Neil Kirby, Eric Dybsand, and myself have had the honor of moderating a series of AI roundtables at the Game Developer’s Conferences for the past six years. These roundtables grew out of a sense of frustration at the 1996 GDC when we found ourselves shut out of AI session after AI session. Further, the sessions we could get into were packed to standing-room only capacity and just plain uncomfortable. Knowing that we could do better and that the field seemed ripe for greater levels of discussion, we submitted a proposal to the GDC folks for a series of three simultaneous roundtables. The idea was that with these sessions we would both accommodate the massive numbers of people who seemed to want to talk about the topic and be able to focus the sessions on problems developers were facing day to day. GDC agreed, and the GDC AI Roundtables were born.

This year’s sessions followed essentially the same format as the 2001 roundtables did. Day One focused on general AI topics—whatever the attendees wanted to discuss. Day Two broke up the roundtables into specific topical areas. My session focused on wargames and strategy games, while Eric’s was aimed at RPGs and Neil’s discussed FPS games of all types. This devotion of sessions to specific arenas of game development were intended to help satisfy attendees who had questions or comments specific to a current project, which might otherwise not come up in the more general sessions.

Day Three changed the format slightly with a combined roundtable we dubbed "AI for Newbies". In this session we attempted to address the needs of developers new to the field of game AI who often didn’t know where to start. We also tried to reach out to artists, producers, and others involved in the game development process who were interested in the field of game AI but who really didn’t know much about it. Our hope with this session was that we would be able to both help developers new to the field and aid in bridging the "understanding gap" that can crop up between AI developers and others on their team. Several more experience AI developers assisted in answering questions and providing examples of why a particular technology or problem was good or bad based on their real world experiences.

Overall the sessions went very well and seemed to be quite popular. The following are notes from the sessions that I moderated. Neil and Eric's moderator’s reports can be found online wherever the reader found this report.

We kicked off the first session with a bang, all things considered. The session was well attended with 24 people representing a wide variety of experience, from folks who had worked on big titles such as Empire Earth, Kohan and MOO3 to writers (such as myself) who had done articles for the Game Programming series and AI Game Programming Wisdom.

As in previous years I asked the attendees about their resources for game AI and the number of developers per project. Twenty of the 24 attendees were actual game AI developers, two of the remainder represented the academic community, and at least one was there purely to learn about techniques that could be useful in military simulations. We also had representatives from two different AI middleware/SDK vendors (more on that later), which sparked some discussion later in the roundtable.

Average CPU utilization was somewhat down from previous years, though everybody thought they had plenty of resources::

Every developer reported that they had plenty of programmers dedicated to working game AI:

A brief discussion of why there seemed to be fewer developers with lesser CPU percentages ensued. For the most part developers seemed to feel that overall CPUs and graphics cards were incrementally faster than in previous years and that they thus had more cycles than previously. The number of programmers dedicated to game AI was roughly the same as in recent years, though it should be noted that the turn-based strategy game folks reported both more available CPU (100%) and more dedicated developers (anywhere from 1 to 2) than other developers did.

The roundtable covered a huge variety of topics:

