[{"type":"speech","title":"Towards a generally applicable self-adapting hybridization of evolutionary algorithms","issued":{"date-parts":[["2004"]]},"author":[{"family":"Jakob","given":"W."},{"family":"Blume","given":"C."},{"family":"Bretthauer","given":"G."}],"note":"Deb, K. [Hrsg.] Genetic and Evolutionary Computation : GECCO 2004 ; Conf., Seattle, Wash., June 26-30, 2004 Proc.Part I Berlin [u.a.] : Springer, 2004 (Lecture Notes In Computer Science ; 3102) incl.CD-ROM","abstract":"1\n{wilfried.jakob, georg.bretthauer}@iai.fzk.de\n2\nUniversity of Applied Sciences, Cologne, Campus Gummersbach, Am Sandberg 1,\n51643 Gummersbach, Germany\nAbstract. Practical applications of Evolutionary Algorithms (EA) frequently\nuse some sort of hybridization by incorporating domain-specific knowledge,\nwhich turns the generally applicable EA into a problem-specific tool. To overcome this limitation, the new method of HyGLEAM was developed and tested\nextensively using eight test functions and three real-world applications. One basic kind of hybridization turned out to be superior and the number of evaluations was reduced by a factor of up to 100.\n1\nIntroduction\nWhen applied to real-world problems, the powerful optimization tool of Evolutionary\nAlgorithms frequently turns out to be too time-consuming due to elaborate fitness calculations that are often based on run-time-intensive simulations. Incorporating domain-specific knowledge by problem-tailored heuristics or local searchers is a commonly used solution, but turns the generally applicable EA into a problem-specific\ntool. The new method of hybridization implemented in HyGLEAM (Hybrid GeneraL\npurpose Evolutionary Algorithm and Method) [1, 2] is aimed at overcoming this limitation and getting the best of both algorithm classes: a fast, global searching and robust procedure with the convergence reliability of evolutionary search being maintained. The basic idea of the concept can be summarized in two points:\n1. Usage of generally applicable local search algorithms instead of the commonly\nused problem-specific ones for hybridization.\n2. Usage of a convergence-dependent control mechanism for distributing the computational power between the basic algorithms for suitable kinds of hybridization.\nThe first point may appear simple, but it is a matter of fact that nearly all realworld applications and investigations are based on problem-specific local searchers.\nAppropriate local search algorithms for parameter optimization must be derivativefree and able to handle restrictions in order to be generally applicable. The Rosenbrock procedure and the Complex algorithm, two well-known powerful local searchers [3], were chosen, as they fulfill these requirements. GLEAM (General Learning\nEvolutionary Algorithm and Method) [4] was used as an EA, but it must be noted that\nthe method can be applied easily to every other population-based EA.\n2\nExperiments and Conclusions\nThe test cases comprised real, integer, and mixed parameter optimization, combinatorial and multi-objective optimization as well as parameter strings of dynamic length.\nThey are described in more detail together with references in [2, 5]. In most cases, the\nresults were based on an average of 100 runs per algorithm and parameterization.\nFour basic kinds of hybridization were investigated:\n1. Pre-optimization of the start population: The idea is that the evolution can start\nwith solutions of more or less good quality. It works pretty well (up to 24 times less\nevaluations) in some cases, but not always and more evaluations may be required.\n2. Post-optimization of the EA results: As EAs are known to converge slowly, an improvement may result from stopping the evolution after approaching the area of attraction of the (global) optimum and leaving the rest to the local search. The appropriate switching point is determined by the convergence-dependent control\nprocedure mentioned above. This approach improves the EA results, but does not\nfulfill the expectation of reliably finding the solution.\n3. Direct integration: Optimizing every or the best offspring of one mating only\ncauses the EA to operate over the peaks of the fitness landscape exclusively rather\nthan to treat the valleys and slopes, too. The offspring\u2019s genotype can be updated\n(Lamarckian evoluti","kit-publication-id":"240058125"}]