
Genetic Algorithms Digest   Monday, April 25, 1994   Volume 8 : Issue 13

 - 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 and MLE Optimization: Summary
	- References on GAs for Grouping/partitioning (Re: v8n8)
	- CFP IEEE/IAS INTERNATIONAL CONFERENCE

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

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

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
ECAI-94 Wkshp on Applied Genetic & Other Evol Algs, Amsterdam(v8n5) Aug 9, 94
IEEE/Nagoya Univ WW Wkshp on Fuzzy Logic & NNs/GAs, Japan(v7n33) Aug 9-10, 94
ISRAM94 Special Session on Robotics & GAs, Maui, Hawaii (v7n22) Aug 14-17, 94
Evolution Artificielle 94, Toulouse, France (v8n10)             Sep 19-23, 94
COMPLEX94 2nd Australian National Conference, Australia (v7n34) Sep 26-28, 94
PPSN-94 Parallel Problem Solving from Nature, Israel (v7n32)     Oct 9-14, 94
EP95 4th Ann Conf on Evolutionary Programming, San Diego,CA(v8n6) Mar 1-4, 95
ICANNGA95 Intl Conf on Artificial NNs and GAs, France (v8n10)   Apr 18-21, 95
ECAL95 3rd European Conf on Artificial Life, Granada, Spain(v8n5) Jun 4-6, 95

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

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

From: econec@vax.ox.ac.uk
Date: Fri, 01 Apr 1994 12:32:15 +0100
Subject: GAs and MLE Optimization: Summary

Dear GA List,

Here is a brief history of my attempts  to apply a GA to a practical  problem.
I'll keep it short, but hope that the process may be found interesting as well
as the results!

The problem is the optimisation of  a Maximum Likelihood Estimator that  arose
from the method of  Harless and Camerer (1992).  Here agents in an  experiment
are modelled as  being of  one of  several types  on the  assumption that  the
chance of their  making each  error is  a constant  e<1. Thus  an agent  whose
observed behaviour is [A A A A] is more likely to be of `type' [A A A B]  with
one error, than  of type  [B B  B B] with  four errors.  In our  case the  MLE
function is a set of 15  or so brackets of 6  terms each, each bracket can  be
raised to a  power of  upto 6.  The GA  represents the  allocations of  agents
between types and the error rate such that this MLE (which is also the fitness
function) is maximised.  After getting  the bugs  out, I  discovered that  the
program was  still hopelessly  slow, but  wasn't sure  how much  more I  could
expect. (I reflected that  comparing GA with brute  search even as a  ballpark
concept, isn't much use in real, as opposed to formal time, because all the GA
tries far fewer  solutions, it needs  a lot  nore processing time  to do  each
generation ...)

At that point  I emailed  the GA  Digest and  got loads  of replies,  promptly
spotting two  more stupid  bugs in  the program.  (One of  which was  a  small
procedure that ended up taking up 1/3 of the processor time till rewritten ...
that helped!)

The replies fell into  two classes. The  first three were:  "Don't do it  like
that, do it like this!" I can't comment on these because my supervisor  wasn't
prepared to spare me more time for  coding, except to say that all three  look
interesting and in another life, I might have tried them ...

Emanuel Falkenauer (EMAIL: PIERRE_LECOCQ_CRIF@eurokom.ie) wrote:

Before going into detail, I would like  to check whether I grasp your  problem
properly. Is  your problem  one of  Clustering, i.e.  dividing `persons'  into
groups, in such a way  that the partition you end  up with is the most  likely
one according to ML criteria? [YES, IT WAS!]

If so, you are facing what I call  a `grouping' problem, and I have very  good
reasons to  believe  that  a  classic Holland-style  GA  or  a  Goldberg-style
ordering GA will NOT  work on it [AH!].  I have been working  on that kind  of
problems for some time and have recently submitted a paper on what I call  the
Grouping GA (GGA) to Evolutionary Computation. [IN A NUTSHELL, THE PROBLEM  IS
THAT INDIVIDUAL OBJECTS DON'T MEAN MUCH  WITHOUT REFERENCE TO GROUPS IN  WHICH
THEY  OCCUR  AND  THAT   WE  CAN  DEFINE   USEFUL  OPERATORS  AND   ALGORITHMS
ACCORDINGLY.]

