New Learning Strategy for Prototypes in Linear Vector Quantization

This paper proposes a new strategy for prototypes initialization in linear vector quantization algorithm (LVQ). Three principles which must be satisfied during the learning stage are shown in the paper. These principles are essential to guarantee appropriate learning for the LVQ algorithm. However, all versions of LVQ algorithms try to answer to one of the principles, but unfortunately contradicting with the other ones. The new strategy proposed in the paper aims to solve this issue and consists of two steps: (1) analyse the a-priori data set and (2) apply a pre-learning algorithm to initialize the prototypes. The pre-learned prototypes resulted from step 2 are used by the LVQ algorithm in the learning process. The examples presented in the case study and the criterion used to assess the training performance of the prototypes reinforce that the training strategy of the prototypes proposed in the paper provides better results in certain situations compared to classical LVQ algorithms.