Sei sulla pagina 1di 9

Capacitated Continuous Facility Location Problem With Unknown Number of Facilities

Hctor J. Carlo, University of Puerto Rico Mayagez Francisco Aldarondo, University of Puerto Rico Mayagez Priscilla M. Saavedra, University of Puerto Rico Mayagez Silmarie N.Torres, University of Puerto Rico Mayagez Abstract: This study is concerned with the long-term strategic decision of simultaneously determining the number and location of new facilities to minimize the total (fixed and variable) cost of interacting with a set of existing facilities. The problem has many practical applications, including locating bus stops for public transportation. A non-linear, mixed integer mathematical formulation, a brute-force algorithm, and four heuristics are developed for the problem. Experimental results show that the proposed Greedy Search Heuristic outperforms all other heuristics considered. The proposed solution methods provide engineering managers with new tools to solve this particular facility location problem. Keywords: Free Entry Facility Location, Facility Location Problem, Transportation Management EMJ Focus Areas: Quantitative Methods & Models, Strategic Management areas, the reader is referred to ReVelle and Eiselt (2005) and Klose and Drexl (2005). The FLP problems can be classified into continuous and discrete space. The continuous FLP assumes that facilities can be located anywhere on a plane, whereas the discrete FLP assumes that new facilities will be selected among a set of pre-identified locations. Clearly, the discrete FLP has a much smaller solution space than the continuous FLP. The discrete case is very appropriate for instances where the possible real estate options are limited and one must choose among them. The continuous FLP, on the other hand, does not consider the available real estate locations; therefore, it is likely to select a facility location that is infeasible. For example, the new facilities might overlap with existing facilities or they might be located on a lake; however, the solution to the continuous FLP could be useful, for instance, to identify a region (e.g., city or neighborhood) in which the facility should be located. At that point, all feasible real estate can be considered. This study focuses on a particular version of the FLP problem, the Capacitated Continuous Facility Location Problem with Unknown Number of Facilities (CCFLPU). As the name suggests, we assume that new facilities have a maximum capacity and could be located anywhere in a continuous space. Further, the number of new facilities to select is unknown; therefore, our problem is to simultaneously determine the number of new facilities, the location of these facilities, and the magnitude of interaction between each new facility and the existing facilities so that the existing facilities demands are satisfied. Assuming an unknown number of new facilities is very common for discrete FLPs; however, it is a very uncommon assumption for the continuous FLP. Typically, continuous FLPs assume that the number of facilities m to be located is known (referred to as the m-FLP); however, the CCFLPU is of particular interest in emergent environments such as extreme events emergency management. In emergency planning the CCFLPU could be used to determine where to store first necessity items (e.g., food, water, medicine) to be distributed after an emergency, or could be used to design the gathering areas (how many and where) for a city so it is accessible to most citizens. The CCFLPU could also be useful in humanitarian logistics to determine centric locations to distribute food, water, medicines, or vaccines among inhabitants of a city. Other potential applications include locating energy sources in distributed energy generation systems to minimize the amount of energy lost in transmitting electricity, in banking to determine the number and location of ATMs, and even for strategically locating mobile stands or kiosks (e.g., hot dog or ice cream stands). In general, the CCFLPU is practical for environments where the fixed costs are low and the type of facility to be opened does not require special infrastructure. In this article we argue that the CCFLPU might also be useful for public transportation management. Some examples of the

his article focuses on two of the most important strategic management decisionsselecting the number and location of new facilities. The best location of new facilities depends on the location of interacting (existing) facilities. Typically, the objective is to minimize the total interaction cost between new and existing facilities. This interaction is most commonly measured in terms of distances. On the other hand, the best location of new facilities also depends on the number of new facilities to open. The required number of facilities, at the same time, depends on the supply and demand of existing and new facilities. The traditional Facility Location Problem (FLP) determines the best location of a set of new facilities that are to interact with a set of existing facilities. The typical objective for the FLP is to minimize the total interaction cost, typically associated with fixed (e.g., investment) and variable (e.g., transportation) costs. The FLP has been applied to different problems during the last four decades and has proven most useful for locating businesses and for marketing. Other application areas include supply chain network design (i.e., locating distribution centers, warehouses, or manufacturing plants) in order to minimize the travel cost to facilities in other echelons of their supply chain. Some less common application areas include the location of medical service facilities for large-scale emergency situations, including locating ambulances, firehouses, and hospitals, where minimizing the average response time for emergencies from all residents of cities is desired. Lastly, other application areas include locating schools, hotels, bank branches, or even the location of a church camp. For a thorough review of the existing FLP literature and application

Refereed Research Manuscript. Accepted by Special Issue Editor Long.


