An ADD/DROP procedure for the capacitated plant location problem

ARTIGO TÉCNICO

An ADD/DROP procedure for the capacitated plant location problem

Claudio Thomas BornsteinI, ^*; Manoel Campêlo^II

^IEngenharia de Sistemas e Computação / COPPE Universidade Federal do Rio de Janeiro Rio de Janeiro — RJ [e-mail] ^IIDepartamento de Estatística e Matemática Aplicada Universidade Federal do Ceará Fortaleza — CE

ABSTRACT

The capacitated plant location problem with linear transportation costs is considered. Exact rules and heuristics are presented for opening or closing of facilities. A heuristic algorithm based on ADD/DROP strategies is proposed. Procedures are implemented with the help of lower and upper bounds using Lagrangean relaxation. Computational results are presented and comparisons with other algorithms are made.

Keywords: capacitated plant location problem; ADD/DROP procedures; heuristic methods; Lagrangean relaxation.

RESUMO

O problema de localização de facilidades capacitado com custos de transporte lineares é considerado. Testes exatos e heurísticas para abrir ou fechar facilidades são apresentados. Um algoritmo heurístico baseado em estratégias ADD/DROP é proposto. Os procedimentos são implementados com o auxílio de limites inferiores e superiores provenientes de relaxação lagrangeana. Resultados computacionais são apresentados e comparações realizadas com outros algoritmos.

Palavras-chave: problema de localização capacitado; procedimentos ADD/DROP; heurísticas; relaxação lagrangeana.

1. Introduction

The capacitated plant location problem (CPLP) may be formulated as:

where I is the set of possible plant locations each with a maximum capacity a_k and fixed cost f_k, J is the set of demand centers each with a demand b_j, and c_kj is the unit transportation cost between a facility k and a consumer j. The variable x_kj represents the amount sent from k to j, and y_k means locating (or not) plant k.

The CPLP is a well known combinatorial optimization problem belonging to the class of the NP-Hard problems. For large instances there may be the need of reduction tests, problem relaxation and heuristic methods. Christofides & Beasley (1983), Beasley (1988) and Barcelo et al. (1991) have used problem reduction and Lagrangean relaxation to solve the CPLP. Thizy (1994) has used Lagrangean relaxation for a variant of the CPLP. Beasley (1993) has developed a framework for producing Lagrangean heuristics for the capacitated and the uncapacitated plant location problems. A good comparison of heuristic methods for the CPLP is presented by Cornuejols et al. (1991).

Here we present a heuristic algorithm for the CPLP based on dominance criteria between fixed and variable costs. These criteria may work in an exact way or just as a heuristic method for determining the status of facilities. Jacobsen (1983) and Mateus & Bornstein (1991) have used them within ADD or DROP heuristics. Holmberg & Ling (1995) have applied an ADD heuristic for the CPLP with staircase costs. More recently Bornstein & Azlan (1998) have used these ideas within the simulated annealing framework.

In the next section we present reduction tests and heuristics based on dominance criteria. The tests and heuristics are well known from literature. However, we combine them in a new way producing the algorithm proposed in section 3. We use lower and upper bounds for the cost increments in order to get an efficient implementation. These bounds can be obtained by Lagrangean relaxation (see Azlan, 1995).

In section 4 we present computational results. A comparison with other known algorithms is made. Finally, the last section points out to some conclusions.

2. Dominance Criteria

The dominance criteria between fixed and variable (transportation) costs are often used in reduction tests, cutting down the original size of the problem by an a priori opening or closing of plants. They are also useful to develop heuristics. In order to state the tests and heuristics, we consider the unified approach presented in Bornstein & Azlan (1998).

Given a subset of the potential plant locations, let

represent the solution value of the transportation problem T(K) associated with the CPLP, where

If we consider

The function w( ) has the important property of supermodularity (Proposition 10 in Wolsey, 1983). A function w( ) is said to be supermodular (or equivalently -w( ) submodular) if (see Wolsey, 1983 and Nemhauser & Wolsey, 1988). Thus, supermodularity is a kind of concavity.

For evaluate the increase/decrease in transportation costs if we close/open facility i. Furthermore, let evaluate the balance between fixed and variable costs with respect to facility i. The function gives rise to criteria for opening or closing plant i.

Let be the sets of closed and opened plants, i.e., Facilities whose status is still undefined remain in We can state the following rules (see Akinc & Khumawala, 1977 and Mateus & Bornstein, 1991):

Add-test: If . Drop-test: If .

Starting with we begin using the Add-test, assigning plants to is big enough allowing the evaluation of , we may use the Drop-test, allocating plants to . The optimality of the decision is assured by supermodularity of w(). Increasing set decreases set leading to non-increasing values of . Similarly, increasing yields non-decreasing values of .

