privateboss/Car_Race_AI_V1

                                      #And the Devil said, "Let there be Air"
                                

#TheEnd #Open-Source and free to use dataset. Only ask is to cite if you are using code or dataset for research, publication etc.

This contains 30,000,000 training time-steps using Ubuntu Linux, Tensorflow 2.xx and PPO algorithm to train a single agent car racing utilizing CarRacing-v3 Gymnasium environment which is a maintained fork of OpenAI’s Gym library.

Trained 30000000 but rolled-back to 16558080 time-step datasets as the agent performance is stable till about 17000000 with perfect runs between 900-999 after 20 trials. With 30million timesteps it does Tokyo drifting practice. The reason for the drastic change is due to Policy Collapse, Catastrophic Forgetting.

This agent learns to uses the FISO (Fast In Slow Out) cornering strategy in racing very well