмотров59 страниц10.1.1.151.4931.pdf

Attribution Non-Commercial (BY-NC)

Вы находитесь на странице: 1из 59

manuscript XXX

Revenue Management with Limited Demand

Information

Yingjie Lan, Huina Gao, Michael Ball*, Itir Karaesmen

Robert H. Smith School of Business, University of Maryland, College Park, MD 20742

*Corresponding author

{ylan, hgao, mball, ikaraes} @ rhsmith.umd.edu

In this paper, we consider the classical multi-fare, single resource (leg) problem in revenue management

for the case where demand information is limited. Our approach employs a competitive analysis, which

guarantees a certain performance level under all possible demand scenarios. The only information required

about the demand for each fare class is lower and upper bounds. We consider both competitive ratio and

absolute regret performance criteria. For both performance criteria, we derive the best possible static policies,

whose booking limits remain constant throughout the booking horizon. The optimal policies have the form

of nested booking limits. Dynamic policies, whose booking limits may be adjusted at any time based on the

history of bookings, are also obtained. We provide extensive computational experiments and compare our

methods to existing ones. The results of the experiments demonstrate the eectiveness of these new robust

methods.

Key words : revenue management, robust optimization, competitive analysis

History : submitted July 28, 2006; revised March 2007; revised July 2007; revised December 2007

1. Introduction

The majority of successful revenue management (RM) implementations rely heavily on the use

of demand information. Experts advise against rushing to optimization in RM without rst fully

implementing accurate demand forecasting systems (Lahoti, 2002). Likewise, research models

require a reasonably accurate characterization of demand to optimize revenues. This requirement

can be met in situations where demand patterns are stable and historical demand information is

available. On the other hand, in the case of new products or situations where, for a variety of rea-

sons, demand patterns might deviate substantially from past history, characterization of demand

is dicult, if not impossible.

Lennon (2004) mentions lack of data and naive forecasts based on inadequate data as two

important factors limiting the realistic application of RM within new industries. Even in the

1

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

2 to appear in Management Science; manuscript no. XXX

airline and hospitality industries where RM has been eectively used for decades, these two factors

commonly cause problems for new products (e.g., new routes own by an airline, new properties

added to a hotel chain). Especially in the hotel industry where franchising is very common, the

diculty with new products is more pronounced: A major hotel company reports signing 200-300

new franchise contracts in the rst half of 2005 (hotel-online.com, 2005). While new properties

added to the system may be similar to existing ones in terms of market and demand characteristics,

franchising agreements or anti-trust laws (because some properties compete in the same market)

may prohibit sharing of demand information. Hence, the hotel chain may not eectively use its

well-established RM system at least in the rst year for new developments. On the other hand,

RM of even old products in airlines and hotels is not easy; obtaining high quality aggregate or

disaggregate forecasts from censored data remains a challenge (see, for example, Weatherford and

Polt, 2002).

Despite the need for robust methods and approaches that do not rely heavily on demand infor-

mation, research in that direction has been scarce. Traditional research models and analyses rely

on several restrictive and possibly unrealistic assumptions about demand such as independence

and stationarity (see Talluri and van Ryzin, 2004a). In this paper, we consider the classical sin-

gle resource (leg) problem in RM and provide alternative forms of controlling the bookings when

demand information is limited. We do not make any assumptions about the demand or the arrival

process beyond upper and lower bounds, and do not need a risk neutrality assumption. Our

approach relies on competitive analysis of online algorithms and we call the policies derived based

on competitive analysis robust since they guarantee a certain performance level under all possible

demand scenarios. We focus on cases where the only information available is lower/upper bounds

on demand. In many practical cases, lower and upper bounds on demand are easier to obtain than

a probability distribution of demand or a characterization of an arrival process over time.

Our contributions can be summarized as follows. We propose new static and dynamic booking

control policies for the single-leg, multiple-fare class problem when only upper/lower bounds on

demand are available. We formulate two dierent models that dier in their objective functions.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 3

The rst model maximizes the competitive ratio (a measure related to relative regret) and the

second minimizes the maximum absolute regret. We show that these two models can be analyzed

in a unied manner and we obtain closed-form optimal solutions for both. The resulting optimal

booking control policies preserve the nesting structure that is widely used in practice. We also show

that the optimal booking limits obtained under the absolute regret criterion are no greater than

the ones obtained under the competitive ratio. Based on computational experiments, we show that

the average revenues obtained by our policies are comparable to other well-known procedures even

though our policies use less information. Further, our policies are robust and not prone to errors

in modelling demand. Finally, our work builds on prior research by Ball and Queyranne (2006)

who used competitive analysis of online algorithms to analyze the single-leg problem and assumed

no demand information is available. Our policies, which assume demand bounds, generalize theirs,

and we also provide alternative proofs for some of their results.

The rest of the paper is organized as follows: We provide a literature review in Section 2. Section

3 introduces our notation and describes our approach. In Section 4, we derive static nested booking

control policies when the objective is to maximize the competitive ratio. Section 5 extends this

analysis to the problem with an absolute regret criterion and Section 6 shows how the policies may

be dynamically updated. Section 7 presents an extensive set of computational results and Section

8 provides conclusions and suggestions for further research.

2. Literature Review

The booking (or capacity) control problem for a single resource has long been studied in the RM

literature. Littlewood (1972) was the rst to analyze the stochastic, single-leg problem with two

fare-classes and assumed product is sold in a low-before-high (LBH) manner; that is, demand in

lower fare class precedes demand in higher fare classes. Belobaba (1987,1989) discussed heuristic

extensions of Littlewoods rule to multiple fare classes, again assuming (i) LBH arrivals and (ii)

the probability distribution of demand in each fare class is known. Curry (1990), Wollmer (1992)

and Brumelle and McGill (1993) make similar assumptions on demand. Robinson (1995) relaxed

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

4 to appear in Management Science; manuscript no. XXX

the LBH assumption in his analysis but assumed requests for dierent fare classes arrive at non-

overlapping intervals. Lee and Hersh (1993) introduced a dynamic programming formulation for

the multiple fare class problem in which they relax the LBH assumption and assume the demand in

each fare-class is characterized by a stochastic process. Lautenbacher and Stidham (1999) address

the static (i.e., arrivals are LBH) and dynamic problems (i.e., arrival sequence is not ordered by

fare-class) for the multiple-fare case by analyzing the underlying discrete time Markov Decision

Process (MDP). More recently, Talluri and van Ryzin (2004b) analyzed the multiple-fare class

problem based on consumer choice. While this approach is more sophisticated and realistic than

many, it requires information not only on the arrival process but also on the choice behavior. There

are several other papers on the single-leg problem with various assumptions on demand or arrivals

(see Brumelle and Walczak, 2003, McGill and van Ryzin, 1999, and Talluri and van Ryzin, 2004a).

Until very recently, the adaptive method proposed by van Ryzin and McGill (2000) for the

single-leg, multi-fare problem was the only model in RM that relied on no demand information.

This approach updates booking limits iteratively based on past observations. van Ryzin and McGill

(2000) prove that the nested booking limits computed using their adaptive procedure converge

to the optimal. Huh and Rusmevichientong (2006) and Kunnumkal and Topaloglu (2007) suggest

dierent iterative approaches to compute the nested booking limits. All of these methods require

LBH arrivals, learn or adapt from ight to ight but not during the booking control horizon of

a particular ight, and may suer from poor performance during the transient period until the

booking limits converge to the optimal.

The objective in all the above papers is to maximize revenues given the risk-neutrality of the

decision-maker. Such assumptions have been questioned in the context of pricing in the RM lit-

erature (e.g., Feng and Xiao, 1999). In addition, there is a growing body of literature on pricing

(e.g., Farias and van Roy, 2006, Rusmevichientong et al., 2006, Lim and Shanthikumar, 2006) that

challenge the traditional assumptions on demand models, availability of demand information, and

also the risk-neutrality assumption.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 5

In this new stream of research, the work of Birbil et al. (2006), Eren and Maglaras (2006),

Perakis and Roels (2006), and Ball and Queyranne (2006) are the only ones - to the best of our

knowledge - that focus on capacity allocation (or booking control) in the single-leg RM problem.

Birbil et al. (2006) assume there are inaccuracies associated with discrete probability distributions

that characterize the demand in each fare class. They propose a robust framework where the

probability value associated with the total demand of each fare-class lies in a polyhedral set (as

opposed to the realized demand lying within a given interval as in our model and that of Perakis

and Roels, 2006). Eren and Maglaras (2006) propose using the maximum entropy approach to

update the booking limits while obtaining demand information. This approach makes use of past

observations and accordingly adds fractile constraints to a convex optimization problem that is

used to obtain an estimate of a discrete probability distribution. That estimate is then used to

determine the booking limits, and this forecasting/optimization cycle is repeated throughout the

booking horizon. Perakis and Roels (2006) derive a limited demand information model similar to

ours and provide a general approach to both the single-leg and the network RM problems. They

consider two objectives: maximizing the minimum revenues and minimizing the maximum regret.

They also restrict their analysis to particular policies (e.g., nested policies for single-leg, partitioned

allocations for network). In contrast, we provide a unied framework for both absolute regret and

competitive ratio criteria in the single-leg problem. Our analysis inspired by competitive analysis

of online algorithms is signicantly dierent than theirs. We provide closed-form optimal solutions

and prove optimality of the nested booking limits in our problem.

Ball and Queyranne (2006), which we refer to as BQ, were the rst to adapt the notion of com-

petitive analysis of online algorithms in the RM context. They derive static, nested booking limits

for the single leg problem and also consider bid-price controls where a booking request does not

belong to a particular fare class but comes with a proposed fare. Their policies require no infor-

mation on demand and come with performance guarantees (achieving maximum competitive ratio

as dened in the next section). They also show for two fare-classes how the booking limits can be

updated during the booking period to dynamically improve worst case performance. Our approach

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

6 to appear in Management Science; manuscript no. XXX

is more general than theirs in that we assume the availability of limited demand information and

show that the analysis extends to problems with both relative and absolute regret criteria. Our

optimal policies reduce to theirs in special cases, hence we obtain alternate proofs for some of their

results as well. Finally, our policies are practically more eective compared to BQ as we show in

our computational experiments.

3. Problem Denition

We consider the single-leg RM problem from the perspective of the competitive analysis of online

algorithms (see the survey in Albers, 2003). Using this approach, the performance of an online

algorithm is evaluated relative to the performance of an oine algorithm that considers the entire

input sequence simultaneously. An oine optimal solution is a solution obtained by an oine

algorithm (with hindsight) that optimizes the objective function of interest. In competitive analysis,

it is common to use the competitive ratio (CR) as a measure of an algorithms eectiveness. There

are two other performance metrics of interest: absolute regret, which is the dierence between the

objective function values of the oine and online algorithms, and relative regret, which is the ratio

of the absolute regret to the objective function value of the oine algorithm. Specically, we are

interested in having a guarantee on these measures under all possible demand scenarios (or input

sequences as we call them in our analysis).

CR is dened as the minimum of the ratio of revenues obtained by the online algorithm to the

oine revenues. If we let

and, for any I

, let R(I; ) be the objective value achieved by the online algorithm for input I

and let R

(I) be the objective value achieved by an optimal oine algorithm. Then, CR is dened

as:

CR of = inf

I

R(I; )

R

(I)

.

Likewise, the maximum absolute regret (MAR) of the online algorithm is dened as

MAR of = sup

I

{R

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 7

By denition, CR is related to maximum relative regret (MRR): CR = 1 MRR. Clearly, the

algorithm that maximizes the CR also minimizes the MRR. We use CR and MAR in our analysis.

Note that these performance measures as dened above apply to deterministic algorithms, i.e.,

algorithms that apply the same decision rule and yield the same output for a given input sequence

as opposed to randomized algorithms that make some choices based on the draw of a random

number. While not as practical as deterministic ones, randomized algorithms are of special interest

from a technical standpoint; discussion on this topic is deferred until Section 4.4. For now, we are

interested in determining the best deterministic algorithm that maximizes the CR (or minimizes

the MAR). Based on the denition of robustness in Kouvelis and Yu (1997), maximizing the CR

(minimizing the MAR) leads to decisions that are relative-robust (deviation-robust).

Throughout this paper, let n denote the total capacity of the resource (seats, rooms, etc.) avail-

able and let m 2 denote the number of fare-classes. Let f

i

denote the fare for class i, where

f

1

> f

2

> > f

m

0. There are no restrictions on the demand or arrival process except that

requests arrive one-by-one and cancellations are not allowed. In our analysis, we assume each

request demands only one unit but this can be generalized as we discuss below. We are interested in

determining nested booking limits n b

1

b

2

b

m

0 that maximize the CR or minimize the

MAR. Note that nesting by revenue order remains the core of booking control in many traditional

RM implementations. Although Talluri and van Ryzin (2004b) show that these types of nested

policies may not be optimal in general, we show later in the text that they are optimal relative to

the CR and MAR criteria.

Nested policies can be implemented in two dierent ways: standard- and theft-nesting (see Talluri

and van Ryzin, 2004a). We only consider the standard implementation of the nested policies in

this paper: The booking limit b

i

denes the maximum number of booking requests to be accepted

in classes i to m. At any point in time, a request in class i is accepted as long as the number of

on-hand reservations in classes i to m is less than b

i

and remaining (available) capacity is positive.

We use an additional variable called bucket size, denoted by x

i

, in our analysis; we dene it as

x

m

= b

m

and x

i

= b

i

b

i+1

for i = 1, ..., m 1. The notation b and x refer to the vectors of the

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

8 to appear in Management Science; manuscript no. XXX

corresponding variables. Note that either of these vectors is sucient to characterize a nested

booking control policy. The protection levels that are commonly used in single-leg RM are easily

derived from the booking limits (see Talluri and van Ryzin, 2004a). The number of seats protected

for classes j =1, .., k from classes k +1 to m is equal to nb

k+1

, k =1, ..., m1. While our analysis

is carried out with b and x, we interpret some of our results with respect to protection levels, as

well.

We note that, in the remainder of the text, we use competitive analysis to evaluate the perfor-

mance of a policy, with policy referring to the decision rule and parameter values, as opposed to an

algorithm, which is a specic iterative procedure. In simple terms, competitive analysis is based on

worst-case analysis: one can think of an adversary in charge of generating booking requests. The

adversary is aware of the policy (nested booking limits in our case) and chooses an input sequence

(the number of requests and the arrival sequence) to create the worst possible policy performance

based on the criterion of interest, i.e. CR or MAR. Given a characterization of the input sequences

an adversary would choose, we determine the optimal parameters for the policy to maximize the

CR (alternatively, to minimize the MAR).

