This paper presents a control architecture with a neural controller and a conventional linear controller for nonminimum phase systems. The objective is to minimize overall position errors as well as to maintain small undershooting. These attributes make it difficult to obtain the optimal solution which satisfied all individual objectives. Moreover, heuristic attempts of a proper combination of several objectives may produce a feasible solution but not necessarily an optimal one. With the concept of Pareto optimality and evolutionary programming, we train the controller more effectively and obtain a valuable set of optimal solutions. According to the preference, we can easily determine the most suitable solution from a pool of optimal candidates.