state type

[FeiZhou02]

please,Can PFRL handle dictionary-type observations returned by a Gym environment?

[prabhatnagarajan]

I think it should be able to. You would need to preprocess your dictionary to be passed into a neural network, though. Do you have an example?

[FeiZhou02]

def _to_recursive(batched: Any, device: torch.device) -> Any: if isinstance(batched, torch.Tensor): return batched.to(device) elif isinstance(batched, list): return [x.to(device) for x in batched] elif isinstance(batched, tuple): return tuple(x.to(device) for x in batched) else: raise TypeError("Unsupported type of data")

I found that in the batch_states.py file, it cannot handle dictionary-type data.

[prabhatnagarajan]

If you look here:

pfrl/pfrl/agents/dqn.py

Line 170 in c8cb332

batch_states (callable): method which makes a batch of observations.

, you can see that a batch_states function can be passed into the agent. You can write your own function (not using PFRL's default) and pass it into the agent for usage.

[FeiZhou02]

Thank you for your reply.I’ll give it a try.