Request Expert

Expert's book, "Pattern Recognition Using Neural Networks, published by Oxford University Press in 1997, is considered an excellent reference in this area. Expert also has some papers in the research literature and has developed a powerful algorithm (the radial basis functional link net). The applications he has worked with include edge detection and sharpening of images, recognition of characters, ice core dating and data compression, and others.

Expert regularly develops algorithms for computation. Often the applications are to pattern recognition and clustering algorithms, neural network algorithms, fuzzy systems algorithms, and other decision-making and simulation algorithms. A recent example was an algorithm to cluster the spectra from laser signals reflected from human skin for the purpose of detecting malignant, benign, or normal skin cells. Other algorithms include the radical basis functional link net that is much faster and more accurate than backpropagation type neural networks.


Much of Expert's work in pattern recognition involves clustering data into subclasses and then clustering the subclasses into classes that are not linearly separable. Applications include geological (mineral) exploration and age dating of ice cores. Clustering is more useful than is commonly realized as it discovers relationships in data.

A significant portion of Expert's work involves decision making. Any type of classification and pattern recognition involves decision making. In all cases, in a given situation, one of a set of alternatives must be selected as a response action and it is chosen to optimize some cost function. Much decision making involves statistics and probability to handle uncertainty, and uses both empirical and theoretical distributions. His two articles in the CRC "Handbook of Electrical Engineering" (1997) deal with noise and stochastic processes, but simpler Bayesian methods often suffice. Decisions can be made using logic, fuzzy logic, rule based expert systems or trained neural networks, especially in nonlinear problems.


Expert's students have developed expert systems under his instruction over the past several years using both propositional logic and fuzzy logic. His current approach is to use subsystems to build an overall expert system. Certain subsystems use crisp and/or fuzzy rules on data put out by neural networks that interpolate and approximate data. Other subsystems may use case based reasoning to select candidate decisions (response actions) that are then refined by crisp or fuzzy rules and neural networks. Additionally, Expert has used fuzzy expert systems to estimate the value of homes and other similar processes.

Feature engineering is discussed in Expert's book, "Pattern Recognition Using Neural Networks," Oxford University Press, 1997. A standard method that he uses is to obtain a large set of candidate features, compare them to obtain the greatest separation between pairs of classes, and then remove the ones that have high correlation with others. Expert has extracted features from $5 chips (tokens) of the various casinos so that a neural network could be trained to separate the chips by casino. Other applications that he has worked with are features for optical character recognition.


Expert does research in fuzzy logic, both for decision making and intelligent control of automated systems. He has implemented a trainable fuzzy rule based controller that was inserted in the loop of an electric DC motor simulation, and the fuzzy system learned to control the motor very accurately. His publications involve applications of fuzzy logic and mathematics.


Expert has worked in the theory of machine learning for several years. A system learns when its experience changes its parameters so that is subsequent behavior changes. One of his research papers (published in the IEEE Trans. Knowledge and Data Engineering, 1996) was on learning controllers that adapt by changing the parameters of their rules so their behavior is more optimal with respect to some criterion. Learning requires a process of decision making to be optimized with regard to some measure of performance. This process may be done with genetic algorithms or other types of search, but in some cases it must be done with a small set of trials to keep costs reasonable. Such cases require more analysis into the search space to obtain mathematical relations that determine the learning.


Expert uses neural networks for various types of decision making, nonlinear interpolation and approximation, and supervised learning. His book, "Pattern Recognition Using Neural Networks," gives a number of examples. Recently Expert used neural networks for exploring mineral (based on a set of conditions present), approximating the ages of layers in Antarctic ice cores, recognition of characters, edge detection in images, metallurgical process control, and other applications.


A recent paper of Expert's followed up on some other papers to show that many types of fuzzy expert systems are equivalent to radial basis function neural networks. In such neural networks, each radial basic function is a Gaussian fuzzy set membership function that does the fuzzification while the training at the output layer does the decision making and defuzzification. Expert has incorporated the best features in his radial basis functional link net, which is extremely quick in supervised learning and very robust in operation. Expert uses this type of neural network in most of his applications, but not all (there are cases where fast operation is more desirable than quick learning, and here he usually uses backpropagation type multiple layered perception networks). In some applications, Expert uses fuzzy rules with inputs from neural networks, or neural networks with inputs from fuzzy systems to form a complete decision making system.


Pattern recognition is one of Expert's strongest areas of interest, from feature engineering and extraction to clustering and supervised training of neural networks. Expert is very interested in finding new applications and applying the most powerful current tools. His new fuzzy clustering algorithm is under development in software. He also uses other methods of fuzzy clustering. Expert's new method of recognizing to which class an object belongs is to not use the minimal distance assignment, but to use a maximal fuzzy membership that depends upon the weighted fuzzy expected value for a prototype rather than the expected value.