Core Concepts
IIE proposes a novel method for efficient multi-agent exploration in complex scenarios using a transformer model to imagine critical states and trajectories before exploration.
Abstract
"IIE" introduces an innovative approach to multi-agent exploration by leveraging a transformer model to imagine how agents reach critical states. The method involves imagining trajectories from initial states to interaction states, utilizing prompts like timestep-to-go, return-to-go, influence value, and one-shot demonstrations. By initializing agents at critical states identified through imagination, IIE significantly enhances the likelihood of discovering important under-explored regions. Empirical results demonstrate that IIE outperforms existing methods on challenging tasks like the StarCraft Multi-Agent Challenge (SMAC) and SMACv2 environments. The proposed method bridges sequence modeling and transformers with MARL, offering promising results in complex cooperative scenarios."
Stats
"Empirical results demonstrate that our method outperforms multi-agent exploration baselines on the StarCraft Multi-Agent Challenge (SMAC) and SMACv2 environments."
"IIE shows improved performance in sparse-reward SMAC tasks."
"Our method produces more effective curricula over initialized states than other generative methods."