Initial variation learning control is executed when learning condition is satisfied each time of its satisfaction, to absorb initial variations of manufacturing errors or variations of components of second clutch (CL2). When the number of times of total learning is counted and this becomes predetermined number (e.g. five times), initial variation learning control is judged to be converged. On the basis of condition of this convergence, so-called vertical sliding speed control returning belt (50) of variator (V) to LOWEST transmission ratio position is performed. With this, vertical sliding speed control is not performed in state in which excessively large torque acts on second clutch, and durability of belt and pulley is increased. It is therefore possible to prevent vertical sliding speed control from being performed in state in which input torque is predetermined value or greater, and possible to improve response of accelerator operation after performing vertical sliding speed control.