• Technologies in Use – If anybody was using anything "exotic" they just weren’t talking about it. Every developer present said they were using straightforward Finite State Machines (FSMs) and Fuzzy State Machines (FuSMs) to build their AIs. These were combined with scripting to handle missions and the general personalities of the computer opponents. A couple of developers had been considering using Neural Networks (NNs) for racing games to control the computer cars but that didn’t seem to go over well with other developers who had already dabbled in them.
• Layered AI – Surprisingly most of the developers seemed to be using (or have built) layered AIs in their products, usually breaking out decision making along strategic and tactical levels. The strategic-level AI seemed to get the broader problems—deciding general strategies, performing map analysis, setting goals, and managing resources. The tactical-level AI got the "grunt" jobs—pathfinding, handling formations, and fighting. It was interesting that some developers thought that computer opponent behaviors were a strategic-level function while others pushed it down to the tactical AIs. The former argued that it affected goals and resource management while the latter said it mattered more in formations and combat. It was interesting to note that while most developers used two levels when they layered their AI, the developers of Empire Earth spoke of using up to seven (!) layers for their game.
• Terrain Analysis - Most games were doing terrain analysis in one fashion or another, though few were spending enormous amounts of CPU on it. Terrain analysis helped most strategic-level AIs in identifying locations of resources or chokepoints on the map. A few developers were also analyzing maps for connectivity and building portals or waypoints for later use by the pathfinder algorithms. One game (C&C Renegade) used a realtime analysis capability to allow that game to learn from the player and create pathfinding nodes into the list of valid paths when the player found some way to do something not otherwise spelled out (such as jumping out a window).
• Pathfinding – A perennial favorite which didn’t spark too much controversy this time around. The discussion was less about the ways in which one found the best or shortest path and more about ways to move units intelligently. Most developers seemed to use some kind of terrain analysis to either identify good paths prior to the scenario or to update valid paths as the game progressed. One developer noted that their game was continually doing a low-level terrain analysis, feeding the revised data to the AIs. There was general agreement that all games should be designed to use hierarchical pathfinding, with one pathfinding for high-level paths and one for lower-level obstacle avoidance issues (both Starcraft and Age of Kings do this, for example). Everybody was interested in ways to minimize the CPU hit of pathfinding and do it only when absolutely necessary; at least one developer of an unamed racing game cited pathfinding as his second biggest CPU hit (the first being collision detection).
• Formations – A natural followup to the pathfinding discussion, it was interesting (to this moderator) how many different approaches developers had to this problem. Many games require units to move in some kind of formation, but this desire is greatly complicated by obstacles and other units. It can also be expensive in terms of CPU utilization, since the naïve approach is to invoke the pathfinder for each unit being moved—often a big hit on available resources. Kohan’s developers noted that they found one path, then gave all the other units in the formation a slightly offset path that roughly corresponded to the formation desired. Individual units recognized when they were out of position or lagging, often cheating to get back into position (speeding up slightly was one option presented). Some games used variations of flocking to select a "leader" for the rest of the group to follow (Kohan) while others found that in their cases the leader often got blocked by his followers and they had to do individual unit pathfinding to get around obstacles (Warcraft III and Starcraft).
• Learning – This topic generated a lot of discussion, if not a lot of optimism. It was generally agreed that learning counted for a lot when it was done right, but most developers reported that they didn’t bother with it outside of limited use during development proper. The C&C Renegade example cited above was one limited example of realtime learning but that was used in a very specific, very special case (i.e., to add pathfinding nodes). Teaching the AI to learn from the player or over the course of many games just wasn’t something anybody was bothering to do. Most seemed to think it would be dandy to have an AI that could ratchet itself up to meet the ability of the player but nobody had a clear idea on how to keep the AI from learning something dumb (just like a player). Some developers used limited learning during development and playtesting, generally along the lines of tweaking existing parameters over the course of hundreds of runs. These parameters were always locked down before the game shipped however.
• Planning/Scheduling – Many of the attendees said that their AIs handled planning and scheduling of resources in some fashion (usually in the strategic-level AI). Most were content with the strategic AI setting high-level goals ("go there") and letting the units work out the details on a cycle-by-cycle basis, since plans were going to change a lot anyway in reaction to changes in the environment. Decision Inertia (an AI tending to stick with a given plan) was encouraged in various ways, usually through the weighting of a selected plan over others in future consideration cycles. Influence Maps (IMs), a topic much discussed at previous roundtables, were much used by strategic planners to determine where to build buildings, find minerals, or deploy troops. A few developers said that they used basic "personality scripts" to get particular AIs started during the first few cycles and then let the game’s strategic AI take over from there.
• Cheating – As usual cheating was a major topic. Most developers were quite comfortable with the idea of cheating, however, and it didn’t bother most of them. If an AI wasn’t too good handling the "fog of war", for example, developers often let the AI overlook it for resources and such. Others let their AIs respect fog of war but gave extra weight to plans that uncovered new areas or which revisited areas not seen in some time by the AI (Kohan does this). At higher difficulty levels many AIs shared their information at the strategic level. Everybody agreed that the precise ways in which the AIs should cheat was basically a design decision and, if it were going to be blatant, the player had to know. Keeping cheating a secret never works, and just makes the AI look dumb when somebody figures it out.
• Difficulty Levels – The discussion of cheating quickly crossed over into the issue of difficulty levels. How did developers handle adjusting game difficulty levels—Easy, Medium, Hard—when the player selected them? The choices seemed to come down to a.) giving smarter AIs more resources at game start (very common), b.) letting smarter AIs view more information (ignoring fog of war, etc.), and c.) letting smarter AIs have more time to do their processing. Kohan, for example, gives more CPU time to the AIs at the higher difficulty levels so they can consider more options more deeply. A few developers noted that they were tending away from multiple levels of difficulty in future designs as their feedback indicated that there were only two modes players really cared about for solitaire play—easy or hard.