24 Engineering Management Journal Vol. 24 No. 3 September 2012

applicability of CCFLPU to public transportation management include determining the number and location of bus stops, train stations, or public parking areas. Consider, for example, the problem of establishing the best location for bus stops in an urban area. The CCFLPU could be used to determine the number and location of such bus stops so that the mean travel distance between residents (potential bus users) and bus stops is minimized. In this case, the number and location of such bus stops (i.e., the new facilities) should be determined. Too few bus stops would require users to walk further, while too many bus stops would require investment and recurrent operational costs. Both extremes could affect the service quality of the transportation system. For the bus stop problem, the existing locations are the residences of the potential users; however, given the complexity of identifying the location of the residence of each individual in an urban area, these locations may be aggregated and represented by one point, which could be the centroid of the location of the residences. The fixed cost for opening a new facility could be the investment cost (e.g., building the bus stop) and the incremental cost of adding a new bus stop, which could be measured in terms of economic impact. On the other hand, the variable costs would include the present value of the operational costs (e.g., maintenance). The interaction between the new and existing facilities would be the number of residents in each residence (or aggregate existing facility). This number could be multiplied by a probability of the residents using the public system to get an expected number of residents using the system. The total objective function would then be to minimize the fixed costs by the number of new facilities, plus the variable costs times the interaction costs. Notice that for the CCFLPU both the location of new facilities and the flow between these facilities are considered. Hence, when applying the CCFLPU to public transportation management, one would need to specify the capacity of a bus station. This capacity could be set to infinity (or a very large number) if desired. Given the location of bus stops serving the customers, one may then construct bus routes. Alternatively, if bus routes are already fixed, the solution to the CCFLPU could be used to determine the best number and location for bus stops within each route. Comparing methods for establishing the number and location of bus stops in a rural area is beyond the scope of this article. It is worth mentioning that engineering management professionals need to be careful regarding the implicit biases when identifying the input data and solution method as highlighted in Purvis et al. (2004). The reminder of this article focuses on the general CCFLPU, which can be easily mapped to public transportation management. The remainder of this article is organized as follows. A review of the relevant literature is presented, followed by a mathematical formulation for the CCFLPU. Next, we introduce a brute-force optimal solution algorithm for the CCFLPU based on the m-FLP, then four heuristics for the CCFLPU. Lastly, the experimental results and the conclusions are presented. Literature Review The majority of continuous FLPs focus on locating one facility, known as the Minisum and Minimax problems (e.g., Tompkins et al., 2010). Owen and Daskin (1998), Klose and Drexl (2005), ReVelle and Eiselt (2005), and Melo et al. (2009) all present survey papers including the continuous FLP. ReVelle and Eiselt (2005) acknowledge that in the FLP the number of facilities may or may not be known. They refer to the FLP with unknown
Engineering Management Journal Vol. 24 No. 3 September 2012

number of facilities as free entry. An extensive literature search shows that there is a dearth of studies related to the CCFLPU. The only published study related to the CCFLPU is Qui et al. (2009) who developed a hybrid heuristic that combines variable-length Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). In their approach, the GA method yields the number of facilities and PSO optimizes the location for each facility. This article will present a new mathematical formulation of the CCFLPU, which to the best of our knowledge, is the only existing formulation for the problem. Further, four heuristics, including a PSO-based heuristic are proposed. A Formulation for the CCFLPU In the simplest version of the CCFLPU a set of predefined existing facilities with known locations and demand volumes are given. A set of unknown number of new facilities with known capacities are to be located on a continuous plane in order to meet the demand of the existing facilities. The objective function is to minimize the total fixed and variable costs. The fixed costs are independent of the magnitude of interaction with existing facilities. On the other hand, the variable costs include the transportation costs per unit, which depend on the interaction. The most used interaction metric is the cost of travel between new and existing facilities, which is dependent on the distance between them. In the continuous space FLP distances are commonly measured using the Minkowski distance (ReVelle and Eiselt, 2005). The Minkowski distance (dpij) between two locations i and j with coordinates (ai , bi) and (aj , bj), respectively, are presented in Equation 1. dpij = [| ai aj |p +| bi bj |p ] 1/p (1) The Minkowski distance depends on the parameter p. When p = 1 the Minkowski distance is the Rectilinear (or Manhattan) distance shown in Equation 2. When p = 2 the Minkowski distance is the Euclidian distance shown in Equation 3. The other wellknown Minkowski distance is when p = , called the Chebyshev distance shown in Eq. (4). di1j = | ai aj | +| bi bj | (2)

di2j = | ai aj |2 +| bi bj |2

d = Max{| ai aj | , | bi bj |} ij

(3)

