The Analysis And Research Of Exploration Strategies And Algorithms In Reinforcement Learning

Reinforcement learning is a kind of important Robot Learning. Unlike most of machine learning, Reinforcement Learning is not order agent to know how to pick up the correct action, but through Trial-and-Error to find out what actions will produce maximum reward. Such imprecise information feedback form, improve the control requirements as lacking of information environment. Reinforcement Learning is not only in artificial intelligence, also in certain areas, such as motion control, mobile robot path planning with fairly strong applicability.In an unknown environment, Reinforcement Learning basic idea is to mimic the natural human way of learning. The goal oriented, in the process of interaction with the environment, through the heuristic learning on convergence to the optimal strategy. Reinforcement Learning is also the process of interaction with the environment necessary knowledge completeness, and action selection mechanism both need to exploit learned knowledge and explore of the unknown knowledge. Exploitation and Exploration both have advantages and disadvantages. The balance between them affected agent, which is vital to understand the environment and the optimization of algorithm performance.After clarifying the basic theories of Reinforcement Learning, there is a detailed analysis of the Reinforcement Learning classic exploration strategies including: Random Walk,ε-Greedy Strategy, Softmax Method and Probabilistic Action Selection Mechanism. Reinforcement Learning is aimed at reducing the uncertainty of the selection operation. Due to the lack of tools analyzing search strategies, introducing of Exploration Entropy concept to quantitatively describe and analyze the exploration strategy. And based on Exploration Entropy, the conditions deciding to terminate the algorithm are presented.In the experimental part, this paper verified the applicability of the Exploration Entropy in both macro and micro aspects. We adopted different search strategies to simulate a Markovian state transition problem and a more complex Learning Control problem of micro 1/2 spin quantum systems. Further proved, the Exploration Entropy as a tool can provide a reference index for analyzing the algorithm performance.