Tag: #marl5 results
HyPOLE is a new framework that leverages hyperproperties and HyperLTL temporal logic to guide Multi-Agent Reinforcement Learning (MARL) under partial observability. It integrates Centralized Training for Decentralized Execution (CTDE) to synthesize robust decentralized policies.
Google's Paradigms of Intelligence team discovered that training AI agents against diverse, unpredictable opponents induces cooperation without hardcoded rules. Using decentralized reinforcement learning, LLM agents adapt via in-context learning to mixed-motive scenarios. This scalable method suits enterprise multi-agent systems.
R2D-RL is a new environment that integrates the RoboCup 2D Soccer Simulation (RCSS2D) with modern Python-based multi-agent reinforcement learning (MARL) workflows. It provides a synchronized, high-performance interface for training agents in complex, adversarial soccer scenarios.
Reusable RL decision structures depend on agent-world boundary. In multi-agent settings, peer policy updates create non-stationary MDPs, shrinking invariant state-action cores. Frames this as endogenous continual RL challenge from boundary instability.
DrIGM introduces distributionally robust IGM for MARL, ensuring decentralized actions align under uncertainties via robust value factorization. Compatible with VDN/QMIX/QTRAN without reward shaping. Boosts OOD performance in SustainGym and StarCraft.