Multi-UAV-aided power-up and data collection: Multi-agent DQL with genetic algorithm approach

초록

Recently, unmanned aerial vehicles (UAVs) are frequently used in wireless sensor networks (WSNs) as mobile power supplies and information collectors, due to their flexible deployment, high mobility, and low cost. In this paper, we consider multi-UAVs equipped with wireless power transfer (WPT) functionality, which uses array antennas to charge the batteries of sensor nodes (SNs) and collect data from them in WSNs. We aim to minimize the maximum mission completion time for all UAVs by jointly optimizing the trajectories of multiple UAVs, transmitting power, and beam-forming. To solve this complex problem, we provide two solutions, the minimum flying time algorithm and the multi-agent deep Q-learning (MA-DQL) algorithm. For the minimum flying time algorithm being used for benchmarking, we first group the SNs by using the K-means algorithm, and find the UAVs’ hovering locations. After that, we utilize the alternative optimization technique, which separates the main problem into two sub-problems. Each sub-problem is solved repeatedly until convergence using ant colony optimization (ACO) and convex optimization tools (CVX), respectively. More importantly, we develop the MA-DQL algorithm to solve this problem, a type of deep reinforcement learning (DRL) that involves multiple agents, with each agent corresponding a UAV. Furthermore, we enhance the MA-DQL framework with a genetic algorithm (MA-DQL with GA) that evolves complete decision trajectories in replay memory through crossover and mutation. This structured experience evolution improves exploration efficiency, learning stability, and constraint satisfaction during training. Simulation results demonstrate that the proposed method reduces mission completion time, improves energy efficiency, and ensures fairness across UAVs in decentralized settings.

키워드