Design of Reward Functions for Autonomous Driving Based on Reinforcement Learning: Balancing Safety and Efficiency

Author:
Open Access
Publisher website
Google Scholar
Abstract
Autonomous driving, leveraging artificial intelligence and reinforcement learning (RL), has made significant strides in improving traffic efficiency and safety. However, current RL-based approaches often focus on single-objective optimization, such as maximizing either efficiency or safety. In real-world driving, multiple conflicting objectives—such as safety, efficiency, and comfort—must be balanced simultaneously, which remains underexplored. This paper proposes a multi-objective reward function design to balance safety and efficiency in autonomous driving. Using the Proximal Policy Optimization (PPO) algorithm, we train seven autonomous driving models with varying collision penalty strategies in the MetaDrive simulation environment. The results show that dynamic collision penalties outperform fixed penalties in balancing safety and efficiency, with Model 5 achieving the best overall performance. Despite this, all models underperform in left-turn scenarios, highlighting the need for further optimization in lateral control. This work provides insights into effective reward design for multi-objective reinforcement learning in autonomous driving.

This publication has 6 references indexed in Scilit: