Genetic Reinforcement Learning Algorithms for On-line Fuzzy Inference System Tuning "Application to Mobile Robotic"

Abstract

In the last decade, fuzzy logic has supplanted conventional technologies in some scientific applications and engineering systems especially in control systems, particularly the control of the mobile robots evolving (moving) in completely unknown environments. Fuzzy logic has the ability to express the ambiguity of human thinking and translate expert knowledge into computable numerical data. Also, for real-time applications, its relatively low computational complexity makes it a good candidate. A fuzzy system consists of a set of fuzzy if-then rules. Conventionally, the selection of fuzzy if-then rules often relies on a substantial amount of heuristic observation to express the knowledge of proper strategies. Recently, many authors proved that it is possible to reproduce the operation of any standard continuous controller using fuzzy controller L. Jouffe, C. Watkins, P. Dayan Dongbing Gu, Huosheng Hu, Libor Spacek . However it is difficult for human experts to examine complex systems, then it isn't easy to design an optimized fuzzy controller. Generally the performances of Fuzzy inference system (FIS) depend on the formulation of the rules, but also the numerical specification of all the linguistic terms used and an important number of choices is given a priori, also it is not always easy or possible to extract these data using human expert. These choices are carried with empirical methods, and then the design of the FIS can prove to be long and delicate vis-a-vis the important number of parameters to determine, and can lead then to a solution with poor performance. To cope with this difficulty, many researchers have been working to find learning algorithms for fuzzy system design. These automatic methods enable to extract information when the experts’ priori knowledge is not available. The most popular approach to design Fuzzy Logic Controller (FLC) may be a kind of supervised learning where the training data is available. However, in real applications, extraction of training data is not always easy and become impossible when the cost to obtain training data is expensive. For these problems, reinforcement learning is more suitable than supervised learning. In reinforcement learning, an agent receives from its environment a critic, called reinforcement, which can be thought of as a reward or a punishment. The objective then is to generate a policy maximizing on average the sum of the rewards in the course of time, starting from experiments (state, action, reward). O pe n A cc es s D at ab as e w w w .ite ch on lin e. co m