1998 CGDC AI Roundtable Moderator's Report

by Steven Woodcock

Approach

The 1998 CGDC AI roundtable sessions were somewhat different in feel and overall "flavor" than the 1997 sessions. Whereas in 1997 we were having a lot of "how to" type discussions, this year our focus seemed split on getting more out of the techniques we had and trying to determine how much AI was enough for a given project. Pathfinding was discussed only in the context of faster 3D pathfinding (such as might be used in a typcial first person shooter).

Wednesday, May 6th

My first session of the conference had a total attendance of 25 people, a nearly perfect sized group for a good roundtable discussion. I began by asking everyone to briefly introduce themselves and solicited topics for discussion. By my estimate roughly 2/3 of the attendees were "professional" AI developers, with the remainder being either interested developers or producer/designer types wanting to understand the field better.

The foremost topic on the minds of most of the people present was that of building a "cheating AI"....when is it appropriate, and when is it bad? The general consensus is that while we all strive to build an AI that doesn't have to "cheat", it's okay to do so if it makes for a better game. Everybody was in agreement that the player must never suspect that the AI is cheating or the game will be ruined; this in turn led to a wide-ranging discussion of good ways to cheat that wouldn't be noticed.

The topic then moved on to a general discussion of resources available to the AI developer, since available CPU in large part guides what is possible and what isn't in the context of a game. Most developers claimed that they were getting anywhere from 1% - 5% of the CPU, a figure somewhat lower than last years' numbers. Out of the 25 people present, 9 were on projects with at least one dedicated AI developer.

The discussion then shifted to what levels of AI should be emphasized given a resource poor environment. Most of the attendees agreed that it made more sense to focus on making smart unit AI rather than smart strategic AI, since in most games players spent far more time watching small groups of units than thinking about "the big picture". Various strategies for making unit AI smart were discussed, from sensible pathfinding (the A* algorithm is now well known) to navigation of narrow terrain passages (the infamous Warcraft 2 peon problem) to conduct during combat (a unit should retreat when injured, not attack against overwhelming odds, etc.). There was general agreement that by focusing on these "up close and personal" areas of AI that the player would think more highly of the game as a whole.

Our final topic was a general discussion of what constituted the "state of the art" in game AI today. Nobody in the room was using any of the more exotic AI technologies such as neural networks or fuzzy logic. Two developers were making use of genetic algorithms to "breed" smarter AIs offline during development, but they weren't yet sure of the success of the experiment. Most developers are sticking with tried and true methods of finite or fuzzy state machines combined with rules-based approaches.

This second session had rather light attendance in my room, with only 10 people present at the outset and two more trickling in during the course of the discussion (Eric and Neil report similarly low numbers on the second day). Six of the 10 were dedicated developers, and most of the conversation was between these attendees....the other people did not contribute much to the conversation despite prompting. The developers estimated that they received anywhere from 1% - 10% of the CPU.

Despite the low turnout, however, we discussed a wider range of topics than on the first day. The first topic of discussion was believable group behavior-that is, methods of handling larger groups of units in a fashion that both made sense and which would be viewed as "organic". We discussed using various A-Life techniques such as flocking and swarming for movement, which in turn led to a side discussion on dynamically selecting groups of units from larger disorganized masses.

How to ensure cooperation between units or groups of units towards a common goal followed from this discussion and led to quite a debate about which methods were most appropriate. Several developers present had worked on sports games and stated that they'd simply used "playbooks" of pre-scripted behaviors to get good coordination between units, while others favored a more dynamic, goals-based approach using variations of blackboard technologies. There was little consensus as to which provided more realistic behavior, however.

The conversation then turned to a discussion of scalable AI-building an AI engine in such a fashion as to ensure reasonable behavior across a variety of platforms (console, low-end PC, high-end PC). Everyone agreed it was a very difficult problem given the state of modern games (which maximize CPU utilization on whatever machine they're running on) and nobody seemed to have satisfactory solutions. One gentlemen from the Department of Defense world suggested that portions of the AI processing could be "farmed out" from lower- end PC to more powerful PC if a network game was being played, but indicated that there could be severe communications problems in an environment such as the Internet that could make such an approach shaky at best.

The final topic concerned learning AIs. Most of those present had tried at one point or another to build an AI that actually learned the preferred moves and strategies of the player, but all had abandoned the effort due to unpredictable results. AIs that can modify themselves can also get themselves into a "stupid state", which is an absolute disaster game-wise. There was also a spirited debate about whether building a learning AI was even desirable...many developers felt that player really only want a "hard fought win", not the possibility of actually losing. Three developers indicated that they were in fact attempting to build learning AIs into their current projects, and promised to report on the results at next years' CGDC.

At that point we ran out of time.

This session was interesting because all three moderators felt that the per room attendance was too low to have effective roundtable discussions (my room and Eric's had 9 people each; Neil had 5). We therefore decided to combine the three roundtables into one larger group, which turned out to be a very good ideas as it led to our best roundtable of the entire convention. People drifted in over the course of the discussion to eventually raise total attendance to 40. Normally this would have been too many for a good discussion, but with three moderators handling things it worked out rather well. Eric and I did most of the moderation while Neil was kept busy taking notes on the whiteboard. The session was very balanced between experienced and inexperienced AI developers, with a couple of producer/design types thrown in for good measure.

The first topic was "Where is the next advance in game AI coming from?", and that led to some interesting opinions. Many present felt that it was difficult telling where the last advances had come from, much less where the industry was heading. Most felt that the real-time strategy game glut of late '97 had led to some interesting advances (Dark Reign and Age of Empires were both spoken of fondly), but that generally advances in the field had been incremental rather than evolutionary. The A-Life approach used so successfully in Creatures was viewed more as an interesting anomaly than the beginning of a trend. There is much interest in the AI being promised in several of the upcoming first- and third-person shooters, a genre notorious for its brain-dead monsters. Everyone felt that progress would continue to be a step at a time, with no major changes to the state of the art as a whole.

This discussion of where the industry is heading in turn led to the subject of AI programmer salaries coming up, and the impact of games like Deer Hunter on future development. (Most everybody present felt that Deer Hunter was something of an anomaly-out of the 40 people present, there were only two people (myself included) who actually hunted.) While no specific numbers were discussed, it was agreed that a.) top talent was worth paying for and b.) formal degrees were far less important that demonstrated talent. Knowledge of the AI normally associated with the field of robotics was held to be more important than formal academic training.

There was a brief discussion as to whether or not dedicated AI hardware, rather like dedicated 3D accelerator cards, would be useful developments in the gaming industry. While most present thought that there were some functions which could logically be hardware based (such as the ubiquitous A* algorithm), most felt that faster CPUs would prove to be more reliable and cheaper solutions overall. There was widespread agreement that too much CPU horsepower is being given over to cutting edge graphics, not surprising considering the audience.

The last major topic concerned bringing the AI into the design process of a game early on. Too many games are still being developed without much consideration to the implementation of the computer opponents until very late in the production cycle, mere weeks from shipping. While this is in some part driven by the realities of the environment (it's difficult to build an AI if the game infrastructure doesn't exist yet), most felt that by working together from the beginning the AI developer and the game's producer could build a better, more stable product. Things worked better when a developer was also well versed in something less technical, like writing, and/or if a producer had some programming experience under their belt.