The battlefield environments being complex, dynamic, characterized by high dynamics, incomplete information, and uncertainty, the deep reinforcement learning (DRL) is enabled to provide a new way of thinking about task assignment in modern information warfare. Aimed at the problem that the agent system is inadequate in generalization ability under condition of uncertain scenario, this paper proposes an event-based reward mechanism to reasonably guide the learning of the agent, and the problem that in deep reinforcement learning, a single reward function is difficult to train an agent of being in keeping with human decision logic, this paper proposes an event-based reward mechanism to reasonably guide the learning of the agent. And this paper proposes a multi-agent architecture for different decision styles, enhancing the ability of the agent to adapt to complex environments. Finally, the feasibility and superiority of the proposed method are verified on a digital battlefield.