KDD Nuggets 95:21, e-mailed 95-09-05 Contents: * KDD-95, KDD-95 Proceedings available from AAAI Press http://www.aaai.org/Publications/Press/Catalog/uthurusamy.html * GPS, Montreal Gazette on KDD-95 * X. Wu, Opinion: Expert Systems and KDD * A. Freitas, Parallel KDD Technical Report * T. Hu, Ph.d thesis: KDD: An Attribute-Oriented Rough Set Approach Job Ad: * S. Troccolo, US-CA-Palo Alto-AI jobs/Data mining/programming * E. Rigdon, Session on Data Mining at Fall 1996 INFORMS, Nov 96 * C. Kadie, Predicting Learning Curves (Ph.D. thesis, online) * D. Martin, IBM's data mining home page: http://www.almaden.ibm.com/cgi-bin/stss/get/data-mining/ The KDD Nuggets is a moderated mailing list for information relevant to Data Mining and Knowledge Discovery in Databases (KDD). Please include a DESCRIPTIVE subject line and a URL, when available, in your submission. Nuggets frequency is approximately weekly. Back issues of Nuggets, a catalog of S*i*ftware (data mining tools), references, FAQ, and other KDD-related information are available at Knowledge Discovery Mine, URL http://info.gte.com/~kdd by anonymous ftp to ftp.gte.com, cd /pub/kdd, get README E-mail add/delete requests to kdd-request@gte.com E-mail contributions to kdd@gte.com -- Gregory Piatetsky-Shapiro (moderator) ********************* Official disclaimer *********************************** * All opinions expressed herein are those of the writers (or the moderator) * * and not necessarily of their respective employers (or GTE Laboratories) * ***************************************************************************** ~~~~~~~~~~~~ Quotable Quote ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Thanks to Gavin Meggs for compiling this selection of quotable quotes from KDD'95 conference. GPS --------------- Tej Anand: "Discovery of knowledge is not an algorithmic endeavor. It is a human endeavor." David Haussler: "[the human genome project] gives us an unprecedented opportunity to peek at what makes us what we are." Ramasamy Uthurusamy: "[following the KDD process without using feedback is] like hitting golf balls at night" Ramasamy Uthurusamy: "While mining beware of mines" Gregory Piatetsky-Shapiro: "Data Mining is Statistics + Marketing" >~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Date: Thu, 24 Aug 95 09:05:04 PDT From: kdd95@aig.jpl.nasa.gov (KDD-95 Account) To: kdd@gte.com Subject: KDD95 Proceedings available from AAAI Press Proceedings of the First International Conference on Knowledge Discovery and Data Mining (KDD-95) Edited by Usama M. Fayyad and Ramasamy Uthurusamy Published by the AAAI Press 348 pp. $50.00 paperback ISBN 0-929280-82-2 Further information about this book, including contents can be obtained from http://www.aaai.org/Publications/Press/Catalog/uthurusamy.html for AAAI Press see [http://www.aaai.org] See http://www-aig.jpl.nasa.gov:80/kdd95/finalprogram.html for the conference contents. ---------- US Shipping Rates Add $4.50 for first book Add $1.00 each additional book Shipping Rates Outside USA: Add $5.50 per book for surface mail Add $16.50 per book for airmail Sales Tax: California residents add 7.75% sales tax. AAAI members in good standing may deduct 20% from the list price of the book. (No discount allowed on tax or shipping, however.) To order, call (415) 328-3123 or fax (415/328-3123) or mail: AAAI 445 Burgess Drive Menlo Park, CA 94025 MasterCard and Visa accepted. An online order form will be available later on this summer. Books will be shipped in late August or early September, 1995. >~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Date: Tue, 29 Aug 1995 16:00:21 -0400 From: gps@gte.com (Gregory Piatetsky-Shapiro) Subject: Montreal Gazette: Knowledge Nuggets Montreal Gazette, Tuesday, August 22, 1995 (page A3) had a story on the KDD-95 and IJCAI-95, which was taking place at Palais de Congres in Montreal. I, and several other people at KDD-95, were interviewed for the story, entitled "KNOWLEDGE NUGGETS", and subtitled HUGE DATA BANKS ARE FERTILE GROUND FOR CYBER-MINERS by Lynn Moore, Montreal Gazette. The electronic age of bar coders and data banks has created fertile ground for a new breed of miners who work, not with pick and shovel, but with sophisticated programs and computers with artificial intelligence. They are researchers engaged in what high-tech folks call knowledge discovery and data mining, or KDDM, and many of the prominent ones are in Montreal this week strutting their stuff and calculating the payoff. "There is a lot of false gold (in the huge data banks that have been compiled in the past decade), but there is also some very valuable stuff,", Gregory Piatetsky-Shapiro, a