Center of Gravity-Guided Focusing Influence Mechanism for Multi-Agent Reinforcement Learning

ICLR 2026 Conference Submission21363 Authors

Keywords: Multi-Agent Reinforcement Learning, Intrinsic Motivation, Sparse Rewards, Credit Assignment, Influence Estimation, Coordination, Counterfactual Reasoning, Centralized Training with Decentralized Execution

Abstract: Cooperative multi-agent reinforcement learning (MARL) under sparse rewards presents a fundamental challenge due to limited exploration and insufficiently coordinated attention among agents. To address this, we introduce the Focusing Influence Mechanism (FIM), a framework that drives agents to concentrate their influence to solve challenging sparse-reward tasks. FIM first identifies Center of Gravity (CoG) state dimensions, inspired by Clausewitz’s military strategy, which are prioritized because when they include task-relevant variables, their low variability can block learning unless agents sustain influence. To encourage persistent and synchronized influence, FIM then focuses agents’ attention on these CoG dimensions using eligibility traces that accumulate credit over time. These mechanisms enable agents to induce more targeted and effective state transitions, facilitating robust cooperation even under extremely sparse rewards. Empirical evaluations across diverse MARL benchmarks demonstrate that FIM significantly improves cooperative performance over strong baselines.

Primary Area: reinforcement learning

Submission Number: 21363