A structured neural network for power-system state evaluation is proposed. The network has a generalized structure that is independent of applications. The network is built only by input-output example patterns, unlike the ordinary structured networks. The interpolation ability of the network is controlled by changing the coefficients relating to activation of units. The focus on one portion of an input vector can be controlled by changing the coefficients relating to relative contributions of subnetworks. This mechanism realizes the feedback from logical knowledge to recognition. The performance of the network has been compared to that of back propagation, and the result shows the importance of the feedback