An Ant Colony Optimization Approach to the Probabilistic Traveling Salesman Problem

Transcription

1 An Ant Colony Otimization Aroach to the Probabilistic Traveling Salesman Problem Leonora Bianchi 1, Luca Maria Gambardella 1, and Marco Dorigo 2 1 IDSIA, Strada Cantonale Galleria 2, CH-6928 Manno, Switzerland {leonora,luca}@idsia.ch htt:// 2 Université Libre de Bruxelles, IRIDIA Avenue Franklin Roosevelt 50, CP 194/6, 1050 Brusselles, Belgium mdorigo@ulb.ac.be htt://iridia.ulb.ac.be/ mdorigo/ Abstract. The Probabilistic Traveling Salesman Problem (PTSP) is a TSP roblem where each customer has a given robability of requiring a visit. The goal is to find an a riori tour of minimal exected length over all customers, with the strategy of visiting a random subset of customers in the same order as they aear in the a riori tour. We address the question of whether and in which context an a riori tour found by a TSP heuristic can also be a good solution for the PTSP. We answer this question by testing the relative erformance of two ant colony otimization algorithms, Ant Colony System (ACS) introduced by Dorigo and Gambardella for the TSP, and a variant of it (ACS) which aims to minimize the PTSP objective function. We show in which robability configuration of customers ACS and ACS are romising algorithms for the PTSP. 1 Introduction Consider a routing roblem through a set V of n customers. On any given instance of the roblem each customer i has a known osition and a robability i of actually requiring a visit, indeendently of the other customers. Finding a solution for this roblem imlies having a strategy to determine a tour for each random subset S V, in such a way as to minimize the exected tour length. The most studied strategy is the a riori one. An a riori strategy has two comonents: the a riori tour and the udating method. The a riori tour is a tour visiting the comlete set V of n customers; the udating method modifies the a riori tour in order to have a articular tour for each subset of customers S V. A very simle examle of udating method is the following: for every subset of customers, visit them in the same order as they aear in the a riori tour, skiing the customers that do not belong to the subset. The strategy related to this method is called the skiing strategy. The roblem of finding an a riori tour of minimum exected length under the skiing strategy is defined as the Probabilistic Traveling Salesman Problem (PTSP). This is an NP-hard roblem [1,2], and was introduced in Jaillet s PhD thesis [3]. J.J. Merelo Guervós et al. (Eds.): PPSN VII, LNCS 2439, , c Sringer-Verlag Berlin Heidelberg 2002

2 884 Leonora Bianchi, Luca Maria Gambardella, and Marco Dorigo The PTSP aroach models alications in a delivery context where a set of customers has to be visited on a regular (e.g., daily) basis, but all customers do not always require a visit, and where re-otimizing vehicle routes from scratch every day is unfeasible. In this context the delivery man would follow a standard route (i.e., an a riori tour), leaving out customers that on that day do not require a visit. The standard route of least exected length corresonds to the otimal PTSP solution. In the literature there are a number of algorithmic and heuristic aroaches used to find subotimal solutions for the PTSP. Heuristics using a nearest neighbor criterion or savings criterion were imlemented and tested by Jézéquel [4] and by Rossi-Gavioli [5]. Later, Bertsimas-Jaillet-Odoni [1] and Bertsimas-Howell [6] have further investigated some of the roerties of the PTSP and have roosed some heuristics for the PTSP. These include tour construction heuristics (sace filling curves and radial sort), and tour imrovement heuristics (robabilistic 2-ot edge interchange local search and robabilistic 1-shift local search). Most of the heuristics roosed are an adatation of a TSP heuristic to the PTSP, or even the TSP heuristic itself, which in some cases gives good PTSP solutions. More recently, Laorte-Louveaux-Mercure [7] have alied an integer L-shaed method to the PTSP and have solved to otimality instances involving u to 50 vertices. No alication of nature-insired algorithms such as ant colony otimization (ACO) [8] or genetic algorithms has been done yet. This aer investigates the otentialities of ACO algorithms for the PTSP. In the remainder of the aer we first introduce the PTSP objective function and notations (section 2), then we describe the ACO algorithms which we tested (section 3), and finally we show the exerimental results obtained (section 4). 2 The PTSP Objective Function Let us consider an instance of the PTSP. We have a comletely connected grah whose nodes form a set V = {i =1, 2,..., n} of customers. Each customer has a robability i of requiring a visit, indeendently from the others. A solution for this instance is a tour over all nodes in V (an a riori tour ), to which is associated the exected length objective function E[L ]= (S)L (S). (1) S V In the above exression, S is a subset of the set of nodes V, L (S) is the distance required to visit the subset of customers S (in the same order as they aear in the a riori tour), and (S) is the robability for the subset of customers S to require a visit: (S) = i (1 i ). (2) i S i V S Jaillet [3] showed that the evaluation of the PTSP objective function (eq.(1)) can be done in O(n 2 ). In fact, let us consider (without loss of generality) an a riori tour =(1, 2,...,n); then its exected length is

