This paper presents a response evaluator as a semi-active control method for a base isolated building in long-period earthquake response analysis. The absolute acceleration of each story and the relative displacement of an isolated layer are the input signals and the switching parameter for the damping coefficient is the output signal of the response evaluator. The response evaluator is designed by a multilayered neural network, and a genetic algorithm is used to adjust the neural network parameters which determine the performance of the response evaluator. Four oil dampers are installed in the isolated layer and each damping coefficient of the dampers is set to a different value to correspond to the various characteristics of earthquake motions, then the total damping force of the isolated layer is multi-step values, regardless of whether using oil dampers with two-step damping coefficients. By the numerical simulation with the proposed semi-active control method, both the acceleration of each story and the displacement of the isolated layer, which are in a trade-off relationship with each other, are reduced compared with the acceleration in the case of all damping coefficients being small (Soft) and the displacement of the isolated layer in the case of all damping coefficients being large (Hard).