Learning Adaptive Horizon Maps Based on Error Forecast for Model Predictive Control

We present a model predictive control framework that uses varying prediction horizons according to the current forecasted uncertainties and estimated distance of the terminal state from its desired state. Our results suggest that the space of such optimal horizons, which we call horizon maps, is well structured for linear systems, meaning that it can be easily learned using tools from machine learning. Our approach is well suited for real-time control and can scale to higher dimensional systems. We also perform an analysis on the required quality of the datasets used to learn the horizon maps and conclude with results of this framework using an externally-driven, constrained linear quadratic regulator problem.