In this article we assume Rectilinear distances, which is the most commonly used Minkowski distance in the FLP literature. Rectilinear distances have the benefit that they can be easily linearized in most mathematical models. A Formulation for the m-FLP Before presenting a mathematical formulation for the CCFLPU, we discuss the formulation of the m-FLP. Recall that unlike the CCFLPU, for the m-FLP the number of new facilities will be exactly m. We define the following notation in order to discuss the m-FLP formulation. Indices i = Index for new facilities (i 1,.., m) j = Index for existing facilities (j 1,.., J )

(4)

Parameters m = Number of new facilities k = Fixed cost of a facility (dollars) cij = Travel Cost between new facility i and existing facility
25

j (dollars per kilometer) Dj = Demand for existing facility j (in number of trips) Ai = Capacity of new facility i (in number of trips) Variables T = Total cost to be minimized (xi, yi) = x and y coordinates of new facility i dij = Distance between new facility i and existing facility j Sij= Supply (flow) from new facility i to existing facility j (number of trips) (5)

Parameters (, ) = Coordinates of the right upper corner point (, ) = Coordinates of the left bottom corner point M = a very large number Variables Zi =
1 if facility i is outside the virtual box (i.e., not used) 0 if facility i is inside the virtual box (i.e., new facility)

Min T = k * m + i j cij * dij * Sij s.t | xi aj | +| yi bj | dij i Sij Dj j i,j

(6) (7) (8) (9) (10)

j Sij Ai i Sij , dij 0i i,j xi , yi 0 i

The objective function in the m-FLP is to minimize Tthe total cost. Unfortunately, the objective function for this problem is inevitably non-linear. In the model, constraint set Equation 6 defines the Rectilinear distances. One can easily linearize Equation 6 or use any other type of distances instead (e.g., Euclidian, Chebyshev, etc.). We will later discuss a linearization scheme. Constraint sets in Equations 7 and 8 consider the demand and supply of existing and new facilities. Constraint sets in Equations 9 and 10 are non-negativity constraints. Flows between facilities (i.e., Sij) needs to be an integer, as this represents the number of trips; however, the formulation turns out to be totally unimodular, which automatically yields integer results. Clearly, the m-FLP formulation cannot be used for the CCFLPU as in the latter the maximum value of the index i (i.e., m) is an unknown. Not knowing the number of new facilities to open makes the mathematical formulation very challenging. In the next subsection we propose an interesting way of modeling the problem. A Mathematical Formulation for the CCFLPU In this section we present a new mathematical formulation of the CCFLPU. To the best of our knowledge, it is the only mathematical model for the CCFLPU. To avoid confusion with the previous model, we redefine index i for the CCFLPU so that its maximum value is I (i.e., I = m for the m-FLP). Since the number of new facilities (i) is unknown for the CCFLPU, we assume that we have a high enough number of new facilities to locate (i.e., I is large). Our CCFLPU modeling strategy is to define a virtual box (rectangle) within our facility location space such that facilities that are inside the box will open and facilities outside this box will not open. In other words, facilities might be located inside or outside the virtual box, yet only facilities located inside the box are considered. We will discuss the desired characteristics of the number of new facilities (I) considered in the model and the dimensions of the box after presenting the formulation. The following additional notation is defined for the virtual box nonlinear mixed integer model.
26

s.t xi a MZi i (12) xi MZi i (13) yi MZi i (14) yi MZi i (15) dij = | xi aj | +| yi bj | i,j (16) i Sij * Zi Dj j (17) i Sij * Zi Ai i (18) dij , Sij 0 i,j (19) Zi {0,1} i (20) The new constraint sets shown in Equations 12 to 16 are used to activate indicator variable Zi if facility i is inside the virtual box. Unfortunately, like the m-FLP, the CCFLPU model has a non-linear objective function (Equation 11) and constraints (Equation 15). Rectilinear distance constraints in Equations 6 and 15 can be easily linearized by adding a set of new variables and constraints. One way of linearizing the constraints is as follows. Let, dixj = horizontal distance between new facility i and existing facility j diyj = vertical distance between new facility i and existing facility j Then, constraint sets (6) and (15) can be substituted for the following constraints: xi aj dixj i (21) aj xj dixj i (22) yi bj diyj i (23) bj yj diyj i (24) dixj + diyj = dij i,j (25) dixj , diyj 0 (26) Unfortunately, the remaining formulation for both problems is still non-linear because of the objective function. The formulations can be solved by optimization software; however, the solution provided in most cases is only guaranteed to be locally optimal. In these cases, the solution from the optimization software serves as an upper bound for problem.
Engineering Management Journal Vol. 24 No. 3 September 2012

