Toward learning time-varying functions with high input dimensionality

Adaptive control problems in which the control law changes over time are considered. Such problems arise in robotics applications in which unanticipated variations in sensors, effectors, and the work environment change the desired input/output behavior of the controller. The problems are characterized in terms of learning an input/output function, and algorithms are presented for quick learning of such time-varying functions. The techniques presented are particularly effective for problems with input spaces of high dimensionality. The authors discuss why many existing algorithms are unsuitable for this type of problem and propose additional techniques for reducing the dimensionality of input spaces.<<ETX>>