Sie sind auf Seite 1von 21

int. j. prod. res., 1998, vol. 36, no.

6, 1549 1569

A genetic algorithm approach for multiple criteria facility layout design


A. A. ISLIER
This paper presents a genetic algorithm-based model for facility layout. Layout of departments consisting of nite elements is modelled in gene structures. Better and better solutions that satisfy multiple objectives are produced by employing genetic operations to these genes. Better quality layouts are obtained by this method on the test problems available in the literature. These results show the potential for solving complex facility layout problems.

1.

Introduction

Cutting out garments from material, printing wallpaper, lling a chocolate box or programming a robot to manipulate components without collision all involve a tricky task, tting and moving irregular shapes economically (Green 1989). Dyckho (1990 ) states that the following problems are characterized essentially by the same `logical structure although they appear under di erent names in the literature: Cutting stock, trim loss, and assortment problems, Packing problems (bin, dual bin, strip, vector, and knapsack ), Loading problems (vehicle, pallet, container, and car), Other problems of commercial interest (capital budgeting, change making, line balancing, scheduling, depletion, nesting, memory allocation, design, partitioning, dividing, and layout ). The underlying common logical structure is based on two groups of basic data that de ne the shapes of the `large objects and `small items. The large objects may be the stock of materials for cutting problems, station times for line balancing or oor area for layout problems, while the small items are the ordered parts to be cut, work elements to be assigned to the stations, and departments to be placed on a site respectively. Determining the desirable patterns of geometric combinations for these small items on the large objects (as cutting and layout plans), is the essence of those problems. Plant layout, being one of these problems, deals with the arrangement of the most valuable assets of the enterprises, the facilities. A facility in this context is a physical entity that facilitates the performance of any job. It may be a machine tool, a work-centre, a manufacturing cell, a machine shop, a department, a warehouse, etc. Although the characteristics di er widely depending on the type of the facility, Raoot and Rakshit (1993a ) state that the area is the unique characteristic that is common for all facilities. The area which facilities occupy could be represented as:

Revision received July 1997. Department of Industrial Engineering, Osmangazi University, Bademlik, 26030, Eskisehir, Turkey.
0020 7543/98 $12. 00

1998 Taylor & Francis Ltd.

1550

A. A. Islier Single points where the facilities are assumed to be concentrated. (This way of modelling is suitable for location problems, where distances between facilities are quite large compared with dimensions of the facilities ), Distributed over a certain region. (In general this region is considered as a rectangle and the distribution is uniform for the sake of convenience ) , or Made up of nite elements. (A planar site is divided into a grid of square blocks, for example 1 1 metres in dimension, as the well known computer aided layout technique CRAFT is modelled. The space allocated to a facility is represented by a collection of adjacent blocks whose total area is equal to that of the facility. These squares are often called cells, although they are not related to the `integrated group of machines concept of cellular manufacturing that is also known as a cell).

Irrespective of the presentation type being used, it is essential that the location of machines/workstations be such that the distance travelled by personnel or material handling carriers between each pair of these facilities be minimized (Heragu 1992). Immer (1953 ), informs us of the fact that up to 40% of the wage bill in manufacturing is spent on materials handling. Matson et al. (1992 ) declare the gure to be 60%. Cutting even 25% of this cost would lead to tremendous savings. Fitting and moving the facilities economically is the most e ective way to cut these transportation costs. Fitting is placing the generally irregularly shaped facilities (small items) into available sites (large objects) by optimizing some objectives while satisfying the restrictions. Moving the facilities means considering time as an additional dimension. The internal and external environment of the enterprise changes as time passes, so rearranging the facilities to maintain their tness is necessary. Rosenblatt (1986) and Conway and Venkataramanan (1994), propose dynamic models to provide adaptation power against those changes in the long term. Nevertheless much of the research on the problem is devoted to the static case. Domschke and Drexl (1985) note that over 2000 articles on the facility layout eld have appeared in academic journals. However, relatively few of these have employed multicriteria decision making techniques (Current et al. 1990). Real life facility layout problems involve the selection of the most e ective arrangement of physical facilities to allow greatest e ciency with a given combination of resources to produce a product or service. At the same time it is necessary to satisfy multiple objectives such as the overall integration of all functions, minimum material movement, smooth work ow, e ective space utilization, employee satisfaction, safety, exibility, etc. ( Raoot and Rakshit 1993b). Although the type of facilities to be placed may vary in each context, the steps to a solution remain largely the same. The area of each facility and the interconnection cost of every facility pair are estimated by using the data provided from production engineers. The interconnection cost may either be speci ed as a quantative measure (e.g. the weighted ow of materials) or a qualitative adjacency requirement. The goal is to minimize the total interconnection cost while satisfying all speci ed constraints (Tam and Li 1991). Facilities must not overlap, as a rst constraint. Additionally, areas of individual departments and the total oor area should be held within pre-de ned limits. The resultant solution should give reasonable shapes for both the departments and the overall layout. Some additional precautions might be given to the xed facilities

