HIERARCHICAL NEURO-FUZZY MODELS BASED ON REINFORCEMENT LEARNING FOR AUTONOMOUS AGENTS

This work introduces a new class of neuro-uzzy systems for intelligent agents, called ReinfolrenteUt Learning - Hierarchical Neuro-Puzzy Systent. This new class combines a hierarchical partitioning method of the input space with a Reinforcement Learning algorithm to achieve the following important characteristics: automatic creation of the model's structure; self-adjustment of the pammeters; autonomous learning of the actions; capacity to deal with a greater number of inputs; and automatic generation of linguistic fuzzy rules. The proposed model was devised to overcome limitations of traditional reinforcement learning methods based on lookup tables, particularly in applications involving continuous environments and/or environments considered to he high dimensional. The paper details the hierarchical neuro-fuzzy architecture, its basic cell, and the learning algorithm. The performance of the proposed system was evaluated in four benchmark applications the Mountain Car Problem, the Cart-Centering Problem. the Inverted Pendulum and the Khepera Robot Control. The results obtained demonstrate the capacity of the novel hierarchical neuro-fuzzy system to automatically extract knowledge from the agent's direct interaction with large and/or continuous ClItliVOTIMCIttS. This knowledge is in the form of fuzzy linguistic rules, with no prior definition of the number and position of the fuzzy sets.