A second approach came fron Michael De La Maza (EMAIL: mdlm@ai.mit.edu):

It is our experience  that genetic algorithms are  poor optimizers and  that a
simple hill climbing algorithm  often outperforms them.  (see, e.g., Davis  in
ICGA91). Also, it has been argued that genetic algorithms were not designed to
optimize  functions  (see,   e.g.,  "Genetic  Algorithms   are  Not   Function
Optimizers" by Ken DeJong in FOGA93).

Deniz Yuret and  I have  developed an optimization  algorithm, called  dynamic
hill climbing, that outperforms genetic algorithms and simulated annealing  on
all the continuous optimization problems that we have tried. Although the code
is not  quite  ready for  distribution,  we were  wondering  if you  would  be
interested in  trying it  out on  your  problem? [I  HAVE TO  SAY THAT  I  WAS
UNCONVINCED BY  THIS ONE.  I  WAS WORKING  IN  PARALLEL WITH  A  "TRADITIONAL"
COMPUTER SCIENTIST  WHO TRIED  NAG  ALGORITHMS AND  ALL  SORTS OF  OTHER  HIGH
POWERED HILLCLIMBERS AND THEY  ALL GAVE UP IN  DISGUST OR GAVE SILLY  ANSWERS.
THERE'S SOMETHING HORRIBLE ABOUT OUR PROBLEM  SPACE, BUT WE STILL AREN'T  SURE
WHAT ...]

The final comment,  the nearest to  a specific previous  attempt to solve  the
problem came from Donald Szarkowicz (EMAIL: DSZARKO@iunhaw1.iun.indiana.edu):

The paper discusses a hybrid GA which combines a simple genetic algorithm with
a real-value  parameter encoding  and a  systematic search  region  reduction.
Three examples are discussed: parameter  estimation for a gamma  distribution,
trajectory estimation, and linear system identification.

The examples I considered are  simpler than the problem  you pose, but I  have
had good  results using  this  hybrid technique  on other  problems  involving
several dozen variables,  linear and  nonlinear constraints,  and mixtures  of
real, integer,  and  logical problem  variables.  [ALTHOUGH THIS  IS  A  FULLY
DEVELOPED PAPER WITH DETAILED ECONOMETRIC DISCUSSION, I SIMPLY WASN'T SURE HOW
EASY IT WOULD BE  TO ADAPT IT  TO THE DISCRETE  GROUPING PROBLEM DISCUSSED  BY
HARLESS AND CAMERER.]

The next  line  of  attack,  which  I was  able  to  make  use  of,  suggested
alternative techniques  WITHIN  the  GA framework.  The  most  important  from
Francisco Gordillo (EMAIL: gordillo@etsii.us.es):

I am  interested  in one  of  the  points that  you  made: how  to  deal  with
optimization when the sum of the parameters must be equal to 1. I am trying to
solve this problem using a GA. I am  using a method that I describe below  and
that seems to work. But I am quite a novice in the subject of GAs so I am  not
sure if I am not doing anything  wrong. So look at this method critically.  [I
DID, AND IT SEEMS GOOD TO ME TOO, BUT I WOULD BE GLAD TO HEAR FROM ANYONE WITH
A MORE FORMAL BENT WHETHER THIS MIGHT BE DONE BETTER.]

The most  important  part of  my  method  is that  I  am trying  to  make  the
parameters sum to 1 by means of the codification of the individuals: the  code
guarantees that every individual is valid,  that is, its parameters sum to  1.
This is  done by  normalising when  decoding. By  means of  this  codification
system you needn't to worry about normalizations after crossover or  mutation.
[MY MISTAKE WAS  TO WASTE TIME  NORMALISING AND RENORMALISING  THE POOL.  THIS
CAUSED ROUNDING ERRORS TOO.]

On the  other hand,  I thought  that imposing  this equality  restriction  the
number of freedom degrees is equal to the number of parameters (n) minus 1, so
I must look for  (n-1) variables. In  short, the decoding  method is that  the
string  of  each  individual  represents  the  unnormalised  value  of   (n-1)
parameters (which can be between 0 and M). The unnormalised value of the  last
parameter is  fixed to  sqrt(M). Every  parameter is  divided by  M to  have a
number less than 1. Finally they are normalised dividing by the sum of them.