Multiple criteria facility layout design: a genetic algorithm

1551

(staircases, washrooms, etc. ), to the connections with the outer world (entrance and delivery points) and to the third dimension (for multistore buildings ). According to Kusiak and Heragu (1987), facility layout algorithms can be classi ed into two main categories: optimal (like branch and bound or cutting plane algorithms ) and suboptimal methods. Since the problem falls into the class of NPcomplete problems, the optimal methods can only be used to solve smaller problems involving up to 15 facilities (Welgama and Gibson 1993). This fact has lead researchers to focus on developing heuristic algorithms. These algorithms can in turn be assessed as four main types: (1) construction methods, (2) improvement methods, (3) hybrid methods, and (4) graph theoretic methods. The construction process involves two important phases. The rst phase is to determine the selection order, which is to decide the sequence of placing of the departments into the oor area. The second phase is their installation in the area to ful l certain objectives (Wang et al. 1991). Improvement methods try to swap the locations of the departments either randomly or by using a sort of hill-climbing technique to increase the solution quality. The swapping stops when there is no further improvement. Thus an optimal (or more likely a suboptimal) point is reached. Approaches that combine optimal and suboptimal techniques or a blending of construction and improvement methods are called hybrid methods. In graph theoretic methods, a layout is prepared as a weighted uni-directional network. The solution method is to seek for the dual of the maximal planar subgraph to determine the nal layout of the facilities. Sirinaovakul and Thajchayapong (1994 ) emphasize that these conventional heuristic techniques do not consider enough possible outcomes in the solution process to arrive at the optimal point, and they are very sensitive to the initial solution. Glover and Greenberg (1989 ) point out the emerging approaches of the heuristic search for solutions to combinatorially complex problems. These problems are not sensitive to initial solutions and some of them allow the application of parallel processing. So more than one solution at a time is developed. Thus lots of possible outcomes are considered. Additionally, the searching process is not restricted to the certain domains of the solution space, since they are insensitive to initial solutions. Therefore the probability of being trapped into inferior solutions is much lower. Simulated annealing, being one of these arti cial intelligence techniques, is shown to be independent of the initial solution (Tam 1992a, Jajodia et al. 1992). But it seems inapplicable to larger problems, since its ability to nd good solutions is at the expense of a reduced cooling rate. As the cooling rate is reduced, solution time increases, making the solution of larger problems impractical. Another arti cial intelligence technique, known as the genetic algorithm (GA) is a strong candidate for solving layout problems. First of all, its guided search characteristic provides the ability to nd good solutions by searching only a very small fraction of the solution space. In addition, parallel processing tremendously increases the processing speed. The main di culties in using genetic algorithms are modelling the information of the solution alternatives as gene structures, devising e ective solution strategies, and nding out appropriate parameters which are highly problem speci c (Grefenstette 1986). Furthermore the central problem in the application of genetic algorithms is constraint handling (Michalewicz 1994). Glover and Greenberg (1989 ) state that genetic algorithms have found their major applications in optimization of unrestricted functions, but research on the types of penalty functions is worthwhile in

1552

A. A. Islier

increasing the applicability of genetic algorithms. Michalewicz et al. (1996 ), in a recent work, categorize the constraint handling techniques as: rejecting, penalizing, retaining, and repair techniques. Rejection of infeasible solutions from consideration immediately is known as the `death penalty . This method could be considered as the most fundamental treatment. Adding a penalty term to the objective function to transform the constrained problem into an unconstrained one, is a more exible way. The penalty is zero if all constraints are satisifed, and non-zero if some of them are violated. Thus casual withdrawal of infeasible solutions is anticipated. The main issue in application of a penalty term is providing the balance between premature convergence due to large penalties and avoiding a lot of violations due to weak penalization (Homaifar et al. 1994). The success of the penalty approach depends on the way in which the penalties are handled. A method that identi es the worst feasible solution as better than the best infeasible solution is appreciated. Homaifar et al. (1994 ) proposed a method based on the multi-level assignment of penalty factors. Here, a stepwise increasing penalty function depending on severity of violation is employed. A metric that measures the distance from the feasible region is necessary to evaluate this kind of severity. Some researchers suggest using the number of constraints violated, regardless of the magnitude of the violations, as a criterion to judge the violation degree, where such a metric is not available. The survival power of infeasible solutions is another important topic for handling constraints dynamically. The life-span of infeasible solutions, that is the number of generations they are perpetuated might be added to the penalty term to reduce their durability. Tate and Smith (1995a ) used both the number of generations and the attributes of the best solution currently found to construct their penalty function. In addition to these static and dynamic approaches, self-adjusting methods are also emerging. Coit and Smith (1996) suggest such a technique where an adaptive penalty function learns to adapt itself based on the severity of the constraints of a particular problem instance. Coit et al. (1996) furnish widespread information on the general form of adaptive penalty functions. Penalty mehods in genetic algorithms force the infeasible solutions to disappear or repair themselves naturally. Directly repairing them either partly or completely is an alternative approach. Extensive research to develop a general solution for the constraint handling problem of genetic algorithms is observed in the literature. Genecop III is such a system, developed for constrained numerical optimization problems with non-linear constraints (Michalewicz 1994). Repairing an infeasible solution is relatively easy even though these techniques are problem dependent. So these types of algorithms enjoy a particular popularity in the evolutionary computation community for many combinatorial optimization problems (Michalewicz et al. 1996). On the other hand, retaining only the feasible solutions during computation is a direct way to safeguard against constraint violation. But an elaborated design of coding schema and operators to maintain feasibility is a prerequisite for con ning to the feasible solutions at each stage of the solution process. Tate and Smith (1995b ), state that `for a given problem famility, particularly clever encoding can restrict the necessary search area to an extremely small subset of all feasible solutions; at the same time computationally e cient breeding and mutation of new feasible solutions could be allowed .

Multiple criteria facility layout design: a genetic algorithm

1553

In order to apply the genetic algorithm on the constrained layout problem, this study tries to employ the most direct way wherever possible. That is, the aim is to design a meticulous encoding schema and operators which are capable of producing and keeping the feasible solutions. Additionally a greedy repair algorithm is devised, where infeasible solutions are unavoidable. Any penalty method is not considered, due to the weak performance of such methods on highly constrained cases such as layout problems. Section 2 of this paper gives the formulation of the layout problem in its general nature. In 3 the fundamentals of genetic algorithms are presented. The basic structure of the genes designed, together with the genetic operators devised, is introduced and the constraint handling strategies installed are explained in detail. Section 4 describes the relation between the genes and layout plans, then explains their usage in nding better and better solutions as the generations pass. Computational experience on the benchmark problems found in the literature is given in 5. Finally a discussion of the proposed method, and its potential usage as a layout problem solving tool, is discussed in 6.
2. Problem formulation

