An enhanced concave program relaxation for choice network revenue management

The network choice revenue management problem models customers as choosing from an offer-set, andthe firm decides the best subset to offer at any given moment to maximize expected revenue. The resultingdynamic program for the firm is intractable and approximated by a deterministic linear programcalled the CDLP which has an exponential number of columns. However, under the choice-set paradigmwhen the segment consideration sets overlap, the CDLP is difficult to solve. Column generation has beenproposed but finding an entering column has been shown to be NP-hard. In this paper, starting with aconcave program formulation based on segment-level consideration sets called SDCP, we add a class ofconstraints called product constraints, that project onto subsets of intersections. In addition we proposea natural direct tightening of the SDCP called ?SDCP, and compare the performance of both methodson the benchmark data sets in the literature. Both the product constraints and the ?SDCP method arevery simple and easy to implement and are applicable to the case of overlapping segment considerationsets. In our computational testing on the benchmark data sets in the literature, SDCP with productconstraints achieves the CDLP value at a fraction of the CPU time taken by column generation and webelieve is a very promising approach for quickly approximating CDLP when segment consideration setsoverlap and the consideration sets themselves are relatively small.

Network revenue management with product-specific no-shows

Revenue management practices often include overbooking capacity to account for customerswho make reservations but do not show up. In this paper, we consider the network revenuemanagement problem with no-shows and overbooking, where the show-up probabilities are specificto each product. No-show rates differ significantly by product (for instance, each itinerary andfare combination for an airline) as sale restrictions and the demand characteristics vary byproduct. However, models that consider no-show rates by each individual product are difficultto handle as the state-space in dynamic programming formulations (or the variable space inapproximations) increases significantly. In this paper, we propose a randomized linear program tojointly make the capacity control and overbooking decisions with product-specific no-shows. Weestablish that our formulation gives an upper bound on the optimal expected total profit andour upper bound is tighter than a deterministic linear programming upper bound that appearsin the existing literature. Furthermore, we show that our upper bound is asymptotically tightin a regime where the leg capacities and the expected demand is scaled linearly with the samerate. We also describe how the randomized linear program can be used to obtain a bid price controlpolicy. Computational experiments indicate that our approach is quite fast, able to scale to industrialproblems and can provide significant improvements over standard benchmarks.

Demand and revenue implications of an integrated public transport policy: the case of Madrid

One of the most popular options for promoting public transport use is the provision of an integrated and high quality public transport system. This was the strategy adopted by the regional government in Madrid in 1986 and since then public transport patronage has increased by more than 50%. This paper has two objectives. The first is to identify the factors underlying the significant increase in the demand for public transport in Madrid. To do this we estimate an aggregate demand function for bus and underground trips, which allows us to obtain the demand elasticities with respect to the main attributes of public transport services and also to calculate the long-term impact of changes in those explanatory variables on patronage. The second objective is to evaluate the impact on revenue derived from the introduction of the travel card scheme, and to discuss the consequences on revenue of changes in the relative fare levels of different types of ticket without substantially affecting patronage. This latter issue is addressed by estimating a matrix of own and cross-price elasticities for different ticket types.

Weak and Strong Altruism in Trait Groups: Reproductive Suicide, Personal Fitness, and Expected Value

A simple variant of trait group selection, employing predators as the mechanism underlying group selection, supports contingent reproductive suicide as altruism (i.e., behavior lowering personal fitness while augmenting that of another) without kin assortment. The contingent suicidal type may either saturate the population or be polymorphic with a type avoiding suicide, depending on parameters. In addition to contingent suicide, this randomly assorting morph may also exhibit continuously expressed strong altruism (sensu Wilson 1979) usually thought restricted to kin selection. The model will not, however, support a sterile worker caste as such, where sterility occurs before life history events associated with effective altruism; reproductive suicide must remain fundamentally contingent (facultative sensu West Eberhard 1987; Myles 1988) under random assortment. The continuously expressed strong altruism supported by the model may be reinterpreted as probability of arbitrarily committing reproductive suicide, without benefit for another; such arbitrary suicide (a "load" on "adaptive" suicide) is viable only under a more restricted parameter space relative to the necessarily concomitant adaptive contingent suicide.