Paolo Agati (EMAIL: agati@uts340.univ.trieste.it) suggested:

I'm developing a GA for fractal interpolation of signals and I had problems in
getting a good solution until I decided to try other selection techniques than
the classic roulette wheel. I've  found that a good  approach is the one  of a
so-called local selection as presented in "Selection in massively parallel GA"
by Collins  R.J. &  Jefferson D.R.,  Proc. of  5th ICGA.  [I DIDN'T  TRY  THIS
BECAUSE IT WASN'T CLEAR WHY A ROULETTE WHEEL SELECTION WAS UNACCEPTABLE FOR MY
PROBLEM.]

Eric  Breakenridge's   (EMAIL:  eric@music.gla.ac.uk)   experience  with   the
"difficulty" of the problem and the complexity of the solution space parallels
my own:

I have used  GAs for  MLE of ARMA  parameters from  time series  observations.
Results have been fairly good. The results have shown that the parameter error
is minimised by the GA. Unfortunately the algorithm takes such a long time  to
run that I have not had time  to fully investigate with all the parameters  of
the GA. The  number of experiments  I have run  will not yet  give a  reliable
estimate of the estimation error variance.

I do agree with you that the problem terrain is very flat. In my application I
have awarded a "large value" to illegal  solutions. I had been using a  binary
representation but I'm now starting to use a real valued GA.

Thanks for replies by:

Frank Zanotto  (zanotto@ripper.enet.dec.com) who  is  trying to  find  maximum
likelihoods for vectors in BP for neural nets. I'm not sure if he means by  ML
what I mean though!

Ivan Mann  (if he  is reading  this)  saw parallels  between his  problem  and
grouping in classes. It sounds like the FALKENAUER work would be relevant.

I hope that this reply  will constitute a (late)  substitute reply to all  the
individuals  mentioned  here,  though   I  am  still   happy  to  enter   into
correspondence. The GA has speeded  up by a factor of  about 5 from the  first
draft and is  now just  slow, rather  than uselessly  slow. The  next step  is
translation to C and installation  on a PC where it  can run all night  rather
than being limited to 15 minutes of VAX time per batch!

REFERENCES

For an  explanation  of  the  basic  idea  behind  MLE  see  any  intermediate
econometrics textbook ...

Falkenauer, Emanuel  and  Gaspart, P  (1993)  "Creating Part  Families  With A
Grouping Genetic Algorithm" in M  Vidyasagar (Editor) PROCEEDINGS OF ISIR  '93
INTERNATIONAL SYMPOSIUM ON INTELLIGENT ROBOTICS  Jan 7-9 1993 Bangalore  India
Tata McGraw-Hill Publishing Co Ltd New Dehli India pp: 375-384

Falkenauer, Emanuel  "New  Representation And  Operators  For GAs  Applied  To
Grouping Problems" Technical Report  CRIF Industrial Automation (Under  Review
For Evolutionary Computation)

Harless,  David  and  Camerer,  Colin   (1992)  "The  Predictive  Utility   Of
Generalized Expected Utility Theories" DRAFT PAPER

Szarkowicz, Donald S "A Multi-Stage  Genetic Algorithm for Maximum  Likelihood
Parameter  Estimation",  Published  in:  proceedings,  1993  Summer   Computer
Simulation Conference, Boston, MA, July 19-21, 1993, pp. 984-989.

Yuret, Deniz and  De La  Maza, Michael "Dynamic  Hillclimbing: Overcoming  The
Limitations  Of  Optimization  Techniques"   (I'm  afraid  the  reference   to
conference proceedings, though not the article itself is in TURKISH, so if you
are interested, perhaps you should email them!)

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

From: Pierre Lecocq CRIF <M3730@eurokom.ie>
Date: Sat, 02 Apr 1994 14:04 +0200 (CET-DST)
Subject: References on GAs for Grouping/partitioning (Re: v8n8)

Hello,

In the recent GA-List v8n8, two contributions looked for references on GAs
for routing, and layout and grouping ("GAs for Manufacturing Systems") and
set covering/partitioning problems ("set covering/partioning problems").

These problems are of considerable practical importance, and I have been
working on them for some time now, which led to the introduction of the
Grouping GA (GGA), i.e. a GA heavily modified to suit the structure of the
grouping/partitioning problems.

