Multi-agent Reinforcement Learning for Coordinated Emergency Building Sweeps

Abstract

We develop a Multi-Agent Deep Deterministic Policy Gradient (MADDPG) framework for coordinating emergency responders during building sweeps. The problem is formulated as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) that captures building topology, occupant vulnerability, dynamic hazards, and partial observability. In the six-room office baseline, MADDPG reaches a 134-second sweep time after 5,000 episodes, within 6.3% of the 126-second theoretical optimum and outperforming PPO and DQN. Extensions to a two-story school and a U-shaped warehouse achieve 120 seconds with five responders and 105 seconds with four responders, respectively, with diminishing returns beyond these staffing levels. Under communication failure, coverage remains about 98% while sweep time rises 15–20%; reduced staffing triggers workload redistribution, and dynamic hazards cause threatened areas to be prioritized. Occupancy sensing, mesh networks, and environmental sensing are projected to reduce sweep time by 35–50%. Sensitivity analysis confirms robust performance across realistic parameter ranges.