Feature Engineering for Predicting Frags in Tactical Games
Dymitr Ruta, Ming Liu, Ling Cen, Di Wang · 2023
In tactical games players command troops of soldiers to fight against human or AI opponents in a variety of 2/3D maps involving specialized units and weapons. In turn-based games one of the key objective of each turn is to score a frag i.e. kill at least one unit of the opposing player. Efficient prediction of frags could be used to design a high quality AI-controlled agents that moderate and fine-tune the game in real-time to make it challenging, immersive and attractive to the human players. In the context of ICME 2023 Grand Challenge to predict frags in Anthracite Shift game turns, we have engineered efficient and robust families of predictive features that combined with state-of-the-art tree-based boosting algorithms deliver highly robust classification models that can be retrained in seconds and make accurate frag predictions in real-time. The feature families are derived from health-points (HP), distances (DS) and line-of-sight (LS) obstruction among opposing player units, consistently sorted by HP/LS to properly correlate various relational properties of all pairs of opposing units. With just over 70 fine-tuned features our shallow and actively diversified ensemble of boosting model variants achieves the 2ndhighest preliminary leader-board AUC score of 0.8533 and the 3rdplace in the final evaluation.