The following are some of my references pertaining to the GGA (some appear
several times, as they report on several applications of the algorithm).
They are available on request.


Bin Packing and Line Balancing:

   - Falkenauer E. - The Grouping Genetic Algorithm and OR: the Case of Bin
Packing, submitted to Annals of Operations Research.

   - Falkenauer E. - Setting New Limits in Bin Packing with a Grouping GA
Using Reduction, Technical Report, CRIF Industrial Automation, Brussels.

   - Falkenauer E. - New  Representation and Operators for GAs Applied to
Grouping Problems, provisionally accepted to Evolutionary Computation.

   - Falkenauer E. - The Grouping Genetic Algorithms - Widening the Scope of
the GAs, in JORBEL - Belgian Journal of Operations Research, Statistics and
Computer Science, Vol. 33 (1, 2), pp.79-102.

   - Falkenauer E., Delchambre A. - A Genetic Algorithm for Bin Packing and
Line Balancing, in Proceedings of the 1992 IEEE International Conference on
Robotics and Automation, May 10-15, 1992, Nice, France. IEEE Computer
Society Press. pp.1186-1192.


Group Technology, Conceptual Clustering:

   - Falkenauer E., Gaspart P. - Creating Part Families with a Grouping
Genetic Algorithm, in M.Vidyasagar (Ed) Proceedings of ISIR'93 -
International Symposium on Intelligent Robotics, January 7-9, 1993,
Bangalore, India, Tata McGraw-Hill Publishing Co. Ltd, New Delhi, India.
pp.375-384.

   - Falkenauer E. - The Grouping Genetic Algorithms - Widening the Scope of
the GAs, see above.

   - Falkenauer E. - A Genetic Algorithm for Conceptual Clustering, Report
FMS39, CRIF Industrial Automation, Brussels, December 1991.


Grouping, Economies of Scale:

   - Falkenauer E. - New  Representation and Operators for GAs Applied to
Grouping Problems, see above.

   - Falkenauer E. - The Grouping Genetic Algorithms - Widening the Scope of
the GAs, see above.

   - Falkenauer E. - Achieving Economies of Scale with a Grouping Genetic
Algorithm, communication presented at ORBEL6, Brussels, January 24-26, 1992.

   - Falkenauer E. - A Genetic Algorithm for Grouping, in R.Gutierrez and
M.J.Valderrama (Eds) Proceedings of the Fifth International Symposium on
Applied Stochastic Models and Data Analysis, April 23-26, 1991, Granada,
Spain. World Scientific Publishing Co. Pte. Ltd. Singapore. pp.198-206.


Workshop Layouting:

   - Falkenauer E., Pommerenke B. - Grouping 750 Machines into Optimal Cells
in Minutes, in preparation.


Sincerely,

Emanuel Falkenauer
CRIF Industrial Automation
CP 106 - P4
50, av. F.D.Roosevelt
B-1050 Brussels
Belgium
E-mail: PIERRE_LECOCQ_CRIF@eurokom.ie

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

From: VEMURI@icdc.llnl.gov
Date: 06 Apr 1994 13:33:19 -0800 (PST)
Subject: CFP IEEE/IAS INTERNATIONAL CONFERENCE

			SECOND ANNOUNCEMENT
			CALL FOR PAPERS
		IEEE/IAS INTERNATIONAL CONFERENCE
			HYDERABAD, INDIA
			JANUARY 5-7, 1995

		IEEE INDUSTRIAL AUTOMATION AND CONTROL

COSPONSORED BY

IEEE INDUSTRIAL APPLICATIONS SOCIETY
IEEE HYDERABAD SECTION
IEEE INDIA COUNCIL

TOPICS  COVERED:

Expert Systems and Fuzzy logic Controllers
Genetic Algorithms in industrial control
Neural Nets in Industrial Automation and control
Intelligent Sensors, transducers, and actuators
Programmable controllers and Industrial Controllers
Distributed Computer Control Systems
Flexible Automation Systems
Real-Time Communications Networks and Controls
Energy Conservation Tools and Technologies
.Other topics can be added.

