ebook img

Evolutionary algorithms, simulated annealing and tabu search PDF

15 Pages·2001·0.38 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Evolutionary algorithms, simulated annealing and tabu search

EngineeringApplicationsofArtificialIntelligence14(2001)167–181 Evolutionary algorithms, simulated annealing and tabu search: a comparative study Habib Youssef*, Sadiq M. Sait, Hakim Adiche DepartmentofComputerEngineering,KingFahdUniversityofPetroleumandMinerals,KFUPMBox#673,Dhahran-31261,SaudiArabia Received1July1999;receivedinrevisedform1June2000;accepted1July2000 Abstract Evolutionary algorithms, simulated annealing (SA), and tabu search (TS) are general iterative algorithms for combinatorial optimization. The term evolutionary algorithm is used to refer to any probabilistic algorithm whose design is inspired by evolutionarymechanismsfoundinbiologicalspecies.Mostwidelyknownalgorithmsofthiscategoryaregeneticalgorithms(GA). GA,SA,andTShavebeenfoundtobeveryeffectiveandrobustinsolvingnumerousproblemsfromawiderangeofapplication domains. Furthermore, they are even suitable for ill-posed problems where some of the parameters are not known before hand. Thesepropertiesarelackinginalltraditionaloptimizationtechniques.InthispaperweperformacomparativestudyamongGA, SA, and TS. These algorithms have many similarities, but they also possess distinctive features, mainly in their strategies for searchingthesolutionstatespace.Thethreeheuristicsareappliedonthesameoptimizationproblemandcomparedwithrespectto (1)qualityofthebestsolutionidentifiedbyeachheuristic,(2)progressofthesearchfrominitialsolution(s)untilstoppingcriteria aremet,(3)theprogressofthecostofthebestsolutionasafunctionoftime(iterationcount),and(4)thenumberofsolutionsfound atsuccessiveintervalsofthecostfunction.Thebenchmarkproblemusedisthefloorplanningofverylargescaleintegrated(VLSI) circuits.Thisisahardmulti-criteriaoptimizationproblem.Fuzzylogicisusedtocombineallobjectivecriteriaintoasinglefuzzy evaluation(cost)function, which isthenused torate competing solutions. # 2001 Elsevier ScienceLtd. All rightsreserved. Keywords: Geneticalgorithms;Simulatedannealing;Tabusearch;Fuzzylogic;Floorplanning;Combinatorialoptimization;VLSI 1. Introduction 2. They are blind, in that they do not know when they reached an optimal solution. Therefore they must be This paper is concerned with one class of combina- told when to stop. torial optimization algorithms: general iterative al- 3. They have ‘‘hill climbing’’ property, i.e., they gorithms, which constitute a class of approximation occasionally accept uphill (bad) moves. algorithms. The growing interest in this class of 4. Theyaregeneral,i.e.,theycaneasilybeengineeredto algorithms is attributed to their generality, ease of implement any combinatorial optimization problem; implementation, and mainly, their robustness in solving all that is required is to have a suitable solution a wide variety of problems from a range of application representation, a cost function, and a mechanism to domains. We shall limit ourselves to three popular traverse the search space. iterative algorithms,(1)geneticalgorithm,(2)simulated 5. Under certain conditions, they asymptotically con- annealing, and, (3) tabu search. All three optimization verge to an optimal solution. heuristics have several similarities, namely (Sait and The question of how to compare two constructive Youssef, 1999a): algorithms has long been settled. Two measures are 1. They are approximation (heuristic) algorithms, i.e., typically used: (1) the time complexity of the algorithm, they do not guarantee finding an optimal solution. and, (2) the quality of solution, i.e., how far would the solution be from optimal in the worst or best case. In most cases, the time complexity of a constructive algorithm is usually easy to derive; however, determin- *Corresponding author. Tel.: +966-3-860-4286; fax: +966-3-860- ing the quality of solution relative to the optimum is 3059. E-mailaddress:[email protected](H.Youssef). usually very difficult, and in some cases not possible. 