The model predictive control algorithm uses a nonlinear model, input domain sampling, and a graph search technique without dependence on gradients. The nonlinear model is obtained by using input and output data from the system to tune a neural network model. The initial neural network can be trained using open loop data. Once the predictive control is turned on, the neural network continually adapts to represent time varying changes in the system. This is the first approach to adaptive nonlinear model predictive control that simultaneously performs online adaptation and model predictive control without the calculations of gradients for the predictive control.
This technology provides, in a single software package, a very general means of simultaneously identifying and controlling nonlinear systems without computing gradients, which leads to lower computational requirements than methods that are currently commercially available.
The technique of sampling the input domain guarantees satisfaction of hard constraints on input commands. Multiple core processing will give the proposed method increasingly greater computational speed advantage over current alternative methods since parallel computing hardware continues to become more widespread and more capable.
