Neural-network-based modeling and dynamic policy synthesis for model predictive control of nonlinear systems

A dynamic control policy with optimized dynamics is explored for its use in a model predictive control (MPC) algorithm for a nonlinear system modeled with a feedforward neural network. The nonlinear system is expressed as a polytopic quasi-linear-parameter-varying (quasi-LPV) system over a region of the state-input space and the dynamics of the policy are allowed to depend on the time-varying parameter of the quasi-LPV model. The policy dynamics are optimized off-line to obtain an enlarged domain of attraction which matches with the state-input region over which the polytopic approximation of the system holds good. A complete MPC algorithm using the dynamic policy as the terminal policy ensures stabilization and improved performance over a larger domain without a larger horizon length.