Malakooti (1989 ) commented that any facility layout problem has multiple criteria in selecting the best layout and it is very speci c in domain and problem. Static, multicriteria, two dimensional layout problems are also known to be the dominant type from experience. So this study is aimed at developing a solution method for those problems, by making use of newly emerging arti cial intelligence techniques. Parts are moved from one department to another to be processed, in a manufacturing environment causing transportation costs. The concepts of unit load and centroids greatly simplify the calculation of the total transportation cost of any layout alternative. A unit load might be either a physical load that consists of a pallet, any standard sized container, etc. or a conceptual equivalent load determined by making some assumptions and conversions. The number of unit loads carried between two facilities within a period (a month or a year) is the ow between them. Calculation of the distances between the facilities is generally based on the centroids (centre of areas) of those departments. The Minkowski metric, as a general device to measure the distance between two facilities i and j , is given by Shafer and Rogers (1993 ) as follows: dij ( p) =

k= 1

xik -

xjk

1 /p

( 1)

where k is the dimension and p > 0 determines the particular metric used. Wellknown special cases include the rectilinear (city block or Manhattan ) metric obtained with p = 1, the Euclidean metric is obtained with p = 2, and the Chebychev metric is obtained with p = . For layout problems, N = 2 and p = 1 is appropriate since these problems are two dimensional in general and rectilinear lanes are present in the shops. Thus leading to: dij = xi1 - xj1 + xi2 - xj2

( 2)

for the rst and second (horizontal and vertical ) dimensions. The products of these distances with the ows and unit costs lead to the total transportation cost within a period. Unit costs are the costs of moving one unit load

1554

A. A. Islier

from one facility to another by one unit of length (the metre for example ). Unit costs are di cult (if not impossible ) to obtain. Furthermore, absolute costs are not necessary in order to compare the layout alternatives. So all the unit costs are assumed to be equal (even all equal to unity) in practice. The data collected are placed in distance, ow and unit cost matrices. These matrices are not necessary symmetrical. If the entries for the distance matrix are sorted in descending order while those for the ow matrix are sorted in ascending order, two vectors are obtained. A dot product of one of these vectors by the transpose of the other, gives a lower bound for the total transportation load. One of the objectives of the proposed method is to optimize the ratio of this ideal load to the real load. This ratio will remain between one and zero irrespective of the problem parameters being used. Increasing this ratio shows a reducing transportation load, hence better solutions. Another problem common to layout techniques is reaching unreasonable department forms at the end. A resulting solution, containing some departments divided into two or more parts and/or some odd shaped (very slender, multi-cornered, void, etc. ) departments is not exceptional. A well-known concept from mechanics, the moment of area, is an indicator of these undesired shapes. As the area becomes more compact, this value reduces. To obtain a manageable scale, the calculated moments are divided by the areas to get the shape factors. The second objective of the proposed model is to optimize this factor, thus providing non-separated, compact departmental areas. The systematic layout planning technique of Muther (1976 ), seeks a balance between required and available areas. That means that areas of the departments are not xed; instead a trade-o between requirements and sources is essential. Accordingly the required areas for the departments should be stated within a certain tolerance, instead of insisting on exact values. Hence the third and the last objective of the designed model is to minimize the total deviation from the most desired areas while they are kept between prescribed upper and lower bounds. The factors used to construct the objective function of the model can be de ned more formally as follows: Load factor: t= V

n- 1 i= 1

n j = 2 cij

f ij dij

( 3)

where V n cij f ij dij ideal load (obtained by multiplying sorted distance and ow vectors) number of departments cost of carrying one unit load for a unit length between facilities i and j number of unit loads to be carried between facilities i and j within a period distance between facility i and j (rectilinear).

Shape factor: s= where Sk

b i= 1

n k= 1 rk e n j= 1 k= 1

aijk

( 4)

the set of cells of department k

Multiple criteria facility layout design: a genetic algorithm

1555

upk rectilinear distance from centroid of department k to centroid of cell p aijk an indicator which is equal to 1 if the cell on the i th row and j th column on the site plan is dedicated to the kth department (otherwise 0) b number of rows on the site plan e number of columns on the site plan rk moment of area of the kth department as: rk = Deviation factor: h= 1p Sk

u2 pk .

( 5)

n k= 1

Ak -

e j=1

b i= 1

e j = 1 aijk

b i= 1

n k= 1

aijk

( 6)

where Ak is the most desired area for department k. Then the model becomes: t Max s h s. t.

( 7) ( 8)

aijk
k= 1

1,

" i and j
Ak ,

Ak

aijk
n

" k

( 9)

i= 1 j= 1

aijk

e b.

( 10)

i= 1 j = 1 k = 1

The rst constraint prohibits the sharing of any cell by more than one department. So overlapping of the departments is avoided. The second constraint keeps the total number of cells (area) of any department between predetermined boundaries. Here Ak and Ak are lower and upper bounds for the numbers of cells for the kth department. The last constraint checks the total number of the cells of all the departments (this total must not exceed the available area). Representing the areas as nite elements (cells) is both meaningful from the point of view of practice and makes it attractive to employ a genetic algorithm approach.
3. Genetic algorithms, ope rations on genes and constraint handling

Genetic algorithms (GA) are basically search techniques. They emulate the natural process of evolution by processing towards the optimum (Al-Sultan et al. 1996). Tam (1992) also indicates the parallelism of the evolution process simulated by GA and the search process conducted in solving a function optimization problem as: structure fitness evolution solution, objective function value, search.