0952-1976/01/$-seefrontmatter # 2001 ElsevierScienceLtd.Allrightsreserved. PII:S0952-1976(00)00065-8 168 H.Youssefetal./EngineeringApplicationsofArtificialIntelligence14(2001)167–181 For iterative approximation algorithms, the time * wi;hi: width and height of bi, which are complexity is mostly a function of the number of constants for rigid blocks and variable for iterations allowed for the algorithm as well as the flexible blocks. complexity of the evaluation function. The various * ri: desirable aspect ratio about each variable- operators of the algorithm as well as the problem to be shape block where 1=r4h=w4r; r ¼1 if i i i i i solved can also have determining effect on the time block i is rigid. complexity of the algorithm. However, typically, all * ai: area of bi (i.e., ai ¼wi(cid:5)hi), ai is constant. implementations of approximation iterative algorithms 2. A set of nets N ¼fn ;n ;...;n;...;n g describ- 1 2 i k have polynomial time complexity. Thus the time ing the connectivity information. complexity isnot of concernwhen comparing heuristics 3. Desirable floorplan aspect ratio r such that in this category. Of major importance however is the 1=r4H=W4r; where H and W are the height quality ofsolution(s) found by a particular heuristic for and width of the floorplan, respectively. the problem class at hand. That is, if for example, a 4. Timing information. giveniterativeapproximationalgorithmA1performson average better than a second algorithm A2 over a Output: A legal floorplan is a floorplan satisfying the numberofprobleminstances,wethenconcludethatA1 following constraints and objectives: implementation is better than that of A2. Another 1. Each block b is assigned to a location (x;y); legitimate comparison would be with respect to the i i i 2. no two blocks overlap; quality of the solution subspace examined by each 3. for each flexible block b; a ¼w (cid:5)h; algorithm. That is, if the number of solutions obtained i i i i 4. meet aspect ratio constraints, for each flexible with algorithm A1 with a cost lower than a particular block b and for the overall floorplan; and value C is higher than that of algorithm A2, we then i 5. minimize floorplan area, wire length and circuit conclude that A1 exhibits superior performance than delay. A2. Yet another comparison would be with respect to the progress of the best solution versus the number of Fig. 3(a) is a floorplan with 7 blocks. iterations. In this paper we report a number of experiments performed to compare the performance of 3. Solution encoding and cost evaluation three well-known approximation iterative algorithms with respect to criteria such as those described above. 3.1. Solution encoding The test problem used is floorplanning, a hard problem encounteredduringthephysicaldesignofVLSIcircuits. To have a simple and efficient encoding scheme the Itisencounteredinseveralotherareasofengineeringas floorplanner constrains the structure to be slicing, but well. In VLSI design, a floorplanning tool helps decide allows the circuit modules to have flexible shapes and several important questions such as dimensions, orien- free orientations. The layout is represented as a tationsandlocationofcells(orblocks),overallrequired repetitive division into basic rectangles by horizontal areaofthecircuit,andspeedperformance requirements and vertical cut-lines. Such a layout is called a slicing (Sait and Youssef, 1995). structure.Aslicingstructurecanbemodeledbyabinary The paper is organized as follows. Section 2 tree with n leaves and n(cid:6)1 nodes, where each node formulates floorplanning as a combinatorial optimiza- represents a vertical cut-line or horizontal cut-line, and tion problem. In Section 3 we describe how a floor- each leaf a basic rectangle. This binary tree is called planning solution is encoded and evaluated. The same slicing tree, and the corresponding floorplan is a slicing solutionencodingandevaluationfunctionareusedwith floorplan. Letters H and V refer to horizontal and all three approximation algorithms. Section 4 briefly vertical cut operators, respectively. A postorder traver- discusses the main features and operators of each of sal of a slicing tree produces a Polish expression with GA, SA, and TS algorithms. In Section 5, experimental operators H and V, and with operands the basic results are presented and the performance of the three rectangles 1;2;...;n (Wong and Liu, 1986). Fig. 3 gives heuristics are compared. We conclude in Section 6. a rectangular dissection, two possible slicing trees, and thecorrespondingPolishexpressions.Sincethereisonly 2. Floorplanning onewayofperformingapostordertraversalofabinary tree, there is one-to-one correspondence between floor- A possible formulation of the floorplanning problem plan trees and their corresponding Polish expressions. of VLSI circuits is as follows (Sait and Youssef, 1995): The operators H and V carry the following meanings: Given: ijH means rectangle j on-top-of rectangle i; ijV means rectangle i to-the-left-of rectangle j. In this work each 1. A set of n rectangular blocks B¼fb ;b ;...; solution is encoded as a Polish