Training from checkpoints / executing trained policies

[kamivao]

Hello,

Thank you for the great library!

Is it possible to continue training from checkpoints or executing a trained policy in BenchMARL?

A bit of context: I'm experimenting with policy transfer/reuse across multi-agent teams and would like to test the scenario, when a MARL team kicks-off with a model learned for a related task in the same environment, and adapts it for the target task (i.e. in MPE, reuses the model learned for simple_spread to solve simple_adversary).

Thanks in advance!

[Xmaster6y]

Hey, don't know if you've found a solution, but for the record, it is easily achievable as is (should the task be similar wrt inputs/outputs). Basically, you could train on task1, then reload the experiment from the desired checkpoint, set the task as task2 and run experiment.evaluate().