Solving the Capacitated Single Allocation Hub Location Problem Using Genetic Algorithm Faculty of Mathematics University of Belgrade Studentski trg 16/IV 11 000, Belgrade, Serbia (e-mail: zoricast@matf.bg.ac.yu) Zorica Stanimirović Abstract. The aim of this study is to present a Genetic Algorithm (GA) for solving the Capacitated Single Allocation Hub Location Problem (CSAHLP) and to demonstrate its robustness and effectiveness for solving this problem. The appropriate objective function is correcting infeasible individuals to be feasible. These corrections are frequent in the initial population, so that in the subsequent generations genetic operators slightly violate the feasibility of individuals and the necessary corrections are rare. The solutions of proposed GA method are compared with the best solutions presented in the literature by considering various problem sizes of the AP data set. Keywords: Evolutionary computation, genetic algorithms, capacitated hub location problems, discrete location. 1 Problem formulation Hub networks are often used in transportation and telecommunication networks where traffic (such as mail, telecommunication packets or airline passengers) must be transported from an origin to a destination point, but where it is expensive or impractical to use direct origin-destination links. In order to facilitate consolidation of traffic, hubs can be used as intermediate switching points. When designing a hub network, different types of constraints may be involved. For example, the number of hubs to be located may be fixed to p, the allocation of non-hub nodes may be either to a single hub (single allocation scheme) or to multiple hubs (multiple allocation scheme), different types of capacity restrictions on hubs may be assumed, as well as fixed costs of establishing hubs. An exhaustive survey of hub location problems and their classification can be found in [Campbell et al., 2002]. The goal of CSAHLP is to locate a set of hub nodes and to allocate the non-hub nodes to the hubs from the chosen set so as to minimize the sum of transportation costs between origin-destination pairs and the fixed costs of establishing hubs. The CSAHLP also assumes a single allocation scheme and limited amount of flow collected at each hub. The problem is NPcomplete, since its subproblem, the Ucapacitated Single Allocation Hub Location Problem-USAHLP, is proved to be NP-hard in [Kara and Tansel, 1998].
2 Stanimirović Z. In the literature, CSAHLP has only been considered by Ernst and Krishnamoorthy in [Ernst and Krishnamoorthy, 1999]. The authors presented new mixed integer LP formulation for the CSAHLP and described two heuristic algorithms for its solution based on simulated annealing (SA) and random descent (RDH). The upper bounds obtained by the heuristics are used in an LP-based branch-and bound method, which provides optimal solutions for smaller size problem instances with n 50 nodes. For realistic sized problems n = 100, 200 that could not be solved exactly, the proposed RDH and SA heuristics provided solutions in a reasonable amount of computer time. The formulation of the CSAHLP from [Ernst and Krishnamoorthy, 1999], which is used in this study, has the following notation: C ij = distance between nodes i and j (in metric sense); W ij = the amount of flow (number of units of flow) between an origin-node i and destination-node j; Γ k = the collection capacity of hub k; F k = the costs of establishing hub k; O i = the amount of flow that departs from node i; i.e. O i = n j=1 W ij; D j = the amount of flow that is distributed to node j, i.e. D j = n i=1 W ij; χ, α, δ =parameters that reflect the unit rates (costs) for a collection (originhub), transfer (hub-hub), and distribution (hub-destination) respectively. The decision variables Z ij {0, 1} have value 1 if node i is allocated to a hub node j, 0 otherwise (Z kk = 1 implies that the node k is a hub), while represent the amount of flow that is originated from node i, collected at hub k and distributed via hub l. Using the notation mentioned above, the problem can be written as: Ykl i n n n n n min C ik Z ik (χo i + δd i ) αc kl Ykl i + i=1 k=1 with constraints: j=1 i=1 k=1 l=1 n F k Z kk (1) k=1 n Z ik = 1 for every i = 1,..., n (2) k=1 Z ik Z kk for every i, k = 1,..., n (3) n n n W ij Z jk + Ykl i = Ylk i + O i Z ik for every i, k = 1,..., n (4) l=1 l=1 n O i Z ik Γ k Z kk ; for every k = 1,..., n (5) i=1 Y i kl 0 for every i, k, l = 1,..., n (6) Z ik {0, 1} for every i, k = 1,..., n. (7) The objective function (1) minimizes the sum of transportation cost between all origin-destination pairs via hub nodes and the fixed costs of locating the set of hubs. Constraint (2) states that each node is allocated to exactly one hub,
