Model of neurocontrol of redundant systems (Q1917893)

A new method of differential control of redundant systems is developed. It is based on the analogy with the brain function in visual-motor coordination. The proposed scheme simulates both learning and operation procedures in living creatures. Design of a controller generally requires generation of the inverse transfer function of the plant. In the case of control by velocity it involves calculation of pseudoinversion of Jacobian of the plant. This process may be rather complicated in systems with many parameters. The proposed method does not require calculation of matrix inverses. The model of pseudoinversion of Jacobian is generated as a layered neural network involving a hidden layer with higher-order neurons. It is developed during the learning process based on the error back-propagation algorithm. Its convergence is analyzed for a linear case. The procedure uses differential control, local linearization, and control error estimation by a special neural network taught by Hebb's rule. The proposed scheme involves two neural control blocks with higher-order neurons that are adapted in the learning process. It is designed for control of mechanical systems by velocity, though it can be generalized for differential control of arbitrary systems. It was tested on a model of a multijoint manipulator.