expression as in Wong 1 2 b;...;b g; for each b 2B we have and Liu (1986). i n i H.Youssefetal./EngineeringApplicationsofArtificialIntelligence14(2001)167–181 169 3.2. Cost evaluation A¼fðx;m ðxÞÞj all x2Xg, where X is a space of A points and m ðxÞ is a membership function of x2X A Unlike constructive algorithms for physical design, being an element of A. In general the membership which produce a solution only at the end of the design functionm ð:ÞisamappingfromX totheinterval[0,1]. A process, iterative algorithms operate with design solu- Ifm ðxÞ¼1,or0,forallx2X,thefuzzysetAbecomes A tionsdefinedateachiteration.Anevaluationfunctionis an ordinary set (Zadeh, 1965). used to compare results of consecutive iterations and to select a solution based on the maximal (minimal) value Example. As an example, let h refer to height of an of the objective function. athlete,and‘‘tall’’consideredasaparticularvalueofthe The cost function is a measure of how good is a fuzzy variable ‘‘height’’. Then for each h in the range particularfloorplan.Forfloorplanning,threecriteriaare from 0 to h , m ðhÞ2½0;1(cid:13) indicates the extent to max tall ofprimeimportance:area,wirelength,andtiming.Area which h is considered to be tall; m ðhÞ is called the tall is estimated by the size of the smallest bounding membership function. rectangle of the floorplan. Wire length is set equal to thelengthofallthenets(orinterconnectsoftheblocks) Example. As another example, consider the possible of the circuit. The length of an individual net or heights of sportsmen in feet to be H ¼ interconnect is estimated using the Steiner tree approx- f3:5;4:0;4:5;5:0;5:5;6:0;6:5g. Heights around 4.5 feet imation of Youssef et al. (1995). The delay along the are considered short (S), around 5.5 feet are considered longestpathofthecircuit,takingintoconsiderationthe moderate (M), and heights around 6.0 feet are estimated interconnect length, is used as a measure of considered tall (T). Thus, short, moderate, and tall are the timing performance of the floorplan (Youssef et al., not crisply defined. Fuzzy sets for short, moderate, and 1995). tall may be expressed as sets of ordered pairs The cost function should reflect all objectives. fðx;m ðxÞÞjall x2Xg, where the first element of the A Traditionally, multiple objectives have been combined pair is the height and the second is its membership in intoascalarobjectivefunction,usuallythroughalinear thatset.Fromourpreviousknowledgewecandefinethe combination (weighted sum) of the multiple attributes, fuzzy sets S, M and T as follows: or by turning objectives into constraints. One way is to assign a constant weight to each of the multiple S ¼fð3:5;1:0Þ; ð4:0;1:0Þ; ð4:5;1:0Þ; ð5:0;0:3Þ; objective functions whose value will depend on the importance of that objective. Assuming that all objec- ð5:5;0:3Þ; ð6:0;0:0Þ; ð6:5;0:0Þg tives are to be maximized, the cost of a floorplan M ¼fð3:5;0:0Þ; ð4:0;0:0Þ; ð4:5;0:1Þ; ð5:0;0:5Þ; solution x can be expressed as follows: ð5:5;1:0Þ; ð6:0;0:5Þ; ð6:5;0:1Þg costðxÞ¼w (cid:9)f ðxÞþw (cid:9)f ðxÞþ(cid:9)(cid:9)(cid:9)þw (cid:9)f ðxÞ; ð1Þ 1 1 2 2 n n where n is the number of objective functions, and w is T ¼fð3:5;0:0Þ; ð4:0;0:0Þ; ð4:5;0:0Þ; ð5:0;0:1Þ; i the weight of the ith objective f. The problem with this i ð5:5;0:4Þ; ð6:0;1:0Þ; ð6:5;1:0Þg weighted sum approach is the difficulty in determining suitable weights. Adhoc methods are sometimes em- ployed. Fig. 1 illustrates thethree membershipfunctions.For Fuzzy logic. The logic used to infer a crisp outcome simplicity we have used piecewise linear membership from fuzzy input values is called fuzzy logic (Zadeh, functions.Membershipfunctionscanalsobecontinuous 1965; Zimmermann, 1996; Kartalopoulos, 1996). Using curves of many different shapes. For example, Fig. 2 fuzzy logic approach, optimization of a vector-valued illustrates a continuous membership function of a fuzzy function is replaced by the optimization of a scalar set: An individual’s weight around 50kg. function, which is constructed from the levels of Fuzzy operators: As seen above, in fuzzy logic, the satisfaction of decision-makers by values of the compo- values are not crisp, and their fuzziness exhibits a nents of a vector function. Below, we present a brief distribution described by the membership function. In introductiontofuzzylogic,andthenshowhowitcanbe ordinary set theory, operations such as union ([), applied in optimization problems with multiple objec- intersection (\), and complementation (:) are used. tives. Whatistheresultoftheseoperationsonfuzzysets?This Fuzzy set theory: