Скачать или смотреть Independent Study Results - Deep RL - PPO - Selfplay

This video is showcasing the progress I have made during my independent study. Each of the six agents in this video have learned to play this homemade game using stable-baselines3's implementation of PPO (a reinforcement learning method using a deep neural network). Each agent trained independently for 10,000,000 steps using selfplay (playing against their checkpointed copies).

The top left window shows agents "1_rs" (red) and "2_rs" (blue), each trained against their most recent checkpoint of themselves during selfplay. Top right show "3_rs_buf" (red) and "4_rs_buf" (blue), each playing against their 10 previous checkpoints. Finally, bottom left shows "5_rs_bigbuf" and "6_rs_bigbuf" which played against their 100 previous checkpoints. The chart shows their relative performance after repeated tournaments using the ELO rating system. "2_rs" is the top performer.

Checkout my GitHub page for the latest update:
https://github.com/urciuolim/TankTwin...

stable-baselines3:
https://stable-baselines3.readthedocs...

Unity (underlying game engine):
https://unity.com/