GTE Laboratories Inc., scientist said yesterday. And the gold rush is on. Computer giant IBM "believes there is a billion-dollar market for them" in selling KDDM hardware and software, he said. Other private companies and government agencies have already dug in, anxious to find -- in relatively short order -- useful information in huge data banks, which contain trillions and trillions of bits of information, he said. The U.S. National Aeronautics and Space Administration, for instance, has so many visual images of the billions of objects in space that "graduate students could spend several hundred years looking at all of them" to prepare for classification and analysis. NASA recently developed a KDDM system that quickly classified the objects and, in process, found several new quasars. The U.S. Department of the Treasury is employing high-tech miners to identify potential money laundering. Since its system began processing about 200,000 suspect financial transactions a week in March 1993, it has directed human investigators to over $1 billion in potentially laundered funds and led to one criminal conviction. Private companies are also tilling their data bases looking for rules that predict good customers -- and bad ones -- as well as digging into mountains of data in search of "knowledge nuggets" that would be helpful for their marketing departments, Piatetsky-Shapiro and other researchers said. Alex Tuzhilin and a research partner at the Stern School of Business at New York University have been working for about a year on creating a system whereby computers can be prepared for knowledge mining by being programmed with a number of beliefs and then told to look for something interesting in relation to those beliefs. Foe example, a telephone company might believe most of its customers make lengthy long-distance calls when rates are lowest and that people tend to repeat their long-distance calls to the same numbers over the time, he said. Using those beliefs as tools, the computer could dig up material for service providers and marketing people, he said. Piatetsky-Shapiro, who has worked with data banks containing medical information, conceded that one area of concern in high-tech mining revolves around privacy issues. "Many types of data mining do not affect personal data ... but companies will have to be careful", he said. The venue for this cyber-mining convention is the Palais des Congres, where about 2,500 scientists and researchers from around the world have gathered for a series of cutting-edge conferences. In addition to the First International Conference on Knowledge Discovery and Data Mining, there is the Seventh Annual Conference on Innovative Applications of Artificial Intelligence and the 14th International Joint Conference on Artificial Intelligence. >~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ From: xindong@insect.sd.monash.edu.au Date: Mon, 28 Aug 1995 18:01:11 +1000 (EST) Subject: Expert Systems and KDD To: kdd@gte.com X-Mailer: ELM [version 2.4 PL23] Content-Transfer-Encoding: 7BIT Content-Type: text Content-Length: 5530 The issue of an expert system component in knowledge discovery and data mining (KDD) was raised at the KDD-95 conference, and I would like to explore this topic a bit further here. KDD is more knowledge processing than data retrieving. Establishing the link between the KDD community and the database community is certainly important and this will surely help the growing of the KDD community. However, over-emphasizing the importance of database technology and belittling the expert system component in KDD are both inappropriate. Database technology is more concerned with storing and retrieving data than handling/discovering knowledge. Statements like ``data mining = database and/or knowledge base querying'' are a good indication of some database people's misunderstanding of KDD. KDD deals with knowledge discovery in databases. It starts with databases and can well take advantage of existing database technology, but its focus is on mining data and discovering knowledge. On the contrary, expert systems are knowledge processing systems, and people in this field distinguish knowledge processing from data processing. Along with the recognition of the so called knowledge bottleneck problem [1] in transforming knowledge from human experts to knowledge-based systems, research on automatic knowledge acquisition from databases has been expanding rapidly in recent years in the expert systems community. Expert systems research has also resulted in many significant results and topics for KDD research, such as structured induction [2] and noise handling at different KDD