min T = k i Zj + i j cij dij sij

(11)

The value for the number of facilities considered in the CCFLPU formulation (i.e., I) needs to be at most the maximum number of facilities that could be optimal to open. We define mmax as the theoretical maximum number of facilities that should be opened. Dj mmax = Jj=1 A (27)
i

Greedy Heuristic (H1) The greedy heuristic (referred to as H1) will start at the minimum number of facilities from Equation 28 and continue solving the m-FLP until increasing the value of m results in an increase in the value of the objective function. The proposed greedy heuristic is presented next. Step 1: Find the minimum and maximum number of facilities, mmin and mmax from Equations 28 and 29, respectively. Let m = mmin. Solve the m-FLP and store the objective function value T as the best solution found (i.e., Tbest= T(mmin)). Step 2: Let m = m+1 and solve the m-FLP. If the resulting T is greater than Tbest then End and return Tbest as the solution. Otherwise, go to Step 3. Step 3: Update Tbest (Tbest = T). If m < mmax return to Step 2; otherwise End. A variant of the greedy heuristic could allow the objective function value to exceed Tbest several times before terminating. In this study we considered the case where the Greedy Heuristic stops after the objective function value exceeds Tbest twice. We labeled this heuristic as the modified greedy heuristic (H2). Search Heuristic (H3) The Greedy Heuristics are not expected to perform well for instances where the optimal solution is close to mmax, or when the CCFLPU is not convex with respect to m. The proposed Search Heuristic (H3) will explore multiple values of m (breadth) before exploring its neighbors (depth). The proposed Search Heuristic is presented below: Step 1: Find the minimum and maximum number of facilities, mmin and mmax from Equations 28 and 29, respectively. Let n = ||(mmaxmmin +1) 1/2||; = (mmaxmmin +1) / n; j = 0; mj = mmin; Tbest = ; a = mmin. (Note: The operator ||x|| is the round function; i.e., ||x|| = Round(x)). Step 2: Solve m-FLP using mj new facilities to obtain the objective function value T. If T < Tbest, then Tbest = T and win = j. If mj = mmax, then m* = mwin-1 +1 and go to Step 4; otherwise go to Step 3. Step 3: Increase a by , increase j by 1, and mj = ||a||. If mj mmax, then mj = mmax and go to Step 2. Otherwise return to Step 2. Step 4: Solve m- FLP using m* new facilities. If T < Tbest then Tbest = T and mbest = m*. Step 5: Increase m* by 1. If m* = mwin+1 then END, return Tbest as the result, and return mbest as the optimal number of new facilities. Otherwise return to Step 4. Particle Swarm-Based Heuristic (PSO) Particle swarm optimization (PSO) is a parallel population-based stochastic optimization method motivated by the organisms behavior such as schooling of fish and flocking of birds. In PSO, simple software agents, called particles, move in the search space of an optimization problem. The search space is constituted by the position of the particles. Particles change their position in
27

Note that this theoretical maximum assigns as many new facilities as required to satisfy the demand of the new facilities. We propose using I = mmax when solving the formulation. On the other hand, we propose using the size of the virtual box as a = mini (ai), = mini (bj), = m ax (ai ), = m ax (bj). Note that i i it is implicitly assumed that the centroids of all existing facilities are non-negative. Lastly, the proposed formulation can be simplified by assuming that the (a,) is located at the origin. Also, the model can be simplified if the problem is uncapacitated. A Brute-Force Algorithm to Solve the CCFLPU Based on the mathematical formulation for the m-FLP presented earlier in the paper, one can find the optimal solution to the CCFLPU by using a brute-force approach. The proposed approach solves the m-FLP for each possible value of m from a theoretical minimum value of m (mmin) to a theoretical maximum value of m (mmax). The brute-force algorithm is presented as follows: Step 1. Given a set of existing facilities, the minimum and maximum number of facilities required can be determined using the formulas in Equations 28 and 29, respectively.
J D = j mmin = j A 1 i

