tuxracer / cars-rl

Reinforcement Learning (DQN) from scratch with zero dependencies applied to a simple car driving game

cars-rl

Reinforcement Leanring (DQN) from scratch with zero dependencies applied to a simple car driving game. Check it out here!

demo.mp4

Details

The agent has 3 inputs which define the vicinity of walls to the sides of the car, each in the range of [0, 1]. At each step the reward is:

• 1 if the car reaches a new point on the map (the map is divided into a grid of 10x10 cells)
• 0.01 if the car does not crash
• -10 if the car crashes

References

This project is heavily based on:

• cars-ai
• tsgrad

License

MIT