Closed vaidas-sl closed 8 months ago
@vaidas-sl I fixed the noise added to the action at training time, as defined in the paper. I also removed the change to the log_all_multi_policy_metrics, as it is supposed to be a method for evaluating Pareto fronts, and I believe that the metric you want should be logged inside PCN's code only.
Original paper suggests that PCN could be used in continuous action space. This paper uses PCN in continuous action space. This PR changes PCN implementation so it would be usable in cont actions space. Also updated constructor to accept model_class, to make it easier to pass different models, as original paper uses different models for different tasks