Research On Reinforcement Learning Problem Based On Artificial Potential Field

Reinforcement learning (RL) has attracted most researchers in the area of robotics, because of its strong on-line adaptability and self-learning ability for complex system. But with the development of robot, more challenges come up, such as environment perception, generalization of RL, reactive control in local environment, large scale and partially observable environments, etc. It is difficult for common algorithms to obtain a satisfied solution. In this thesis, we try to research the mobile robot navigation in the large scale and partially observable environments, with the artificial potential field (APF) and reinforcement learning.Firstly, the introduction reviews the research on RL, its relative aspects in the world, the background and practical significance. Then temporal difference learning and Q-learning, two relatively mature kinds of algorithms are analyzed.Secondly, the repulsion force function and the gravitation force function of the potential field are introduced. And the excellent and shortcoming of the artificial potential field method have analyzed at the same time. A reinforcement learning problem is transferred to a path planning problem by using artificial potential field is the main contents of this thesis. That is, an efficient reinforcement potential field model (RPFM) is presented, with a virtual water-flow concept.Finally, the performance of RPFM is tested by the three well-known gridworld problems, and also the experiment with HQ and Q-learning for comparison had been done. Experimental results show that the RPFM is simple and effective to give an optimal solution for observable and partially observable RL. Compared with HQ and Q learning, our model is more stable and effective.