Dynamic parameter tuning in a particular branch of soft computing specially designed for mechanical systems' control

A novel and efficient approach invented for the adaptive control of approximately and partially known mechanical systems under unmodeled external dynamic interaction is presented. The method overcomes the limitations of classical feedforward neural network-based approaches via applying uniform structures derived from the Euler-Lagrange equations in the most general and formal level. Being a compromise between the classic hard computing (HC) and soft computing (SC) the typical difficulties as the a priori unknown number of the necessary nodes and free parameters, the scaling problems regarding the applicable range of the parameters are evaded. On this basis a relatively simple uniform structure of reduced number of parameters appropriate for real time tuning can be obtained. Several ancillary procedures also independent of the details of the particular task to be solved are applied to support machine learning, too. The operation of the method is illustrated via simulation in the case of a 3 active and one passive DOF SCARA arm used for polishing the surface of a bell-shaped work-piece.