Genetic Algorithm for the CSAHLP 3 while constraint (3) enforces that flow is only sent via open hubs, preventing direct transmission between non-hub nodes. Constraints (4) represents flow conservation equality in the network. The amount of flow that is collected in a hub is limited by (5). Finally, constraints (6) and (7) specify variables Ykl i and Z ij to be non-negative and binary respectively. 2 Proposed GA Since each hub problem has its own specific structure (objective function, decision variables and constraints), there is no general solution approach for solving all hub problems, or at least a smaller group of them. Few additional constraints or a slight modification of the problem structure can substantially change the computational behavior of the designed solution approach. Exact methods can not provide solutions for large-scale hub location problems, which arise from practice, in a reasonable amount of time. Therefore, genetic algorithms, as robust heuristic methods (see [Back et al., 2000]) are very promising approaches for solving hub location problems. Some successful applications of the GA for hub location problems can be found in the literature: [Abdinnour-Helm, 1998], [Topcuoglu and et al., 2005], [Kratica et al., 2006], etc. Representation of individuals: Genetic code of an individual consists of n genes, each referring to one network node. First bit in each gene takes value 1 if the current node is located hub, 0 if not. Considering these bit values, the array of opened hub facilities is formed. Remaining bits of the gene are referring to the hub that is assigned to the current node. Hub nodes are assigned to themselves. For each non-hub node, the array of located hub facilities is created and arranged by non-decreasing order of their distances from the current node. This strategy, named nearest neighbour ordering, ensures that closer hubs have higher priority than the distant ones in assigning them to non-hub nodes. For example, genetic code: 00 10 10 02 10 corresponds to the following solution: the first bits in each gene (0, 1, 1, 0, 1) denote established hubs (nodes 1, 2 and 4), while the remaining bits of the genes (0, 0, 0, 2, 0) show assignments. Non-hub node 0 is assigned to its closest established hub, while node 3 is assigned to the third hub from the corresponding array of established hubs. Hub nodes 1, 2 and 4 are obviously assigned to themselves. Objective function: The indices of established hubs are obtained from the first bits of each gene. For each non-hub node, the array of established hubs is arranged in non-decreasing order with respect to distances from the current node. The index of a hub that is assigned to the current non-hub node is obtained from the remaining part of the gene (if its value is r, the r-th hub is taken from the previously arranged array). Arranging the array of established hubs is performed for each individual in every generation
4 Stanimirović Z. After the assigning procedure described above has been performed, the objective value is simply evaluated only by summing distances origin-hub, hub-hub and hub-destination, multiplied with flows and corresponding parameters χ, α, δ and by adding the sum of fixed costs for established hubs. It may happen that a non-hub node is allocated to a hub whose remaining capacity is not enough to satisfy the node s demand. In this case, the next hub from the array of established hubs for the current node that satisfies the capacity constraint is taken. If there is no such a hub, we consider the individual infeasible by setting its fitness to 0. This case is very rare in practice and it usually happens if the sum of hub capacities is less than the overall flow in the network. So, the infeasible individuals (in the sense of insufficient hub capacities) will be generated in the initial population with very small probability. The applied strategy of correcting individuals