The following figure illustrates how the Add- and Drop- tests work. The arrows show how move while the tests are used.

If we get closing the gap meaning that the use of the Add- and Drop- tests leads to an optimal solution of the CPLP. Unfortunately this seldom happens. There is usually a situation where In order to proceed, the following heuristics may be used (see Jacobsen, 1983):

Now supermodularity does not guarantee optimality anymore. Moreover, for a plant i satisfying the rules above allow both Trying to escape from this dilemma we look for a plant giving minor max . The former seems to be the best one to be opened whereas the latter seems to be the most advantageous one to be closed.

One of the difficulties with the Add- and Drop- tests and heuristics is the requirement of solving a great number of transportation problems, which may be very time consuming. The use of approximations for is an alternative for efficient implementation. Using the dual solutions of problems and , Azlan (1995) applies Lagrangean relaxation to calculate lower bounds . Thus, upper and lower bounds are obtained (see also Bornstein & Azlan, 1998).

3. The Algorithm

The algorithm uses the dominance criteria studied in the previous section. It starts with and repeatedly applies the tests and heuristics defined above. The algorithm consists of two phases. Phase 1 defines the status of all plants, ending with . Plants are assigned to , where is the set of temporarily closed plants. Phase 2 refines the solution reexamining the plants placed in . The number of plants assigned to depends on a parameter e. Smaller values of e result in a greater number of plants to be only temporarily closed.

Phase 1 begins with the Add-test being used at step 2. Facilities still remaining in K₂ have leading to the use of the Drop-heuristic. At step 4 facility s with maximum is chosen. For greater accuracy the exact value is calculated at step 5 and examined at step 6. If (this is possible because now we consider the exact value) then we open facility s. If then we put s in . Finally if we close facility s definitively. Since the Drop-heuristic does not work in an optimal way, plants are definitively closed only if the gain is very evident.

The convergence of phase 1 can be shown very easily. There are two loops. The first consists of steps 1 to 6 and the second consists of steps 3 to 6. At each iteration of phase 1 one of the two loops is executed and a facility is discarded from set K₂ at step 6. This finishes emptying set K₂ leading to phase 2.

Phase 2 uses the Add-heuristic meaning the evaluation of or its lower bound . Therefore we need big enough to guarantee . Let us show that this will always happen at the end of phase 1. One should just consider that at step 6 an element s will only be removed from . This means . Thus, we never get . As phase 1 finishes with we can assure that .

At phase 2 each plant in is reexamined. It may be opened, interchanged with a plant from or closed definitively. The procedure can be divided in two main parts. Part 2.1 consists of steps 7 to 11 and uses the Add-heuristic. At step 9 we choose the best plant (plant r) from to be opened. If it is opened at step 11 according to the Add-heuristic. Otherwise, we proceed to Part 2.2 (steps 12 to 17) which embodies an interchange heuristic. We examine whether the opening of plant r can still be advantageous by simultaneously closing a plant from . In case of disadvantage plant r is closed definitively.

Let us examine the interchange heuristic with a little more detail. Plants that will be further examined in order to be closed are placed in . Initially, at step 12, we make . At step 13 we choose plant with the largest upper bound . The next steps compare the cost increase with the cost reduction .

At step 14 if then an interchange between r and s will not lead to an improvement. Neither will an interchange between r and any other plant from , because s has a maximum upper bound. Thus, and plant r can be definitively closed.

Otherwise and a refinement is necessary. We calculate the exact value at step 15. If we still have the interchange is accomplished at step 17, opening plant r and closing plant s.

Finally, means that although an interchange with s does not improve the solution, another plant in could do it. Plant s is discarded from and a new attempt is made. The result is the loop between steps 13 and 16.

With respect to the convergence of phase 2 let us first consider the innermost loop which runs from step 13 to step 16. Each time this loop is executed one element is discarded from . In the worst case we get , exiting the loop at step 13. The other loops all include step 9 where one facility r is discarded from set . We finish by getting bringing phase 2 to an end.

At this point a general comment relating to upper and lower bounds should be made. They are only used where a great number of transportation problems would have to be solved otherwise. Where plants are opened or closed, even temporarily closed, exact values are calculated.

4. Computational Results

The algorithm was implemented in FORTRAN 77 using a transportation procedure based on Ahrens & Finke (1980). When solving a transportation problem we generally start from the solution of the problem solved previously. One should not forget that most problems differ from the previous one in a single closed or opened plant.

With respect to upper and lower bounds, is evaluated following Bornstein & Azlan (1998) and is calculated according to in Mateus & Bornstein (1991). The value of e was fixed in 0.05.

