Spatial-Temporal Heterogeneous Graph Modeling for Opponent’s Location Prediction in War-game

In War-Game, the opponent’s location is the key information for decision-making, but it is hard to be completely known because of "the fog of war". Traditional location prediction methods model sequential behavior of a single target and predict the next location based on its recent locations. This restricted form of modeling benchmark limits the location prediction accuracy in War-game since explicitly omitting the fact that the movement of a target unit is affected by the situation and other units. In this paper, we propose a spatial-temporal heterogeneous graph neural network to encode semantic relation between combat units and integrate other units’ information into opponent's location prediction. The historical situation is modeled as a spatial-temporal graph, of which heterogeneous edges are used to describe different semantic relation of combat units. Furthermore, we apply spatial graph convolution block and temporal gated recurrent units for situation representation learning. To validate the effectiveness of our method, we construct two benchmark datasets with different scenarios and conduct comprehensive experiments on them. Experiments show that our proposed method extracts the correlated information of location prediction from other combat units and achieve 5.13% and 18.70% improvements over the strongest baselines.