with insufficient capacities to feasible ones may slightly affect the quality of the GA solution, but it preserves the diversity of the genetic material. If all the infeasible individuals were encountered in the population, they may become dominant in the following generations and the algorithm might provide no solution or finish in a local optimum. If the incorrect individuals were excluded from the population, the possibility of premature convergence would rapidly increase. Genetic operators: The GA method uses an improvement of standard tournament selection, named fine-grained tournament selection - FGTS. It is used in cases when the average tournament size F tour is desired to be fractional (see [Filipovic, 2003]). In this GA implementation F tour = 5.4. After a pair of individuals is selected, modified one-point crossover operator is applied to them producing two offsprings. A bit position i (crossover point) is randomly chosen in the genetic code and whole genes are exchanged starting from the gene that contains the chosen crossover point. Crossover is performed with the rate p cross = 0.85, which means that around 85% individuals take part in producing offsprings. Offsprings generated by crossover operator are subject to modified twolevel mutation with frozen bits (see [Kratica et al., 2006]). Basic mutation rates used in the GA implementation are: 0.4/n for the bit on the first position in a gene, 0.1/n for the bit on the second position in a gene, while the following bits have repeatedly two times smaller mutation rate (0.05/n, 0.025/n,...). The appearance of frozen bits during the GA generations may increase the possibility of premature convergence significantly ([Kratica et al., 2006]). Therefore, comparing to basic mutation rates, frozen bits are muted with: 2.5 times higher rate (1.0/n instead of 0.4/n) if they are positioned at the first place of the gene and 1.5 times higher rate (0.075/n, 0.0375/n,...) otherwise. Note that smaller values of mutation rate and frozen factor for the remaining part of the gene are used, because it is important to have many zeros there (each zero corresponds to the closest hub facility for particular non-hub node).
Genetic Algorithm for the CSAHLP 5 Other GA characteristics: The initial population numbers 150 individuals. Each individuals of the initial population is generated with following strategy: the first bit in each gene takes value 1 with 5/n probability; the second bit in each gene is generated with 2.5/n probability and the following bits take value 1 with two times smaller probability than the previous ones (1.25/n, 0.625/n,...). Since closer hubs for each non-hub node are favoured, it is desirable that the second segment in initial genetic code contains many zeros. One third of the population is replaced in every generation, except the best 100 individuals that are directly passing into the next generation. These elite individuals preserve highly fitted genes of the population. Their objective values are calculated only in the first generation. The infeasible individuals (in the sense of insufficient hub capacities) in the initial population are corrected to be feasible. The applied genetic operators preserve their feasibility, so the infeasible individuals do not appear in the following generations. If an individual with same genetic code appears again in population, its objective value is set to zero, which prevents it to enter the next generation. The appearance of individuals with the same objective value, but different genetic codes is limited to constant N rv = 40. Described strategy helps in preserving the diversity of genetic material and in keeping algorithm away from local optima. The running time of GA is also improved by caching technique (see [Kratica, 1999]). The number of cached objective values is limited to N cache = 5000 in GA 3 Computational results The proposed GA approach was tested on the AP hub instances [Beasley, 19996], with n 200 nodes and parameters χ = 3, α = 0.25 and δ = 2. Two types of capacities and fixed costs on the nodes are assumed: tight (T) and loose (L), which gives four types of problems LL, LT, TL and TT for each problem size n. The GA was coded in C programming language and run on an AMD Athlon K7/1.33GHz with 256 MB RAM memory. On each AP instance, the GA method was run 