TACTIC: Task-Aware Sparse Coordination Graphs for Multi-Task Multi-agent Reinforcement Learning

Value factorization eases non-stationarity in MARL, but its static coordination assumptions hinder generalization on long-horizon tasks with shifting dependencies. Prior VQ-VAE methods abstract trajectories yet miss time-varying inter-agent dependencies. We present TACTIC, a CTDE framework with three components: (i) VQ-VAE-based trajectory abstraction that learns discrete task-semantic classes; (ii) semantic-conditioned sparse coordination graphs that adapt dependencies by pruning edges according to variance-based pairwise payoff sensitivity; and (iii) a pretrained, frozen trajectory-class predictor that conditions local policies while decoupling task recognition from control. On SMAC and SUMO, TACTIC shows strong overall competitiveness and adaptive coordination under sparse rewards and dynamic task structures.

This paper studies how a team of learning agents can cooperate across different tasks. In many real-world settings, such as games or traffic control, agents need to change who they coordinate with as the situation changes. Existing methods often use fixed or overly simple coordination patterns, which can make them less effective when tasks vary. We propose TACTIC, a method that learns high-level patterns from agents’ past behavior and uses these patterns to build a sparse and adaptive coordination structure. This helps agents focus on the most relevant teammates instead of considering all possible interactions. Experiments in StarCraft-based multi-agent tasks and traffic control simulations show that TACTIC improves coordination, generalization, and stability across changing task settings.