Multiagent Coordination Without Communication in Evaluation

In complex scenarios, multiagent coordination is a challenging task. we introduce a multiagent coordination algorithm that integrates a generative strategy network with a coordinative strategy network to address the cooperation challenge among multiple agents tasked with reaching different target positions from random initial positions, without communication during the evaluation phase. During the training process, the generative strategy network utilizes images collected from the agent in four directions to generate the agent-centered top view. Concurrently, the coordinative strategy network uses both the first-person view and the agent-centered top view as inputs to determine the corresponding actions. In the evaluation phase, images of the generated agent-centered top view and first-person view are input to the coordinative strategy network, which relies solely on internally collected images for strategic decision-making. Our method achieves coordination by seamlessly integrating local and global information, utilizing a distinctive combination of neural network architectures. This approach successfully addresses the multiagent coordination challenge without the need for direct communication in diverse and complex environments.