1556

A. A. Islier

In a genetic algorithm, an individual (any feasible solution to the problem) is an element of a population. Here the population is a subset of the solution space that is held in the hand at any instant of the solution process. The magnitude of the population depends on the size and the nature of the problem, and the size of the computer memory being used. Each element of the population is called a chromosome (an individual of the population or a solution to the problem ). Chromosomes are combinations of symbols, known as genes. In each generation the best chromosomes are selected by the roulette principle to act as parents. The roulette principle is basically a process by which a good parent (solution) gets a higher probability of being selected compared to the bad ones. Four genetic operations known as reproduction, crossover, mutation, and inversion are applied to those parents to generate the o spring (new and better solutions). These o spring inherit good qualities from the parents. The fundamental theorem of genetic algorithms states that `schemata with high tness values and small de ning lengths grow exponentially with time (Srinivas and Patnaik 1994). A schema in this sense is a similarity template describing a subset of strings with similarities at certain positions. The process of applying genetic operators to the populations is continued until there is no further improvement in the o spring. Reproduction is the main genetic process in which two parents are selected from the total parent solution pool by using the roulette rule. From these parents an o spring is generated which inherits the qualities from both parents. Population size is held at 50, and a reproduction rate of 0. 95 is used in this study. That is, the selection of the parents to be crossed-over is made amongst the best 95% of the population. The crossover operator operates on two chromosomes (parent solutions) and generates the o spring. Since it is an inheritance mechanism, the o spring generated inherits some characteristics of the parents. Crossing-over is the exchanging of substrings between randomly selected parents. A chromosome structure with three segments of g1, g2, and g3 is used in this study (this structure will be explained in detail in the following section). The proposed algorithm crosses-over the rst, second, and third segments of the parent chromosomes simultaneously. Single point crossingover is applied for all segments from randomly selected in ection points. Figure 1 shows an example of this operation. It would be helpful to think of the existence of two lists that are kept throughout the solution procedure. The rst list holds the names of the departments to be placed and the latter holds their current magnitudes respectively. These lists for the example

Figure 1. Cross-over operation.

Multiple criteria facility layout design: a genetic algorithm

1557

problem are given as AB CDEF and 537654 8 . The area of department A is 5 units, the area of B is 3 units, and so on, for this example. In ection points for the segments are after the fourth, again fourth and third digits, respectively, between the normal and boldfaced charcters in gure 1. As Glover and Greenberg (1989 ) argue, the resultant o spring genes must be modi ed to permit certain constraining relations to hold. Segment 1 of the rst o spring is DEACBE. As is easily seen, there are two Es but no F in this structure. On the other hand, representing each department on the chromosome structure once is the logical constraint of the problem. A modi cation is needed to treat that issue. So the rst repetition of department E is replaced with the rst absent department of F. Then this segment becomes DEACBF by replacing the second E by F. A similar procedure is repeated for the second segment. The third segment needs a di erent type of modi cation, since a di erent constraint (the site width is equal to b) is considered. So the sum of the elements in segment 3 (parameter b) should be equal to 2 + 3 + 1 + 4 = 3 + 1 + 1 + 5 = 10 throughout the solution process. Repaired and unrepaired forms of the rst o spring (upper o spring on gure 1) are given in gure 2. The formal algorithm A1 employed to repair segments 1 and 2 is as follows: Step 0. Get a crossed chromosome, list 1, list 2 and prepare n null sets ( n is the number of departments to be placed and number of digits of segment 1). Set pointer p1 to one. Step 1. Set pointer p2 to one. Step 2. If element p1 is of list 1 equals to digit p2 of segment 1, then add the value of p2 to set p1 as an element. Step 3. If the number of elements of set p1 is one then replace this element with a zero. Step 4. If p2 < n then increase p2 by one and go to 2. Step 5. If p1 < n then increase p1 by one and go to 1. Step 6. Keep the sequence number of null sets as vi and non-zero elements on the other sets as zi successively. Step 7. Replace the digits zi of segment 1 with elements vi of list 1. Then replace the digits zi of segment 2 with elements vi of list 2 and STOP. The sets produced by the second and third steps of this algorithm are: {0 },{0 },{0 }, {0 },{0, 6},{ }. Here non-zero elements (reooccurrences ) and null sets (vacancies ) are apparent. There is a pair of violations where v1 = 6, z1 = 6. That means there are repetitions on the sixth digits of g1 and g2, at the same time that the sixth elements of list 1 (F) and list 2 (8) are not used at all.

Figure 2.

Repaired and unrepaired forms of the rst o spring.

1558

A. A. Islier

Figure 3.

An example of mutation operation.

By applying the rule: Replace the foremost reoccurrence with the foremost vacancy , at step 7 of A1, the repaired segments appear as: DEACBF and 645738 as explained before. Thus both violations are removed. Algorithm A2 repairs the third segment of the chromosomes as follows: Step 0. Get the crossed chromosome and the value of b. Set p3 to the sequence number of the last digit of segment 3. Step 1. Sum up the current values of the digits of this segment as s. If s = b, then STOP. Step 2. Set the value of digit p3 of segment 3 as r. Step 3. If s< b, then increase the value of r by ( b - s). Then STOP. Step 4. If ( s - b) r then reduce r by ( s - b) . Then STOP. Step 5. Set the digit p3 of segment 3 to zero and reduce p3 by one. Then go to 1. Some substrings are destroyed by these modi cations, but the overall e ect is not so great, as the test problems verify, since the schema is kept as small as possible. Mutation normally brings random changes in a parent solution to generate an o spring. The mutation mechanism utilized in this study on segment 2 is demonstrated in gure 3. Randomly selected digits for mutation are the fourth and the last digits shown as boldfaced in gure 3. The randomly chosen di erence is 2 (the maximum allowable di erence should be less than the minimum departmental area). So the fourth digit of the second segment (seven) is increased by two, while the last digit (three) is decreased by two. Since the net change is kept at zero, the constraint on the total area is not forced. Inversion is another operator that increases the diversity of a population. It may be thought of as self-crossing. Genes of any individual are exchanged internally by this process. The inversion mechanism used in the proposed method is exempli ed in gure 4. Boldfaced symbols E and C, together with 4 and 7, are exchanged. No modi cation is needed after this process, since the existence of the departments is not a ected by this operation. The constraint handling techniques applied to segments of the chromosomes to remove the adverse e ects of genetic operators are outlined in table 1.

Figure 4.

An example for inversion operation.

