Abstract: The influences of the number of rolled strips,the quality of measured data and the tolerance of rolling force prediction were taken into account for building a self-learning speed optimization model of rolling force.The grades and values of thickness and width were considered in the determinant condition of long-term self-learning to reduce the frequency of size change.The information of equipment states which was separated from the data of the last strip was used into the calculation of long-term self-learning factor to improve the continuity of the self-learning model.Offline simulation results show that the accuracy of the rolling force model is improved after the self-learning algorithm is optimized.