Dynamic Attention Network for Multi-UAV Reinforcement Learning

Recent methods for multi-agent reinforcement learning problems make use of Deep Neural Networks and provide stateof-the-art performance with dedicated neural network architectures and comprehensive training tricks. However, these deep reinforcement learning methods suffer from reproducibility issues, especially in transfer learning. Since the fixed size of the network input, it is difficult for the existing network structure to transfer the strategies learned from a small scale to a large scale. We argue that proper network architecture design is crucial to the cross-scale reinforcement transfer learning. In this paper, we use transfer training with attention network to solve multi-agent combat problems from aerial unmanned aerial vehicle (UAV) combat scenarios, and extend the small-scale learning to large-scale complex scenarios. We combine the attention neural network with the MADDPG algorithm to process the agent observation. It started training from a small-scale multi-UAV combat scenario and gradually increases the number of UAV. The experimental results show that methods for multi-agent UAV combat problems trained by attention transfer learning can achieve the target performance faster and provide better performance than the method without attention transfer learning.