Approximating Arbitrary Reinforcement Signal by Learning Classifier Systems using Micro Genetic Algorithm

摘要: Learning Classifier Systems are Evolutionary mechanisms which combine Genetic Algorithm and the Reinforcement paradigm. try to evolve state-action-reward mappings propose best action for each environmental state maximize achieved reward. In first versions of learning classifier systems, state-action pairs can only be mapped a constant real-valued So model fairly complex environment, LCSs had develop redundant different reward values. But an extension well-known LCS, called Accuracy Based System or XCS, was recently developed able map linear function. This new extension, XCSF, more compact population than original XCS. some further researches have shown that this is not proper when input parameters from certain intervals. As solution issue, in our previous works, we proposed novel inspired by idea using evolutionary approach approximate landscape. The results seem promising, but approach, XCSFG, converged goal very slowly. paper, XCSFG employs micro-GA its needed extremely smaller simple GA. expect help converge faster. Reported show assumed as alternative XCSF family with respect convergence speed, approximation accuracy compactness.