This list definitely is not doing justice to the wide variety of topics discussed but it was difficult for this moderator to jot everything down! The session ran a few minutes long with the discussion going strong on several fronts. It was difficult to keep up with things at times!

The sessions on the second day focused in on specific areas of game AI. My session covered RTS games, wargames, and strategy games in general.

The room was overflowing with 37 people attending and we started a few minutes earlier as a result. We started quickly and as the roundtable progressed there seemed to be two main themes of conversation—implementation of AIs that could handle diplomacy and implementation of personalities.

Diplomatic AIs and the problems of building them were something not often covered in previous roundtables. Engaging the player in a believable "give and take" of positions and moves that one gets when playing with many humans is something that current AIs don’t do very well. This was of particular interest to some folks present who were very interested in building computer opponents for the old Avalon Hill boardgame Diplomacy—it depends heavily on how you can convince your opponents to support your moves towards winning. Many topics were touched upon:

• Behavior – One developer noted that people act differently in real life than they do in games. In real life people tend to follow a leader; in games they tend to gang up on him. That made some aspects of diplomacy awfully fleeting, especially when it was further noted that many multiplayer games start out with "let’s gang up on the AIs" thinking.
• Communication – Games right now aren’t really doing a very good job of providing ways for the AI to communicate with the player, or vice versa. Civ III, for example, allows the player to view basic diplomatic relations on a big chart but gives him limited tools to make proposals. Much of the art of diplomacy lies in the art of the "give and take" of negotiation between players, whereas games tend to only provide the most bare bones "take it or leave it" options. Building a strategic AI that recognized and could communicate effectively the "coincidence of interests" between itself and a player was very difficult. It was further noted that in some games (such as Diplomacy) the move itself is a form of communication—England moving an Army to Norway almost certainly signals a northern strategy which both Russia and Germany might view with alarm. Getting this kind of coordination working between AIs, or AIs and the player, is very difficult. Some games (such as Kohan: Ahriman’s Gift) have crude facilities for the player to "tell" an AI something but it’s difficult for him to judge how the AI responds to it.
• Coordination – Coordination between players and the AIs was also a major topic. AIs have a hard enough time getting their own units to work together, much less interfacing with the players. Some games try to work around this by letting allies see each other’s units (such as in Age of Kings) while others even let players take over their units completely. Age of Kings and a couple of others let one player "draw" on the strategic map to indicate proposed goals or attack sites and most developers thought that getting the AI to do this itself would be a big step.I also noted that computer AIs that did allow communication didn’t have any good to ways to communicate bungled communication. In Diplomacy, for example, a player can plausibly lie about misunderstanding an order ("Oh, did you mean support Norway? I wrote attack…sorry!") and get away with it, all in support of his strategic ends. Computer AIs barely handle truth gracefully, much less lies.
• Goals – A big chunk of communication and coordination deals with deciding on goals. One developer pointed out that there are few games that provide facilities for a diplomatic AI to work with the player to set a goal for their alliance, such as "you attack there and I’ll defend our side". Some discussion on the merits of Nash’s Theorem (a gameplay theorem for optimizing moves based on perception of diplomatic moves with imperfect communication) ensued between the academic attendees and the Diplomacy researchers but it was clear that this was far beyond the problem most developers were worried about.

