Learning sequential and continuous control

An experiment is described in which a sequential, goal-seeking learning system is combined with a CMAC neural network controller to produce smooth continuous movement. Our robot learning system PURR-PUSS, or PP, is designed to learn complex tasks and is not suited to the smooth control of movement. This is because logical reasoning cannot normally assume that stimuli or actions with similar representations have a similar significance, so it cannot easily take advantage of situations where such similarity can be exploited. CMAC, a model of the functioning of the cerebellum, is designed to operate with analogue variables, mapping them into a series of neurons with overlapping dendritic fields. In this paper we demonstrate that the logical PP can be coupled to CMAC so as to combine the benefits of each in a 2 level hierarchy