In its natural form, the single-leg RM problem has discrete demand. However, the analysis in

many research papers assumes continuity. In this paper, we distinguish between the continuous

problem and discrete problem. In the continuous problem, the demands, and therefore the booking

limits as well, may be any nonnegative real numbers. In this case, a request in a fare-class may be

partially accepted (split). Although the continuous problem is less realistic, its analysis is simpler.

The continuous case easily generalizes to multi-unit and multi-fare requests (as in batch arrivals) by

allowing splitting of requests. Our analysis carries through as long as each request demands a non-

negative and nite amount. Note that group bookings cannot be enforced/guaranteed because of

splitting. In the discrete problem, the demands are integral and the protection levels are restricted

to being integer. Our analysis of the discrete problem extends to multi-unit and multi-fare requests

as long as requests are split into integral quantities. In the rest of the paper, we assume, without

loss of generality, that input sequences consist of requests that demand one unit from a fare-class.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 9

4. Optimal Static Policies for the Competitive Ratio Problem

In this section, we analyze the multi-fare single-leg problem and derive optimal booking control

policies under the CR criterion assuming upper and lower bounds are available on demand in each

fare class. Consider an input sequence which consists of a nite stream of fare requests during the

booking horizon. For each fare class i, we assume the input sequences are restricted so that the

total number of units demanded in class i falls between L

i

and U

i

, where L

i

and U

i

are integers

with 0 L

i

U

i

for i = 1, ..., m. We use the notation L and U to denote the respective vectors

(L

1

, ..., L

m

) and (U

1

, ..., U

m

). Let (L, U) be the set of all input sequences where the total demand

in each fare-class falls between the upper and lower bounds, R

from sequence I, and R(I; b) be the online revenue gained by a standard nested booking limit

policy b when the policy is used to process the requests in sequence I. The problem of nding an

optimal booking control policy that maximizes the CR can be stated as follows:

z

CR

= max

b

z : z

R(I; b)

R

(I)

, I (L, U) (1)

where b is the vector of decision variables and z

CR

is the optimal CR. There are some potential

challenges in solving this maximization problem: 1) R(I; b) might not have a closed-form expression;

2) |(L, U)| grows exponentially with m, prohibiting any serious direct attempt with even small

problems. However, since there will be redundant constraints, which means the corresponding

sequences are redundant as well, a sequence reduction approach seems natural, and, as we will

show in the next two sections, this maximization problem can be rewritten in a very compact form.

The solution to the above problem will result in a static control policy.

4.1. Sequence Reduction

A sequence includes information on the amount of booking requests and their order of arrival.

Since each request in a sequence demands one unit, we can characterize the sequences based on

the fare-classes and units demanded: Let I[j] be the total number of class j requests in sequence I.

The prole of I, denoted by [I], is the m-dimensional vector, (I[j] : j =1, ..., m). Observe that the

oine optimal revenue realized in sequence I only depends on the prole, while the online revenue

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

10 to appear in Management Science; manuscript no. XXX

(of a policy) also depends on the arrival order of requests. However, given a prole, we can ignore

permutations of requests that do not yield the lowest online revenue (see Appendix A of the Online

Addendum for the proof).

Proposition 1 Relative to all input sequences with the same prole, a nested booking limit policy

b generates the least revenue when applied to the unique LBH sequence with that prole. This is

true for both the continuous and discrete problems.

From the standpoint of the maximization problem in (1), all non-LBH sequences can be discarded

based on Proposition 1. In the remaining sequences, there is a one-to-one relationship between

sequences and proles. This yields a substantial reduction in the size of the problem. Yet, the total

number of proles can be prohibitive. Therefore, we pursue further sequence reductions to dene a

tractable problem. We now introduce a categorization of input sequences, which is even a broader

concept than proles. The ultimate goal is to choose only one sequence in each category. For

mathematical completeness, we now introduce a virtual fare-class m+1 where b

m+1

=0, f

m+1

=0,

and L

m+1

, U

m+1

0 are arbitrary, non-negative numbers.

Denition 1 (Input Sequence Category) Given a nested booking limit policy b, an input sequence

category A

b

j

is the set of sequences such that an input sequence I belongs to category A

b

j

if and

only if j has the highest fare (hence the smallest index) among the classes whose booking limits are

reached after executing b on sequence I.

To better understand the notion of a category, consider the following: First of all, the virtual class

has b

m+1

=0, so at least the booking limit of class m+1 is always reached, and j is well dened.

By construction, A

b

m+1

includes all sequences for which no request is rejected when the booking

limit vector is b. Therefore,

R(I;b)

R

(I)

=1 for all I A

b

m+1

. Second, some categories may be empty. For

instance, in the trivial case of b

i

=n for all i =1, ..., m and

m

i=1

U

i

<n, no booking limit (except

for the virtual class) would be reached when policy b is applied to the sequences in (L, U). In

that case, all A

b

j

for j = 1, .., m would be empty, and the only category that would be non-empty

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 11

is A

b

m+1

. In fact, A

b

m+1

= (L, U) in this case. Finally, the categorization is complete, meaning

m+1

j=1

A

b

j

=(L, U) for any policy b.

Appending more requests in classes k j to a sequence in category j will not change the online

revenue of policy b, since all such requests would be rejected. In addition, any permutation of the

order of requests in a sequence will not change its category, which means the category is totally

determined by the prole of the sequence. Note that the notion of a category applies to both

discrete and continuous problems, and so does our next result (see Online Addendum, Appendix

B for the proof).

Denition 2 (CAtegory-dominant-STream, or CAST) Given j = 1, ..., m, a LBH sequence I is

called a category-dominant-stream (CAST) if I has the prole of I[k] =U

k

for k j and I[k] =L

k

for k <j.

There are m CASTs in total. We denote each one by CAST

j

, j = 1, .., m. By denition,

CAST

j

[k] =U

k

for k j and CAST

j

[k] =L

k

for k <j.

Proposition 2 (Dominance of CAST) Consider a nested booking limit policy b and all input

sequences in some category A

b

j

. CAST

j

A

b

j

dominates the other sequences in category A

b

j

, i.e.,

R(I;b)

R

(I)

R(CAST

j

;b)

R

(CAST

j

)

for all I A

b

j

.

For any nested policy, Proposition 2 eectively reduces the number of sequences to be considered

down to m, disregarding the virtual class m+1 (which is not critical since

R(I;b)

R

(I)

= 1 for all I

A

b

m+1

). With these reductions, problem (1) can be reformulated as

z

CR

= max

b

z : z

R(CAST

k

; b)

R

(CAST

k

)

for k =1, ..., m. (2)

We next show how the optimal solution to the reduced formulation can be obtained by dening

an appropriate linear programming model. We rst focus on the continuous problem.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

12 to appear in Management Science; manuscript no. XXX

4.2. A Linear Programming Formulation for the Continuous Problem

In determining the optimal booking control policy, we use the bucket sizes x

i

, i = 1, ..., m, as the

decision variables. We make two observations regarding the optimal policy below.

First, if

m

i=1

U

i

n, then the optimal bucket sizes are x

i

=U

i

, i and z

CR

=1. Second, if there

is a class k such that (n

k

i=1

L

i

) <0, then any reasonable policy would have x

j

=0 for all j >k

and the problem size can be reduced by ignoring classes j >k in that case. This second observation

relies on the following: Remember that the lower bounds imply that all sequences have at least

L

i

requests of class i. Consider the quantity

k

i=1

L

i

; this represents the number of seats that are

guaranteed to be sold to classes 1 to k. Consequently, any reasonable policy would protect at least

that many seats for classes 1 to k. Hence, the remaining seats, i.e., (n

k

i=1

L

i

)

+

, are the total

number of seats that would be considered for sale for classes k + 1 to m. Therefore, the m-fare

problem is only interesting when

m

i=1

U

i

>n and

m1

j=1

L

j

<n. We impose these two conditions on

the problem parameters to keep our problem statement general, the notation simpler and solutions

to the problem non-trivial.

We study the properties of nested policies under CASTs so that we can rewrite the right hand

side of (2) in a form suitable for analysis. Note that any reasonable policy would have x

1

U

1

so

as not to protect any seats for class-1 that would denitely remain unsold. Likewise, x

i

U

i

for

all i =1, .., m; otherwise there would be a class k 1 whose bucket would have a slack (i.e., the

booking limit of class k would not be reached). Naturally, we require that

m

i=1

x

i

n so as not to

exceed the capacity.

By denition, the prole of each CAST is known. We further dene (k) = min{j :

j

i=1

CAST

k

[i] n, 1 j m} which denotes the index of the lowest-fare class to be accepted

in the oine optimal solution (requests of classes with lower fares are rejected) when the input is

CAST

k

. Then, we can express R

(CAST

k

), the oine optimal revenue for CAST

k

, as

R

(CAST

k

) =

_

n

(k)1

j=1

CAST

k

[j]

_

f

(k)

+

(k)1

j=1

CAST

k

[j]f

j

.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 13

Next, we focus on the online revenue obtained from the CASTs. For any non-trivial m-fare

problem, the online revenue generated from CAST

k

can be considered in two parts: 1) from the

L

i

requests in classes i < k, with a subtotal revenue at most

k1

i=1

f

i

L

i

(because not all L

i

may

be accepted based on the bucket sizes), and 2) from the U

i

requests in classes i k, from which

revenue is

m

i=k

x

i

f

i

(note x

i

U

i

, from our discussion above). Then, R(CAST

k

; b), the online

revenue obtained by policy b and corresponding bucket sizes x, satises R(CAST

k

; b)

k1

i=1

f

i

L

i

+

m

i=k

f

i

x

i

. Therefore, we have an upper bound on the ratio of online revenues to that of oine,

and this upper bound is linear in the bucket sizes x

j

:

R(CAST

k

; b)

R

(CAST

k

)

k1

i=1

f

i

L

i

+

m

j=k

f

j

x

j

R

(CAST

k

)

.

Combining these observations, we can formulate a linear program (LP) that nds a continuous,

nested policy. We call this formulation the General Bounded Model (GBM):

GBM: z

CR

= max

x

z

s.t. R

(CAST

j

)z

j1

i=1

f

i

L

i

+

m

i=j

f

i

x

i

, j =1, ..., m (3)

m

i=1

x

i

n, (4)

0 x

j

U

j

, j =1, ..., m. (5)

Although GBM has been developed based on a necessary set of constraints and relations that

constitute an upper bound on the optimal CR, the solution to GBM provides the optimal nested

policy for the CR problem introduced in (1). We formalize this statement and discuss how a

closed-form solution to GBM can be derived in the next section.

Note that the actual CR problem expressed in (1) is not only reduced to a problem with a

small number of constraints based on CASTs, but its optimal policy in the continuous version

is obtained by solving a linear programming problem with 2m + 1 constraints. The multi-fare

continuous booking problem discussed in BQ can be represented using GBM. Theirs is a special

case: they use no information about the demand, hence L

i

=0 and U

i

= (or eectively U

i

n)

for i =1, ..., m in their problem.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

14 to appear in Management Science; manuscript no. XXX

4.3. Optimal Policy for the Continuous Problem: Solution to GBM

Our analysis of GBM relies heavily on an investigation of the intrinsic structure of the LP and the

relationship among the parameters f

i

, U

i

, L

i

for i =1, ..., m, and n; details are provided in Appendix

C in the Online Addendum. An important step in the analysis is to determine the critical fare

class u {1, ..., m} such that classes j >u would be closed (i.e., booking limit of j >u would be

zero) in the optimal solution. Once u is known, the solution to GBM can be determined by solving

a set of linear equations made up of the binding constraints in the LP. We derive a closed-form

solution to the GBM based on these observations. The proof of the next result is also available in

the Online Addendum, Appendix C.

Proposition 3 (a) The optimal solution of GBM is

z

CR

=

R

+

u

/f

u

+N

u

R

(CAST

u

)/f

u

+

u1

i=1

g

i

(6)

x

CR

j

=

_

_

_

g

j

z

CR

+L

j

j <u

(R

(CAST

u

) z

CR

R

+

u

)/f

u

j =u

0 j >u

(7)

u = max{j m: R

+

j

j1

i=1

g

i

<N

j

R

(CAST

j

)} (8)

where the index u denotes the critical fare-class such that all classes k >u are closed, and g

i

, N

j

, R

+

j

are auxiliary parameters dened as

g

m

=

R

(CAST

m

)

f

m

, g

i

=

R

(CAST

i

) R

(CAST

i+1

)

f

i

, i =1, , m1, (9)

N

j

=n

j1

i=1

L

i

, R

+

j

=

j1

i=1

f

i

L

i

j =1, .., m. (10)

(b) The nested booking limits dened by

b

CR

j

=

m

i=j

x

CR

i

for j =1, , m (11)

maximize the CR in problem (1) and the optimal CR is z

CR

= z

CR

.

Note that our analysis so far yields the optimal solution within the class of nested policies. One

question that remains to be answered is whether nested policies are the best. The next result shows

that, in fact, this is true (see the Online Addendum, Appendix D for the proof).

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 15

Theorem 1 For the continuous m-fare problem with demand bounds, the nested booking control

policy with booking limit vector b

CR

dened by (11) has a CR of z

CR

given by (6) and this is the

best possible among deterministic policies.

If all lower bounds are zero, the above result can be simplied, and the optimal booking limits

can be expressed in terms of n and g

j

for j = 1, ..., m. Although this is only a special case, we

provide a formal proof of the next result in Appendix E in the Online Addendum.

Corollary 1 For the continuous m-fare problem with all lower bounds equal to zero, the nested

booking control policy with booking limits dened by:

b

CR

i

=n

m

j=i

g

i

m

j=1

g

j

for i =2, ..., m

has a CR of n/(

m

j=1

g

j

) and this is the best possible among deterministic policies.

Note that for the special case of L

j

= 0, U

j

n for all j = 1, ..., m, we have R

(CAST

j

) =f

j

n,

and g

i

=n(1 f

i+1

/f

i

). This, in fact, denes the optimal nested policy obtained in BQ.

