Multi-Objective H∞ Control

• Multi-objective H∞ control is a robust synthesis method that combines multiple performance objectives using structured weighting functions to manage trade-offs like tracking and disturbance rejection.
• It utilizes advanced techniques such as LMIs, Riccati equations, and augmented plant constructions to derive controllers that meet strict frequency-domain norms.
• The approach finds practical applications in areas like grid-connected converters, vehicle platoons, and digital PID design, enabling decentralized and scalable control solutions.

Multi-objective $H_{\infty}$ control is a robust control methodology in which multiple, sometimes conflicting, performance objectives are posed as constraints or cost functions within a single $H_{\infty}$ synthesis framework. The technique utilizes generalized plants with structured weighting functions to simultaneously regulate competing channels—tracking, disturbance attenuation, coupling, or safety—and often employs Linear Matrix Inequality (LMI) or Riccati equation methods for controller construction. Recent advances also enable multi-norm and multi-channel formulations, decentralized certification, model-free synthesis, and integration of time-domain constraints. Applications span grid-connected converter networks, vehicle platoon string stability, digital PID parameter space design, and distributed systems with integral quadratic constraints (IQCs).

1. General Problem Formulation

Multi-objective $H_{\infty}$ control starts from a generalized plant $P(s)$ partitioned as

$$\begin{pmatrix} z \ y \end{pmatrix} = \begin{pmatrix} P_{11}(s) & P_{12}(s) \ P_{21}(s) & P_{22}(s) \end{pmatrix} \begin{pmatrix} w \ u \end{pmatrix}$$

where $w$ contains reference/disturbance signals, $u$ are control inputs, $y$ are measured states/regulation errors, and $z$ are performance outputs, often including regulated, coupling, or safety channels (Huang et al., 2019). Multi-objectivity is introduced through $n$ frequency-domain weighting functions $W_i(s)$ shaping specific input-output transfers, which are embedded in a weighting matrix $\mathcal{W}(s) \in \mathbb{R}^{p \times m}$. The central objective is to design a controller $K$ so that each weighted transfer $\mathcal{W}_{ij}(s) P^{\mathrm{cl}}_{ij}(s)$ is bounded in the $\infty$-norm.

The canonical synthesis problem becomes: $$\min_K \left\| \mathcal{W}(s) \circ T_{zw}(s) \right\|_\infty = \min_K \max_{\omega} \overline{\sigma} \left[ \mathcal{W}(j\omega) \circ T_{zw}(j\omega) \right]$$ where $T_{zw}(s)$ is the closed-loop transfer from $w$ to $z$ under feedback $u=K\,y$. An auxiliary scalar $\gamma$ is introduced to enforce

$$\| \mathcal{W}(s) \circ T_{zw}(s) \|_\infty < \gamma$$

across all frequencies (Huang et al., 2019). Multi-objective extensions admit additional performance channels ($\ell_2$, energy-to-peak, peak-to-peak gains), multi-norm cost functions, and decentralized or structural constraints (Schwenkel et al., 28 Mar 2025, Kayacan, 2021).

2. Augmented Plant Construction and Weighting Strategies

The plant is augmented by pre-multiplying each channel with its weighting function: $$P_a(s) = \begin{pmatrix} W(s) & 0 \ 0 & I \end{pmatrix} \begin{pmatrix} P_{11}(s) & P_{12}(s) \ P_{21}(s) & P_{22}(s) \end{pmatrix} = \begin{pmatrix} W(s)P_{11}(s) & W(s)P_{12}(s) \ P_{21}(s) & P_{22}(s) \end{pmatrix}$$ where $W(s)=\operatorname{diag}(W_1(s),\ldots,W_n(s))$ (Huang et al., 2019). Frequency-domain weights are tuned so that high gain at a particular frequency band enforces tight regulation or disturbance rejection, while low gain relaxes the requirement and may improve robustness to unmodeled dynamics. In parameter-space digital PID design, similar weightings generate mixed-sensitivity bounds $\|W_S S + W_T T\|_\infty \leq \gamma$ mapped directly to controller gain regions (Wang et al., 2020).

Trade-offs are managed by adjusting cut-off poles/zeros in the weights to allocate bandwidth between regulation, tracking, noise attenuation, and inter-channel coupling (e.g., voltage/power and admittance-shaping in grid converters). In multi-objective settings, weights are selected per performance channel to produce feasible intersection regions in controller parameter space.

3. Solution Methodologies: LMI and Riccati Approaches

Controller synthesis in the multi-objective $H_{\infty}$ framework is accomplished either via nonsmooth optimization for static controllers or LMI (Linear Matrix Inequality) formulations and coupled generalized Riccati equations for dynamic ones. The LMI-based design for a state-space augmented plant typically seeks matrices $X > 0$, $Y > 0$ and controller parameters satisfying: $$\begin{pmatrix} A X + X A^\top + B_2 \hat{C} + (B_2 \hat{C})^\top & B_1 & X C_1^\top + \hat{C}^\top D_{12}^\top \ B_1^\top & -\gamma I & D_{11}^\top \ C_1 X + D_{12} \hat{C} & D_{11} & -\gamma I \end{pmatrix} < 0$$ and dual constraints, with controller realization $K(s) = \hat{C} X^{-1}$ (Huang et al., 2019, Kayacan, 2021).

