Adaptive time horizon optimization in model predictive control

Whenever the control task involves the tracking of a reference signal the performance is typically improved if one knows the future behavior of this reference. However, in many applications, this is typically not the case, e.g., when the reference signal is generated by a human operator, and a remedy to this can be to try and model the reference signal over a short time horizon. In this paper, we address the problem of selecting this horizon in an adaptive fashion by minimizing a cost that takes into account the performance of the underlying control problem (that prefers longer time horizons) and the effectiveness of the reference signal model (that prefers shorter time horizons). The result is an adaptive time horizon controller that operates in a manner reminiscent of Model Predictive Control (MPC).