Improving genetic algorithms: an approach based on multi-elitism and Lamarckian mutation
Domenico Bellomo, David Naso, Biagio Turchiano · 2003
This paper describes a genetic algorithm for multi-modal optimization combining fitness sharing and local search (Nelder and Mead's Simplex). While conventional fitness sharing tends to distribute the population across all the optima locating the peaks with low accuracy, the proposed algorithm aims at clearly discovering the position of each optimum with a higher accuracy. To pursue this objective, the algorithm uses fitness sharing to locate the basins of attraction of the various optima, and a special mutation operator (Lamarckian mutation) assuring an accurate local search. Furthermore, differently from conventional elitist genetic algorithms, rather than copying only the single best solution, our algorithm replicates the best solution found in every peak (residents) in the next population (multi-elitism). The simulated benchmark used to evaluate the approach encompasses both well-known multi-modal functions and the application to a fuzzy controller design problem.