3 An Ant Colony Otimization Aroach 885 (0) (1) (2) Fig. 1. The lengths of these sets of (sub)tours, (0), (1) and (2), constitute the first three terms of the exected length for the homogeneous PTSP. From left to right, the total length of each set of (sub)tours gives the terms L (0) (4)., L(1) and L (2) of equation E[L ]= n n i=1 j=i+1 d ij i j j 1 k=i+1 (1 k )+ n i 1 d ij i j i=1 j=1 n k=i+1 j 1 (1 k ) (1 l ). (3) This exression is derived by looking at the robability for each arc of the comlete grah to be used, that is, when the a riori tour is adated by skiing a set of customers which do not require a visit. For instance, an arc (i, j) is actually used only when customers i and j do require a visit, while customers i +1,i+2,..., j do not require a visit. This event occurs with robability j 1 i j k=i+1 (1 k) (when j n). In the secial class of PTSP instances where i = for all customers i V (the homogeneous PTSP), equation (3) becomes n 2 E[L ]= 2 (1 ) r L (r) (4) r=0 where L (r) n j=1 d(j, (j +1+r) modn). The L(r) s have the combinatorial interretation of being the lengths of a collection of gcd(n,r + 1) sub-tours 1 (r) l=1, obtained from tour by visiting one customer and skiing the next r customers. As an examle, Fig. 1 shows (0) (i.e., the a riori tour), (1) and for a PTSP with 8 customers. (2) 3 Ant Colony Otimization In ACO algorithms a colony of artificial ants iteratively constructs solutions for the roblem under consideration using artificial heromone trails and heuristic information. The heromone trails are modified by ants during the algorithm execution in order to store information about good solutions. Most ACO algorithms follow the algorithmic scheme given in Fig The term gcd stays for greatest common divisor.

4 886 Leonora Bianchi, Luca Maria Gambardella, and Marco Dorigo rocedure ACO metaheuristic for combinatorial otimization roblems Set arameters, initialize heromone trails while (termination condition not met) ConstructSolutions (AlyLocalSearch) UdateTrails end while Fig. 2. High level seudocode for the ACO metaheuristic. ACO are solution construction algorithms, which, in contrast to local search algorithms, may not find a locally otimal solution. Many of the best erforming ACO algorithms imrove their solutions by alying a local search algorithm after the solution construction hase. Our rimary goal in this work is to analyze the PTSP tour construction caabilities of ACO, hence in this first investigation we do not use local search. We aly to the PTSP Ant Colony System (ACS) [9,10], a articular ACO algorithm which was successfully alied to the TSP. We also consider a modification of ACS which exlicitly takes into account the PTSP objective function (we call this algorithm robabilistic ACS, that is, ACS). In the following, we describe how ACS and ACS build a solution and how they udate heromone trails. 3.1 Solution Construction in ACS and ACS A feasible solution for an n-city PTSP is an a riori tour which visits all customers. Initially m ants are ositioned on their starting cities chosen according to some initialization rule (e.g., randomly). Then, the solution construction hase starts (rocedure ConstructSolutions in Fig. 2). Each ant rogressively builds a tour by choosing the next customer to move to on the basis of two tyes of information, the heromone τ and the heuristic information η. To each arc joining two customers i, j it is associated a varying quantity of heromone τ ij, and the heuristic value η ij =1/d ij, which is the inverse of the distance between i and j. When an ant k is on city i, the next city is chosen as follows. With robability q 0, a city j that maximizes τ ij η β ij is chosen in the set J k (i) of the cities not yet visited by ant k. Here, β is a arameter which determines the relative influence of the heuristic information. With robability 1 q 0, a city j is chosen randomly with a robability given by τ ij η β ij, if j J k (i, j) = k (i) r J k (i) τir ηβ ir (5) 0, otherwise. Hence, with robability q 0 the ant chooses the best city according to the heromone trail and to the distance between cities, while with robability 1 q 0 it exlores the search sace in a biased way.

