PG642 / multi-sample-factory

High throughput reinforcement learning on clusters
MIT License
2 stars 0 forks source link

Test MSF with own gym wrapper and a mixed action space #25

Open KonstantinRamthun opened 2 years ago

KonstantinRamthun commented 2 years ago

General Description

A rocket league agent has naturally 3 continuous and 5 multi-discrete actions. Being able to represent this structure in our model may be beneficial for the agent's performance. The environment team implemented a gym unity wrapper, which supports mixed actions spaces (the wrapper of ml-agents does not support this feature). With this wrapper we can test, whether MSF supports mixed action spaces and whether we get any performance boost out of it.

Tasks