Restaurant Revenue Management

RESTAURANT REVENUE MANAGEMENT DIMITRIS BERTSIMAS Sloan School of Management, E53-363, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, dbertsim@mit.edu ROMY SHIODA Operations Research Center, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, romy@mit.edu We develop two classes of optimization models to maximize revenue in a restaurant (while controlling average waiting time as well as perceived fairness) that may violate the first-come-first-serve (FCFS) rule. In the first class of models, we use integer programming, stochastic programming, and approximate dynamic programming methods to decide dynamically when, if at all, to seat an incoming party during the day of operation of a restaurant that does not accept reservations. In a computational study with simulated data, we show that optimization-based methods enhance revenue relative to the industry practice of FCFS by 0.11% to 2.22% for low-load factors, by 0.16% to 2.96% for medium-load factors, and by 7.65% to 13.13% for high-load factors, without increasing, and occasionally decreasing, waiting times compared to FCFS. The second class of models addresses reservations. We propose a two-step procedure: Use a stochastic gradient algorithm to decide a priori how many reservations to accept for a future time and then use approximate dynamic programming methods to decide dynamically when, if at all, to seat an incoming party during the day of operation. In a computational study involving real data from an Atlanta restaurant, the reservation model improves revenue relative to FCFS by 3.5% for low-load factors and 7.3% for high-load factors. Received January 2001; revisions received September 2001, June 2002; accepted June 2002. Subject classifications: Dynamic programming: revenue management, stochastic optimization. Industries: restaurant. Area of review: Financial Engineering. 1. INTRODUCTION from the belief that restaurants can increase their revenue by optimizing their nesting decisions, i.e., when to save tables in anticipation for larger parties, even when there are smaller parties currently in queue. We develop two classes of optimization models to maximize revenue in a restaurant, while controlling average waiting time as well as perceived fairness, that may violate the FCFS rule. In the first class of models, we use integer programming, stochastic programming, and approximate dynamic programming methods to decide dynamically when, if at all, to seat an incoming party during the day of operation of a restaurant that does not accept reservations. In the second class of models, we use a stochastic gradient algorithm to decide when, if at all, to accept a reservation for a future time and also incorporate reservations in a dynamic model. We illustrate, using both simulated and real data, that our models lead to significant revenue enhancements relative to FCFS. Maximizing efficiency is of utmost importance in order for large and popular restaurants to increase their profits and to remain competitive. This can explain the surge in the usage of point of sales (POS) softwares that track the arrival time, size, and order of each customer. Although floor managers can utilize these tools as an aid to better estimate the remaining service time of the customers and to see which tables need to speed up their service, their seating policy is mainly based on intuition brought on by experience. In most cases, they follow a simple first-comefirst-serve (FCFS) policy. The challenge of a floor manager is to decide when and where to seat each arriving customer. If there are only tables of four available and a party of two enters, does he seat the party at the larger table or reserve it for a larger, more revenue-producing party? In addition, if the restaurant takes reservations, he needs to further decide how to seat walk-in customers so that they would not take tables away from the reservation customers while considering the possibility of no-shows. These are important practical issues for restaurant managers, where in some cases a good floor manager can make the difference of couple of hundred dollars per night (Kosugi 1999). Thus, a tool that can help floor managers better make these decisions would be of significant value to a restaurant. With all the data that are collected by the POS software, a revenue-maximizing seating policy can be utilized. We believe the key lies in nesting—where parties are seated at tables that can seat larger parties. The present paper stems Operations Research © 2003 INFORMS Vol. 51, No. 3, May–June 2003, pp. 472–486 Literature Revenue management in general is the practice of maximizing a company’s revenue by optimally choosing which customers to serve. It has been used extensively in the airline, hotel, and car rental industries. McGill and van Ryzin (1999) give a comprehensive overview of the history of revenue management in transportation, where it has had the greatest impact. In the case of restaurants, restaurant managers want to allocate their tables by seating the largest possible party at each table—assuming the total bill increases with party size. However, they need to also consider seating small parties at large tables when the larger parties are not expected 472 0030-364X/03/5103-0472 $05.00 1526-5463 electronic ISSN Bertsimas and Shioda to arrive in the near future, because they would rather seat them than have an empty table. Thus, the challenge is to understand the demand flow throughout the day of each type of customer and optimize the allocation of the tables among them. Unlike the widespread application of revenue management methods in airlines, hotels, and rental cars, the number and depth of studies on revenue management in restaurants have been comparatively slim. Kimes (1998, 1999) and Kimes et al. (1999) have been among the first to directly address the issue of restaurant revenue management. They built a strategic framework for applying revenue management for restaurants to increase demand, and thus revenue, by effective duration management and demand-based pricing. They proposed using the revenue per available seat hour (RevPASH
= revenue accrued in a given time interval divided by the number of seats available during that time) as the restaurant’s performance metric. This value is calculated for each time period of each day, to identify times that the RevPASH is low. Similarly, Sill et al. (1999, 2000) propose the use of CapacityManagement Science (CMS)® as a systematic method of assessing the restaurant’s capacity potential and process efficiency. CMS® involves monitoring every component of the service and production delivery process with quantifiable measurements to improve customer satisfaction, enhance employee work life, and increase profit. Although not specified under the name of revenue management, many other approaches have been proposed to increase the revenue of restaurants. Vakharia et al. (1992) developed models and heuristics to find the best tradeoff between wages and hour preferences to minimize the cost of employees while maintaining employee satisfaction. Quain et al. (1998) and Muller (1999) addressed managerial factors that may improve the efficiency of restaurants, such as realizing profit centers, dispersing demand, decreasing operating hours, and decreasing service time by making the restaurant operational procedures as efficient as possible. Most of the studies in this area address issues concerning the overall management of the restaurant. However, to the best of our knowledge there are no studies on a mathematically rigorous dynamic seating model. Our objective in the present paper is to develop and test, using both simulated and real data, several increasingly sophisticated optimization-based approaches to restaurant revenue management (RRM) that address trade-offs between expected revenue, average waiting time, and perceived fairness, that may violate the FCFS rule. Contributions Our contributions are: (1) Because an exact dynamic programming approach to RRM is infeasible due to “the curse of dimensionality,” we develop mathematical-programming-based approaches of increasing sophistication to maximize expected revenue, taking into account expected waiting time and associated / 473 fairness issues. We develop integer programming, stochastic programming, and approximate dynamic programming models to provide efficiently implementable policies for RRM. These approaches are interesting in their own right and may find application in revenue management in other contexts. A particularly promising method is the approach of approximate dynamic programming based on an integer programming model. (2) We propose a stochastic gradient algorithm motivated by the work of Karaesmen and van Ryzin (1998) to address RRM with reservations. We also apply approximate dynamic programming to make optimal seating decisions for both reservation and walk-in customers online. (3) In computational studies involving both simulated and real data, we show that our models improve revenue, often substantially, relative to the industry standard of seating parties in a FCFS manner, without increasing the average waiting time. Structure The structure of this paper is as follows: §2 describes the structure and components of the basic model. Section 2.1 describes the integer programming approach, while §2.2 outlines the simulation model. Section 3 elaborates on extensions to the basic model. Two variations of the model are introduced in this section: §3.1 describes the stochastic programming version of the basic model; §3.2 introduces an approximate dynamic programming approach for the basic integer programming model; §3.3 describes FCFS models and the bid-pricing model used as a performance baseline for the previous models. Section 4 introduces reservations to our previous models. Section 4.1 describes the reservation-booking model that determines the optimal booking level using a stochastic gradient algorithm, while §4.2 describes a modified version of the approximate dynamic programming model introduced earlier. In §§5 and 6 we report computational results for models without and with reservations, respectively. Section 7 summarizes our findings and contains some concluding remarks. 