Game design intelligence

About game generation

Game generation is the creation of parts of games (e.g., levels) or entire games automatically. This could be for a variety of purposes — for example, to make a game endless, or to adapt it to the player, or to make it tractable for non-specialists to create.

Just like hand-authored games, generated games create experiences and embed ideas. When creating with a generator, we want to be able to author these experiences and ideas, iteratively, rather than accept the likely flat and biased output of a black box. That is, we want to be able to author the generator at a meta level — to author the generative author. This desire has informed the approaches taken in EIS research in this area.

Investigation of game generation started as soon as EIS was founded — before I joined, in 2008. EIS student Mark Nelson began, in 2006, with a system that automatically "skinned" WarioWare-style minigames to portray different activities. The system had a library of minigames, because not all games are appropriate for communicating all activities.

This thread was picked up by Mike Treanor, who began to investigate what particular games could mean. For example, looking at ways that the game Kaboom! could be skinned with a range of meanings — creating political, child rearing, or religious meanings — but not every meaning. Treanor was investigating the "rhetorical affordances" of these relatively-simple games.

In parallel with this was Adam Smith's work (in which I also had limited involvement) inspired by humanistic game studies and the practices of game designers, looking at ways to computationally model game rules. He developed models that can be used for generation, reasoning, and discovery. His work focused on building a framework for specifying such models backed by Answer Set Programming and using that framework to enable previously-impossible projects.

As part of this research, Smith was using these representations to create an art game — a game that changed its rules every time it was played — discovering ways this work could be used generatively. Simultaneously, University of Washington researchers had a major project in educational gaming, with one of the goals being the generation of customized levels for individual learners, reflecting the understandings players had demonstrated (and failed to demonstrate) in their play thus far. But this turned out to be impossible when approached as applied research. Standard techniques, such as depth-first search, either took too long or failed at the task completely.

Luckily, Smith was personal friends with a PhD student at UW, and arranged to spend the summer working there with her. And it turned out that the techniques developed in our basic research at UCSC enabled fast generation of levels for the UW game that were guaranteed to be solvable by players, that matched player progress (demonstrated mathematical and spatial skills) to that point, with prescribed visual aesthetics in element placement, and with guarantees that they could not be solved without students employing the intended concepts. This combination of applied and basic research was so successful that we published a joint paper about it.

At the same time, Gillian Smith (a member of EIS and Jim Whitehead's lab) was creating systems that understood the rhythms of 2D platformer games. In one line of investigation, with contributions from Treanor, she used her approach to create the intelligent authoring tool Launchpad. (I was not part of this work, either.) Launchpad's AI could create a platformer level using different rhythms, based on human-authored tilesets, and with human-adjustable style parameters. Another line of Smith's investigation was in the automatic generation of levels. This resulted in the game Endless Web, in collaboration with EIS MFA student Alexei Othenin-Girard.

Endless Web was an experimental 2D adventure/platformer game in which the world is procedurally generated as the player moves through the game, shifting according to choices that the player makes. In Endless Web, players must search through the procedural rather than the physical space of the game in order to complete their objectives. A complete version of this game was created, but it never saw public release.

I also worked with Smith and EIS MS student Martin Jennings-Teats on the creation of Polymorph. This system was one of the first attempts to use game generation to customize the challenge level of a game — in our case, a 2D platformer — while it was being played (aka "dynamic difficulty adjustment"). It had a library of level segments (combinations of gaps to jump, coins to collect, enemies to defeat, etc) and we collected data on the difficulty of each. Then we monitored the player's behavior during play, modulating the difficulty of future level segments.

The next major piece of research in this area was publicly released: Game-o-Matic. This was led by Treanor, working with fellow EIS student Bryan Blackford, and in collaboration with Ian Bogost and his students at Georgia Tech. I was not involved in this project, but it is an important part of the story of EIS research in this area.

Game-o-Matic is a proposal, in the form of a system, for a radical vision of media creation: making media by specifying what it is about.

When using most video game creation tools, you start by specifying spaces, objects, and behaviors. Tools aimed at novices make this process easy — for example, making it possible to drag and drop common behaviors (collision, navigation) onto objects. But the process of making a game begins with constructing the elements that will eventually appear in the final product. Obviously, at some time before this, you need to have figured out how your ideas for the game's topic will translate into largely-spatial behaviors and structures.

In Game-o-Matic, instead, the first stage is to tell the system what play will be about. Then you never need to construct spaces and objects. The system itself produces a wide range of proposals for possible games that might communicate, through play, the situation you specify. You browse through them and tweak them — perhaps replacing their abstract shapes with specific images. This can be remarkably successful, but it can also result in nonsense. It's up to the human author to separate Game-o-Matic's successful design leaps from its total misunderstandings.

Our attempt to go further was led by EIS students Adam Summerville and Joseph C Osborn, with EIS post-doc Chris Martens, Sarah Harmon from Arnav Jhala's lab, and Jill Denner and her collaborators at ETR Associates.

It began with prototyping small minigames about climate change and looking for a way to represent them with more rigor than Game-o-Matic's ad hoc approach. We looked at the most popular approach (VGDL) and found some aspects of our hand-authored prototypes were impossible to represent, while others were awkward. So we developed a new game representation language called Cygnus. We also developed a framework (on top of the Phaser game engine) for automatically translating Cygnus games for play on web pages.

From there, we developed a system that can use the Cygnus representation to reason about games. This system, called Gemini, can automatically interpret games to understand, for example, what actions players are likely to take and what conclusions they are likely to draw about ideas presented in the game. The system can then be run in the "opposite direction" (because it is specified logically, using Answer Set Programming) to create candidate games to express a particular design intent.

Our publications include a journal article about how the Gemini system performs its automatic game interpretation, in IEEE Transactions on Games, and three years of publications at the Artificial Intelligence and Interactive Digital Entertainment conference — on Gemini's game generation (with Ben Samuel); on the evaluation approach we developed for Gemini, its results, and lessons for future system development (with Barrett Anderson); and on a game authoring system for naive users, built on top of the Gemini technology with funding from CITRIS (led by Max Kreminski and Melanie Dickinson).

Noah Wardrip-Fruin