Lane Change Decision Planning for Vehicles Based on Reinforcement Learning

To alleviate urban traffic congestion, this project aims to utilize Deep Reinforcement Learning (DRL) techniques to guide autonomous agents (vehicles) in making optimal lane-changing decisions within complex traffic environments. This project is built on the Highway-env simulation platform and implemented using the PyTorch framework. Four mainstream deep reinforcement learning algorithms, including DQN, PPO, SAC, and DDPG, are implemented and systematically compared. Based on this, supervised learning is introduced to optimize PPO’s lane-changing decision strategy. Experimental results show that integrating supervised learning improves the cumulative reward of PPO while reducing the collision rate. Overall, the hybrid supervised learning and PPO approach demonstrates better policy stability, particularly in complex traffic scenarios with highly dynamic conditions.