Multi-Agent Collaborative Behavior Decision-Making based on Deep Reinforcement Learning

Abstract

For the problem of collaborative decision-making, we propose a multi-agent deep reinforcement learning collaborative behavior decision-making algorithm. Firstly, a discrete state space and a greedy strategy-based action space are established in the context of multi- agent collaborative attack, the conditions for successful collaborative siege are given for the requirements of rapidity and collocation. Secondly, the Markov Decision Process (MDP) framework is established based on the multi-agent collaborative behavior decision algorithm, we introduce the experience replay to train the neural network using gradient descent. Finally, a centralized training and distributed execution architecture is used to complete the training of collaborative behavioral decision making, in which the agents share the same strategy and execute actions independently. The simulation shows that the deep reinforcement learning algorithm is able to realize the multi-agent collaborative decision. It can be placed in a real environment.