2. THE BASIC MODEL 2.1. An Integer Programming Approach In this section, we develop an integer programming (IP) model that aims to maximize the expected future revenue, while controlling for expected waiting time by deciding when and where to seat each incoming party. We use a discrete time horizon of N equal-length periods so that the number of decision variables and constraints remain tractable. We first introduce the data on which the model is based. Data. We consider a restaurant that can seat parties of size k = 2 4     K, with K even. For simplicity, parties of sizes 1 3     K − 1 can be considered to be one person larger. There are k′ = 2 4     K different table sizes such that a party of size k can sit at a table of size k′ , 474 / Bertsimas and Shioda for k′
k. We assume that the total revenue generated by a party increases with the party size. Note that we do not necessarily assume that the revenue per person is affected by party size. We model the service time as a collection of SP service phases (for example: first course, second course, dessert). By breaking up the service time and keeping track of how many parties are in each phase, we can make a better estimate of the remaining service time. Our model only uses information about expected values. In this respect, our model is similar to network airline revenue management systems that use a linear programming model (for example, the widely used bid-price control approach) to decide seat allocation. We assume the following information is known: ck
= the number of tables of size k available, Dt k
= the expected number of size k parties arriving at time, t, Sk s
= the expected duration of phase s for a size k party,
Sk
= SP s=1 Sk s , the expected total service duration of a size k party,
SP ′ Sk n=s Sk n , the expected remaining service duras
= tion for a size k party entering phase s, Rk
= the revenue expected from a party of size k, CostQt k
= the cost of postponing service to a party of size k that arrived at time t and is currently in queue, and CostX t k
= the cost of postponing service to a party of size k that is expected to arrive at time t. The above data can be collected by many POS softwares. For example, Dt k and Rk can be estimated using historical data on customer arrivals and bills, which is often kept track of by POS softwares. We have found there are several POS products that have the ability to track service phases of customers with an electronic ordering pad used by waiters when taking orders. Currently, these pads are used to send orders to the kitchen, but we believe they can also be used to notify the floor manager of the progress of the customers in their meals. With this added feature, it would be easy for the POS software to estimate average service-phase durations. Floor managers would need to record arrival and departure times of parties, but that can be readily linked to seating the parties and processing their bills, respectively. Finally, we set the values for CostQt k and CostXt k in our computational experiments (see §5) to capture the trade-off between quality of service (excessive waiting times and denied service) and revenue. In addition to revenue, important considerations in managing restaurants are the control of waiting time and associated fairness issues. Thus, in the IP model we explicitly address these issues. For this purpose, we introduce the following parameters that we adjust to achieve the “right” trade-off between revenue, waiting time, and fairness. Max
= the maximum number of periods that a party will wait, M
= a user-defined parameter that controls the tradeoff between revenue and waiting time (the higher M, the higher the importance of the waiting time; see Equation (1)), and 
= a user-defined parameter that controls the tradeoff between revenue and allowing flexibility in allocating a size k party to a table of size k′
k; see Equation (1). In practice, the waiting time behavior of customers is complex. The amount of time a party may wait in queue may follow a probabilistic distribution and may be dependent on the size of the party and time of arrival. However, to simplify the model, we assume that every party is willing to wait up to Max periods. If they are not seated within Max periods, the model assumes that they automatically renege from the queue.1 It further assumes that parties cannot renege earlier or later than Max periods. Appropriate values for Max, M, and  are discussed in §§5 and 6. State. The state of the system is described as follows. now
= the current time (in periods), Qt k
= the number of size k parties that arrived at time t currently waiting in queue, and Nks k′
= the number of size k parties in service phase s seated at a size k′ table. To better utilize capacity, the IP model allows parties of size k to be seated at a table of size k′
k. The service state parameters Nks k′ keep track of these parties. Decision Variables. We introduce the following decision variables: qt t′  k k′
= the number of parties of size k, who arrived at time t and are currently in the queue, that should be seated at a size k′ table at time t ′ , qdenyt k
= the number of parties of size k, who arrived at time t and are currently in the queue, that are not currently allocated a seat (i.e., not seated within Max periods), xt t′  k k′
= the number of parties of size k, out of Dt k , that should be seated at a size k′ table at time t ′ , xdenyt k
= the number of parties of size k, out of Dt k , that are not currently allocated a seat (i.e., not seated within Max periods), zqt k
= the auxiliary variables that allow us to model fairness in the seating decisions, captured in Equations (5) and (6) below, and zxk
= the auxiliary variables that allow fairness in the seating decisions, captured in Equations (7) and (8) below. Each time the IP model is formulated, there are two disjoint sets of customers that it must consider, those that have already arrived and those that are expected to arrive. qt t′  k k′ and qdenyt k are seating-decision variables corresponding to the former type of customers, and xt t′  k k′ and xdenyt k are the seating-decision variables for the latter type. The parties represented by the qt t′  k k′ s and qdenyt k s are parties that have already arrived and are thus known. However, the parties represented by xt t′  k k′ s and xdenyt k s have not yet arrived and are thus uncertain. Partitioning the decision variables in such a way allows for more dynamic and flexible modeling. The model does not explicitly deny service to the customers, but it may not be able to seat some parties within Max periods—the maximum number of periods that a party would be willing to wait in queue. These parties that are not currently allocated a table are captured by the decision Bertsimas and Shioda variables qdenyt k and xdenyt k . The use of the auxiliary variables zqt k and zxk will be described in more detail in the IP formulation. The Integer Program Formulation. We formulate and solve an integer program at each customer arrival and departure, and determine from the optimal solution whether a party should be seated. Our objective is to maximize expected revenue, while maintaining reasonable waiting times and fairness in our seating decisions, by considering the state of the restaurant. We can also maximize occupancy by assuming that the revenue per person is independent of the number in the party. Although we solve the IP to optimality, the data we use, such as expected demand, expected service time, and the simplified waitingtime behavior, are only estimations. Thus, the expected future state of the restaurant modeled by the IP are also estimations. However, capturing the states of the system exactly in a stochastic dynamic programming sense would be computationally intractable, thus making our IP model a viable heuristic. The proposed model is as follows. Objective Function. The objective is the maximization of expected future revenue while controlling waiting time:   Rk −Mt ′ −t−k′ −kqtt′ kk′ max t=1now k′ =kK k=2K t ′ =nowminN t+Max−1 −  CostQtk qdenytk  t=1now k=2K   Rk −Mt ′ −t−k′ −kxtt′ kk′ t=nowN k′ =kK k=2K t ′ =tminN t+Max−1 −  CostXtk xdenytk  (1) 475 arrive in the future. Also, CostXt k for t = now is higher than that for t > now. If it were the same, the model would consider postponing service to a party that arrives in the current period versus seating that party and postponing service to an equal-sized future party as the same. We prefer the former solution, since there is a possibility that a party will arrive in the current period, and thus we want the model to give us the ability to seat them. The exact numerical value of these weights are determined by computational experimentation. Constraints. (1) Demand Constraints.  