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

摘要

In this chapter, we used the algorithm of reinforcement learning, Fuzzy Q-Learning (FQL) L. Jouffe, A. Souissi which allows the adaptation of apprentices of the type FIS (continuous states and actions), fuzzy Q-learning is applied to select the consequent action values of a fuzzy inference system. For these methods, the consequent value is selected from a predefined values set which is kept unchanged during learning, and if an improper value set is assigned, then the algorithm may fail. Also, the approach suggested called Fuzzy-Q-Learning Genetic Algorithm (FQLGA), is a hybrid method of Reinforcement Genetic combining FQL and genetic algorithms for on line optimization of the parametric characteristics of a FIS. In FQLGA we will tune free parameters (precondition and consequent part) by genetic algorithms (GAs) which is able to explore the space of solutions effectively. However, many times the priory knowledge about the FIS structure is not available, as a solution, the suggested approach called Dynamic Fuzzy Q-Learning Genetic Algorithm (DFQLGA), which is a hybrid learning method, this method combines the Dynamic Fuzzy Q-Learning algorithm (DFQL) Meng Joo Er, Chang Deng and the genetic algorithms to optimize the structural and parametric characteristics of the FIS, with out any priori knowledge, the interest of the latter (GA) is to explore the space of solutions effectively and permits the optimization of conclusions starting from a random initialization of the parameters. This chapter is organized as follows. In Section II, overviews of Reinforcement learning, implementation and the limits of the Fuzzy-Q-Learning algorithm is described. The implementation and the limits of the Fuzzy-Q-Learning algorithm are introduced in Section III. Section IV describes the combination of Reinforcement Learning (RL) and genetic algorithm (GA) and the architecture of the proposed algorithm called Fuzzy-Q-Learning Genetic Algorithm (FQLGA). In section V we present the DFQL algorithm, followed by the DFQLGA algorithm in section VI. Section VII shows simulation and experimentation results of the proposed algorithms, on line learning of two elementary behaviours of mobile robot reactive navigation, "Go to Goal" and "Obstacles Avoidance" is presented with discussion. Finally, conclusions and prospects are drawn in Section VIII.