4.4. Randomized Policies

Our focus so far has been on deterministic policies. However, the performance can sometimes be

improved by allowing the use of randomization strategies. We dene a randomized policy by

starting with a given set of deterministic policies {

1

,

2

, ...}. The same input sequence may result

in dierent outputs when a randomized policy is used. This is because a deterministic policy is

randomly chosen from the given set, prior to observing and processing the input sequence. The

random choice prior to each execution is a random variable that is independent and identically

distributed with a known probability distribution. Let be a random variable dened over the

set {

1

,

2

, ...} given the probabilities P(=

j

). The expected objective function value of the

randomized policy is then E[R(I; )]. Consequently, the CR of a randomized policy is then

dened based on the expected performance:

CR of = inf

I

E[R(I; )]

R

(I)

.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

16 to appear in Management Science; manuscript no. XXX

We recognize that randomized policies may not be very desirable in practice (confusing the

users and possibly the buyers). However, we believe our analysis is of theoretical interest and also

provides further insight into the performance obtained by rounding the continuous booking limits.

Given the class of randomized policies for the single-leg RM problem, we can show that no policy

can achieve a higher CR than z

CR

given in (6). While this may seem to be a trivial extension of

our previous results, analysis of randomized policies generally requires an approach dierent from

the analysis of deterministic policies. In fact, we use a dierent proof technique for the next result

(see Appendix F in the Online Addendum).

Theorem 2 For the continuous m-fare problem with demand bounds, no randomized booking con-

trol policy has a CR larger than z

CR

given in (6). Therefore, the deterministic nested booking limit

policy with booking limits dened by b

CR

in (11) is the best possible among all policies.

This result implies that nesting by revenue order provides the optimal policy for the single-leg

RM problem where the objective is to maximize the CR under limited demand information.

4.5. Solution to the Discrete Problem

In most practical settings, integral quantities of a product are sold indicating that a realistic analy-

sis should consider the discrete problem rather than the continuous one. On the other hand, we can

view the continuous problem as a (possibly very close) approximation for the discrete one in that,

rounding the booking limits up or down will have a relatively small impact on overall performance.

We note that with integer upper and lower bounds rounding always produces feasible booking

limits. To develop a randomized policy for the discrete problem, we rst consider the solution x

CR

of GBM given in equation (7). If this solution is non-integral, we create deterministic policies by

rounding up or down each x

CR

i

, i = 1, ..., m. Based on the rounding scheme, we dene probabili-

ties associated with each policy. We provide the details on how the policies and probabilities are

determined along with a proof of existence of a probability distribution in the Online Addendum,

Appendix G. Based on the analysis in Appendix G, we can show that the expected performance

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 17

of the randomized policy is exactly the same as the performance of the best deterministic policy

for the continuous problem, which is our desired result:

Proposition 4 For the discrete m-fare problem with demand bounds, an appropriately dened

randomized policy achieves a CR of z

CR

, which is given in (6).

While this result shows that a (randomized) policy exists for solving the discrete problem, it (pos-

sibly more importantly) provides justication for creating discrete policies by selectively rounding

continuous policies.

5. Optimal Static Policies for the Absolute Regret Problem

The analysis in Section 4 easily extends to the problem with the absolute regret criterion where

the objective is to minimize the MAR. Let us rst express this problem in its general form:

z

AR

= min

b

z : z R

where b is the vector of decision variables and z

AR

is the optimal MAR. We can reformulate this

problem by reducing the number of input sequences using Propositions 1 and 2. We only have to

consider the CASTs as inputs. The following LP (called GBM-AR) is designed to provide a lower

bound on the minimum MAR in the continuous problem:

GBMAR: z

AR

= min z

s.t. R

(CAST

j

) z

j1

i=1

f

i

L

i

+

m

i=j

f

i

x

i

, j =1, ..., m (13)

and (4), (5).

This model is actually easier to analyze than the GBM, and a closed-form solution is given below

(proof is omitted):

z

AR

= R

(CAST

u

) R

+

u

f

u

x

AR

u

(14)

x

AR

j

=

_

_

_

g

j

+L

j

j < u

N

u

u1

i=1

g

i

j = u

0 j > u

(15)

u = max{j m:

j1

i=1

g

i

<N

j

} (16)

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

18 to appear in Management Science; manuscript no. XXX

where u is the critical fare class such that classes k > u are closed, and the auxiliary parameters

g

i

, N

j

, R

+

j

for j =1, .., m are as dened in equations (9) and (10).

Following an argument similar to that for the CR problem, we can show that the optimal solution

of GBM-AR provides the optimal MAR of the problem in (12). Perakis and Roels (2006) derive

the same closed-form solution for the MAR problem using a dierent approach. In addition to

providing the optimal solution within the class of nested policies, our analysis in the previous

section can be used to prove that the optimal nested policy obtained by GBM-AR is the best

possible under the MAR criterion (details omitted). Hence, nesting leads to both deviation-robust

and absolute-robust decisions in the single-leg RM problem with limited demand information.

Theorem 3 For the continuous m-fare problem with demand bounds, the nested booking control

policy dened by the booking limits

b

AR

j

=

m

i=j

x

AR

i

for j =1, ..., m (17)

has minimum MAR of z

AR

(where x

AR

i

and z

AR

are given in (15) and (14)). No other policy,

deterministic or randomized, has a lower MAR.

A close look at the optimal policies for the CR and MAR problems, given in equations (6)-(8)

and (14)-(16), respectively, reveals that u u. Therefore, MAR policies tend to deny bookings to a

higher number of fare classes. Furthermore, x

CR

i

x

AR

i

for i < u u and

u

i=1

x

CR

i

=

u

i=1

x

AR

i

=n

at optimality. Our next result combines these observations:

Proposition 5 In the continuous problem with demand bounds, the optimal nested booking limits

obtained by (11) and (17) satisfy: b

AR

i

b

CR

i

for i =1, ..., m.

While competitive analysis using the CR or MAR criterion provides conservative solutions to

guarantee worst-case performance, the above result shows that the MAR criterion is more aggressive

in protecting seats, i.e., fewer seats are available for lower-fare classes and more seats are protected

for higher fares.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 19

Our analysis of the randomized policies and the discrete problem as presented in Sections 4.4

and 4.5, respectively, can be extended to the MAR problem. Details are omitted.

6. Dynamic Policies

In this section, we show how dynamically adjusting a static policy can improve the CR (or MAR)

in our problem. Our analysis extends the discussion in BQ (where a sample analysis for the CR

problem in the case of m=2 is provided) to multiple fare classes with demand bounds. Using the

imaginary adversary paradigm, the CR (or MAR) of a policy eectively assumes that the adversary

always adheres to an optimal strategy (i.e., the adversary chooses one of the CASTs as an input

sequence). On the other hand, if the adversary makes a mistake, we can consider a new problem

scenario based on the remaining capacity and requests accepted so far, and create a new policy for

the remaining requests in order to guarantee better overall performance.

Consider a dynamic scenario where a dynamic policy has been executing so that a partial input

sequence has already been processed. Suppose h

i

bookings have been accepted for fare class i for

i =1, ..., m by processing the partial sequence I

0

. This accumulates revenue of

m

i=1

h

i

f

i

from the

m

i=1

h

i

sold seats. The question is whether the booking limits can be adjusted to improve the CR

(MAR) achievable under a future input sequence

I, which yields the complete input sequence of

I =I

0

I. The concatenation of I

0

and

I into the complete input sequence I produces the following

prole: L

j

I[j] =I

0

[j] +

I[j] U

j

, j =1, ..., m. Since I

0

[j] h

j

, j =1, ..., m, it follows that I[j] h

j

must hold, so that we can update lower and upper bounds for

I as follows:

L

j

=max(L

j

, h

j

) h

j

,

U

j

=U

j

h

j

, j =1, ..., m.

Let n =n

m

i=1

h

i

denote the remaining number of available seats,

b the nested booking limits

for allocating the remaining n seats, and x the corresponding bucket sizes.

Given a partial policy

b and partial input sequences I

0

,

I, the overall CR can be improved by

solving

z

CR

= max

b

z : z

m

i=1

h

i

f

i

+R(

I;

b)

R

(I

0

I)

,

I (

L,

U). (18)

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

20 to appear in Management Science; manuscript no. XXX

where z

CR

is the new guarantee on the worst-case CR performance. Similarly, the improved mini-

mum MAR of z

AR

is determined by:

z

AR

= min

b

z : R

(I

0

I) z

m

i=1

h

i

f

i

+R(

I;

b),

I (

L,

U). (19)

Applying Propositions 1 and 2 to the partial

I sequences, we only have to focus on m sequences

that dominate the others in (

L,

U). These sequences have the same structure as CASTs. Following

the corresponding steps in Sections 4.2 and 5, the optimal CR (MAR) can be computed by solving

a LP. This LP has the same structure as GBM (GMB-AR). Closed-form solutions are obtained by

replacing appropriate parameters in (6)-(8) and (14)-(16). The closed-form solutions and discussion

of special cases are provided in Appendix H in the Online Addendum.

These dynamic policies are easy to implement; specically, the h

j

should be updated each time

a request is accepted, LPs are re-solved and the booking limits adjusted accordingly. These tasks

require minimal computational resources as we have closed-form optimal solutions of the LPs. The

idea of such dynamic adjustments is reminiscent of re-optimization in traditional RM where a

static model is used in a rolling horizon fashion to mimic the performance of a dynamic program.

The major dierence in our dynamic approach is that we consider the remaining capacity and the

past performance (as represented by the revenue component

m

i=1

h

i

f

i

) from a partial input to

guarantee a better worst-case performance for the entire booking horizon.

7. Computational Results

While the policies obtained above come with performance guarantees, such worst-case guarantees

provide only a limited indication of their eectiveness in practical situations. We have designed

computational experiments (i) to quantify the performance of policies obtained using the CR and

MAR criteria, (ii) to illustrate the dierences between static and dynamic policies, (iii) to compare

our policies to other well-known procedures in single-leg RM, and (iv) to show the robustness of

our policies under various demand scenarios and parameter settings.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 21

7.1. Performance of Our Policies: CR vs. MAR, Static vs. Dynamic

In this section, our goal is to test the eectiveness of CR vs. MAR and static vs. dynamic poli-

cies. The policies that use demand bounds are denoted BSTAT-CR, BSTAT-AR, BDYN-CR and

BDYN-AR, and are obtained by solving GBM, GBM-AR, dynamically adjusted GBM and dynami-

cally adjusted GBM-AR, respectively. The policies STAT-CR, STAT-AR, DYN-CR, and DYN-AR,

respectively, are special cases of these policies where no demand information is used. We only

implement the deterministic versions of our policies, and use continuous booking limits in these

experiments. We use simulation to study the (experimental) performance of the policies. We also

use FCFS and OFFLINE policies as benchmarks; FCFS accepts all arrivals up to the capacity and

OFFLINE computes the hindsight optimal solution and revenue at the end of each simulation run.

The demand is uniformly distributed, m = 2, and n = 100 in all the examples in this section.

We vary the fare values, demand parameters, arrival regimes with respect to fare classes (LBH vs.

time-homogeneous arrivals), and the demand mix (mean demand of a fare class relative to the mean

demand of other classes) in our experiments. Requests arrive one-by-one in all the experiments.

When arrivals are LBH, the sequence of arrivals is known, and the total number of arrivals in each

class is computed by sampling from the demand distribution in each simulation run. When the

arrivals are time-homogeneous, the total number of arrivals in each class is determined by sampling

from the corresponding (aggregate) demand distribution. Then, the arrival times of requests within

each class are randomly generated from a Uniform(0,1) distribution. The corresponding arrival

times determine the sequence of arrivals in the time-homogeneous case.

Example-1a. The fares are (f

1

, f

2

) =(500, 100), and demand in each fare class is discrete uniform

distributed between L

i

=40 and U

i

=80, i =1, 2. We use 6000 simulation runs with LBH arrivals in

this example. The booking limit of class-2 for each of the policies, the theoretical CR of each policy

computed given the demand bounds, average policy revenues, relative performance (computed by

taking the ratio of policy revenues to OFFLINE in each simulation run, and averaging this ratio

across all simulation runs) of policies, and the average number of seats sold are displayed in Table

1. The static and dynamic policies are equivalent in this example because of LBH arrivals (see

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

22 to appear in Management Science; manuscript no. XXX

the discussion in Appendix H in the Online Addendum). Note that the dierences between the

booking limits of CR and MAR policies are signicant when no demand information is used. As

expected, the worst-case CR is not an indication of the average performance of the policies. In this

example, STAT-CR achieves signicantly lower average revenues compared to other policies based

on these booking limits and fares: note that STAT-CR achieves a higher load (as indicated by the

average number of seats sold) compared to the other policies, but the other policies accept more

class-1 requests leading to higher average revenues.

We analyzed the performance of the policies more closely by studying the distribution of revenues

(i.e., we computed estimates of percentiles). We split the 6000 simulation runs into 30 samples of

size 200 each, computed the percentiles in each sample, and took the averages of the percentiles of

the samples to obtain a good estimate. This information is provided in Figure 1. Notice that the

ranking of the policies with respect to the 10

th

, 50

th

and 90

th

percentiles are dierent. In fact, the

ranking of our policies in the 10

th

percentile is reversed in the 90

th

. Hence, no policy (OFFLINE

and FCFS excluded) stochastically dominates the others. Note that our use of the term stochastic

dominance is only based on numerical observations; we do not provide a theoretical foundation

here. STAT-CR has the highest 10

th

percentile value. The use of demand information degrades

the performance of CR policies exhibited in the results for the 10

th

percentile. However, demand

information provides signicant gains for the 90

th

percentile results: the likelihood that BSTAT-

CR achieves a revenue of at least $150,000 is about 10% whereas STAT-CR can only guarantee

that the revenues will be at least $124,000 10% of the time. The opposite relationship is observed

between STAT-AR and BSTAT-AR; STAT-AR has a higher (lower) 90

th

(10

th

) percentile values

compared to BSTAT-AR. Table 1 indicates that these dierences are related dierences in the

booking limits and dierences in aggressiveness/conservatism (in terms of the number of seats

protected for class-1).

This example shows the main dierence between CR and MAR policies and also the eect of

demand information. The MAR policies set a lower booking limit for class-2 and protect more seats

for class-1. Therefore, the 90

th

(10

th

) percentile values of their revenue distributions are higher

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 23

Booking Limit Theoretical Average of ratio Average

of class-2 CR Average of policy revenues number of

Policy b

2

(%) revenues to OFFLINE (%) seats sold

BSTAT-CR, BDYN-CR 31.51 89.04 129,235 95.37 89.3

BSTAT-AR, BDYN-AR 28 87.69 129,501 95.28 86.93

STAT-CR, DYN-CR 55.5 66.19 114,302 85.84 98.79

STAT-AR, DYN-AR 20 84.61 128,013 93.82 79.58

FCFS 100 42.86 101,864 76.63 98.99

OFFLINE 100 135,663 100 98.99

Table 1 The booking limit, theoretical CR, and average performance of policies in Example-1a.

