Research Areas The Design Reasoning Lab follows a use-inspired basic research philosophy. We’d like to reflect in our software systems a broad theory of design (often focused on game design), but we’ll always do it in a way that is also shaped to be of practical use on the 1-2 year timescale. We work with in-the-wild and in-development games that have either already reached audiences in the millions or are being built on platforms that reach similar audiences.
Procedural content generation (PCG) is the trendy term for one slice of the broader category of generative methods. Our research in generative methods yields new methods (such as those based on constraint solving and machine learning) and associated modes of control (declarative rules and artist-crafted example targets). At the same time, it offers new interpretations of what we are doing when we use PCG (perhaps we are writing down design space models) or what we are searching for when we apply constraint and optimization techniques (implementing approximate inference in a probabilistic formulation of the design process where forms lead to functions and functions lead to subjective interpretations of appropriateness to contexts).
Our work in machine playtesting is concerned with automatically assessing properties of videogames at the level of play experience rather than the level of their hardware/software implementation. What can the player do? Do they always have to do what I think to achieve some goal? A machine’s answers to these questions should come quickly enough to influence the design process, be expressed in a familiar enough notation (e.g. a heatmap rather than a logical proof), and be available in ongoing design projects without too much additional engineering effort. We believe the regularizing tendencies of design-for-test (that you tend to produce better designs when you match a design goal at the same time as being efficiently testable) represent a significant but oft overlooked benefit of testing in the design process. Our work in machine playtesting aims to make the space of play for a game (under relevant constraints and speculative assumptions) visible so that we can see how the shape of that space responds to changes in our design.
In the history of AI, automatically playing games is a perennial challenge taken on with the idea that games provide well-defined test-beds for general intelligence. Technical methods for success in some game, it is hoped, will transfer to real-world applications. We see specific games as messy technical and artistic artifacts with complex relations to other games and broader culture – navigating them and answering questions about them are already real world applications. Our lab tries to produce automated gameplay methods that, rather than embodying any form of general intelligence, can be bolted onto in-the-wild games with ill-defined notions of success and still yield interesting coverage of their spaces of play. If there is a general theory behind our methods, it is one that is trying to explain why we like to go down the slides in a playground rather than one that explains why we make a certain opening move in Chess. We don’t think you need to have a score function or win state to define optimal play. We think the answer is the topology of the space of play, and one of the biggest challenges is even figuring out what that space looks like.
Information Retrieval for Interactive Media
The way we search for games, apps, or other interactive media is analogous to pre-Google web search engines. We trust publishers to tag their work with all and only the appropriate metadata, and we imagine we can group all of the relevant content into stable categories and lists. Our lab considers the problem of automatically crawling to discover new games and new possibilities for play within then, we consider how to index these spaces to accelerate useful queries, and we consider how to serve up results more useful than linear result lists. As more and more of the web becomes dynamically generated and each page adds more interactive elements, the situation begins to resemble an ever-growing app market more than a finite collection of texts. Our research puts automated gameplay to to work in mapping the space of play and integrates computer vision and natural language processing techniques to index the interactive moments uncovered.