20 times. The maximal number of GA generations N gen = 5000 is used as a stopping criterion. The GA also terminates if the best individual or the best objective value remained unchanged through N rep = 2000 successive generations. On all instances that were tested, these stopping criterions allowed the GA to converge to high quality solutions. The results of the proposed GA on smaller size AP instances with n 50 nodes are presented in Table 1, while Table 2 contains results on the larger AP instances n = 100, 200. The columns of Table 1 and Table 2 contain following data: instance s dimension, fixed cost and capacity type; optimal solution for current instance (only in Table 1) obtained with LP-based branch and bound method - BnB [Ernst and Krishnamoorthy, 1999] (for instance 50TT
6 Stanimirović Z. Inst. Opt.sol. GA name LP BnB Bestsol. t[s] t tot[s] gen gap[%] σ[%] eval cache 10LL 224250.055 opt 0.019 0.940 2032 0.128 0.574 35547.6 65.1 10LT 250992.262 opt 0.064 1.105 2125 0.000 0.000 49641.4 53.3 10TL 263399.943 opt 0.127 1.067 2255 0.120 0.446 47350.6 58.3 10TT 263399.943 opt 0.028 0.967 2051 0.046 0.137 45390.3 55.8 20LL 234690.963 opt 0.043 2.169 2035 0.000 0.000 66000.7 35.3 20LT 253517.395 opt 0.456 2.406 2461 0.074 0.227 75465.8 37.9 20TL 271128.176 opt 0.451 2.565 2438 0.000 0.000 78903.6 35.6 20TT 296035.402 opt 0.211 2.160 2205 0.297 0.415 53469.3 51.5 25LL 238977.95 opt 0.402 3.066 2288 0.779 1.221 76972.7 32.9 25LT 276372.5 opt 0.127 2.969 2077 0.980 0.562 67922.3 34.8 25TL 310317.64 opt 0.251 3.077 2174 2.247 2.300 72906 33 25TT 348369.15 opt 0.548 3.003 2425 0.845 0.526 64927.9 46.5 40LL 241955.71 opt 0.226 5.266 2082 0.082 0.214 73973.1 29.1 40LT 272218.32 opt 1.15 6.463 2394 3.538 2.686 76834.2 35.9 40TL 298919.01 opt 0.399 5.619 2146 0.000 0.000 78009 27.5 40TT 354874.10 356509.86 0.842 5.750 2321 4.249 3.113 71647.6 38.3 50LL 238520.59 opt 0.521 7.448 2139 0.267 0.652 76537.9 28.6 50LT 272897.49 opt 1.066 8.534 2269 0.518 1.009 80235.6 29.5 50TL 319015.77 opt 1.875 8.866 2526 0.720 0.934 92312.4 26.8 50TT 417440.99* 422794.56 4.348 10.981 3100 2.793 1.591 110612.1 28.7 Table 1. GA results on smaller AP instances only the best value of BnB method is presented, since BnB found no optimal solution in this case); the best value of GA (Best.sol) with mark opt in cases when GA reached optimal solution; average time t (in seconds) needed to detect the best GA value; total time t tot (in seconds) needed for finishing GA; the average number of generations gen; a percentage agap = 1 20 20 i=1 gap i where gap i = 100 soli Opt.sol Opt.sol is evaluated with respect to the optimal solution Opt.sol, or the best-known solution Best.sol, i.e. gap i = 100 soli Best.sol Best.sol in cases where no optimal solution is found (sol i represents the GA solution obtained in the ith execution); standard deviation of the average gap σ = 1 20 20 i=1 (gap i agap) 2 (in percent); the average number of evaluations eval; savings achieved by using caching technique cache (in percent). GA concept cannot prove optimality and adequate finishing criteria, that will fine-tune solution quality, does not exist. Therefore, as column t tot in Table 1 and Table 2 shows, our algorithms run through additional t tot t time (until finishing criteria is satisfied), although they already reached optimal (best) solution. As it can be seen from Table 1, the proposed GA approach quickly reaches all optimal solutions on AP instances with n 50 nodes in less than
Genetic Algorithm for the CSAHLP 7 Inst. Best.sol. t[s] t tot[s] gen gap[%] σ[%] eval cache 100LL 246713.97 5.300 56.052 5520 1.071 0.914 193743.9 29.3 100LT 256207.52 29.020 81.965 7814 3.194 1.791 279284.7 27.5 100TL 364515.24 15.184 66.293 6484 0.773 2.158 246441.4 23.8 100TT 475156.75 24.164 75.963 7298 4.391 4.442 270017.9 25.8 200LL 241992.97 168.966 424.517 8295 0.698 1.323 340043 18 200LT 270202.25 142.640 410.405 7588 1.864 2.227 302818.1 20.2 200TL 273443.81 80.872 325.878 6621 3.63 2.96 246438 25.2 200TT 291830.66 195.174 427.568 8968 0.81 0.8 349444.5 22.1 Table 2. GA results on large AP instances 1.875 seconds. The exception is the instance 40TT where GA has 4.249% average gap from the optimal solution. For instance 50TT, where no optimal solution is known in advance, GA has 2.793% average gap from the best known solution obtained