Analysis and design of most tolerant logical neural networks

Neural networks with the sub-most and the most tolerant ability for input data are designed on the basis of the fact that sub-most and/or the most tolerant ability can only be obtained by classification hyperplanes which are just through the middle point of the connective line of any adjacent vertices of different logic values in an n-dimensional hypercube and/or orthogonal with the line. The design rules of connective weights and bias values are presented. It is proved that the sub-most and most tolerant network is in n-k-1 and n-n-k-1 scale (where k⩽2^n-1), and the connective weights can only be 0, 1, -1, which results in the easiest realization of the net. Finally, computer simulation results are presented