When the variation of the knock control system learning value is large, an idle speed learning control is performed in a learning acceleration mode. In this mode, the idle speed learning update increases and idle speed learning is performed at a high speed. By performing the idle speed learning in the learning acceleration mode, the idle speed learning can immediately follow large variations of the knock control system learning value. Accordingly, the idle speed learning control performed during idling of a hybrid car can be completed in a short time, and fuel consumption during idling can be reduced.