The computational results mainly compare the algorithm proposed in the previous section with the algorithms presented by Bornstein & Azlan (1998), Van Roy (1986) and Beasley (1993). The first one also builds on dominance criteria, but uses these ideas within a framework of simulated annealing. The cross decomposition algorithm proposed by Van Roy (1986) is quite effective for small problems (see Beasley, 1988). Finally, we compare the proposed algorithm with Lagrangean heuristics which have proved to work quite well even with large problems. The comparison is made with an algorithm due to Beasley (1993).

The proposed algorithm (PA) and the SAE-best algorithm considered by Bornstein & Azlan (1998) were run on a PC 486, 100 MHz. Results of the cross decomposition algorithm were extracted from Table I in Van Roy (1986). They were generated on a IBM 3033U. The Lagrangean heuristic in Beasley (1993) was run on a Cray X-MP/28 using CFT compiler with maximum optimization.

Table 1 presents the computational results for Kuehn & Hamburger (1963) test problems. The data was obtained from Beasley's (1990) OR-Library. Problems IV-VI are 16 x 50, problems VIII and IX are 25 x 50 and problems XI and XII are 50 x 50. The first number relates to potential plant locations and the second to the number of demand centers. Van Roy (1986) uses a different numbering.

The second column of Table 1 shows that the relative difference between the solution value obtained by PA and the optimal value was always under 0.4%. For 22 out of 25 problems, i.e. 88%, the optimal solution was found. Times were always under 0.55 seconds.

Although numbers from Table 1 give an idea about the efficiency of each algorithm, comparisons have to be made very carefully. To measure the quality of the solution, PA, Bornstein & Azlan (1998) and Beasley (1993) use the relative difference (ERR%) between the solution value obtained by the algorithm and the optimal value. However, for Bornstein-Azlan, ERR% is calculated taking the best solution of a 5-run of each problem. Considering an average value instead of the best value for the 5-run, shifts the average ERR% for the 25 problems from 0.03% to 0.13% ! On the other hand, Van Roy (1986) evaluates his solutions using the relative duality gap (GAP%) which only provides an upper bound for ERR%. The fact that Van Roy does not give the value of the generated solutions unfortunately makes it impossible to calculate ERR% for his results at Table 1.

With respect to times only the first two algorithms can effectively be compared, since similar machines were used. The SAE-best algorithm presented by Bornstein & Azlan (1998) expends much more time than PA. Also Beasley (1993) and Van Roy (1986) are computationally more expensive than PA. Although times of the first two algorithms are almost equivalent to times spent by PA, one has to consider that they were run on more powerful machines. One should also take into account that both Beasley (1993) and Van Roy (1986) provide upper and lower bounds for the optimal solution.

Summing up, a good trade-off seems to exist for PA between the quality of the solution and the computational time. This is what should be expected from a heuristic.

To test the performance of PA for larger instances, we considered 100 x 100 randomly generated problems available at Beasley's (1990) OR-Library. Table 2 compares PA only with Beasley (1993) because for the other algorithms data was not available for problems of this dimension. The relative error of PA was always less than 0.75%, but for only 4 out of 12 problems the optimal solution was obtained. Times were less than one minute. Again, PA obtains solutions with quality similar to those obtained by Beasley (1993). Times, however, are much smaller.

5. Conclusions

A central dilemma of the CPLP is the conflict between the minimization of fixed and variable costs each driving the solution in a different direction. The minimization of variable/transportation costs in the extreme case forces the solution towards putting a plant near each demand center i.e. represents a tendency towards decentralization. On the other hand, due to economies of scale, fixed costs are generally minimized by placing a central unit which supplies all demand centers. Of course, the joint minimization of fixed and variable costs means that there needs to be a compromise between the centralization and decentralization tendencies.

The algorithm presented in this paper moves around this dilemma, detecting dominances among the two types of costs and locating plants according to these dominances. The computational results show that it is able to tackle large scale problems finding almost always near optimal solutions at very low cost.

However not everything is as positive as it may seem. The Achilles' heel is the shortsightedness of the algorithm, i.e. the tendency to act only locally, determining whether each plant should be opened or not sequentially. One should not forget that the dilemma mentioned above is a global question where all the complex interrelations between the various plants and demand centers play an important role. To counterbalance this tendency the algorithm has at phase 2 an interchange heuristic which tries to detect relations between plants.

Detecting dominances locally may lead to distortions with respect to determining a global optimal solution. These distortions may heavily rely on data. This is the reason for introducing parameters which may correct these distortions. In the sense of greater transparency only one parameter was introduced in the algorithm. Greater values of e make it easier to close plants at phase 1. Closed plants are not further examined at phase 2. On the other hand smaller values of e result in putting plants in instead of . Plants in will be examined again at phase 2 and are object of the interchange heuristic. As a general principle one should have in mind that decreasing e generally leads to better solutions at a higher computational cost. For the set of tested problems, using e equal to 0.05 seems to lead to the best value solution, i.e. leads to the best possible solution within a reasonable computational cost. Decreasing e further down may only lead to a very slight improvement of the solution at a much higher computational cost.

