applications of soft computing – some personal reflections

2
E.P. Klement Fuzzy Logic Laboratorium Linz-Hagenberg Institut fu ¨ r Mathematik Johannes Kepler Universita ¨ t A-4040 Linz, Austria Original paper Soft Computing 2 (1998) 16 17 ( Springer-Verlag 1998 Applications of soft computing some personal reflections E.P. Klement Abstract Some experiences of the cooperation of the Fuzzy Logic Laboratorium Linz-Hagenberg (Austria) with industrial partners are described. Special emphasis is given to the intelligent combination of various mathematical methods. 1 Industrial cooperation Cooperating with the industry, as it is done quite sucessfully by the FLLL (Fuzzy Logic Laboratorium Linz-Hagenberg), a division of the Department of Mathematics of the Johannes Kepler University in Linz, for a number of years with Austrian and international companies, is quite different a task as compared with doing research, even with doing applied research. This little note therefore intends to discuss some of our experience in such cooperations in a way which may be useful for other colleagues in the field. The main goal of an industry-university cooperation is usually f to solve a real world problem f in the best possible way, f but also at the lowest possible cost, therefore making usable the scientific knowledge and power of a research institution to a commercial company. A strong motivation (and, therefore, almost a necessary condition) for a company to look for university support is that the problem they have hurts them severely (in terms of high costs or low quality or both), and that it cannot be solved within the company itself. There are many other conditions to be met in such a projects which shall be discussed below. Issues such as tractability and robustness of the solution provided play an important role. Issues which are crucial for discussions within the scientific community such as theoretical elegance, priority, or simply how to name an object, play a much smaller role. This is the reason why we shall not enter here the (albeit necessary) discussion what soft computing should be ideally or whether it should better be called computational intelligence. 2 FLLL experience Since 1993, the FLLL has been involved in a number of projects, most of which resulted in software packages for quality control (see, e.g., [1][10]). In each of these expert systems, the intelligent interpretation of signals was the key issue. The signals we dealt with so far are of very different type, including f optical data (images), f music data, f acoustic data, f data from various measurements. The industrial partners include international companies such as SONY or the Swedish paper company SCA, but also smaller Austrian enterprises. More recently, we also started a control project with an electric power plant on the Danube river. 3 Shift of paradigm The last twenty years or so witnessed a remarkable shift of paradigm. Formerly, the general point of view was that everything (most prominent example: the path of a satellite) could be computed at arbitrary precision provided one has f enough and sufficiently precise data, f sufficient computer power. Many (physical or mathematical) models proved to be so successful that people used larger and larger models (a good example are the world models as used, e.g. by the Club of Rome), expecting better and better results. Unfortunately, this was not always the case. The reasons for this are manyfold: sometimes the necessary data simply are not available, quite often the models themselves are rather coarse approximations of reality or the reality simply does not behave exactly as the (somehow ideal) models would suggest (key word chaos). As a consequence, now it is at least partially acknowledged that in many real situations f precision is sometimes artificial (it is usually not necessary to measure the temperature in a room up to 1/1000 of a centigrade), f precision is always expensive (high-precision measurement devices cost a lot of money), 16

Upload: e-p-klement

Post on 14-Jul-2016

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Applications of soft computing – some personal reflections

E.P. KlementFuzzy Logic Laboratorium Linz-HagenbergInstitut fur MathematikJohannes Kepler UniversitatA-4040 Linz, Austria

Original paper Soft Computing 2 (1998) 16—17 ( Springer-Verlag 1998

Applications of soft computing — some personal reflectionsE.P. Klement

Abstract Some experiences of the cooperation of the FuzzyLogic Laboratorium Linz-Hagenberg (Austria) with industrialpartners are described. Special emphasis is given to theintelligent combination of various mathematical methods.

1Industrial cooperationCooperating with the industry, as it is done quite sucessfully bythe FLLL (Fuzzy Logic Laboratorium Linz-Hagenberg),a division of the Department of Mathematics of the JohannesKepler University in Linz, for a number of years with Austrianand international companies, is quite different a task ascompared with doing research, even with doing appliedresearch.

This little note therefore intends to discuss some of ourexperience in such cooperations in a way which may be usefulfor other colleagues in the field.

The main goal of an industry-university cooperation isusually

f to solve a real world problemf in the best possible way,f but also at the lowest possible cost,

therefore making usable the scientific knowledge and power ofa research institution to a commercial company.

A strong motivation (and, therefore, almost a necessarycondition) for a company to look for university support is thatthe problem they have hurts them severely (in terms of highcosts or low quality or both), and that it cannot be solvedwithin the company itself.

There are many other conditions to be met in such a projectswhich shall be discussed below. Issues such as tractability androbustness of the solution provided play an important role.

Issues which are crucial for discussions within the scientificcommunity such as theoretical elegance, priority, or simplyhow to name an object, play a much smaller role. This is thereason why we shall not enter here the (albeit necessary)discussion what soft computing should be ideally or whether itshould better be called computational intelligence.

2FLLL experienceSince 1993, the FLLL has been involved in a number ofprojects, most of which resulted in software packages forquality control (see, e.g., [1]—[10]).

In each of these expert systems, the intelligent interpretationof signals was the key issue. The signals we dealt with so far areof very different type, including

f optical data (images),f music data,f acoustic data,f data from various measurements.

The industrial partners include international companiessuch as SONY or the Swedish paper company SCA, but alsosmaller Austrian enterprises.

More recently, we also started a control project with anelectric power plant on the Danube river.

3Shift of paradigmThe last twenty years or so witnessed a remarkable shift ofparadigm. Formerly, the general point of view was thateverything (most prominent example: the path of a satellite)could be computed at arbitrary precision provided one has

f enough and sufficiently precise data,f sufficient computer power.

