The Evolution of Fun: Automatic Level Design through Challenge Modeling Nathan Sorenson and Philippe Pasquier School of Interactive Arts and Technology, Simon Fraser University Surrey, 250 -13450 102 Avenue, Surrey, BC {nds6, pasquier}@sfu.ca Abstract. A generative system that creates levels for 2D platformer games is presented. The creation process is driven by generic models of challenge-based fun which are derived from existing theories of game design. These models are used as fitness functions in a genetic algorithm to produce new levels that maxi- mize the amount of player fun, and the results are compared with existing levels from the classic video game Super Mario Bros. This technique is novel among systems for creating video game content as it does not follow a complex rule- based approach but instead generates output from simple and generic high-level descriptions of player enjoyment. Key words: video games, challenge-based fun, level creation, genetic algorithms 1 Introduction Existing processes for creating video game levels are time consuming and expensive, and result in environments that are static and cannot be easily adjusted. Clearly, leverag- ing a creative system that automatically designs levels would enable independent game developers to generate content that would otherwise require the resources of larger com- panies. If effective, this process would also result in environments that are not static but are instead amenable to endless variety and modification, ideally creating more inter- esting and enjoyable experiences for game players. Furthermore, a dynamic and unsu- pervised method of level creation would allow a virtually limitless amount of content which could be produced off-line and distributed to clients as expansions to the base game, or generated on-the-fly, whereby levels could adapt to individual players as the game is being played. Although automated methods for creating content are occasionally seen in exist- ing games [1–3], current approaches follow a bottom-up, rule-based approach. This method requires a designer to embed aesthetic goals into a collection of rules, resulting in systems just as difficult to construct as hand-designed levels [4]. Genetic algorithms instead allow developers to specify desirable level properties in a top-down manner, without relying on the specifics of the underlying implementation. However, any effec- tive fitness function for automated level creation must correctly identify the levels that are “fun.” To this end, a model of what precisely constitutes a fun level must be de- veloped. In this paper, two such models are proposed: one based on Csikszentmihalyi’s concept of “flow” [5] and the other on the notion of “rhythm groups,” as described by