Due to the very narrow beam used in millimeter wave communication (mmWave), beam alignment (BA) is a critical issue. In this work, we investigate the issue of mmWave BA and present a novel beam alignment scheme on the basis of a machine learning strategy, Bayesian optimization (BO). In this context, we consider the beam alignment issue to be a black box function and then use BO to find the possible optimal beam pair. During the BA procedure, this strategy exploits information from the measured beam pairs to predict the best beam pair. In addition, we suggest a novel BO algorithm based on the gradient boosting regression tree model. The simulation results demonstrate the spectral efficiency performance of our proposed schemes for BA using three different surrogate models. They also demonstrate that the proposed schemes can achieve spectral efficiency with a small overhead when compared to the orthogonal match pursuit (OMP) algorithm and the Thompson sampling-based multi-armed bandit (TS-MAB) method.