Using Qualitative Constraints In Ozone Prediction

We describe a case study in which we applied Q learning (qualitatively faithful quantitative learning) to the analysis and prediction of ozone concentrations in the cities of Ljubljana and Nova Gorica, Slovenia. We used program QUIN to induce a qualitative model from numerical data that include the measurements of several meteorological and chemical variables. The resulting qualitative model consists of tree-structured monotonic qualitative constraints. We show how this model for Nova Gorica enables a nice interpretation of complex meteorological and chemical processes that affect the level of ozone concentration. For Ljubljana, in addition to inducing a qualitative model from data, we extended the qualitative model to also enable numerical prediction. In this case, we used in addition to measured data also data from the European meteorological prognostic model ALADIN which itself does not model pollutants. Program QCGrid was used to induce a numerical prediction model which respects the constraints in the qualitative model and fits the data well. We show that qualitatively constrained numerical model improves numerical prediction in comparison with some standard numerical learning methods.