Cooperative Overtaking Strategy Based On Multi-agent Reinforcement Learning

Recently,with the development of sensing,information processing and computation technology,the Intelligent Transportation System has also developed rapidly which also brought huge economic benefits to the society.From the perspective of drivers,the emergence of agent cars enables people to travel by themselves without being constrained by cars,and can reduce traffic accidents caused by human factors,such as drunk driving and fatigue driving.Can effectively improve the safety of the journey,improve the comfort of passengers.From the perspective of the whole intelligent transportation system,it will help solve the two major problems of safety and traffic congestion,and greatly improve the current low traffic efficiency.Overtaking is a common driving behavior,but also one of the most likely to cause safety accidents and traffic jams.This is because the process of overtaking is tedious and the environment is complex and changeable,each process is faced with two serious potential accident factors from human factors and the changing road environment.Therefore,overtaking has become one of the difficulties of intelligent driving in many possible operation s in the process of driving.The emergency agents include ambulances,fire engines,police vehicles and other agents that perform urban emergency rescue services.When carrying out emergency rescue missions,it is necessary to ensure that time is raced ag ainst time,otherwise it is very likely that serious delays due to delays of just a few minutes will cause unavoidable losses.Ensuring that emergency agents evade routinely in the process of emergency vehicles surpassing conventional vehicles is the key t o the rapid deployment of urban emergency rescue forces.Therefore,it is extremely important for multiple agents to actively cooperate with rescue agents to overtake.This thesis mainly studies the whole process of autonomous overtaking on multiple highways in the same direction on a straight two-lane road,including lane changing,overtaking,and returning to the original lane.First,the classic Q learning algorithm is used to complete the single agent overtaking control in the non-cooperative scene.Because of its limitations in dealing with complex roads,a Q-learning algorithm based on multi-agent joint control and a Q-learning algorithm based on multi-agent independent control are proposed,and two cooperative modes are constructed in the same environment by constructing a model of two-car cooperative overtaking.Make an analysis and comparison.It provides an important basis for the complex and variable environment and the collaborative overtaking strategy among N agents.The independent cooperative Q learning algorithm is applied to the agent car cooperative overtaking in the emergency rescue scene,combined with the actual rescue vehicle overtaking scene and traffic rules to design a suitable state set,action set and from the perspective of rescue vehicles and conventional vehicles.Reward and punishment,let the rescue car smoothly overtake,the conventional vehicle actively avoids.The ultimate goal is to have the rescue vehicle successfully overtake and arrive at the destination in the shortest p ossible time.Finally,the collaborative reinforcement learning algorithm with shared experience is introduced to further improve the safety and efficiency of the entire overtaking process.