Unlike in ordinary set theory where question has been addressed by various fuzzy logics. an element is either in a set or not in a set, in fuzzy set These are logics that have been defined for operations theory, an element may partially belong to a set. Lotfi on fuzzy sets. One such logic defined by Zadeh is Zadeh defined a fuzzy set as a class of objects with a calledthemin–maxlogic(Zadeh,1965).Thereareother continuum of grades of membership. Formally, a fuzzy fuzzy logics, one of which we used is discussed later. In set A of a universe of discourse X ¼fxg is defined as min–max logic, the ‘‘union’’, ‘‘intersection’’ and 170 H.Youssefetal./EngineeringApplicationsofArtificialIntelligence14(2001)167–181 For floorplanning, the designer seeks to find floor- plansoptimizedwithrespecttotheareaoccupiedbythe floorplan (area occupied by the cells and wires), wire length required to interconnet all the cells, and delay along the longest path of the corresponding circuit (delays due to the cells and wires). Therefore, the objective function is not a scalar, but a vector function F(cid:2)ðxÞ¼ðAreaðxÞ;LengthðxÞ;DelayðxÞÞ; where x is a particular floorplan solution with area AreaðxÞ, wire length LengthðxÞ, and delay performance Fig.1. Membershipfunctionsforshort,moderate,andtall. DelayðxÞ (Fig. 3). To obtain a fuzzy logic definition of the above multicriteriaobjectivefunction,threelinguisticvariables area, length, and delay are defined Shragowitz et al., 1997). In this implementation, only one linguistic value is defined for each variable. That is, small area, short length,and low delay. Theselinguistic values character- ize the degree of satisfaction of the designer with the values of objectives Area, Length, and Delay of the floorplan.Thesedegreesofsatisfactionaredescribedby membership functions m ; m , and m on fuzzy sets of A L D linguisticvaluessmallarea,shortlength,andlowdelay. Membership functions for small area, short length Fig.2. Continuousmembershipfunctionforthefuzzyset:individual’s andlowdelayareeasytobuild.Theyarenon-increasing weightaround50Kg. functions,since thesmaller thearea,thelength,andthe delay,thehigheristhedegreeofsatisfactionm ,m ,and A L ‘‘complementation’’ are defined as follows: m of the expert and vice versa (see Fig. 4). D m ðxÞ¼minðm ðxÞ; m ðxÞÞ; To make the membership functions applicable to A\B A B different designs, the base variables floorplan area, m ðxÞ¼maxðm ðxÞ; m ðxÞÞ; length and delay are normalized to the interval [0,1] A[B A B (see Fig. 4). The values A ;L , and D are lower min min min m:AðxÞ¼1:0(cid:6)ðmAðxÞÞ: boundsonthearea,totallengthandclockperiodofthe Fuzzy rules: Approximate reasoning can be made circuit. Amin is the sum of areas of all cells; Lmin is the basedon linguisticvariablesandtheir values.Rulescan smallest possible length, computed assuming all con- be generated based on previous experience. These rules nectedcellsareplacedadjacenttoeachother;Dministhe can then be expressed with If ... Then statements. smallest possible delay of the circuit, which is the delay Connectives such as AND and OR can be used in of the longest path due to cells and assuming minimum approximate reasoning to join two or more linguistic interconnect delay for the nets. The values of Amax, values. Lmax, and Dmax are values corresponding to the initial As mentioned above, in min–max logic, the fuzzy solution. AND is realized by the function min. If more than one The most desirable floorplan is the one with the rule is used to perform decision-making, each rule can highest membership in the fuzzy subsets small area, be evaluated to generate a numerical value. Then, these numerical values from various evaluations of different rules can be combined to generate a crisp value on a higher level of hierarchy. Fuzzification. In this work, we rely on the expressive power of fuzzy logic (Zadeh, 1965, 1975, 1973) to expressthedesignerobjectivesintheformoffuzzylogic rules. The evaluation of a floorplanning solution consists of the evaluation of the fuzzy rules which returnavaluerepresentingthedegreeofmembershipof thatsolutioninthefuzzysubsetofgoodfloorplans.The Fig. 3. (a) Slicing floorplan. (b) Slicing tree with Polish expression fuzzy logic rules are expressed on linguistic variables of VH21HHV67V453. (c) Another possible slicing tree with Polish the problem domain. expressionVH21HV67HV453. H.Youssefetal./EngineeringApplicationsofArtificialIntelligence14(2001)167–181 171 Fig.4. Normalizedmembershipfunctionsforarea,lengthanddelay. short length and low delay. Such a solution most likely aggregation technique. According