Are rules-based government programs shielded from special-interest politics? Evidence from revenue-sharing transfers in Brazil

Manipulation of government finances for the benefit of narrowly defined groups is usuallythought to be limited to the part of the budget over which politicians exercise discretion inthe short run, such as earmarks. Analyzing a revenue-sharing program between the centraland local governments in Brazil that uses an allocation formula based on local population estimates,I document two main results: first, that the population estimates entering the formulawere manipulated and second, that this manipulation was political in nature. Consistent withswing-voter targeting by the right-wing central government, I find that municipalities withroughly equal right-wing and non-right-wing vote shares benefited relative to opposition orconservative core support municipalities. These findings suggest that the exclusive focus ondiscretionary transfers in the extant empirical literature on special-interest politics may understatethe true scope of tactical redistribution that is going on under programmatic disguise.

Resolving inconsistencies in utility measurement under risk: Tests of generalizations of expected utility

This paper explores biases in the elicitation of utilities under risk and the contribution that generalizations of expected utility can make to the resolution of these biases. We used five methods to measure utilities under risk and found clear violations of expected utility. Of the theories studies, prospect theory was most consistent with our data. The main improvement of prospect theory over expected utility was in comparisons between a riskless and a risky prospect(riskless-risk methods). We observed no improvement over expected utility in comparisons between two risky prospects (risk-risk methods). An explanation why we found no improvement of prospect theory over expected utility in risk-risk methods may be that there was less overweighting of small probabilities in our study than has commonly been observed.

The customer valuations game as a basis for revenue management

I describe the customer valuations game, a simple intuitive game that can serve as a foundation for teaching revenue management. The game requires little or no preparation, props or software, takes around two hours (and hence can be finished in one session), and illustrates the formation of classical (airline and hotel) revenue management mechanisms such as advanced purchase discounts, booking limits and fixed multiple prices. I normally use the game as a base to introduce RM and to develop RM forecasting and optimization concepts off it. The game is particularly suited for non-technical audiences.

A randomized concave programming method for choice network revenue management

Models incorporating more realistic models of customer behavior, as customers choosing froman offer set, have recently become popular in assortment optimization and revenue management.The dynamic program for these models is intractable and approximated by a deterministiclinear program called the CDLP which has an exponential number of columns. However, whenthe segment consideration sets overlap, the CDLP is difficult to solve. Column generationhas been proposed but finding an entering column has been shown to be NP-hard. In thispaper we propose a new approach called SDCP to solving CDLP based on segments and theirconsideration sets. SDCP is a relaxation of CDLP and hence forms a looser upper bound onthe dynamic program but coincides with CDLP for the case of non-overlapping segments. Ifthe number of elements in a consideration set for a segment is not very large (SDCP) can beapplied to any discrete-choice model of consumer behavior. We tighten the SDCP bound by(i) simulations, called the randomized concave programming (RCP) method, and (ii) by addingcuts to a recent compact formulation of the problem for a latent multinomial-choice model ofdemand (SBLP+). This latter approach turns out to be very effective, essentially obtainingCDLP value, and excellent revenue performance in simulations, even for overlapping segments.By formulating the problem as a separation problem, we give insight into why CDLP is easyfor the MNL with non-overlapping considerations sets and why generalizations of MNL posedifficulties. We perform numerical simulations to determine the revenue performance of all themethods on reference data sets in the literature.

A tractable consideration set structure for network revenue management

Models incorporating more realistic models of customer behavior, as customers choosing from an offerset, have recently become popular in assortment optimization and revenue management. The dynamicprogram for these models is intractable and approximated by a deterministic linear program called theCDLP which has an exponential number of columns. When there are products that are being consideredfor purchase by more than one customer segment, CDLP is difficult to solve since column generationis known to be NP-hard. However, recent research indicates that a formulation based on segments withcuts imposing consistency (SDCP+) is tractable and approximates the CDLP value very closely. In thispaper we investigate the structure of the consideration sets that make the two formulations exactly equal.We show that if the segment consideration sets follow a tree structure, CDLP = SDCP+. We give acounterexample to show that cycles can induce a gap between the CDLP and the SDCP+ relaxation.We derive two classes of valid inequalities called flow and synchronization inequalities to further improve(SDCP+), based on cycles in the consideration set structure. We give a numeric study showing theperformance of these cycle-based cuts.

