An object of the present invention is to carry out air-fuel ratio learning to a high accuracy even in a region such as the idling region wherein the number of acquisitions of air-fuel ratio feedback correction values within a predetermined period is small. Average values A for one fluctuation period of an air-fuel ratio feedback correction coefficient .alpha. are computed and stored in order of computation in a RAM, and an arithmetical mean value B obtained. A weighted average value C of a previous learning correction coefficient .sub.KL and the arithmetical mean value B is then obtained (C=X.multidot.B+(1-X).multidot.K.sub.L). The weighting proportion X is set larger the larger the number of acquisitions of the average value A. C is updated/set as a new learning correction coefficient K.sub.L. In this way, since the updating/setting treatment of the learning correction coefficient K.sub.L is carried out in accordance with the acquisition conditions of .alpha., then even in a region such as the idling region wherein the number of acquisitions of the average value A per predetermined period is small, updating/setting of the learning correction coefficient K.sub.L can be expedited with minimal influence from an unstable .alpha. immediately after commencing air-fuel ratio feedback control. Therefore learning opportunities similar to those for the high speed region can be provided. Moreover due to learning being expedited, the learning correction coefficient K.sub.L can have a higher reliability than for a conventional arrangement.