TheButlah / makrl

makrl - modular algorithm kit for reinforcement learning
4 stars 1 forks source link

Get a better implementation of environment batching #24

Closed TheButlah closed 5 years ago

TheButlah commented 5 years ago

We need an alternative to baselines's SubprocVecEnv. We don't need the entirety of baselines as a dependency, and the environments cannot be individually reset which is very important to our approach, as we want the batch size to remain the same size for convenience purposes and to avoid over fitting on longer episodes as smaller batches mean biased gradients.

TheButlah commented 5 years ago

I spent all of today working on this and have finished the implementation.