Aerial multi-agent collaboration is an important research direction in the field of Multi-Agent Systems (MAS), whose core goal is to achieve efficient collaboration and autonomous decision-making among multiple agents in a complex aerial environment. In this paper, a multi-agent intensive learning (MARL) model based on centralized training and decentralized execution is proposed, which provides a more effective scheme for multi-machine cooperative tasks in air game. By comparing various advanced algorithms in many task scenarios, the advantages of the model in improving collaboration efficiency and task completion rate are verified, which can provide reference for the application of multi-agent system in other complex environments.