Symbiotic Evolution Genetic Algorithms for Reinforcement Fuzzy Systems Design

The advent of fuzzy logic controllers has inspired the allocation of new resources for the possible realization of more efficient methods of control. In comparison with traditional controller design methods requiring mathematical models of the plants, one key advantage of fuzzy controller design lies in its model-free approach. Conventionally, the selection of fuzzy if-then rules often relies heavily upon the substantial amounts of heuristic observation to express the strategy's proper knowledge. It is very difficult for human experts to examine all the input-output data from a complex system, and then to design a number of proper rules for the fuzzy logic controllers. Many design approaches for automatic fuzzy rules generation have been developed in an effort to tackle this problem (Lin & Lee, 1996). The neural learning method is one of them. In (Miller et al., 1990), several neural learning methods including supervised and reinforcement based control configurations are studied. For many control problems, the training data are usually difficult and expensive, if not impossible, to obtain. Besides, many control problems require selecting control actions whose consequences emerge over uncertain periods for which training data are not readily available. In reinforcement learning, agents learn from signals that provide some measure of performance which may be delivered after a sequence of decisions being made. Hence, when the above mentioned control problems occur, reinforcement learning is more appropriate than supervised learning. Genetic algorithms (GAs) are stochastic search algorithms based on the mechanics of natural selection and natural genetics (Goldberg, 1989). Since GAs do not require or use derivative information, one appropriate application for their use is the circumstance where gradient information is unavailable or costly to obtain. Reinforcement learning is an example of such domain. The link of GAs and reinforcement learning may be called genetic reinforcement learning (Whitley et al., 1993). In genetic reinforcement learning, the only feedback used by the algorithm is the information about the relative performance of different individuals and may be applied to reinforcement problems where the evaluative signals contain relative performance information. Besides GAs, another general approach for realizing reinforcement learning is the temporal difference (TD) based method (Sutton & Barto, 1998). One generally used TD-based reinforcement learning method is Adaptive Heuristic Critic (AHC) learning algorithm. AHC learning algorithm relies upon both the learned evaluation O pe n A cc es s D at ab as e w w w .ite ch on lin e. co m