A support vector machine based synthesis of suboptimal feedbacks for linear optimal control problems (Q6180997)

The artificial neural networks, learning methods on labeled data, reinforcement learning, and other machine learning methods recently are used for solving control problems and processes management. In this paper, related to the above direction, two problems of optimal control for continuous linear non-stationary system with terminal constraints and bounded discrete controls is considered. The authors present a method for constructing suboptimal feedback with respect to state, based on the support vector machine usually used to classify data. In particular, it is established that in the problems under consideration the optimal feedback is piecewise affine, which allows the use of support vector machines for linearly separable data samples and effectively build approximation of regions in which feedback is affine. The solutions to the analyzed examples are illustrated with plots.