qtt′ kk′ +qdenytk = Qtk t ′ =nowminN t+Max−1 k′ =kK ∀ t = 1 now k = 2 K  (2) xtt′ kk′ +xdenytk = Dtk t ′ =tminN t+Max−1 k′ =kK ∀ t = now N  k = 2 K (3) Constraints (2) insure that the model does not seat more parties than those currently in the queue. Constraints (3) insure that the model does not seat more parties than currently expected. (2) Seating-Capacity Constraints. This constraint characterizes the capacity constraint at each time  for each table size k′ . It also incorporates a nesting concept that allows parties of size k to sit at a size k′ table, where k  k′ :    qt t′  k k′ k=2 k′ t ′ ∈T t k t=max1 t ′ −Max +1 now + t=nowN k=2K The first set of summations corresponds to the expected revenue that can be attained from parties currently in the queue. The second set corresponds to the cost of not allocating a table to parties currently in the queue. The third set corresponds to the revenue from parties expected to arrive in the future, and the last is the cost of not allocating a table to those customers. The term −Mt ′ − t in the first and third coefficients is a cost against excessive waiting time. The cost −k′ − k encourages seating from the smallest available table first. Without this term, the model may, for example, seat a party of size two at a table of size eight if it does not expect larger-sized parties in the near future. However, because our expected demand data are only estimations and the actual demand is uncertain, we want to be on the safe side and seat from the smallest available table first. The costs CostQt k and CostXt k are necessary for the model to give preference to customers in the queue, as well as those expected to arrive in the current period, over those expected to arrive in the future. Thus, the value of CostQt k is higher than CostXt k for all t for each k, because postponing service to a customer already in queue should be higher than postponing service to a customer expected to /  xt t′  k k′ t = maxnow t ′ −Max +1 t ′ +  IN   ck′  k = 2 k′ s = 1 SP ∀  = now     N  k′ = 2     K (4) where T t k = t ′
t  t ′    t ′ + Sk  t ′ ∈
  s ′ Nk k′  if   now + Sk s, IN  = 0 otherwise. Note that the parameters Nks k′ allow for a more accurate estimation of the remaining service time of the parties currently being served. The last summation term in Equation (4) corresponds to the expected number of parties currently seated at table k′ who are still being served at time . (3) Fairness Constraints. The model uses parameter M to control waiting time in the objective function. However, when this cost is added, the model would rather seat lastcome-first-serve within the same party size because the last customer to arrive would have less waiting time, and thus a higher objective value. To avoid this problem, we add constraints in the model that would seat the customers within the same party size in the order that they arrived. By seat- 476 / Bertsimas and Shioda ing parties in queue by FCFS within the same party size, the waiting time of the parties decreases and fewer parties will leave from the queue. In particular, because of perceived unfairness, customers in a restaurant would strongly complain if a same-sized party that arrived after them were seated before them. The following constraint insures a FCFS seating policy within the same party size for those in the queue that are able to be seated at the current time. It is not extended to future periods due to the uncertainty of the future state of the queue.    Qt̃k −qt̃nowk k̃ t̃=max1now−Max+1t−1 k̃=kK   Qt̃k zqtk   (5) t̃=max1now−Max+1t−1  qtnowk k̃  L1−zqtk  k̃=kK for t = max2now−Max+2 now k = 2 K  t̃=max1now−Max+1now k̃=kK    Qt̃k −qt̃nowk k̃  (6)  Qt̃k zxk  (7) t̃=max1now−Max+1t−1  xnownowkk̃  L1−zxk  for k = 2 K. (8) k̃=kK Here zqt k , zxk are auxiliary binary decision variables associated with each set of constraints. L is some large positive constant. Equations (5) and (6) state that if there are parties of size k that arrived before time t still in queue (i.e., the LHS of (5) is positive), then zqt k has to be equal to 1. This implies that the LHS of (6) must be zero. In other words, Equations (5) and (6) insure that a party of size k that arrived at time t can be seated only when there are no parties of k that arrived before them still in queue. Constraints (7) and (8) are analogous to parties of size k expected to arrive at the current period. (4) Integrality Constraints. For t = 1     N  t ′ = t     N  k = 2     K k′ = k     K; qt t′  k k′ ∈
+  zqt k ∈ 0 1 xt t′  k k′ ∈
+  zxk ∈ 0 1 (9) 2.2. The Simulation The IP described in Equations (1)–(9) will be formulated and solved whenever the floor manager needs to make a seating decision. When incorporated into a POS product, the system would optimize the IP model with the current system data each time a new customer arrives or when a customer exits. If the optimal solution of the model indicates that a party should be seated or not seated at that moment, the result is conveyed to the floor manager, who can act accordingly. The implementation of the model is simple. There is little or no added work for the floor manager who uses most POS products. He or she needs only to input the size of a newly arrived party and take note when a party exits. All the information needed to run the model is already being collected in many large restaurants. To summarize, there are three events that drive the model: (1) customer arrivals, (2) service completion, and (3) customer attrition from the queue. The model works in the following way after each of these events: (1) Customer Arrives: • Input the party size and time of arrival. • Update appropriate Dt k . • Formulate and solve the IP model using current system data. • Act according to the IP solution and update system data, i.e., — If the party is to be seated, seat them at the specified table. Update Nks k′ accordingly. — If the party is not to be seated, put them at the end of the line. Update Qt k accordingly. (2) Customer Exits After Service Completion: • Exit the party from the table. • Update appropriate Nks k′ . • Formulate and solve the IP model using current system data. • Act according to the IP output and update system data, i.e., — If a party currently in the queue is to be seated, take them off the queue and seat them at the specified table. Update Nks k′ and Qt k accordingly. (3) Customer Attrition from the Queue • Update appropriate Qt k . We update Nks k′ and Qt k by incrementing or decrementing the appropriate term. We do the same for Dt k , the expected future demand, instead of using any probabilistic approach to forecasting demand. The initial value of Dt k is set to the expected demand for each time period t and party size k calculated by a simple average of the historical data. As arrivals are seen for a given t and k, Dt k is decremented. If Dt k is 0 when an unexpected customer arrives, the simulation model automatically puts them in the queue, the queue state is updated, and the IP is optimized. Although the party is theoretically put in the queue from the model’s perspective, they often do not have to wait to be seated. Though this is a simplified estimate of future demand levels, it seems sufficient from our computational experimentations. 3. EXTENSIONS OF THE BASIC MODEL In this section, we present several increasingly more sophisticated extensions of the basic model. Bertsimas and Shioda 3.1. A Stochastic Integer Programming Model One of the shortcomings of the basic model is its sole use of the expected demand values as an indication of future customer arrivals. The stochastic version of the IP introduces various demand scenarios that the IP model can work with. Several demand scenarios are generated similarly to the process of generating arrivals in simulations. By using the generated scenarios with the expected demand values, the IP attempts to more accurately capture the future demand characteristics. The stochastic IP model works with demand scenarios (one of them being the expected case), each with probability !" , where " = 1 2     . The objective function is modified to maximize the expected revenue over the possible scenarios. Additional and Modified Data and Parameters. To summarize the new parameters and data required:
= the total number of scenarios, !"
