Motion control using support vector machine-based learning feed-forward

The tracking performance of a motion system increases drastically, if in addition to a feedback compensator, a feed-forward controller is applied. This feed-forward controller should be equal to the (stable part of the) inverse of the plant. However an accurate inverse is generally difficult to obtain, e.g. due to plant non-linearities. Learning Feed-Forward Control can help to overcome this difficulty. For the learning a function approximator is required. Support Vector Machines can be used as a function approximator in the LFFC setting. When using a SVM, a non-linear function is approximated by a linear approximation in the feature space, which is a (non-linear) mapping of the input-space. The mapping from input space to feature space determines the set of functions from which the function approximator can choose. The perfor- mance of LFFC is influenced by this set of functions. Mappings that give rise to a set of function that have a support on the complete input-space are preferred over mappings that have a local support