How do agents learn to cooperate?

hard

Answer

$(1) Shared reward \to both agents want same outcome (value decomposition, MAPPO).$
$(2) Communication learning (RIAL, DIAL, CommNet — learn what/when to communicate).$
$(3) Learn joint value function (QMIX).$
$(4) Difference reward: reward each agent for its marginal contribution.$
$(5) Curriculum from easier sub-tasks to full coordination.$
$(6) Social dilemmas (Iterated PD): reciprocity via memory / punishment.$

Check yourself — multiple choice

Random
Shared reward + comm learning (RIAL/DIAL) + joint value (QMIX) + difference reward (marginal) + curriculum + social dilemmas (reciprocity) — many paths
Same as PPO
Not real

Coop: shared reward + comm + joint value + difference + curriculum + reciprocity.

#multi-agent

214 interview questions in this topic.