Local Sensitivity Analysis for Kernel-Regularized ARX Predictors in Data-Driven Predictive Control

1. Problem Statement

The paper addresses a fundamental challenge in Data-Driven Predictive Control (DDPC): the propagation of parameter uncertainty from system identification to closed-loop control performance.

• The Core Mismatch: In structured ARX-based DDPC, the identification of predictor parameters is a linear problem. However, the multi-step predictor used within the Model Predictive Control (MPC) optimization is nonlinear and implicit with respect to these parameters (due to the inversion of a matrix involving the parameters).
• The Consequence: This nonlinearity makes it difficult to analytically propagate parameter uncertainty (covariance) into control-relevant prediction uncertainty. Existing methods often rely on linear approximations that ignore the implicit nature of the predictor or use computationally expensive Monte Carlo methods.
• The Goal: The authors aim to derive a tractable, first-order linearization of the implicit predictor map to bridge the gap between linear identification and nonlinear control sensitivity. This allows for:
  1. Approximating control-relevant uncertainty penalties (Final Control Error, FCE).
  2. Designing task-aware regularization kernels that penalize parameter errors most harmful to the specific control objective.

2. Methodology

A. Local Linearization of the Implicit Predictor

The authors model the system using a state-space representation in innovations form, leading to an ARX predictor. The multi-step predictor $\hat{y}_f$ is expressed as:
$$\hat{y}_f(u_f; \theta) = (I - \Phi_y(\theta))^{-1} \left( \Psi_u(\theta)u_p + \Psi_y(\theta)y_p + \Phi_u(\theta)u_f \right)$$
where $\theta$ represents the ARX parameters, and the matrices $\Phi, \Psi$ are block Toeplitz matrices constructed from $\theta$.

To handle the nonlinearity, the authors perform a first-order Taylor expansion around a nominal estimate $\bar{\theta}$:
$$\hat{y}_f(u_f; \theta) \approx \hat{y}_f(u_f; \bar{\theta}) + J(\bar{\theta}, u_f)(\theta - \bar{\theta})$$
They derive the Jacobian matrix $J(\bar{\theta}, u_f)$ analytically. Crucially, they show that the Jacobian depends affinely on the future control input sequence $u_f$. This affine structure allows the uncertainty term to be incorporated directly into the quadratic MPC objective.

B. Uncertainty Propagation (FCE Approximation)

Using the derived Jacobian, the authors approximate the covariance of the predicted output trajectory:
$$\Sigma_{y_f}(u_f) \approx J(\bar{\theta}, u_f) \Sigma_\theta J(\bar{\theta}, u_f)^\top$$
This leads to an approximate Final Control Error (FCE) cost term, $L_\Omega$, which adds a penalty to the standard MPC cost based on the expected variance of the prediction error:
$$L_\Omega(u_f) \approx \text{tr}\left( J(\bar{\theta}, u_f)^\top Q J(\bar{\theta}, u_f) \Sigma_\theta \right)$$
This term quantifies how parameter uncertainty degrades control performance.

C. Sensitivity-Based Kernel Regularization

The primary contribution is using the Jacobian to shape the regularization kernel during the identification phase.

1. Sensitivity Metric: They define a local sensitivity matrix $W(\xi) = J(\bar{\theta}, \xi)^\top Q J(\bar{\theta}, \xi)$, where $\xi$ represents the operating condition (past data and nominal future input).
2. Task-Aware Shaping: Instead of using a generic prior (like standard TC/SS kernels), they construct a task-dependent metric $\bar{W}$ by averaging $W(\xi)$ over representative task realizations.
3. Two-Stage Estimation:
   • Stage 1: Identify a nominal model $\bar{\theta}$ and baseline kernel $K$ using standard Empirical Bayes (EB) tuning.
   • Stage 2: Refine the estimate by adding a sensitivity-based penalty term $\mu(\theta - \bar{\theta})^\top \bar{W} (\theta - \bar{\theta})$ to the cost function. This penalizes parameter variations in directions that are locally most sensitive to control performance, while preserving the computational benefits of quadratic regularization.

3. Key Contributions

1. Analytical Jacobian Derivation: The paper provides a closed-form derivation of the Jacobian for the implicit ARX multi-step predictor, explicitly separating the dependence on past data and future control inputs.
2. Dual Utility of the Jacobian: The same Jacobian is used to:
   • Approximate the uncertainty penalty in the MPC cost (FCE formulation).
   • Construct a task-dependent sensitivity metric for shaping the identification kernel.
3. Practical Regularization Strategy: The authors propose a "sensitivity-shaping" kernel that targets parameter directions critical for control, offering a refinement over standard identification-oriented priors.

4. Experimental Results

The framework was tested on a second-order linear system under two regimes:

• Informative-Data Regime (Strong Excitation):

  • When data is abundant and informative, standard ARX identification (OLS) and baseline kernel regularization (SS) perform similarly well.
  • The proposed sensitivity shaping and FCE terms offered only marginal improvements, as the identification problem was already well-conditioned.

• Weak-Excitation Regime (Low Excitation/High Uncertainty):

  • This is the critical test case where predictor uncertainty is high.
  • Baseline SS Regularization: Provided substantial robustness gains over plain OLS, stabilizing the closed-loop system where OLS failed.
  • Sensitivity Shaping (SS+W): Provided further improvement over the baseline SS regularization. The results showed reduced output variance and lower closed-loop costs.
  • FCE Term: Adding the FCE cost term directly to the MPC objective improved performance over OLS but did not significantly improve upon already regularized predictors.
  • Posterior Covariance: Table I shows that the sensitivity-shaping method (SS+W) significantly reduced the trace of the posterior covariance matrix ($\text{tr}(\Sigma_\theta)$) compared to SS, confirming that the method successfully shrunk uncertainty in the most critical directions.

5. Significance and Conclusion

• Bridging Identification and Control: The paper successfully bridges the gap between linear system identification and nonlinear control performance analysis. It demonstrates that "identification for control" requires not just accurate parameter estimation, but estimation of parameters in directions that matter for the specific control task.
• Weak-Excitation Robustness: The method is particularly valuable in data-scarce or weakly excited regimes, where standard regularization is insufficient. The proposed sensitivity shaping acts as a "targeted" regularization, shrinking uncertainty where it hurts control performance most.
• Computational Efficiency: By maintaining a linear regression structure with a modified quadratic penalty, the approach remains computationally efficient (closed-form solution) compared to non-convex or sampling-based alternatives.
• Limitations: The current approach relies on a local linearization, which may be less accurate under large parameter deviations or strong nonlinearities. Future work is suggested for automated hyperparameter tuning of the sensitivity weight $\mu$.

In summary, this work provides a rigorous mathematical framework to quantify and mitigate the impact of identification uncertainty on control performance, offering a practical method to enhance DDPC robustness in challenging data regimes.