A VEHICLE ROUTING PROBLEM WITH STOCHASTIC DEMAND DIMITRIS J. BERTSIMAS Massachusetts Insitute of Technology, Cambridge, Massachusetts (Received September 1988; revisions received February, November 1990; accepted May 1991) We consider & natural probabilistic variation of the classical vehicle routing problem (VRP), in which demands are stochastic. Given only a probabilistic description of the demand we need to design routes for the VRP. Motivated by applications in strategie planning and distribution systems, rather than resolving the problem when the demand becomes known, we propose to construct an a prion sequence among all customers of minimal expected total length. We analyze the problem using a varity of theoretical approaches, We find closed-form expressions and algorithms to compute the expected length of an a priori soquence under general probabilistic assumptions. Based on these expressions we find ‘upper and lower bounds for the probabilistic VRP and the VRP re-optimization strategy, in which we find the optimal route at every instance. We propose heunstics and analyze their worst case performance as well a their average behavior using techniques from probabilistic analysis. Our results suggest that our approach is a strong and useful altemnatve to the strategy of re-optimization in capacitated routing problems, he deterministic vehicle routing problem (VRP) is a well studied problem in the operations research literature, In this paper, we study an impor- tant variation of the VRP, in which demands are probabilistic in nature rather than deterministic. In particular, a single vehicle of limited capacity must meet demands at 1 fixed locations, returning periodi- cally to the depot to empty its current load. The objective is to minimize the total distance traveled. In this paper, we consider the situation in which demand at each location is unknown at the time when the tour is designed, but is assumed to follow a known proba- bility distribution. This situation arises in practice whenever a company (e.g., UPS), on any given day, is faced with the problem of deliveries (collections) to (from) a set of customers, which have random demand. An obvious approach to this problem is to redesign the routes when the demand becomes known. There are, however, several difficulties with this approach; the system's operator might not have the resources for doing so; or, it may be that such redesign of tours is, not sufficiently important to justify the required effort and cost, Even more importantly, the operator may have other priorities, such as regularity and persona ization of service, by having the same vehicle and driver visita particular customer every day. Moreover, it might be very difficult to learn the demand on a particular day before actually visiting the customer. Instead of redesigning the routes every day we pro- pose a different strategy to update the routes: Deter- mine a fixed a priori sequence among all potential customers. Depending on when information about a customer's demand becomes available we can define two different strategies for updating the routes. Strategy a Under Strategy a the vehicle visits all the customers in the same fixed order as under the a priori sequence, but serves only customers requiring service that day. The total expected distance traveled corresponds to the fixed length of the a priori sequence plus the expected value of the additional distance that must be covered whenever the demand on the sequence exceeds vehicle capacity. Strategy b Strategy b is defined similarly to Strategy a with the sole difference that customers with no demand on a particular instance of the vehicle route are simply skipped. To illustrate the difference between the two strategies consider the example in Figure 1. If the a priori sequence is (0, 1, 2, 3, 4, 5, 6, 0), the depot is node 0, the vehicle has capacity 3, and the demand of the customers is D, = 0, D; = 2, Ds = 1, Di = 0, Ds = 2, De = 0, then under Strategy a the resulting routes are shown in Figure I (left) and under Strategy b the resulting routes are shown in Figure 1 (right). Note that at node 3 the capacity is reached and the vehicle is forced to return to the depot. Sulyectclasucations. Networks/graphs stachasc applications Probsbihiy: stochastic model applications Transportation, vehicle routing stochashe ‘eh routing Arca of rest DISTRIB HON, TRANSPORTATION AND LOGIHCS Operations Research Yor 20, No.3 May-June 1992 374 "CS ETEETE BOOT AN Rights Resenved 7 ‘0n30-364X /92/4005.0874 $01.28, © 1992 Operations Rewearch