A new hybrid neuro-fuzzy model with reinforcement learning for intelligent agents

This work presents a new hybrid neuro-fuzzy model for the automatic learning of actions taken by agents. The objective of this model is to provide intelligence for an agent, making it capable of acquiring and retaining knowledge, as well as thinking (infer an action), by interacting with its environment. This new model, named Reinforcement Learning Hierarchical Neuro-Fuzzy Politree (RL-NFHP), descend from the hierarchical BSP neuro-fuzzy models, which employ supervised learning and BSP partitioning (Binary Space Partitioning) of the input space. By using this hierarchical partitioning method, together with the Reinforcement Learning methodology, a new class of Neuro-Fuzzy Systems (SNF) was obtained, which automatically learns its structure as well as the actions that must be taken by an agent. These characteristics represent an important differential when compared to existing intelligent agents learning systems. The RL-NFHP model was tested in different benchmark problems, as well as in a robotic application (Khepera robot). The results obtained demonstrate the potential of the proposed model, which does without information as number of rules, rules' format and number of partitions that the input space should have.