Clustering and Classification of Time Series in Real-Time Strategy Games - A machine learning approach for mapping StarCraft II games to clusters of game state time series while limited by fog of war

Abstract

Real-time strategy (RTS) games feature vast action spaces and incomplete information, thus requiring lengthy training times for AI-agents to master them at the level of a human expert. Based on the inherent complexity and the strategical interplay between the players of an RTS game, it is hypothesized that data sets of played games exhibit clustering properties as a result of the actions made by the players. These clusters could potentially be used to optimize the training process of AI-agents, and gain unbiased insight into the gameplay dynamics. In this thesis, a method is presented to discern such clusters and classify an ongoing game according to which of these clusters it most closely resembles, limited to the perspective of a single player. Six distinct clusters have been found in StarCraft II using hierarchical clustering over time, all of which depend on different combinations of game pieces and the timing of their acquisitions in the game. An ongoing game can be classified, using neural networks and random forests, as a member of some cluster with accuracies ranging from 83% to 96% depending on the amount of information provided.