In this paper an improved Simulated Annealing (SA) based Learning Vector Quantization (LVQ) algorithm has been proposed and applied to two benchmark classification problems. LVQ is a prototype based heuristic supervised algorithm that has been successfully applied to classification problems in medical imaging and for speech and image compression. However LVQ sometimes suffers from stability and convergence issues. In order to alleviate these problems a new globally convergent LVQ based on SA has been proposed. In this paper SA is used to update the prototype vectors in LVQ to minimize the classification error in addition to the classical competitive learning based LVQ scheme. The performance of the proposed LVQ-SA algorithm is tested on two benchmark classification problems. © Research India Publications.