A Dynamic LVQ Algorithm for Improving the Generalisation of Nearest Neighbour Classifiers

A new dynamic strategy for Kohonen’s LVQ algorithms using growing and pruning methods is introduced. Once the learning system converges to a solution using a fixed number of prototypes, the growing method incrementally adds new prototypes in those local regions where the misclassification error is greater. The initial locations of the new prototypes are viewed as estimates of the new equilibrium points of the learning algorithm with the augmented number of prototypes. This constructive process is repeated until the classification accuracy measured with a validation set stops decreasing. Then, a pruning algorithm is executed to remove all those prototypes that do not form class borders. Experimental results using NIST hand-written databases are promising since we show that dynamic LVQ algorithms outperform their static counterparts for the same codebook size.