Newsgroups: rec.games.corewar,comp.ai.alife
From: Robert@buchanan.demon.co.uk (Robert Macrae)
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news3.near.net!noc.near.net!bigboote.WPI.EDU!news.mathworks.com!news.duke.edu!zombie.ncsc.mil!paladin.american.edu!howland.reston.ans.net!pipex!peernews.demon.co.uk!buchanan.demon.co.uk!Robert
Subject: Re: Evolving CoreWar Warriors (was Pointers in pMARS)
References: <3lbjnp$cbh@uuneo.neosoft.com>
Organization: Buchanan Partners Limited
Reply-To: Robert@buchanan.demon.co.uk
X-Newsreader: Demon Internet Simple News v1.29
Lines: 40
X-Posting-Host: buchanan.demon.co.uk
Date: Fri, 31 Mar 1995 18:47:09 +0000
Message-ID: <796675629snz@buchanan.demon.co.uk>
Sender: usenet@demon.co.uk

In message <9503311346.AA18033@deep-thought.ericsson.se> Mark Sadler writes:

> Another idea is to run separate evolutions, score one population against each
> type of warrior, this should tend to develop 'paper-killers' or 'imp-killers'
> etc.  This would only required one set of battles for each individual.

Sounds good, but very optimistic. How familiar are you with corewars? This
issue of paper-scissors-stone is key; yes, in theory develop one of each by
feeding it its natural prey, but paper and scissors will be harder to grow
than stones because they require many more co-ordinated instructions. At 
a rough guess 10 opcodes x 5^2 address modes x 20^2 addresses => 100 000
possible instructions. A stone needs 4-5, a scanner (scissors) needs 7-10,
a replicator (paper) 6-10, so the search space is much more tractable for 
stones, at 10^20 rather than 10^40.

Note I'm conservative on addresses, assuming 20. This factors in a strong
preference for small addresses (-3,-2,-1,0,1,2,3 etc). Not compulsory
I guess, but if you fancy searching a (10x5^2x8000^2)^10 space, you must
_really_ believe in GAs! 

Further idea on sources of codes, in spirit of preselection above. Why not
form a large list of handwritten warriors end to end, and select most of your 
random codes from this list of instructions? It would bias you towards 
combinations of opcode, mode, address which appear more often. I guess I'm 
betraying my "Results over purity" bias again, but thats what you get for 
being an engineer in a finance company. How hard does academia push the 
purity line?

-- 
Robert Macrae
