
Genetic Algorithms Digest    Monday, 18 June 1990    Volume 4 : Issue 10

 - Send submissions to GA-List@AIC.NRL.NAVY.MIL
 - Send administrative requests to GA-List-Request@AIC.NRL.NAVY.MIL

Today's Topics:
	- using GA to generate hypothesis valid against historical data
	- Summary of the Artifical Life conference
	- GA Research Assistantships and a GA Professorship

******************************************************************************

CALENDAR OF GA-RELATED ACTIVITIES: (with GA-List issue reference)

7th Intl. Conference on Machine Learning (Austin)             Jun 21-23, 1990
Workshop Foundations of GAs (v3n19)                           Jul 15-18, 1990
Conference on Simulation of Adaptive Behavior, Paris (v3n21)  Sep 24-28, 1990
Workshop Parallel Prob Solving from Nature, W Germany (v4n5)  Oct 1-3,   1990
2nd Intl Conf on Tools for AI, Washington, DC (v4n6)          Nov 6-9,   1990
4th Intl. Conference on Genetic Algorithms (v4n9)             Jul 14-17, 1991

(Send announcements of other activities to GA-List@aic.nrl.navy.mil)

******************************************************************************

Date: Sun, 17 Jun 90 10:53:32 EDT
From: hall@ishmael.leis.bellcore.com (Michael R. Hall)
Subject: Re: using GA to generate hypothesis valid against historical data

   Mr. Havener wrote:

     [Intro munched]
     >
     >  I need to apply GA to learn from a vast database of historical
     >  readings. One tough part is to design a payoff function that will
     >  reward a GA for generating a hypothesis that can be demonstrated
     >  valid against the historical data.
     >
     >  Example:
     >  Goal:  explain occasions when product attribute X > value Y
     >  GA generated hypothesis:  That happens when variable Z > a constant C.
     >
     [Outro munched]

   I advise against using a GA (or neural net for that matter) to learn
   such concepts. There are a family of algorithms that are tuned to
   solve this problem, and you will be hard pressed to beat them with a
   reasonable amount of CPU. Some examples of split-based concept
   learning algorithms are C4, ID3, PLS1, and CART. Split-based concept
   learning algorithms have many advantages over GA's and neural nets,
   but their key selling points are their reasonable computational
   complexity (both parallel and nonparallel), their robustness, ease of
   use, and their ability to output human readable concept definitions
   (as decision trees or rules.) 

   An example of a concept learning algorithm beating the pants off a
   genetic algorithm can be found in "An Empirical Comparison of Genetic
   and Decision-Tree Classifiers" by Quinlan in the Proceedings of the
   Fifth International Conference on Machine Learning (1988). The
   comparison was perhaps not entirely "fair", because the C4 algorithm was
   batch, while the GA was incremental, and the concept taught in the
   experiment is rather untypical. Also, the GA was a brain-damaged
   classifier system, and so with a different architecture, one might
   hope that GA's would do better. 

   When I was a naive undergrad, I applied a full-blown classifier system
   to a rather difficult database of tax audit data. Parasitic rules
   quickly attacked any productive rules that arose, and after a great
   many generations, the system collapsed to a bunch of overly-specific
   (and thus parasite-resistent) classifiers that would only output
   "True", and thus the performance of the system was equal to the
   percentage of positive instances in the database (70%) - pretty
   pathetic, especially considering that the concept learning algorithm I
   later applied did so much better. The split-based concept learning
   algorithm output the degenerate "always true" hypothesis when given
   just a few dozen examples (out of the few thousand available
   examples), but when given more examples it rose to about 90% accuracy,
   and it did find some interesting and useful classifications (e.g.,
   when the company is located in a certain area and has gross wages
   between X and Y, then it is very likely cheating on its taxes.)

   As a result of this experience, I believe that classifier systems are
   not good for learning concepts. I have personally tried to devise a
   non-classifier GA-based concept learner, but I have not come up
   with anything that works well. However, you might want to look at
   Larry Rendell's paper in the proceedings of the first GA conference
   for an example of combining a GA with a split-based concept learner (PLS2).

   If you are able to send me your data, I would be willing to run it
   through PLS1 and my own unpublished algorithm (YACL) and let you
   know what pops out. Even if you do use a GA to analyze your data, you
   had better compare its performance against that of a split-based
   concept learning algorithm.

   Michael R. Hall
   Bell Communications Research
   hall@ishmael.leis.bellcore.com
   {rutgers!} bellcore!nvuxh!hall

-------------------------------------

Date: Thu, 14 Jun 90 16:48:57 EDT
From: Gary Berg-Cross <garybc@potomac.ads.com>
Subject: Summary of the Artifical Life conference

   Notes on the Artificial Life II (A-Life) Conference (Santa Fe, Feb 5-9)
   Jointly sponsored by the Santa Fe Institute and the Los Alamos National 
   Laboratory Center for Non-Linear Studies

   Gary Berg-Cross
   Advanced Decision Systems

                       ----------------------------
   Background

	To say the least this was a wide-ranging conference attended by
   200-300 professionals (computer scientists, biologists, philosophers,
   psychologists, mathematicians etc.) The conference was organized around
   the theme that artificial systems  (those with same dynamics, but
   different medium from natural ones) exhibit "lifelike" behaviors such
   as learning, adaptation, collective actions etc. and are worthy of
   scientific investigation even if they do not structurally mimic
   existing life.  Because it offers the promise of scientific
   manipulations of variables Alife is  an extremely fruitful research
   area in complex domains.   In particular Alife is seen as an area
   useful to the establishment of general theoretical biologies, ecologies
   economies and even social systems. 

