Deep Reinforcement Learning for Dynamic Resource Allocation : Generalizing Defensive Strategies Based on Pairwise Unit Analysis

Managing resource allocation dynamically across varying scenarios is a complex problem that necessitates robust and flexible approaches. In this paper, we propose a deep reinforcement learning approach based on an effector-target modeling framework to solve the dynamic weapon-to-target assignment problem (WTA). Using StarCraft II as a simulation environment, we created various WTA scenarios. Our approach includes a method utilizing MLPs and another leveraging the attention mechanism. This method is input-size independent, enables knowledge generalization, and allows agents to perform well across diverse scenarios, while also providing an initial level of explainability for decision-making.