search

Gerryflap / RL_continuous_action_spaces

An exploration of continuous action spaces in Reinforcement Learning
8 stars 1 forks source link

Implementation in OpenAI Gym #1

Open sunnyshi0310 opened 2 years ago

sunnyshi0310 commented 2 years ago

Hi,

Thanks for sharing the algorithms. I am trying to implement the algorithm (e.g. AAC) with model in OpenAI Gym (e.g., Walker2d) while it meets some dimensional problem. I am curious about whether we could implement the algorithms directly in other environments or we need to make essential changes. If the modifications are required, could you please give some hints on how the design works and which parts we should pay more attention to? Much appreciated for any help.

Bests, Lu

Gerryflap commented 2 years ago

Hello Lu,

I'm not entirely sure what you meant with "other environments". Do you mean environments that don't fit the OpenAI Gym format, or environments that use different dimensions of the observation tensor (i.e. larger/smaller vector or images) and different action spaces? I'd also advise you to take caution when using the implementations in this repo specifically. They definitely converge to something, but there's no guarantee that my implementations are completely correct because I was essentially just playing around and randomly mixing ideas from different algorithms.

Kind regards, Gerben

sunnyshi0310 commented 2 years ago

Hi Gerben,

Thanks for the reply. "other environments" means those are not defined in your folder "environments". I think I got your meaning while thanks anyway.

Bests, Lu