LVQ combined with simulated annealing for optimal design of large-set reference models

Learning Vector Quantization (LVQ) has been intensively studied to generate good reference models in pattern recognition since 1986, and it has some nice theoretical properties. However, the design of reference models based on LVQ suffers from several major drawbacks for the recognition of large-set patterns, in which good reference models play an important role in achieving high performance. They are due in large part to the following facts: (1) it may not generate good reference models, if the initial values of the reference models are outside the convex hull of the input data, (2) it cannot guarantee optimal reference models due to the strategy to accept new reference models in each iteration step, and (3) it is apt to get stuck at overtraining phenomenon. In this paper, we first discuss the impact of these problems. And then, to cope with these, we propose a new method for the optimal design of large-set reference models using an improved LVQ3 combined with Simulated Annealing which has been proven to be a useful technique in many areas of optimization problems. Experimental results with large-set handwritten characters reveal that the proposed method is superior to the conventional method based on averaging and other LVQ-based methods.