Use of Machine Learning Techniques in Real-Time Strategy Games
Martin Čertický, Martin Sarnovský, Tomas Varga · 2018
Despite excessive amount of research done in the field of automated RTS gameplay, subtle changes in strategies are often ignored leading to non-optimal results. Researchers often consider obvious strategical decisions in order to cover as many gameplay scenarios as possible. In this paper, we focus on creating a dataset of atypical strategies which are often overlooked and using machine learning methods for detection of these strategies during the gameplay. Such information could be used to predict the strategies before they occur, or to correspond which adaptive behavior is able to answer them. In work presented in this paper, we approached the strategy recognition in StarCraft game using a set of classifiers trained on data obtained from the various replays. As there was not a proper dataset available to solve such task during our work, the dataset was created and annotated to cover four selected strategies for our experiments. Binary classification models were trained to detect each particular strategy and evaluated in a set of replay data using cross-validation technique. Then the overall platform architecture to train the models, export them and use in runtime during the gameplay was designed. Best performed models were then applied to real games to detect the covered strategies in replays or matches by bots or human players.