PRIME boosts UAV emergency network RL by 24.9% with neuron recovery
New RL method solves neuron death in multi-agent drone networks, boosting performance 24.9%.
PRIME is a plasticity recovery method for multi-agent reinforcement learning in UAV-assisted emergency communication networks. It identifies and reinitializes only truly silent neurons using bidirectional activation and gradient statistics, preserving useful representations while restoring learning capacity. On a phase-switching UAV simulator, PRIME improves interquartile mean return by 24.9% over MAPPO and maintains dormant neuron fractions at 10–20% (vs. 40–45%). The method is safe for shared-parameter multi-agent training, with a dynamic regret bound showing that perturbation cost scales with the silent-subspace dimension rather than the full parameter count.
- PRIME achieves 24.9% improvement in interquartile mean return over MAPPO in UAV emergency communication simulations.
- Dormant neuron fraction reduced from 40–45% to 10–20%, preserving learning capacity under non-stationarity.
- Method safely resets neurons by verifying both activation dormancy and gradient silence across the multi-agent team.
Why It Matters
Enables resilient multi-agent AI for disaster response networks, where continuous learning under shifting conditions is critical.