Game Theory and Reinforcement Learning based Smart Lane Change Strategies 2022-01-0221
With the development of science and technology, breakthroughs have been made in the fields of intelligent algorithms, environmental perception, chip embedding, scene analysis, and multi-information fusion, which together prompted the wide attention of society, manufacturers and owners of autonomous vehicles. As one of the key issues in the research of autonomous vehicles, the research of vehicle lane change algorithm is of great significance to the safety of vehicle driving. This paper focuses on the conflict of interest between the lane-changing vehicle and the target lane vehicle in the fully autonomous driving environment, and proposes the method of coupling kinematics and game theory and reinforcement learning based optimization, so that when the vehicle is in the process of lane changing game, the lane-changing vehicle and the target lane vehicle can make decisions that are beneficial to the balance of interests of both sides. Firstly the conditions for judging whether the lane-changing vehicle and the target lane vehicle are in the lane-changing game are provided. According to the actual vehicle driving situation, the type of payoff in the game between the two vehicles is then determined. The payoff function is designed by using the kinematic method, and the corresponding total payoff value under the different strategy combinations of the two vehicles is obtained. Then, the game payoff matrix is analyzed, and the optimal strategy combination and the corresponding acceleration are obtained. Finally, the optimal strategy combination of the two vehicles is determined, which can effectively avoid the safety problem, and the overall payoff of the two sides is relatively balanced. The results of experience indicate that two vehicles will choose the most beneficial strategy combination and then make their own decisions according to the proposed model, at the same time, this paper also respectively analyzes the payoffs of the vehicles in the velocity and the relative position of different cases to verify the feasibility and rationality of the model.