Reinforcement Learning for Autonomous Vehicle using MPC in Highway Situation

Path planning for Autonomous Vehicle(AV) is a challenging problem, as the vehicle is required to obey the traffic rules while avoiding the collision with the other vehicles. Model Predictive Control(MPC) is one of the popular approach for proposing a feasible and stable path by reflecting vehicle dynamics in solving objective function and constraining the expected future control input. However, one of the drawbacks with this approach is that the demanded computational power increases proportionally to the number of considered future inputs. This paper presents a path planning algorithm using Reinforcement Learning(RL). RL is similar to MPC in finding the optimal solution that maximizes the reward function which can be seen as intrinsic objective function. In that respect, adequate employment of MPC path in training resulted in improved efficiency and performance. Through the simulations, proposed method showed 98% of similarity with path of MPC and reduced computation time by 91.13% on average, thus it is qualified for real-time path planning.