Mobile games represent a killer application that is attracting millions of subscribers worldwide. One of the aspects crucial to the commercial success of a game is ensuring an appropriately challenging artificial intelligence (AI) algorithm against which to play. However, creating this component is particularly complex as classic search AI algorithms cannot be employed by limited devices such as mobile phones or, even on more powerful computers, when considering imperfect information games (i.e., games in which participants do not a complete knowledge of the game state at any moment). In this paper, we propose to solve this issue by resorting to a machine learning algorithm which uses profiling functionalities in order to infer the missing information, thus making the AI able to efficiently adapt its strategies to the human opponent. We studied a simple and computationally light machine learning method that can be employed with success, enabling AI improvements for imperfect information games even on mobile phones. We created a mobile phone-based version of a game calledGhostsand present results which clearly show the ability of our algorithm to quickly improve its own predictive performance as far as the number of games against the same human opponent increases.
Limited lookahead in imperfect-information games
