A modified interior point method for supervisory controller design

The design of supervisory controllers entails two phases: set point and model identification and the determination of appropriate control strategies. Many methods exist for designing controllers given linearized models of a plant at a set point. This paper presents a learning procedure that identifies models and set points using interior point techniques for solving linear programming (LP) problems. The learning procedure is an alternating minimization (AM) technique which optimizes the mean square prediction error of a multiple model stochastic supervisor. The algorithm is fast, requiring O(√n) iterations to find a local optimum, and computationally efficient requiring O(n^3.5) computations to solve the problem. The principal results of the paper are two bounds on how aggressively each leg of the alternating minimization may be performed to achieve this efficiency