Research Of Multi-Agent System's Coordination Machanism Based On Planning Amalgamation

In the Multi-Agent System, coordination is an extremely important factor for multi-Agents'intelligence. By the way of sharing information, communicating with each other, and making reasonable arrangement for agent's tasks, the goal of coordination is to obtain the overall best performance. Among those methods to resolve Multi-Agent System coordination problem, Markov Decision Process is a key effective theory model of intelligence planning, which based on the following thought:though interactive with each other and the surrounding environmentm, agents can continue to comprehend the state of environment and rectification their own actions.In this paper, based on the module of Markov Decision Process and the module of partial observable multi-agent Markov Decision Process, we study two types of coordination problems:task coordination and action coordination. Here task coordination especially refers to dynamic task allocation problem. So our research includes the following aspects.First of all, to the centralized dynamic task allocation problem, we construct an Markov Decision Process module to meet the need of maximized-profits, the uncertainty of allocated tasks, and the feature of sequential decision-making. In the module, we give abstract module's essentials the actual meaning associated with task allocation problem, and prove the existence of optimal allocation policy. The computation of policy synthesizes two existed algorithm:the value iteration and the policy iteration. The result of experiment indicates this module can reduce the number of iteration and improve converge speed, which also on the basis of getting the optimal or sub-optimal policy.Secondly, we analyze reallocation problem by classifying the reason of real location tasks, then we take some adjustments for initial allocation strategy aiming to prove that all the tasks which not be execute success can arrange to new task queue and be successly completed finally. The module's robustness is enhanced.Besides, to behavior coordination between agents'in Multi-Agent System, a behavior coordination mechanism based on planning amalgamating is promoted by considering these situations:partial observable feature of environment, an amount of history information, limited communication resource. In this mechanism, history scale is limited by equivalence definition and history merging theorem. Conflict detective and communication delay are be used to get maximized-profits. The simulation result indicates that the mechanism increases the profit using history information to system decision, and gets the better effectiveness and performance in the condition of limited communication resource.