Learning a context-aware weapon selection policy for Unreal Tournament III

Abstract

Modern computer games are becoming increasingly complex and only experienced players can fully master the game controls. Accordingly, many commercial games now provide aids to simplify the player interaction. These aids are based on simple heuristics rules and cannot adapt neither to the current game situation nor to the player game style. In this paper, we suggest that supervised methods can be applied effectively to improve the quality of such game aids. In particular, we focus on the problem of developing an automatic weapon selection aid for Unreal Tournament III, a recent and very popular first person shooter (FPS). We propose a framework to (i) collect a dataset from game sessions, (ii) learn a policy to automatically select the weapon, and (iii) deploy the learned models in the game to replace the default weaponswitching aid provided in the game distribution. Our approach allows the development of weapon-switching policies that are aware of the current game context and can also imitate a particular game style.