A learning controller makes learning correction based on a time tST between start of a shift control routine and detection of start of a change in rotational speed of the transmission input shaft and further based on the amount of rotational speed change &ohgr;S′ at the detected start of the rotational speed change. When the time tST is longer than a basic time and when the rotational change amount &ohgr;S′ is too large, it is determined that a servo piston has not been sufficiently advanced and the shift control is corrected so that the time tST is lengthened. When the time tST to the detected start of the rotational speed change &ohgr;S′ is too small, it is determined that, while the piston has been advanced properly, the target engagement pressure PTA in torque phase control is too low because of a miscalculation, and the shift control is corrected so that the target engagement pressure is raised.