to this orlike opera- does not exist since some of the criteria conflict with tion, the fuzzy logic rule R.0 evaluates to the following: each other. Therefore, one has to tradeoff these individual criteria against each other. This tradeoff is 1X3 m ðxÞ¼b(cid:5)maxðm ;m ;m Þþð1(cid:6)bÞ(cid:5) mðxÞ; conveniently specified in linguistic terms in the form of ðSÞ 1 2 3 3 i i¼1 one or several fuzzy logic rules. ð2Þ As mentioned earlier, a fuzzy logic rule is an If–Then where m is the membership function of the fuzzy rule.TheIfpart(antecedent)isafuzzypredicatedefined ðSÞ subsetofgoodsolutions,andbisaparameterbetween0 in terms of linguistic values and fuzzy operators (AND and 1 indicating the degree of nearness of this orlike andOR).TheThenpartiscalledtheconsequent.Inour operator to the strict meaning of the logical OR case, the linguistic value used in the consequent part operator. When b¼1, the orlike operator behaves like identifies the fuzzy subset of good solutions. Therefore, a regular max operator, and for b¼0 it behaves like a the result of evaluation of the antecedent part identifies weighted averaging operator. the degree of membership in the fuzzy subset of good Individual objective criteria can be given relative floorplan solutions according to the fuzzy rule in rankingtoreflecttheirimportance.Thiscanbeachieved question (Shragowitz et al., 1997). by assigning weights as exponents to individual mem- The fuzzy subset of good floorplan solutions is bership functions. When the membership function of a characterized by the following fuzzy rule: particular criterion is assigned a positive exponent less R:0 If ðsmall areaÞ OR ðshort lengthÞ OR ðlow delayÞ than 1, we say that the corresponding fuzzy subset is dilated. This corresponds to giving more empha- Then good solution: sis=importance to that criterion. Concentration corre- According to the traditional min–max fuzzy logic, the spondstothecaseofassigninganexponentgreaterthan above rule R.0 evaluates to the following: 1 (Mendel, 1995). For example, in the case of floor- planning problem, Area is a more significant criterion, m ðxÞ¼maxðm ;m ;m Þ; ðSÞ A L D whereas Length and Delay are of equal but less importance. Such preferences can be conveniently where m is the membership function of the fuzzy ðSÞ expressedbyusingdilationandconcentrationoperators. subset of good solutions, and m , m , and m are the A L D With these preferences, Eq. (2) becomes membershipfunctionsinthefuzzysubsetsofsmall-area solutions, short wire-length solutions, and small delay m ðxÞ¼b(cid:5)maxðm1=2;m2;m2Þþð1(cid:6)bÞ solutions, respectively. The min and max operators are ðSÞ A L D non-compensatory, i.e., the weaker elements cannot (cid:5)1ðm1=2þm2 þm2Þ: ð3Þ 3 A L D compensate for the stronger element. Such property is In the above equation, the fuzzy subset representing undesirable in the case of multi-criteria optimiza- solutions with small area is dilated, whereas the fuzzy tion=decision problems. For this category of problems subset of solutions with short length and the fuzzy it is important to make use of all the information subset of solutions with small delay are concentrated. sources. In this case it is recommended to use a compensatory interpretation of the fuzzy OR operator, which allows all the criteria elements to influence the outcome of the operator. One such interpretation is the 4. GA, SA, AND TS algorithms orlike ordered weighted averaging (OWA) operator suggested by Yager in Yager (1994); Yager and Filev Anyiterativeapproximationalgorithm,beitGA,SA, (1994) in the context of the softening of the rule or TS, requires decisions with respect to the following 172 H.Youssefetal./EngineeringApplicationsofArtificialIntelligence14(2001)167–181 elements: expressions with n operands and n(cid:6)1 operators. The fitness function is a measure of how good each * a solution encoding, individual of current generation is with respect to the * amechanismtogenerateinitialsolutionsfrom where desired features. For each chromosome x; fitness the iterative search will proceed, ðxÞ¼m ðxÞ, where m is as defined in Eq. (3). The * an evaluation function to rate solutions encountered ðSÞ ðSÞ choice of parents from the set of individuals that during the search, comprise the population (also known as the mating * perturbationoperatorstocreatenewsolutionsfroma pool) is probabilistic. In keeping with the ideas of given current solution, natural selection, we assume that stronger individuals, * assignment to the parameters of the algorithm, and that is, those with higher fitness values, are more likely * stopping criteria. to mate than the weaker ones. One way to simulate this Below, we provide a brief description of each of GA, is to select parents with a probability that is directly SA, and TS, and illustrate how each algorithm is proportional to their fitness values. This method is adapted to the floorplanning problem. called the roulette-wheel method (Goldberg, 1989). Crossover.Crossoverisamechanismforprobabilistic 4.1. Genetic algorithms inheritance of useful information from two fit indi- viduals to offsprings. The main idea is that the genetic Genetic algorithms (GAs) are robust and effective information of a good solution is spread over the entire optimization techniques inspired by the mechanism of population. Thus, the best solution can be obtained by evolutionandnaturalgeneticsGoldberg,1989;Holland, thoroughly combining the chromosomes in the popula- 1975. They are characterized by a parallel search of the tion.Crossoveroperationachievesrecombinationofthe state space as against a point-by-point search by the genetic material. The recombination process includes conventional optimization techniques. The parallel domain specific knowledge to enforce the inheritance of search is achieved by keeping a set of possible solutions desirable features from individuals of current popula- to the optimization problem, called population. An tion. The followingthree crossoveroperators havebeen individual in the population is a string of symbols and used in this work (Cohoon et al., 1991). is an abstract representation of the solution. The In the first crossover scheme, the operands from symbols are called genes and each string of genes is parent P are copied into the offspring O. Then, the 1 termedachromosome.Theindividualsinthepopulation operatorsfromparentP arecopiedbymakingaleft-to- 2 are evaluated by some fitness measure. The population right scan to complete the chromosome. Hence, each of chromosomes evolves from generation to the next offspring inherits the block structure of one of the through the use of two types of genetic operators: (1) parents. For example, if P ¼VH21HHV67V453 and 1 unary operators such as mutation and inversion which P ¼VH12HV47HV653, then, according to this cross- 2 alter the genetic structure of a single chromosome, and over, the offspring would be VH21HVH67V453. (2)higher-orderoperator,referredtoascrossoverwhich The second crossover operates by copying the consists of obtaining new individual by combining operators from parent P into the corresponding 1 geneticmaterialfromtwoselectedparentchromosomes. positions in offspring O. Then, it completes the Then the new population is selected out of the indi- construction of O by copying the operands from parent viduals of the current population and its offsprings. P by making a left-to-right scan. By propagating 2 Based on the fitness value, two individuals (parents) are the groups of operators from P to O, this cros- 1 selected at a time from the population. The genetic sover produces an offspring having the same overall operators (crossover and mutation) are applied on the slicing structure as that of P . We call this crossover 1 selectedparentstogeneratenewpossiblesolutionscalled ‘‘slicing inheritance crossover’’. For example, if offsprings. The structure of basic GA is given in Fig. 5. P ¼VH21HHV67V453 and P ¼VH12HV47HV653, 1 2 Theapplication ofgeneticalgorithm foranyproblem then,accordingtothiscrossover,theoffspringwouldbe requires a representation of the solution to the problem VH12HHV47V653. Both the above crossoversproduce as a string of symbols, a choice of genetic operators, an valid Polish expressions. evaluation function, a selection mechanism, and deter- ThethirdcrossoveremployedinthisworkisthePMX mination of genetic parameters (population size and crossover (partially mapped crossover) which has been probabilities of crossover, mutation, and inversion). widely used in other genetic implementations for Each of these greatly influences the performance of the combinatorial optimization problems (Shahookar and genetic algorithm. Mazumder,1991).PMX isapplied onlyonthemodules In this work, a chromosome in the population (operands), the operatorsand thepositionsfrom one of (solution)isaPolishexpressionrepresentingaparticular the parents are used in the offspring (see Figs. 6 and 7). slicing floorplan (see Section 3.1, Fig. 3). The initial Mutation. Mutation is a means of introducing new population consists of randomly generated Polish informationintothepopulation.Threesimplemutation H.Youssefetal./EngineeringApplicationsofArtificialIntelligence14(2001)167–181 