Multi-agent coordination failures fixed by targeted exploration injection
Anchor and Perturb: Lazy Agent Remediation by Exploration Injection
Multiagent Systems
Summary
In situations where teams of AI agents struggle to work well together, some agents can get 'lazy' and stop exploring new strategies, causing poor overall results. The authors present Anchor and Perturb (AnP), a technique that selectively encourages these underperforming agents to try new actions while keeping well-performing teammates steady. This focused approach helps the group recover from failings and reach better team performance without changing the whole system’s design. The approach was tested to improve win rates significantly in complex cooperative tasks.
What this means in practice
- •For multi-agent system developers: Improve team performance by selectively enhancing exploration only for underperforming agents while keeping others stable.
- •For ai game developers: Prevent cooperative game AI agents from getting stuck in poor coordination by applying targeted exploration recovery techniques.
Authors
Chengxi Zhong, Yongzhe Chang
Abstract
Anchor and Perturb (AnP) is a lightweight framework that resolves multi-agent coordination failures by decoupling exploratory variance injection from recurrent manifold stability. Existing remediation strategies predominantly alter mixing network architectures or enforce simultaneous exploration across the collective, which inevitably precipitates severe temporal-difference penalties in non-monotonic reward spaces. Specifically, AnP isolates underperforming lazy agents and injects an asymmetric exploratory pulse into targeted coordinates whilst anchoring converged teammates to nominal greedy exploitation. Empirical telemetry benchmarks demonstrate that AnP successfully rescues collapsed joint policies (recovering from a 5% evaluation win rate nadir back to 85%) and facilitates escape from suboptimal coordination plateaus, sustaining peak win rates of 90% without requiring structural network modifications.