Equivalence of piecewise-linear approximation and Lagrangian relaxation for network revenue management

The network revenue management (RM) problem arises in airline, hotel, media,and other industries where the sale products use multiple resources. It can be formulatedas a stochastic dynamic program but the dynamic program is computationallyintractable because of an exponentially large state space, and a number of heuristicshave been proposed to approximate it. Notable amongst these -both for their revenueperformance, as well as their theoretically sound basis- are approximate dynamic programmingmethods that approximate the value function by basis functions (both affinefunctions as well as piecewise-linear functions have been proposed for network RM)and decomposition methods that relax the constraints of the dynamic program to solvesimpler dynamic programs (such as the Lagrangian relaxation methods). In this paperwe show that these two seemingly distinct approaches coincide for the network RMdynamic program, i.e., the piecewise-linear approximation method and the Lagrangianrelaxation method are one and the same.

A new compact linear programming formulation for choice network revenue management

The choice network revenue management model incorporates customer purchase behavioras a function of the offered products, and is the appropriate model for airline and hotel networkrevenue management, dynamic sales of bundles, and dynamic assortment optimization.The optimization problem is a stochastic dynamic program and is intractable. A certainty-equivalencerelaxation of the dynamic program, called the choice deterministic linear program(CDLP) is usually used to generate dyamic controls. Recently, a compact linear programmingformulation of this linear program was given for the multi-segment multinomial-logit (MNL)model of customer choice with non-overlapping consideration sets. Our objective is to obtaina tighter bound than this formulation while retaining the appealing properties of a compactlinear programming representation. To this end, it is natural to consider the affine relaxationof the dynamic program. We first show that the affine relaxation is NP-complete even for asingle-segment MNL model. Nevertheless, by analyzing the affine relaxation we derive a newcompact linear program that approximates the dynamic programming value function betterthan CDLP, provably between the CDLP value and the affine relaxation, and often comingclose to the latter in our numerical experiments. When the segment consideration sets overlap,we show that some strong equalities called product cuts developed for the CDLP remain validfor our new formulation. Finally we perform extensive numerical comparisons on the variousbounds to evaluate their performance.

Revenue management under general discrete choice model of consumer behavior

Customer choice behavior, such as 'buy-up' and 'buy-down', is an importantphe-nomenon in a wide range of industries. Yet there are few models ormethodologies available to exploit this phenomenon within yield managementsystems. We make some progress on filling this void. Specifically, wedevelop a model of yield management in which the buyers' behavior ismodeled explicitly using a multi-nomial logit model of demand. Thecontrol problem is to decide which subset of fare classes to offer ateach point in time. The set of open fare classes then affects the purchaseprobabilities for each class. We formulate a dynamic program todetermine the optimal control policy and show that it reduces to a dynamicnested allocation policy. Thus, the optimal choice-based policy caneasily be implemented in reservation systems that use nested allocationcontrols. We also develop an estimation procedure for our model based onthe expectation-maximization (EM) method that jointly estimates arrivalrates and choice model parameters when no-purchase outcomes areunobservable. Numerical results show that this combined optimization-estimation approach may significantly improve revenue performancerelative to traditional leg-based models that do not account for choicebehavior.