173 Fig.5. Structureofgeneticalgorithmforfloorplanning. operators are used in this work (Wong and Liu, 1986): Inversion. This operator consists of randomly select- (1) inversion of a randomly selected chain of operators, ing two positions within the string (a chromosome) and (2)swappingoftworandomlyselectedoperands,and(3) then laterally inverting the string between them. In this swapping an operand with an adjacent operator work,itisimplementedsoastomakeafullright-to-left (provided it results in a valid Polish expression). or left-to-right lateral inversion on the entire string of 174 H.Youssefetal./EngineeringApplicationsofArtificialIntelligence14(2001)167–181 Fig.6. Crossoverthatpropagatesslicingstructuretotheoffspring. Fig.7. IllustrationofPMXcrossover. operands(only).Forexample,ifthechromosomebefore (higher fitness translates to higher survival). The inversion is 268HH7H5V4H13VV, then after inversion probability of mutation is set to 5%, and the prob- thenewsolutiongeneratedwillbe:314HH5H7V8H62VV. ability of crossover to 80%. Inversion is applied A number of methods have been proposed to select with a probability of 30%. Population size chosen is individuals that can survive and be used in the next 10, and in each generation the number of offsprings generation (Goldberg, 1989). In this work, among the producedisthesameasthenumberofparents.Stopping parents and their offsprings, the chromosomes for the criterion in this case was set to a fixed number of next generation are chosen based on their fitness values generations. H.Youssefetal./EngineeringApplicationsofArtificialIntelligence14(2001)167–181 175 4.2. Simulated annealing peratures.TemperatureisinitializedtoavalueT atthe 0 beginning of the procedure (explained below), and is Simulated annealing (SA) is an iterative search slowly reduced (in a geometric progression); the par- methodinspiredbytheannealingofmetals(Kirkpatrick ameter a is used to achieve this cooling. The amount of et al., 1983; Cerny, 1985). Starting with an initial time spent in annealing at a temperature is gradually solution and armed with adequate perturbation and increased as temperature is lowered. This is done using evaluation functions, the algorithm performs a stochas- the parameter b>1. The variable Time keeps track of tic partial search of the state space. Uphill moves are the time being expended in each call to the Metropolis. occasionallyacceptedwithaprobabilitycontrolledbya The annealing procedure halts when Time exceeds the parameter called temperature ðTÞ. The probability of allowed time. acceptanceofuphillmoves decreasesasT decreases.At The Metropolis procedure is shown in Fig. 9. It uses hightemperature,the search isalmostrandom,whileat the procedure Neighbor to generate a local neighbor low temperature the search becomes almost greedy. At NewS of any given solution S. The function Fitness zerotemperature,thesearchbecomestotallygreedy,i.e., returns the fitness of a given solution S. If the fitness of only good moves are accepted (Kirkpatrick et al., 1983; the new solution NewS is better than the fitness of the Cerny, 1985). current solution S , then the new solution is accepted, c ThebasicSAalgorithmisshowninFig.8.Thecoreof and we do so by setting CurS ¼NewS. If the fitness of the algorithm is the Metropolis procedure, which the new solution is better than the best solution (BestS) simulates the annealing process at a given temperature seen thus far, then we also replace BestS by NewsS. If T (Fig. 9) Metropolis et al., 1953. This procedure is thenewsolutionhasalowerfitnessincomparisontothe named after the scientist who devised a similar scheme original solution CurS, Metropolis will accept the new to simulate a collection of atoms in equilibrium at a solution on a probabilistic basis. A random number is giventemperature.TheMetropolisprocedurereceivesas generated in the range 0 to 1. If this random number is input the current temperature T, and the current smaller than eDfitness=T, where Dfitness¼fitness solution CurS which it improves through local search. ðNewSÞ(cid:6)FitnessðcurSÞ, and T is the current tempera- Metropolis must also be provided with the value M, ture, the inferior solution is accepted. This criterion for whichistheamountoftimeforwhichannealingmustbe accepting the new solution is known as the Metropolis applied at temperature T. The procedure Simulated_an- criterion. The Metropolis procedure generates and nealing simply invokes Metropolis at decreasing tem- examines M solutions. Fig.8. Procedureforsimulatedannealingalgorithm. 