SC Colloquium: "Learning in Nonlinear Model Predictive Control"
Emmanuel G. Collins
Department of Mechanical Engineering
Center for Intelligent Systems, Control, and Robotics
Florida A&M University – Florida State University College of Engineering
Systems with a priori unknown, and time-varying dynamic behavior pose a significant challenge in the field of Nonlinear Model Predictive Control (NMPC). When both the identification (or learning) of the nonlinear system and the optimization of control inputs are done robustly and efficiently, NMPC may be applied to control such systems. This seminar describes a novel method for adaptive NMPC (i.e., NMPC that continually learns the model), called Adaptive Sampling Based Model Predictive Control (SBMPC) that combines a radial basis function neural network identification algorithm with a nonlinear optimization method based on graph search. Unlike other NMPC methods, it does not rely on linearizing the system or gradient based optimization. Instead, it discretizes the input space to the model via pseudo-random sampling and feeds the sampled inputs through the nonlinear model, producing a searchable graph. An optimal path is found using LPA*, an efficient graph search method. This optimization is complete, meaning that if the prediction horizon is sufficiently long, SBMPC is guaranteed to find the global minimum subject to the sampling. In line with this, SBMPC is shown using transparent examples to avoid the local minima to which gradient-based methods tend to converge. Some results are compared with an adaptive version of Neural GPC, an existing NMPC algorithm based on Newton-Raphson optimization and a back propagation neural network model. Experimental results are described for flow separation control and simulation results are given for application to a complex power plant combustion process. Finally, new results show how to speed up on-line computations using learning of the heuristic function in SBMPC.