= the probability of scenario ", and Dt" k
= the number of currently expected size k parties arriving at time t in scenario ". Modified Decision Variables. To summarize the modified decision variables: xt t′  k k′  "
= the number of size k parties, out of Dt" k , that should be seated at a table of size k′ at time t ′ in scenario ", xdenyt k "
= the number of size k parties, out of Dt" k , that are not allocated a table in scenario ", and zxk "
= the auxiliary variables that allow fairness in the seating decision, captured in Equations (14) and (15). Stochastic Integer Programming Formulation.  max  ′ Rk −Mt −t−k −kqtt′ kk′ t=1now k′ =kK k=2K t ′ =nowminN t+Max−1 −  CostQtk qdenytk  t=1now k=2K   !" Rk −Mt ′ −t−k′ −kxtt′ kk′ " t=nowN k′ =kK k=2K t ′ =tminN t+Max−1 "=1 −  !" CostXtk xdenytk"  (10) t=nowN k=2K "=1 Constraints. (1) Demand Constraints. The first set of constraints insures that the IP model does not seat more customers than those in the queue. The second set insures that the IP model does not seat more parties than the demand for each scenario:  qt t′  k k′ + qdenyt k = Qt k t ′ =now minN  t+Max −1 k′ =k K ∀ t = 1     now k = 2     K ∀ t = now     N  k = 2     K " = 1      (12) (2) Seating-Capacity Constraint.    qt t′  k k′ k=2 k′ t ′ ∈T t k t=max1 t ′ −Max +1 now  + (11) xt t′  k k′  " t=maxnow t ′ −Max +1 t ′ +  IN   ck′ k=2 k′ s=1 SP  ∀  = now     N  k′ = 2     K " = 1      (13) where T t k′  and IN  are as in (4). (3) Fairness Constraint. Same as (5) and (6). In addition,    Qt̃ k − qt̃ now k k̃ t̃=max1 now−Max +1 now k̃=l K    Qt̃ k zxk "  (14) t̃=max1 now−Max +1 t−1  xnow now k k̃  L1 − zxk "  k̃=k K for k = 2     K, " = 1     . (15) (4) Integrality Constraints. For t = 1     N  t ′ = t     N  k = 2     K k′ = k     K " = 1     zqt k ∈ 0 1 ′ 477 t ′ =t minN  t+Max −1 k′ =k K qt t′  k k′ ∈
+  Objective Function. / xt t′  k k′  " + xdenyt k " = Dt" k  xt t′  k k′  " ∈
+  zxk " ∈ 0 1 (16) The Simulation for the Stochastic Model. The simulation using the stochastic IP model is similar to that of the basic model. The model follows the procedure described in §2.2 using the stochastic programming model instead of the deterministic IP model. 3.2. An Approximate Dynamic Programming Model The approximate dynamic programming model solves for the maximum revenue-producing seating policy for each customer. We formulate and solve the IP under each possible seating decision for a particular customer, and the decision that results in the maximum revenue value is chosen. Let S be a vector describing the current states of the system that can be affected by a seating decision. In particular, S = Qt k  Nks k′  for t = max1 now − Max +1     now k = 2     K k′ = k     K s = 1     SP  (17) 478 / Bertsimas and Shioda A decision time is when a party arrives or a party exits after service completion. If an arrival of size k occurs, the following decisions are available: (1) Do not seat the incoming party. The new state is such that: Qt k ← Qt k + 1. (2) For k′ = k     K, seat the incoming party at a table of size k′ . The new state is such that: Nk1 k′ ← Nk1 k′ + 1. If a party of size k occupying a table of size k′ exits from the restaurant, the following decisions are available: (1) Do not assign the table to a party from the queue. The new state is such that: NkSPk′ ← NkSPk′ − 1. (2) Assign the table to a party of size k′′  k′ , who has been in the queue since time t ′′ . The new state is such that: Qt′′  k′′ ← Qt′′  k′′ − 1, Nk1′′  k′ ← Nk1′′  k′ + 1, NkSPk′ ← NkSPk′ − 1. Let u be the decision taken, so that  0 don’t seat now, u= (18) k′  seat at a table of size k′ , for each of the events. Let Su be the updated state after a decision u is taken. Let IP S be the expected future revenue resulting from the IP model (Equations (1)–(9)) as a function of the state S. Under an approximate dynamic programming (ADP) framework, we choose the decision u, corresponding to seating a size k party that arrived at time t, by solving max max %Rk − u − k + IP Su & IP S0   u=k K table of size u available (19) The first term in the maximization corresponds to the expected revenue of seating the party minus a nesting cost, and the second term corresponds to the expected revenue of not seating the party. This method uses the ADP methodology by approximating the true value function (in a dynamic programming sense) by the value of the IP model. We note that the IP in (19) need not include the fairness constraints (5)–(8) required in the previous models because ADP evaluates the value function of each party in the order of their arrival and thus ensures FCFS seating within the same party size. 3.3. Comparison Models First-Come-First-Serve. We develop three FCFS models to compare their revenue with the revenue generated by using the above models. The FCFS model seats the customers in the order of arrival if there are tables available for them. The “Full Nesting” FCFS model incorporates full nesting of capacity, where a party of size k is allowed to sit at a size k′ table for k′
k. If there are several possible table sizes, the model would seat them at the smallest of those tables. The “1 Up Nesting” FCFS model allows a party of size k to sit at either a table of size k or at the next largest table if size k tables are saturated. The “No Nesting” FCFS model allows no nesting. If there are no tables for a particular customer of size k, but there are tables for smaller customers that arrived after them, the model will seat those customers in order of arrival. Customers who have waited longer than Max will automatically leave the queue. The average of the simulated revenues over several iterations are compared to the revenue of the IPs. Bid-Pricing Model. The bid-pricing heuristic commonly applied in airline revenue management is run as a performance benchmark. The linear programming relaxation of the basic IP model is solved and the seating decisions are made based on the difference between the immediate revenue and the sum of the dual prices corresponding to the utilized capacity corresponding to the party’s stay. Suppose pt k′ is the dual value of the seating-capacity constraint corresponding to a table of size k′ at time t. The bid price for seating a party of size k to a table of size k′
k is as follows: exit−1 Rk −  pt k′ −  now − nowpnow k′ t= now − exit − exitpexit k′  (20) where exit = now + Sk . The term  now − nowpnow k′ and exit − exitpexit k′ are to prevent overestimation of the expected stay of the customer due to the discreteness of time.  now − now is the fraction of period now andexit − exit is the fraction of period exit that the party is expected to be in service. If there are positive bid prices, the party is seated at the table corresponding to the maximum price. Otherwise, the party is not seated. 4. MODELS WITH RESERVATIONS The models in the previous two sections assume that the restaurant does not accept any reservations. Many highend restaurants, however, do indeed accept reservations for their customers’ convenience. Such a policy introduces several decisions which need to be addressed. For example, how much should they overbook and how should they service walk-in customers? We will incorporate reservations into our previous models by using two models: (1) a static reservation-booking (RB) model, and (2) a dynamic seat allocation model with reservations (DSAR). The static reservation-booking model is an offline model that is optimized using a stochastic gradient approach as in Karaesmen and van Ryzin (1998). The dynamic seat allocation with the reservation model is an online model solved using an ADP approach as in §3.2. 4.1. A Reservation-Booking Model In this section, we develop a model that determines how many reservations to accept in prior days for a particular day in the future, given data regarding the expected number of reservation requests, the expected number of walk-ins, and the rates for no-shows. As before, we assume that there are N time periods in which reservations can be accepted. We do not, however, Bertsimas and Shioda consider nesting of table sizes; i.e., a party of size k can only be assigned to a table of size k. Thus, the seating decisions are independent across the party sizes so that the problem can be broken down into smaller subproblems for each party size. Because we are building a planning, as opposed to operational, model in this section, we feel that this is justified. Methodologically, motivated by the work of Karaesmen and van Ryzin (1998) in a very different context, we solve the optimization problem using a steepestdescent algorithm with stochastic gradient estimation. Data. The following are the data that we expect the restaurant to have for each party size, R
= the expected revenue of the given party size, p
= the probability that a reservation party will show up, ERt
