Newsgroups: comp.ai.alife
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!newsfeed.internetmci.com!in1.uu.net!world!kipb
From: kipb@world.std.com (Kip Bryan)
Subject: Re: Evolutionary environmental decision making
Message-ID: <Do8LBq.83o@world.std.com>
Organization: The World, Public Access Internet, Brookline, MA
References: <4i6icc$r4q@maze.ruca.ua.ac.be>
Date: Thu, 14 Mar 1996 02:52:38 GMT
Lines: 31

aliekens@zorro.ruca.ua.ac.be (Anthony Liekens) writes:

>Hiyall,

>I'm stuck with this problem of environments, decision making and evolution.
>I'm trying to make an evolutionary pacman. A game where the computer
>learns to play the game of pacman.
...
>Maybe I can try to replace the programs with simple neural networks which 
>evolve (not the classical learning process...). Are there any other
>possibilities to start with a completely random walk through the labyrinth
>to evolve to better movement?

If your ghosts move using an algorithm as opposed to randomly, you could
do it in a simple way, but the pacmen and pacwomen will just know
how to move against those ghost algorithms and would be useless against
similar but modified algorithms of ghosts.  The simple way would be
"go right 8 cells, then left 4 cells, then right 8 cells..."

A better way (or a more interesting way) would be to have the pacpeople
know about where the ghosts are, and what their states are, just as
humans do when operating the pacman.  "If there is a hungry ghost
that could get me in 8 cells' movement nearby, then I should either
run away or else move to the nearest 'turn it around' cookie and
then chase him."

It'd be fun to have both the ghosts and the pacpeople evolve.  Another
variation that might be interesting is "more than one pacman per game"
where they can run over each other, but they might cooperate by having
one of them taunt the ghost while the other eats up (symbiosis).

