Entropy-constrained generalized learning vector quantization neural network and soft competitive learning algorithm

According to the generalized learning vector quantization (GLVQ) network and the maximum-entropy principle, an entropy-constrained generalized learning vector quantization (ECGLVQ) neural network is proposed. A learning algorithm of the network, a generalization of the soft-competition scheme (SCS), is derived via the gradient descent method. Because the loss factor and the corresponding scaling function are defined as the same fuzzy membership function, it can overcome the problems for fuzzy algorithms of GLVQ network possess. Many important properties of the ECGLVQ network and its soft competitive learning algorithm are given. Thereby, the rule for choosing the Lagrangian multiplier is designed.