5 An Ant Colony Otimization Aroach Pheromone Trails Udate in ACS and ACS Pheromone trails are udated in two stages. In the first stage, each ant, after it has chosen the next city to move to, alies the following local udate rule: τ ij (1 ρ) τ ij + ρ τ 0, (6) where ρ, 0<ρ 1, and τ 0, are two arameters. The effect of the local udating rule is to make less desirable an arc which has already been chosen by an ant, so that the exloration of different tours is favored during one iteration of the algorithm. The second stage of heromone udate occurs when all ants have terminated their tour. Pheromone is modified on those arcs belonging to the best tour since the beginning of the trial (best-so-far tour), by the following global udating rule τ ij (1 α) τ ij + α τ ij, (7) where τ ij = ObjectiveF unc 1 best (8) with 0 <α 1being the heromone decay arameter, and ObjectiveF unc best is the value of the objective function of the best-so-far tour. In ACS the objective function is the a riori tour length, while in ACS the objective function is the PTSP exected length of the a riori tour. In the next section we discuss in more detail the differences between ACS and ACS. 3.3 Discussion of Differences between ACS and ACS Differences between ACS and ACS are due to the fact that the two algorithms exloit different objective functions in the heromone udating hase. The global udating rule of equations (7) and (8) imlies two differences in the way ACS and ACS exlore the search sace. The first and most imortant difference is the set of arcs on which heromone is globally increased, which is in general different in ACS and ACS. In fact, the best tour in eq. (8) is relative to the objective function. Therefore in ACS the search will be biased toward the shortest tour, while in ACS it will be biased toward the tour of minimum exected length. The second difference between ACS and ACS is in the quantity τ ij by which heromone is increased on the selected arcs. This asect is less imortant than the first, because ACO in general is more sensitive to the difference of heromone among arcs than to its absolute value. The length of an a riori tour (ACS objective function) may be considered as an O(n) aroximation to the O(n 2 ) exected length (ACS objective function). In general, the worse the aroximation, the worse will be the solution quality of ACS versus ACS. The quality of the aroximation deends on the set of customer robabilities i. In the homogeneous PTSP, where customer robability is for all customers, it is easy to see the relation between the two objective functions. For a given a riori tour L we have

