So I’m doing some experiments with curriculum learning using an environment similar to the SoccerTwos example. I have already trained up a model with self-play that I now want to use to train another model. The idea is to train up a new striker from scratch, but with the opponents and the agent’s teammates using the already trained model during training. The other players won’t have their models retrained, just the new guy.
So far I’ve figured out how to set the other agents to only use their old model (Set behaviour type to Inference only), but I still need to keep their behaviour in the trainer_config.yaml file to keep the training from crashing. I suspect that this means that I’m still training these models in addition to my new striker, which seems like a huge waste of computational resources. Does anyone know how to “turn off” training for a certain behaviour name?