Abstract: A statistical wear model of backup roll was built up based on a large number of measured data. The model is a multiple extremes and high non-linearity semi-theoretic formula whose pattern and magnitude are hardly defined with classical optimization methods because the computation is large and complex. By using the simulated annealing algorithm with memory, the coefficients in the roll wear model are optimized. The calculation results indicate a higher precision for prediction. After the model was applied in a 2800 4-roll mill of Wuhan Iron and Steel (Group) Co, the plate shape had been notably improved.