CGDC '97 Roundtable Report

Artificial Intelligence in Computer Games
Report #2

In each session I asked how much CPU the AI in games were allotted. I also asked how many full time staff were writing AI code.

Sunday, 27 April, 1987

This session was characterized by having far more newbies than experienced AI people.

Percent CPU: CPU numbers listed were 2% and 10-20%.

Autonomous Agents: The thrust here was opponents that felt "analog" and not "digital." Notably that having a smooth range of behaviors felt more realistic.

"Real AI": Artificial opponents that could play by themselves.

Scripts & Synthetic AI: The discussion was about granularity of controls. And about how fragile a script could be.

Thinking and Emoting computers: Learning so as to improve decisions by modifying the knowledge base. This is not an easy thing and no one offered examples of its use in commercial games.

Natural Language Processing: Both input and output NLP were discussed. Nearly all of the group had extremely little experience with this topic. Those that had quickly disabused the rest any notion that this is an easy thing to do.

AI Planning: no notes

Cheating: This annual topic generated good discussion. A new twist on it was that sometimes the entertainment experience is better if the AI "cheats" by making realistic looking artificial mistakes. This is especially true in sports games where the program knows a priori what the player plans and intends. A good player should be able to fake out the computer AI, even if the computer AI does have faster "reflexes" than the player. The current take on cheating in general is that the state of the art is good enough that it usually can be avoided, but it is held in reserve for careful use. One such careful use is stopping bug-exploitation. When a user finds an offense that the AI has only a 1:20 chance of defending, the user may use this offense over and over. To stop the player from racking up massive scores, the AI cheats and pulls the correct defense whenever it sees exploitive behavior. This is realistic behavior, thus providing a better entertainment experience, but the AI had to cheat to do it.

Genetic Algorithms and Genetic Programming: Some work is being done here. The areas of use were planning decisions and creating music solos. Music solos in a jazz context proved quite amenable to GA/GP.

Monday, 28 April, 1997

Percent CPU: CPU numbers listed included 20% for a console sports game. Console games benefit from having good graphics throughput. The other figure mentioned was 1/n. This was a multiplayer (and multi-AI) game. Each player or AI got an equal share of CPU.

Natural Language Processing: We were strong on people with NLP experience. Speech processing is among my own areas of expertise so I was among the speakers. The consensus was "model it the same as anything else." NLP requires substantial semantic as well as syntactic processing for both input and output. We also talked about Text-To-Speech. I mentioned the TTS "slit-your-wrist-test." The measurement is to time how long users can stand to hear TTS before they slit their wrists to put themselves out of their misery. (Larger numbers are better). I moved us to the next topic before discussion was over to allow other topics that might be thought of as closer to AI to have some time.

Synthetic Characters: The two points that came out of this were thought trees and semantic processing.

AI for Story Control: Reacting to player input. While everyone would like a Holy Grail, for this, currently none exists. Chris Crawford's Erasmatron got mentioned, but the jury is still out.

AI for Cinematics: This area showed great potential for markedly increasing the entertainment value of games using existing technology. People with Hollywood experience had useful input that was probably new to game developers. Notably was the idea of "where's the focus?" Tomb Raider and others showed that there is some gold to be mined here. The other notable idea was the use of one-shot triggers and cut scenes. For example, the first time the player turns a corner, the camera cuts to a long shot showing the player entering a long corridor from the far end of that corridor. Or upon climbing a ladder, the view changes to a 360 degree pan around the character showing them and the stunning view from up here. These cut scenes only happen one time (to avoid boredom), and only if the game is not about to threaten the player (to keep them from being unfair).

Real-Time Collision Avoidance: This one boiled down to anticipation and allowing realistic effects. The AI in a sports game should allow a skilled player to "fake it out." It should also have collisions when some entity deviates from predictable behavior.Non-Human AI: This lively discussion was best described as, "What does the cat want?" Conclusions came down to, "New goals, same algorithms."

Individual vs. Society: Or group behavior modeling. The first point was that "invisible means not needed." Schroedinger's cat came up. If no one sees it, why do it? The enemy crawling along a trench pops up and fires. If all the user sees is the pop up and fire, why model the crawl? The instant response was that at any time the box is opened, we have to know what the cat did. And so if artillery is rained down upon that trench, the crawl must be modeled. Sports games with multi-angle instant replay require that everything be modeled at all times because the box is never closed. The final point was that, "close enough is good enough." If a group can be accurately modeled, there might not need to be a need to model all of its aggregate elements.

World Modeling and knowledge representation

Character development

What do game writer's want?

Items below were generated as we went

Tuesday, 29 April, 1997

Percent CPU: CPU numbers listed blew earlier amounts off the chart. There was a strong showing of 25-30%. These numbers could be inflated and they sounded like shots in the dark. The 30% number comes from a turn based game. The real curve breaker was a game server with 45%. Having no user interface code, the server has far more cycles free for AI. The number of dedicated AI staff numbers were 0.5, 1.5, and 2.5 staff normalized over the life of the project.

Strategy Games: This was the context for most of the discussion in other topics.

Learning & Adapting AI: This takes tracking the success or failure of tactics (in other words, tons of data to go through). It also takes recognizing success or failure when it happens. People are trying Neural Nets to help with the recognition task. War being chaotic, these nets are trying to deal with a chaotic situation full of noise and it's no wonder that it is a hard problem. Even if all of that were solved, closing the feedback loop is inherently dangerous. An AI that learns the wrong things will act as many a nation has that takes the wrong lesson from history. A perfectly decent-but-not-brilliant rules based system can easily be made less effective this way.

Laboratory and Industry Stuff: The only useful thing mentioned was a hybrid system. It used three parts. The first was programmed heuristics similar to classical game AI. Second was a rules based expert system. And third was a neural net. The systems were polled, in order. If the heuristic felt it had a good answer, it was always preferred. If it came up unsure, the expert system was consulted. If both other systems were stymied, the neural net usually gave a decent response to an ambiguous situation.

Player aid AI: It was the case for a good AI to relieve the tedium of micro-management. This was to provide ease of use in the game. It needs to have good player control so that the player is confident that things will go well. Also mentioned were the idea of being able to provide stim->response scripts and settings.

Multi-agent Planning and Communication: Blackboarding came up as a way of dealing with this. The fragility of planning amidst chaos and the ability to react were mentioned as well. The ability to fix broken plans was a clear need.

Mapping Personality onto Good AI: Given a good AI, how do you give it personality? The user is looking for a social experience. One way is to modify a rules based system. Everything depends on game design, but the idea is to avoid a total rebuild.

Group vs. Individual: Multilevel AI came up as a way to deal with the problems here. Grouping units was the minimalistic way of dealing with the problem.

Realistic Human Replacement

Creatures and Emergent Behavior

Items below were generated as we went

Avoiding Artificial Stupidity: The goal was to place a minimum floor that has no holes.

Serial Execution of Parallel Unit Actions: There is a problem when planning one unit in isolation when many units need to be planned as a group. There is difficulty in avoiding N squared explosion of complexity.

SHIP IT NOW! SYNDROME: Is AI ever the critical path? When was the last time playtest and AI refinement stopped shipment of an otherwise finished game? Will we ever get time to do it right?

Good Abstractions Are Hard To Find: But everybody is looking hard for them.