Machine learning for Smartphone games

Details

Supervisors

Dupont, Pierre

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques

Abstract

We investigate different approaches in the creation of artificial players on smartphone. Instead of implementing a fixed way of playing, we want the artificial player to be able to learn how to play a game efficiently by itself. Learning for an artificial player implies that the more it plays games, the more efficient the artificial player should become. Our first idea was to choose a simple game where the AI and the player had to choose among 4 actions. The goal of the player was to guess AI’s move and to play it while the AI had to try not to play the same action. Unfortunately, the best strategy in this game is to play randomly. Even if we know humans can’t play perfectly randomly, although we know they always have a strategy, it is almost impossible to identify it. We then selected another game, a little more complex and for which there existed better strategies than pure chance. We will expose our conclusion about this game using different models. We’ll analyse how implement various AI’s that are able to learn this game and examine the performance and advantages of each of them.