= the expected number of reservation requests for time t, and EWt
= the expected number of walk-in parties that arrive at time t. Decision Variables. The key decision is the number of reservations ut to accept for time period t; t = 1     N . Random Variables. We define zt
= the number of reservation parties booked for time t that show up, and wt
= the number of walk-ins that show up at time t. We assume that the probability of no-shows is independent and identically distributed across parties of the same size, as shown empirically by Martinez and Sanchez (1970). We further assume that zt obeys a binomial distribution with parameters (ut , p). We approximate this distribution for time t as a nonhomogeneous Poisson distribution of rate put for ease in numerical computation. We also model wt as a nonhomogeneous Poisson with rate EWt for time t. Objective Function. We maximize the following objective function RBu, where u, ER, z, and w are N dimensional vectors of elements ut , ERt , zt , and wt , respectively, t = 1     N : max RBu u
0 (21) such that 1 RBu = E %V z w& − - E u − ER2  2 (22) The first term in Equation (22) corresponds to the expected maximum revenue resulting from having z reservation customers and w walk-in customers arrive on the requested date. We will refer to the term V z w as the the optimal OTD (on-the-day) revenue. This value is calculated by solving a simplified version of the previously discussed seating models of §§2.1–3.2 generalized for reservations. We describe this term in further detail in the following section. The second term is a regularizing term that discourages the decision variables from straying too far from / 479 the expected reservation requests. Without this term, the model may allocate no seats for the 6 o’clock time period and many for the 10 o’clock period, though the expected requests are mostly for 6 o’clock and hardly any are for 10 o’clock. Thus, a policy that allocates significantly more (less) seats for a particular period with few (many) reservation requests overestimates (underestimates) the possible expected revenue. The parameter - > 0 represents the tradeoff between the two terms. Static On-the-Day Model. We calculate the value of V z w for a given vector z and w by solving an integer program that optimally allocates the tables to reservation and walk-in customers. The OTD model is similar to the previous models with an additional class of reservation parties. However, it does not consider nesting of party sizes and assumes that there are no queues. By making each period large enough (e.g., half an hour), it is reasonable to assume that after one period, both walk-in and reservation parties will renege. Also, most reservation requests are on the half-hour. Thus, this simplification will not significantly jeopardize the accuracy of the model. We will see that these assumptions make the problem a single-commodity maximum flow problem which aids us in the convergence analysis of the stochastic gradient algorithm. The OTD model is also a static model that is not updated dynamically because it is solved before the day in question. It decides offline how many of the reservation and walk-in parties to seat and how many to deny service for each time period, t = 1     N . We need to introduce some additional and modified notation. Data. S
= the expected service time for the given party size, CostW
= the cost of denying service to a walk-in party, CostR
= the cost of denying service to a party with a reservation, and C
= the total capacity (number of tables) for the given party size. We assume that the expected revenue and service-time distributions are the same for walk-in and reservation customers of the same party size. CostW and CostR are qualitative values of the loss of goodwill and customer dissatisfaction. CostR should be large relative to R, because not being able to serve a customer with a reservation would most likely entail losing that customer and garnering a negative reputation. However, CostW can be negligible because walk-in customers’ expectation of being served is significantly less. Decision Variables. xwt
= the number of walk-in parties that are seated at time t, xrt
= the number of reservation parties booked for time t that are seated, xwdenyt
= the number of walk-in parties that arrive at time t that are denied service, and xrdenyt
= the number of reservation parties booked for time t that are denied service. 480 / Bertsimas and Shioda Problem Formulation. Given a particular vector z and w, the OTD problem is as follows: (OTD)  R · xwt + xrt  − CostW  · xwdenyt t=1 N − CostR · xrdenyt (23) subject to xwt + xwdenyt = wt  t = 1     N  (24) xrt + xrdenyt = zt  t = 1     N   xw + xr   C t = 1     N  (25) (26) =max1t−S+1 t xrt
0 t = 1     N  t = 1     N  (27) (28) The objective value is the net revenue of seating xwt walk-ins and xrt reservation parties and denying service to xwdenyt walk-ins, and xrdenyt reservation parties for t = 1     N . Equations (24) and (25) are the demand constraints for walk-ins and reservations, respectively, for the particular realizations of demand zt and wt for t = 1     N . Equation (26) represents the capacity constraint for each time period, t = 1     N . Equations (23)–(28) can be characterized as a singlecommodity minimum cost flow problem; thus, the value of V z w can be calculated by solving the OTD problem as a linear program. It follows by the network structure of OTD and Karaesmen and van Ryzin (1998) that the function V z w is submodular and jointly concave with respect to zt and wt , t = 1     N , and RBu is componentwise concave, continuous, and differentiable with respect to ut , t = 1     N . As shown in Karaesmen and van Ryzin (1998), these properties allow convergence of the stochastic gradient algorithm described in the next section. Stochastic Gradient Algorithm. We solve the stochastic optimization problem (21) by estimating the stochastic gradient of RBu as done in Karaesmen and van Ryzin (1998). We apply this gradient to maximize RBu using a method similar to the steepest descent algorithm. We first calculate 0 RBu 0ut =   k=1 V z w = max xwt
0 The step sizes bk  satisfy 0 E%V z w& − -ut − ERt  0ut t = 1     N  (29) When zt ∼ Poissonput  and wt ∼ PoissonEWt , t = 1     N , Karaesmen and van Ryzin (1998) show that an unbiased estimate of 0/0ut E%V z w& is given by 0 E%V z w& 0ut = pV z + et  w − V z w t = 1     N  (30) bk = +   bk2 <  (31) k=1 Let MaxIt be the maximum number of iterations we allow. The following is the stochastic gradient algorithm. Step 0. Initialization. k = 1, ukt = ERt for t = 1     N . Step 1. Generate 4RBuk : • Generate new vectors zkt ∼ Poissonpukt  and wt ∼ PoissonEWt  for t = 1     N . • Evaluate V zk  w using a network flow algorithm (for example, the network simplex algorithm). • Evaluate V zk + et  w for t = 1     N . • Derive 4RBuk  such that, for t = 1     N : 0 RBuk  0ut = pV z + et  w−V z w−-ut − ERt  Step 2. Compute uk+1 = Projectuk + bk 4RBuk  where Project· is the projection function to the feasible space uk
0. Step 3. If k > MaxIt, Exit. Else, set k
= k + 1 and return to Step 1. Theorem 1 (Kushner, Clark, Karaesmen, and van Ryzin). Let KT be the set of Kuhn-Tucker points of (21). If KT is a connected set and uk  is the sequence determined by the previous algorithm with step sizes satisfying (31), then uk  → KT in probability as k → . The proof is given in Kushner and Clark (1978) and Karaesmen and van Ryzin (1998), Theorem 3 and Theorem 6.3.1, respectively. Kusher and Clark also show that a weaker convergence follows when KT is not a connected set. 4.2. Dynamic Seat Allocation with Reservations The DSAR model is the same as the previous dynamic seating models of §§2.1–3.2, but with reservation customers taken into account. It takes the reservations accepted as input and determines the optimal seating policy for each arriving party given the current state of service and queue. It does not, however, update the maximum reservationbooking number produced by the RB model. As before, each time the state changes (i.e., due to a customer arrival or customer exit), the state parameters are updated and the model is continually reoptimized online. We apply the the ADP approach as in §3.2, with a slight modification in the IP formulation. Bertsimas and Shioda The following are additional and modified notations for the DSAR model: Input. vt k
= the number of reservations booked of size k for time t. vt k is such that vt k  u∗t k , where u∗t k is the optimal reservation booking value outputted from the RB model. Data. CostQWt k
= the cost of postponing service to a walkin party of size k that arrived at time t and currently waiting in queue, CostQRt k
= the cost of postponing service to a party of size k with reservations that arrived at time t and is currently waiting in queue, CostXWt k
