TL;DR
This paper introduces a novel, convex formulation for "minimal regret" state estimation in time-varying systems. By adapting System Level Synthesis (SLS) techniques, they enable practical implementation of observers that can adapt to various noise and disturbance patterns, outperforming traditional Kalman (H2) and H-infinity (H∞) filters in numerical experiments.
What Problem Does This Paper Address?
Estimating the internal state of a dynamic system from noisy measurements is a fundamental problem in many engineering fields. For linear systems, the Kalman filter (H2) is optimal when noise is Gaussian, while the H∞ filter offers robustness against worst-case disturbances. However, real-world systems often exhibit noise and disturbance patterns that don't fit these idealized assumptions, leading to suboptimal performance. Existing "minimal regret" approaches aim to bridge this gap by minimizing the difference between an observer's performance and that of an ideal, clairvoyant observer that knows the future. However, these methods have been computationally challenging and difficult to implement, especially when dealing with both process and measurement noise. This paper tackles the challenge of making minimal regret estimation practical and effective for linear time-varying systems.
Methods
The core contribution of this paper is a new convex formulation for the minimal regret state estimation problem. They achieve this by:
- System Level Synthesis (SLS) Adaptation: They adapt the System Level Synthesis framework, originally developed for controller design, to the state estimation problem. This involves a novel parametrization of the system dynamics that can handle both disturbance and measurement noise simultaneously.
- Convex Optimization: The key insight is that by using this SLS parametrization, the finite-horizon minimal regret problem can be transformed into a convex optimization problem. This is a significant improvement over previous approaches that were often iterative or non-convex.
- Semi-Definite Programming (SDP): The convex formulation is then solved using semi-definite programming, a well-established and computationally tractable optimization technique. This makes the design of minimal regret observers practical.
- Benchmark Observers: The paper also details how to derive standard H2 (Kalman) and H∞ observers within the same SLS framework, providing a consistent basis for comparison.
Datasets
The paper does not introduce new datasets. Instead, it uses:
- Complib Benchmarks: Numerical experiments are conducted on the "NN4" system from the Complib library (Leibfritz, 2004) and average performance is evaluated on three fifth-order aircraft models ("AC1", "AC2", "AC3") from the same library.
- Nine Disturbance Patterns: The performance is tested across nine different disturbance patterns, including stochastic (uniform and Gaussian) and deterministic (sine wave, sawtooth, step, stairs, worst-case) scenarios.
Key Insights and Ideas
- Regret as a Practical Metric: The paper highlights that minimizing regret offers a way to design observers that adapt to observed noise patterns, rather than relying on a-priori assumptions about noise distributions. This can lead to better performance when those assumptions are violated.
- Convexity is Key: Transforming the minimal regret problem into a convex optimization problem is a major breakthrough, enabling practical implementation using standard SDP solvers.
- SLS as a Unifying Framework: The use of SLS provides a unified framework for designing and comparing H2, H∞, and minimal regret observers, simplifying the analysis and design process.
- Adaptability Beyond H2/H∞: The minimal regret observer demonstrates a capability to outperform both H2 and H∞ filters in scenarios where noise is neither purely Gaussian nor strictly worst-case, suggesting it's a valuable alternative for a broader range of applications.
Key Results
- Significant Performance Improvement: Numerical experiments show that the minimal regret observer can significantly outperform both H2 and H∞ filters across various disturbance patterns, particularly when noise is not Gaussian or adversarial.
- Tractable Implementation: The proposed method allows for the practical computation of minimal regret observers using semi-definite programming.
- Comparison with Benchmarks: Table 1 in the paper quantifies the relative improvement. For instance, on the "NN4" system with a "worst" disturbance pattern, the minimal regret observer shows a relative difference of 0.26% compared to the H∞ observer (which is near optimal for this pattern), while for a "step" disturbance, it achieves a 4.37% improvement over H2 and 4.02% over H∞. Similar improvements are observed for other patterns and the AC models.
Limitations and Future Work
- Time-Varying Systems: While the framework is presented for time-varying systems, the numerical experiments primarily focus on time-invariant systems for simplicity. For time-varying systems, the observer must be recomputed at each time step if parameters change.
- Finite Horizon: The current formulation is for finite-horizon estimation. Extending this to infinite-horizon problems, potentially through Moving Horizon Estimation (MHE), is a natural next step.
- Nonlinear Systems: The paper focuses on linear systems. Extending the minimal regret concept to nonlinear systems and hybrid systems is suggested as future work.
- Constraints: While the authors mention that similar SLS formulations for controller design allow for constraint introduction, this aspect is not explored for the observer design in this paper.
Summary updated