| The research work covered several topics within the field of artificial neural networks, discriminant analysis, multivariate calibration and signal processing in chemometrics, which constitute the three parts of the thesis.; Part one consists of two chapters. Chapter 1 emphasizes the investigation of the influence of multi-layered feed-forward neural network parameters on the performance characteristics in supervised pattern recognition In Chapter 2, autoencoding networks and Kohonen networks are used to reduce the dimensionality of multivariate data sets, producing a two-dimensional display of the data. The plots obtained by these two networks are compared with results from two conventional methods, principal component analysis and non-linear mapping. Advantages and drawbacks of these four methods are discussed.; A large part of the work is devoted to the application of repeated cross-validation in discriminant analysis and multivariate calibration with a quantitative comparison between repeated cross-validation, repeated evaluation set, single cross-validation and single evaluation set procedures. Several different pre-processing steps for NIR spectral data, three types of discriminant analysis methods and several regression methods are also compared based on repeated cross-validation, This work appears in Chapters 3 to 5.; The last part of this thesis is centred around smoothing of noisy data by principal component analysis and least-squares splines. In Chapter 6, details of a new PCA filter are described. The effectiveness of the PCA filter is verified by simulated and real data sets. Chapter 7 is devoted to application of least-squares splines for smoothing noisy data. A new scheme for the automatic selection of smoothing parameters used in fitting experimental curves is proposed. |
