A symbiotic genetic algorithm with local-and-global mapping search for reinforcement fuzzy control

Abstract

This paper proposes a Symbiotic Genetic Algorithm with Local-and-Global mapping search (SGA-LG) for fuzzy controller design under reinforcement learning environments. The objective of the proposed SGA-LG is to increase the reinforcement fuzzy controller design efficacy and efficiency. SGA-LG operates in two concurrently evolving searches: the local mapping search and the global mapping search. The local-mapping search helps to find the well-performed local rules. A population is created in this search, and each individual in the population encodes only one fuzzy rule. An elite strategy is adopted, where the top-half best-performing individuals, the elites, are reproduced directly to the next generation, and parents are selected from the elites only. For global-mapping search, another population is created, where each individual encodes a whole fuzzy network as opposed to a single rule. The objective is to determine which local rules designed in the local-mapping search should be combined together to achieve a good fuzzy network. To demonstrate the performance of SGA-LG, it is applied to cart-pole and ball-and-beam system controls. The efficacy and efficiency of SGA-LG are verified by comparing with other GAs, evolution strategy and evolutionary programming based fuzzy controller designs.