Neural networks are used for extracting information on the sensitivity of surface ozone to changes of ozone precursors directly from the data of routine air quality measurements. Three neural network-based (perceptron type) empirical models are created following original methodology and using independent data sets of long-term (1980--1995) air quality monitoring collected at three sites (Azusa, Riverside-Rubidoux, and Reseda) at South Coast Air Basin, California. The qualitative features of the nonlinear relationships between ozone and its precursors captured by the empirical models are compared with theoretical ones under approximated zero-wind conditions. Fair agreement with the photochemical theory is found for the results demonstrated by two models, and the results of the third model (for Reseda) are found to be in poor agreement with the theory. It is found also that in accordance with the theoretical expectations all the empirical models demonstrate a seasonal transition from NOx-sensitive to NOx-saturated regimes of ozone photochemistry. A comparison between the results obtained with and without the zero-wind approximation shows that windy conditions usually are a benefit to the NOx-sensitive regime. With regard to the problem of reducing ozone concentrations, it is assessed that a dominant fraction of days on which 8-hour-averaged ozone concentrations exceeded a level of 80 ppb is associated with the NOx-sensitive regime at Riverside and Reseda sites (89% and 93%, respectively), while at Azusa, both regimes appear on those days with the same frequency.