Continuous Function Approximation Using Feature Projections

Abstract: In terms of their application areas, machine learning algorithms can be grouped into two categories as classification learning and continuous (real valued) function approximation learning algorithms. The difference between these categories is in the target values of the examples; while the in the former target values are nominal, the latter assumes continuous target values. This project proposal is about learning, from a set of examples, approximations of such continuous functions that their mathematical model either do not exits or unknown. Using such an approximation function, it is possible to predict the values of the target function at new data points. Due to its importance in many application areas, studies in this area have begun by statisticians long before the development of machine learning approaches.The first approaches to this problem, called as regression by statisticians, were about to determine the parameters of a linear function that fits the training examples as close as possible. In recent years, studies in machine learning have developed non-parametric approaches to the function approximation problem. The knowledge representation technique based on feature projections and the prediction scheme based on voting among feature predictions that we have developed recently have proven to be very successful in classification problems. In this project, the application of knowledge representation based on generalizations of feature projections and the voting based prediction to the problem of continuous function approximation will be investigated. The approximation algorithms that will be developed in the context of this project will be evaluated on the standard data sets available at UCI-Repository and the statistical data sets compiled by some governmental institutions.

Keywords: Machine Learning, Feature Projections, Continous Functions

Principal Investigator: H. Altay Guvenir, Ph.D.
Investigator: Ilhan Uysal
Investigator: Tolga Aydin, BSc.

Duration: February 1999 - August 2000.

Sponsor: Scientific and Technical Research Council of Turkey

Grant No: 198E015

Budget: 1,005,550,000 TL (USD 2,873 in Feb. 1999).