Various aspects of the problem of designing linear, multivariable controllers from nonparametric plant models via search-based methods are reviewed and discussed. In particular, three new mathematical programming algorithms suitable for solving this problem are developed and evaluated. It is shown through theoretical analysis and experimental results that the key issues determining the performance of these algorithms are the choice of a measure of algorithm progress and the corresponding parameter search direction. Methods for parameterizing controllers are developed, and the effects of these parameterizations upon the dimension and topology of the underlying parameter space are investigated. One parameterization in particular, called the state feedback parameterization, is shown to provide a straightforward method for fixing specified controller poles while still allowing the use of a state-variable controller representation. Expressions for the gradients of functions describing many common control system design specifications, such as constraints on singular value frequency responses, damping ratios of controller poles and zeros, are derived. Efficient methods for calculating these gradients are also discussed. Several successful applications are presented, including controller design and implementation for a large aerospace structure ground test facility and a controller redesign for the Hubble Space Telescope Pointing Control System.
展开▼
