Browse Teach Curate

Communicating Unexpectedness in Out-of-Distribution Multi-Agent Reinforcement Learning

A decentralized MARL algorithm in which each agent predicts its next observation, measures the discrepancy against the observation actually seen, and broadcasts that prediction-error 'unexpectedness' as a message alongside reward-driven communication, so cooperating agents adapt robustly to situations outside their training distribution.

KITT Text

Communicating Unexpectedness for Out-of-Distribution Multi-Agent Reinforcement Learning Min Whoo

The Unexpected Encoding Scheme is a decentralized multi-agent reinforcement learning method for adapting to out-of-distribution situations. Beyond the usual reward-driven messages, each agent maintai…