search

google-deepmind / meltingpot

A suite of test scenarios for multi-agent reinforcement learning.
Apache License 2.0
577 stars 116 forks source link

Help Request: Decreasing size of local observations #175

Closed AsadJeewa closed 11 months ago

AsadJeewa commented 11 months ago

I am working in a scaled-down version of cleanup to generate quick baselines. Please advise on how I can decrease the local observation size to be smaller than an 11x11 window

AsadJeewa commented 11 months ago

I was able to do this via clean_up.py config

duenez commented 11 months ago

For posterity, you can change the avatar view parameters: https://github.com/google-deepmind/meltingpot/blob/main/meltingpot/configs/substrates/clean_up.py#L696

However, doing so will make your agent incompatible with test scenarios. A more principled way would be to add a wrapper to your agent so that it would do this post-processing on its observations, that way it could also be placed in a scenario without problems. Something like meltingpot/utils/substrates/wrappers/base.py for a substrate at train time and an equivalent policy wrapper at test time.