What MARL is: Multi-Agent Reinforcement Learning (MARL) enables independent autonomous agents to learn optimal policies through trial-and-error interactions in a shared environment.

The Coordination Bottleneck: Coordinating multiple agents is difficult due to environmental non-stationarity: when all agents update their policies simultaneously, the transition dynamics shift unpredictably, causing policy oscillations and training instability.

Our Solution: We introduce dependency-aware activation scheduling, which models agent interdependencies as a directed graph. Instead of concurrent updates, agents are scheduled sequentially based on their causal coupling weight.

Efficiency & Convergence: Sequential scheduling reduces training updates and policy conflict, stabilizing the learning path and accelerating convergence compared to standard simultaneous update models.