Research And Application On Reinforcement Learning And Communication Technology In Agent

Research on the agent and multi-agent system is an important branch in the artificial intelligence and computer science, where communication and learning are two important technologies that Agent should possess. Agent must have learning ability to deal with dynamic and complex environments, and have an effective communication method to interact and cooperate with other agents. Reinforcement learning is the main kind of learning method, which is recognized as an ideal technology to construct the intelligent agent. Communication protocol is the basis for agents to exchange information and knowledge effectively.In this dissertation, the research was focused on some key problems about reinforcement learning and communication protocol in agent. Some improvements and useful experiments were made on the basis of early research. The main work is summarized as follows:(1) Based on good understanding of the reinforcement learning theory and Q-learning algorithm, a knowledge-based Q-learning (KBQL) algorithm was proposed. In order to accelerate the convergence of learning algorithm, the learning state space is contracted by using knowledge in agent. Using learning mechanism in Agent to correct some inaccurateness of knowledge, the applicability and robustness of the algorithm is boosted. After making simulation experiments on a complex Grid-World problem, we came to a conclusion based on the experiments result as follow: As compared standard Q-learning, the KBQL has predominance in convergence speed despite giving some imprecise knowledge.(2) An improved Sarsa(λ) algorithm was advanced after making a thorough analysis on trace eligibility, with time complexity 0(|A|) per update. A Heuristic Reward Function was introduced in the improved Sarsa(λ) algorithm. The Heuristic reward function would not influence the optimal policy any more, so the learning efficiency and convergence speed were improved by using knowledge to lead agent searching in the expected state space.(3) We had done some research on keepaway soccer that was built on RoboCup simulated soccer, which was regarded as a benchmark for machine learning. The reinforcement learning model with prior knowledge was proposed for the keeper's learning problems according to some common sense on soccer. The...