Newsgroups: bionet.population-bio,comp.theory.cell-automata,comp.ai.alife
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!swrinde!sgiblab!uhog.mit.edu!news.kei.com!ub!galileo.cc.rochester.edu!prodigal.psych.rochester.edu!stevens
From: stevens@prodigal.psych.rochester.edu (Greg Stevens)
Subject: Re: Simulated Environments Parameter Question
Message-ID: <1994Dec9.170002.11514@galileo.cc.rochester.edu>
Sender: news@galileo.cc.rochester.edu
Nntp-Posting-Host: prodigal.psych.rochester.edu
Organization: University of Rochester - Rochester, New York
References: <1994Dec5.205450.28653@galileo.cc.rochester.edu> <HIEBELER.94Dec8204448@hershey.harvard.edu>
Date: Fri, 9 Dec 94 17:00:02 GMT
Lines: 58
Xref: glinda.oz.cs.cmu.edu comp.theory.cell-automata:3154 comp.ai.alife:1466

In <HIEBELER.94Dec8204448@hershey.harvard.edu> hiebeler@husc.harvard.edu (Dave Hiebeler) writes:
>In article <1994Dec5.205450.28653@galileo.cc.rochester.edu> stevens@prodigal.psych.rochester.edu (Greg Stevens) writes:

>>[problems with REGEN_RATE, REPROD_THRESH, WORLD_SIZE, etc.]  
>> Does anyone know of any hints, other than methodical trial and error,
>> for discovering good combinations of values for these parameters?

>  What are the goals of the simulation?  Are you just trying to get
>stable populations?  Something biologically plausible?

Well, I wantto get populations stable enough to evolve.  Right now the
chromosome contains two factors.  First, the food-searching strategy
of the agents is controlled by a little neural net, which takes as input
the angle and distance of nearest food (and its own previous output) and
(through 7 hidden units) outputs a 2-unit, binary coded output for
movement (00 halt, 01 turn right, 10 turn left, 11 forward).

This part of the expierment was inspired by on done by Elman and Nolfi where
they started off with the chromosom controling thr weights, didn't have
learning of any kind, (therefore in initial generations having random
search patterns), and saw that through time more and more efficient strategies
for getting food were evolved (where fitenss was # foods collected).

I'm doing a spin-odd on that, except I have an additional gene and the 
following question I want to ask: "Can it be computationally shown that there
is an evolutionary benefit to there being a weaning period during which the
child learns from the parent's behaviors?"  The additional gene controls
the length-in-time-steps of the weaning period, and when a new agent is
born, the counter moves through that many time steps and for that 
intercval, the agent copies the parent's movements and uses a supervised
learning algorithm to hange weights.  After that, no learning takes place.

I want to see if, when length-of-weaning-period is included in the
chromosome, there is evolutionary pressure on the length of the weaning
period.

But in order to do this, I want the buggers to reproduce for enough generations
for them to evolve.

>....  One
>difference is that in my simulation the "food" (plants) were not
>deposited randomly down on the grid; new plants appeared when the
>current ones decided to reproduce (using age-based life-history
>probability tables -- although after learning something more about
>plants I realize I should have used size-based tables).  And in fact
>the new plants were deposited fairly near the parent.  This gave the
>simulation more spatial structure and heterogeneity.  I got lots of
>local extinctions in different parts of the lattice, but global
>extinctions were less common (how much less common depended on the
>size of the lattice).

This is interesting and sounds like something I may want to try -- I've
actually considered it at one point already.

Greg Stevens

stevens@prodigal.psych.rochester.edu

