Some orthogonal weights are selected from pseudorandom numbers and accumulated successively to cover whole learning space in the two-layer feedforward topology for pattern classification.