A Simple Soft Computing Structure for Modeling and Control

Abstract

Using the interpolation/extrapolation skills of the core function of an iterative adaptive controller, a structurally simple single essential layer neural network-based topological structure is suggested with fast and explicit single-step teaching and data-retrieving abilities. Its operation does not assume massive parallelism, therefore it easily can be simulated by simple sequential program codes not needing sophisticated data synchronization mechanisms. It seems to be advantageous in approximate model-based common, robust, or adaptive controllers that can compensate for the effects of minor modeling imprecisions. In this structure a neuron can be in either a firing or a passive (i.e., producing zero output) state. In firing state its activation function realizes an abstract rotation that maps the desired kinematic data into the space of the necessary control forces. The activation function allows the use of a simple and fast incremental model modification for slowly varying dynamic models. Its operation is exemplified by numerical simulations for a van der Pol oscillator in free motion, and within a Computed Torque type control. To reveal the possibility for efficient model correction, a robust Variable Structure/Sliding Mode Controller is applied, too. The novel structure can be obtained by approximate experimental observations as e.g., the fuzzy models.