| Reinforcement learning is an unsupervised learning technology, by which the agent can find optimal behavior sequence and perform on-line learning. So reinforcement learning is recognized as an ideal technology to construct intelligent agent.Q learning algorithm is the most popular reinforcement learning algorithm, but the algorithm exist some problems. Firstly, Q learning algorithm can't be used in the learning problem of continuous state space and continuous action space. Secondly, searching and storing of Q table will be a difficult problem when state space is very large. Finally, the learning speed of Q learning algorithm is very slow.In this dissertation, we improve and extend Q learning algorithm by combining with fuzzy inference system. The improved algorithm can be used in agent's decision-making in complex environment. The main works are as follows.A fuzzy Q learning algorithm is proposed in this dissertation, which map continuous state spaces to continuous action spaces by fuzzy inference system and then learn a rule base. The rule base provides prior knowledge for agent's action selection and dynamic programming.For ball intercept problem in RoboCup, a multiple rewards fuzzy Q learning algorithm is proposed. The algorithm is applied to the learning problem of continuous state space and discrete action sequence. The short-term interests and long-term rewards of agent are well balanced because the rewards signals are considered from different views.A novel modular fuzzy Q learning algorithm with prior knowledge is proposed which is used to solve the MAS's learning problem. Prior knowledge is adopted to improve the performance in the initialization stage in this algorithm. Furthermore, we divide the complex problem into many sub problems in order to resolve the learning problem in complicated environment. We consider other agents' action in the environment for implementing the optimization of decision-making when an agent selects its action. |
PDF Full Text Request