Among the results of the PA, Bornstein-Azlan, Van Roy and Beasley's algorithms, only the first two can be compared effectively due to the fact that similar machines were used. Van Roy and Beasley's algorithms were run on more powerful machines. Results for Bornstein-Azlan were similar to those obtained by PA at a much more expensive computational cost. Although average percentage errors presented at Table 1 for Bornstein-Azlan are slightly better than those of PA, one should not forget that the former considers the best solution of a 5-run for each problem. When considering an average value for the 5-run, the performance of PA is much better than the performance of Bornstein-Azlan. For the larger instances of Table 2, PA presents similar results to those obtained by Beasley's algorithm. However PA results in a lower computational cost.

Acknowledgments

The authors are partially supported by CNPq (grants nr. 523158/94-7 and 300251/00-9) and FUJB (grant nr. 5561-1).

Recebido em 06/2003; aceito em 10/2003

Received June 2003; accepted October 2003

• (1) Ahrens, J.H. & Finke, G. (1980). Primal Transportation and Transshipment Algorithms. Zeitschrift fur Operations Research, 24, 1-32.
• (2) Akinc, U. & Khumawala, B.M. (1977). An Efficient Branch and Bound Algorithm for the Capacitated Warehouse Location Problem. Management Sci., 23, 585-594.
• (3) Azlan, H.A.B. (1995). Aplicação do Algoritmo Simulated Annealing ao Problema de Localização Capacitado, Usando Testes de Redução (The Use of Simulated Annealing for the Capacitated Plant Location Problem Using Reduction Tests). D.Sc. Thesis, COPPE/UFRJ, Rio de Janeiro.
• (4) Barcelo, J.; Fernandez, E. & Jörnsten, K.O. (1991). Computational Results from a New Lagrangean Relaxation Algorithm for the Capacitated Plant Location Problem. European J. Operational Research, 53, 38-45.
• (5) Beasley, J.E. (1988). An Algorithm for Solving Large Capacitated Warehouse Location Problems. European J. Operational Research, 33, 314-325.
• (6) Beasley, J.E. (1990). OR-Library: distributing test problems by electronic mail. J. Operational Research Soc., 41, 1069-1072.
• (7) Beasley, J.E. (1993). Lagrangean Heuristics for Location Problems. European J. Operational Research, 65, 383-399.
• (8) Bornstein, C.T. & Azlan, H.B. (1998). The Use of Reduction Tests and Simulated Annealing for the Capacitated Plant Location Problem. Location Sci., 6, 67-81.
• (9) Christofides, N. & Beasley, J.E. (1983). Extensions to a Lagrangean Relaxation Approach for the Capacitated Warehouse Location Problem. European J. Operational Research, 12, 19-28.
• (10) Cornuejols, G.; Sridharan, R. & Thizy, J.M. (1991). A Comparison of Heuristics and Relaxations for the Capacitated Plant Location Problem. European J. Operational Research, 50, 280-297.
• (11) Holmberg, K. & Ling, J. (1995). A Lagrangean Heuristic for the Facility Location Problem with Staircase Costs. Technical Report LiTH-MAT-R-1995-24, Linköping University, Sweden.
• (12) Jacobsen, S.K. (1983). Heuristics for the Capacitated Plant Location Problem. European J. Operational Research, 12, 153-261.
• (13) Kuehn, A.A. & Hamburger, M.J. (1963). A Heuristic Program for Locating Warehouses. Management Sci., 9, 643-666.
• (14) Mateus, G.R. & Bornstein, C.T. (1991). Dominance Criteria for the Capacitated Warehouse Location Problem. J. Operational Research Soc., 42, 145-149.
• (15) Nemhauser, G.L. & Wolsey, L.A. (1988). Integer and Combinatorial Optimization. J. Wiley & Sons.
• (16) Thizy, J.M. (1994). A Facility Location Problem with Aggregate Capacity. INFOR, 32, 1-18.
• (17) Van Roy, T.J. (1986). A Cross Decomposition Algorithm for Capacitated Facility Location. Operations Research, 34, 145-163.
• (18) Wolsey, L.A. (1983). Fundamental Properties of Certain Discrete Location Problems. In: Locational Analysis of Public Facilities [edited by J.-F. Thisse & H.G. Zoller], North-Holland, 331-355.

*

Corresponding author / autor para quem as correspondências devem ser encaminhadas

Publication Dates

• Publication in this collection
  04 June 2004
• Date of issue
  Apr 2004

History

• Accepted
  Oct 2003
• Received
  June 2003