This thesis initially overviews the general methodologies and techniques of databased models design and learning algorithms for pattern recognition. New modelling approaches are then proposed as original contributions of thi~ thesis with the aim of deriving the models with enhanced recognition performance. The proposed approaches are rigorously investigated through analysis and numerical examples. A new combined approach is proposed to reduce the complexity of support vector regression (SVR) model by using a forward orthogonal least squares (FOLS) with predicted residual sum of squares (PRESS) statistic algorithm. This new approach is called SVR-PRESS. t' A new modelling approach is developed to improve the model's generalization capability based on multi-objective optimization (MOO). It is proposed that an improved generalization capability can be achieved by splitting a training set and '. treating the model performance on the training subsets as conflicting objectives, followed by training the model that simu1tan~ouslyoptimizes these objectives.. A new improved stochastic discrimination (SD) is proposed to improve the classification accuracy of the standard SD based on the idea of important space containing all the misclassified training samples. The improved SD reduces the misclassification I error rate over the important space by defining a new discriminant function with smaller variance than that of the standard SD. Three new techniques are developed to modify the probabilistic neural network (PNN). A new PNN learning algorithm reduces the complexity of the conventional PNN by utilizing a forward constrained selection (FCS) that chooses the most significant neurons. This new method is called PNN-FCS. A reweighted PNN (RPNN) achieves a higher classification accuracy than that of the traditional PNN via tuning the weights of the neurons. A multi-level PNN (MLPNN) reduces the misclassification error rate of the conventional PNN by using new PNNs with improved classification performance over the clusters containing all the misclassified training samples.