Reinforcement learning based on modular fuzzy model with gating unit

Abstract

In order to realize intelligent agent such as autonomous mobile robots, reinforcement learning is one of necessary techniques in behavior control system. However, applying the reinforcement learning to actual sized problem, the ldquocurse of dimensionalityrdquo problem in partition of sensory states should be avoided maintaining computational efficiency. Furthermore the robot task is desired to be decomposed automatically in learning process for achievement of good performance. We tackle these two issues by applying modular fuzzy model with gating unit to reinforcement learning. The modular fuzzy model extending SIRMs architecture is formulated to apply Q-learning algorithm. The gating unit that is constructed as a neural network model or simple learning parameters is installed to switch the use of the modular model for task decomposition. Through numerical examples, we found that the proposed method has fair convergence property of learning compared with the conventional model structure.