94000

104000

114000

124000

134000

144000

154000

164000

10 30 50 70 90

percentile

p

o

l

i

c

y

r

e

v

e

n

u

e

OFFLINE BDYN-CR, BSTAT-CR DYN-CR, STAT-CR

BDYN-AR, BSTAT-AR DYN-AR, STAT-AR FCFS

Figure 1 Distribution of revenues for each policy in Example-1a

(lower) than the CR policies. The use of demand information increases the average revenues of

both the CR and MAR policies and decreases the dierences between the performances of CR and

MAR. We repeated Example-1a with dierent demand parameters and arrival regimes. The results

are reported in the Online Addendum (see Example-1b through Example-1f in Appendix I).

Example-2a. We next investigate the eect of the fare ratio r =f

2

/f

1

. The demand has the same

distribution as in Example-1a. We vary the fare ratio from 0.1 to 1. The arrivals are LBH. Figure 2

shows how the average of the ratios of policy revenue to OFFLINE revenue changes with r. Notice

that in this case the CR and MAR policies result in similar average performance when demand

information is used. Use of demand information not only improves the average performance but it

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

24 to appear in Management Science; manuscript no. XXX

0.85

0.875

0.9

0.925

0.95

0.975

1

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

ratio of fares (r)

a

v

g

.

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

s

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

s

DYN-CR, STAT-CR DYN-AR, STAT-AR

BDYN-AR, BSTAT-AR BDYN-CR, BSTAT-CR

FCFS

Figure 2 Performance of each policy relative to OFFLINE in Example-2a

also makes the performance of policies less sensitive to changes in the fare ratio r. MAR policies

perform better than CR policies when there is no demand information and r is low: STAT-AR is

signicantly better than STAT-CR for r = 0.1; this is expected because STAT-AR accepts more

class-1 requests than STAT-CR. We repeated this experiment with a dierent demand mix (see

Example-2b in Appendix I of the Online Addendum). In that experiment the mean demand of

class-2 is twice the mean demand of class-1: the relative performance of the policies with demand

bounds are unaected, however, STAT-CR and STAT-AR perform dierently, with the former

having higher revenues than the latter for low values of r.

Next, we look at the eect of the quality of demand information on our policies.

Example-3. The arrivals are LBH and the true demand in each fare class is uniformly distributed

between 40 and 80 as in Example-1a. In addition to the true range of the demand, we use our policies

with a narrow range of demand where the upper and lower bounds of demand are estimated to be

55 and 65, also with a wide range where the upper and lower bounds of demand are estimated to

be 20 and 100. This setup results in three versions of each of our policies, and we denote them as T

(for true bounds), N (for narrow range) and W (for wide range). The average relative performance

of the policies are reported for dierent fare ratios in Figure 3. Note that the results with true

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 25

0.92

0.93

0.94

0.95

0.96

0.97

0.98

0.99

1

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

fare ratio (r)

a

v

g

.

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

BDYN-CR-T, BSTAT-CR-T, BDYN-AR-T,BSTAT-AR-T

BDYN-AR-N, BSTAT-AR-N

BDYN-CR-N, BSTAT-CR-N

BDYN-CR-W, BSTAT-CR-W

BDYN-AR-W, BSTAT-AR-W

Figure 3 The relative performance of the policies with inaccurate demand information in Example-3

bounds are no dierent than the results in Example-2a, and the MAR and CR policies have almost

exactly the same relative performance regardless of the fare ratio. This is still true when the range is

narrow. However, the eect of wide range on MAR and CR policies diers. When the range is wide

and the fare ratio is low, MAR policies have poor performance because they expect higher class-1

demand and (aggressively) set lower booking limits. At higher fare ratios, this eect is minimal

and the performance of the two policies is similar. To more easily understand the dierences in

these policies, we display the various booking limits at three dierent fare ratios in Table 2. One

interesting observation in this example is the following: The policies do not perform perfectly when

r = 1 if a narrow range is used. This is very surprising given the fact that all requests have the

same fare, and any reasonable policy (including FCFS) would perform as well as OFFLINE when

r =1. The main reason for this is that both the MAR and CR policies allow for no more than 45

class-2 requests when the range is narrow, because the estimated lower bound on class-1 demand is

55 (see the optimal bucket sizes in equation (7) of Section 4.3). Hence they reject too many class-2

requests.

We repeated this example with a dierent demand mix (mean demand of classes 1 and 2 were

40 and 80). The main observations remained the same. Narrow bounds make the MAR and CR

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

26 to appear in Management Science; manuscript no. XXX

BSTAT-CR-T, BSTAT-CR-N, BSTAT-CR-W, BSTAT-AR-T, BSTAT-AR-N, BSTAT-AR-W,

r BDYN-CR-T BDYN-CR-N BDYN-CR-W BDYN-AR-T BDYN-AR-N BDYN-AR-W

0.2 31.5 37.19 32.78 28 37 16

0.5 42.5 40.15 50 40 40 40

0.9 56.15 44.01 72.58 56 44 72

Table 2 The booking limit of class-2 for each policy in Example-3

policies almost indistinguishable, wide bounds aect the performance of the MAR policy when the

fare ratio is low, and narrow bounds lead to clearly suboptimal choices when the fare ratio is high.

Nonetheless, our policies remain very robust: they achieve at least 92% of the OFFLINE revenues

on the average in all the experiments. Note that this experiment only studies the eect of demand

information when the demand distribution has the correct mean, but has a range that is either too

narrow or too wide. The performance of our policies can be very poor if the range of demand is

chosen arbitrarily, e.g., when there is no overlap between the estimated range and the true range.

However, such instances are not very realistic and/or are problematic for any method that relies

on demand information.

The observations from this set of experiments can be summarized as follows: (1) CR and static

policies are better in terms of the 10

th

percentile of the revenue distribution, except when class-2

demand is low: these policies are more conservative protecting too few high fare seats. (2) MAR

and dynamic policies are better in terms of the 90

th

percentile of revenues, i.e. the chance of

achieving higher revenues is higher with these policies, except when class-1 demand is low. (3)

Time-homogeneous arrivals are in general better for all the policies, but can hurt the performance

of dynamic MAR policies when class-1 demand is low. (4) The use of correct demand information

decreases the variance in revenues of MAR and CR policies, increases the average performance,

and makes the relative performance of these policies less sensitive to changes in the fare ratio. (5)

The dierence between MAR and CR policies is negligible when correct demand information is

used and class-1 demand is low.

7.2. Comparison to other well-known policies

In this section, we compare our policies to other well-known methods. We use (i) Littlewoods rule

(we call it EMSR as in Belobaba, 1989) for m = 2, (ii) EMSR-b heuristic for m > 2 (see Talluri

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 27

and van Ryzin, 2004a), (iii) the dynamic programming (DP) model of Brumelle and McGill (1993)

when m>2 and arrivals are LBH, and (iv) the Markov Decision Process (MDP) model of Lee and

Hersh (1993) for Markovian arrivals. These policies are benchmarks for our methods and represent

the ideal situations where one has complete information about arrival processes or probability

distribution of demand. From a practical point of view, these methods serve another purpose: these

are methods that are popular in practice and research. However, their performance with forecast

errors is not widely reported. We also perform sensitivity analysis on these methods and compare

them to our methods under scenarios where either the assumptions in these models are violated or

there is an error in model parameters. In this section, we only report the performance of the policies

relative to OFFLINE. Each instance in each experiment involves 300 simulation runs. The capacity

is n =100 unless noted otherwise. The demand parameters for our policies are computed by setting

the lower and upper bounds on demand to two standard deviations away from the (estimated)

mean of each fare class. This is a simple, heuristic rule that does not necessarily provide true

bounds on demand. In some cases, this choice of bounds may be a relatively poor characterization

of the range of demand, e.g., when asymmetric demand distributions are involved. As will become

evident from our results, our policies perform well, even when poor estimates of demand bounds are

used to compute the policy parameters. Our policies that use no demand information are excluded

from the discussion in this section unless noted otherwise, because the focus is on the methods

that use some form of demand information.

Example-4. In this example, the fares are (f

1

, f

2

) =(500, 100). The demand for each fare class

is independent and follows a Poisson distribution. We denote the mean demand of class i as

i

,

i =1, 2. The demand factor (ratio of total mean demand to capacity) is 1.2 while the mean demand

of the fare classes [

2

,

1

] ranges from [120,0] to [0,120]. When arrivals are time-homogeneous (see

Figure 4), MDP is the optimal policy and it performs remarkably well compared to OFFLINE. Note

that CR policies perform comparably to EMSR except in the extreme case where class-2 demand

is negligible and class-1 demand exceeds capacity (which is not realistic from a RM perspective,

and/or not interesting from a segmentation and capacity control point of view). There is less than

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

28 to appear in Management Science; manuscript no. XXX

95%

96%

97%

98%

99%

100%

[120,0] [100,20] [80,40] [60,60] [40,80] [20,100] [0,120]

mean demand [

2

,

1

]

a

v

g

.

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

BDYN-CR BSTAT-CR BDYN-AR

BSTAT-AR EMSR MDP

Figure 4 Average relative performance of policies in Example-4

1% dierence in the relative performance of any of our policies. We repeated this example with

LBH arrivals. The arrival regime does not aect the relative performance of the policies, and their

performance relative to EMSR does not change (EMSR is optimal when arrivals are LBH). The

results with LBH arrivals and the performance of policies that do not use demand information are

presented in Example-4 in Appendix I of the Online Addendum.

We utilized the same example, keeping the mean demand at [60,60] but varying the fare ratio.

The main observations remain the same: MDP (as the ideal solution) is close to OFFLINE when

arrivals are time-homogeneous. Our policies are as good as EMSR at all fare ratios regardless of

the arrival regime (except when class-2 demand is negligible). The relative performance of our

policies is no worse than 95% in any of the experiments. In these experiments, the additional

benet of using the dynamic versions of our policies is not signicant. This is because the demand

is stationary and demand mix is balanced. Next, we look at an example with more than two fare

classes.

Example-5. This is adapted from the example used in Talluri and van Ryzin (2004a), Sec-

tion 2.2.3.4. We have m= 4, fares are f = (1050, 567, 527, 350), n = 124, the demand is Normally

distributed and independent across fare classes. The arrivals are LBH and mean demand is 17.3,

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 29

DP EMSR-B BDYN-CR BSTAT-CR BDYN-AR BSTAT-AR FCFS

0.94 0.91 0.92 0.93 0.94 0.94 0.87

Table 3 Average of ratio of policy revenue to OFFLINE revenue in Example-5 when CoV=0.4.

45.1, 73.6 and 19.8 for classes 1 through 4, respectively. The demand for all fare classes has the

same coecient of variation (CoV) and we set CoV=0.4. Note that static and dynamic policies

are not equivalent in this example. The average relative performance is reported in Table 3. The

optimal policy in this case is given by DP. BSTAT-AR and BDYN-AR are almost as good as DP

in this example. All of our policies dominate EMSR-b. The performance of our policies with no

demand information is given in Example-5 in Appendix I of the Online Addendum. Appendix I

also provides the results of further experiments obtained by increasing the CoV of the distribution

(and using truncation in the simulations): the relative performance of the various policies remains

the same as CoV increases, while the deviation from the best policies from DP becomes slightly

larger.

The results of the last two experiments are very encouraging: our methods with limited demand

information are practically as good as EMSR and even better than EMSR-b which is commonly

used in airline RM practice. Note that the main motivation for our work on robust methods is lack

of data and accurate forecasts. In real life, not only will the demand forecasts be wrong, but also

some of the assumptions in models such as EMSR, DP or MDP may fail to hold. The next set of

experiments shows how these methods perform when booking limits of the policies are computed

with incorrect demand models and/or data.

Example-6. In this example, m = 2, demand is Poisson distributed, arrivals are time-

homogeneous, and parameters of all control policies are computed assuming a demand mix of

[40,40] for the booking horizon. However, there is an unexpected surge and the demand doubles in

the second half of the booking horizon, bringing the demand mix to [60,60]. The average relative

performance of the policies as a function of the fare ratio r is given in Figure 5. EMSR performs

signicantly worse when r is small because the optimal protection level chosen by Littlewoods rule

decreases with r, and EMSR is not able to benet from the unexpected surge in the demand of

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

30 to appear in Management Science; manuscript no. XXX

86%

88%

90%

92%

94%

96%

98%

100%

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

fare ratio (r=f

2

/f

1

)

a

v

e

r

a

g

e

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

BDYN-CR BSTAT-CR BDYN-AR BSTAT-AR

FCFS MDP EMSR

Figure 5 Average relative performance of policies in Example-6

class-1. There is not a signicant dierence among our policies; they all dominate EMSR and are

dominated by MDP which is able to adjust booking limits based on remaining time and capac-

ity. Note that the policies that protect the highest number of seats for class-1 (e.g., STAT-AR,

DYN-AR) and can dynamically update the booking limits (e.g., DYN-AR) perform well in this

experiment. See Example-6 in Appendix I of the Online Addendum for the performance of our

policies with no demand information.

We also looked at what happens when the parameters of the demand distribution are misspec-

ied. See Example-7 in Appendix I of the Online Addendum. In that example, we have m = 2,

the true demand distribution of class-2 is known, whereas only an estimate of the parameter of

the demand distribution of class-1 is available. All policy parameters are determined based on this

estimate. The sensitivity of our results is no more or less than that of EMSR to forecast errors in

that two-fare example. Our experiments conrm that our policies with limited demand information

perform as well as EMSR under various scenarios. We see that MDP is quite robust to the error

in parameter estimation in Example 7 where the arrivals are stationary and only the mean arrival

rate for the MDP model is misspecied. In another experiment (see Example-8 in Appendix I of

the Online Addendum), we test whether MDP is aected by stationarity vs. non-stationarity of the

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 31

arrivals. In fact, MDP performs very poorly if it assumes the arrival rate is time-invariant whereas

the actual arrival rate is time-dependent. In contrast, our policies are only sensitive to aggregate

demand information, and do not require a complete characterization of the arrival process.

8. Conclusions

In this paper, we have analyzed the traditional single-leg RM problem from the perspective of

competitive analysis of online algorithms. Our models make use of limited demand information

and we derive static and dynamic policies for both competitive ratio and absolute regret criteria.

