Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!rutgers!usc!ucsd!ucbvax!tut.cis.ohio-state.edu!cwjcc!ukma!xanth!ginosko!uunet!ncrlnk!wright!joh From: joh@wright.EDU (Jae Chan Oh) Newsgroups: comp.ai Subject: Re: GA's, references, this 'n that, etc... Summary: Pattern Recognition using GAs Message-ID: <590@thor.wright.EDU> Date: 14 Jul 89 07:26:35 GMT References: <1020@cb.ecn.purdue.edu> <EYeUmVy00V46I12El4@andrew.cmu.edu> Organization: Wright State University, Dayton OH, 45435 Lines: 80 In article <EYeUmVy00V46I12El4@andrew.cmu.edu>, dg1v+@andrew.cmu.edu (David Greene) writes: > This post contains: > 1. an address for a Genetic Algorithm mailing list/ bboard > 2. a list of general GA references > 3. a response to: 26-Jun-89 Discover Patterns in GAs Ioannis > Androulakis@cb.e > ... original text has been removed between the above and the below... > 3) > *Excerpts from ext.nn.comp.ai: 26-Jun-89 Discover Patterns in GAs Ioannis* > > *Androulakis@cb.e (620)* > > I would like to know if there has been any work done attempting > > to discover patterns in GA search. My basic concern is how I can > > draw knowledge from the system and maybe I could do that if I > > were able to study how patterns behave, while they are formed. > > This might contradict the basic notion of "implicit parallelism" > > in GA search, but could also help in understanding why a particular > > system evolved that way or another. > > > This is rather unclear as to what patterns you are looking for (ie. patterns in > some underlying object of study or patterns in the GA search itself). For the > former, you'll find many references to pattern recognition problems including > vision systems (see above sources). For the pattern recognition and discovery using genetic algorithms, there are several works one might want to consider. 1) Englander, A. C., ``Machine learning of visual recognition using genetic algorithms'', First G.A. conference proceedings. 1985 2) Schaffer, J. D., ``Some Experiments in Machine Learning Using Vector Evaluated Genetic Algorithms'', Ph. D. Dissertation, Dept. of Electrical Engineering, Vanderbilt University. 3) Gillies, A. M. ``Machine Learning Procedures for Generating Image Domain Feature Detectors'', Computer and Communication Science in The University of Michigan. And one can find some of works from The second G.A. conference by Schaffer and others. I, myself have done some research on image learning and classifications using genetic algorithms for my thesis and published couple of papers. I used Holland type classifier system architecture to attack the problem domain that deals with large class image learning and classification. To accomplish the task, I needed to add some new strategies that prevent good rules from being replaced (some rules can be good to classify some objects but the same rules can be bad for the other images. In this case, we may want to keep the rules.) and the techniques that can delve into the interested area to distinguish similar image objects but in different classes. I have succeeded on training the system to learn and recognize up to 26 different classes (all the alphabet images) with the new strategies. Currently, I'm starting to use CRAY-XMP to perform bigger problems such as 50 classes or even more. If anyone interested in my work, one could contact my department for the copy of the thesis or request an inter-library loan. The thesis is entitled as ``Improved Classifier System Using Genetic Algorithms Applied to Image Learning''. One can send e-mail to `nblair@wright.edu' or contact the Department of Computer Science and Engineering, Wright State Univ., Dayton, OH, 45435. One of my paper on a conference entitled ``Image Learning Classifier System Using Genetic Algorithms'' by McAulay and Oh in the proceedings of IEEE NAECON conference 89' might be interesting too. Hope this helps. -- Jae Chan Oh (Rm. 109, Computer Sci. Dept.) | Disclaimer : Wright State University Research Building | All mine ... 3171 Research Blvd., Kettering, Ohio 45420 | As far as I CSNET: joh@CS.wright.EDU UUCP: ...!osu-cis!wright!joh | type..