
Genetic Algorithms Digest   Thursday, October 14, 1993   Volume 7 : Issue 27

 - Send submissions to GA-List@AIC.NRL.NAVY.MIL
 - Send administrative requests to GA-List-Request@AIC.NRL.NAVY.MIL
 - anonymous ftp archive: FTP.AIC.NRL.NAVY.MIL (Info in /pub/galist/FTP)

Today's Topics:
	- GAs To Optimise A Maximum Likelihod Estimator?
	- Reprint available
	- Technical Report available
	- "generic" GA package available
	- Call for papers
	- CFP: 1994 IMACS Int. Symp., Lille, France
	- Industrial/commercial apps of GAs - references needed

----------------------------------------------------------------------
****************************************************************************

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

IEE/IEEE Workshop on Nat Alg in Signal Processing, Essex (v7n5) Nov 15-16, 93
AI'93 Workshop on Evolutionary Computation, Melbourne, Aust(v7n16) Nov 16, 93
EP94 3rd Ann Conf on Evolutionary Programming, San Diego (v7n7) Feb 24-25, 94
IEE94 Colloquium on Molecular Bioinformatics, London, UK (v7n21)   Feb 28, 94
SPIE, Neural & Stoch. Methods in Image & Sig Proc, Orlando(v7n18) Apr 5-8, 94
FLAIRS-94 Workshop on Artif Life and AI, Pensacola Beach, FL(v7n23) May 4, 94
The IEEE Conference on Evolutionary Computation, Orlando(v7n26) Jun 26-30, 94
FOGA94 Foundations of GAs Wkshop, Estes Park, Colorado(v7n26)Jul 30-Aug 3, 94
SAB94 3rd Intl Conf on Sim of Adaptive Behavior, Brighton(v7n11) Aug 8-12, 94
ECAI-94, 11th European Conference on AI, Amsterdam (v7n23)       Aug 8-12, 94
ISRAM94 Special Session on Robotics & GAs, Maui, Hawaii (v7n22) Aug 14-17, 94
PPSN-94 Parallel Problem Solving from Nature, Israel (v7n9)      Oct 9-14, 94

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

****************************************************************************
------------------------------

From: econec@vax.ox.ac.uk
Date: Tue, 21 Sep 1993 12:54:55 +0100
Subject: GAs To Optimise A Maximum Likelihod Estimator?

Dear GA Persons,

I have written a GA to optimise a Maximum Likelihood Estimator and it doesn't
work very well. Has anybody else tried to do this? What problems did you
encounter? (The MLE here is of the form MLE=(A1**P1)*(A2**P2)*... where Pn are
numbers from 2-15 and An are terms of the form (M1*(e**X1)*((1-e)**Y1)) +
(M2*(e**X2)*((1-e)**Y2) + ... The set of Ms is the same for each term and must
sum to 1 ... they are probabilities that a person is of a certain type. The
powers in X and Y can be different over _all_ terms but each corresponding
X, Y set must sum to a fixed number usually about 5 or 6. The task is to
maximise the function by varying the Ms and e. (e is the chance of making an
error not the natural log base!)

This is a good GA problem, it has a very large search space which is pretty
irregular due to the large number of terms. Grid search is just unfeasible and
the nature of the problem space seems to cause hillclimbers to fail or reach
local optima.

I use three operators: CROSSOVER, MUTATION and something I call RELOCATION
which involves "taking a bit off one M and adding to another". Because the
terms in M must sum to one, CROSSOVER and MUTATION solutions need to be
renormalised and I'm not sure how this will affect the GA. Again somebody must
have done that before. (I know another possibility is to penalise solutions
that don't sum to 1 but I expect the number of legal solutions to be very small
relative to the number of illegal ones.)

The other main problem is that the problem terrain is quite "flat" with lots
of solutions of equivalent fitness below the optimum. I'm not sure if this may
be having some effect since I also have difficulties with premature
convergence.

I'm using a simple Holland type algorithm with offspring proportional to
individual/total fitness and I'm going to try toning this down as it seems to
produce too many copies of a few very fit solutions. (Does this signal
anything significant about the problem in itself?) I'm also wondering about a
better problem representation. At the moment it's just [M1 M2 .... e]. I guess
this isn't giving the GA too much chance to work on "separable" subassemblies
of the problem but one difficulty is that with probabilities these are not so
obviously available.  

I would be grateful to anyone who has tried to solve this problem or any
similar one before and could provide some tips. My grasp of GA theory is not
strong and I'm not even sure that my operators are "sensible" ones.

If anyone sends me any information that would be generally useful I'll pass it
back to the list

Many thanks,

Edmund Chattoe

PS Given the difficulties I have experienced with it, I can commend MLE's as
"hard" problems for testing GA efficiency and efficacy. Any takers?