	The conference days were sequenced with a bottom up approach.
   Thus we started with physical models and transitioned to evolutionary
   chemical, simple organisms, complex organisms, ecologies and ended with
   society and culture.  In keeping with this, early conference work
   emphasized finite automata technologies and
   complexity/chaos/autocatalytic theory and Lindenmayer systems for
   botanical growth.  GAs were more typical in the middle and later
   portions of the conference.  Most of these  presentations were in terms
   of interesting applications of GAs rather than insights into GA design.
   The following  represent simple summaries of talks citing or inspired
   by GAs.

   John Holland described the Echo system - a simulated, closed world with
   very simple presentations for physics, geography, biochemistry etc.
   Organisms (phenotypes) interact with the local environment and survive
   based on fitness (partially geotype based).  Genes include one for
   offense and others for defense.  Environmental competitions for food
   match one organism's offenses against others defenses. The winner gets
   he food.  Once an Echo world is started there is no outside control or
   intervention.  Echo worlds exhibits sophisticated ecological processes
   such as biological arms races and speciation.  For example "trees"
   develop a  "quinine" substance that kills its predator.  A new
   "predator" evolves that is dependent on the "quinine". More advanced
   tests of selective mating, co-adapted alleles are planned and better
   models including realistic models of fitness. 

   Related work on simulated ecologies includes David Jefferson 's (UCLA)
   Genesys system evolving populations of up to 100K+ organisms each
   represented by a finite state automata.  RAM also at UCLA represents
   organisms as separate Lisp programs with capabilities of interactions
   with the environment, learning, communication and reproduction.  ArtAnt
   (Robert Collins, rjc@cs.ucla.edu) simulating a colony of food foraging
   and reproducing organisms is also out of UCLA.   ArtAnt uses a NN to
   represent haploid chromosomes that guide behavior.  Evolution is not
   yet implemented.

   Stephanie Forest discussed using GAs to study the evolution of
   Cooperative Behavior.  This talk reviewed older work on the iterated
   prisoner's dilemma problem using it to understand international relations.

   There were several efforts that combined neural network NNs learning
   techniques with evolutional approaches.  For example David Ackley
   (Bellcore) strategy of evolutionary reinforcement learning to select
   generational survivers (up to 300 generations were studied)..  More
   familiar and well received was Richard Belew's (rik%cs@ucsd.edu) work
   on evolving networks.  This was an expansion of the work, reported at
   the 89 GA conference at George Mason University, exploring how GAs can
   be used to find local and efficient learning parameters for back prop
   networks.  He uses GAs for global sampling and back prop networks for
   local sampling.  GAs are also  used to find likely staring points in
   weight space for these nets.

   David Stork, Bernie Jackson, Mark Burns  &  Scott Walker (Stanford)
   described preadaptation research on crayfish swimming-tail flipping.
   They programmed biologically motivated GAs, and NNs for learning and
   simulated 150 generations of "crayfish" (no pun intended they used a
   Connection Machine-2 for implementation) to show the development of
   relevant neural circuits suitable.

  Peter Todd and Geoffrey Miller (also at Stanford's Psych Dept.)
  described a general framework for the evolution of adaptive simulated
  creatures moving in 1 and 2 D worlds.  Again the phenotype's nervous
  system is represented by NNs and a GA is used to "design" the back prop
  architectures.  Sensory (color, smell), cognitive and effector systems
  are evolved.  the environmental situations studied include  avoiding
  traps and predators and optimal foraging (trade off movement costs
  versus nutrition).  For example they studied population survival as a
  function of smell fitness/accuracy and learning efficiency.  One strong
  conclusion of there work is to confirm the primacy of motor fitness -
  you must get it right first then can evolve computational efficiency and
  internal representations.  These authors explored GA issues of the use
  of fitness functions (against a direct use as I understand it) and a
  bucket brigade approach to credit assignment. 

-------------------------------

Date: Thu, 14 Jun 90 15:03:38 CDT
From: "Dave Goldberg" <DGOLDBER@UA1VM.ua.edu>
Subject: GA Research Assistantships and a GA Professorship

  I seek BS, MS, and PhD students interested in investigating genetic
  algorithms at the University of Illinois at Urbana-Champaign.  Good
  computations and mathematics background and capability are a must; prior
  exposure to GAs through courses or research is a plus.  Candidates must
  be admitted or admissable to the University of Illinois in an
  appropriate academic discipline (science, engineering, business,
  mathematics, etc.), although assistantships will be administered through
  the Department of General Engineering.  Before 1 August, send inquiries
  to me at the following address: Department of Engineering Mechanics,
  University of Alabama, Tuscaloosa, AL 35487, dgoldber@ua1vm.bitnet or
  dgoldber@ua1vm.ua.edu.  After 1 August, send inquiries to my new
  address: Department of General Engineering, 117 Transportation Building,
  University of Illinois at Urbana-Champaign, Urbana, IL, 61801, (217)333-
  2730 (email address unknown).

  My department at Alabama seeks to fill the position I'm vacating with a
  person who can teach in an engineering mechanics department and do
  research in genetic algorithms.  A more detailed position announcement
  is available from Dr. James L. Hill, Professor and Head, Department of
  Engineering Mechanics, University of Alabama, Tuscaloosa, AL 35487,
  (205)348-7241.

  Dave Goldberg
  dgoldber@ua1vm.ua.edu
  dgoldber@ua1vm.bitnet

--------------------------------
End of Genetic Algorithms Digest
********************************


