Overview
This research introduces a model predictive control (MPC) method termed "observed control." The core of this approach lies in identifying and utilizing the duality between state estimation and model predictive control. The primary objective of observed control is to efficiently compute control actions while achieving linear scalability with respect to prediction horizon length.
The method is characterized by its computational efficiency, the incorporation of adaptive time horizon lengths, and the provision of early optimization termination criteria. For its backend optimization framework, observed control utilizes Kalman smoothers, which are described as providing a familiar implementation coupled with strong theoretical guarantees.
The scope of observed control extends beyond linear systems. A formulation is presented that separates linear model predictive control into distinct reactive and anticipatory components. This separation facilitates the application of observed control at any time and with any horizon length, while simultaneously ensuring controller stability even for short time horizons. Furthermore, the methodology is extended to address nonlinear systems and non-quadratic cost functions. This extension aims to achieve locally-optimal control for complex systems, maintaining the benefits of linear prediction horizon scalability and adaptive-horizon capabilities.
Research Context
The work positions itself within the field of model predictive control, a control strategy that computes control actions by optimizing a model of the system over a prediction horizon. A central theme is the concept of duality between state estimation and model predictive control. State estimation, particularly through tools like Kalman smoothers, involves inferring the internal state of a system from noisy measurements. The research highlights that this relationship can be exploited to design efficient control algorithms.
Approach
The proposed "observed control" method fundamentally leverages the duality between state estimation and model predictive control. The practical implementation of its backend optimization relies on Kalman smoothers. These smoothers provide the computational framework for determining control actions. Key algorithmic features include:
- Linear prediction horizon length scalability.
- Adaptive time horizon lengths.
- Early optimization termination criteria.
For linear systems, the method introduces a specific formulation that disentangles linear model predictive control into two constituent parts: a purely reactive component and an anticipatory component. This structural separation is designed to enable flexible application of observed control, allowing it to operate "any-time any-horizon," while preserving controller stability even when short time horizons are employed.
The methodology is then extended to accommodate more complex scenarios:
- Nonlinear systems.
- Non-quadratic cost functions.
This extension aims to achieve locally-optimal control for these complex systems. Crucially, the extension is developed such that the benefits observed in the linear case—specifically, linear prediction horizon scalability and adaptive-horizon capabilities—are maintained.
Findings
The primary findings of this work concern the characteristics and capabilities of the introduced "observed control" method:
- **Duality Exploitation**: The research demonstrates that the duality between state estimation and model predictive control can be effectively utilized to design control algorithms.
- **Computational Efficiency**: The presented algorithms are noted for providing exceptional computational efficiency.
- **Scalability**: The method achieves linear prediction horizon length scalability. This means the computational effort scales linearly with the length of the prediction horizon.
- **Adaptive Horizons**: It incorporates adaptive time horizon lengths.
- **Optimization Termination**: The algorithms include early optimization termination criteria, contributing to efficiency.
- **Kalman Smoother Foundation**: Kalman smoothers serve as the backend optimization framework, offering a familiar implementation and theoretical guarantees.
- **Linear MPC Decomposition**: A formulation is presented that separates linear model predictive control into purely reactive and anticipatory components. This decomposition supports "any-time any-horizon" observed control.
- **Stability for Short Horizons**: The decomposed formulation ensures controller stability even for short time horizons.
- **Nonlinear Extension**: The method is extended to nonlinear systems and non-quadratic cost functions. This extension enables locally-optimal control for complex systems.
- **Maintained Capabilities**: For nonlinear systems, the method maintains linear prediction horizon scalability and adaptive-horizon capabilities.
Why This Matters
The development of observed control offers a computationally efficient approach to model predictive control, capable of managing both linear and nonlinear systems while adapting to varying time horizons. Its utilization of Kalman smoothers grounds the method in a theoretically robust and familiar framework. The ability to separate reactive and anticipatory components in linear MPC, along with stability guarantees for short horizons, suggests enhanced flexibility and reliability in control applications.