Risk-Aware Reward Shaping of Reinforcement Learning Agents for Autonomous Driving

Author: Lin-Chi Wu (l.wu@student.tue.nl)

This git repository contains the programs for risk-aware reinforcement learning for autonomous driving. It is a supplementary for the following publication,

Wu, Lin-Chi, Zengjie Zhang, Sofie Haesaert, Zhiqiang Ma, and Zhiyong Sun. "Risk-aware reward shaping of reinforcement learning agents for autonomous driving." In IECON 2023-49th Annual Conference of the IEEE Industrial Electronics Society, pp. 1-6. IEEE, 2023.

Environment and setup

• The running program is running with a Python environment. The suggested Python installed base: Python 3.8.15.
• For the nicest experience and observation, one can use the Jupiter notebook.
• Conda environments and the basic package version are presented alongside the repository.

Training file

The training files for 3 algorithms are displayed in Train_DDPG.ipynb, Train_DQN.ipynb and Train_PPO.ipynb.

Demonstrate (test the trained agent) file

One can observe the trained agent's (after training with 1000 episodes) behavior with LoadModelNTest.ipynb

The out-of-track (O2T) counts and the on-edge (OE) counts.

According to the review, there is still a lack of evaluation with a fair standard for different agents' experiences in the same environment. The new evaluation method is proposed to evaluate the agent's performance in the same environment and they are evaluated not just by the reward but also by agents' behaviors counteract to the environment, i. e. the strategy.

We think that the agent has a good strategy of being inside the track and can survive longer during the playing round by showing that the game is less likely to be terminated due to out of the track. In addition, the agent can drive smoothly so that it is less frequently turning angles while remaining on the track. We propose the evaluation method by counting the number of out of track and the number of touching the track's edge. We consider an agent cannot drive stable when it consistently turns driving directions avoiding out of the track near the edge. The lower the number of out-of-track and driving stable, the better the strategy of the agent.

Two parameters to evaluate the agent's strategy

We want these two factors as low as possible.

• Out of the track.
• The instance of the agent approaching the edge zone.

The edge zone area the place is between the green area and the yellow dash line.