In this paper, we study an intelligent surface-aided transmit-array antenna architecture which deploys intelligent transmitting surface (ITS) near the active antenna in mmWave downlink communication system. We aim to maximize the average achievable rate at the user by optimizing the phase shift matrix of ITS. To overcome the difficulty of acquiring channel state information (CSI), we propose a stochastic gradient descent (SGD)-based algorithm that requires only historical channel observations. Simulation results show that our proposed scheme with single active antenna and ITS outperforms conventional MISO system.