(28) (29)

mmax = jJ= 1

Dj Ai

Step 2. For m = mmin to mmax, solve the m-FLP from Section 3.1. Report the minimum value of T obtained. The bounds obtained in Equations 28 and 29 are theoretical bounds which are not necessarily tight. The minimum number of facilities required to meet the overall demand, mmin, can be computed by dividing the total demand by the capacity of each new facility. If m < mmin then it would be impossible to meet the necessary demand. On the other hand mmax, the maximum number of facilities required to meet the overall demand, is found by dedicating as many new facilities to each existing facility as needed to meet its demand. It can easily be shown that and m value greater than mmax will never be optimal. Unfortunately, the m-FLP formulation is non-linear; therefore, the brute-force algorithm will only provide an optimal solution if the optimization software guarantees a global optimal solution. In general, for the particular structure of the m-FLP formulation, the optimization software was unable to guarantee optimal solutions; therefore, although the brute-force algorithm provides a scheme to find the optimal solution to the problem, it is only used as a heuristic for our experiments. Heuristics for the CCFLPU Four additional heuristics are proposed to solve the CCFLP. These heuristics are described next.
Engineering Management Journal Vol. 24 No. 3 September 2012

the search space by changing their velocity according to rules emulating swarm intelligence. In the CCFLPU there are two main decisions to make the location of the new facilities and their respective flows. We will use PSO to change the position of the new facilities in the search space. Given the position of the new facilities, the problem of determining the flows from new to existing facilities can be solved as a transportation problem; therefore, given the location of new facilities from PSO, we use the Minimum-Cost Method (Winston, 2004) to assign the flows. The Minimum-Cost Method finds the shortest distance from an existing facility with unfulfilled demand to a new facility with uncommitted capacity and assigns (i.e., commits) as much flow as possible. The PSO-based heuristic used for our problem is described below. Let f(xi) be the fitness function given a particle position obtained using Equation 5. Note that it is assumed that the greedy heuristic has been used to distribute the flows given the particles position. Only particles with positive flow are considered new facilities (i.e., opened). Initialization: For each particle i = 1, ..., S For each dimension d = 1, 2 (representing x and y coordinates) Initialize the particles positions (xi,d) randomly within the virtual box (note that xi,2 is the y-coordinates of each particle) Initialize the particles best known position to its initial position: pi,d xi,d If (f(pi,d) < f(gd)) update the swarms best known positions: gd pi,d Initialize the particles velocity: vi = 0 Next Dimension dimensions Next particle Improvement: Do For each particle i = 1, ..., S-1 For each dimension d = 1, 2 Update the particles velocity: vi,d a*vi,d + b*Rnd()* (pi,d - xi,d) + c*Rnd()*(gd - xi,d) Update the particles position: xi,d xi,d + vi,d Next dimension If f(xi,d) < f(pi,d)) Update the particles best known position: pi,d xi,d End If If (f(pi) < f(g)) update the swarms best known position: gd pi,d End If Next Particle For particle i = S we perform a local optimization Let xS,d gd For each particle with a positive flow determine the weighted average x and y coordinate (i.e., centroid) considering the existing facilities that receive the flow (weighted by flow) Move each new facility to the corresponding centroid If f(xi,d) < f(pi,d)) Update the particles best known position: pi,d xi,d End If If (f(pi) < f(g)) Update the swarms best known position: gd pi,d End If
28

Loop Until the MaxIter iterations have occurred without an improvement or MaxTime expires The parameters a, b, and c are selected to control the behavior and efficacy of the PSO method. Rnd() is a uniform random number between 0 and 1. The Sth particle is used to run a greedy heuristic in order to find a local optimal solution for the global best particle during each iteration. Preliminary experiments indicated that the best choice of parameters for our problem was: a = 0, b = -0.5, c = 1, MaxIter = 500, and S = 25. Experimental Results The full factorial experiment shown in Exhibit 1 was performed to understand the behavior of the proposed solution methods under different instances of the problem. The number of existing facilities (J) was varied from 10, 25, and 50. Three levels of demand-tocapacity ratios were considered (1:3, 5:2, and 80/20). The 80/20 level for demand-to-capacity ratio assigned 80% of the demand to 20% of the existing facilities. The average distance between existing facilities factor has three levels. The first two levels (1002, 10002) uniformly located the existing facilities within a rectangle with dimensions 100100 and 10001000, respectively. The third level (80/20), forced 80% of the existing facilities to be clustered in a square sub-area of 5050, while the remaining 20% of the existing facilities are located on a concentric rectangular area of 100100, outside of the clustered sub-area. The levels corresponding to the fourth factor, the ratio of fixed costs to variable costs, are 1:10, 1:100, and 1:1000. For simplicity, it was assumed that all new facilities are identical in terms of capacity.
Exhibit 1. Factorial Experiments Factor and Levels Levels 1 Factors 2 3 4 No. of Existing Facilities Demand : Capacity Ratio Average distance between existing facilities Fixed : Variable Costs Ratio 10 1:3 1002 1:10 25 5:2 10002 1:100 50 80/20 80/20 1:1000

