Prediction of early stage opponents strategy for StarCraft AI using scouting and machine learning

Abstract

StarCraft is one of the most famous Real-Time Strategy Games and there have been several competitions on AI bots. In order to win StarCraft, players have to predict their opponents strategy and respond properly. Human players used to scout their opponent territory using a unit and gathering information through direct observation to predict their opponents strategy. The accurate prediction of an opponents strategy gives players a big advantage in the early stage of a game. Usually, strategies of StarCraft can be divided into two parts: fast and slow attack strategies. Initial attack timing is an important factor of game strategies. In this paper, we apply a scouting algorithm and various machine learning algorithms to predict an opponents attack timing (strategy). Training data are collected from the games between our Xelnaga bot with the scouting algorithm and various online human players. Experimental results show that the machine learning approach based on realistic scouting data can be beneficial in predicting the opponents early-stage strategy.