Learning and Design of Fuzzy Logic Controllers For Pursuit-Evasion Differential Games

This thesis presents different learning techniques for robotic applications. We show how to automatically and adaptively tune the input and the output parameters of a fuzzy logic controller in different situations using different types of learning. We focus on three robotic applications; the wall-following case, the single pursuit-evasion case and the multiple pursuit-evasion case.;The second learning type is reward-based learning in which we do not have an output data set or an expert to learn from. Instead, we need a reward or a punishment function for evaluating the learning model. Under this type, we use genetic algorithms and reinforcement learning methods to tune the input and the output parameters of fuzzy logic controllers. We propose a new technique that combines genetic algorithms with reinforcement learning methods to tune the input and the output parameters of fuzzy logic controllers. The proposed technique is applied to different pursuit-evasion differential games starting from a single pursuit-evasion case in which only the pursuer self-learns its control strategy and ending with decentralized learning in which n pursuers and n evaders with different capabilities can self-learn their control strategies. We also propose a new technique that uses reinforcement learning methods to tune the input and the output parameters of fuzzy logic controllers. We also present a convergence proof for the model-building Q-learning algorithm with eligibility traces.;In this thesis, we focus on two learning types. The first type is supervised learning in which we need input/output data or an expert to learn from. Under this type, we use genetic algorithms and gradient-based techniques to tune the input and the output parameters of fuzzy logic controllers. We propose a new technique that uses input/output training data obtained from a PD controller" an expert" to iteratively tune the input and the output parameters of fuzzy logic controllers using genetic algorithms. The number of rules in the rule-base is also reduced. The proposed technique is applied to a wall-following mobile robot. We also propose another technique that uses input/output training data obtained from a PD controller to iteratively tune the input and the output parameters of fuzzy logic controllers using gradient-based techniques. The proposed technique is applied to a pursuit-evasion game.