Socsety of Americao® © Apprior! sequence Vehicle Routing With Stochastic Demand $75 ©o Strategy a Strategy b Figure 1. ‘The two updating strategies a and b. There is an important difference in the philosophy of the two updating strategies. Strategy a models situ- ations in which the demand (if any) of any particular customer becomes known only when the customer is Visited. The vehicle is then forced to return to the depot when its capacity is reached. Under Strategy b, however, the actual demand is known before the tour starts (customers call or the operator calls them or in the case of package deliveries the addresses are known), so that savings can occur by skipping cus- tomer locations with zero demand. OF course, in Strategy b we assume that the demand is known before the vehicle starts its route, so itis clearly better to use a re-optimization strategy. As mentioned, however, there might be practical considerations that make the re-optimization strategy less attractive (computing resources are not available, itis time consuming even iff resources are available, regularity of service, etc.) The natural question that arises is how to choose the a priori sequence. We propose to choose an a priori sequence of minimal expected total length, which corresponds to the expected total length of the fixed set of routes plus the expected value of the extra distance that might be required by a particular reali- zation of the demand. The extra distances will be due to the fact that demand on the route may occasionally exceed the capacity of the vehicle and force it to go back to the depot before continuing on its route. We call the problem of selecting the a priori sequence of minimum expected length, when we use updating Strategy a (b), the probabilistic vehicle routing prob- Jem (PVRP) under Strategy a (b). We now review some situations in which vehicle routing problems with stochastic demand arise. Application Areas In a strategic planning scenario, consider a delivery and collection company which has decided to begin service in a particular area. The company has carried out a market survey and identified a number 1 of potential major customers who during any collection/ distribution period have a significant probability of requiring a visit. The company wishes to estimate the resources necessary to serve these customers. At this stage of planning, the company assigns the same probability distribution of demand to all potential Copyright © 2001 All Rights Reserved3576 / BeRTsimas customers, To address the planning problem the com- pany will wish to estimate approximately the expected amount of travel that will be necessary on a typical day to serve the subset of n customers that will require a visit. Ina routing context, consider, for example, a prob- Jem in which a central bank has to collect money on a daily basis from several but not all of its branches. ‘The capacity Qof the vehicle used may not correspond to any physical constraint but to an upper bound on the amount of money that a vehicle might carry for safety reasons. The distribution of demand at each particular branch may be different depending on the amount of money it handles. In the same way, there is.a similar problem when the bank wishes to deliver money to automatic teller machines that are located in several locations in each area Similarly, the distribution of packages from a post office can be modeled as a PYRP, where the proba- bility that a certain building requires a visit is given and the capacity Q corresponds to the physical con- straint that a truck can carry only a fixed weight or volume. Other examples reported in the literature include a “hot meals” delivery system (Bartholdi et al 1983) and routing of forklifts in a cargo terminal or in a warehouse. Brief Literature Re\ w The scientific literature concerning the VRP has been expanding at a very rapid pace, see, for example, the three excellent review volumes on the traveling sales- man problem (Lawler et al. 1985), on routing and scheduling (Bodin et al. 1983) and on vehicle routing, (Golden and Assad 1988), each of which offers several hundreds of references. Except for an isolated result in the 1970s (Tillman 1969), VRPs with stochastic elements in their definitions have received attention only recently. Stewart and Golden (1983), Dror and Trudeau (1986), Laporte and Louveau (1987), and Laporte, Louveau and Mercure (1987) use techniques from stochastic programming to solve optimally small problems and find bounds for the problems. Com- pared with this approach, our