= the cost of postponing service to a walkin party of size k that is expected to arrive at time t, CostXRt k
= the cost of postponing service to a party of size k with reservations that is booked for time t, pk
= the probability that a party of size k with reservations will show up, MaxW
= the maximum number of periods that a walkin party will wait, MaxR
= the maximum number of periods that a reservation party will wait, Mw
= a user-defined parameter that controls the trade-off between revenue and waiting time for walk-in customers, and Mr
= a user-defined parameter that controls the tradeoff between revenue and waiting time for reservation customers. The quantities CostQWt k , CostQRt k , CostXWt k , and CostXRt k are analogous to CostQt k and CostXt k in §2.1 which are split into walk-ins and reservation customers (denoted by W and R, respectively). Similarly, MaxW and MaxR are analogous to Max, and Mw and Mr are analogous to M in §2.1. State. The following are modified state parameters in addition to those described in §2.1. Qwt k
= the number of size k walk-in parties currently in queue that arrived at time t, Qrt k
= the number of size k reservation parties currently in queue that arrived at time t, Dwt k
= the expected number of walk-in parties of size k that are going to arrive in time t who have not already arrived, and Drt k
= the expected number of reservation parties of size k that are going to arrive in time t who have not already arrived. When now < t, Dwt k = EWt k and Drt k = pk vt k . Each time a walk-in or reservation party of size k arrives at time t, Dwt k or Drt k is decremented, respectively. Decision Variables. qwt t′  k k′
= the number of walk-in parties of size k that arrived at time t and are currently in queue, that should be seated at a size k′ table at time t, / 481 qrt t′  k k′
= the number of reservation parties of size k that arrived at time t and are currently in queue, that should be seated at a size k′ table at time t, qwdenyt k
= the number of walk-in parties of size k that arrived at time t and are currently in queue, that are not allocated a table (i.e., are not seated within MaxW periods), qrdenyt k
= the number of reservation parties of size k that arrived at time t and are currently in queue, that are not allocated a table (i.e., are not seated within MaxR periods), xwt t′  k k′
= the number of walk-in parties of size k out of Dwt k , that should be seated at a size k′ table at time t ′ , xrt t′  k k′
= the number of reservation parties of size k out of Drt k , that should be seated at a size k′ table at time t ′ , xwdenyt k
= the number of walk-in parties of size k out of Dwt k , that are not allocated a table (i.e., are not seated within MaxW periods), and xrdenyt k
= the number of reservation parties of size k out of Drt k , that are not allocated a table (i.e., are not seated within MaxR periods). IP Formulation. The DSAR uses a slightly modified version of the ADP model described in §3.2 due to its superior performance compared to the other models. The following describes these modifications: Objective.  max Rk −Mw t ′ −t−k′ −kqwtt′ kk′ t=1now k=2K t ′ =nowminN t+MaxW −1 k′ =kK Rk −Mr t ′ −t−k′ −kqrtt′ kk′  + t=1now k=2K t ′ =nowminN t+MaxR−1 k′ =kK +  Rk −Mw t ′ −t−k′ −kxwtt′ kk′  Rk −Mr t ′ −t−k′ −kxrtt′ kk′ t=nowN k=2K t ′ =tminN t+MaxW −1 k′ =kK + t=nowN k=2K t ′ =tminN t+MaxR−1 k′ =kK −    CostQWtk qwdenytk +CostQRtk qrdenytk t=1now k=2K −  t=nowN k=2K   CostXWtk xwdenytk +CostXRtk xrdenytk  (32) The first four sets of summations correspond to the expected revenue, waiting-time cost, and nesting costs of seating (1) the walk-in parties currently in queue, (2) the reservation parties currently in queue, (3) the walk-in parties expected to arrive in the future, and (4) the reservation parties expected to arrive in the future, respectively. The last two sets of summations correspond to the cost of postponing service to (1) walk-in parties and reservation parties currently in queue, and (2) walk-in parties and reservation parties expected to arrive in the future, respectively. 482 / Bertsimas and Shioda Constraints. (1) Demand Constraints. The following constraints are analogous to Constraints (2) and (3) in §2.1:  qwt t′  k k′ + qwdenyt k = Qwt k t ′ =t   minN  t+MaxW −1 k′ =k K ∀ t = 1     now k = 2     K  qrt t′  k k′ + qrdenyt k = Qrt k  xwt t′  k k′ + xwdenyt k = Dwt k  xrt t′  k k′ + xrdenyt k = Drt k (33) t ′ =t minN  t+MaxR −1 k′ =k K ∀ t = 1     now k = 2     K (34) t ′ =t minN  t+MaxW −1 k′ =k K ∀ t = now     N  k = 2     K (35) t ′ =t minN  t+MaxR −1 k′ =k K ∀ t = now     N  k = 2     K (36) (2) Seating-Capacity Constraints. The following constraint is analogous to the (4) in §2.1 with reservation parties:    qwtt′ kk′ k=2k′ t ′ ∈T tk t=max1t ′ −MaxW +1now  + qrtt′ kk′ t=max1t ′ −MaxR+1now  + xwtt′ kk′ t=maxnowt ′ −MaxW +1vt ′ +  xrtt′ kk′ t=maxnowt ′ −MaxR+1t ′ +  IN   ck′ ∀  = now N k′ = 2 K (37) where T t k′  and IN  are as in (4). (3) Integrality Constraints. For t = 1     N  t ′ = t     N  k = 2     K k′ = k     K qrt t′  k k′  xwt t′  k k′  xrt t′  k k′  qwdenyt k  qrdenyt k  xwdenyt k  xrdenyt k ∈
+  Incorporating the ADP Model. In the DSAR model, the state vector S is characterized by S = Qwtw  k  Qrtr  k  Nks k′  for (38) tw = max1 now − MaxW +1     now tr = max1 now − MaxR +1     now k = 2     K k′ = k     K s = 1     SP  Nks k′ • Reservation arrival at time t of size k: (1) Do not seat the incoming reservation party. The new state is such that: Qrt k ← Qrt k + 1. (2) For k′ = k     K, seat the incoming party at a table of size k′ . The new state is such that: Nk1 k′ ← Nk1 k′ + 1. • Walk-in arrival at time t of size k: (1) Do not seat the incoming walk-in party. The new state is such that: Qwt k ← Qwt k + 1. (2) For k′ = k     K, seat the incoming party at a table of size k′ . The new state is such that: Nk1 k′ ← Nk1 k′ + 1. • Party of size k exits table for k′ at time t: (1) Do not assign the table to a party from the queue. The new state is such that: NkSPk′ ← NkSPk′ − 1. (2) Assign the table to a reservation party of size k′′  k′ , who has been in the queue since time t ′′ . The new state is such that: Qrt′′  k′′ ← Qrt′′  k′′ − 1, Nk1′′  k′ ← Nk1′′  k′ + 1, NkSPk′ ← NkSPk′ − 1. (3) Assign the table to a walk-in party of size k′′  k′ , who has been in the queue since time t ′′ . The new state is such that: Qwt′′  k′′ ← Qwt′′  k′′ − 1, Nk1′′  k′ ← Nk1′′  k′ + 1, NkSPk′ ← NkSPk′ − 1. Finally, the DSAR model uses the objective value of the IP described by Equations (32)–(38) for IPS in (19).  k=2k′ s=1SP qwt t′  k k′  Thus, the decision times are when there is a reservation party arrival, walk-in arrival, and a customer service completion. The decisions available at each event are illustrated below: remains the same as The service state parameter in §3.2. There are two types of arrivals: reservation customer arrivals and walk-in arrivals. There is, however, no distinction between reservation and walk-in service completions. Comparison Model First-Come-First-Serve Model with Reservations. We incorporate reservation customers into the FCFS model in §3.3, which uses a heuristic to seat walk-in parties within reservations by taking the optimal reservation-booking data output from the RB model of §4.1. When a walk-in party arrives, the model will check whether seating them would take away a table from any outstanding reservations booked within a 15-minute (1 period) interval of the current time. The three different nesting models (Full Nesting, 1 UP Nesting, and No Nesting), as described in §3.3, are tested. 5. COMPUTATIONAL RESULTS FOR MODELS WITHOUT RESERVATIONS In this section, we present the performance of the nonreservation models of §2 and 3 on simulated data. 5.1. Data The test data for the capacity, service time, demand, and revenue are taken from a contrived restaurant. The data is constructed from its dinner-time operation, which runs from 6 pm to 10 pm. We divide this time into 16 equal periods of 15-minute durations. Bertsimas and Shioda Table 1. Expected duration (in minutes) of service phases. Party Size Phase 2 4 6 8 1 2 3 Total 6 39 6 51 9 45 6 60 12 57 8 77 18 75 9 102 Capacity Data. This is a small-scale restaurant with four tables for two, two tables for four, one table for six, and two tables for eight. The restaurant allows nesting of the capacity, i.e., a party of two can be seated at a table of two, four, six, or eight. Service Time Data. The meal duration is split up into three phases: appetizer (phase 1), entrée (phase 2), and coffee and dessert (phase 3).2 The expected service durations are illustrated in Table 1. For this example, we have in mind a restaurant with faster than usual turnover time. Demand Data. The restaurant does not accept reservations. We have tested the various methods for three levels of demand; low, medium, and high with average load factors 0.68, 0.93, and 1.54.3 Each demand level is split up into two demand distributions: constant, where all parties arrive uniformly throughout the day, and varied, where the larger parties arrive mainly in the later part of the evening.4 We simplify the data so that there are only four possible types of customers: parties of sizes two, four, six, and eight. We simulate the customer arrivals as a nonhomogeneous Poisson process with rate 6t k
