Decision-making Research On Safe Lane Change Of Intelligent Vehicle Based On Deep Reinforcement Learning

" Merging in the lane " is called " changing lanes to the left or right ".When there are two or more motor vehicle lanes in the same direction of the road,the driving road can be changed without affecting the normal driving of the motor vehicle traveling in adjacent lanes.As a research focus in the field of vehicle engineering in the world today,the research on the improvement of traffic safety level and the decision-making behavior research is of great significance for relieving traffic congestion,improving road traffic capacity and improving the green ecological driving environment.The traditional rule-based driving decision algorithm has strong dependence on the intelligent car and the environment model.The space for lane-changing differs in the complex and dynamic traffic environment.The dynamic and uncertainty information in the lane-changing process is also poseing a huge challenge for autonomous driving decisions.In order to ensure the security of intelligent car’s driving decisions,this paper studies and designs two autonomous driving decision models.First,the hybrid driving decision planning model of hierarchical finite-state machine is designed.Based on some certain priori rules,different driving states and sub-behaviors can be changed by hierarchical finite-state machine and the complex decision logic can be clearly organized.The trajectory planning module adopts the local path planning method of receding horizon control(RHC).The algorithm integrates the results of behavior planning into the surrounding environment perceived by sensor.For different condition constraints,the method of high-order polynomial fitting is adopted,and the principles of safety and comfort are considered,real-time planning.The optimal path is obtained,and finally the effectiveness of the algorithm is verified by simulation of continuous lane-changing.Then,the action-reactive decision model of deep reinforcement learning DQN is designed.Firstly,the traditional memory replay method is improved,and priority sampling is introduced,which solve the problem of sparse sample return during the initial training period and improves the utilization of samples with large TD error.In the road conditions of multi-lane free lane-changing and ramp lane-changing,combined with mission requirements,the state set,action set and reward set in the MDP process of safe driving behavior are designed respectively.Finally,the relevantsimulation proves that deep reinforcement learning is intelligent.The effectiveness of the car safety and lane-changing problem.Compared with the hybrid decision-making planning model,deep reinforcement learning is modeled by the interaction and reward and punishment of the environment and the agent,and is fitted through the neural network,which not only conforms to the original decision-making thinking,but also fully considers the uncertainty of the environment and finally completes certain strategic of gaining more and avoiding harm.Deep reinforcement learning is the key to general artificial intelligence,and research that advances deep reinforcement learning also helps autonomous driving decision technology take a big step forward.