Automated Discovery of Search-Extension Features P´ almi Skowronski 1 , Yngvi Bj¨ornsson 1 , and Mark H.M. Winands 2 1 Reykjav´ık University, School of Computer Science, Kringlan 1, Reykjav´ık 103, Iceland {palmis01,yngvi}@ru.is 2 Games and AI Group, Department of Knowledge Engineering, Maastricht University, Maastricht, The Netherlands m.winands@micc.unimaas.nl Abstract. One of the main challenges with selective search extensions is designing effective move categories (features). This is a manual trial and error task, which requires both intuition and expert human knowl- edge. Automating this task potentially enables the discovery of both more complex and more effective move categories. In this work we in- troduce Gradual Focus, an algorithm for automatically discovering in- teresting move categories for selective search extensions. The algorithm iteratively creates new more refined move categories by combining fea- tures from an atomic feature set. Empirical data is presented for the game Breakthrough showing that Gradual Focus looks at two orders of magnitude fewer combinations than a brute force method does, while preserving good precision and recall. 1 Introduction The αβ algorithm is one of the fundamental and most effective search techniques used by game-playing programs for playing two-person adversary board games, such as chess and checkers. Over the years many enhancements have been pro- posed to further improve its efficiency. In particular, it has long been evident that the standard strategy of exploring all alternatives to the same fixed depth is not the most effective. Instead various techniques have been proposed for searching the game tree more selectively, where some lines of play are terminated prema- turely whereas others are explored more deeply. The former scenario is referred to as search reductions (or speculative pruning) and the latter as search exten- sions. In chess, for example, it is common to resolve forced situations, such as checks and recaptures, by searching them more deeply. The move-decision quality of the alpha-beta algorithm is greatly influenced by the choices of which lines are investigated deeply [1,2]. Therefore, the design of an effective search-extension scheme is fundamental to any high-performance αβ-based game-playing program. The typical approach for incorporating search extensions into a game-playing program is to predefine a set of move categories (e.g. checks and recaptures), and then associate a different cost weight to each