| Reinforcement learning is an important machine learning method. Reinforcementlearning could learn the optimal policy of the dynamic system through environmentstate observation and improve its behavior through trial and error with the environment.Multi-agent reinforcement learning is the improvement of traditional reinforcementlearning. It obtains the effect of parallel processing, reduces learning time and acceleratesthe speed of seeking for the optimal strategy as well by multiple Agents workingtogether.The main research achievement is to design a multi-agent reinforcement learninghierarchical model, including the task layer, the work layer, the communication layer andthe decision layer. On this basis, the role and responsibilities of each layer and thedetailed implementations of the layers are illustrated. Further, it proposed implements ofthis in multi-core environment with multi-core technology and the evaluation standardsof performance of this model.Based on this model; it also proposed task-distribution andsubtask-assignment multi-agent Q-learning methods. The two methods are same inmodel but different in specific implementations. The former implements the decisionlayer based on information fusion and communication layer based on wait-lock mode;the latter implements the decision layer based on information arbitration andcommunication layer based on lock-free mode. They are used to solve the robot pathplanning and traffic signal control of multi-intersections in multi-core environment. Theresults of robot path planning simulation with multi-agent Q-learning method based ontask-distribution demonstrate that learning speeds, convergence cycle decreases, and thefulll use of computing resources, compared wiht single-agent Q-learning method.Thisproves effectiveness of this method. The results of traffic signal control ofmulti-intersections simulation with multi-agent Q-learning method based onsubtask-assignment demonstrate that the average waiting time and numbers of queuingvehicles reduce, and it maintains good traffic, compared wiht non-arbitration Q-learningcontrol and Timing Control.This proves effectiveness of this method. |
