System Identification and Control Using Quadratic Neural Networks

This letter proposes convex formulations of system identification and control for nonlinear systems using two layer quadratic neural networks. The results in this letter cast system identification, stability and control design as convex optimization problems, which can be solved efficiently with polynomial-time algorithms. The main advantage of using quadratic neural networks for system identification and control as opposed to other neural networks is the fact that they provide a smooth (quadratic) mapping between the input and the output of the network. This allows one to cast stability and control for quadratic neural network models as a Sum of Squares (SOS) optimization, which is a convex optimization program that can be efficiently solved. Additionally, these networks offer other advantages, such as the fact that the architecture is a by-product of the design and is not determined a-priori, and the training can be done by solving a convex optimization problem so that the global optimum of the weights is achieved. It also appears from the examples in this letter that quadratic networks work extremely well using only a small fraction of the training data.