Our optimal booking control policies have signicant practical advantages: the nesting property

of booking limits is preserved while the need for information is reduced, and the optimal policy

parameters are obtained in closed-form, hence the computational burden is minimal. Our compu-

tational experiments show that, when coupled with good estimates of the bounds on demand, our

policies have average performance that is similar to EMSR and EMSR-b, which are two widely

used methods. At the same time, our policies also perform well when subjected to errors in demand

information and/or deviations from standard assumptions used in traditional RM models.

From a research perspective, the competitive analysis of online algorithms approach is very

promising in RM. In this paper, we took a rst step in using limited demand information to increase

the practical eectiveness of online algorithms in RM. Adding more demand information to the

single-leg problem (e.g., the use of time-varying bounds) and applying the competitive analysis

ideas to the network RM problem remain as challenging future research topics. In this paper,

analysis was carried out assuming the demand information is given and static. However, it is worth

investigating how demand bounds can be estimated, what types of estimation procedures and/or

choice of demand bounds make our policies most eective, and how robust methods can be adapted

to changes in demand information.

Acknowledgments

The rst three authors acknowledge the support of the National Science Foundation (NSF) under grants:

DMI0205489 and DMI0540312. We thank the Associate Editor, two anonymous referees, and the Department

Editor, Candace Yano, for their comments that signicantly improved the exposition of the text.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

32 to appear in Management Science; manuscript no. XXX

References

Albers, S. 2003. Online algorithms: a survey. Math. Programming, Ser. B, 97, 3-26.

Ball, M. O. and M. Queyranne. 2006. Toward robust revenue management: competitive analysis

of online booking. Working paper, available at SSRN: http://ssrn.com/abstract=896547.

Belobaba, P.P. 1987. Airline yield management: an overview of seat inventory control. Trans. Sci.

21, 63-73.

Belobaba, P.P. 1989. Application of a probabilistic decision model to airline seat inventory control.

Oper. Res. 37, 183-197.

Birbil, S. I., J.B.G. Frenk, J. A.S. Gromicho and S. Zhang (2006). An integrated approach to

single-leg airline revenue management: the role of robust optimization. Working paper, Erasmus

University, Erasmus Research University Inst. of Mmgt. (ERIM), Rotterdam, The Netherlands.

Brumelle, S.L. and J.I. McGill. 1993. Airline seat allocation with multiple nested fare classes. Oper.

Res. 41, 127-137.

Brumelle, S. and D. Walczak. 2003. Dynamic airline revenue management with multiple semi-

Markov demand. Oper. Res. 51, 137-148.

Curry, R.E. 1990. Optimal airline seat allocation with multiple nested fare classes. Trans. Sci. 24,

193-204.

Eren, S. and C. Maglaras. 2006. Revenue management heuristics under limited market information:

a maximum entropy approach. presented at The 6th Annual INFORMS Revenue Management

Conference, June 5-6, Columbia University.

Farias, V. and B. Van Roy. 2006. Dynamic pricing with a prior on market response. presented at

The 6th Annual INFORMS Revenue Management Conference, June 5-6, Columbia University.

Feng, Y. and B. Xiao. 1999. Maximizing revenues of perishable assets with a risk factor. Oper. Res.

47, 337-341.

hotel-online.com. 2005. Choice hotels international signed 276 new franchise contracts in the rst

six months of 2005, up 19% over the same period in 2004. URL:

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 33

http://www.hotel-online.com/News/PR2005 3rd/July05 ChoiceMidYear.html

Huh, T. and Rusmevichientong, P. 2006. An adaptive algorithm for multiple-fare-class capacity

control problems. Presented at the 6th Annual INFORMS Revenue Management Conference,

June 5-6, Columbia University.

Kouvelis, P. and G. Yu. 1997. Robust Discrete Optimization and Its Applications. Kluwer Academic

Publishers, Nonconvex Optimization and Its Applications, V.14, Boston.

Kunnumkal, S. and H. Topaloglu. 2007. A stochastic approximation method for the revenue man-

agement problem on a single ight leg with discrete demand distributions. Working paper, School

of Operations Research and Industrial Engineering, Cornell University, Ithaca, NY.

Lautenbacher, C.J. and S. Stidham, Jr. 1999. The underlying Markov Decision Process in the

single-leg airline yield management problem, Trans. Sci. 33, 136-146.

Lahoti, A. 2002. Why CEOs should care about revenue management. OR/MS Today, February

2002, http://www.lionhrtpub.com/orms-2-02/rm.html.

Lee, T.C. and M. Hersh. 1993. A model for dynamic airline seat inventory control with multiple

seat bookings. Trans. Sci. 27, 252-265.

Lennon, J.J. 2004. Revenue management and customer forecasts: a bridge too far for the UK visitor

attractions sector? J. Revenue and Pricing Manag. 2, 338-352.

Lim, A.E.B. and J.G. Shanthikumar. 2006. Relative entropy, exponential utility, and robust

dynamic pricing. forthcoming in Operations Research.

Littlewood, K. 1972. Forecasting and control of passengers. 12th AGIFORS Symposium Proceed-

ings, 95-128.

McGill, J. and G. van Ryzin. 1999. Revenue management: research overview and prospects. Trans.

Sci. 33, 233-256.

Perakis, G. and G. Roels. 2006. Robust controls for network revenue management. Working paper,

UCLA, Anderson School of Management, Los Angeles, CA.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

34 to appear in Management Science; manuscript no. XXX

Robinson,L. W. 1995. Optimal and approximate control policies for airline booking with sequential

nonmonotonic fare classes. Oper. Res. 43, 252-263.

Rusmevichientong, P., B. Van Roy, and P. W. Glynn. 2006. A non-parametric approach to multi-

product pricing. Oper. Res. 54, 82-98.

Seiden, S. S., 2000. A guessing game and randomized online algorithms. Proc. of the 42nd ACM

Symposium on Theory of Computing (STOC 2000), 592601.

Talluri, K. and G. van Ryzin. 2004a.The Theory and Practice of Revenue Management. Kluwer

Academic Publishers, Boston.

Talluri, K. and G. van Ryzin. 2004b. Revenue Management under a general discrete choice model

of customer behavior. Manag. Sci. 50, 15-33.

van Ryzin, G. J. and J. McGill. 2000. Revenue management without forecasting or optimization:

an adaptive algorithm for determining airline seat protection levels. Manag. Sci. 46, 760-775.

Weatherford, L.R. and S. Polt. 2002. Better unconstraining of airline demand data in revenue man-

agement systems for improved forecast accuracy and greater revenues. J. Revenue and Pricing

Manag. 1, 234-254.

Wollmer, R.D. 1992. An airline seat management model for single-leg route when low fare classes

book rst. Oper. Res. 40, 26-37.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 1

Online Addendum to

Revenue Management with Limited Demand Information

Yingjie Lan Huina Gao Michael Ball* Itir Karaesmen

Robert H. Smith School of Business, University of Maryland

College Park, MD 20742 USA

{ylan, hgao, mball, ikaraes}@rhsmith.umd.edu

*Corresponding author

Appendix A: Proposition 1

Proposition 1 Relative to all input sequences with the same prole, a nested booking limit policy b gener-

ates the least revenue when applied to the unique LBH sequence with that prole. This is true for both the

continuous and discrete problems.

Proof We prove this rst for the case of a discrete (integer) booking limit policy b. The main idea behind

the proof is that, starting with any initial input sequence I that is not LBH, we can iteratively swap the

order of requests in this sequence (eventually to reach a LBH sequence) such that the iterations/swaps will

never lead to online revenue gains for the nested policy b.

Let I(k) {1, ..., m} to be the fare class of the k

th

request in sequence I. If an input sequence, I, is

not LBH, then we can nd some index k such that I(k) < I(k +1); we construct a new sequence I

such

that I

following observation regarding (standard) nested policies: If any two consecutive requests are both accepted

or rejected by a nested policy b, then a swap in their arrival order will not aect the decision on any request

in the entire sequence. For policy b, if I(k) and I(k +1) were both rejected/accepted in I, then the same

would happen in I

Since the application of a static policy does not allow re-opening of closed fare classes, it is impossible

that I(k) is rejected and I(k +1) is accepted in I under (standard) nested policy b. So the only remaining

case is when I(k) is accepted and I(k +1) is rejected in I. There are two possibilities for I

to reject I

I, and so we will have the same online revenue for both I and I

I

(k +1). We show below that policy b would generate less revenue executing sequence I

in this case.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

2 to appear in Management Science; manuscript no. XXX

We keep track of the number of accepted reservations while processing an input: We denote the length

of a sequence (i.e., the total number of requests in a sequence) by |I|. By denition, |I| =

m

i=1

I[i]. Let

B

j

(t), t =0, 1, , |I| be the total number of accepted reservations of classes j to m by policy b after the t

th

request of sequence I has been processed, and B

j

(t) be the analagous quantity for the sequence I

. We refer

to B

j

(t) and B

j

(t) as the booking record of class j under sequences I and I

a nested booking limit policy b works. If a request of fare class j is accepted, it increases the booking record

of each of the classes 1 to j. A request of class j is rejected if the booking record is already equal to the

booking limit for class j. Notice that the rst k 1 requests of sequences I and I

B

j

(k 1) =B

j

(k 1), j =1, ..., m.

The case we are investigating is the one where both I(k) and I

j

(k) =

B

j

(k 1) +1 for j I(k) and B

j

(k) =B

j

(k 1) +1 for j I

j

(k 1) =

B

j

(k 1), we have B

j

(k) = B

j

(k) for j I(k). Notice that the next request I

(k + 1) = I(k) is rejected,

which means for some i I(k), the booking limit has been reached: B

i

(k) =b

i

. As i I(k), we must have

B

i

(k) =B

i

(k) =b

i

. Thus only requests with fares strictly higher than f

i

may be accepted in the future for

both sequences, and so only the identical booking records B

j

(k) = B

j

(k) for j i matter in the future.

As I(k +1) =I

(k +1) are both rejected, resulting in no change in the booking records, we know that all

the future decisions would be the same for both I and I

processing requests I(k) and I(k +1), and a lower online revenue is generated from sequence I

.

We conclude from the analysis above that a swap of two consecutive fare requests into LBH order will

never lead to a gain in revenue. From any input I, after nite number of swaps, we will eventually end reach

a LBH sequence. Since each swap gains no revenue, that means no other input sequences can have lower

revenue than the LBH input, which is the desired result for the discrete case. For a continuous b, the same

logic applies. The only technical detail that diers involves the swap operation. The swap must be applied

to consecutive requests of arbitrary sizes; the details of this generalization are omitted.

Proposition 2 (Dominance of CAST) Consider a nested booking limit policy b and all input sequences in

some category A

b

j

. CAST

j

A

b

j

dominates the other sequences in category A

b

j

, i.e.,

R(I;b)

R

(I)

R(CAST

j

;b)

R

(CAST

j

)

for all

I A

b

j

.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 3

Proof The framework we use here is similar to the one in the proof of Proposition 1. We show that

from any given sequence I in the desired category A

b

j

, a nite series of input sequences can be constructed,

with non-increasing values of the ratio of the online to oine revenues, leading to a sequence satisfying the

proposition. We present the proof for the discrete case; the proof for the continuous case is similar and is

omitted.

Remember that CASTs have the following properties by denition: they are LBH sequences with a particu-

lar prole. In this proof, we focus on LBH sequences and describe the impact of certain iterative adjustments

to those sequences. This is without loss of generality since any sequence that is not LBH can be transformed

into a new sequence that is LBH using the techniques from the proof of Proposition 1 in Appendix A. The

transformation does not change the prole of the sequence (hence oine revenues are unaected) and the

resulting sequence has online revenues that are no more than the revenues of the original sequence (see

Proposition 1 above).

Consider a LBH sequence I A

b

j

whose prole does not match that of a CAST. First we will make

adjustments to I to ensure I[k] =U

k

for all k j. Suppose there is a k j such that I[k] <U

k

. Consider the

new sequence I

category as I by denition of A

b

j

and it follows that the additional request in I

b (because k j). Therefore, policy b obtains the same online revenues from sequences I and I

. However,

oine revenues for I

will be no less than that of I. Therefore the resultant ratio of online revenues to oine

revenues cannot increase if sequence I is adjusted to I

until I[k] =U

k

for all k j. The resulting sequence may not be LBH, but can be converted to one without

loss of generality: see Proposition 1 and our discussion above.

Now consider I[k] =L

k

with k <j. Suppose there is a k <j such that I[k] >L

k

, and consider removing a

class k request from I, resulting in a sequence I

b

j

where j >k, the

resultant online revenue is simply reduced by f

k

: R(I

; b) =R(I; b) f

k

. As for the oine optimal revenue,

we obviously have R

(I

) R

(I) f

k

. Given that R

(I) R(I; b) f

k

>0, the resultant ratios satisfy

R(I

; b)

R

(I

)

R(I; b) f

k

R

(I) f

k

R(I; b)

R

(I)

.

We continue creating new sequences based on these adjustments until we obtain I[k] =L

k

for all k <j. Thus,

after nite number of adjustments without ever increasing the CR, the desired LBH sequence is reached

from any sequence within the given category.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

4 to appear in Management Science; manuscript no. XXX

Appendix C: Parameter Relationships in the GBM and Proof of Proposition 3

In this Section we examine the coecients in the constraints of GBM and derive some useful relationships

that will serve as the basis for our ability to solve GBM in closed-form. These relationships reect the

structure of the oine optimal revenues of the CASTs.

For the sake of mathematical completeness in investigating the relationship among the problem parameters,

we utilize a virtual fare class m+1 with f

m+1

=0. So, instead of looking at an m-fare problem, we are now

looking at an m+1-fare problem. We refer to the GBM model that includes the virtual class m+1 as the

virtual extension of GBM. We rst clarify the notation for the virtual extension of GBM and introduce the

shorthand notation used in this section.

We set L

m+1

=0 for the virtual class. Let N

j

=n

j1

i=1

L

i

, j =1, 2, , m+1. Remember that

m1

i=1

L

i

<

n (see the discussion in Section 4.2). Therefore, we have N

m

>0. Note that L

m

does not aect any parameters

in the original GBM, so we can set L

m

= min(U

m

, N

m

) > 0, so that N

m+1

0. Also, we set U

m+1

= n, so

we will always have at least n requests in CAST

j

, for all j = 1, 2, , m + 1. Let R

j

= R

(CAST

j

), and

R

+

j

=

j1

i=1

L

i

f

i

for all j =1, 2, , m+1. We have R

m+1

=R

+

m+1

>0 because N

m

>0, L

m

>0 and f

m+1

=0.

