Support for mixed discrete-continuous actions

[CloudyDory]

Hi, I have a custom environment that needs to input both discrete and continuous actions. The continuous action should be a Gaussian policy with learned mean and standard deviation. I hope to train a PPO agent to solve the task. Is it still possible to use Tianshou's default Collector and OnPolicyTrainer, or should I write my custom Collector and OnPolicyTrainer?

[Trinkle23897]

imo you only need to inherit PPOPolicy and rewrite parts of them (forward, learn) to let it support combined action input. No need to rewrite Collector and Trainer

[CloudyDory]

Hi, I guess the collector should also be rewritten as well?

[Trinkle23897]

I don't think so. You can pack your action as a dict (similar to dict action space) each time, which will be converted to Batch automatically in Collector.