Research On Cooperation And Coordination Of Multi-robot System Based On Reinforcement Learning And Swarm Intelligence Method

In the 21^st Century , the distribution, intelligence and the coordination requests are brought forward for the multirobot system .Reasonable architecture and effective cooperation algorithm are paid more and more attention. These two fields are studied in three parts in the dissertation, which are the study of multirobot architecture , the Reinfocement Learning algorithm of multirobot system and the Swarm Intelligence algorithm of multirobot system. All these researches can satisfy the low communication, variety, distribution and decentralization needs of the system.The architecture which determines the relationship of the robots and the task assigned is the basis of the multirobot study. Facing the Reinforcement Learning algorithm, the level architecture of the multirobot system is brought forward .At the same time, the potential grid method, action fuzzy control and the blackboard communication are studied. This architecture has the concurrent , real-time and flexible ability. Pointing to the Swarm Intelligent algorithm, the intention-behavior architecture is brought forward. And the group structure, robot ability and the communication are studied. The cooperation based Markov Game without communication is discussed. At one time, the intention competing method and the behavior restrained-exhausted method are investigated. Moreover, the robot behavior assigned mode and the communication based on pheromone spreading method are researched. This architecture has the advantages of distributed control and dispersible data. These two architectures can be widely used in the similar system.The Reinforcement Learning theory is attached importance for itsself-learning and the self-adaption. But the problems of state compressed, structure credit assigned and the task partition still prevent the theory extending .In the dissertation the self organization method is brought forward to solve the state compressed problem and the self and the group signal method is put forward to solved the structure credit assigned problem. Compressing state speeds up the ergodic so that the learning speed can be increased. Assigning the credit maps the state to behavior reasonably so that the system's effect swing can be avoided. After the algorithm ameliorated , the self-adaption and the robust ability of the system increased rapidly.The swarm intelligent theory offers the thought that the soul of the intelligence exists more in the group than in the single part. The intelligence of the system can also be increased even though the intelligence of the single part is low. But the thought still exists the exploring fields such as lacking the transplant ability and the practical field. The group behaviors algorithms are designed based on the swarm intelligence in the dissertation. The simple interaction rules are set down to realize the complex task. The pheromone spreading method is constituted to realize the communication, which can decrease the information flowing and strengthen the stability of the system. The stability of the system is discussed and the conclusion is when the functions of the robot are chosen reasonably , the stability of the team formation can be guaranteed.At last, the compare of these two algorithms is presented and the capabilities are also analyzed. These conclusions gives the guidance of corresponding algorithms' application in the different environment.Facing the multirobots' team formation task, the application models of the two algorithms are discussed . In the simulation of the experiment , the feasibility of these technologies is verified further. The expands ofthe methods are strong and can be used in the similar system.