by BnB method. In Table 3, a comparison of the Inst. RDH and SA heur. GA The best name Best. sol. of RDH or SA Alpha 200 MHz (sec) Best. sol. AMD 1.33 GHz (sec) method 100LL 246713.97 18.55 246713.97 56.052 same 100LT 256638.38 24.71 256207.52 81.965 GA 100TL 362950.09 30.16 364515.24 66.293 SA 100TT 474680.32 34.83 475156.75 75.963 SA 200LL 241992.97 136.01 241992.97 424.517 same 200LT 268894.41 437.21 270202.25 410.405 RDH 200TL 273443.81 195.28 273443.81 325.878 same 200TT 292754.97 190.12 291830.66 427.568 GA Table 3. Comparisons on large AP instances results for eight large AP instances obtained by proposed GA method and RDH and SA heuristics [Ernst and Krishnamoorthy, 1999] is presented. The second column of Table 3 contains better solution of both SA and RDH for the current AP instance. The third column shows the DEC 3000/700 (200MHz) computational time for which the SA/RDH heuristic obtained the corresponding solution. In the next two column the best GA solution and corresponding AMD (1.33GHz) computational time are presented. The last column specifies which of the three heuristic methods gave the best overall solution for the current AP instance. As it can be seen from Table 3, the proposed GA gave better solution in comparison with other two heuristics in two cases, the SA also provided better solution on two AP instances, while the RDH was better than other methods in one case. In remaining three cases all three heuristics gave the same solution.
8 Stanimirović Z. 4 Conclusions In this paper, a new genetic algorithm has been introduced to the CSAHLP. Applied objective function ensures that infeasible individuals do not appear in the generations of the GA. Arranging located hubs in non-decreasing order of their distances for each non-hub node directs GA to promising search regions. By using mutation with frozen bits, the diversibility of genetic material has been increased. For the same reason, the number of individuals with same objective function value, but different genetic code is limited. Caching technique additionally improves computational performance of the GA implementation. Computational experiments reveal that the performance of the proposed GA is quite satisfactory. For smaller AP instances, the GA is highly effective in finding high quality solutions that match with the optimal ones in a very short time. Its performance on large AP instances with regard to both solution quality and computation time shows the potential of this algorithm as an useful metaheuristics for solving CSAHLP and other capacitated hub problems, as well as more complex hub location models. The future work could be also concentrated on the parallelization of the GA and its hybridization with exact methods. References [Abdinnour-Helm, 1998]S. Abdinnour-Helm. A hybrid heuristic for the uncapacitated hub location problems. European Journal of Operational Research, 106:489 499, 1998. [Back et al., 2000]T. Back,, and et al. Basic algorithms and operators. In Evolutionary Computation 1, 2000. [Beasley, 19996]J.E. Beasley. Obtaining test problems via internet. Computers and Artificial Intelligence, 18:429 433, 19996. [Campbell et al., 2002]J.F. Campbell,, and et al. Hub location problems. In Drezner Z. and Hamacher H., editors, Facility Location: Applications and Theory, pages 373 407, 2002. [Ernst and Krishnamoorthy, 1999]A.T. Ernst and M. Krishnamoorthy. Solution algorithms for the capacitated single allocation hub location problem. Annals of Operations Research, 86:141 159, 1999. [Filipovic, 2003]V. Filipovic. Fine-grained tournament selection operator in genetic algorithms. Computing and Informatics, 22:143 161, 2003. [Kara and Tansel, 1998]B.Y. Kara and B.C. Tansel. On the allocation phase of the p-hub location problem. Technical Report, Dpt. of Ind. Eng., 1998. [Kratica et al., 2006]J. Kratica,, and et al. Two genetic algorithms for solving the uncapacitated single allocation p-hub median problem. European Journal of Operational Research (to be published), 2006. [Kratica, 1999]J. Kratica. Improving performances of the genetic algorithm by caching. Computers and Artificial Intelligence, 18:271 283, 1999. [Topcuoglu and et al., 2005]H. Topcuoglu and et al. Solving the uncapacitated hub location problem using genetic algorithms. Computers and Operations Research, 32:967 984, 2005.