approach is entirely different, We propose to find an a priori solution for the problem, which iseasily updated when the demand is realized. Moreover, we derive worst case and average case bounds for the performance of the strategy we propose. We also compare the strategy of finding the a priori solution with the re-optimization strategy, ‘which is the best one can possibly do. The idea of using an a priori sequence for the solution of traveling salesman problems when "Sopvright © 2001 All Rights Reserved instances are modified probabilistically was first intro- duced in the Ph.D. thesis of Jaillet (1985) (see also Jaillet 1988). This idea was generalized to other com- binatorial optimization problems in the Ph.D. thesis of Bertsimas (1988), in which the probabilistic mini- mum spanning tree, the probabilistic traveling sales- man problem, the probabilistic vehicle routing problem, and facility location problems were ana- lyzed. In all these investigations the demand distribu- tion is assumed to be binary, ie., customer / has a unit demand with probability p, or does not have any demand with probability 1 — p.. In this paper, we consider arbitrary discrete-demand distributions. ‘The paper is organized as follows. In Section 1, we introduce the problem formally and establish the nota- tion. In Section 2, we address the question of finding closed-form expressions and algorithms to compute the expected length of an a priori sequence under general probabilistic assumptions. In Section 3, we prove upper and lower bounds for the PVRP and the VRP re-optimization strategy, in which we find the optimal route after the demand is realized. We further use these bounds to propose heuristics with provable ‘worst case performance. In Section 4, we examine the asymptotic behavior of the PVRP using techniques from probabilistic analysis for the case that customer locations are randomly distributed in the Euclidean plane, In the final section, we summarize the contri- butions of the paper and discuss the limitations and applicability of our model as well as future research, 1. FORMAL DEFINITION AND NOTATION In this section, we formally define the PVRP and establish the notation we will use. Given a complete network, let the nodes be {0, 1, ..., m}, where node 0 denotes the depot and the set = {1, 2, ..., 1] denotes the set of customer locations. The distances ali, j) are assumed to be symmetric (although our results can easily be modified to hold even in the nonsymmetrie case) and they satisfy the triangle inequality: d(i, j) < d(i, k) + dk, j), Let the capacity of the vehicle be Q (Q is an integer) and let D,, i= 1, ..., be the random variable that describes the demand of customer i. We assume that the probability distribution of D, is discrete and known. Let p(k) = PriD, = Ki, i=1,..., mand k=0,1,..., K. We further assume that K < Q, that is, no single location has demand exceeding the capacity Q. We further assume that the demands are independent. There are (K + 1)" possible reali jons of thedemand and therefore (K+ 1)" possible instances of the problem. Ifwe solve the underlying VRP optimally at every problem instance, ie., we find the route that minimizes the total distance traveled, let Reno(D1, Ds, «.., Dx) be the optimal route length if the demand is Dy, D:, ..., Dy. Note that since the demand is stochastic this is a random variable. We call the expectation of this random variable the expected length under the re-optimization strategy because we redesign (re-optimize) the routes at every problem instance. This expected length is thus given by EIR wo] = Lilie. PaltndReeellas vor dds a) where the summation is over all demand instances for the nodes, Clearly, the exact estimation of E[Ri¢r] is a very difficult problem because it involves (K + 1)” terms. Moreover, in order to evaluate each term (Ri wii... t)) We need to solve exactly a VRP. So, in a strategic planning scenario, in which a company needs to have an estimate of the expected travel cost, ‘the expected length of the re-optimization strategy is not a realistic alternative computationally. Related to the vehicle routing re-optimization strategy, we can define a traveling salesman re- optimization strategy in which, at every instance of the problem, the vehicle visits all the customers with nonzero demand according to the optimal traveling salesman tour. We denote ElRise] = X II PriD, > 0} TI PriD, = OLS), @) where Lysp(S) is the length of the optimal TSP tour among customers in the set S. Let us now consider the ovo a priori strategies we are