SUMMARYEven though model order reduction (MOR) techniques for linear dynamical systems are developed rather properly, there are still quite a lot of issues to be considered. This paper addresses a novel MOR technique for multi-input multi-output system with dominant eigenvalue preservation, which leads to controller cost minimization. The new technique is formulated based on an artificial neural network (ANN) prediction of an upper triangular form of the system state matrix A. Using the new system state matrix along with the linear matrix inequality (LMI) optimization method, a permutation matrix is obtained which leads to the new formulation of the complete system considered for MOR. Utilizing the non-projection state residualization technique, a reduced model order is obtained. The proposed ANN-LMI-based MOR method is compared with well-known reduction techniques such as the balanced Schur decomposition, proper orthogonal decomposition (POD), and state elimination through balanced realization.