Why do agents trained by the reinforcement learning PPO algorithm get different results each time they load？

In the process of reinforcement learning, a problem will be encountered. During the training process, an effective agent will appear. At this time, the training will be finished in advance, but the result of the saved agent running out will be worse

Answer 1

[Shivansh]

Hello Ye,

The behaviour shown by the agents trained by the reinforcement learning PPO algorithm suggests there might be some randomness involved in the environment. Since, the performance worsens for saved agents, it is also possible that the agent might be overfitted on certain episodes while training and the performance might not be same for other episodes. Try to remove any stochasticity in the model using deterministic seeds. You can also try to evaluate the model on various episodes to get a better understanding of the issue.

If the issue still remains unclear, share more information regarding the model and the environment.

I hope it helps!