A Study Of Multi-robot Information Fusion And Coordination Based On Agent

With the rapid progress of modern science and technology, robots' development and application extend ceaselessly. For different tasks and environment, especially to large complex tasks and environment, the problems of capability limitation of single robot become more and more obvious, such as information gathering and processing, controlling and so on. Therefore, the multi-robot coordination and cooperation are required to improve these abilities.Nowadays, the multi-robot system has been paid much attention. It has the prominent property of multidisciplinary intercrossing and fusion. Based on multi-agent system, and with the introduction of multi-agent system's architecture, coordination and cooperation, this dissertation gives a thorough and systematic research on the information fusion of multi-robot system, the tasks distribution and programming of multi-robot coordination, and the multi-robot under oppositional environment. The main contributions are as follows:1. On the basis of analysis and research of multi-agent architecture, an information fusion model based on multi-agent is proposed, in which, each agent is responsible to observe the environment and provide the belief of local environment information, and the fusion center will combine all the believes to provide a final decision. In multi-agent coordination methods, the multi-agent learning and Markov games are researched and analyzed.2. The distributed decision-making based on multi-agent is studied comprehensively. An agent information model based on evidential reasoning and a multi-agent decision frame based on the transferable model are put forward and the corresponding algorithm is also provided. With its application to robot soccer, the game information and opponent team situation are analyzed and the final decision is effective.3. A new reinforcement function based on knowledge is put forward. In the physical system, the reinforcement learning algorithm is improved, in which, with the introduction of experience information and domain knowledge, thereinforcement function combines the reward of global goal and the agent's action strategy. With the application and experiment in robot soccer, the results show that the learning with the reinforcement function based on knowledge has better performance than traditional reinforcement function.4. A multi-agent coordination model and corresponding algorithm based on distributed reinforcement learning are proposed. We analyze and research the distributed reinforcement learning in multi-agent environment and adopt a multi-level model to resolve the multi-agent coordination: the system task is decomposed and distributed in coordination level and the sub-tasks are accomplished in task level. The experiment results in robot soccer show that this method is better than the conventional reinforcement learning.5. A multi-agent coordination frame and corresponding algorithm based on Markov games are proposed. We analyze the Markov game and its properties in multi-agent environment and pay much attention to the multi-agent learning algorithm based on adversarial equilibrium and coordination equilibrium in multi-agent system. By means of the competition and cooperation of multi-agent, a layered frame is applied to multi-agent coordination: the zero-sum Markov games are adopted to resolve the agent teams' competition; the team Markov games are adopted to accomplish member agents' coordination and cooperation. The application and experiments in robot soccer show that this method is of satisfactory performance.