Plasticity-Enhanced Multi-Agent Mixture of Experts for Dynamic Objective Adaptation in UAVs-Assisted Emergency Communication Networks

A research team has published a paper on arXiv detailing PE-MAMoE, a novel AI framework designed to solve a critical problem in emergency drone networks. When drones (UAVs) act as flying cell towers after a disaster, user movement and shifting data demands create a highly unstable environment that causes standard deep reinforcement learning (DRL) models to fail. This failure, known as 'plasticity loss,' occurs when AI agents' internal representations collapse and neurons become dormant, crippling their ability to adapt to new conditions.

PE-MAMoE tackles this by giving each drone agent a 'Mixture of Experts' actor network. A smart router selects the best specialist 'expert' for each decision. The key innovation is a 'Phase Controller' that injects carefully calibrated stochastic noise and resets key parameters whenever the mission phase changes. This 're-plasticizes' the AI, waking up dormant neurons and restoring adaptability without destabilizing safe, learned behaviors. The team proved its effectiveness with a dynamic regret bound and extensive simulation.

In a realistic simulator with mobile users and 3GPP-style wireless channels, PE-MAMoE delivered substantial gains over existing baselines. It improved the normalized interquartile mean return—a key performance metric—by 26.3%. The system also increased the network's served-user capacity by 12.8% and dramatically reduced the risk of drone-on-drone collisions by approximately 75%. Diagnostics confirmed the method successfully maintained higher feature diversity in the AI and periodically recovered dormant neurons at critical regime switches.