search

oxwhirl / pymarl

Python Multi-Agent Reinforcement Learning framework
Apache License 2.0
1.82k stars 380 forks source link

Why add agent's previous action to agent's input? #153

Closed tolyan3212 closed 1 year ago

tolyan3212 commented 1 year ago

Hello,

In this implementation agents' network is a recurrent network, and this should allow for agents to remember necessary previous actions and observations. But the multi-agent controller also appends to agents' inputs their previous action.

What is the reason for that, and is it really necessary?

Thanks!

samvelyan commented 1 year ago

If you are using the SMAC benchmark, there is a special flag for this called obs_last_action. See SMAC code for more details on all flags that specify the observation function.