proposing. Given an a priori sequence + let Liliy, «+5 dq) be the length of the a priori sequen that will result under strategy / = a, b if the demand pattern is f),..., in. We denote with FIL = D piliy) PablndLiCss == 25 tne eB) the expected length of the a priori sequence + under Strategy a and X paid ELLY) = «Palin Ely ves de), (AY Vehicle Routing With Stochastic Demand J $77 the expected length of the a priori sequence r under Strategy b. Our goal is to find the a priori sequences r, and + that minimize the expected lengths in (3) and (4) respectively. Although there are (K+ 1)” terms in both (3) and (4) we will be able to compute the expected length of an a priori sequence efficiently in the next section, 2. THE EXPECTED LENGTH OF AN A PRIORI ‘SEQUENCE In this section, we propose an algorithm for finding the expected length of an a priori sequence under Strategies a and b. We first consider the case of binary demand, where a customer cither has a demand of fone unit or it does not have any demand, 2.1. Binary Demand We first examine the important special case in which all the demand is binary, either 0 or 1. There are several motivations for examining this case separately. From an applications viewpoint, it is important in situations in which the randomness in the demand can be modeled by the presence (or not) of a customer. For example, in the distribution of packages by the post office, there is a potential set of customers each requiring a visit with probability p.. From a theoretical viewpoint, the usual traveling salesman problem can be viewed as a special case of the PVRP under Strategy a, ifthe capacity Q= n,i.c., the problem is uncapacitated. Moreover, the probs bilistie traveling salesman problem (PTSP) introduced and analyzed in Jaillet (1988) and further explored in Bertsimas (1988) can be viewed as a special case of the PVRP under Strategy b, for which the capacity Q is equal to 1, i, the capacity of the vehicle is not a binding constraint. Moreover, the insights gained from the binary case carry over to the general case. Our initial goal then is to compute E[L‘], FUL") efficiently for a given a priori sequence 7. Let p, be the probability that customer i has a demand of one unit and 1 ~ p, of not having any demand independently of any other customer. Then we can compute the expected length of an a priori sequence as follows. ‘Theorem 1. Ifthe a priori sequence is 7 =O. 1... nn +10), then BUA] =Y awit ys Dysei+ vy, (5) Copyright © 2001 Alll Rights Reserved578 / BeRTsimas BILE] = X 400. tp, TA ~ pe) + Sau Op, I -p) + z E dior, MW -p) +3 z s(t, DyPs ta —P) 6) where si, j) & d(i, 0) + dQ, j) ~ di, J), y= 0, (= 0, ,Q- n= PD fim LkO- iF, a and fm, r) & Priexactly r customers among the customers 1, ..., m have nonzero demand\ are computed from the recursion: For m = 1, ..., m, ra lem SL(m, 1) = Poflm = lyr 1) += pedflm = 1,9, @) with the initial conditions fm, m)= Th ps fom, 0) = TL 0 ~ p. Proof. Consider first Strategy a. The expected length of the sequence is a summation of the length of the a priori sequence plus the expected value of the extra distance when the vehicle reaches its capacity. To evaluate this second term, let i be a node on the sequence, where the vehicle reaches its capacity. The vehicle will then go to the depot before going back to the following node in the route, which is i + 1 under Strategy a, even if node i + 1 has no demand. The extra distance traveled is then s(i, i + 1) = d(i, 0) + dO, i+ 1) ~ di + 0). In (5), 7 is the probability that the vehicle reaches its capacity Q at node i. Clearly, 7, = 0 for i= 0,..., Q~ 1. Consider now node 1. For the vehicle to reach its capacity at node 1 (1 = Q), node i must have a unit demand and from the previous i ~ 1 nodes exactly kQ — 1 must be present for some k= 1,.... l/Ql, so that the capacity is reached with the addition of node i, From this observation (7) follows. The proba- bilities (mm, r) are computed recursively from (8) by conditioning on the event that node m has a demand. re ASC een Under Strategy b, the first three terms in (6) are simply the expected length of the tour r in the prob- abilistic traveling salesman sense. In particular, the distance d(/, j) contributes to the expectation if nodes and J are present and all the intermediate nodes have zero demand. The fourth term is identical with Strategy a, except that when the vehicle reaches its capacity at node #, it goes back, after a visit to the depot, to the first node j with a nonzero demand, skipping nodes i+ 1, i+ j= Lwith no demand, As an application