Multiple criteria facility layout design: a genetic algorithm


Operator Crossing-over Mutation Inversion Table 1. Segment g2 {g1 and g3 Constraint handling strategy Use algorithm A1 Use algorithm A2 No mutation is made Retain by carrying on the net change as zero No mutation is made Swapping has no adverse e ect No inversion is made

1559

g2 {g1 and g3

g1 g2 g3

Operators and constraint handling techniques employed in the study

Austin (1989) remarks on some advanced strategies to increase the e ciency of genetic algorithms. Two of them, the elitist and the adaptive strategies, are used in this study. The elitist strategy is employed to prevent the destruction of the good solutions by genetic operators. The best solution of each population is copied to the following generation. The duplicates of the best solution are detected and they are changed by the inversion operator. Some of the adaptive strategies are related to the parameters that trigger the genetic operators. The timing used to trigger the operators highly a ects the solution quality. Consequently, selection of a triggering strategy is an outstanding issue. As the circumstances change, the optimum frequency with which to apply a mutation operator also changes. Therefore an adaptive strategy is advantageous for the performance of the algorithm. The adaptive strategy developed for the designed method is built on the concept of coe cient of variation. The tness function of a GA, which corresponds to the objective function of optimization models, is an indicator of the survival power of individuals. The mean and standard deviation of the value of this function for all the individuals of a generation are calculated. The ratio of the standard deviation to the mean (coe cient of variation ) indicates the diversity degree of the genes of a generation. As cross-over operations are performed, individuals become more and more similar. This event increases the danger of being trapped into suboptimal solutions. So mutation and inversion operators are employed to increase the diversity. Conversely, these operators partly destroy the good solutions. Thus providing a good balance for the diversity is an important topic in designing a genetic algorithm. As the value of the coe cient of variation drops below a certain level (5 to 20% threshold values gave good results on test problems) either a mutation or an inversion operation is applied. Solving the layout problems by making use of genetic algorithms necessitates developing a method to model any feasible solution as a chromosome as well as to devise a method for converting these chromosomes into layout plans and vice versa.
4. Layout plans, genes and the proposed GA

Determination of the tness function, string representation, parameters like population size, number of generations, and genetic operations are the main tasks in constructing any GA (Chan and Tansri 1994 ). Determination of the tness function of the proposed GA is based on the objective function of the model given in section 2. The parameters are settled by experiments made on the model. The genetic

1560

A. A. Islier

operators to be used are decided as explained in section 3. This section explains the modelling of the strings. Symbolizing the information as genes may be the most di cult task in employing a GA to solve any problem. So a `divide and conquer strategy is used. The problem is separated into two subproblems of sequencing and placing. Sequencing is made randomly for the initial solution and new sequences of departments for new layout alternatives are obtained by using genetic operators. This sequence is shown on the rst segment of the chromosomes. Departments are symbolized by letters, a sequence of these letters gives the rst segment g1 of the chromosomes. The second segment of the chromosomes g2 is parallel to the rst part. The areas of the departments are given in the same order with the rst segment. This time, areas are given as integer numbers which are the number of cells of the corresponding departments. Areas of the departments are allowed to change between their upper and lower bounds, so every solution (individual ) should contain this information on its chromosomes. The second subproblem (placement) is quite di cult to solve, and structuring of the third segment is shaped by that solution. The facility placement procedures should construct a layout and present the plan as a two dimensional matrix. All the basic unit areas of the plan that constitute the facilities are expressed as matrix elements. Placement procedure should generate these feasible layouts with a minimum computational requirement ( Raoot and Rakshit 1993a ). These conditions point to highly systematic placement procedures, which are presented in gure 5. The area of each department consists of a certain number of cells which is proportional to the area of the departments. If the sequence of the departments is determined, their cells are concatenated as a string in the same sequence. This string could be laid o in one of the ways shown in gure 5. The procedures x-oscillatory and y-oscillatory are the same in principle. The string is laid o from one end of the site to the other, then the direction is reversed. This procedure continues until all the cells are lled. A spiral procedure begins from the centre and lls up the area spirally. The xy-oscillatory procedure rst divides the site into bands or strips, then applies a single dimensional oscillation to these strips. The band widths are not necessarily the same. The ability to use multi-bands with di erent widths increases the diversity of the layout alternatives. So this procedure is favoured in this study, and the third segments of the chromosomes g3 carry this information. The layout information is kept either in layout plans or in genes. The convertibility from one model to the other is very important. A layout plan is required to evaluate the layout alternatives. The shapes of the departments, the positions of their centroids, and the distances between the centroids are exclusively calculated if such a

Figure 5. Systematic placing procedures.

Multiple criteria facility layout design: a genetic algorithm

1561

layout plan exists. Thus the value of the tness function could be determined, and the algorithm is used. Conversely, genetic operations are easy to perform on the chromosomes. So a means of transformation is essential between two presentations. The xy-oscillatory laying-o procedure provides an excellent tool for this transformation. Additionally, a larger number of alternatives are produced via di erent bandwidths. (The well-known computerized layout technique ALDEP, also uses the xy-oscillatory laying-o procedure, but sweeping widths are kept equal in this technique. The proposed method generates these widths randomly. The sum of the bandwidths used in any plan should be equal to the width of the oor area:

where

M i= 1

Bi = b,

( 11)

Bi width of the ith band, M the total number of the bands in a layout. The random numbers on which bandwidths are based are generated so that the inequality As

Bi

A t

for " i

( 12)

is held, to get reasonable widths. The boundary values are: As the number of cells of the smallest department, At the number of cells of the largest department. The last bandwidth ( Bm ) is recalculated to equate the sum of the widths to the number of columns on the site plan. This adjustment is made both for the initial solution generating process and after each crossing-over. A sample layout plan, lled by the xy-oscillatory process, is given in gure 6, together with its corresponding gene structure.

Figure 6.

A sample layout plan and the corresponding gene structure.

1562

A. A. Islier

The genetic algorithm based on these two interchangeable models is established as follows: Generate the initial population randomly. Calculate the tness values of the individuals. Save the best solution of the population in a separate place. Perform the inversion operations to duplicated best solutions. Apply the roulette principle to select the pairs to be crossed-over. Apply the cross-over operator to all pairs selected. If the value of the coe cient of variance is lower than the threshold value, then perform inversion and mutation operations alternately. Step 7. Replace the last individual of the population with the best solution saved. Step 8. Go to step 1, if a satisfactory solution is not obtained yet. Step Step Step Step Step Step Step Experience shows thirty to a hundred generations is enough to reach a satisfactory solution in the sense of step 8. Examining the tendencies of the best and average values of the tness functions of the generations, gives an idea of that point. After a certain point (a generation ) is passed, no more progress on the best values is witnessed while the average value oscillates around a lower value continuously, as seen in gure 7. The resultant layout given in gure 6 is obtained as the second individual of generation 57 within 2 min and 2. 43 s on an IBM compatible. The best value obtained is 1. 58 while the average of that population is 1. 06. The PASCAL program that generates gures 6 and 7 has been written to check the proposed algorithm against some benchmark problems found in the literature. The results of these comparisons are given in the following section.
5. Computational experience

0. 1. 2. 3. 4. 5. 6.

