A Maximum Entropy Inverse Reinforcement Learning Algorithm for Automatic Parking

Abstract

Reinforcement learning is applied to automatic parking in order to eliminate errors caused by dead reckoning and path tracking during automatic parking. However, the reinforcement learning method has the problem of manually setting the reward function, and the learning network is difficult to converge and easily falls into local optimum. By maximum entropy inverse reinforcement learning, the reward function that maximizes the probability of expert trajectories is learned, and the reward function is applied to reinforcement learning to accelerate learning, improve the stability of learning, and learn driving strategies that are more consistent with human driving habits. The vehicle experimental results also show that this method learns faster and learns expert trajectories better than traditional reinforcement learning.