stages [3]. I agree that not every KDD researcher needs to use expert systems or has to be familiar with expert systems technology, but this is also the case with database technology. For large, realistic KDD applications, both technologies are important. A very useful point from Hayes-Roth and Jacobstein [4] about knowledge-based systems, which I believe applies to KDD perfectly well, is that ``Practical experience has made it clear that KBS are rarely more than 20% of a complete solution to real-world problems'' ([4], page 37). [ In my experience, KBS part in KDD systems is perhaps 5% or less of the total system, but it is the critical part, like the human brain, which is also less than 5% of weight of the body. -- GPS ] To be effective, both KBS and KDD must be integrated into those technologies which have already found wide application in many fields. In line with this point, I have been exploring for several years [5,6,7] a concept called an intelligent learning database (ILDB) system. I define an integrated learning system, which combines machine learning techniques with both database and knowledge base technology, as an intelligent learning database (ILDB) system if it provides mechanisms for: 1. Translating standard (e.g., relational) database information into a form suitable for use by its induction engines. 2. Using induction techniques to extract rules from databases. 3. Interpreting the rules produced to solve users' problems. With an ILDB system, one can, for example, produce a small number of conjunctive rules for some diseases from a large medical database of these diseases. The ILDB system can then use the rules in two different ways: keeping these rules instead of the original cases because the original cases might take up a large space, and using these rules to diagnose new cases. ([5], pages 9-10) Generally speaking, ILDB systems have wide potential application to various classification problems, such as diagnosis, pattern recognition, prediction, and decision making, where there are large amounts of data sets to be sorted. In pattern recognition, for example, which is concerned with the identification of objects by means of their features, there are two problems to be solved: (a) acquiring and representing the features for each object pattern at first, and (b) finding out which object pattern is most similar to the unknown object in question. One can of course use statistical methods to describe the features of patterns and unknown objects and identify an unknown object by computing the distances between its features and those of known objects. However, ILDB systems research provides a feasible way to both acquire the features of known patterns by using machine learning techniques and identify unknown objects by logical inference on the acquired features. ([5], pages 13-14) References [1] E.A. Feigenbaum, Expert Systems in the 1980s, Infotech State of the Art Report on Machine Intelligence, A. Bond (Ed.), Maidenhead: Pergamon--Infotech, 1981. [2] A.D. Shapiro, Structured Induction in Expert Systems, Turing Institute Press in association with Addison-Wesley, 1987. [3] X. Wu, J. Krisar, and P. Mahlen, Noise Handling with Extension Matrices, Proceedings of the 7th IEEE International Conference on Tools With Artificial Intelligence, Hyatt Dulles, Washington, D.C., USA, November 5-8, 1995. [4] F. Hayes-Roth and N. Jacobstein, The State of Knowledge-Based Systems, Communications of the ACM, 37(1994), 3: 27--39. [5] X. Wu, Knowledge Acquisition from Databases, 211 + xvi pp., Ablex Publishing Corp., U.S.A., 1995 (forthcoming). [6] X. Wu, Inductive Learning: Algorithms and Frontiers, Artificial Intelligence Review, 7(1993), 2: 93--108. [7] X. Wu, KEshell2: An Intelligent Learning Data Base System, Research and Development in Expert Systems IX, M.A. Bramer and R.W. Milne (Eds.), Cambridge University Press, U.K., 1992, 253--272. >~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ From: Freitas A A Date: Mon, 21 Aug 95 19:51:39 BST To: gps@gte.com Subject: Parallel KDD Technical Report Content-Type: text Content-Length: 1763 Dear Prof. Piatetsky-Shapiro, Maybe you could announce in KDD Nuggets (and/or make available in your Knowledge Discovery Mine) that the Technical Report "A data-parallel primitive for high-performance knowledge discovery in large databases" is available by ftp from: ftp.essex.ac.uk cd pub/csc/technical-reports get CSM-242.ps.Z The Report has 20 pages (in double space), and its abstract is as follows. "Efficiency is crucial in KDD (Knowledge Discovery in Databases), due to the huge amount of data stores in current databases. We argue that high efficiency in KDD