An algorithm and software for knowledge-base reduction in software computing expert systems

Technical cybernetics. Information technology. Computer facilities


Аuthors

Abdulkhakov A. R.

Kazan National Research Technical University named after A.N. Tupolev, KNRTU-KAI, 10, Karl Marks str., Kazan, 420111, Russia

e-mail: aidar_abdulhakov@mail.ru

Abstract

Expert systems are used to solve complex practical problems in different spheres of human activity. Methods and data mining algorithms are widely used to create knowledge-bases for expert systems. However, this mechanism often generates thousands of rules, which makes it difficult for experts to analyze and interpret data, and also makes knowledge-base rules excessive and very often conflicting with each other.
In order to improve the efficiency of expert systems’ usage, it is necessary to combine similar rules. This optimization is performed by structuring the data and minimizing the amount of rules without losing data completeness.
The search for similar rules is performed for rules with identical values of consequents. That is why we need to divide sets of the rules into disjoint subsets.
Let’s consider the fuzzy-production rule represented as the Takagi-Sugeno mode. We will proceed from the fuzzy sets to their quantitative value using a defuzzification procedure. This procedure uses the method of gravity center. Then we need to normalize all the values obtained. The resulting system of vectors can be used as points in the high dimensional Euclidean space. Our task is limited to solve the problem of taxonomy of these points. I chose the k-means algorithms to solve this task.
In the issue of a knowledge-base optimization, the criteria of optimal work can be the generalization error obtained by the expert system during the work on a selection of test data. The minimum value of the generalization error corresponds to the optimal cluster solution.
A series of experiments have been made to confirm a prior assumption of increasing effectiveness. For testing purposes random sets of data have been selected as «spam» and regular mail messages.
Series of experiments show that the classification accuracy of the system using optimized knowledge-base has improved in comparison with the initial results. Consequently, optimization of the knowledge-base has improved generalization capability of the expert system during the work with test set of data.

Keywords:

knowledge-base, expert system, reduction of fuzzy rules, knowledge-base optimization in expert systems

References

  1. Gavrilova T.А., Khoroshevskij V.F. Bazy znanij intellektual’nykh sistem (Knowledge Base of Intellectual Systems), Saint-Petersburg, Piter, 2001, 384 p.
  2. Zagorujko N.G. Prikladnye metody analiza dannykh i znanij (Applied methods of Data and Knowledge analysis), Novosibirsk, Institut matematiki, 1999, 270 p.
  3. Bukhnin А.V., Bazhanov Yu.S. Nejrokomp’yutery: razrabotka, primenenie, 2007, no. 11.
  4. Shhurevich E.V., Kryuchkova E.N. Vestnik Аltajskogo gosudarstvennogo tekhnicheskogo universiteta im. I.I. Polzunova, 2007, no. 2, pp. 173-177.
  5. Shhurevich E.V. Informatsionnye tekhnologii, 2009, no. 2, pp. 25-29.
  6. Takagi T., Sugeno M. Fuzzy identification of systems and its application to modeling and control, IEEE Transactions, Systems, Man and Cybernetics, 1985, Vol. 15, pp. 116-132.
  7. Jang J.R., Sun C.T. ANFIS: Adaptive-Network-based Fuzzy Inference Systems, IEEE Tranc. on Systems, Man and Cybernetics, 1993, vol. 23, pp. 665-685.
  8. Katasyov А.S., Kornilov G.S. Materialy 7oi mezhdunarodnoj nauchno-prakticheskoj konferentsii «Infokommunikatsionnye tekhnologii global’nogo informatsionnogo obshhestva», Kazan, 2009, pp. 507-512.

Download

mai.ru — informational site MAI

Copyright © 2000-2021 by MAI

Вход