6 888 Leonora Bianchi, Luca Maria Gambardella, and Marco Dorigo n 2 = L E[L ]=(1 2 )L (1 ) r L (r), (9) which imlies O(q L ) (10) for (1 ) =q 0. Therefore the higher the robability, the better is the a riori tour length L as an aroximation for the exected tour length E[L ]. In the heterogeneous PTSP, it is not easy to see the relation between the two objective functions, since each arc of the a riori tour L is multilied by a different robabilistic weight (see eq.(3)), and a term with L cannot be isolated in the exression of E[L ], as in the homogeneous case. ACS and ACS also differ in time comlexity. In both algorithms one iteration (i.e., one cycle through the while condition of Fig. 2) is O(n 2 ) [11], but the constant of roortionality is bigger in ACS than in ACS. To see this one should consider the rocedure UdateTrail of Fig. 2, where the best-so-far tour must be evaluated in order to choose the arcs on which heromone is to be udated. The evaluation of the best-so-far tour requires O(n) time in ACS and O(n 2 ) time in ACS. ACS is thus faster and always erforms more iterations than ACS for a fixed CPU time. r=1 4 Exerimental Tests 4.1 Homogeneous PTSP Instances Homogeneous PTSP instances were generated starting from TSP instances and assigning to each customer a robability of requiring a visit. TSP instances were taken from two benchmarks. The first is the TSPLIB at htt:// From this benchmark we considered instances with a number of city between 50 and 200. The second benchmark is a grou of instances where customers are randomly distributed on the square [0, 10 6 ]. Both uniform and clustered distributions where considered in this case, and the number of cities varied between 50 and 350. For generating random instances we used the Instance Generator Code of the 8 th DIMACS Imlementation Challenge at htt://research.att.com/dsj/chts/download.html. 4.2 Comutational Environment and ACS Parameters Exeriments were run on a Pentium Xeon, 1GB of RAM, 1.7 GHz rocessor. In order to asses the relative erformance of ACS versus ACS indeendently from the details of the settings, the two algorithms were run with the same arameters. We chose the same settings which yielded good erformance in earlier studies with ACS on the TSP [10]: m = 10, β = 2, q 0 = 0.98, α = ρ = 0.1 and τ 0 =1/(n Obj ), where n is the number of customers and Obj is the value of the objective function evaluated with the nearest neighbor heuristic [10].

7 An Ant Colony Otimization Aroach (<E_acs> - <E_acs>) / <E_acs> Fig. 3. Relative erformance of ACS versus ACS for the homogeneous PTSP. The vertical axis reresents (E[L (ACS)] E[L (ACS)])/E[L (ACS)]. On the horizontal axis there is the customer robability. Each oint of the grah is an average over 21 symmetric homogeneous PTSP instances. Error bars reresent average deviation, defined as n i=1 xi <x> /n, with n = 21. Note that for = 1 ACS outerforms ACS, since for a fixed CPU stoing time ACS makes more iterations. The stoing criterion used in both algorithms is CPU time in seconds, chosen according to the relation stotime = k n 2, with k =0.01. This value of k lets ACS erform at least iterations on roblems with u to 100 customers, and at least iterations on roblems with more than 100 customers. For each instance of the PTSP, we ran 5 indeendent runs of the algorithms. 4.3 Results For each TSP instance tested, nine exeriments were done varying the value of the customer robability from 0.1 to 0.9 with a 0.1 interval. Fig. 3 summarizes results obtained on 21 symmetric PTSP instances, one third of the instances were taken from the TSPLIB, the others were random instances (half of them uniform and half clustered). The figure shows the relative erformance of ACS versus ACS, averaged over the tested instances. A tyical result for a single instance is reorted in Fig. 4. As it was reasonable to exect, for small enough robabilities ACS outerforms ACS. In all roblems we tested though, there is a range of robabilities [ 0, 1] for which ACS outerforms ACS. The critical robability 0 at which this haens deends on the roblem. The reason why ACS does not always erform better than ACS is clear if we consider two asects: the comlexity (seed) of ACS versus ACS, and

8 890 Leonora Bianchi, Luca Maria Gambardella, and Marco Dorigo 0.05 eil76.ts 0 (<E_acs> - <E_acs>) / <E_acs> Fig. 4. Relative erformance of ACS versus ACS for the eil76.ts instance of the TSPLIB. Error bars reresent the average deviation of the result over 5 indeendent runs of the algorithms. the goodness of the PTSP objective function aroximation as aroaches 1. When is near to 1, a good solution to the TSP is also a good solution to the PTSP; therefore, ACS, which erforms more iterations than ACS, has a better chance to find a good solution. 4.4 Absolute Performance For the PTSP instances we tested, the otimal solution is not known. Therefore, an absolute erformance evaluation of the ACS heuristic can only be done against a theoretical lower bound, when this is available and tight enough. A lower bound to the otimal solution would give us an uer bound to the error erformed by the ACS heuristic. In fact, if LB is the lower bound and E[L ] is the otimal solution, then by definition we have E[L ] LB. (11) If the solution value of ACS is E[L ], then the following inequality holds for the relative error E[L ] E[L ] E[L ] LB. (12) E[L ] LB For the homogeneous PTSP and for instances where the otimal length L TSP of the corresonding TSP is known, it is ossible to use the following lower bound to the otimal exected length [6] LB = L TSP (1 (1 ) n 1 ). (13)

