Unapređenje hibridizacijom metaheuristika inteligencije rojeva za rešavanje problema globalne optimizacije

Abstract

Te²ki optimizacioni problemi, nere²ivi u prihvatljivom vremenu izvr²avanja deterministi £kim matemati£kim metodama, uspe²no se poslednjih godina re²avaju populacionim stohasti£kim metaheuristikama, me u kojima istaknutu klasu predstavljaju algoritmi inteligencije rojeva. U ovom radu razmatra se unapre enje metaheuristika inteligencije rojeva pomo¢u hibridizacije. Analizom postoje¢ih metaheuristika u odre enim slu£ajevima uo£eni su nedostaci i slabosti u mehanizmima pretrage prostora re²enja koji pre svega proisti£u iz samog matemati£kog modela kojim se simulira proces iz prirode kao i iz nedovoljno uskla enog balansa izme u intenzikacije i diversikacije. U radu je ispitivano da li se postoje¢i algoritmi inteligencije rojeva za globalnu optimizaciju mogu unaprediti (u smislu dobijanja boljih rezultata, brºe konvergencije, ve¢e robustnosti) hibridizacijom sa drugim algoritmima. Razvijeno je i implementirano vi²e hibridizovanih metaheuristika inteligencije rojeva. S obzirom da dobri hibridi ne nastaju slu£ajnom kombinacijom pojedinih funkcionalnih elemenata i procedura razli£itih algoritama, ve¢ su oni utemeljeni na sveobuhvatnom izu£avanju na£ina na koji algoritmi koji se hibridizuju funkcioni²u, kreiranju hibridnih pristupa prethodila je detaljna analiza prednosti i nedostataka posmatranih algoritma kako bi se napravila najbolja kombinacija koja nedostatke jednih neutrali²e prednostima drugih pristupa. Razvijeni hibridni algoritmi verikovani su testiranjima na standardnim skupovima test funkcija za globalnu optimizaciju sa ograni£enjima i bez ograni£enja, kao i na poznatim prakti£nim problemima. Upore ivanjem sa najboljim poznatim algoritmima iz literature pokazan je kvalitet razvijenih hibrida, £ime je potvr ena i osnovna hipoteza ovog rada da se algoritmi inteligencije rojeva mogu uspe²no unaprediti hibridizacijom...

Hard optimization problems that cannot be solved within acceptable computational time by deterministic mathematical methods have been successfully solved in recent years by population-based stochastic metaheuristics, among which swarm intelligence algorithms represent a prominent class. This thesis investigates improvements of the swarm intelligence metaheuristics by hybridization. During analysis of the existing swarm intelligence metaheuristics in some cases deciencies and weaknesses in the solution space search mechanisms were observed, primarily as a consequence of the mathematical model that simulates natural process as well as inappropriate balance between intensication and diversication. The thesis examines whether existing swarm intelligence algorithms for global optimization could be improved (in the sense of obtaining better results, faster convergence, better robustness) by hybridization with other algorithms. A number of hybridized swarm intelligence metaheuristics were developed and implemented. Considering the fact that good hybrids are not created as a random combination of individual functional elements and procedures from dierent algorithms, but rather established on comprehensive analysis of the functional principles of the algorithms that are used in the process of hybridization, development of the hybrid approaches was preceded by thorough research of advantages and disadvantages of each involved algorithm in order to determine the best combination that neutralizes disadvantages of one approach by incorporating the strengths of the other. Developed hybrid approaches were veried by testing on standard benchmark sets for global optimization, with and without constraints, as well as on well-known practical problems. Comparative analysis with the state-of-the-art algorithms from the literature demonstrated quality of the developed hybrids and conrmed the hypothesis that swarm intelligence algorithms can be successfully improved by hybridization...