of (5) and (6) we find the closed- form expressions derived in Jaillet and Odoni (1988) for the case in which all points have the same proba- bility p of requiring a visit. Then expressions (8) imply that fmm, 7) = (2")p"(1 =p)", and thus FIL = 3 a ine DG i+t 1 = py, UL! EL) + z 'S Ue ip OL = pyre Xs Dea = pyres where E[L,] denotes the expected length of the a priori tour 7 in the probabilistic traveling salesman sense (the first three terms in (6)). ‘An important consequence of (5) and (6) is that they provide an algorithm of O(*) to compute ELL!), E[L?) for the general case of unequal proba- bilities, because the computation of the probabilities f(m, r) can be done recursively from (8) in O(n”), and there aren = Q nonzero probabilities »,. The computation of each of these probabilities from (7) requires the evaluation of a sum of at most [n/Q1 terms. Thus, we can compute all the y, in O(n = Q)n/Q + m) = O(n"). Finally, the expecta- tion of the length of the sequence, given that we have already computed the probabilities 7, is done in O(n) for Strategy a and O(n") for Strategy b, which means that we can compute the expected length of an a priori sequence in O(n*) for both strategies. In the next subsection, we generalize these expressions for the case of general discrete-demand distributions 2.2. General Demand In this section, we find expressions for the expected length of a priori sequences under Strategies a and b.Let Pr{D, = k] = pk) be the probability that the demand of customer i, D,, is equal to k, for k ranging, from 0 to K. As mentioned, we assume that K < Q, ive,, the vehicle's capacity is greater than the largest demand of a customer on a given day. This assump- tion removes the consideration of multiple returns to the depot from the same node. Similarly, as before, we define +, to be the probabil- ity that the vehicle exactly reaches its capacity at node and 6, to be the probability that the vehicle exceeds its capacity at node 7, Then the expected length of an a priori sequence is computed as follows. Theorem 2. if the a prior sequence is + = (0, 1, nen 120) then, ELL!) = ¥ di, i+) + S sti) + yh + dD) (9) BIL = ¥ dO, dp. Th ( - + Ydiow. iL Up) +3 3 dion, ila -p) + 3 (00 D+ Y vsti) ha- ). (10) where © pk) = 1 pO), s(i, j) = dli, 0) + dO, j) — di. i), m=0 y ‘x {3 pk) fi ~ 1, aa - »} . 2sien, 5 =0, wie) pres 7 =z {3 (3 ne) 1. a0- 2 Km f= a forO2Y dri, Now since the capacity is never exceeded on any subtour, we have S, E[Rrse). Moreover, from the triangle inequality, si, j) > 0. Therefore, from (9) E[L2,] = L,, > Lise. Similarly, EL!) > EUL,,] > ELL,,), and henee, using (11) we obtain (13) and (14), In Proposition 2, if the distance matrix is asymmet- ric, then we should replace the term 28 gd d0.neP] with 33 [d(0.7)+d¢r, OED] 3.3. Upper Bounds In this subsection, we concentrate on finding upper bounds for the two probabilistic Strategies a and b. We first consider the case of identically distributed demand and for convenience we assume that the Gistance matrix is symmetric. We consider the cyclic heuristic introduced in Haimovitch and Rinnooy Kan (1985) in the context of the deterministic VRP.Cylic Heuristic Step 1. Given an initial sequence + & 1, <7, 0), consider the sequences Oriel 1,0), Step 2. Compute E(L)] for all 2, Peet Step 3. The sequence with the minimum expected length among EL‘, i= 1. ..., 1 is the proposed solution +), to the PVRP under Strategy a. We now analyze the worst case behavior of the cyclic heuristic under the assumption that each customer has the same demand distribution, Proposition 3. Ler D be the random variable deserib- mg the demand of each customer. If the initial sequence to the eyelic heuristic is the optimal traveling salesman tour and ris the tour proposed by the evelte heuristic, then under the triangle inequality FILS IS ELS] stonft +214 Meld) 2 (0, i) as Q n Proof. If the initial sequence to the cyclic heuristic is the optimal deterministic tour, then let LAY ditt) +din1). With this definition the lengths of the sequences 7, become’ L,,=L+d(0, 1) +d(0,n)=d(1,n), L,,=L+d(0,i)+d(0,i=1)=d(i,i- i= 2.00 Asa result, LD L,=2¥ dO, i) += ve Clearly, FUL J < FUL) <2 BLL From (9) and the triangle inequality EULS] $ Le, + 2D (y, + 5) dO. ii), where s(i) = (J + i — 2)mod(n) + 1. Therefore, res) <2] E1423 O48) ¥ ao.a). Vehicle Routing With Stochastic Demand $81 since the probabilities y, + 4, multiply every term (0, i) in Yt, E[L*). Therefore, Ly] iPr = \2 + Q~rfor some I < Y PriD,=r} 0}. For the PTSP the following results are known, Then with probability 1 (Jaillet 1988): i Flfsuta D Base VD, (@2) Vehicle Routing With Stochastic Demand | $83 (23) where Bsr is the constant appearing in the celebrated Beardwood, Halton and Hammersley (1959) theorem and 8(p) is a constant that only depends on p for which Brse Vb < 8(p) = min{0.92 Yp, Brsel The asymptotic behavior of the expected length of the re-optimization strategy and the PVRP depends critically on the dependence of the capacity Q on the number of customers n. This dependence is also ical for the asymptotic behavior of the VRP examined in Haimovitch and Rinnooy Kan. Let Q, denote the capacity of the vehicle to indicate its dependence on n. We prove the following theorem. Theorem 5. The asymptotic behavior of the three updating strategies is: 1, fQ isa constant, then almost surely tim ERA XN i, ELL] mn an = tim EUELO) _ 2EUIE(D] 2 = lim Fj 0 (24) 2. ff lity. (Qu/-¥n) = 0, then almost surely tim QeELRe AA pg, QoE TLE X] ee 7" on = tim SALE opt ELD) 05) 3. A lity sx Qn/Vn = &, then almost surely sin EE ya mn jim Pee CO 6, mn ELK) a lim Se = ap) 26) Proof. Let? 8 1 a0, i)/n Case 1. Assume that @ is constant. Combining the bounds from Propositions 2 and 5 we obtain 2E(D] E[Rs wo X"")) @ isn [Ravel XK) 7 (1 a) Copyright © 2001 Alll Rights Reserved584. / BeRTsIMas But E[r] <% implies that F + E{r] almost surely by the strong law of large numbers. Taking limits and using (22) we obtain (24) for the re-optimization strategy. We obtain (24) for the PVRP strategies in a very similar manner using the bounds from Propositions 1, 2, and 3. Case 2, Similarly, Q DEWDIF < =P ERs AX) 4 AER reo en Since with probability 1, tim QETR AK ying Oo ELR roe Xi" n men vn 0 (25) for the re-optimization strategy follows by taking, limits. Again, we obtain (25) for the PVRP strategies in a very similar manner using the bounds from Propositions 1, 2, and 3. Case 3. From Propositions 2 and 5 ELR isl X")) < ELR rs X)] va vn “Dye ~O/sn From (22) and since lim,» (Q,/V7) = %, we obtain (26) for the re-optimization strategy by taking limits With regard to Strategy a from Propositions 2 and 3 we obtain: Lesol vn Lis X"”) eg Sia Olin vn va Since limy.. Lyse(X)/Vn = Bysy almost surely and Timy sn (Qx/Win) = %, we obtain (26) for Strategy a by taking limits. Finally, for Strategy b using Propositions 1, 4, and. 6 we obtain: ELL. (XY) S ELLE (XY) = ELL, XY] +2 dO, Ky +5) <£ LX) + 2d(0, finan) Y Cvs + 4) Copyright © 2001 All Rights Reserved “ Dividing with Vi, taking the imit as n +o and using (23), we obtain (26) for Strategy b. The case Q, = @(,/n) is not covered in the previous theorem. The reason is that in this case neither the radial collection term, (2E{DJn?/Q,), nor the local collection term, E[Rrse(X)], dominate as was the case in Theorem 5, where the radial collection term dominated in Cases 1 and 2 and the local calculation term dominated in Case 3. Another interesting observation is that in both Cases | and 2 the sequence produced by the eyelic heuristic with initial tour any tour of length O(Wn) is asymptotically optimal for Strategy a. Moreover, for Strategy a the sequence produced by the cyclic heuris- tic with initial sequence the optimal TSP is asymptot- ically optimal in Case 3 as well. In Case 3 the PTSP solves the PVRP under Strategy b optimally. An even more interesting con- sequence of Theorem $ is that in Cases | and 2 both the probabilistic Strategies a and b are asymptotically equivalent to the re-optimization strategy. Further- more, as we argued before, in these cases the tour produced by the cyclic heuristic with initial sequence any tour of length O(vn) is asymptotically equivalent to the re-optimization strategy. In Case 3, Strategy b is asymptotically within a constant factor from the strategy of re-optimization, Indeed, we conjecture that the constant factor is 1, i.e., Strategy b is asymptoti- cally optimal, Moreover, in this case Strategy a is also close to the re-optimization strategy if pis large. Finally, we only considered the case where customer locations are uniformly distributed in the unit square. Similar asymptotic theorems can be proved in the ddimensional Euclidean space and, furthermore, for the case where the distribution of customer locations has a continuous part with density f. We chose dimen- sion two in the exposition because the geometry is clearer, and the uniform distribution for customer locations, because it is more intuitive. 