Project information
- Category: AI/ML
- Project date: 2022
- Project URL: github.com/jrtivald/EEC-289A-Final-Project
TrackMania RL Agent
My partner and I developed a Reinforcement Learning (RL) agent to race a car in TrackMania Nations Forever (TMNF). We used the double Q-learning algorithm as that has better interim performance along the asymptotic curve of total reward per episode. The agent's actions were accelerating, turning left, turning right, and braking. The feedback presented to the agent was a distance to the track walls along 16 radially spread lines and the car's speed. Both the wall distance and car speed were detected with image recognition algorithms. The reward function received a positive reward for higher speed and a negative reward for a longer time. Our agent was able to successfully learn a simple oval track but was significantly slower than a human player.