Predicting Frags in Tactic Games using Machine Learning Techniques and Intuitive Knowledge

High-quality prediction models can be used for game analysis and help to create challenging and human-like artificial intelligence (AI) agents. In this Predicting Frags in Tactic Games grand challenge, we apply our intuitive knowledge of video game for feature engineering, and propose a sequential floating forward and backward method combined with gradient boosting decision tree (GBDT) algorithm for feature selection. We generate different feature subsets which represent different aspects of the problem to train several GBDT models, and use the ensemble of these models for frag prediction. Our methods can be considered as a collection of heuristic rules and have well interpretability. They can act as heuristic evaluation functions that properly approximate the outcomes, which are critical to accelerating solution discovery in artificial intelligence search algorithms. The effectiveness of our methods was proven by experiments and competition results.