Inverse training scheme for MS_CMAC neural network to handle random training data

An MS_CMAC neural network is a modular CMAC model designed to model smooth functional mapping. By using a small number of grid-state training data points, the MS_CMAC can use a tree structure network composed of one-dimensional CMAC nodes to decompose a complex multi-variable problem into several simple one-variable sub-problems. However, the grid-state training data points are not always available because most practical applications only obtain random training data points. To overcome this flaw, this work develops an inverse training scheme to supplement the original MS_CMAC. The inverse training scheme is based on a mathematical optimization approach which can employ random training data points to construct the virtual grid-state training data points. The computational results indicate that the ratio 3:1 of random training data points to virtual grid-state training data points can lead to an acceptable prediction accuracy in MS_CMAC.