PRIME: Plasticity Recovery in Multi-Agent Environments for UAV-Assisted Emergency Communication Networks

In multi-agent reinforcement learning (MARL) for UAV-assisted emergency communication, the environment is inherently non-stationary due to shifting disaster phases. The authors investigate how this sustained non-stationarity affects the internal state of shared-parameter policy networks, specifically focusing on the phenomenon of 'plasticity loss'—where neurons become dormant and the network loses its capacity to learn new tasks.

The researchers analyze the behavior of dormant neurons in a standard Multi-Agent Proximal Policy Optimization (MAPPO) controller. They identify that forward-only dormancy detection is insufficient because many neurons that appear inactive are still receiving significant gradient updates. To address this, they introduce PRIME (Plasticity Recovery in Multi-agent Environments). PRIME aggregates activation and gradient statistics across the entire team batch and only resets neurons that are simultaneously forward-dormant and gradient-silent. This ensures that the system does not destroy useful representations that are still being actively optimized.

The study finds that in non-stationary UAV networks, dormant neurons accumulate and persist, significantly degrading performance. The authors demonstrate that forward-only reset criteria (like those used in single-agent methods) are prone to high false-positive rates, as they discard neurons that the optimizer is still actively steering. By using a bidirectional criterion that incorporates the actual training gradient, PRIME maintains a lower fraction of dormant neurons (10–20% compared to 40–45% in vanilla MAPPO) and achieves a 24.9% improvement in interquartile mean return. Theoretical analysis further shows that this selective reset approach incurs a lower perturbation cost than global weight resetting.

As MARL is increasingly deployed in dynamic, real-world scenarios, maintaining the 'plasticity' of neural networks is critical. This paper provides a robust, architecture-agnostic method for ensuring that shared-parameter policies remain adaptable over long training horizons, preventing the silent erosion of decision-making capabilities in critical infrastructure applications.