Airplane learning to fly through a race track using Deep Reinforcement Learning and Unity(Using MLAgents 0.11.0). Each time the agent goes through the checkpoint prefab it is rewarded with +1, and each time it hits a rock or the map it gets punished with -1.
Built with C# and Python.
Dependencies: MLAgents, TensorFlow, InputSystem
Every video is done with the same neural network brain. I trained them with 4 different racetracks so they can generalize their knowledge (and prevent overfitting) to tracks they have never seen before.
Training the agent in a loop:
Training the agent in a Loop with sparse rewards:
Training the agent in a 8 loop:
Training the agent in a difficult course with obstacles:
Made By: Samuel Boylan-Sajous