A total of 81 instances are tested with each heuristic. Each instance of the problem was solved using the CCFLPU formulation, the brute-force algorithm, and PSO. The non-linear models were solved using Lingo 11.0s non-linear solver. Meticulous preliminary experiments were used to fine-tune the software parameters for the experiment. (It is important to highlight that due to the nonlinearity of the formulation, Lingo could not guarantee global optimal solutions.) Given the stochastic nature to the PSO, three runs were performed for each instance. For practical purposes, the PSO was forced to stop after one hour of runtime. Also, CCFLPU solved with Lingo was stopped after 3 hours of runtime. Since H1, H2, and H3 evaluate a sub-set of the m-FLPs used in the bruteforce algorithm, their objective function values were obtained by studying the solution of the brute force algorithm instead of rerunning the respective m-FLPs. Exhibit 2 summarizes the results by presenting the average and standard deviation of the percent difference in terms of total
Engineering Management Journal Vol. 24 No. 3 September 2012

cost (T), with respect to the best found solution in each instance, over the 81 instances. The individual percent differences in terms of total cost (T) are computed as follows, % diff = THeu/TBest1. The columns in Exhibit 2 represent the six solution methods tested. For example, the first result Exhibit 2 can be interpreted as, on average, the CCFLPU formulation solved with Lingo was 277.8% higher than the best solution found in each instance. For the PSO, only the best of three runs is reported.
Exhibit 2 Average and Standard Deviation of the Percent Differences with Respect to the Best Solution Found
Brute Force 25.7% 101.5%

1. H3 found the best solution from the brute-force algorithm in 75 of the 81 instances (92.59%). The instances in which the H3 did not find the optimal solution the percent difference ranged between 1% and 15%. 2. PSO and CCFLPU formulation tend to use fewer new facilities than the brute-force algorithm. Exhibit 4 presents the average runtimes for each method in seconds. The average runtime of the three PSO runs was considered as the runtime for the instances. It can be observed from Exhibit 4 that the runtimes for H3 were much less than the brute-force algorithm; therefore, we conclude that H3 is the best of all the heuristics tested. Hence, out of the solution methodologies considered, engineering managers should focus on using H3 to solve the CCFLPU.
Exhibit 4. Average Runtimes Over All Instances in Seconds
CCFLPU 442.7 H1 310.4 H2 568.7 H3 1676.2 Brute Force 3029.2 PSO-Ave 2483.4

CCFLPU Average StdDev 277.8% 875.3%

H1 156.8% 277.1%

H2 122.6% 235.1%

H3 36.9% 158.0%

PSObest 2682.9% 9247.0%

Exhibit 3 shows the percent of times, over 81 instances, that each method found the best found solution.
Exhibit 3. Percent of Instances Each Method Found the Best Found Solution
CCFLPU % best 20.99% H1 4.94% H2 6.17% H3 55.56% Brute Force 59.26% PSObest 17.28%

It can be seen from Exhibits 2 and 3 that in terms of the average percent difference with respect to the best solution found, the PSO had the worst performance, followed by CCFLPU solved with Lingo, H1, H2, H3, and brute-force; however, although PSO did not have a good average performance, it found the best solution in 17.29% of the instances. This indicates that the PSO either had very good or very bad performance depending on the instance. This situation could be a consequence of the 1 hour runtime limit imposed. In terms of the percent of times that each method yields the best solution, the worst performance was from H1, followed by H2, PSO, CCFLPU solved with Lingo, H3, and brute-force. Clearly, the brute-force algorithm and H3 were the best performers in terms of solution quality. By analyzing each instance separately, the following observations can be made:

