Goal-directed property of online direct inverse modeling

In order to learn an inverse system of a plant, the learning method which inputs the output of the plant to the learner and uses the input of the plant as the desired output signal for the learner has been used in many researches. This learning method for learning control by using a neural network is called “direct inverse modeling” (DIM). Jordan (1995) regarded DIM as a purely off-line learning method and pointed out that DIM is not goal-directed. However, some researches have proposed a different way of using DIM. We deal with online DIM, which is the simultaneous or alternate execution of the plant control and the inverse model learning by DIM. The learning properties of online DIM is completely different from that of off-line DIM. This paper shows that online DIM becomes goal-directed under certain conditions