Moderators Report

AI Roundtables 2004

Moderator: Neil Kirby, Lucent Technologies

nak@lucent.com

 

Wednesday, March 24, 2004: General Topics

Longitudinal Data:

Full Time AI Programmers:

½

1

1.5

2

3

5

Number of responses:

4

7

2

2

1

1

%CPU

"tiny"

5%

10%

10-15%

20%

25-30%

40%

near infinite

Number of responses

1

3

4

3

1

1

1

1

We had 33 participants in this session. Our first topic was the goals of game AI. Most notably, it was noted that "good" AI might not always be the same as "fun" or "exciting." Since games are an entertainment product, "fun" would be the best kind of AI. So when an AI in an MMO figured out that it should kill the mages and healers first, it was retuned to be less efficient.

Once again there was a call for "middleware that doesn’t suck." Many of the techniques being used are reasonably well understood and people are tired of rewriting them. Programmers would enjoy having a configurable off-the-shelf AI.

Scripting was another topic of some depth. Scripting should be tunable, layered, and provide good debugging support.

We touched on cheating. The general consensus was that cheating should be avoided if possible, but always invisible and impossible for the player to prove.

Visible AI was seen as a requirement. Visible AI is one where the player can somehow appreciate what the AI is thinking about. Thief clearly showed how feedback to the player provided more value to the game play because the player knew that the AI was thinking. In general, more AI is good if the player gets to see it in action. Conversely, any AI that did not somehow become visible to the player was probably wasted effort.

Another growth area was friendly AI. Cooperative agents in the virtual space enhance the game play and have different issues than competitive agents.

Thursday, March 25, 2004: AI in RPG

Longitudinal Data:

Full Time AI Programmers:

1/8

½

1

1.5

Number of responses:

1

2

4

3

%CPU

<1%

10%

10-15%

25

50

100%+ or near infinite

Number of responses

1

1

3

1

1

2

We had 16 people in this session. We talked about a number of topics that have come up over the years and some newer ones. Smart terrain, 3-layer control structures (ala Craig Reynolds) and goal selection met up with path finding. The new twist was path finding influenced by goals and other altering factors. Similar to signs on freeway exits saying what was available and how far away it was causing people to change where they drive when they take trips. If there was something "close" along the path that met other goals, the pathfinder would alter the path to visit those things.

Simulations LOD got a lot of discussion. Fable got mentioned as we talked about how to effectively get lossless compression and expansion. For example, if the player walks into a tavern, a tavern has to be instantiated. The tavern keeper might have a daughter helping out, and she might have blonde hair. The player leaves and goes far away. The tavern gets "compressed" or more accurately said, simulated at lower level of detail, while the player is away. When the player returns, the tavern needs to be recreated with the same set of details, such as the blonde hair. If the player walks through the whole world, everything has been instantiated and there are storage issues to be dealt with if the illusion is to be maintained.

Player Character AI got mentioned and we very carefully asked to make sure that we had heard correctly. We had. For example, say a character in an MMO gets killed, possibly more than once in a row, by orcs. It would be reasonable for the character to have a fear of orcs, even if the player does not. This situation would be handled by the player character AI, which would influence how the character behaved. Explained that way, the concept made sense.

 

Friday, March 26, 2004: AI For Beginners

This session went well considering the surprising number of non-beginners who showed up. Some of them seemed disappointed that the session was not aimed at them, but most were very helpful in insuring that all of the questions were answered. I moderated from the center of the big circle with the ever helpful assist from Eric Dybsand who provided me with the equivalent of eyes in the back of my head. Steve Woodcock took notes; his write-up should be the most detailed. As I recall, we had over forty people in the room so I made it a point to walk around so that everyone felt like their questions were being addressed personally and so that I did not have my back to any one group for long.

 

Friday, March 26, 2004: AI Programmers Dinner

Since some people can not get enough time to talk about AI at GDC, there is the AI Programmers Dinner. This year fifty AI types enjoyed dinner and completely unmoderated, no holds barred conversations. Eulipia restaurant provided scrumptious food, but it was the people there that provided the real value. Pictures can be found at the www.gameai.com website.