Facilities layout is an immense problem in nature, so researchers can consider only a limited number of the facets of the problem at any time. Therefore any

Figure 7.

The best and average values of the generations.

Multiple criteria facility layout design: a genetic algorithm

1563

solution procedure o ered so far, is able to cover it only partly. Environments, requirements, approaches, models, objectives, constraints and the data structure vary from one application to the other. Consequently, testing the performance of a newly developed algorithm by using previously published benchmark data is a tough issue, since this data is hardly applicable without any alteration. So some modi cation of the existing data is required to adapt it to a newly designed model. Hillier 1963, Rosenblatt 1986, Khare et al. 1988, Golany and Rossenblatt 1989, Imam and Mir 1993 worked on simple problems with equal areas. Since the number of the departments they concerned does not exceed twelve, the optimum solutions are also known. Comparisons of these optimum and heuristic solutions with the results of the GA proposed, are summarized in table 2. Only materials handling costs are compared because all the departments of the test problems contain single unit cells. The initial solution obtained by using the data on the rst group, gave a result almost as good as Golany s best solution. The solution generated at the fourth generation was only 0. 03% worse than the optimum with a deviation 20 times lower than Golany s solution of 0. 56% defect. For the data of Hillier, the best solution of Golany is outperformed at the third generation. And the solution obtained within 2 minutes is 5% closer to optimum while Golany s solution is two times worse (10%). Both Khare s and Imams solutions are reached at the tenth generation with data given by Khare. The result obtained with 1. 5 minutes is 2% better than both. Cases 1 and 2 of Banerjee et al. (1992 ) are also solved, to test the behaviour of the proposed method in larger problems. Flows and desired departmental areas for these cases are given by the authors. Additionally, lower and higher bounds for perimeters are also stated. The areas of the squares that have these perimeters are calculated and
The best layout of generation 4 2 5 4 3 1 6

Rosenblatt 1986 (set 1) Optimum solution Best solution of Golany 1989 Initial solution of the algorithm The best solution of generation 4

Load Computation time 12 822 12 894 12 898 12 926 Load 297 327 323 313 practically 0 1. 5 s

Comparisons with data of Hillier 1963 Optimum solution The best solution of Golany 1989 The best solution of generation 3 The best of generation 92 Comparisons with data of Khare et al. 1988, Imam and Mir 1993 Solution of Khare et al. 1988 Solution of Imam and Mir 1993 Initial solution of the algorithm Solution on the generation 10 Solution on the generation 76

Computation time 4. 9 s 1 min 46 s

Layout 3 10 5 6 2 9 12 7 1 11 4 8 (generation 92)

Load 505 506 512 505 495

Computation time practically 0 12 s 1 min 28 s

Layout 6 3 7 4 10 11 12 9 6 1 2 8 (The best solution of generation 76

Table 2. Test results of the GA with benchmark data for equal areas.

1564
Case 1. (Hypo manufacturing )

A. A. Islier
Load 43 700 39 270 37 780 Load 147 799 143 600 112 100 Computation time 2 min 2. 4 s 7 min 23 s Computation time 9 min 40. 3 s 21 min 52. 17 s

Final solution of Banerjee et al. (1992) recalculated for centroidal distances (1) The best solution on generation 57 (2) For a site of 13 21, on generation 3201 Case 2. (Har Bal Company) Final solution of Banerjee et al. (1992) recalculated for centroidal distances (3) The best solution on generation 14 (4) The best solution on generation 31
(1) (2)

Layout is given in gure 6. The best result obtained, using di erent site alternatives. (3) The rst generation that outperforms the solution of Banerjee et al. (1992 ). (4) The layout and gene structure is given in gure 8. Table 3. Comparisons of the results for two larger problems.

accepted as upper and lower bounds for corresponding departments. Then the problems are solved by using the proposed algorithm. The resultant layout of Banerjee et al. (1992 ) had to be reworked to be compared with the results obtained, since their calculations of the loads were based on the pick-up and drop-o points concept adopted by the authors. The recalculated load, based on the centroids, is compared with the values obtained by the proposed method in table 3. Here, Case 1 could be accepted as a middle sized plant with its 12 departments, while Case 2 is a quite large plant for real life with its 32 departments.
6. Discussion and conclusions

This study has been conducted to devise an advanced method to solve the plant layout problems encountered in real life. An applicable, meaningful, fast, adequate, and re ned method is intended. Consequently, the newly emerged technique of GA is combined with prevailing concepts such as material ows, unit loads and grid structures. So the suitability of the accepted, common and conventional approaches is enhanced with the power and speed of an AI technique. The problem is perceived in association with its multi-criterial structure. Minimizing the transportation load, maximizing the compactness of the departmental areas and minimizing the di erence between requested and available areas are recognized as signi cant criteria in evaluating the layout alternatives. Total transportation load is calculated by using the ow information of unit loads and rectilinear distances between the centroids of the departments. The proposed method is also capable of adding an alternative way of calculating the total load between the pick-up and drop-o points. (This approach might be advantageous especially for cellular manufacturing systems). Although minimizing the transportation load is known to be the dominant factor in many of the cases, determination of these quantitative ow values is quite di cult. So some layout techniques use qualitative closeness ratings instead. These ratings are expressed as vowels to represent the opinion of experts. Their opinions re ect the degree of desirability for the neighbouring of department pairs. The neighbouring of two

Multiple criteria facility layout design: a genetic algorithm

1565

Figure 8.