Many (physical or mathematical) models proved to be sosuccessful that people used larger and larger models (a goodexample are the world models as used, e.g. by the Club ofRome), expecting better and better results.

Unfortunately, this was not always the case. The reasons forthis are manyfold: sometimes the necessary data simply are notavailable, quite often the models themselves are rather coarseapproximations of reality or the reality simply does not behaveexactly as the (somehow ideal) models would suggest (keyword chaos).

As a consequence, now it is at least partially acknowledgedthat in many real situations

f precision is sometimes artificial (it is usually not necessaryto measure the temperature in a room up to 1/1000 ofa centigrade),f precision is always expensive (high-precision measurementdevices cost a lot of money),

16

Page 2: Applications of soft computing – some personal reflections

f precision and complete information are not always neededto make proper decisions (in human decision making theavailable information is rarely precise and often quiteincomplete).

This means, however, that the standard analytical andstatistical methods alone are not always adequate to dealwith such problems. It often requires entirely newmethods, and more often a clever combination of severalmethods.

4Combination of methodsOne important aspect of commercial partners is their anti-fundamentalistic attitude concerning the use of specificmethods. Usually, the commercial partners will not care whichmethod is used to solve a given problem, as long as the solutionis adequate (in the sense of the criteria mentioned earlier, i.e.,concerning quality and costs) and in some sense optimal.

According to our experience, no company will paya university partner just for using method A or algorithmB — the only thing which matters is the quality of the solutionprovided (and the cost involved, of course).

This makes it necessary to search for the best method whichis available under the given restrictions (it may very well bethat this method is still to be invented, making out a lot of thescientific satisfaction connected with the solution of real worldproblems).

As already mentioned, there are usually a number ofrestrictions and limitations in such problems. The mostimportant ones concern

f time: development and computation time, 2f complexity: amount of data, algorithmic complexity, 2f disturbations: noisy data, instabilities, 2f hardware: memory space, speed, 2f money: development cost, hardware cost, 2

This and the high complexity of a typical real problem quiteoften implies that the final method is itself a combination ofseveral methods.

Here the techniques which often are considered to be in thecore of soft computing, namely, fuzzy logic, neural networks,and genetic algorithms [12, 13], can play a key role.

Their flexibility and adaptivity allows them to be combinedamong themselves, with more classical ones such as

f analytical methods,f algebraic methods,f numerical methods,f stochastic methods,

but also, e.g., with machine learning.Each of these methods has its own advantages and

disadvantages, and it requires a broad knowledge and some

experience to find out the adequate mix to solve a givenproblem.

5Confidence and prototypingThe absolutely necessary basis for a successful cooperation,however, is that the company has sufficient confidence in thequality and professionality of the university partner. This isbest proven by reference projects (if they exist).

In our experience, a strategy of evolutionary prototyping,starting with a feasibilty study and defining milestones fordecisions whether or how to continue, was always a good wayto build up a good level of confidence step by step.

References1. Bauer P, Bodenhofer U, Klement EP (1996) A fuzzy system for

image pixel classification and its genetic optimization. In: R.Trappl (ed.), Cybernetics and Systems ’96. Austrian Society forCybernetic Studies, Wien, Vol. 1, pp 285—290

2. Bauer P, Bodenhofer U, Klement EP (1996) A fuzzy method forpixel classification and its application to print inspection. In:Proceedings IPMU ’96. Information Processing and Managementof Uncertainty in Knowledge-Based Systems, Granada, Vol. III,pp 1301—1305

3. Bauer P, Bodenhofer U, Klement EP (1996) A fuzzy algorithm forpixel classification based on the discrepancy norm. In:Proceedings Fifth IEEE International Conference on FuzzySystems FUZZ-IEEE ’96, New Orleans. IEEE, Piscataway, N.J.,Vol. 3, pp 2007—2012

4. Bauer P, Klement EP (1996) A self-tuning fuzzy system for smoothcurve fitting. In: T. Yamakawa, G. Matsumoto (eds.),Methodologies for the Conception, Design, and Application ofIntelligent Systems. World Scientific, Singapore, Vol. I, pp 452—455

5. Bauer P, Klement EP, Leikermoser A, Moser B (1993)Approximation of real functions by rule bases. In: ProceedingsFifth IFSA World Congress ’93, Seoul, pp 239—241

6. Bauer P, Klement EP, Leikermoser A, Moser B On minimizing thenumber of rules of Sugeno controllers. Submitted for publication

7. Bodenhofer U, Klement EP (1997) Pixel classification: A fuzzy-genetic approach. In: Proceedings Seventh IFSA World Congress’97, Prague. Vol. IV, pp 38—43. Academia, Prag

8. Bodenhofer U, Klement EP Genetic optimization of fuzzyclassification systems — A case study. Submitted for publication

9. Klement EP, Koczy LT, Moser B (1995) Approximation and fuzzycontrol. In: Proceedings Sixth IFSA World Congress ’95, Sa8 oPaulo, Vol. I, pp 625—628

10. Klement EP, Koczy LT, Moser B Are fuzzy systems universalapproximators? Intl. J. Gen. Syst. (in press)

11. Klement EP, Slany W (1996) Fuzzy logic in artificial intelligence.In: A. Kent, J.G. Williams (eds.), Encyclopedia of ComputerScience and Technology, Vol. 34, Suppl. 19, Marcel Dekker, NewYork, pp 179—190

12. Kruse R, Gebhardt J, Klawonn F (1994) Foundations of FuzzySystems, (J. Wiley & Sons, Chichester)

13. Zimmermann H-J Fuzzy Set Theory and its Applications (Kluwer,Boston, 1991)

17