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

From: Hugh Cartwright <hcart@vax.ox.ac.uk>
Date: Thu, 23 Sep 1993 08:41:12 +0100
Subject: Reprint available

    The following reprint is available: "Analysis of the
Distribution of Airborne Pollution using Genetic Algorithms",
published in Atmospheric Environment, Vol 27A, 1993. For copyright
reasons, this reprint is not available by anonymous ftp,
but hardcopy reprints are available on request to
Hugh Cartwright (HCART@vax.ox.ac.uk).

ABSTRACT:  Increasing environmental awareness has been an important
factor behind the  development of receptor models, which attempt to
identify the  sources of  airborne pollution  reaching a monitoring
station by analysis of the profile of pollutants collected. Despite
the central  role of  such models in  understanding the movement of 
air  pollution,  they  yield  results  which  contain   significant
uncertainty.   This paper reports  the  application of the  Genetic
Algorithm (GA) to this source apportionment problem. Implementation
of a matrix  formulation  for the  GA  allows it to tackle the many
source/many  receptor  problem,  with  results which suggest the GA 
approach is a significant advance over current models.

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

From: ww@neuroinformatik.ruhr-uni-bochum.de (Willfried Wienholt)
Date: Mon, 27 Sep 93 12:02:35 MET
Subject: Technical Report available ...

A postscript copy of the following Internal Report 93-07 
can be obtained by anonymous ftp (BINARY Mode) at 
ftp.neuroinformatik.ruhr-uni-bochum.de 
with path /pub/manuscripts/FuzzyRbfnES.ps.gz:

"Optimizing the Structure of Radial Basis Function Networks 
by Optimizing Fuzzy Inference Systems with Evolution Strategy"

Willfried Wienholt
Ruhr-Universit\"at Bochum
Institut f\"ur Neuroinformatik ND 04/584
44780 Bochum, Germany 
ww@neuroinformatik.ruhr-uni-bochum.de

ABSTRACT
This report takes advantage of Neural Networks (NN) and
Fuzzy Inference Systems (FIS) in order to design a system suited to
predict time series. 
We choose the solution of the Mackey--Glass time delay differential
equation in the chaotic domain as a sample problem. 
Fuzzy rules are generated from the sample data.
The system performance is improved by means of Evolution Strategy (ES).
The rules of the FIS are diminished in number due to a heuristic
approach. 
The optimization process is convenient for the structural design
of Radial Basis Function Networks (RBFN). The so far predetermined
RBFN is further optimized by gradient descent.
The system exhibits a good prediction accuracy and generalization.

Comments, suggestions, and questions are welcome.

If there are any problems with the gzip-format, the file may also
be transfered as an uncompressed text file just typing 
ftp>get FuzzyRbfnES.ps

Regards,

Willfried Wienholt
ww@neuroinformatik.ruhr-uni-bochum.de

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

From: janikow@radom.umsl.edu (Cezary Janikow)
Date: Mon, 20 Sep 93 17:03:19 CDT
Subject: "generic" GA package available

I have been using a "generic" GA package for a while. Few people that received
a copy suggested I make it available to others. I also decided to do so 
since the package is also intended as a growing library of representations 
and operators - it separates these from other mechanisms, following the 
AI ideas of separation of knowledge and inference. It also speeds up 
development of applications and experiments with comparison of different 
models, etc.

WHAT IS IT?
It is a "generic" GA package. It is not generic in the normal sense,
which usually implies some standard (binary) representation and its operators.
It is generic in the sense that all problem independent mechanisms have been 
implemented and can be used regardless of application domain.
Using the package forces (or allows, however you look at it) concentration
on the problem: you have to suggest the best representation, and the best
operators for such space that utilize your problem-specific knowledge. 
You do not have to think about possible GA models
or their implementation. This speeds up problem-specific GA applications
by about 95% for researchers w/o extensive GA libraries and about 50%
for those with complete good libraries. The package provides some libraries
of representations and operators. The user can either create new rep/opers,
use the existing, or implement rep/operators along with the problem. 
Problem is implemented by specifying only the evaluation and initialization.
Representations are defined by functions to allocate storage and display
chromosomes. etc... All such interfaces have standard definitions.

The package, in addition to allowing for fast implementation of applications
and being a natural tool for comparing different models and strategies,
is intended to become a depository of representations and operators.
Currently, only FP representation is implemented in the library with few
operators. 

The algorithm provides a wide selection of models and choices.
For example, population models range from 
generational GA, through steady-state, to (n,m)-EP and (n,n+m)-EP
models (for arbitrary problems, not just parameter optimization).
(Some are not finished at the moment).
Choices include automatic adaptation of operator probabilities and
a dynamic ranking mechanism, etc.

