Fuzzy logic applications

Abstract

Fuzzy logic originated in the mid-1960s with the work of Lotfi Zadeh, as described in Chapter 6. However, it wasn't until the 1990s that it was widely recognized as a valuable tool in the atmospheric and other environmental sciences. One of fuzzy logic's first successful applications in the atmospheric sciences was the Machine Intelligence Gust Front Detection Algorithm (MIGFA) developed at the Massachusetts Institute of Technology's Lincoln Laboratory (Delanoy and Troxel 1993). Since then, a wide range of environmental science problems have been successfully addressed using fuzzy data analysis and algorithm development techniques. The purpose of this chapter is to supplement the introduction to fuzzy logic presented in Chapter 6 by describing a few selected applications of fuzzy logic that provide a flavor of the power and flexibility of this approach. The unique contribution of fuzzy logic is that it provides a practical approach to automating complex data analysis, data fusion, and inference processes that are usually performed by human experts with years of formal training and extensive experience. For instance, in developing a weather forecast, meteorologists consult data from various observations and models, each having a different level of relevance and reliability. Telltale patterns are sought in satellite and radar images, with contamination identified and disregarded. Soundings and surface data are analyzed and compared for consistency. Numerical weather models are consulted and weighed, with attention given to their past performance. Each bit of information provides a new piece of the puzzle, and may suggest a reanalysis of other data in an iterative process. Eventually, the available information is synthesized to form an overall consensus view, and perhaps also an assessment of confidence in the resulting forecast. This is exactly the type of complex procedure that a fuzzy logic expert system might automate. A fuzzy logic algorithm solution would first develop modules for analyzing and performing quality control on the various sources of information. For example, standard image processing techniques might be used to measure local characteristics in the radar or satellite data, or to identify features such as fronts by convolving appropriate templates with the image. Physical models or function approximation, including trained neural networks or other empirical models, might be used to relate raw sensor measurements to a quantity of interest-for example, determining temperature profiles from radiometer measurements. Statistical analyses might be employed to help determine data quality, or to determine conditional probabilities based on historical data. Human input such as hand-drawn fronts or boundaries might also be incorporated. The evidence supplied by these multiple data sources would be weighted by quality and by relevance, and used to perform fuzzy inference. The inference steps might be multi-layered or even iterative, feeding back into the data analysis; they could be tuned based on a training set or dynamically adjusted online using recent verification data. As this description suggests, a fuzzy algorithm can be quite complex-enough so, in fact, to represent many aspects of a human expert's reasoning process. Like human experts, the best fuzzy logic algorithms make use of all available information, saving any hard thresholding or binary decisions until the last step whenever possible. They also often provide a confidence value that lets downstream decision processes know how trustworthy the results are likely to be. © 2009 Springer Netherlands.