The best solution found for Har Bal Company.

departments could be considered as absolutely (A ) or extremely (E) necessary, important (I), ordinary (O) or unimportant (U) . Some neighbouring alternatives like the neighbouring of painting and welding departments are even avoided (X ), since they might have adverse e ects on each other. A more reasonable way to declare the interconnection cost of every facility pair might be an amalgamation of qualitative and quantitative values. Output of any method that combines these values is readily accepted as the input of the proposed solution technique. Compactness of the departmental areas is another important topic in evaluating the layout alternatives. The ratio of the perimeter to the area should be kept at a minimum, to get adequate departmental areas. Separated or isolated areas as well as non-rectangular areas are not desirable. Moments of areas indicate the degree of condensation of the e ective room around centroids. So this study recommends using of these easy to calculate moments as a new basis for evaluating the compactness of the departmental areas. The trade-o among requested and available areas was the third argument considered while designing the proposed technique. Di erences between requested and assigned departmental areas are attempted to be minimized while keeping them within predetermined limits. The objective function of the problem composed of these three criteria is accepted as the tness function of the GA. Overlapping of departments and falling outside of the siting area are not allowed. Areas of the departments are kept within predetermined limits. New encoding schemes and repair techniques are developed to achieve these necessities. Additionally, receiving and shipping departments are desired to be placed in contact with the outer world. Some existing facilities like staircases, transformers, etc. might

1566

A. A. Islier

require a de nite place in the siting area. Some departments should be placed as distantly as possible. These and other potential restrictions can be added to the proposed method with small alterations. The layout alternatives to be evaluated by those criteria, are generated by the GA design. The second task in building a GA following the determination of the tting function is developing an appropriate string representation. That is, to design a system to model the information on the layout plans as a gene structure while considering pertinent genetic operators. Separating the sorting and placing procedures and preferring the xy-oscillatory installing procedure, simpli es the designing of that system. Once the sequence of the departments, the numbers of cells that make them (their areas) and the bandwidths to be used are decided, then the layout plan and its corresponding gene structure are promptly constructed. The information of sequence and magnitude of the areas together with bandwidths are coded as three distinct sectors of the chromosomes. This information is also the subject of modelling as layout plans. Both of these equivalent models are necessary to achieve the designing of an appropriate system. The performance of this system adheres on the balance of the diversity and tness power of the generations. Using an elitist strategy together with the coe cient of variation provided this trade-o . The best solution of each generation is copied to the following generation. Duplicates of that best value are destroyed by using an inversion operator. Diversities of the generations are checked continuously by making use of the coe cient of variation. As the value of this coe cient drops below a certain threshold, either a mutation or an inversion operator is applied to increase the diversity. So, trapping into local minima is avoided as far as possible. This system is tested by using two categories of available data sets. Small sized problems with known optimum solutions were in the rst category. The proposed technique reached near-optimum solutions within seconds. The di erence from the optimum was in the order of ten thousandths. The solutions of the previous heuristics are outperformed many times. Medium and large sized problems fall into the second category of the problems tested. The foregoing results of cases 1 and 2 were recalculated for centroids to ascertain a common basis for comparison purposes. For case 1 (a middle sized problem ) an 11% better solution is obtained in 2 min. A kind of sensitivity analysis is made by slightly adjusting the dimensions of the site area. Then a 16% better solution is obtained within 7 1 2 min. Case 2 could be considered as a large sized problem. The rst solution to outperform the antecedent result is obtained within 10 min. A 32% better solution is attained in 22 min (at 31st generation ). Then no improvement on the best value of the following generations is witnessed. Only the rst criterion, the total load, is checked against the existing solutions, since the other criteria are not included in the models found in the literature. Furthermore the cases used as benchmark solutions utilize the spine concept. This notion uses the space generously, although it seems convenient for extension and enlargement projects of the facilities. On the contrary, new production philosophies like JIT, advocate the economical usage of valuable sources like space. And compact overall layouts are preferable both for space utilization and transportation load reduction. The computer code developed to solve the problem for rectangular siting areas could be easily modi ed to cover non-rectangular siting areas as well.

Multiple criteria facility layout design: a genetic algorithm

1567

The chromosome structure as designed, motivates the use of a discrete grid con guration to model the space arrangement. Although the recent studies seem to be focused on continuous models, discrete models are easier to use on computers. The increasing capacities of computers are also able to approximate these models to continuous counterparts more closely. These facts anticipate the future usage of discrete models. The information handled by the proposed technique can be visualized as a framework modelled in seven dimensions: two geometrical dimensions of siting area, sequence of the departments, the magnitude of the areas of these departments, the bandwidths used to place the departments, the individuals within a generation and the generations as computation proceeds. Processing of this framework of information with the technique proposed, gave e cient results for both small and large problems. The results contiguously obtained for small problems were reasonably close to the optimum and they were better than the results obtained by outdated techniques. The results from larger problems were superior. Improvements of more than 30% are attained over the results of old techniques within a fraction of an hour. However, the actual value of a technique is particularly appreciated when it is employed to solve professional problems. The natural basis and adaptation ability of the GA increase the competitive power of the proposed approach over opponent techniques for solving real-life problems. The proposed technique might be improved to cover dynamic problems too, since plant layout is a continuous (indeed an endless) trial in practice. Furthermore, that technique might be augmented with production cell design and scheduling modules to get a more integrated problem solving tool. The combination of this technique with arti cial neural networks and more powerful computers in the prospective studies would increase its pro ciency. A layout plan has some apparent and some obscured features. Separating and modelling these features as recessive and dominant genes on homolog chromosomes would improve the e ectiveness of dynamic models seeking trade-o s between criteria in the future. As a result, the proposed technique passed the examination in hypothetical problems and is su ciently complete to be tested on actual problems of the moment. The tness power of this proposal in real life will indicate its actual worth.

