This work proposes a novel approach for improving the control performance of Unmanned Aerial Vehicles (UAVs) through cooperative reinforcement learning. By sharing their experience, it is shown that multiple UAVs can work together to converge on a set of optimal Model Predictive Control (MPC) parameters faster than when working on their own. In order to benefit from this shared experience, the UAVs must coordinate their learning strategies. Here, we proposed a Leader-Follower approach, whereby the Leader ensures all trials are drawn from the same distribution and contribute to a common payoff game of Learning Automata. Experimental results show that this approach results in faster learning without any loss of performance.

这项工作提出了一种通过协作强化学习提高无人驾驶飞行器 (UAV) 控制性能的新方法。通过分享他们的经验，表明多架无人机可以协同工作以比单独工作时更快地收敛到一组最佳模型预测控制 (MPC) 参数。为了从这种共享经验中受益，无人机必须协调他们的学习策略。在这里，我们提出了一种领导者-跟随者方法，即领导者确保所有试验都来自相同的分布，并有助于学习自动机的共同收益游戏。实验结果表明，这种方法可以在不损失任何性能的情况下实现更快的学习。