1Biologically and Cognitively Inspired General Game Playing: Self-Motivation in Sokoban, Gambler’s Problem and Pacman

Abstract—We use empowerment, a recently introduced biolog-ically inspired measure, to allow an AI player to assign utility values to states within a previously un-encountered game where it has no knowledge of any existing goal states. We demonstrate how an extension to this concept, open-ended empowerment, allows the player to group together candidate action sequences into a number of ‘strategies’. This grouping is determined by their strategic affinity, and these groupings are used to select a repertoire of action sequences that aim to maintain anticipated utility. We show how this method provides a proto-heuristic for non-terminal states in advance of specifying the concrete game goals, and propose it as a principled candidate model for “intuitive” strategy selection, in line with other recent work on generation of “self-motivated agent behaviour”. We in particular demonstrate that the technique, despite being generically defined indepen-dently of scenario, performs quite well in relatively disparate scenarios, such as a Sokoban-inspired box-pushing scenario, in Dubin and Savage’s Gambler’s Problem, and in a simplified Pac-man game, and it suggests novel and principle-based candidate routes towards more general game-playing algorithms. Index Terms—Artificial intelligence (AI), information theory