Attention-Based Initiative In Real Time Strategy

Abstract

Initiative is a concept which describes certain behaviors in collaborative or competitive play. Due to the broad usage and qualitative nature of the initiative concept, quantitative modeling poses several challenges. We propose to model competitive initiative by measuring what players pay attention to during gameplay. For this purpose, we decompose player actions into discrete types, Voronoi spaces and time- ranges. We test and analyze our model empirically on a Real Time Strategy (RTS) dataset. As part of the analysis, we use our model to predict game outcomes through time with the Random Forest algorithm. Results show that a Pareto front can be established between game time and the predictive accuracy of game outcomes, which starts at 50%, followed by an exponential growth towards 80%. We conclude that there is empirical support for attention-based initiative. Future work can be directed towards refining and expanding on the model for analytical and/or predictive usecases. For reproducibility, we share data and corresponding results in a public repository.