Genetic-based reinforcement learning for fuzzy logic control systems

This paper proposes a genetic-based reinforcement learning for fuzzy logic control systems (GR-FLCS) to solve reinforcement learning problems. The proposed GR-FLCS is constructed by integrating a real-coded genetic algorithm with a time accumulator as the fitness evaluator, a success criterion, a fuzzy logic controller (FLC), and a parameter adapter for the FLC. In this simple but powerful architecture, restrictions, usually met in reinforcement learning for FLCs, can be taken off completely. They are, the FLC must be implemented by a neuronlike network, the shapes of the membership functions in the FLC must be in some form, e.g., bell-shaped, the fuzzy operators must be modified, or only the consequent part of the rule base in FLC can be learned. Finally, the applicability and efficiency of GR-FLCS are demonstrated by an simulation example of the cart-pole balancing problem.