can be achieved by combining two approaches, namely encapsulating KDD functionally within standard DBMS operations and using parallel processing. Hence, KDD tasks can be executed on a back-end SQL server, e.g. a parallel DB machine. We propose a KDD primitive (a set of basic operations) which underlies the candidate-rule evaluation procedures of many KDD algorithms. We compare and analyse the time required to carry out this primitive on three different computational architectures, viz. a conventional workstation and two parallel DB machines. The main advantages of encapsulating a KDD primitive in a parallel DB server are automatic parallelization and the run-time speed which can be achieved through parallel processing." Contact: Alex Alves Freitas (freial@essex.ac.uk) Thanks, Alex ================================================== Alex Alves Freitas University of Essex Dept. of Computer Science, PhD student Wivenhoe Park, Colchester, CO4 3SQ, United Kingdom phone: (44) (01206) 87-2290 e-mail: freial@essex.ac.uk ================================================== >~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Date: Fri, 25 Aug 1995 09:21:00 -0400 From: "tony (t.) hu" To: gps@gte.com, kdd@gte.com Subject: KDD Ph.d thesis abstract Xiaohua Hu Dept. of Computer Science, Univ. of Regina, Canada xiaohua@cs.uregina.ca (univ), tonyhu@bnr.ca (company) Doctoral thesis: Knowledge Discovery in Databases: An Attribute-Oriented Rough Set Approach June 1995 Supervisor: Dr. Nick Cercone Thesis Outline This dissertation describes knowledge discovery in databases using rough set theory in three aspects: (1) In this thesis, we develop an attribute-oriented rough set approach for knowledge discovery in databases. The method adopts the ``learning from examples" paradigm combined with rough set theory and database operations. The learning procedure consists of two phases: data generalization and data reduction. In data generalization, our method generalizes the data by performing attribute-oriented concept tree ascension, thus some undesirable attributes are removed and a set of tuples may be generalized to the same generalized tuple. The generalized relation contains only a small number of tuples, which substantially reduces the computational complexity of the learning process and, furthermore, it is feasible to apply the rough set techniques to eliminate the irrelevant or unimportant attributes and choose the ``best" minimal attribute set. The goal of data reduction is to find a minimal subset of interesting attributes that have all the essential information of the generalized relation; thus the minimal subset of the attributes can be used rather than the entire attribute set of the generalized relation. By removing those attributes which are not important and/or essential , the rules generated are more concise and efficacious. (2) A generalized rough set model is formally defined with the ability to handle statistical information and also consider the importance of attributes and objects in the databases. We construct a theoretical model to explain the mechanism of multiple knowledge bases (or redundant knowledge) in the context of rough sets theory . A decision matrix is used to construct a multiple knowledge bases system in a dynamic environment. (3) Our method integrates a variety of knowledge discovery algorithms, such as DBChar for deriving characteristic rules, DBClass for classification rules, DBDeci for decision rules, DBMaxi for maximal generalized rules, DBMkbs for multiple sets of knowledge rules and DBTrend for data trend regularities, which permit a user to discover various kinds of relationships and regularities in the data. This integration inherits the advantages of the attribute-oriented induction model and rough set theory. A prototype system DBROUGH was constructed under a Unix/C/Sybase environment. Our system implements a number of novel ideas. In our system, we use attribute-oriented induction rather than tuple-oriented induction, thus greatly improving the learning efficiency. By integrating rough set techniques into the learning procedure, the derived knowledge rules are particularly concise and pertinent, since only the relevant and/or important attributes (factors) to the learning task are considered. In our system, the combination of transition network and concept hierarchy provides a nice mechanism to handle dynamic characteristic of data in the databases. For applications with noisy data, our system can generate multiple sets of knowledge rules through a decision matrix to improve the learning accuracy. The experiments using the NSERC information system and some other experimental