Back to the Game AI SIG

Toward Standardization of AI Interfaces
Report on the Roundtable Discussion of the 2002 Game Developers Conference

Moderated by Alexander Nareyek

(San Jose, CA, USA, March 23, 2002)

(Please follow this link for the call for participation for the meeting)


The roundtable was attended by 23 participants - a healthy mix of mostly game programmers, tool producers, and academics. Even people from the military showed up as they are pursuing similar approaches.
After an introduction of the topic presented by me, a number of people gave initial kick-off statements, namely Bryan Stout, Paul Kruszewski, Steven Woodcock and Jozsef Toth. Advantages of potential standards were mentioned, such as risk reduction by using third-party tools and being enabled to do rapid prototyping, focusing on creative tasks instead of having to re-write A* for the 100th time, and providing some guidance for academic research. On the downside, producers may fancy reducing a game's AI production schedule to the basics of "proven" functionality, and a company becomes dependent on third-party developers if they use their tools that implement such standard interfaces.
Many issues that need to be addressed for standards were mentioned, such as what kind of sensing/acting possibilities an AI-guided character has, whether the AI is called via triggers or has a regular heartbeat, whether the AI is given a time threshold until which it has to end its computation, if sensing information is pushed to the AI or the AI has to pull/query it, and how the objects/world are represented in general.
The discussion part of the roundtable was planned to start with a discussion of the pros and cons of standards. However, I was faced with a very encouraging, uniform hand-waiving "We know it won't be easy, we want it, let's do it!" and we moved on quickly.
One issue that popped up at various points of the discussion was the level at which standards should be established. One option is to provide only a very general service infrastructure interface, by which AI tools can register and handle their services. Another option would be a general AI character/agent interface with sensing/acting sub-interfaces. Or - which seems most promising as a first step and without sacrificing performance too much - establishing interfaces for a set of specialized functions, among which pathfinding, steering, influence mapping, decision trees, finite state machines, rule-based systems, goal-oriented action planning, and resource handling were considered to be most important and manageable.
We started with discussing standardization options for pathfinding in order to dig deeper into related problematics. Pathfinding - even if we restrict ourselves to conventional single-unit movement - is a group of different techniques, and computer games often make use of many methods, e.g., different versions for long, medium and short distances. Again, the discussion about the level began, i.e., should interfaces for very basic routines be provided, like for handling open/closed lists for A*, or interfaces for complete path-planning procedures? There are arguments for both, and indeed, the solutions don't exclude each other.
Another very important aspect (not only for pathfinding interfaces) was identified as how the game world is modeled, how world information is passed to the AI procedures, and in which formats it should be returned. Depending on the world's structure, different interfaces may be necessary. Thus, a close cooperation with people building world physics is important.
The overall discussion was very energetic, and many topics that were on my list to be discussed had to be omitted due to time shortage. For example, we were not able to discuss how such interfaces should generally look like with respect to implementation issues. Throughout the discussion, other standardization initiatives like DirectX and OpenGL were often mentioned, and we definitely have to take a close look at these.
The time available for discussing topics like potential hardware, such as AI accelerator cards, was limited. A kind of standards are a prerequisite for such developments. However, it's a complicated process to get this going, as the spread of installed hardware and the number of games that support this heavily depend on each other. The console area might be a domain where we may see AI hardware developments first.
Institutions in other contexts (such as the military) are interested in establishing AI standards as well. Joint work with them is probably useful, however, one should keep in mind that the goals are sometimes different, and the solutions should not be enforced to be the same. For example, for academia and military, it is probably not appropriate to allow for cheating options as well as their target hardware platforms' capabilities may be drastically different.
The roundtable was very successful and an excellent starting point toward establishing standards for game AI interfaces. We've clarified directions for a potential standards initiative, identified potential areas, and touched many important related topics. Further developments will be coordinated by the IGDA and working groups will be set up soon.


Last update:
April 18, 2002 by Alexander Nareyek