A neural-net controller for multivariable systems is presented. The neural network for this controller has a structure in which small neural-net controllers for SISO systems are assembled and offers a unique path from the controlled variable to the manipulated variable. By using such a structured assembly, the interactions among the controlled and manipulated variables can be evaluated. The simulation results for both a linear three-input and three-output system and a crystal-growth process indicate that the proposed controller has the ability to learn the interactions between the control variables and to disclose the features of the interactions.