= expected demand of size k customers at time t. Revenue Data. The expected revenue for parties of sizes two, four, six, and eight are $50, $120, $210, and $320, respectively. We simplify the revenue function so that it is time invariant. 5.2. Algorithms Tested and Parameter Settings Using the above data, we tested the following algorithms: the FCFS models (FullNest, 1Up, and NoNest), the bidpricing model (BidP), the basic integer programming model (IP), the stochastic programming model using three scenarios (STOCH), and the approximate dynamic programming model (ADP). We used CPLEX to solve the optimization models. The models were run on Dell Pentium II workstation operating under LINUX. Parameter Settings. For the STOCH model, we used a three-scenario model (i.e., = 3) in which one of the scenarios is the expected demand and the other two are randomly generated. We assign the probability of 0.5 to the expected demand scenario and the probability of 0.25 to each of the generated scenarios. We tested models with larger and with variations in the scenario probability, but / 483 the significant increase in the computation time was not worth the small increase in the average revenue. The value of M is set to 5 for all of the models. After running the models for different values of M, M = 5 produced the highest revenue, on average, for most of the models. Setting this value too high makes the model averse to seating customers that have been waiting in the queue for a long time. When M is set to zero, the model would give no consideration to waiting time when solving for the optimal solution. For values of M between 0 and 5, the resulting average revenue actually increases because the model is forced to seat a customer at a table which it was reserving for a future customer. This may be beneficial because the model sometimes incorrectly forecasts future demand and keeps tables idle when they could have been used. The appropriate values for  seem to vary according to the model and the demand level. For the IP and STOCH model we set  = 4, and for ADP we set  = 025. These values were chosen from empirical testing, thus they may not be the optimal. However, the differences in revenue and average waiting time were not significant for slight changes in . The value for Max is set to 3 for all models; thus we are assuming customers do not wait for more than 45 minutes. The models AUTOMATICALLY EXIT customers who have been waiting for more than three periods. For the IP and STOCH model, the values of CostQt k are 2.5 for k = 2, 6.0 for k = 4, 10.5 for k = 6, and 16.0 for k = 8, ∀t. CostXt k when t = now are 0.5 for k = 2, 1.2 for k = 4, 2.1 for k = 6, and 3.2 for k = 8, and when t > now, CostXt k is set to 0 for all k. The ADP model performed better with significantly higher values of CostQt k and CostXt k than IP and STOCH. Thus, these costs for ADP were set to 10 times that of IP and STOCH. Appropriate values for these parameters are clearly restaurant dependent. They would depend on the demand level and characteristics, expected revenue, expected service time, capacity, and waiting-time behavior of the customers of each restaurant. We suggest testing for the right values for these parameters by simulation, using the data and characteristics of each restaurant. 5.3. Results Table 2 contains the average daily revenue, average waiting time per customer, and the percentage of customers served for all the algorithms we tested. Each subtable corresponds to a demand scenario, low and constant, low and varied, medium and constant, medium and varied, high and constant, and high and varied. Table 3 contains the average run times per party for each model. For smaller load factors, Full Nesting is better than the other FCFS models and performs similarly to the IP, in general. This implies that the phenomenon that the IP captures to increase revenue for small load factors is nesting. The No Nesting model loses revenue by unnecessarily saving large tables for large parties. As demand level increases, 484 / Bertsimas and Shioda Table 2. Revenue, average wait, and percent served resulting from static capacity models. Load = 068 Constant Revenue ($) % Difference Average Wait (min) % Served Load = 068 Varied Revenue ($) % Difference Average Wait (min) % Served Load = 093 Constant Revenue ($) % Difference Average Wait (min) % Served Load = 093 Varied Revenue ($) % Difference Average Wait (min) % Served Load = 154 Constant Revenue ($) % Difference Average Wait (min) % Served Load = 154 Revenue ($) % Difference Average Wait (min) % Served FullNest 1Up NoNest BidP IP STOCH ADP 235194 000% 50 9323% 239032 163% 53 9308% 212546 −963% 107 8325% 232810 −101% 60 9089% 235446 011% 57 9181% 237592 098% 57 9194% 240156 211% 49 9367% 206654 000% 77 8806% 203118 −171% 75 8811% 188390 −884% 131 7852% 199786 −332% 60 9089% 208714 100% 83 8651% 207730 052% 78 8735% 211234 222% 75 8858% 296462 000% 99 8782% 292898 −120% 98 8769% 270236 −885% 144 7830% 282212 −481% 113 8335% 297684 041% 111 8548% 296950 016% 104 8626% 303418 235% 99 8784% 262494 000% 108 8316% 259636 −109% 104 8396% 239100 −891% 159 7535% 243018 −742% 126 7822% 261006 −057% 108 8271% 265106 100% 113 8208% 270271 296% 105 8331% 340982 000% 123 6884% 346766 170% 242 6870% 34988 261% 259 6036% 320310 −606% 237 5785% 367674 783% 237 6565% 370876 877% 240 6402% 385736 1313% 229 6896% 314676 000% 240 6783% 315498 026% 237 6850% 305106 −304% 275 5716% 290252 −776% 243 5677% 339336 784% 240 6425% 338742 765% 237 6433% 354540 1267% 234 6656% the nesting decisions become more complex. This explains the decrease in the revenue gap across the FCFS models with higher demand. We observe that increasing the sophistication of the models results in monotonically increasing revenue. There is a marginal revenue improvement of using stochastic programming versus the deterministic IP. The ADP model outperforms the deterministic and stochastic model in all demand scenarios, with about 2% improvement from FCFS in low and medium load factors and 12% improvement in high load factors in revenue. We also observe that the models do not sacrifice waiting time for higher revenue. The optimization models had comparable waiting times to the FCFS models, and in most of the larger-load cases had lower waiting time than the FCFS. Thus, higher revenue was achieved without any sacrifices and some improvements in the average waiting time. When examining the waiting time for each party size separately, we see that the optimization models have significantly lower waiting times for parties of sizes six and eight, while slightly higher for parties of two compared to the Table 3. Run time per party in seconds for static capacity models. FCFS BidP Average Run Time (sec) 0.00 0.67 IP 0.21 STOCH ADP 0.66 1.07 FCFS models. Thus, our models have a smaller range of waiting times across party sizes. In addition, IP and STOCH have a lower percentage of parties served than the best performing FCFS models, while producing higher revenue. Overall, the ADP model seems to be the best performing method: It serves about the same percentage of parties as the FCFS models, does not increase and occasionally decreases waiting time, and produces significantly higher revenue. This implies that the optimization models are able to seat more of the “right” (higher-revenue-producing) customers. The run times are also in a practical range for all of the models. Thus, any of these models can be run online with a POS system. 6. COMPUTATIONAL RESULTS FOR RESERVATION MODELS In this section, we report computational results for the models with reservations. We first run the static RB model to decide a priori how many reservations to accept, and we then run the DSAR model. 6.1. Data The test data was taken from Soto’s, a Japanese restaurant in Atlanta, Georgia (Kosugi 1999). Similar to the previous data set, the data are taken from its dinner-time operations Bertsimas and Shioda Table 4. 