data set illustrate the promise of attribute-oriented rough set learning for knowledge discovery in databases. The thesis contains nine chapters organized as follows: An introduction is described in Chapter 1. An overview of the current knowledge discovery systems are discussed in Chapter 2 and several typical systems such as ID3, the AQ family, the KDW workbench, INLEN and ITRULE are briefly discussed. We describe in Chapter 3 an attribute-oriented induction system (DBLEARN) and our extension to this system. In Chapter 4, the general concept of a rough set is introduced and a general rough set model is proposed to handle uncertainty and vague information in databases. Chapter 5 is devoted to rough set based data reduction, along with some illustrative examples. Multiple sets of knowledge rules and a proposed decision matrix approach to constructing multiple sets of knowledge rules are the topic of Chapter 6. In Chapter 7 the experimental results of our system using the NSERC information system (Natural Science and Engineering Research Council of Canada) and some other test data such as IRIS Data, Appendicitis Data, Thyroid Data , are presented and demonstrated and a discussion of our methods is given in Chapter 8. Some concluding remarks are presented in Chapter 9 with a summary of the major thesis findings and with suggestions about the directions for future progress. ========================================================================== If you want a copy of the thesis, please send an e-mail to tonyhu@bnr.ca or xiaohua@cs.uregina.ca, then I will send you a postcript version of the thesis >~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ From: auguri@ix.netcom.com (Susan R. Troccolo ) Newsgroups: ba.jobs.offered,ba.jobs.misc.,misc.jobs.offered,su.jobs,ucb.jobs,com p.lang.prolog,comp.lang.tcl Subject: US-CA-Palo Alto-AI jobs/Data mining/programming Date: 27 Aug 1995 23:14:50 GMT Organization: Netcom Positions Available - AI jobs/Data mining/programming Job code: BL/Recon Join a dynamic team in Palo Alto, CA, developing leading-edge Data Mining software to extract patterns and other useful information from large quantities of of data. The software integrates deductive database technology with several machine learning techniques. This position requires a can-do attitude and a keen interest in data mining and related technologies. Requirements: We have employment opportunities in several disciplines. We are looking for motivated programmers, researchers and data analysts from new graduates to seasoned professionals. Desired skills include: machine learning, applied statistics, SAS programming, deductive databases, Prolog programming, relational database management systems, UNIX and C programming and GUI development in Tcl/Tk. WHO ARE WE? The Research and Software Technology Divisions of a Fortune 20 Company are seeking bright and talented Software professionals to add to their workforce on a Full-time basis. We have had an influx of project work and have both junior and senior positions available. If you are a recent graduate,but have several years of proven programming skills and are now interested in working in a research setting, we want to talk with you!!! *WHAT IS OUR WORKPLACE LIKE? We are researchers in an entrepreneurial environment. You will have flexibility to make a lot of your own goals - and work with other talented people in nice offices in Palo Alto. We require permanent residency status for most of our work. However,if you have an F1 (Educational visa), you must have ONE YEAR MINIMUM remaining on your practical training. NO EXCEPTIONS. *HOW TO REACH US? For immediate consideration, fax or Email your resume to my attention. PLEASE REMEMBER TO REFERENCE THE JOB CODE and let us know the best times and places to reach you. We will review your background and contact you to discuss the positions available. Send your resume via Email to: Susan Troccolo (ASCII or WinWord format) DANTE Associates Auguri@ix.netcom.com Via FAX on HIGH RES to: (415) 326-5969 or (415) 354-5235 No phone calls please. Principals only. >~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Date: Mon, 28 Aug 1995 17:02:32 -0500 From: ED RIGDON To: kdd@gte.com Subject: Session on Data Mining Content-Type: text Content-Length: 1724 I am a member of the Marketing faculty at Georgia State University, in Atlanta. Naveen Donthu, the source of the following message, is a colleague of mine: >>>>>>>>>>>>> >>> NAVEEN DONTHU 08/17/95 02:01pm >>> The Fall 1996 INFORMS (originally ORSA/TIMS) conference will be held in Atlanta from November 3-6, 1996. I will be chairing the Marketing track of this conference. We will probably have 9 to 10 special Marketing sessions. Please contact me if you would like to organize a special session for this conference. You will basically have 3 to 4 papers in each session. You pick a topic for your session (e.g., Pricing Models, or Location Models, or ....), and invite 3 to 4 researchers to present their work. If you are interested: 1. I need your name, address, phone and fax number, e-mail address, along with the tentative title for the session you propose to organize by AUGUST 25, 1995. 