A comparative evaluation of task allocation strategies for multi-agent drone swarms in wildfire management

Abstract

This thesis presents a simulation-based comparative study of task allocation strategies in multi-agent systems (MAS) for wildfire detection and suppression using autonomous drone swarms. Three task assignment strategies are implemented and evaluated: a heuristic method based on spatial proximity, an auction-based allocation mechanism relying on bid coordination, and a reinforcement learning (RL) approach trained with Proximal Policy Optimization (PPO). The wildfire environment is modeled as a cellular automaton that incorporates the parameters of wind, terrain, and moisture. Drone agents operate under resource constraints, such as limited battery life and restricted sensor range. The study evaluates each method in terms of task completion time, energy consumption, and workload distribution. The results indicate that the auction-based method achieves a better workload balance, while the RL-based strategy demonstrates adaptability in dynamic scenarios, albeit with higher training requirements. The heuristic approach, although computationally simple, shows significant performance degradation in larger fire spread situations. These findings provide insights into the practical trade-offs between coordination cost and task efficiency in MAS-based wildfire management.