485 Revenue, percent served, and average waiting time of reservation models for demand level 90. Demand = 90 FullNest 1Up NoNest DSAR Revenue ($) % Difference % Reservation Served Average Wait Reservation (min) % Walk-in Served Average Wait Walk-in (min) 694468 000% 9596% 30 8119% 111 691718 −040% 9595% 32 8090% 113 677702 −241% 9602% 33 7907% 123 718220 342% 9797% 40 8502% 86 of 16 periods, from 6 pm to 10 pm. The service time and revenue data are also identical to that of §5.1. Capacity. Soto’s has many small tables that can be put together to accommodate large parties. For the purpose of our model, we assume that the restaurant takes only one table configuration. In this configuration, Soto’s has 16 tables for two, 7 tables for four, 3 tables for six, and 1 table for eight. We allow nesting of capacity as before. Demand Data. Soto’s gets a total of around 90 customers on weekdays and 120 to 130 customers on weekends. We test the models on these two demand levels, with corresponding loads of 0.93 and 1.24, respectively. Around 30% of these customers are reservation customers with a no-show rate of 3% to 15%. We use a constant no-show rate of 10%. Out of the walk-in parties, 55% are of size two, 30% are of size, four and 15% are of size six. Out of the reservation parties, 40% are of size two, 43% are of size four, 10% are of size six and 7% are of size eight. The distribution of reservation customers throughout the day is as follows: 20% during 6 pm to 7 pm, 40% during 7 pm to 8 pm, 30% during 8 pm to 9 pm, and 10% during 9 pm to 10 pm. The distribution of walk-in customers is as follows: 30% during 6 pm to 7 pm, 35% during 7 pm to 8 pm, 25% during 8 pm to 9 pm, and 10% during 9 pm to 10 pm. 6.2. Parameter Settings We use the following values for CostQWt k : 2.5 for k = 2, 6.0 for k = 4, 10.5 for k = 6, and 16.0 for k = 8, ∀t. The values for CostQRt k are 150 for k = 2, 300 for k = 4, 500 for k = 6, and 700 for k = 8, ∀t. Distinction for the values for both CostXWt k and CostXRt k for t = now are set higher than for t > now. This implies that parties expected to arrive in the current period are given more priority than those expected to arrive later because the state Table 5. / of the distant future is more uncertain than the near future. The values for CostXWt k when t = now are 0.05 for k = 2, 0.12 for k = 4, 0.21 for k = 6, and 0.32 for k = 8. When t > now, the values are set to 0 for all k. The values for CostXRt k when t = now are 101 for k = 2, 241 for k = 4, 421 for k = 6, and 641 for k = 8. When t > now, the values are 100 for k = 2, 240 for k = 4, 420 for k = 6, and 640 for k = 8. The value of Mr was set to 8 and Mw was set to 3, reflecting higher cost for keeping reservation customers waiting  was set to 0.5. MaxR was set to 4 and MaxW was set to 2, which reflects the customer behavior at Soto’s. 6.3. Computational Results The average revenue, percentage of customers served, and average waiting time for reservation and walk-in customers are illustrated in Tables 4 and 5. Table 4 uses the demand data of 90 customers and Table 5 uses the demand data of 120 customers. In the low-demand scenario of 90 customers, the DSAR outperforms all of the FCFS models by 3.5% to 6.9%. The DSAR also serves a larger percentage of both reservation and walk-in customers than the FCFS models. The average wait for reservation parties is slightly higher for the DSAR, but 0.27 periods (4.05 minutes) is still a reasonable length of wait. The average wait for walkin customers is lowest using the DSAR. The results for 120 customers are similar. The DSAR outperforms the FCFS model by 6.43% to 8.29%. It again has the best percentage seated for both the reservation and walk-in customers. It also has the highest average waiting time (0.40 periods, or 6 minutes) for reservation parties and the lowest average waiting time for walk-in parties. Thus, the DSAR produces more revenue and serves more customers than FCFS Revenue, percent served, and average waiting time of reservation models for demand level 120. Demand = 120 FullNest 1Up NoNest DSAR Revenue ($) % Difference % Reservation Served Average Wait Reservation (min) % Walk-in Served Average Wait Walk-in (min) 8,210.70 0.00% 94.57% 3.6 53.97% 25.4 8,274.82 0.78% 94.42% 3.75 53.75% 25.5 8,132.2 -0.96% 93.92% 4.2 50.39% 26.7 8,806.60 7.26% 97.39% 6.0 65.83% 19.5 486 / Bertsimas and Shioda for both low and high demands. The results also imply that the DSAR has a higher revenue impact with higher demand. 7. SUMMARY AND CONCLUDING REMARKS We feel we gained the following insights from the computational study: (1) For models without reservations, optimization-based strategies outperform FCFS-based strategies for all low and medium load factors and significantly for high load factors. Somewhat surprisingly, optimization-based strategies do not adversely affect the service quality (waiting times either remain unchanged or decrease somewhat, while FCFS is maintained within parties of the same size). (2) Increasing the sophistication in the models results in higher revenue without sacrificing waiting time. We believe that the performance of the ADP model represents the best trade-off between maximizing revenue and maintaining low average waiting time and run time. (3) The reservation models we propose (using a stochastic gradient approach to decide the reservations a priori, and ADP to implement it online) result in a significant improvement relative to the FCFS models for both low and high demand levels. The RB and DSAR models result in both higher revenue and lower customer attrition. (4) Overall, we feel that optimization-based models have a role to play in restaurant revenue management. There are many areas for future research: (1) Extending our model to support dynamic capacity— that is, allowing restaurants to move their tables around to better accommodate the demand at each time. (2) Incorporating balking and reneging. (3) Further empirical testing; this might be facilitated by combining algorithms from this paper with online restaurant reservation providers. ENDNOTES 1. In the language of queueing theory, reneging is when a customer prematurely leaves the restaurant, after waiting in line, before receiving any service. 2. The probability distribution of the length of each phase for each party size is approximated by a discrete distribution illustrated in an online appendix. 3. The load factor 7k corresponding to party size k was calculated as 7k = 6̄k Sk /c̄k, where 6̄k
= the expected rate of arrival of size k parties, Sk
= the expected service time of size k customers, and c̄k
= the average number of tables of size k over all three configurations. 4. The online appendix contains an illustration of the expected demand for each period and party size corresponding to each of the demand scenarios (http://or.pubs. informs.org/Pages/collect.html/). ACKNOWLEDGMENTS This research was supported in part by the Singapore-MIT alliance. The authors thank Sotohiro Kosugi, owner and chef of Soto’s, a Japanese restaurant in Atlanta, Georgia, for sharing data from his restaurant with them and helping the authors to better understand restaurant operations. They also thank the reviewers of the paper and the associate editor for many insightful suggestions. REFERENCES Ahuja, Ravindra K., Thomas L. Magnanti, James B. Orlin. 1993. Network Flows: Theory, Algorithms, and Applications. Prentice Hall, Englewood Cliffs, NJ. Karaesmen, I., G. van Ryzin. 1998. Overbooking with substitutable inventory classes. Working paper, Columbia Graduate School of Business, New York. Kimes, Sheryl E. 1999. Implementing restaurant revenue management: A five step approach. Cornell Hotel Restaurant Admin. Quart. 40(3) 16–21. , Deborah I. Barrash, John E. Alexander. 1999. Developing a restaurant revenue-management strategy. Cornell Hotel Restaurant Admin. Quart. 40(5) 18–29. , Richard B. Chase, Sunmee Choi, Philip Y. Lee, Elizabeth N. Ngonzi. 1998. Restaurant revenue management: Applying yield management to the restaurant industry. Cornell Hotel Restaurant Admin. Quart. 39(3) 32–39. Kosugi, Sotohiro. 1999. Owner and manager of Soto’s, personal communication. Atlanta, GA. Kushner, H. J., D. S. Clark. 1978. Stochastic Approximation Methods for Unconstrained Systems, AMS Vol. 26. SpringerVerlag, New York. Martinez, R., M. Sanchez. 1970. Automatic booking level control. AGIFORS Sympos. Proc., Vol. 10. American Airlines, New York. McGill, Jeffrey I., Garrett J. van Ryzin. 1999. Revenue management: Research overview and prospects. Transportation Sci. 33(2) 233–256. Muller, Christopher C. 1999. A simple measure of restaurant efficiency. Cornell Hotel Restaurant Admin. Quart. 40(3) 31–37. Quain, Bill, Michael Sansbury, Stephen M. LeBruto. 1998. Revenue enhancement, part 1: A straightforward approach for making money. Cornell Hotel Restaurant Admin. Quart. 39(5) 41–48. Sill, Brian. 2000. Capacity management: Engineering the balance between customer satisfaction, employee satisfaction and company profit. The Consultants (second quarter) 72–81. , Robert Decker. 1999. Applying capacity-management science. Cornell Hotel Restaurant Admin. Quart. 40(3) 22–30. Vakharia, A. J., H. S. Selim, R. R. Husted. 1992. Efficient scheduling of part-time employees. Omega 20(2) 201–213.