Newsgroups: comp.ai.games,comp.ai.neural-nets,comp.ai.alife
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!gatech!newsfeed.internetmci.com!in1.uu.net!world!eshepard
From: eshepard@world.std.com (Eliot Shepard)
Subject: Realtime AI design
Message-ID: <Do8zDJ.7KE@world.std.com>
Organization: The World @ Software Tool & Die
X-Newsreader: TIN [version 1.2 PL2]
Date: Thu, 14 Mar 1996 07:56:07 GMT
Lines: 43
Xref: glinda.oz.cs.cmu.edu comp.ai.games:4075 comp.ai.neural-nets:30474 comp.ai.alife:5269

I'm in the process of designing an AI for competitors in a realtime
3D game which is similar to one-on-one (or two-on-two) basketball. I
have some questions on the implementation of "personality" in the
decision engine, and on how the sense filters should work. Here's a
black-box description of my design: 

inputs:
o world state (ball position, velocity, possesion, goal proximity, etc.)
o senses (sight, hearing)
o competitor state (position, velocity, fatigue, etc.)
o competitor personality (long-range shooter, aggressive defense, et. al.)

outputs:
o steering
o power
o shoot/pass

Other requirements for the engine include quick updates (fewer than
4e5 x86 instructions/update) and smoothly parameterizable
hearing/sight acuity and overall difficulty.

Ideally, I'd like to train a series of competitors with different
skills and behavioral patterns, and plug in their parameters at run-time.
Neural nets seem promising, but I've never implemented one, and I have
questions about their performance.

questions:
o Is a design like this feasible within my performance constraint?
o Is a neural net the right tool for the determination of competitor
  objective? Can anyone steer me towards a description of a similar
  implementation? How large a net does it sound like I'll need?
o Can anyone provide a reference on a good hearing model? I realize
  it's more or less a triangulation problem, but I'm interested in how
  the placement accuracy goes in 3-space. (I can fudge this without
  too much trouble, I think.)

Thanks for any help you can give. Email or posted replies appreciated.

Eliot Shepard

-- 
Eliot Shepard
eshepard@world.std.com