176 H.Youssefetal./EngineeringApplicationsofArtificialIntelligence14(2001)167–181 anumberofstatetransitionsaremadesoastoreachthe probabilistic steady state. The perturb mechanisms employed are similar to the mutation operators in GA. The stopping criterion is whenthefinalT50:001orA=M50:05,whereA isthe t t t totalnumberofallacceptedmoves(bothgoodandbad) and M is the number of all moves made at a given t temperature. 4.3. Tabu search Tabu search is a higher-level method or meta strat- egy for solving combinatorial optimization problems (Glover, 1989, 1990). It is an iterative procedure that startsfromsomeinitialfeasiblesolutionandattemptsto determineabettersolution.TSmakesseveralneighbor- hood moves and selects the move producing the best solution among all candidate moves for current itera- Fig.9. TheMetropolisprocedure. tion. This best candidate solution may not improve the current solution. Selecting the best move (which may or may not Initial temperature. The SA algorithm needs to start improve the current solution) is based on the suppo- from a high temperature ðTÞ. However, if this initial sition that good moves are more likely to reach the value of T is too high, it causes a waste of processing optimalornear-optimalsolutions.Thesetofadmissible time.Theinitialtemperaturevalueshouldbesuchthatit solutions attempted at a particular iteration forms a allows virtually all proposed uphill or down hill moves candidate list. TS selects the best solution from the to be accepted. The temperature parameter is initialized candidate list. Candidate list size is a trade-off between using the procedure described in Wong and Liu (1986). quality and performance. The idea is basically to use the Metropolis function To avoid move reversals, a device called tabu (eDFitness=T) to determine the initial value of the restrictionisusedthatmakesselectedattributesofthese temperature parameter. Before the start of actual SA moves tabu (forbidden). Tabu restrictions allow the procedure, a constant number of moves, say M, in the search togobeyondthepointsoflocaloptimalitywhile neighborhoodofthecurrent solutionaremade,andthe still making the best possible move in each iteration. respective fitness values of these moves are determined. Taburestrictionsareenforcedbyatabulistwhichstores Thefitnessdifferenceforeachmovei; DFitness isgiven the move attributes to avoid move reversals. Tabu list i by DFitness ¼Fitness (cid:6)Fitness .Let M and M be hasanassociatedsizeandcanbevisualizedasawindow i i i(cid:6)1 u d the number of uphill and downhill moves, respectively on accepted moves. The moves which tend to undo (downhill refers to inferior moves) (M ¼M þM ). moves within this window are forbidden. u d The average DFitness is then given by Aspiration level component is used to temporarily d override the tabu status if the move is sufficiently good. 1 XMd If a move is made tabu in iteration i and its reversal DFitness ¼ DFitness: d Md i¼1 i comes in iteration j, where j5iþc, (where c is the size ofthetabulist),thenitispossiblethatthereversemove Since we wish to keep the probability, say P0, of may take the search into a new region because of the acceptinguphillmoveshighintheinitialstageofSA,we effects of c intermediate moves. The simplest aspiration estimate the value of the temperature parameter by criterion, also known as the ‘‘best cost aspiration substituting the value of P0 in the following expression criterion’’, consists of overriding the tabu status if the derived from the Metropolis function: reversal produces a solution better than the best DFitness obtained thus far. Another approach is to use the same T0 ¼ lnðP Þ d; attribute of the move which is used to identify the tabu 0 statusandassociateanaspirationlevelvaluewithit.The where P (cid:17)1 ðP ¼0:999Þ. reversal has to do better than this historical aspiration 0 0 Parameters, stopping criterion. The parameter a for level. Fig. 10 illustrates the flow of the short-term tabu updating the temperature is user specified. In our search algorithm. implementation a takes on the value of 0.85, and b is Only short-term tabu search was implemented in this setequalto1.Ateachupdatedvalueofthetemperature, work. The program is executed for 250 iterations. The

Description:
Evolutionary algorithms, simulated annealing (SA), and tabu search (TS) are general iterative algorithms for combinatorial optimization. The term evolutionary algorithm is used to refer to any probabilistic algorithm whose design is inspired by evolutionary mechanisms found in biological species.
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.