Even though the implementation is far from optimal, it is quite efficient
- implemented in ATT's C++ (3.0) (functional design) and also tested on gcc.
Along with the package you will get two examples. They illustrate how
to implement problems with heterogeneous and homogeneous structures, 
with explicit rep/opers and how to use the existing library (FP). 
Very soon I will place there another example - our GENOCOP operators
for linearly constrained optimization. This example nicely
illustrates the power of the generic mechanism for adapting the operators,
which nicely allocates the resources (high probabilities) to those operators
that currently show the best promise.
One more example soon to appear illustrates how to deal with complex 
structures and non-stationary problems - this is a fuzzy rule-based
controller optimized using the package and some specific rep/operators.

HOW TO LEARN MORE OR GET THE PACKAGE?
To learn more, you may get the User's Manual, available in compressed
postscript by anonymous ftp from radom.umsl.edu, in /var/ftp/userMan.ps.Z.
The package can be found in /var/ftp/GenET.tar.Z, and it includes
the User's Manual itself.

If you start using the package, please send me evaluations (especially bugs)
and suggestions for future versions.
Have fun.

Sincerely,
Cezary Z. Janikow                        Department of Math and CS, CCB319
tel (314) 553-6352                       UMSL
fax (314) 553-5400                       St. Louis, MO 63121
janikow@radom.umsl.edu

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

From: "Thomas L. Ward" <TLWARD01@ULKYVX.LOUISVILLE.EDU>
Date: Mon, 27 Sep 1993 12:17:33 -0500 (EST)
Subject: Call for papers

                                CALL FOR PAPERS

     THIRD INTERNATIONAL FUZZY SYSTEMS AND INTELLIGENT CONTROL CONFERENCE

                               Louisville KY USA
                               14-16 March 1994

This conference is devoted primarily to computer based feedback control
systems that rely on fuzzy logic, neural network theory, probabilistic
reasoning techniques, genetic algorithms, chaos theory, learning theory, and
other soft computing and artificial intelligence techniques.

The theme of this year's conference is "Industrial Applications of Soft
Computing."

In an abstract of a paper on soft computing, Lotfi Zadeh wrote:

      The past few years have witnessed a rapid growth of interest in
      novel modes of computation which are collectively referred to as
      soft computing.  The distinguishing characteristic of soft
      computing is that its primary aim is to exploit the tolerance for
      imprecision and uncertainty to achieve tractability, robustness
      and low cost.  Thus, in soft computing what is usually sought is
      an approximate solution to a precisely formulated problem or, more
      typically, an approximate solution to an imprecisely formulated
      problem.  A simple case in point is the problem of parking a car. 
      Generally, a car can be parked rather easily because the final
      position of the car is not specified exactly.  If it were
      specified to within, say, a fraction of a millimeter and a few
      seconds of arc, it would take hours or days of maneuvering and
      precise measurements of distance and angular position to solve the
      problem.  What this simple example points to is the fact that, in
      general, high precision carries a high cost.  The challenge, then,
      is to exploit the tolerance for imprecision by devising methods of
      computation which lead to an acceptable solution at low cost. 
      This, in essence, is the guiding principle of soft computing.  At
      this juncture, the major components of soft computing are fuzzy
      logic (FL), neural network theory (NN) and probabilistic reasoning
      techniques (PR), including genetic algorithms, chaos theory and
      parts of learning theory.  It may be argued that it is soft
      computing - rather than hard computing - that should be viewed as
      the foundation for artificial intelligence.  In the years ahead,
      this may well become a widely held position.

Some of the areas of interest include, but are not limited to, the following:

      Adaptive vector quantization        Modeling
      Chaos theory                        Neural control
      Differential competitive learning   Neuro-fuzzy control
      Expert control                      Operating experience
      Fuzzy control                       Process control
      Genetic algorithm control           Rule completeness
      Geno-fuzzy control                  Rule consistency
      Identification                      Rule interaction
      Implementation                      Self-organizing controllers
      Learning theory                     Soft computing
      Machine learning                    Soft computing for control
      Membership function elicitation     Statistical process control
      Membership function scaling         Time series
                                          
This conference is sponsored by the Institution for Fuzzy Systems and
Intelligent Control.

                             ORGANIZING COMMITTEE

Honorary Chair:

      Lotfi A. Zadeh

Local Chairs:

      Waldemar Karwowski
      Patricia A. S. Ralston
      Thomas L. Ward

International Program Advisory Board:

      James C. Bezdek         Charles L. Karr         Elie Sanchez
      Piero Bonissone         Arnold Kaufmann         Philippe Smets
      Christer Carlsson       George Klir             Alice E. Smith
      Augustine O. Esogbue    R. Lowen                Hideo Tanaka
      Donald R. Jones         E. H. Mamdani           Enric Trillas
      Siegfried Gottwald      Masaharu Mizumoto       Tibor Vamos
      William A. Gruver       Rammohan K. Ragade      Jonathan Weiss
      Mohammad Jamshidi       Dan Ralescu             C. K. Wong
      Deng Julong             Dan B. Rinks            R. R. Yager

                                CALL FOR PAPERS

Papers selected for presentation will be published in a PROCEEDINGS to be
distributed at the conference.

