PDF《Ebner, Marc and Levine, John and》

licensed under Creative Commons License NC-ND Dagstuhl Seminar on Arti?cial and Computational Intelligence in Games. Editor: Editor;

pp. 1C17 Dagstuhl Publishing Schloss Dagstuhl C Leibniz-Zentrum für Informatik, Germany

2 Video Game Description Language Generation, Automatic Game Design and Transfer Learning and how the roadmap for GVGP can provide opportunities for these areas. Keywords and phrases Video games, description language, language construction Digital Object Identi?er 10.4230/DFU.xxx.yyy.p

1 Motivation Recent discussions at the Dagstuhl seminar on Arti?cial and Computational Intelligence in Games addressed the challenge of extending the principles of General Game Playing (GGP) to video games. While the rationale and challenges facing research in General Video Game Playing (GVGP) is the focus of another report in these proceedings, the need for a Game Description Language (GDL) designed to suitably express the key concepts and mechanics inherent in video games became apparent. This has led to this report exploring the challenges and potential pitfalls faced in creating a Video Game Description Language (VGDL) that will work as part of the GVGP process. The focus of traditional GDLs is to express components expected in the state of a game, and the rules that induce transitions, resulting in a state-action space. Typically, GDLs for use in GGP competitions conform to a set-theoretic language that expresses atomic facts in predicate logic. One of the more popular examples: the GDL component of the Stanford General Game Playing Competition employs this approach to de?ne discrete games of complete information (4). While this provides means to de?ne and subsequently generate games that conform to this particular set, the language is ill-suited for de?ning video game environments. This is due to a number of factors: Nondeterministic behaviour in video games as a result of non-player characters (NPCs) or elements of chance are common. Decisions for actions in video games can be attributed to player input or prescribed NPC behaviour. These decisions are no longer turn-based and may occur simultaneously at any step of the game. Video game environments may employ continuous or temporal e?ects, real-time physics and context-based collisions or interactions between combinations of players and artefacts. Part of the player'

s task is learning to predict those dynamics. Typical video games use much larger environments than board games for the player to interact with. However, the player-environment interactions are sparse, meaning that a part of the environment complexity is rarely or never in?uencing decision making. These factors among others present an interesting research challenge in its own right: to try and maintain a balance between simplicity and generality in a description language whilst ensuring it is suitably expressive. Driven by these factors and the challenge they represent, we have raised key goals that shape our aspirations: To de?ne a GDL that supports the core mechanics and behaviour expected of classical 2D video games. To ensure the VGDL de?ned will support the work being developed in the GVGP discussion group. De?ne a language that is su?cient not only to represent a game to the human reader, but also for a compiler to generate an instance of the game. Ebner, Levine, Lucas, Schaul, Thompson and Togelius

3 Provide a VGDL that is unambiguous, extensible and tractable that could provide opportunities for Procedural Content Generation (PCG). In this report, we consolidate the key points raised by the authors as we explored the challenges and impact of expanding the current research in Game Description Languages for Video Games to further these design goals. We highlight the current state-of-the-art in the ?eld, the core criteria and considerations for the design and subsequent construction of a VGDL, the challenges the problem domain will impose, the potential for future research and collaboration and ?nally highlight preliminary steps in developing a VGDL prototype.