USING LEARNING MACHINES MORE INTELLIGENTLY AND RIGOROUSLY FOR THE GROUND-LEVEL OZONE PREDICTION

Multilayer perceptron (MLP) and support vector machine (SVM), as two popular learning machines, are used increasingly as alternatives to classical statistical models for the ground-level ozone prediction. However, without the sufficient awareness about their limitations employing learning machines can still lead to unsatisfactory results on modeling ozone evolving mechanism, especially during ozone episodes. With some spirit of review, this commentary discusses, in the ozone prediction context, the recently developed algorithms/technologies on treating the most prominent model-performance-degraded limitations which include, for MLP, the "black-box" property, i.e., hardly proving physical explanation for the trained model, overfitting and local minima problems, and for SVM, parameters identification and class imbalance problems. This commentary will impress the "would-be" environmental modelers that the underlying philosophy of using learning machines is by no means as trivial as simply fitting models to the data, because it has reflected on the difficulties, controversies or unresolved problems; meanwhile this commentary also serves as a reference point for further technical reading for experts in this field.