Proving the performance of a new revenue management system

Revenue management (RM) is a complicated business process that can best be described ascontrol of sales (using prices, restrictions, or capacity), usually using software as a tool to aiddecisions. RM software can play a mere informative role, supplying analysts with formatted andsummarized data who use it to make control decisions (setting a price or allocating capacity fora price point), or, play a deeper role, automating the decisions process completely, at the otherextreme. The RM models and algorithms in the academic literature by and large concentrateon the latter, completely automated, level of functionality.A firm considering using a new RM model or RM system needs to evaluate its performance.Academic papers justify the performance of their models using simulations, where customerbooking requests are simulated according to some process and model, and the revenue perfor-mance of the algorithm compared to an alternate set of algorithms. Such simulations, whilean accepted part of the academic literature, and indeed providing research insight, often lackcredibility with management. Even methodologically, they are usually awed, as the simula-tions only test \within-model" performance, and say nothing as to the appropriateness of themodel in the first place. Even simulations that test against alternate models or competition arelimited by their inherent necessity on fixing some model as the universe for their testing. Theseproblems are exacerbated with RM models that attempt to model customer purchase behav-ior or competition, as the right models for competitive actions or customer purchases remainsomewhat of a mystery, or at least with no consensus on their validity.How then to validate a model? Putting it another way, we want to show that a particularmodel or algorithm is the cause of a certain improvement to the RM process compared to theexisting process. We take care to emphasize that we want to prove the said model as the causeof performance, and to compare against a (incumbent) process rather than against an alternatemodel.In this paper we describe a \live" testing experiment that we conducted at Iberia Airlineson a set of flights. A set of competing algorithms control a set of flights during adjacentweeks, and their behavior and results are observed over a relatively long period of time (9months). In parallel, a group of control flights were managed using the traditional mix of manualand algorithmic control (incumbent system). Such \sandbox" testing, while common at manylarge internet search and e-commerce companies is relatively rare in the revenue managementarea. Sandbox testing has an undisputable model of customer behavior but the experimentaldesign and analysis of results is less clear. In this paper we describe the philosophy behind theexperiment, the organizational challenges, the design and setup of the experiment, and outlinethe analysis of the results. This paper is a complement to a (more technical) related paper thatdescribes the econometrics and statistical analysis of the results.

On bounds for network revenue management

The Network Revenue Management problem can be formulated as a stochastic dynamic programming problem (DP or the\optimal" solution V *) whose exact solution is computationally intractable. Consequently, a number of heuristics have been proposed in the literature, the most popular of which are the deterministic linear programming (DLP) model, and a simulation based method, the randomized linear programming (RLP) model. Both methods give upper bounds on the optimal solution value (DLP and PHLP respectively). These bounds are used to provide control values that can be used in practice to make accept/deny decisions for booking requests. Recently Adelman [1] and Topaloglu [18] have proposed alternate upper bounds, the affine relaxation (AR) bound and the Lagrangian relaxation (LR) bound respectively, and showed that their bounds are tighter than the DLP bound. Tight bounds are of great interest as it appears from empirical studies and practical experience that models that give tighter bounds also lead to better controls (better in the sense that they lead to more revenue). In this paper we give tightened versions of three bounds, calling themsAR (strong Affine Relaxation), sLR (strong Lagrangian Relaxation) and sPHLP (strong Perfect Hindsight LP), and show relations between them. Speciffically, we show that the sPHLP bound is tighter than sLR bound and sAR bound is tighter than the LR bound. The techniques for deriving the sLR and sPHLP bounds can potentially be applied to other instances of weakly-coupled dynamic programming.

Weak and strong altruism in traitgGroups: Reproductive suicide, personal fitness and expected value

A simple variant of trait group selection, employing predators as themechanism underlying group selection, supports contingent reproductivesuicide as altruism (i.e., behavior lowering personal fitness whileaugmenting that of another) without kin assortment. The contingentsuicidal type may either saturate the population or be polymorphicwith a type avoiding suicide, depending on parameters. In addition tocontingent suicide, this randomly assorting morph may also exhibitcontinuously expressed strong altruism (sensu Wilson 1979) usuallythought restricted to kin selection. The model will not, however,support a sterile worker caste as such, where sterility occurs beforelife history events associated with effective altruism; reproductivesuicide must remain fundamentally contingent (facultative sensu WestEberhard 1987; Myles 1988) under random assortment. The continuouslyexpressed strong altruism supported by the model may be reinterpretedas probability of arbitrarily committing reproductive suicide, withoutbenefit for another; such arbitrary suicide (a "load" on "adaptive"suicide) is viable only under a more restricted parameter spacerelative to the necessarily concomitant adaptive contingent suicide.