A study on multi-dimensional fuzzy Q-learning for intelligent robots

Reinforcement learning is one of the most important learning methods for intelligent robots working in unknown/uncertain environments. Multi-dimensional fuzzy Q-learning, an extension of the Q-learning method, has been proposed in this study. The proposed method has been applied for an intelligent robot working in a dynamic environment. The rewards from the evaluation functions and the fuzzy Q-values generated by the neural networks (fuzzy Q-net) are expressed in vector forms in order to obtain optimal behaviors for several different purposes. By applying this learning method, evaluation and learning of fuzzy Q-values for the other behaviors can be carried out simultaneously in one trial. We express fuzzy states as the vector of fuzzy sets for input variables of the fuzzy Q-net. The behavior selection algorithm is also proposed in this study. The simulation results show the effectives of the proposed methods for a mobile robot selects optimal behavior in a dynamic environment.