Introspective Control Systems: Fast Model Predictive Control with Explicit Optimization Search, Nonlinear Models, and On-line Learning

摘要

Maximizing dynamic characteristics of a machine with complex nonlinear behavior can be very challenging. Most traditional control algorithms are quite short-sighted and perform poorly with nonlinearities and hysteresis in the machines they are controlling. In this paper, an advanced form of model-predictive controller is presented, dubbed as introspective control system (ICS), and compared against traditional forms of control algorithms, such as Optimal, Linear, and Adaptive model following algorithms. One essential feature of the ICS algorithm is that it utilizes search-based optimization, allowing it to work with nondifferentiable and noninvertible models, and discrepancies between plant and model can be recorded and used to update the model on-line. This study is centered around a real-life application, attempting to maximize dynamic characteristics of the Cockpit Motion Facility flight simulator, located at NASA Langley Research Center. The proposed methodology demonstrates superior dynamic bandwidth, without sacrificing damping of the system. Compared to the existing controller for this flight simulator platform, the proposed algorithm is a substantial improvement, able to achieve 5x and 2.3x increase in gain and phase bandwidth respectively. Moreover, transient response shows little to no overshoot while also improving the rise and settling time. This paper offers a detailed description of the proposed ICS methodology and compares it against traditional control theory and NASA's existing controller. Some insight is also provided as to how the choice of fitness function affects overall closed loop system performance, allowing additional fine-tuning of ICS-controlled system response.