Orchestrating the generation of game facets via a model of gameplay

Abstract

Computer games are media that weave together many different facets. When the design of games is supported by the automatic creation of game content, a multidisciplinary approach would be expected. Yet, many approaches to procedural content generation tend to focus on a single facet at a time, assuming that a human game designer will guarantee a suitable context. The systems that do create multiple types of content usually rely on expensive game simulations to evaluate their quality and complementarity. However, the complexity of modern games increases the runtime of these simulations to such a degree that at some point this approach becomes infeasible. This thesis proposes a framework for the procedural generation of the level and ruleset components of games via a model of gameplay that can act as a surrogate for expensive game simulations. By combining the level and ruleset components as input and gameplay outcomes as output, deep learning is used to construct a mapping between three different facets of a game. This thesis argues that the learned mapping enables the model to identify the synergies between these facets, which can be used to orchestrate the generation of both level and rules towards desired gameplay outcomes. The experiments support this by demonstrating the ability of a search-based generative approach that uses a surrogate model for quality evaluation to adapt players’ character classes, levels, or both towards designer-specified targets in the domain of shooter games. The findings demonstrate that the proposed method of game facet orchestration can produce improved designs of both facets without the use of simulations and makes less changes to an initial design than traditional single-facet methods.