Research On Multi-Agent Path Planning Based On Deep Reinforcement Learning

With the rapid development of artificial intelligence,reinforcement learning has made great progress and achieved great success in solving various sequential decision problems in machine learning.With the deepening of research,multi-agent reinforcement learning is also full of vitality in the field of reinforcement learning,and has been applied to many fields.The path planning problem has important applications in chip,traffic,logistics and other scenarios.In essence,it is a sequential decision problem,which is very suitable for combining with reinforcement learning.In this paper,from the perspective of global to local and rough to detailed,different multi-agent path planning problems are studied in three scenarios based on deep reinforcement learning.The detailed routing of VLSI is a very important part in the integrated circuit design,which is essentially a global path planning problem.This paper proposes a detailed routing framework,which transforms the traditional detailed routing problem into a multi-agent path planning task based on reinforcement learning,and realizes the asynchronous routing.At the same time,by limiting the observation area of each agent,the size of features is reduced,the difficulty of training is reduced,and with the help of global information,the target vertex of the agent is predicted,and the routing conflict is reducedVehicle platoon is an important traffic operation to reduce vehicle energy consumption and improve air quality.How to control the formation of vehicle platoon from sparse vehicles on the highway is a very meaningful research point.This paper proposes a control planning framework for the formation of vehicle platoon based on federated learning and multi-agent reinforcement learning.We can accurately evaluate the energy consumption during the formation of vehicle platoon by modeling the vehicle energy consumption model.Furthermore,we introduce reinforcement learning algorithm into the formation process of vehicle platoon,and make asynchronous decision for multiple vehicles to save the time of vehicle formation.Federated learning is introduced into training,which greatly reduced the communication traffic of the whole system.Vehicle lane change is a frequent vehicle action at the highway ramp exit,and it is very easy to affect the traffic flow.In order to improve the speed of lane change and reduce the impact on traffic,we use multi-agent reinforcement learning algorithm to control vehicles to make full use of the collaborative ability of vehicles.At the same time,by delimiting the cooperative lane change area,the global information is generated according to the lane change area to cope with the change of the number of agents in the training process.Finally,according to the situation of lane change,the actions of vehicles when changing lanes are studied in detail,and the control of vehicles is more specific.