MACHINE LEARNING OF MARITIME FOG FORECAST RULES

In recent years, the field of artificial intelligence has contributed significantly to the science of meteorology, most notably in the now f amiliar form of expert systems. Expert systems have focused on rules o r heuristics by establishing, in computer code, the reasoning process of a weather forecaster predicting, for example, thunderstorms or fog. In addition to the years of effort that goes into developing such a k nowledge base is the time-consuming task of extracting such knowledge and experience from experts. In this paper, the induction of rules dir ectly from meteorological data is explored-a process called machine le arning. A commercial machine learning program, called C4.5, is applied to a meteorological problem, forecasting maritime fog, for which a re liable expert system has been previously developed. Two datasets are u sed: 1) weather ship observations originally used for testing and eval uating the expert system, and 2) buoy measurements taken off the coast of California. For both datasets, the rules produced by C4.5 are reas onable and make physical sense, thus demonstrating that an objective i nduction approach can reveal physical processes directly from data. Fo r the ship database, the machine-generated rules are nor as accurate a s those from the expert system but are still significantly better than persistence forecasts. For the buoy data, the forecast accuracies are very high, but only slightly superior to persistence. The results ind icate that the machine learning approach is a viable tool for developi ng meteorological expertise, but only when applied to reliable data wi th sufficient cases of known outcome. Tn those instances when such dat abases are available, the use of machine learning can provide useful i nsight that otherwise might take considerable human analysis to produc e.