The other major topic was implementation of computer AI personalities, making the AIs seem "real and human" by having them act in a consistent fashion. Nearly all of the developers were already doing this in one way or another and thought that a wide variety contributed greatly to replayability. The discussions here tended to focus more on neat tricks and enhancements than anything else:

• Implementation – Everybody did things in a similar fashion though using different personality variables. The Stars Supernova folks had perhaps the most detailed implementation discussed, with a five-layer AI using blackboards to post strategic directives that were the fulfilled by tactical "administrator" AIs according to their individual personalities and whims. As with the day before developers noted that they always developed multiple strategic options, then chose the ones that best suited the personality in question. Where two or more options were even, points were given for options that continued previous strategies, or AIs randomly chose from the most plausible options in order to not be predicable.
• Psychology – Not something often seen outside of the academic world, but a field which some developers said they’d looked at. Psychology breaks out personality types into 5 main areas: Openness, Conscientiousness, Extroversion, Apathy, and Neurotisism (the "OCEAN" theory). Some developers said they’d used this breakdown as the basis for the personalities of their AIs, giving them numerical ranges in the various measures and then using those to define actions and goals. This was most often done along a racial basis (i.e., Reptiles were aggressive while Human were more general) mostly for ease of categorization on the part of the player.
• Expectation – There was some discussion of design consideration to meet the expectations of the player. Personalities had to make plausible actions that fit their personality but make them differently each time so as to avoid being predictable. There was some opinion that anything that didn’t directly map to the player’s perception was a waste of time—if he didn’t see it and let it guide his playing style why bother with it? The issue of expectations was also influenced by those games that fed personality into reputation, particularly in RPGs and wargames. If an AI does volatile seemingly random things, it can develop an in-game reputation for being unpredictable. This reputation in turn influenced the kinds of deals it offered or might consider.

As with the previous day’s sessions we were going strong when we hit the end of our time, and it was clear that there were many other issues that could be discussed. Some of these spilled over into a later Strategy AI Roundtable the next day.

Day Three was another "AI for Newbies" session, following up on the popular format we began in 1999. Neil, Eric, and myself combined our separate AI roundtables into one large session devoted to helping folks new to game AI get answers to their questions. The session had right at 42 attendees—up from the similar 2001 session, but in line with the convention averages. The room seemed to have just the right amount of people for a good "give and take" session.

In order to help facilitate the discussion we had put together a handout with pointers to references on a variety of the most common game AI topics. This seemed to go over fairly well with the attendees sparked some of the questions from attendees..

Neil and Eric moderated while I took notes on the whiteboard (thanks to the GDC staff for providing that!). What follows are the questions I got onto paper together with the best summation of our answers (and related discussion) I could get down:

• Question #1: What can’t you do in H/W that you want to do?

• Question #2: What other kinds of AI technologies can games use besides Finite State Machines (FSMs)?

• Question #3: What about those Neural Networks; does anybody use them and train them in realtime?

• Question #4: What is the purpose of game AI?

• Question #5: What amount of randomness is right?

• Question #6: What languages to people use for their AIs? Is LISP any good?

• Question #7: Is a design in which the AI machine is totally independent of the world machine a good one?

• Question #8: What about scalability?

• Question #9: Are there any racing games doing good AI?

• Question #10: Are there any games with good AI? What technologies are they using?

• Question #11: What about middleware?

• Question #12: How can we make AIs look smarter?

• Question #13: Where are we going to be in five years?