Learning Structure from Data and Its Application to Ozone Prediction

  • Authors:

  • Luis Enrique Sucar;Joaquín Pérez-Brito;J. Carlos Ruiz-Suárez;Eduardo Morales

  • Affiliations:

  • Instituto Tecnológico y de Estudios Superiores de Monterrey, Campus Morelos, A.P. C-99, Cuernavaca, Morelos 62050, México. E-mail: esucar@campus.mor.itesm.mx;Instituto Tecnológico y de Estudios Superiores de Monterrey, Campus Morelos, A.P. C-99, Cuernavaca, Morelos 62050, México. E-mail: esucar@campus.mor.itesm.mx;Depto. de Física Aplicada, Cinvestav del IPN, Unidad Mérida, A.P. 73 Cordemex, Mérida, Yucatán 97310, México. E-mail: cruiz@kin.cieamer.conacyt.mx;Instituto Tecnológico y de Estudios Superiores de Monterrey, Campus Morelos, A.P. C-99, Cuernavaca, Morelos 62050, México. E-mail: emorales@campus.mor.itesm.mx

  • Venue:
  • Applied Intelligence
  • Year:
  • 1997

Quantified Score

Hi-index 0.00

Visualization

Abstract

In this paper we propose an algorithm for structure learning inpredictive expert systems based on a probabilistic networkrepresentation. The idea is to have the “simplest” structure(minimum number of links) with acceptable predictive capability. Thealgorithm starts by building a tree structure based on measuringmutual information between pairs of variables, and then it adds linksas necessary to obtain certain predictive performance. We haveapplied this method for ozone prediction in México City, wherethe ozone level is used as a global indicator for the air quality indifferent parts of the city. It is important to predict the ozonelevel a day, or at least several hours in advance, to reduce thehealth hazards and industrial losses that occur when the ozonereaches emergency levels. We obtained as a first approximation atree-structured dependency model for predicting ozone in one part ofthe city. We observe that even with only three parameters, itsestimations are acceptable.A causal network representation and the structure learning techniquesproduced some very interesting results for the ozone predictionproblem. Firstly, we got some insight into the dependence structureof the phenomena. Secondly, we got an indication of which are theimportant and not so important variables for ozone forecasting.Taking this into account, the measurement and computational costs forozone prediction could be reduced. And thirdly, we have obtainedsatisfactory short term ozone predictions based on a small set of themost important parameters.