Various traffic situations on multi-lane highways pose challenges for autonomous driving, requiring adherence to traffic rules. Traditional rule-based decision-making struggles with safety in complex environments, leading to research on deep reinforcement learning (DRL). This paper proposes a decision-making method based on proximal policy optimization (PPO) with hybrid actions. The DRL model inputs the states of the ego vehicle and surrounding vehicles, outputting continuous longitudinal control and discrete lateral lane changes. For lane changes, actuator-level steering is controlled via pure pursuit. Experiments show that agents with hybrid actions are safer than those using only continuous or discrete actions.