2. Names of speakers, title of their paper, and a short 50 word abstract for each paper in your session by OCTOBER 20, 1995. If you know of someone who might be interested in organizing a session, please forward this message to them. Thanks, Naveen Donthu >>>>>>>>>>> I believe that a session on data mining with applications in marketing would be a great addition to this conference. I am a novice in data mining, so I would be pleased to let someone else organize this session. However, I would also be pleased to serve as "bus driver" for this event. If someone would like to take responsibility for organizing this session, or has research on "data mining and applications in marketing" that they would like to present at INFORMS 96 in Atlanta, please contact me (erigdon@gsu.edu) or track organizer, Naveen Donthu, directly at mktnnd@gsu.edu --Ed Rigdon >~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ From kadie Mon Aug 28 15:45:12 CDT 1995 To: ai-stats@watstat.uwaterloo.ca Subject: Predicting Learning Curves (Ph.D. thesis, online) I'm happy to announce that my Ph.D. thesis Seer: Maximum Likelihood Regression for Learning-Speed Curves is available on-line via anonymous ftp from "ftp.cs.uiuc.edu", directory "pub/TechReports", file "UIUCDCS-R-95-1874.ps.Z". The URL is "ftp://ftp.cs.uiuc.edu/pub/TechReports/UIUCDCS-R-95-1874.ps.Z". The abstract is enclosed. - Carl ========================= Seer: Maximum Likelihood Regression for Learning-Speed Curves Carl Myers Kadie, Ph.D. Department of Computer Science University of Illinois at Urbana-Champaign, 1995 David C. Wilkins, Advisor The research presented here focuses on modeling machine-learning performance. The thesis introduces Seer, a system that generates empirical observations of classification-learning performance and then uses those observations to create statistical models. The models can be used to predict the number of training examples needed to achieve a desired level and the maximum accuracy possible given an unlimited number of training examples. Seer advances the state of the art with 1) models that embody the best constraints for classification learning and most useful parameters, 2) algorithms that efficiently find maximum-likelihood models, and 3) a demonstration on real-world data from three domains of a practicable application of such modeling. The first part of the thesis gives an overview of the requirements for a good maximum-likelihood model of classification-learning performance. Next, reasonable design choices for such models are explored. Selection among such models is a task of nonlinear programming, but by exploiting appropriate problem constraints, the task is reduced to a nonlinear regression task that can be solved with an efficient iterative algorithm. The latter part of the thesis describes almost 100 experiments in the domains of soybean disease, heart disease, and audiological problems. The tests show that Seer is excellent at characterizing learning-performance and that it seems to be as good as possible at predicting learning performance. Finally, recommendations for choosing a regression model for a particular situation are made and directions for further research are identified. ========================= -- Carl Kadie -- kadie@cs.uiuc.edu -- University of Illinois at Urbana-Champaign >~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Return-Path: From: (David C. Martin) Organization: IBM Scientific & Technical Solutions Email: dcmartin@almaden.ibm.com Phone: 408/927-2239 Tie: 457-2239 Fax: 408/927-3025 To: kdd-request@gte.com Subject: addition to next KDD Date: Wed, 30 Aug 95 09:01:13 -0800 Sender: Content-Type: text Content-Length: 486 Please announce the availability of IBM's data mining home page: http://www.almaden.ibm.com/cgi-bin/stss/get/data-mining/ Also, please make a link to our page from your page. Thank you. dcm --- David C. Martin mail: dcmartin@almaden.ibm.com IBM Almaden Research Center work: 408/927-2239 (457-2239) 650 Harry Road, PHP/802 fax: 408/927-3025 (457-3025) San Jose, California 95120 http: stss.almaden.ibm.com/~dcmartin >~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~