This study proposes the surrogate modeling by the Gaussian process regression with the transfer learning (TL-GPRSM). The TL-GPRSM can reduce the computational cost by using data with input-output relationships close to those of the target analysis for constructing surrogate models. For validation, the TL-GPRSM was constructed for a Monte Carlo calculation for the live load performance evaluation of a steel-plate girder bridge with damage. Here, transfer learning was used to reduce the computational cost by considering the input-output data from the analysis of the initial state bridge. The results showed that the TL-GPRSM was able to predict higher accuracy than surrogate models without transfer learning. It was also shown that the effectiveness of transfer learning can be determined from the contribution estimated by ARD.