Robot Behaviour Learning Based On Reinforcement Learning

With the continuous development of robot technology, robot navigation technology application areas are expanding.Therefore, research on mobile robot navigation technology into the foreign academic of the important issues topic. Because the mobile robot in which the environment is unknown, so there is a strong uncertainty in robot navigation. This thesis mainly aimed at researching robot navigation and sensor design technology, the purpose is to enable the moblie robot can quickly and effectively to complete the task.Deeply studied the type-2 fuzzy logic system (Type-2 FLS) and apply its to the specific robot navigation, using Type-2 FLS for robot obstacle-avoiding and approaching the target behavior model. Comparing the simulation results show that, compared with traditional of FLS, Type-2 FLS has strong fuzziness and can be through expert knowledge which effectively solve the robot navigation in environmental uncertainty and navigation the dynamic problem. But the Type-2 FLS of fuzzy rules defined by the expert knowledge, so there are still has some limitations.Based on the mechanism analysis of Type-2 FLS, this paper presents a method of combine with the Type-2 FLS and reinforcement learning, the method of which using Type-2 FLS to solve reinforcement learning in the state space to the action space mapping problem. Use of reinforcement learning approach to the fuzzy rules, the state space as input and output obtained is action space, through the way of learning to build a more perfect rules and systems. This method not only improves the convergence of reinforcement learning speed and ability to adapt the environment, while improving the robot's performance and learning efficiency.Finally, the proposed method is applied to specific mobile robot navigation. Through the establishment of simulation platform, the experiment proved that this method can use the Type-2 FLS membership describe of fuzzy information, and then relatively few fuzzy rules can be achieved reinforcement learning process, so the robot can quickly and efficiently to complete the navigation task under unknown.