Evolutionary computation for statistical pattern recognition
Xiao Wang · 2006
Evolutionary computation as a general problem-solving technique has been extensively applied in statistical pattern recognition. Typically, evolutionary algorithms are developed to solve complex optimization and search problems involved in different folds of designing a recognition system, e.g. feature extraction, supervised classification and clustering. These problems are often characterized by high-dimensional search spaces with convoluted landscape, noisy data, and little information about the objective functions. Traditional optimization methods are not efficient in dealing with them and evolutionary algorithms are therefore introduced. In the thesis a brief introduction to evolutionary computation is first presented. The synergetic combination of the two fields of evolutionary computation and statistical pattern recognition are then discussed. The state of the art is surveyed by analyzing those representative works. In our own effort of designing evolutionary systems for pattern recognition, two important topics are selected: ensemble learning and Markov random field (MRF) modelling. First, the technique of constructing classifier ensembles by manipulating the training examples is analyzed. It appears that the technique is essentially searching for appropriate weights associated with the examples. Thereby we adopt the idea of genetic search and develop a novel evolutionary learning algorithm.