Qualitative Models as Indices for Memory-based Prediction

Contact Person: Boi Faltings
Phone: (+41 21) 693-6677
E-mail: Boi.Faltings@epfl.ch
Partners: Nestlé
homepage: http://liawww.epfl.ch/~faltings/nestec/nestec.html

Project Description

Computational modeling is a powerful tool for today's engineers. It is particularly appicable in domains where systematic experimentation allows gathering examples to prove or disprove specific hypotheses. For processes where experimentation is not possible, even a large number of examples is often insufficient to reliably construct models. Such is the case for example for cement kilns or coffee roasters.

The goal of this project is to develop a computational modeling method where learning is replaced by reasoning directly from examples. This case-based or memory-based reasoning mimicks the way that human experts often reason from particular previous experiences:

The difficulty now shifts from modelling the relation between observations and a value to be predicted to deciding when a previous example is similar enough to make a good predictor. This question, intuitively simple, turns out to be very subtle and sensitive to a careful choice of attributes and their weighting. Similarity measures which fully exploit the information in the examples require a lot of knowledge about the system which generated them!

Last modified: Sun Apr 23 04:57:08 2000