Generalized Reinforcement Learning Fuzzy Control with Vague States

摘要

This paper presents a generalized method for tuning a fuzzy logic controller based on reinforcement learning in a dynamic environment. We extend the Generalized Approximate Reasoning-base Intelligent Controller (GARIC) model of Berenji and Khedkar to be able to work in presence of vagueness in states. Similar to GARIC, the proposed architecture, i.e., Generalized Reinforcement Learning Fuzzy Controller (GRLFC), has the self-tuning capability even when only a weak reinforcement signal such a binary failure signal is available. The proposed controller shows a better performance, regarding learning speed and robustness to changes in controlled system dynamics, than similar models even in the presence of uncertainty in states.