MUTE framework cuts multi-agent communication by 80-90% without performance loss
New 'machine unlearning' approach lets AI agents talk less and achieve more.
Multi-agent reinforcement learning (MARL) agents often need to communicate to coordinate under partial observability, but real-world bandwidth constraints demand sparse interactions. Prior approaches used information-theoretic proxies to prune messages, but these can discard valuable information. The new MUTE framework, developed by Rui Zuo and collaborators, reframes communication reduction as a value-guided machine unlearning problem. It introduces a Counterfactual Message Value metric, estimated via an attention mechanism, to quantify each message's contribution to the team's overall return. Messages deemed low-value are systematically 'unlearned' from a policy originally trained with full communication.
MUTE achieves this through a dual-objective loss that enforces communication sparsity while preserving the return of the original joint policy. The authors provide a theoretical upper bound on the performance gap caused by sparsification, guaranteeing controlled degradation. Empirically, across complex multi-agent environments, MUTE reduces bandwidth by 80-90% while matching or exceeding state-of-the-art baselines. This work opens the door to deploying efficient, communication-constrained multi-agent systems in robotics, autonomous vehicles, and sensor networks.
- MUTE uses an attention-based estimator to compute Counterfactual Message Value for each agent's message.
- The dual-objective mechanism enforces sparsity while preserving the original joint policy's return, with a theoretical performance bound.
- Achieves 80-90% bandwidth reduction across complex MARL benchmarks without degrading task performance.
Why It Matters
Enables real-world deployment of multi-agent AI under severe bandwidth constraints, from drone swarms to autonomous vehicles.