About the input of policy

[Liuxueyi]

In the paper, there is a state that

leverage the output state policy in nonmemory environments and the full state policy or hidden state policy within memory environments.

But in the configs.yaml, there is only in the POPGym task where the actor.inputs is set to be [stoch, hidden]. Why? Looking forward to your reply.

[artemZholus]

Hi @Liuxueyi .

the sentence you quoted is a general recommendation for working with new environments. There is another part of the paper where we state that in popgym we use the hidden state policy, so this config reflects that.

[Liuxueyi]

Thanks for your reply!
By the way, for the atari100k environment and pong task, what should be the input of policy(the default is [deter, stoch, hidden])? And in the issue #8 , when set the parameter run.script: train, there isn't any error. Is this feasible and will it affect the final performance of the algorithm?