With this shorthand notation, we can rewrite g

i

= (R

i

R

i+1

)/f

i

0, i = 1, ..., m from (9). For any input

sequence I, let I(k) denote the fare class of the k

th

request in I, and I(k) denote the fare class of the k

th

request counting backwards in I, for all k =1, 2, ..., |I| where |I| =

m

j=1

I[j] is the total number of requests

in sequence I.

Our initial goal is to derive the following relationship:

N

m+1

=n

m

i=1

L

i

m

i=1

g

i

. (20)

We will also use later some of the inequalities derived along the way.

The oine optimal revenue from an input I is simply the sum of the fares from the n highest requests in

I. As any CAST

j

is LBH, we have

R

j

=

n

k=1

f

CAST

j

(k)

. (21)

We rst seek to derive a relationship between R

j

and R

j+1

. Observe that CAST

j+1

can be obtained from

CAST

j

by taking out U

j

L

j

fare class j m requests. As

j

i=1

L

i

<n, we know that some of those U

j

L

j

requests belong to the n highest-fare requests in CAST

j

. Thus, some of the n highest-fare requests in CAST

j

are missing from CAST

j+1

, and some of the succeeding requests must take the place of these missing

requests. Depending on the value of U

j

+

j1

i=1

L

i

, we have two cases:

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 5

Case 1: If U

j

+

j1

i=1

L

i

n, then all of the U

j

L

j

belong to the n highest-fare requests in CAST

j

.

Thus, exactly U

j

L

j

of the n highest-fare requests in CAST

j

are missing from CAST

j+1

, and the next

U

j

L

j

highest-fare requests are used in determining R

j+1

. So we have:

R

j+1

=R

j

(U

j

L

j

)f

j

+

n+U

j

L

j

k=n+1

f

CAST

j

(k)

. (22)

Case 2: If U

j

+

j1

i=1

L

i

>n, then U

j

+(

j1

i=1

L

i

) n of the U

j

L

j

are not in the highest n requests in

CAST

j

. Thus, only U

j

L

j

[U

j

+(

j1

i=1

L

i

) n] =N

j+1

of the highest n requests from CAST

j

are missing

from CAST

j+1

. These are N

j+1

highest requests following the rst U

j

+

j1

i=1

L

i

highest requests in CAST

j

.

But note that, in CAST

j

, the highest requests ranked between n +1 and U

j

+

j1

i=1

L

i

all belong to fare

class j. This gives us a total of U

j

+(

j1

i=1

L

i

) n =U

j

L

j

N

j+1

requests of class j. So we have:

R

j+1

=R

j

N

j+1

f

j

+

n+U

j

L

j

k=n+1

f

CAST

j

(k)

(U

j

L

j

N

j+1

)f

j

, (23)

which turns out to be identical to (22).

Consider the U

j

L

j

requests that follow the n highest-fare requests in CAST

j

. For j m, let Q

j

be the

number of such requests, where requests in the virtual class are excluded. Also let

f

j

denote the average fare

of these Q

j

requests, so we have

n+U

j

L

j

k=n+1

f

CAST

j

(k)

=Q

j

f

j

. (24)

As none of the Q

j

requests has a fare higher than f

j

, we have

f

j

f

j

and

g

i

=(R

i

R

i+1

)/f

i

=U

i

L

i

Q

i

f

i

/f

i

U

i

L

i

. (25)

As we build the relationships between two consecutive CASTs in this way, starting with CAST

1

and

ending with CAST

m+1

, all the requests not among the n highest of CAST

1

can contribute to a Q

j

for some

j at most once, so we have

m

j=1

Q

j

m

i=1

U

i

n. (26)

The preceding analysis implies:

m

i=1

g

i

=

m

i=1

(U

i

L

i

)

m

i=1

Q

i

f

i

/f

i

m

i=1

(U

i

L

i

)

m

i=1

Q

i

m

i=1

(U

i

L

i

) (

m

i=1

U

i

n)

=n

m

i=1

L

i

=N

m+1

.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

6 to appear in Management Science; manuscript no. XXX

We have now proved relationship (20), which plays an important role in our presentation of closed-form

solutions to the GBM below. We use the shorthand dened in this appendix, instead of the explicit notation of

the main text, for convenience of discussion. We leave it for the reader to verify that the following statements

are the same as in the main text.

Proposition 3 (a) The optimal solution of GBM is

z

CR

=

R

+

u

/f

u

+N

u

R

u

/f

u

+

u1

i=1

g

i

x

CR

j

=

_

_

_

g

j

z

CR

+L

j

j <u

(R

u

z

CR

R

+

u

)/f

u

j =u

0 j >u

u = max{j m: R

+

j

j1

i=1

g

i

<N

j

R

j

}

where the index u denotes the critical fare class such that all classes k >u are closed.

(b) The nested booking limits dened by

b

CR

j

=

m

i=j

x

CR

i

for j =1, , m (27)

maximizes the CR in problem (1) and the optimal CR is z

CR

= z

CR

.

Proof (a) Using condition (20), we see that the u obtained in (8) is the same as

u =max{j m: R

+

j

j1

i=1

g

i

<N

j

R

j

}.

This fact and R

m+1

=R

+

m+1

immediately imply u m. We also see that u 1. Observe that R

+

j

j1

i=1

g

i

is

non-decreasing in j and is equal to 0 for j =1; and that N

j

R

j

is decreasing in j, starting from nR

1

>0 for

j =1. So u is well dened in the range 1 u m.

Consider the system of the linear equations below:

R

j

z =R

+

j

+

u

i=j

f

i

x

i

j =1, , u

u

i=1

x

i

=n.

The closed-form solution to this system, denoted z

, x

i

, is given by

z

= (R

+

u

/f

u

+N

u

)/(R

u

/f

u

+

u1

i=1

g

i

)

x

u

= (R

u

z

R

+

u

)/f

u

x

j

= g

j

z

+L

j

, 1 j <u

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 7

This turns out to be the optimal solution to GBM, after we set x

i

=0, i >u. Some of the constraints were

not part of the equation system used to obtain the above solution; we now need to show that they are all

satised. With z

i

0, i =u. We still need to show x

u

0. By the denition

of u, we have

R

+

u

u1

i=1

g

i

<N

u

R

u

.

By adding R

+

u

R

u

/f

u

to both sides of the inequality and regrouping the terms, we get

R

+

u

u1

i=1

g

i

+R

+

u

R

u

/f

u

<N

u

R

u

+R

+

u

R

u

/f

u

R

+

u

(R

u

/f

u

+

u1

i=1

g

i

) <(N

u

+R

+

u

/f

u

)R

u

.

By the denition of z

R

+

u

<z

u

,

and we get

x

u

=(R

u

z

R

+

u

)/f

u

>0.

Next we show that the constraints of R

j

z R

+

j

+

m

i=j

f

i

x

i

, j > u are satised. Since x

i

= 0, i > u, these

constraints reduce to zR

j

R

+

j

, j >u. Observing the monotonicity of R

+

j

/R

j

in j, we see that we only need

to show z

u+1

R

+

u+1

. By the denition of u, we have:

R

+

u+1

u

i=1

g

i

N

u+1

R

u+1

.

We divide both sides by f

u

and add N

u+1

u

i=1

g

i

to both sides to obtain:

(R

+

u+1

/f

u

+N

u+1

)

u

i=1

g

i

N

u+1

(R

u+1

/f

u

+

u

i=1

g

i

).

Regrouping the terms and replacing z

(R

+

u

/f

u

+N

u

)

u

i=1

g

i

N

u+1

(R

u

/f

u

+

u1

i=1

g

i

)

z

u

i=1

g

i

N

u+1

=N

u

L

u

Finally, rearranging the terms and adding R

+

u

/f

u

to both sides of the inequality we have:

R

+

u

/f

u

+N

u

z

u

i=1

g

i

R

+

u

/f

u

+L

u

=R

+

u+1

/f

u

.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

8 to appear in Management Science; manuscript no. XXX

Since R

+

u

/f

u

+N

u

= z

(R

u

/f

u

+

u1

i=1

g

i

), we have: z

(R

u

/f

u

g

u

) = z

u+1

/f

u

R

+

u+1

/f

u

. The proof is

valid for j =m+1, so we have z

R

+

m+1

/R

m+1

=1.

We now show that the upper bound constraints x

i

U

i

, i = 1, , m also hold. First we show that x

u

z

g

u

+L

u

:

z

u+1

R

+

u+1

z

(R

u

g

u

f

u

) R

+

u

+L

u

f

u

x

u

=(z

u

R

+

u

)/f

u

z

g

u

+L

u

Given g

i

U

i

L

i

and z

1, we have x

i

z

g

i

+L

i

U

i

for i u, and x

i

=0 U

i

for i >u. This concludes

the proof of feasibility of the solution.

To prove optimality, we consider the dual of GBM, with variables v, y

j

, w

j

:

min nv +

m

j=1

(R

+

j

y

j

+U

j

w

j

)

s.t.

m

j=1

R

j

y

j

1

w

j

+v

j

i=1

f

j

y

i

0 j =1, ..., m

v 0, y

j

, w

j

0 j =1, ..., m

A dual feasible solution is constructed as

y

j

=

_

v(1/f

j

1/f

j1

) j u

0 j >u

v = 1/(R

u

/f

u

+

u1

i=1

g

i

)

w

j

= 0 j

where f

0

= for convenience. So the solution is optimal for GBM, and we let x

CR

=x

, z

CR

=z

.

(b) We show optimality of the policy b

CR

to the CR problem. As discussed in the main text, GBM provides

an upper bound on the true objective function, z

CR

, of the CR problem. To prove that b

CR

is the optimal

nested booking limit policy that maximizes CR, we need to show that z

CR

=z

CR

. Obviously, when k u,

R

+

k

gives the exact value for the revenues generated by x

CR

from the requests in classes from 1 through k 1

in CAST

k

. A discrepancy occurs only when k >u. Noting that x

CR

k

=0, we have

R(CAST

k

; b

CR

) =R(CAST

u+1

; b

CR

) =f

u

min(x

CR

u

, L

u

) +R

+

u

=min(R(CAST

u

; b

CR

), R

+

u+1

).

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 9

The monotonicity of R

k

with respect to k allows that

R(CAST

k

; b

CR

)

R

k

min(

R(CAST

u

; b

CR

)

R

u

,

R

+

u+1

R

u+1

) z

CR

.

So, z

CR

=z

CR

.

Appendix D: Proof of Theorem 1

Theorem 1 For the continuous m-fare problem with demand bounds, the nested booking control policy with

booking limit vector b

CR

dened by (11) has a CR of z

CR

given by (6) and this is the best possible among

deterministic policies.

Proof Earlier results showed that b

CR

dened by (11) has a CR of z

CR

given by (6). Thus, to prove this

Theorem, we need only show that an arbitrary policy P can do no better. The only requirement we place

on P is that it should accept/reject all portions of a request upon its arrival. We generate an input sequence

by specifying the actions of an adversary that can observe each of Ps decisions and immediately react

by manipulating future input. The performance of P on this input sequence will be no better than z

CR

implying the result. When specifying the adversarys actions, we characterize the policys eect by dening

an eective booking limit vector, b

P

, and bucket size vector, x

P

. Consider the following adversary strategy:

Step 0: let the current fare index

i =m+1 and its eective booking limit b

P

i

=0;

Step 1: set

i =

i 1, send in U

i

class

i requests;

Step 2: set x

i

to the number of class

i requests accepted by P;

Step 3: let the eective booking limit be b

P

i

=b

P

i+1

+x

P

i

;

Step 4: if b

P

i

b

CR

i

and

i >1 then go to step 1);

Step 5: if b

P

i

<b

CR

i

, send in the rest of CAST

i

.

The execution of this strategy will terminate having generated the input sequence CAST

i

. Dene the

vector b

CR

as the optimal nested booking limit vector obtained using GBM, i.e., b

CR

j

=

m

i=j

x

CR

i

for all

j =1, .., m. The conditions in steps 4 and 5 above imply that b

P

j

b

CR

j

for j >

i and b

P

i

b

CR

i

, which, in turn,

imply that the revenue of policy P based on classes k >

the optimal nested policy b

CR

. The revenue from classes k

i 1 would be at most

i1

k=1

L

i

f

i

, which is the

revenue obtained by the optimal nested policy b

CR

. Combining these two revenue portions, we have that the

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

10 to appear in Management Science; manuscript no. XXX

revenue of P cannot be higher than that of the optimal nested policy b

CR

. Hence, policy b

CR

is better (or

no worse) than any other arbitrary policy. This then implies the CR is at most z

CR

, and the nested booking

control policy b

CR

that achieves z

CR

is the best possible among all deterministic policies.

Appendix E: Proof of Corollary 1

Corollary 1 For the continuous m-fare problem with all lower bounds equal to zero, the nested booking

control policy with booking limits dened by:

b

CR

i

=n

m

j=i

g

i

m

j=1

g

j

for i =2, ..., m

has a CR of n/(

m

j=1

g

j

) and this is the best possible among deterministic policies.

Proof We use the shorthand notation introduced in Appendix C above in this proof. When the lower

bounds are all zero (L

i

=0, i =1, , m), it can easily be seen that R

+

i

=0, N

i

=n for all i =1, , m and

the GBM reduces to the Upper Bound Model (UBM) below:

UBM: max z

s.t. R

j

z

m

i=j

f

i

x

i

, j =1, , m (28)

m

i=1

x

i

n (29)

0 x

j

U

j

, j =1, ..., m. (30)

To derive the optimal solution to UBM, the upper bound constraints in (30) are dropped, and we assume

the remaining constraints in (28) and (29) are binding. We solve a linear system of equations with m+1

variables. The solution z

, x

i

to this system is:

z

=n/

m

i=1

g

i

and x

i

=g

i

z

CR

, i =1, , m; (31)

where

g

i

=

R

i

R

i+1

f

i

i =1, , m (32)

with R

m+1

= 0. Note that g

i

U

i

and 0 z

1, hence x

i

=g

i

z

U

i

, which implies that constraint (30)

is satised. Hence, the solution provided in (31) is feasible for UBM. We consider the dual formulation of

UBM in order to show that this solution is optimal:

min nv

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 11

s.t.

m

j=1

R

j

y

j

1

v f

j

j

i=1

y

i

j =1, ..., m

v 0, y

j

0 j =1, ..., m.

