B
Bender
Text
Peer-Incentivized Reward Functions for Exploration in Multi-Agent Reinforcement Learning
A generalized multi-agent social-influence peer reward that rewards an agent for influencing its peers toward novel states, unifying prior causal-influence and value-of-interaction rewards under weighted terms and adding a new term in which an agent shares in a peer's reward proportional to its influence on that peer. Students learn how peer incentivization can drive coordinated exploration in cooperative settings with sparse or deceptive rewards, realized through a communication channel and counterfactual influence estimation on top of any actor-critic algorithm.