Hi,
I’m working on a project that requires agents acting like below.
Step1: Agent A runs a couple of episodes and stops making decisions.
Step2: Agent B takes an action and stops making decisions.
Step3: Agent A uses inference with its network to run an episode.
Step4: Go to Step2.
…
Agent A and Agent B have totally different action spaces and observation spaces.
My question is how to make agent A pause training and use inference while the training session is running.
You can most likely do this through the UnityEnvironment API within python + ML-Agents side channel control. There are some pages in the docs about this sort of thing; though I have not gotten to this point myself. I don’t think you’d be able to easily do this within unity/c#, but I may be wrong.
Thank you. I appreciate it.
I will look into it.
Depends on what you mean by “stops making decisions”, if I’m understanding correctly you’re trying to train two agents at the same time, right? If that’s the case, look at the Striker vs Goalie sample (self-play), I think that might be what you’re looking for.