OFFICIAL LANGUAGE: English

PAPER DEADLINES

15 June 94	Receipt of Extended Abstract from Authors
16 August 94	Notification of Acceptance
14 October 94	Receipt of camera-ready full paper from authors

Prospective authors are requested to forward an extended abstract of their
proposed paper to the Selection Committee. Abstracts should indicate the name,
company affiliation, address, telephone number, FAX number, E-mail address
of each other.

NOTE: I see no guidelines on the length of an extended abstract. I suggest
that we interpret this to mean about two pages. - Rao Vemuri

NOTE: Guidelines on the paper format and visual aids will be sent to you
along with the acceptance letter from the conference committee.

NOTE: Papers with a strong applications orientation, case studies,
implementations, and tutorial topics  dealing with "hands-on" topics,
implementations, and standards are encouraged.

TUTORIALS

Many pre-conference and post-conference tutorials are being
planned. If you have a proposal for a tutorial, please submit
your proposals by

15 June 1994

to the Technical Program Chair (USA) or the Technical Program Chair(INDIA)

PROGRAM COMMITTEE

GENERAL CHAIR:

DR. E. BHAGIRADHA RAO (INDIA)
ENGINEERING STAFF COLLEGE OF INDIA
KHAIRATABAD, HYDERABAD- 500 004, INDIA
EMAIL: ebrao%engsci@uunet.in

Technical Program Co-Chair (INDIA)

Dr. Vijay P. Bhatkar
Director, Center for Development of Advanced Computing
Poona University Campus
PUNE 411 007, INDIA
Email: bhatkar@parcom.ernet.in

Technical Program Co-Chair (USA)

Dr. M. Safiuddin
President, STS International
2451 Wehrle Building, Suite F
Williamsville, NY 14211, USA
Tel: (716) 688-2056
Fax: (716) 688-2056

or,
Technical Program Volunteer (USA)

Dr. V. Rao Vemuri
Professor, Dept. of Applied Science
University of California at Davis
Livermore, CA 94550
Tel: (510) 424-4051
Fax: (510) 422-8681
Email: vemuri@icdc.llnl.gov

IF YOU ARE INTERESTED IN ATTENDING, PRESENTING A PAPER, ORGANIZING
A SESSION OR CONDUCTING A TUTORIAL, PLEASE CONTACT ANY OF THE ABOVE
INDIVIDUALS.

PLEASE PASS THIS INFORMATION TO YOUR FRIENDS. WE WOULD LIKE TO SEE
A LARGE INTERNATIONAL PARTICIPATION. START WORKING ON YOUR ABSTRACT NOW.

HyDERABAD:

Hyderabad is about 700 km from Bombay or Madras and about 1600 km from
Delhi. Goa, a popular tourist attraction is about an hour away by air.

January weather in Hyderabad is very pleasant. It is cool with no rain.
Temperatures range from 27C (81F) to 14C (57F).

If you intend to travel in N. India, it will be cold.

REGISTRATION FEES AND HOTELS:
At this time, I do not have this information. Generally the registration fees
at Indian conferences are nominal (from the US standpoint). But I will send
this information to you as soon as I get it.

INTERNATIONAL CONFERENCE COMMITTEE

V. J. Bhatkar, C-DAC, Puna, India, Conference Co-Chair
M. Safiuddin, STS International, New York, USA, Co-chair
B. D. O. Anderson, Canberra, Australia
K. Baheti, NSF, USA
A. van Cauwenberghe, Zwijnaarde, Belgium
P. Deshpande, Louisville, USA
U. M. Gupta, Saskatchewan, Canada
A. Hossain, W. Lafayette, USA
K. Iizuka, Japan
J. Jaaksoo, Tallinn, Estonia
B. V. Jayawant, Sussex, U. K.
C. R. Kasarbada, Hyderabad, India
H. Kwakernak, The Netherlands
J. K. Pal, New Delhi, India
S. Parthasarathy, France
D. Popovic, Bremen, Germany
C. Satish, Hyderabad, India
M. Sundararajan, Singapore
S. G. Tsafestas, Athens, Greece
V. I. Utkin, Moscow, Russia
M. Vidyasagar, Bangalore, India
N. Viswanathan, Bangalore, India
K. Warwick, Reading, U. K.
T. Yoshikawa, Japan
S. Yurkovich, Columbus, OH, USA

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

