Learning task-dependent input shaping control using radial basis function network

We consider a general problem of learning a connectionist approximation for the dependence of the input shaping control programs on a set of task parameters. The task parameter set may include the initial and desired final coordinates of the controlled system and some other parameters that can take different values from task to task. The pertinent feature of the problem is that the approximation algorithm works with the control program and the time history of the system output, each of which is represented by a vector of large dimension. We demonstrate that despite these large dimensions the problem is still tractable if the number of the task parameters is moderate. We consider a controller architecture that is based on an affine radial basis function network. The network weights are updated depending on the results of the learning trials with different task parameters. The presented paradigm could help to obtain technically feasible solutions in cases that are difficult for classical control approaches. The application of the algorithm to the input shaping control of a manipulator arm with elastic joints demonstrates the high performance of the method.<>