CRADLE: Cooperative Adaptive Decentralized Learning and Execution for UAV network to monitor Wildfire Front

In this paper, we investigate the challenges of continuous wildfire front monitoring using Unmanned Aerial Vehicles (UAVs). While several UAV learning and execution strategies exist to effectively monitor the wildfire progression, they grapple with centralized limitations such as scalability, single-point-of-failure, lack of adaptability, etc. To address these, we introduce a novel decentralized framework— Cooperative Adaptive Decentralized Learning and Execution (CRADLE)— diverging from conventional Centralized Training and Decentralized Execution (CTDE) models. CRADLE promotes adaptive coordination among UAVs by leveraging a double Q-learning (DDQN) algorithm with a weighted experience aggregation mechanism. Our evaluations indicate our CRADLE framework reached performance levels akin to CTDE approaches. In other words, our decentralized framework can exhibit performance on par with centralized frameworks.