Whether state dependent action space will be supported?

[evanspuck]

I'm recently using your project for RL research. However, in my problem, the action space depends on the current state. I wonder whether this feature will be supported or not? Thank you very much.

[YangletLiu]

Please elaborate more? Do you mean "allowed actions at different states"? Any RL paper you want to reproduce?

[evanspuck]

Please elaborate more? Do you mean "allowed actions at different states"? Any RL paper you want to reproduce?

For example, if the action space is {1,2,3,4}, but on state 1, only action {2,3,4} is allowed.

I know I can put a negative reward on the invalid actions, but the convergency will be slow.

Another way commonly used is by "masking" invalid actions in different states. I've searched your project and there does exist the keyword "mask". I wonder if there is any way that I can modify your code to conduct this masking procedure for invalid actions? Thanks a lot!