The neural networks which can not only realize n-dimensional logic functions

but also be withthe sub-most and the most robust ability for input data are designed

based on the fact that the sub-most and/orthe most robust ability can only be obtained by classification hyperplanes which are in the middle point of theconnective hoe of any adjacent vertexes of different logic values in n-dimensional hypercube and/or venical withthe line.ffe design rules for connective weights and bias values between robust neurons are presented

withwhich it is proved that the sub-most and most robust network is in the n-k-1 and n-n-k-1 scale （where k≤2n-1）

and the connective weights between the robust neurons can only be 0

1

-1

which result in the eastestrealization of the net. Finally

computer simulation results are presented.