9 An Ant Colony Otimization Aroach eil51.ts eil76.ts kroa100.ts ch150.ts d198.ts uer bound of relative % error Fig. 5. Uer bound of relative ercent error of ACS for 5 TSPLIB instances. The horizontal axis reresents the customer robability. If we ut this lower bound into the right side of equation (12), we obtain an uer bound of the relative error of ACS. Fig. 5 shows the absolute erformance of ACS, evaluated with this method, for a few TSPLIB instances. From the figure we see that, for instance, ACS finds a solution within 15% of the otimum for a homogeneous PTSP with customers robability Conclusions and Future Work In this aer we investigated the otentialities of ACO algorithms for the PTSP. In articular, we have shown that the ACS algorithm is a romising heuristic for homogeneous TSP instances. Moreover, for customers robabilities close to 1, the ACS heuristic is a better alternative than ACS. At resent we are investigating the heterogeneous PTSP, for different robability configurations of customers. This is an interesting direction of research, since it is closer to a real-world roblem than the homogeneous PTSP. We are also trying to imrove ACS erformance by inserting in the ants tour construction criterion information about the customers robabilities. Work which will follow this aer also comrehends a comarison of ACS with resect to other PTSP algorithms. Moreover, ACS should be imroved by adding to the tour construction hase a local search algorithm. The best choice and design of such a local search is also an interesting issue for the PTSP. Acknowledgments This research has been artially suorted by the Swiss National Science Foundation roject titled On-line fleet management, grant 16R10FM, and by the

10 892 Leonora Bianchi, Luca Maria Gambardella, and Marco Dorigo Metaheuristics Network, a Research Training Network funded by the Imroving Human Potential rogramme of the CEC, grant HPRN-CT The information rovided in this aer is the sole resonsibility of the authors and does not reflect the Community s oinion. The Community is not resonsible for any use that might be made of data aearing in this ublication. Marco Dorigo acknowledges suort from the Belgian FNRS, of which he is a Senior Research Associate. References 1. D. J. Bertsimas, P. Jaillet, and A. Odoni. A riori otimization. Oerations Research, 38: , D. J. Bertsimas. Probabilistic Combinatorial Otimization Problems. PhD thesis, MIT, Cambridge, MA, P. Jaillet. Probabilistic Traveling Salesman Problems. PhD thesis, MIT, Cambridge, MA, A. Jézéquel. Probabilistic Vehicle Routing Problems. Master s thesis, MIT, Cambridge, MA, F. A. Rossi and I. Gavioli. Asects of Heuristic Methods in the Probabilistic Traveling Salesman Problem, ages World Scientific, Singaore, D. J. Bertsimas and L. Howell. Further results on the robabilistic traveling salesman roblem. Euroean Journal of Oerational Research, 65:68 95, G. Laorte, F. Louveaux, and H. Mercure. An exact solution for the a riori otimization of the robabilistic traveling salesman roblem. Oerations Research, 42: , M. Dorigo, G. Di Caro, and L. M. Gambardella. Ant algorithms for discrete otimization. Artificial Life, 5(2): , L. M. Gambardella and M. Dorigo. Solving symmetric and asymmetric TSPs by ant colonies. In Proceedings of the 1996 IEEE International Conference on Evolutionary Comutation (ICEC 96), ages IEEE Press, Piscataway, NJ, M. Dorigo and L. M. Gambardella. Ant Colony System: A cooerative learning aroach to the traveling salesman roblem. IEEE Transactions on Evolutionary Comutation, 1(1):53 66, M. Dorigo, V. Maniezzo, and A. Colorni. The Ant System: Otimization by a colony of cooerating agents. IEEE Transactions on Systems, Man, and Cybernetics Part B, 26(1):29 41, 1996.