By treating the rst two constraint sets as equalities and solving this system of linear equations, we obtain

the following dual solution: v

=z

/n, y

j

=v

/f

j

v

/f

j1

, where we dene f

0

=+ for convenience. It is

easy to verify that this dual solution is both feasible and has the same objective function value as the primal

solution obtained in (31). This proves that x

CR

=x

z

CR

=z

.

Appendix F: Proof of Theorem 2

Theorem 2 For the continuous m-fare problem with demand bounds, no randomized booking control policy

has a CR larger than z

CR

given in (6). Therefore, the deterministic nested booking control policy with booking

limits by b

CR

in (11) is the best possible among all policies.

Proof (Some of the shorthand notation introduced in Appendix C is used in this proof.) Let

D denote the

set of all deterministic policies for this problem. Let be the set of all probability distributions on

D. Any

randomized policy may be viewed as a random choice

D( p) among deterministic policies, dened by some

probability distribution p . On the other hand, the adversary makes a choice among the input sequences

I (L, U). Using as a payo function the expected CR, we dene a zero-sum two-person game between a

seller choosing a randomized policy to maximize her expected CR and an adversary choosing a distribution

of input sequences to minimize this expected ratio. The von Neuman/Yao principle (e.g., see Seiden, 2000)

implies that the best possible CR of any randomized policy satises

z

=sup

p

inf

I

E

p

_

R(I;

D( p))

R

(I)

_

= inf

q

sup

D

E

q

_

R(I(q);

d)

R

(I(q))

_

where denotes the set of all probability distributions on set , and I(q) is a random sequence chosen

according to probability distribution q . The right hand side of this equality may be interpreted as the

adversarys problem of choosing a probability distribution of input sequences to force every deterministic

policy to experience an expected competitive ratio at most z

CR

is the best possible CR for a deterministic policy and a deterministic policy is a special case of a randomized

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

12 to appear in Management Science; manuscript no. XXX

one, it is clear that z

CR

z

z

CR

. This will be accomplished by

showing for a particular q

,

sup

D

E

q

_

R(I(q

);

d)

R

(I(q

))

_

z

CR

If this inequality holds for a particular q

We dene the q

iq

(f

1

i

f

1

i1

)R

i

) and q

i

= q(f

1

i

f

1

i1

)R

i

, i u. We can then

choose input CAST

i

with probability q

i

. Let X

k

i

(

deterministic online policy

d

D is applied to CAST

k

, and R(I;

d) denote the online revenue generated by

d

from input sequence I. Since CAST

k

and CAST

1

are LBH sequences and they are identical before any k 1

class request is seen, we have X

k

j

(

d) =X

1

j

(

k

j

(

d) CAST

k

[j] =L

j

, j <k, so:

R(CAST

k

;

d) =

m

j=1

f

j

X

k

j

(

d) (33)

R

+

k

+

u

j=k

f

j

X

1

j

(

d) +f

u

m

j=u+1

X

1

j

(

d), k u. (34)

The above expression now enables us to derive a relationship between the performance under an arbitrary

d

and the performance of the best nested booking limit policy dened by x

CR

in (7):

E

q

_

R(I(q

);

d)

R

(I(q

))

_

=

u

k=1

q

k

R(CAST

k

;

d)

R

(CAST

k

)

=

u

k=1

q(f

1

k

f

1

k1

)R(CAST

k

;

d)

Using the relationship in (34),

k=1

q(f

1

k

f

1

k1

)(R

+

k

+

u

j=k

f

j

X

1

j

(

d) +f

u

m

j=u+1

X

1

j

(

d)) (35)

=

u

k=1

q(f

1

k

f

1

k1

)R

+

k

+ q

m

k=1

X

1

k

(36)

k=1

q(f

1

k

f

1

k1

)R

+

k

+ q

m

k=1

x

k

(37)

= z

CR

(38)

The term in (36) is obtained by algebraic manipulation. The inequality in (37) results since

m

k=1

X

1

k

n =

m

k=1

x

CR

k

. The nal equation follows because the CR achieved by x

CR

equals z

CR

for CAST

k

, k u.

Therefore no randomized policy has a CR greater than z

CR

; this completes the proof.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 13

Appendix G: Randomized Policies for the Discrete CR Problem

A randomized policy consists of a policy set and a discrete probability distribution over the policy set. The

policy set is a nite set of deterministic, discrete (integral) policies. We will rst use the solution x

CR

given

in equation (7) to construct a policy set. Then we prove that there always exists a discrete probability

distribution over that policy set such that the randomized policy is optimal.

Note that

u

i=1

x

CR

i

=n because we exclude the trivial cases from our analysis. Each of the deterministic

policies will be dened by rounding up or down each fractional x

CR

i

, and any integral x

CR

i

can be simply

left unchanged. For convenience, we assume that x

CR

i

, i = 1, , u are all fractional. Let p

i

=x

CR

i

x

CR

i

denote the fractional part of x

CR

i

, and r =

u

i=1

p

i

< u denote the sum of the fractional parts. Note that

r =n

u

i=1

x

CR

i

must be integral and satisfy u > r >0. For each s {1, , u}, with |s| = r, we dene a

deterministic policy based on x

s

given by:

x

s

i

= x

CR

i

+1{i s}, i =1, , u (39)

where 1{} is the indicator function. Notice each policy dened in this way satises x

s

i

L

i

, i <u, U

i

x

s

i

, i

u, and

u

i=1

x

s

i

= n. All of these deterministic integral policies are feasible for GBM and there are

_

u

r

_

in

total. These policies form the policy set for our random policy. We now need to nd a probability q

s

for each

policy x

s

, subject to

s:is

q

s

= p

i

, i =1, , u (40)

q

s

0 and

s

q

s

=1. (41)

Note that (41) ensures q

s

represents a probability distribution, and (40) implies E[x

s

i

] = x

CR

i

, i = 1, , u.

Let us assume for now that there exists such a set of q

s

, which we will prove shortly in Lemma 1 (see below).

To show that this randomized policy achieves the optimal performance, we examine its expected perfor-

mance on the CASTs. The analysis provided in Section 4.2 shows that the performance of the policy x

s

upon

CAST

j

is exactly

j1

i=1

f

i

L

i

+

u

i=j

f

i

x

s

i

. For all j u, the expected performance can then be obtained as

follows:

E

_

j1

i=1

f

i

L

i

+

u

i=j

f

i

x

s

i

_

=

j1

i=1

f

i

L

i

+

u

i=j

f

i

E[x

s

i

]

= R

+

j

+

m

i=j

f

i

x

CR

i

= z

CR

R

j

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

14 to appear in Management Science; manuscript no. XXX

where the nal equality in the equation is implied by the properties of the optimal solution of GBM (see

Appendix C above). For j >u, the nal equality above is replaced by the inequality .

Therefore, the expected performance of the randomized policy on each of the CASTs is exactly the same

as the performance of the best deterministic policy for the continuous problem, which proves that the

randomized policy on the discrete problem achieves the same CR as the optimal deterministic policy applied

to the continuous problem. It remains to show that the probabilities of q

s

indeed exist; see the next result.

Lemma 1 Given 0 p

i

1 for i = 1, , u and 0 <

u

i=1

p

i

= r <u, there always exists a solution q

s

to

the system of (40) and (41).

Proof We prove the result by induction on the pair ( r, u). Notice that when r = 1, the solution is q

{i}

=

p

i

, i =1, , u. Also, when u r =1 the solution is obviously q

{1, ,u}\i

= p

i

, i =1, , u, where p

i

=1 p

i

>

0, i =1, , u. It remains to show that for any ( r, u), with r >1 and u r >1, a solution can be found if any

system of (u

, r

), with u

<u and r

r, has a solution.

As indicated earlier, without loss of generality, we assume 0 <p

i

<1 for all i. Let t

i

0, i = 1, , u 1

satisfy

u1

i=1

t

i

=p

u

p

u

and 0 t

i

min(p

u

p

i

, p

u

p

i

). Obviously, such t exists if

p

u

p

u

u1

i=1

min(p

u

p

i

, p

u

p

i

). (42)

We will now prove that this inequality holds. First note that min(X, Y ) =X [X Y ]

+

, so we have

min(p

u

p

i

, p

u

p

i

) =p

u

p

i

[p

u

p

i

p

u

p

i

]

+

=p

u

p

i

[p

u

+p

i

1]

+

.

We dene set K such that i K if and only if p

u

+p

i

1 >0. The following sequence of inequalities are all

equivalent to (42):

p

u

p

u

u1

i=1

(p

u

p

i

[p

u

+p

i

1]

+

)

iK

[p

u

+p

i

1]

+

u1

i=1

p

u

p

i

p

u

p

u

iK

(p

u

+p

i

1) p

u

( r 1)

|K|p

u

|K| +

iK

p

i

p

u

( r 1) (43)

Thus, we can now show that (42) holds by proving (43). We consider two cases:

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 15

Case 1: |K| r. Noting that

iK

p

i

r p

u

, it follows that

|K|p

u

|K| +

iK

p

i

|K|p

u

|K| + r p

u

= p

u

( r 1) p

u

(|K| r).

Now inequality (43) follows since the condition of Case 1 implies p

u

(|K| r) 0.

Case 2: |K| < r. We have

iK

p

i

|K|, so

|K|p

u

|K| +

iK

p

i

|K|p

u

p

u

( r 1).

and inequality (43) follows.

For all i = 1, , u 1, let p

0

i

= p

i

+t

i

/ p

u

, and p

1

i

= p

i

t

i

/p

u

, so we can have p

i

= p

0

i

p

u

+p

1

i

p

u

. From

the constraints put on t, we obviously have 0 p

0

i

1, 0 p

1

i

1 and

u1

i=1

p

0

i

= r,

u1

i=1

p

1

i

= r 1. By the

assumption of induction, we know both systems have a solution, and let them be q

0

s

, q

1

s

respectively. Note

that the subscripts for q

0

s

are sets with |s| = r, while those for q

1

s

satisfy |s| = r 1. Let q

s

be dened as

q

s

=

_

p

u

q

0

s

, u / s

p

u

q

1

s\u

, u s.

(44)

We can directly verify that q

s

is a solution to the system with ( r, u), which is omitted here.

Based on the proof of Lemma 1, one can develop an ecient algorithm to generate the randomized policies.

The algorithm chooses the non-zero probability values without the need for complete enumeration of the

policy set; details are omitted. (The algorithm is included in the ongoing doctoral dissertation of the rst

author, Yingjie Lan, at University of Maryland, College Park, MD.)

Appendix H: Dynamic Policies

Given the adjusted bounds

L,

U, remaining capacity n, number of accepted requests h

i

, i = 1, ..., m, and

the partial input sequence I

0

, the new values of nested booking limits

b and bucket sizes x that improve the

minimum CR and maximum MAR are obtained by solving the problems in (18) and (19), respectively.

The structure of the optimal policy is the same as in GBM (GBM-AR), except that some parameters

are replaced. We use n for n,

L

i

for L

i

. The CASTs for the set of partial input sequences are now dened

over the set (

L,

U). We denote the new CAST by CAST

. By denition, CAST

j

[k] =

U

k

for k j, and

CAST

j

[k] =

L

j

for k <j.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

16 to appear in Management Science; manuscript no. XXX

The new bucket sizes x

CR

that improve minimum CR are expressed as:

z

CR

=

(1/f

u

)(

m

i=1

h

i

f

i

+

u1

i=1

f

i

L

i

) +( n

u1

i=1

L

i

)

+

R

(I

0

CAST

u

)/f

u

+

u1

i=1

g

i

(45)

x

CR

j

=

_

_

_

g

j

z

CR

+

L

j

j < u

(1/f

u

)(R

(I

0

CAST

u

) z

CR

m

i=1

h

i

f

i

u1

i=1

f

i

L

i

) j = u

0 j > u

(46)

u = max{j m:

j1

i=1

g

i

<

R

(I

0

CAST

j

)( n

j1

i=1

L

i

)

+

m

i=1

h

i

f

i

+

j1

i=1

f

i

L

i

} (47)

where the index u denotes the critical fare-class such that all classes k > u are closed, and g

i

is an auxiliary

parameter dened as

g

m

=R (I

0

CAST

m

)/f

m

, g

i

=

R

(I

0

CAST

i

) R

(I

0

CAST

i+1

)

f

i

, i =1, .., m1. (48)

Likewise, the new bucket sizes x

AR

that improve MAR are expressed as:

z

AR

= R

(I

0

CAST

u

)

m

i=1

h

i

f

i

i=1

f

i

L

i

f

u

x

AR

u

(49)

x

AR

j

=

_

_

_

g

j

+

L

j

j < u

( n

1

i=1

L

i

)

+

i< u

g

i

j = u

0 j > u

(50)

u

= max{j m:

i<j

g

i

<( n

j1

i=1

L

i

)

+

.} (51)

One key observation pertaining to our dynamic policies is the following: Static and dynamic policies are

identical under LBH input sequences when m= 2 but not when m>2. This is because the only mistake

the adversary can make in LBH sequences is in the total amount of requests of each class (not the sequence

of arrivals because the sequence is LBH). When m = 2, the only inputs of interest, CASTs, call for the

adversary to send U

2

class-2 requests rst. The mistake implies the adversary switches to sending class-1

requests earlier than expected. Resolving the problem and re-allocating the seats between classes does not

improve the worst-case performance in this case, because there cannot be any more class-2 requests in a

LBH input after the arrival of the rst class-1 request.

Appendix I: Numerical Examples

The results of the computational experiments in this section complement the ones in the main text.

Example-1b. This example is the same as Example-1a of Section 7 except for the demand mix: the

lower and upper bounds on demand are L

1

=20, U

1

=60, L

2

=60 and U

2

=100, i.e., the mean demand for

class-1 is lower. The revenue percentiles for the policies are presented in Figure 6. In this case, no policy

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 17

50000

60000

70000

80000

90000

100000

110000

120000

130000

10% 30% 50% 70% 90%

percentile

p

o

l

i

c

y

r

e

v

e

n

u

e

(

$

)

OFFLINE BDYN-CR, BSTAT-CR

DYN-CR, STAT-CR BDYN-AR, BSTAT-AR

DYN-AR, STAT-AR FCFS

Figure 6 Distribution of revenues for each policy in Example-1b

(except OFFLINE) is stochastically dominant (as observed numerically). This particular demand mix aects

STAT-AR most: It has a signicantly lower revenue at the 10

th

percentile, and provides the second highest

revenue value at the 90

th

percentile, following BSTAT-AR. This is expected because STAT-AR protects the

highest number of seats for class-1, for which the demand is lower in this case. The observations regarding

the relative performances of the other policies remain the same.