References A l -Sultan, K. S., Hussain, M. F., and Niz ami, J. S., 1996, A genetic algorithm for the set

covering problem. Journal of the Operations Research Society, 47 , 702 709. A ustin, S., 1989, Introduction of genetic algorithm. Arti cial Intelligence Expert, Mary 49 53. Banerjee, P., Montreuil, B., Moodie, C. L., and Kashyap, R. L., 1992, A modelling of interactive facilities layout designer reasoning using qualitative patterns. International Journal of Production Research, 30 , 433 453. Chan, K. C., and Tansri, H., 1994, A study of genetic crossover operations on the facilities layout problem. Journal of Computers and Industrial Engineering , 26 , 537 550. Coit, D. W., and Smith, A. E., 1996, Penalty guided genetic search for reliability design optimisation. Journal of Computers and Industrial Engineering, 30, 895 904. Coit, D. W., Smith, A. E., and Tate, D. M., 1996, Adaptive penalty methods for genetic optimisation of constrained combinatorial problems. INFO RMS Journal of Computing, 8, 173 182. Conway, D. G., and Venkataramanan, M. A., 1994, Genetic search and the dynamic facility layout problem. Computers and Operations Research, 21, 955 960.

1568

A. A. Islier

Current, J., Min, H., and Schilling, D., 1990, Multiobjective analysis of facility location decisions. European Journal of Operational Research, 49 , 295 307. Domschke, W., and Drexl , A., 1985, Location and layout planning. In L ecture Notes in Economics and Mathematical Systems, 238 (New York: Springer-Verlag). Dyckhoff, H., 1990, A typology of cutting and packing problems. European Journal of Operational Research, 44 , 145 159. Glover, F., and Greenberg , H. J., 1989, New approaches for heuristic search: a bilateral linkage with arti cial intelligence. European Journal of Operational Research, 39, 119 130. Golany, B., and R osenblatt, M. J., 1989, A heuristic algorithm for the quadratic assignment formulation to the plant layout problem. International Journal of Production Research, 27 , 293 308. Green, R. C., 1989, Geometry, shoemaking and milk tray problem. New Scientist , August 12, 50 51. Grefenstette, J., 1986, Optimisation of control parameters for genetic algorithms. IEEE Transactions on Systems, Man, and Cybernetics , 16 , 122 128. Heragu , S. S., 1992, Recent models and techniques for solving the layout problem. European Journal of Operational Research, 57, 136 144. Hillier, F. S., 1963, Quantitative tools for plant layout analysis. The Journal of Industrial Engineering, 4, 33 40. Homaifar , A., Qi, C. X., and Lai, S. H., 1994, Constrained optimisation via genetic algorithms. Simulation , 62 , 242 254. Imam, M. H., and Mir, M., 1993, Automated layout of facilities of unequal areas. Journal of Computers and Industrial Engineering , 24, 355 366. Immer , R. J., 1953, Materials Handling (New York: McGraw-Hill). Jajodia , S., Minis, I., Harhalakis, G., and Proth, J. M., 1992, CLASS: Computerised layout solution using simulated annealing. International Journal of Production Research, 30 (1), 95 108. Khare, V. K., Khare, M. K., and Neema , M. L., 1988, Combined computer aided approach for the facilities design problem and estimation of the distribution parameter in the case of multigoal optimisation. Journal of Computers and Industrial Engineering , 465 476. Kusiak , A., and Heragu , S. S., 1987, The facility layout problem. European Journal of Operational Research, 29 , 229 251. Malakooti, B., 1989, Multiple objective facility layout: a heuristic to generate e cient alternatives. International Journal of Production Research, 27, 1225 1238. Matson, J. O., Mellichamp, J. M., and Swaminathan, S. R., 1992, EXCITE: Expert consultant for in-plant transportation equipment. International Journal of Production Research, 8 , 1969 1983. Michalewicz , Z., 1994, Genetic Algorithms+ Data Structures= Evolution Programs (New York: Springer). Michalewicz , Z., Dasgupta , D., Le R iche, R. G., and Schoenauer, M., 1996, Evolutionary algorithms for constrained engineering problems. Computers and Industrial Engineering, 30 , 851 870. Muther , R., 1976, Systematic L ayout Planning (Boston, MA: Cahnes Books). R aoot, A. D., and R akshit, A., 1993a, An experimental comparison of systematic placement procedures for facility layout design. International Journal of Production Research, 7, 1735 1756. R aoot, A. D., and R akshit, A., 1993b, A `linguistic pattern approach for multiple criteria facility layout problems. International Journal of Production Research, 31 , 203 222. R osenblatt, M. J., 1986, The dynamics of plant layout. Management Science, 32 , 76 86. Shafer , S. M., and R ogers, D. F., 1993, Similarity and distance measures for cellular manufacturing. International Journal of Production Research, 5 , 1133 1142. Sirinaovakul , B., and Thajchayapong, P., 1994, A knowledge base to assist a heuristic search approach to facility layout. International Journal of Production Research, 1, 141 160. Srinivas, M., and Patnaik, L. M., 1994, Genetic Algorithms: A Survey. IEEE Computer, June 17 26.

Multiple criteria facility layout design: a genetic algorithm

1569

Tam, K. Y., 1992a, A simulated annealing algorithm for allocating space to manufacturing cells, International Journal of Production Research, 30, 63 87. Tam, K. Y., 1992b, Genetic algorithms, function optimisation, and facility layout design. European Journal of Operational Research, 63 , 322 346. Tam, K. Y. , and Li, S. G., 1991, A hierarchical approach to the facility layout problem. International Journal of Production Research, 29, 165 184. Wang , M. J., Liu , C. M., and Pan, Y. S., 1991, Computer-aided panel layout using a multicriteria heuristic algorithm. International Journal of Production Research, 6 , 1215 1233. Welgama , P. S., and Gibson, P. R., 1993, A construction algorithm for the machine layout problem with xed pick-up and drop-o points. International Journal of Production Research, 31 , 2575 2590. Tate, D. M., and Smith, A. E., 1995a, Unequal area facility layout using genetic search. IIE Transitions, 27, 465 472. Tate, D. M., and Smith, A. E., 1995b, A genetic approach to the quadratic assignment problem. Computers and Operations Research, 22, 73 84.

Das könnte Ihnen auch gefallen