Three copies of an abstract (in English) of about 300 words should be
submitted by 31 October 1993.  Please include title, author(s) name(s),
affiliation(s), and address of person to whom correspondence should be
directed.  FAX and e-mail submissions of abstracts are acceptable.  Please
send abstracts prior to 31 October 1993 to

      Prof. Patricia A. S. Ralston
      Engineering Mathematics and Computer Science
      University of Louisville
      Louisville KY 40292
      USA

      Telephone:  502-588-0479
      Fax:        502-852-4713
      Bitnet:     PARals01@ULKYVM.BITNET
      Internet:   PARals01@ulkyvm.louisville.edu

Authors will be notified of acceptance by 15 November 1993.  Full camera ready
papers will be due by 15 January 1994.

For additional information regarding the 3rd IFSICC, please contact

      Prof. Thomas L. Ward
      Industrial Engineering
      University of Louisville
      Louisville KY 40292
      USA

      Telephone:  502-588-6342
      Fax:        502-588-5633
      Bitnet:     TLWard01@ULKYVM.BITNET
      Internet:   TLWard01@ulkyvm.louisville.edu

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

From: "N. Saravanan" <saravan@amber.me.fau.edu>
Date: Thu, 30 Sep 93 18:36:33 -0400
Subject: CFP: 1994 IMACS Int. Symp., Lille, France 

                    ******Call for Papers*******

                1994 IMACS International Symposium
                                 on
          Signal Processing, Robotics and Neural Networks
                            Lille, France
                         April 27 - 29, 1994

                     ****INVITED SESSION****

	Evolutionary Computation Applications in Signal Processing

Evolutionary Computation includes stochastic optimization techniques
like Evolutionary Programming, Genetic Algorithms, and Evolution
Strategies.  Papers are solicited in the area of evolutionary
computation applications in signal processing, robotics and neural
networks.

Topics include but not limited to:

	* Identification using Evolutionary Computation
	* Automatic Control using Evolutionary Computation
	* Evolutionary Computation in Robotics
	* Neural networks and Evolutionary Computation
	* Evolutionary Computation in Signal Processing 

Deadlines:

	October 25, 1993	Extended abstract* due 
	November 8, 1993	Notification of Conditional Acceptance**
	February 1, 1994	Deadline for final papers

Abstract Format:

	The abstract may be submitted in any of the following formats. It
	must contain the name, address, and Email (if appropriate) of the
	author to whom all correspondence will be sent.

	-Email submission (ASCII or LaTeX) should be mailed to
	 saravan@amber.me.fau.edu
	-Hard copy submissions should be sent to the Session Chair at the 
	 address given below.
	-The abstracts may be faxed to the Session Chair at (407) 367-2336

	
Please Contact:

	Paul Luebbers (Session Chair)      N. Saravanan (Session C-Chair)
	Department of Elec. Eng.           Department of Mech. Eng.
	Florida Atlantic University        Florida Atlantic University
	P.O. Box 865                       P.O. Box 3091
	Deerfield Beach, FL 33443          Boca Raton, FL 33431
	Tel (407) 367-3476                 Tel (407) 367-2730
	Fax (407) 367-2336                 Fax (407) 367-2825
	luebbers@acc.fau.edu               saravan@amber.me.fau.edu

	
	*Extended abstract ( < 3 pages or 1000 words) or a draft paper.  

	**Acceptance at this stage will be conditional on the acceptance
	of the entire session by IMACS. 
	
Saravan

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

From: Hugh Cartwright <hcart@vax.ox.ac.uk>
Date: Mon, 27 Sep 1993 08:36:59 +0100
Subject: Industrial/commercial apps of GAs - references needed.

    For an upcoming conference I am reviewing applications 
of GAs which have been actually implemented in industry
or commerce, or have a real prospect of being implemented.
Although I am aware of much of the work that has been done
in this field, I am sure there are applications of which I am
unaware (having been published very recently perhaps, or
published in industrial, rather than AI journals).

    If you have completed any work in this area, even some time
ago, I would be pleased to have a note of references to published papers 
and articles. Even better would be a copy of the papers themselves,
or instructions for ftp access to them.  Thanks.

Hugh Cartwright
Physical Chemistry Laboratory
Oxford University, Oxford OX1 3QZ, England.
HCART@vax.ox.ac.uk

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