5. CONCLUDING REMARKS In this paper, we propose a different approach to solve vehicle routing problems when the demand is sto- chastic. We give analytical evidence that this ap- proach, which is based on finding an a priori sequence among the entire set of customers, performs quite well from a worst case perspective, especially if the distri- bution of the demand of the customers is the same. Although we propose worst case bounds for the case of nonidentical demand distributions, our boundswere not tight. Finding tighter bounds is an interesting, open problem deserving further work. We also attempt to give analytical evidence that the a priori strategies we propose for the PVRP are ver close to the strategy of re-optimization, on average. In particular, we show that if customer locations are uniformly distributed in the unit square, both Strategies a and b perform asymptotically very clos to the strategy of re-optimization, which is a strong indication of the usefulness of these strategies. Fur- thermore, our analysis reveals some asymptotically optimal heuristics for both Strategies a and b. It is a fair conclusion to say that a priori and re-optimization strategies perform comparably, on averag From a practical standpoint, we believe that the a priori strategies we are proposing provide a strong, alternative to the strategy of re-optimization, and therefore they can be useful in practice, especially in the absence of intense computational power. Related to this point, an interesting question deserving further research is to computationally compare the a priori and re-optimization strategies. Asa final conclusion, we believe the paper demon- strates that in the context of capacitated vehicle rout- ing problems a priori strategies (PVRP) are a serious, and practical alternative to re-optimization strategies. Our investigation in Bertsimas (1988) reached the same conclusion for other combinatorial optimization problems ACKNOWLEDGMENT I thank Professors Amedeo Odoni and Patrick Jaillet for several useful comments and interesting discus- sions. This work was partially supported by the National Science Foundation under grant ECS- 8717970. REFERENCES ALTINKEMER, K., AND B, Gavistt, 1985. Heuristics for Unequal Weight Delivery Problems With Con- stant Error Guarantees, Working Paper QM8536. Graduate School of Management, University of Rochester, N.Y BartHoLpl. J. J., LK. PLATZMAS, C. R. Lee AND W, Vehicle Routing With Stochastic Demand {585 H, WaRDEN, 1983. A Minimal Technology Routing System for Meals on Wheels, Interfaces 13, 1-8. BraRDWOOD, J., J. HALTON AND J. HAMMERSLEY. I The Shortest Path Through Many Points. Proc Camb, Phil. Soe. 85, 299-327, Benrsimas, D. 1988. Probabilistic Combinatorial Opti- mization Problems. Ph.D. Thesis. Massachusetts Institute of Technology, Cambridge. Mass. Bomn, L., B. GOLDEN, A. ASSAD AND M. BaLt. 1983. Routing and Scheduling of Vehicles and Crews, The State of the Art. Comp. Opns. Res. 10, 169-211 Dox, M., AND P. TRUDEAU. 1986. Stochastic Vehicle Routing With Modified Savings Algorithm. Eur. J Opal, Res. 23, 228-235. GOUDEN, G.. AND A. ASSAD. 1988. Vehicle Routing: Methods and Studies. North-Holland, Amsterdam Hatmoviren, M., aND A. RINNOOY Kan. 1985, Bounds and Heuristics for Capacitated Routing Problems Math. Opns. Res. 10, 527-542. Jautter. P, 1985. Probabilistic Traveling Salesman Prob- lems. Ph.D. Thesis. Technical Report No. 185, Operations Research Center. Massachusetts Institute of Technology. Cambridge, Mass. Jitter, P. 1988. A Priori Solution of a Traveling Sales- ‘man Problem in Which a Random Subset of the Customers are Visited. Opns. Res. 36, 929-936. IAILLET, P., AND A. ODONI. 1988. The Probabilistic Vehicle Routing Problem. In Vehicle Routing: Methods and Studies, B. L. Golden and A. A. Assad (eds). North-Holland, Amsterdam, Laporte, G., AND F. Louveav. 1987. Formulations and Bounds for the Stochastic Capacitated Vehicle Routing Problem With Uncertain Supplies. Report G-87-23, Ecole des Hautes Etudes Commerciale, University of Montreal, Montreal, Canada Larorte, G.. F. LOUvEAU AND H. Meecurr. 1987 Models and Exact Solutions for a Class of Stochastic Location-Routing Problems. Report G-87-14, Ecole des Hautes Etudes Commerciale. University of Montreal. Montreal, Canada Lawten, E. L., J. K. LENStRA, A, RINNOOY KAN AND D. B. Susiovs (eds.). 1985. The Traveling Salesman Problem: A Guided Tour of Combinatorial Optimi- zation. John Wiley, Chichester, U.K. Stewanr, W.. AND B. GOLDEN. 1983. Stochastic Vehicle Routing: A Comprehensive Approach. Eur. J. Opn. Res. 14, 371-385 THLMan, F. 1969. The Multiple Terminal Delivery Problem With Probabilistic Demands. Trans. Sci. 3, 192-204. Copyright © 2001 All Rights Reserved