Another interesting result from our experiments is observing the effect of the experimental factors on the solution. Exhibits 5-7 present the results for the brute-force optimal solution for the CCFLPU for different values of m for the case of J = 50. The legend on Exhibits 5-7 represents the four factors at their corresponding levels. It can be observed that Exhibits 5-7 keep factor 1 at level 1 (i.e., J=50). Exhibits 5-7 consider different values of the second factor (i.e., demand/capacity ratio) and vary the other factors. The third and fourth factors are the average distance between facilities and the variable-to-fixed costs ratio, respectively. From Exhibits 5-7, it is observed that the solution curve for the CCFLPU using the brute-force algorithm is not convex with respect to the number of new facilities. This observation is very important as it is an indicator of the smoothness of the CCFLPU. Perhaps this behavior of having multiple local optimal solutions is more pronounced when the third factor (i.e., the average distance between existing facilities) is at level 3 (80/20). Also, it was observed that when the fourth factor (i.e., Fixed-to-Variable Costs Ratio) was set at Level 1 (1:10) the optimal solution was typically to place the new facilities on top or very close to the existing facilities. On the other hand, when this factor was at Level 3 (1:1000), the optimal solution typically had new facilities

Exhibit 5. T vs. m for J=50 Existing Facilities (Factor 1 at Level 3, Factor 2 at Level 1)

Engineering Management Journal

Vol. 24 No. 3

September 2012

29

Exhibit 6. T vs. m for J=50 Existing Facilities (Factor 1 at Level 3, Factor 2 at Level 2)

Exhibit 7. T vs. m for J=50 Existing Facilities (Factor 1 at Level 3, Factor 2 at Level 3)

further from the existing facilities. Once again, these observations depend on the quality of the solution of Lingo; however, these results are useful for engineering managers by demonstrating how the characteristics of the CCFLPU solution vary with the different experimental factors. Implications for the Engineering Manager The CCFLPU is considered a strategic-level decision problem by engineering managers. Decisions at the strategic level are considered long-term and typically involve high investment costs and are hard to reverse. Furthermore, strategic decisions impose constraints for tactical and operational level decisions. For example, determining the bus routes to be used in a public transportation system (tactical-level decision) depends on the number and location of bus stops (strategic-level decisions). Hence, strategic-level decisions such as the quantity and location of new facilities is one of the most important decisions made by engineering managers. The CCFLPU has many applications for engineering managers, particularly for supply chain and transportation network design. Some examples include locating cross-docks, distribution centers, warehouses, and manufacturing facilities. The CCFLPU is also particularly suited for situations where the investment and operational cost of new facilities is not very high, such as emergency management response and transportation management problems such as the bus stop problem.
30

The CCFLPU can be used as part of a hierarchical methodology to design user-centered public transportation systems. In the hierarchical methodology, the CCFLPU should be solved first, considering the user demographics. The solution from the CCFLPU (i.e., the number and location of bus stops) can then be incorporated to design the bus routes. When designing the bus routes, minor perturbations to the solution of the CCFLPU may be made. Ideally, Geographic Information Systems (GISs) should include the CCFLPU. This article presents new solution methodologies that are easily implementable and could be incorporated into GISs. Experimental results indicate that the proposed solutions methodologies for the CCFLPU are appropriate for small or medium size problems. Conclusions and Future Work This study presents Capacitated Continuous Facility Location Problem with Unknown Number of Facilities (CCFLPU). The CCFLPU is considered a strategic-level decision problem by engineering managers. Although this problem has many potential applications, including transportation management, there is a dearth of research. It is argued that this problem could be used, among other things, to determine the number and location of bus stops for an urban area. This article presents a mathematical formulation for the CCFLPU, a brute-force optimal solution algorithm for the problem based on the m-FLP, and four heuristics. A full factorial experiment considering four factors
Engineering Management Journal Vol. 24 No. 3 September 2012

