My agent finds a good solution but it does not repeat it and turns back to bad behavior again (Reinforcement Learning )

My agent finds a good solution but it does not repeat it and turns back to bad behavior again. I dont know why is that the case! it happens a lot and everytime even after finishig the training it does not learn. Can anyone tell me why is that?

Attached you can see an example. it happend very early and it was almost why expected for but it didnt follow it anymore.

Answer 1

[Emmanouil Tzorakoleftherakis]

This picture shows progress for a very small number of episodes. You should wait for a few hundred episodes before you can evaluate the agent. Even then, there is no guarantee that episode reward will increase monotinically. Please see this answer. If after training for some time, the agent is still not learning anything, you need to go back and evaluate the problem formulation. You may need to increase exploration, adjust neural net architectures, pick different sample time, tune agent hyperparameters, pick a different reward and so on. These are a few suggestions, but as you can see this is unfortunately a trial and error process

Hope this helps