search

nicoladainese96 / RelationalDeepRL

6 stars 1 forks source link

Code to run experiment #2

Open ava6969 opened 2 years ago

ava6969 commented 2 years ago

HI Can you please provide code to run this experiment. looks really cool but seems to have multiple exceptions thrown when i try to use train_agent

nicoladainese96 commented 2 years ago

Hi, I tried to refresh my memory about this project and I think this is a summary of what happened:

  1. I tried to implement the relational architecture from the paper and apply it to BoxWorld (probably with a single-process actor critic), but never got the agent to work.
  2. I then played around a bit and ended up implementing batched A2C and A3C for a simple grid world navigation task (it was more of a sanity check than a proper task to solve)
  3. Finally I migrated the code that worked on my repo https://github.com/nicoladainese96/SC2-RL where I focused on the StarCraft II Learning Environment and ended up re-implementing a baseline architecture and trained it with batched A2C and IMPALA agents.

As far as I remember at the time I never got the relational module to work, but we could look into it if you're interested.