Improved Learning Rule for LVQ Based on Granular Computing

摘要

LVQ classifiers are particularly intuitive and simple to understand because they are based on the notion of class representatives (i.e., prototypes). Several approaches for improving the performance of LVQ in batch-learning scenarios are found in the literature. However, all of them assume a fixed number of prototypes in the learning process; we claim that the quantized approximation to the distribution of the input data using a finite number of prototypes, should not be fixed. Thus, in this paper we propose an improved learning algorithm for batch and on-line variants in LVQ. The proposed algorithm is based on a modified LVQ rule and granular computing, a simple and low cost computational process of clustering. All this, increases the dynamics in the learning process, proposing new prototypes which have a better covering of the distribution of classes, rather than using a fixed number of them. Similarly, in order to avoid an exponential growth in the number of prototypes, an automatic pruning step is implemented, respecting the desired reduction rate.