Example-1c. This example is the same as Example-1a of Section 7 except the demand mix: the lower

and upper bounds on demand are L

1

=60, U

1

=100, L

2

=20 and U

2

=60, i.e., the mean demand for class-

1 is higher. The revenue percentiles for the policies are presented in Figure 7. In this case, STAT-CR is

stochastically dominated (numerically) because its booking limit for class-2 is signicantly higher than those

of other policies. When the arrivals occur homogeneously over time in both fare classes, dynamic and static

policies are not equivalent. If demand mix is balanced as in Example-1a, then the average revenues obtained

by each of the policies except DYN-AR are higher when arrivals are time-homogeneous; DYN-AR sets a

low booking limit for class-2 initially, and this booking limit is updated with each class-1 request leading to

rejecting far too many class-2 requests and having too many empty seats at the end of the booking horizon.

However, DYN-AR is not stochastically dominated (based on numerical observations), i.e., it has a higher

90

th

percentile than, for e.g., BSTAT-AR.

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

18 to appear in Management Science; manuscript no. XXX

100000

120000

140000

160000

180000

200000

10% 30% 50% 70% 90%

percentile

p

o

l

i

c

y

r

e

v

e

n

u

e

OFFLINE FCFS BDYN-CR, BSTAT-CR

DYN-CR, STAT-CR BDYN-AR, BSTAT-AR DYN-AR, STAT-AR

Figure 7 Distribution of revenues for each policy in Example-1c

These experiments are repeated by changing the arrival regime. Our previous observations regarding

stochastic dominance relations among the policies do not change with the arrival regime. The next three

examples provide the revenue percentiles for these experiments when arrivals are time-homogeneous.

Example-1d. This example is the same as Example-1a of Section 7 except that the requests in each fare

class arrive homogeneously over time. The revenue percentiles for the policies are presented in Figure 8.

Example-1e. This example is the same as Example-1b above except that the requests in each fare class

arrive homogeneously over time. The revenue percentiles for the policies are presented in Figure 9.

Example-1f. This example is the same as Example-1c above except that the requests in each fare class

arrive homogeneously over time. The revenue percentiles for the policies are presented in Figure 10.

Example-2b. This example is the same as Example-2a of Section 7 except for the demand mix: demand is

distributed uniformly between 30 and 50 for class-1, and between 70 and 90 for class-2. The average relative

performance of each policy is reported in Figure 11. In this case, the relative performance of policies with

demand bounds is again indistinguishable and is insensitive to the changes in the fare ratio. However, the

relative performance of policies with no demand information is dierent: more demand for class-2 favors

STAT-CR and DYN-CR, and these policies perform as well as the policies with demand bounds when the

fare ratio is low. The performances of STAT-AR and DYN-AR are adversely aected by this demand mix

when the fare ratio is low (e.g. their average relative performance is 84% for r = 0.2 as opposed to 94% in

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 19

90000

100000

110000

120000

130000

140000

150000

160000

10 30 50 70 90

percentile

p

o

l

i

c

y

r

e

v

e

n

u

e

OFFLINE FCFS BDYN-CR

DYN-CR DYN-AR BSTAT-CR

STAT-CR STAT-AR BDYN-AR~BSTAT-AR

Figure 8 Distribution of revenues for each policy in Example-1d

52000

62000

72000

82000

92000

102000

112000

122000

132000

10 30 50 70 90

percentile

p

o

l

i

c

y

r

e

v

e

n

u

e

OFFLINE FCFS BDYN-CR DYN-CR

BDYN-AR DYN-AR BSTAT-CR STAT-CR

BSTAT-AR STAT-AR

Figure 9 Distribution of revenues for each policy in Example-1e

Figure 2). When demand is distributed uniformly between 20 and 60 for class-2, and between 60 and 100

for class-1, then MAR policies have slightly better performance compared to CR when demand information

is used. Low r and high demand for class-1 favor BSTAT-AR (BDYN-AR). The eect of time-homogeneous

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

20 to appear in Management Science; manuscript no. XXX

125000

135000

145000

155000

165000

175000

185000

195000

10 30 50 70 90

percentile

p

o

l

i

c

y

r

e

v

e

n

u

e

OFFLINE FCFS~STAT-CR

DYN-CR DYN-AR

BDYN-AR~BSTAT-AR STAT-AR

BDYN-CR~BSTAT-CR

Figure 10 Distribution of revenues for each policy in Example-1f

0.84

0.86

0.88

0.9

0.92

0.94

0.96

0.98

1

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

fare ratio (r)

a

v

e

r

a

g

e

o

f

r

a

t

i

o

n

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

BDYN-AR, BSTAT-AR DYN-AR, STAT-AR

BDYN-CR, BSTAT-CR DYN-CR, STAT-CR

FCFS

Figure 11 Average relative performance of policies in Example-2b

arrivals on our policies, in general, is the same as in Example-1a.

Example-4 (extended). Here, we report extended results on Example-4 of Section 7, including the

performance of policies that use no demand information. When arrivals are time-homogeneous (see Figure

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 21

0.8

0.84

0.88

0.92

0.96

1

[

1

2

0

,

0

]

[

1

1

0

,

1

0

]

[

1

0

0

,

2

0

]

[

9

0

,

3

0

]

[

8

0

,

4

0

]

[

7

0

,

5

0

]

[

6

0

,

6

0

]

[

5

0

,

7

0

]

[

4

0

,

8

0

]

[

3

0

,

9

0

]

[

2

0

,

1

0

0

]

[

1

0

,

1

1

0

]

[

0

,

1

2

0

]

mean demand [

2

,

1

]

a

v

e

r

a

g

e

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

DYN-CR STAT-CR FCFS EMSR

MDP BDYN-CR BSTAT-CR BSTAT-AR

BDYN-AR DYN-AR STAT-AR

Figure 12 Average relative performance of policies in Example-4 when arrivals are time-homogeneous

12), MDP is the optimal policy and it performs remarkably well compared to the oine optimal. EMSR is

only a heuristic in this case. There are three important observations about the robust policies: First, the use

of demand bounds improves the performance signicantly. Second, there is minimal dierence among the

robust policies that use demand information. Third, robust policies that use demand bounds perform as well

as EMSR except in the extreme case where class-2 demand is negligible and class-1 demand exceeds capacity

(which is unrealistic and/or not interesting from a RM perspective). When the arrivals follow the LBH

regime (see Figure 13), EMSR is the optimal policy. Note that the robust policies with demand information

are again indistinguishable from each other and also from EMSR for practical instances. Note that among

the policies that use no demand information, static (dynamic) policies perform better than dynamic (static)

policies when class-2 (class-1) demand is higher than class-1 (class-2) demand. This is a direct consequence

of the dierences in the booking limits of static vs. dynamic policies.

Example-5 (extended). We now report additional results on Example 5 of Section 7. First of all, when

the CoV=0.4, the average of ratio of policy revenues to OFFLINE is 0.86, 0.87, 0.69, and 0.70 for DYN-CR,

STAT-CR, DYN-AR, and STAT-AR, respectively. The benet of demand information is signicant especially

on MAR policies in this example: DYN-AR and STAT-AR achieve the lowest revenues (signicantly lower

than any other policy) in this experiment, while BSTAT-AR and BDYN-AR are as good as DP which provides

the optimal solution. When the CoV increases, and demand is truncated (to eliminate negative demand

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

22 to appear in Management Science; manuscript no. XXX

0.8

0.84

0.88

0.92

0.96

1

[

1

2

0

,

0

]

[

1

1

0

,

1

0

]

[

1

0

0

,

2

0

]

[

9

0

,

3

0

]

[

8

0

,

4

0

]

[

7

0

,

5

0

]

[

6

0

,

6

0

]

[

5

0

,

7

0

]

[

4

0

,

8

0

]

[

3

0

,

9

0

]

[

2

0

,

1

0

0

]

[

1

0

,

1

1

0

]

[

0

,

1

2

0

]

mean demand [

2

,

1

]

a

v

e

r

a

g

e

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

DYN-CR, STAT-CR FCFS

EMSR BDYN-CR, BSTAT-CR

BSTAT-AR, BDYN-AR DYN-AR, STAT-AR

Figure 13 Average relative performance of policies in Example-4 when arrivals are LBH

values) in the simulations, the relative performance of the policies with respect to OFFLINE decreases, and

MAR policies are not as good as DP. However, among the policies that use demand bounds, BSTAT-AR and

BDYN-AR perform better than BSTAT-CR and BDYN-CR, both of which perform signicantly better than

EMSR-b; see Figure 14. Notice that the x-axis in the gure reports the CoV used in computing the policy

parameters, but this CoV is not valid for the simulated demand since the mean and variance are distorted

when the distribution is truncated. In any case, the results in Figure 14 show the dierences among our

policies and EMSR in a setting where the demand distribution changes.

Example-6 (extended). We present extended results on Example 6 of Section 7 here. We report the

results of our policies with demand information and also their special cases with no demand information. The

average relative performance of the policies as a function of the fare ratio r is given in Figure 15. DYN-CR,

STAT-AR and DYN-AR perform very well in this example. This is not surprising because they protect more

seats for class-1 in general and it pays o in this case. In fact, DYN-AR protects too many and is worse than

STAT-AR. STAT-CR performs relatively poorly especially when r is small (this observation is the same as

in Example-2a and is a factor of the protection level of STAT-CR and the demand for class-1).

Next we test the eect of errors in estimation/specication of parameters of the demand distribution.

Example-7. We have m=2, (f

1

, f

2

) =(500, 100), and the demand is Poisson distributed. The true mean

demand for each fare class is 60 for the entire booking horizon. While the true demand distribution of class-2

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 23

65%

70%

75%

80%

85%

90%

95%

100%

0.4 0.8 1.2 1.6

coefficient of variation

a

v

e

r

a

g

e

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

DP FCFS EMSR-B BDYN-AR

BSTAT-AR DYN-CR STAT-CR DYN-AR

BDYN-CR BSTAT-CR STAT-AR

Figure 14 Average relative performance of policies in Example-5 when CoV increases and demand is truncated.

86%

88%

90%

92%

94%

96%

98%

100%

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

fare ratio (r=f

2

/f

1

)

a

v

e

r

a

g

e

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

DYN-CR STAT-CR BDYN-CR BSTAT-CR

DYN-AR STAT-AR BDYN-AR BSTAT-AR

Figure 15 Average relative performance of policies in Example-6

is known, only an estimate of

1

, the parameter of demand distribution for class-1, is available. All policy

parameters are determined based on the estimates. The arrivals are LBH. The average relative performance

of the policies is presented in Figure 16, where the estimate of mean demand for class-1 varies from 30

to 90. Notice that our policies with demand bounds behave similar to EMSR. They perform relatively

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

24 to appear in Management Science; manuscript no. XXX

76%

78%

80%

82%

84%

86%

88%

90%

92%

94%

96%

98%

100%

30 40 50 60 70 80 90

estimate of

1

a

v

g

.

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

DYN-CR BDYN-CR FCFS

DYN-AR BSTAT-AR EMSR

Figure 16 Average relative performance of policies in Example-7 when arrivals are LBH.

well given a moderate forecast error (e.g., mean in the range 45 to 75). LBH arrivals amplify the eect

of underestimating high-fare demand. Our policies with demand information perform slightly better than

EMSR when class-1 demand is underestimated or highly overestimated. This is natural because (i) our

policies tend to overprotect seats, hence, the adverse aect of underestimating the demand for class-1 is less,

and (ii) the demand information for our policies is only valuable if the upper bound on demand is no more

than the capacity (i.e., overestimation becomes less detrimental to the performance beyond a certain point).

We repeated this example with time-homogeneous arrivals: the average relative performance of policies is

presented in Figure 17.

Example-8. In this example, the demand and fare parameters are the same as Example-4 above. The

demand is Poisson distributed and demand mix varies from [120,0] to [0,120]. The mean demand for the

entire booking horizon is known. In this experiment, MDP assumes stationary arrivals for the entire booking

horizon, whereas all the demand arrives in a rush only in the rst half of the booking horizon. This unexpected

burst of arrivals hampers the performance of time- and capacity-dependent policy of MDP. As seen in Figure

18, MDP has poor performance when

2

is high and

1

is low because it ends up rejecting too many class-2

requests. In contrast, robust policies and EMSR are not aected by modeling/estimation errors regarding

inter-arrival times. While this example may be an extreme one, it illustrates that a control policy obtained

via MDP is sensitive to time-varying rate of arrivals even though it appears to be robust for misspecations

Y.Lan, H.Gao, M.Ball, I.Karaesmen: Revenue Management with Limited Demand Information

to appear in Management Science; manuscript no. XXX 25

86%

88%

90%

92%

94%

96%

98%

100%

30 40 50 60 70 80 90

estimate of

1

a

v

e

r

a

g

e

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

DYN-CR STAT-CR DYN-AR STAT-AR

BDYN-CR BSTAT-CR BDYN-AR BSTAT-AR

EMSR FCFS MDP

Figure 17 Average relative performance of policies in Example-7 when arrivals are time-homogeneous.

84%

86%

88%

90%

92%

94%

96%

98%

100%

[

1

2

0

,

0

]

[

1

1

0

,

1

0

]

[

1

0

0

,

2

0

]

[

9

0

,

3

0

]

[

8

0

,

4

0

]

[

7

0

,

5

0

]

[

6

0

,

6

0

]

[

5

0

,

7

0

]

[

4

0

,

8

0

]

[

3

0

,

9

0

]

[

2

0

,

1

0

0

]

[

1

0

,

1

1

0

]

[

0

,

1

2

0

]

mean demand [

2

,

1

]

a

v

g

.

o

f

r

a

t

i

o

o

f

p

o

l

i

c

y

r

e

v

e

n

u

e

t

o

O

F

F

L

I

N

E

r

e

v

e

n

u

e

BSTAT-CR BDYN-AR BSTAT-AR FCFS

EMSR MDP BDYN-CR

Figure 18 Average relative performance of policies in Example-8

of the arrival rate when arrivals are time-stationary.

## Гораздо больше, чем просто документы.

Откройте для себя все, что может предложить Scribd, включая книги и аудиокниги от крупных издательств.

Отменить можно в любой момент.