is used to compare the solution methods and to understand the behavior of the optimal solution. It is concluded that H1 and H2 were the two worst performing heuristics. PSO and the CCFLPU formulation solved with Lingo were outperformed by H3 and the brute-force algorithm. In general, PSO and the CCFLPU solved with Lingo opened fewer new facilities than the brute-force algorithm. Brute force found the best objective function value among all solution methods in 59.26% of the instances. On the other hand, the CCFLPU found the best solution in 20.99% of the instances, while PSO found the best solution in 17.28% of the instances. The Search Heuristic (H3) found the same solution as the brute-force in 75 out of the 81 instances of the problem (92.59%) with average runtimes of approximately 40% less than the brute-force algorithm. The instances in which H3 did not find the optimal solution the percent difference ranged between 1% and 15%. Hence, it is concluded that H3 was the best performing heuristic when combining solution quality and time. Experimental results indicate that the proposed solutions methodologies for the CCFLPU would be appropriate for small or medium size problems. To promote the widespread use of the CCFLPU by engineering managers, the CCFLPU should be incorporated into Geographic Information Systems (GISs). The best performing heuristic found in this article is easily implementable and could be incorporated into GISs; however, future work is required to develop additional mathematical models and solution approaches to solve the CCFLPU. H3 should be tested with larger problems, and the impact of the number of intervals used can also be investigated. The CCFLPU can also be extended to consider different fixed costs for new facilities and transportation distances / costs that are based on the actual location of the new facilities. Acknowledgments This material is based upon work supported by the U.S. Department of Homeland Security under Award Number 2009ST-062-000037, and by the National Science Foundation under Grant No. EEC-0851879. The views and conclusions contained in this document are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the U.S. Department of Homeland Security or the National Science Foundation. References Klose, Andreas, and Andreas Drexl, Facility Location Models for Distribution System Design, European Journal of Operational Research, 162 (April 2005), pp. 4-29. Mehrez, Abraham, and Alan Stulman, The Maximal Covering Location Problem With Facility Placement on the Entire Plane, Journal of Regional Science, 22:3 (August 1982), pp. 361-365. Melo, Mo T., Stefan Nickel, and Francisco Saldanha-da-Gama, Facility Location and Supply ChMin managementA Review, European Journal of Operational Research, 196 (July 2009), pp. 401-412.

Owen, Susan H., and Mark S. Daskin, Strategic Facility Location: A Review, European Journal of Operational Research, 111:3 (December 1998), pp. 423-447. Purvis, Russell L., Gordon E. McCray, and Tom L. Roberts, Heuristics and Biases in Information Systems Project Management, Engineering Management Journal, 16:2 (June 2004), pp. 19-27. Qiu, Minmin, Liu Liu, Hongwei Ding, Jing Dong, and Wei Wang, A New Hybrid Variable-length GA and PSO Algorithm in Continuous Facility Location Problem with Capacity and Service Level Constraints, Service Operations, Logistics and Informatics IEEE/INFORMS International Conference (July 2009), pp. 546-551. ReVelle, Charles S., and Holst Eiselt, Location Analysis: A Synthesis and Survey, European Journal of Operational Research, 165:1 (August 2005), 1-19. Tompkins, James A., John A. White, Yavuz A. Bozer, and Jose M. A. Tanchoco, Facilities Planning, 4th ed. John Wiley & Sons (2010). Winston, Wayne L., Operations Research Applications and Algorithms, 4th. Ed. (2003). About the Authors Hctor J. Carlo received a BS in industrial engineering from the University of Puerto RicoMayagez (UPRM) in 2001, and MS and PhD degrees in industrial and operations engineering from The University of Michigan, Ann Arbor, in 2003 and 2007, respectively. He is an associate professor in the Industrial Engineering Department at UPRM and Director of the Lean Logistics (LeLo) Lab. His current research interests include the design and optimization of logistics systems. Francisco Aldarondo is an undergraduate industrial engineering student at the University of Puerto Rico Mayagez. His research experience at the Lean Logistics (LeLo) Lab has been funded by the NSF. His research interests include optimization, mathematical modeling, and artificial intelligence. Priscilla M. Saavedra is an undergraduate industrial engineering student at the University of Puerto Rico Mayagez. She is a Boeing, Verizon Foundation, and Department of Homeland Security Scholar. For the past year, she has been working on undergraduate research as a part of the Lean Logistics (LeLo) Lab with focus on nonlinear optimization. Priscillas work experience includes a summer internship with General Mills. Silmarie N. Torres is an undergraduate industrial engineering student at the University of Puerto Rico Mayaguez. She is a Xerox scholar and Department of Homeland Security scholar. Her research interests include optimization and finance. Contact: Hctor J. Carlo, University of Puerto Rico Mayagez, Department of Industrial Engineering, Call Box 9000, Mayagez, PR 00681-9000; phone: 787-832-4040 x3105; hector.carlo@upr.edu

Engineering Management Journal

Vol. 24 No. 3

September 2012

31

Copyright of Engineering Management Journal is the property of American Society for Engineering Management and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.

Potrebbero piacerti anche