In this paper, we propose a beamforming scheme with low complexity in terms of the number of training sequences to be used for the estimation of weight vectors in millimeterwave wireless personal area network. Our proposed scheme is based on the multiple-level procedure to reduce the number of training sequences not to lose the beamforming gain for the time varying channels. In the first level, the coarse beam is utilized to determine the best beam pair using two of transmit antennas and two of receive antennas. The beam search is conducted by using narrower beams with more antennas as the level index increases and the searching space is only the region corresponding to the best beam in the previous level and its two adjacent beams. In addition to the reduction of the number of training sequences, our proposed scheme satisfies a certain cost function such as maximization function of the received signal-to-noise ratio (SNR). To verify the performance we compare our proposed scheme with the one specified in the IEEE802.15.3c standard in terms of the received output SNR and show that our proposed scheme outperforms the one in IEEE802.15.3c standard even using the fewer training sequences.