Artificial
Intelligence Moderator's Report
I began two of my three sessions
with the same format. First I went over the format of the roundtable.
Then I went around the table asking each person to state who they were,
who they worked for, the CPU budget for AI on their projects, the number
of full time AI programmers, and optionally a topic for discussion.
Then I would start conversation on a topic. When a topic died out, I
would start a new one. We never got to every topic listed in any of
the sessions. Throughout I took notes on the board and occasionally
summarized points for these notes.
Trends: The maximum CPU numbers were up from last year. Turn
based games are reporting 50% of CPU. There were claims of higher values;
some for real-time games, as high as 60 to 90% but these numbers seem
inflated. The typical response is still stuck at "whatever does not
interfere with the frame rate."
The AI staff numbers were almost evenly split between 1/n of a person
and 1 full time dedicated programmer. A few responses were as high as
3 AI programmers.
Wednesday - May 6, 1998
This session was characterized by having far more newbies than experienced
AI people. This session had 41 people in it, and it was too large. Introductions
and topic collection took between a third and half of the available
time. Next year the sessions ought to perform some load balancing as
soon as they take up. The session did not go as smoothly as others have.
Multi Agent: The only discussion centered on blackboarding.
Optimization: The discussion split into turn based and real time.
Threads were mentioned, as well as throwing more memory at the problems
to gain speed. Interruptible and restartable methods were mentioned.
Avoiding Artificial Stupidity: TEST! And self test (AI vs. AI).
In particular the interactions with the simulation need to be studied
for boundary conditions and places where the AI does stupid or predictable
things.
World Abstractions: Keep lists.
Emotions Sim: This was the topic one of the members came to discuss.
The group sense was that not many people are doing it, but that the
same rules apply. Namely that data driven methods survive better than
fragile and complex scripting. The painful lesson was that the base
for abstraction in emotional modeling is not as well developed for games
as warfare. Warfare draws on board games and military simulations for
ways to model behavior. Unfortunately, the non-computer games basis
for emotional and character games seems to lack such a concrete basis.
Fantasy Role Playing games rarely control or specify emotional behaviors
in any concrete and numerical way. Those activities are left for the
players to supply, and the players tend not to write down the results
for later use in a database such as, "Thor was mildly pissed off (38%).
Not so much as last time (68%), and so he held his tongue." I moved
the group on to get one last topic covered before we ran out of time,
which may have been a mistake.
Formal AI Techniques: People said that they were lightly using
them. They mentioned rules engines, NN, GA, blackboarding, and Fuzzy
Logic.
Also in My notes: Put AI in early to the design. More than one
person reported that the game designer had made the game artificially
hard for the AI and a redesign had been required. Also, sports AI writers
get to use and re-use algorithms from previous games in the same genre.
They also get to use real-world plays.
-------topics below here were not covered ----------------
Speech generation and text to speech
Real Time pattern recognition
Combat
RTS
Thursday - May 7, 1998
This session had 24 people in it and it went perfectly.
Group Pathing: The knot at the bridge was the typical topic.
A* was mentioned. One person suggested that it resembled a fluid flow
problem and perhaps could be modeled that way.
Group AI: The level designer can help here. The goal is to get
a team instead of a mob. The game Gettysburg was mentioned.
Teamwork: Boiled down to analysis and paradigms.
Adaptive and Emergent: Noted as "interesting" and hard to actually do
well.
Character Influenced AI: "Data drives behavior." This needs support
from the world model. War games have decent models, but characters don't.
Rules based methods might work here.
--------Topics below here were not covered ------------
Desired tools
High level structures
Reducing code bloat
Where is AI going?
Forming and executing plans
NN
Why do people think AI stinks?
Adventure game AI
Teamwork.
Friday - May 8, 1998
When the time arrived, Steve and Eric had 15 attendees between them
so we added my 9 for the perfect 24-person size round table. After we
took up in one room, a small host of people joined us, but it still
went well. With three of us there, I took notes and Eric generally did
the moderating with the occasional assist from Steve or I when we saw
that it was someone's turn and Eric missed them in the crowd.
What is The Next Big Advance in AI: Everyone's first response
seemed to be, "Hmm. Good question!"
Away from Rules based
Learn from the Player - but there is the problem of the small training
set when modeling this way.
Dealing with intangibles
There won't be massive advances
System Modeling and Emergent Behavior: But don't let the sim
get in the way of entertainment. Figuring out how to beat the AI as
a puzzle game.
AI authoring systems: Increasing the standard knowledge base. Some of
this is already evidenced in the industry. A* has become well known,
for example, even if not everybody is using it.
Hardware for AI
It would be nice
Not as standard as 3D graphics: This lead to my best joke of
the conference, when I quipped, "DirectAI?"
Has to be able to beat the speed increases and multi CPU boxes of the
future. My own industry experience with custom speech processing hardware
suggests that to be viable, such hardware must be rather different from
a normal CPU so that it provides an extremely high added value to make
up for the fact that it will come to market later. In the time it takes
to design and field the hardware, Intel will have released a new chip
or MS will have multi-CPU support in game platform OS products.
AI Salaries
Is Deer Hunter a blip or a trend?
The very top is always worth lots.
Grow talent in-house.
Academic degree is not the kiss of death: But you had better do games
well.
Robotics stuff beats formal AI
Communications skills essential
Other Notes
Put AI in place EARLY
Programmers and writers working together
Scenario builders rather than script languages
Use paper game prototypes
Looking for people with 2 hats; AI and writers
Don't want programmer doing design, Subject Matter Expert does design
SME's to be taught object modeling
Shared vision needed with separation of skills and responsibilities.