In the presence of unmodeled stochasticity or plant uncertainties, multi-norm objectives and IQCs are embedded within coupled LMIs utilizing a congruence transformation and affine change of variables. The multi-objective cost

$$\min_K \alpha_1 \|T_{zw}(K)\|_\infty + \alpha_2 \|T_{zw}(K)\|_{\mathrm{e-p}} + \alpha_3 \|T_{zw}(K)\|_{\mathrm{p-p}}$$

is imposed, and finite-horizon time-domain IQCs with terminal cost enable handling dynamic or nonlinear uncertainty blocks (Schwenkel et al., 28 Mar 2025). Sufficient conditions for $\ell_2,\rho$-stability and performance are established via explicit LMIs for each norm.

In mixed $H_2/H_{\infty}$ problems on stochastic discrete-time systems, optimal controllers may be recovered from coupled generalized algebraic Riccati equations (Jiang et al., 2023). Recently, model-free algorithms based on Q-learning and least-squares evaluation permit synthesis without plant matrices, allowing direct multi-objective tuning by sampled trajectory data.

4. Multi-objective Constraints and Trade-offs

Multi-objective $H_{\infty}$ control enables explicit handling of competing requirements:

• Robust stability: Closed-loop poles constrained in left-half plane (continuous time) or within unit disk (discrete time).
• Disturbance attenuation: Worst-case exogenous inputs bounded by $\gamma$.
• Performance shaping: Weighted sensitivity, complementary sensitivity, power regulation, string stability, etc.
• Decentralization and scalability: Synthesis methods certify global stability via local channel-wise small-gain conditions, e.g., in converter networks or vehicle platoons.

Constraints are enforced either via parameter space intersection (PID design), validation of matrix inequalities (LMI/Riccati solutions), or frequency-domain spectral mapping. In decentralized systems, sufficient small-gain conditions guarantee global adherence when each local block satisfies $$\|F^{-1}(s) Y_i(s)\|_\infty < \lambda_{\min}(Q_{\mathrm{red}})$$ (Huang et al., 2019).

The feasible region of controllers is typically characterized by the intersection of stability, phase margin, gain margin, and $H_{\infty}$ bounds, with several design points tuned to maximize one objective while minimally impacting others (Wang et al., 2020).

5. Applications in Power Systems, Vehicle Platoons, and Digital Control

Multi-objective $H_{\infty}$ control underpins robustness and optimality in several active domains:

• Grid-connected converters: Systematic frequency-weighted $H_{\infty}$ schemes enable robust tracking, grid disturbance rejection, and admittance shaping; decentralized small-gain certificates ensure network-wide stability by local design (Huang et al., 2019).
• Cooperative adaptive cruise control (CACC): Controllers are synthesized for each vehicle to simultaneously minimize regulated error, attenuate external disturbances, and enforce string stability ($\|T(j\omega)\|_\infty \leq 1$), preventing disturbance amplification down platoons (Kayacan, 2021).
• Digital PID parameter-space design: Multi-objective constraints map directly to polygonal solution regions in discrete gain space, with time-domain simulation validating plant tracking, stability, and robustness (Wang et al., 2020).
• Stochastic/discrete systems with IQCs: Peak and energy norms are included for safety-critical applications, with convex synthesis LMIs providing explicit guarantees under general uncertainty descriptions (Schwenkel et al., 28 Mar 2025).
• Model-free synthesis: Q-learning delivers mixed $H_2/H_{\infty}$ controllers using trajectory data, with finite-iteration admissibility and convergence guarantees (Jiang et al., 2023).

6. Decentralized and Scalable Certification Techniques

A major advancement is decentralized stability certification via the small-gain theorem. In networked converter systems, the closed-loop admittance map is analyzed: $${\bf L}(s) = -[Q_{\mathrm{red}} \otimes F(s)]^{-1} {\bf Y}(s)$$ and the network’s global $\|{\bf L}(s)\|_\infty < 1$ is asserted if each local block satisfies: $$\|F^{-1}(s) Y_i(s)\|_\infty < \lambda_{\min}(Q_{\mathrm{red}})$$ This enables independent local controller tuning with a guaranteed network-wide stability margin (Huang et al., 2019). Similarly, in vehicle platoons, inclusion of complementary sensitivity constraints in the LMI set guarantees string stability without recourse to ad hoc design cycles (Kayacan, 2021).

7. Limitations, Computational Aspects, and Future Directions

Multi-objective $H_{\infty}$ synthesis presents several computational and methodological considerations:

• Non-convexity: Joint multiplier–controller design for IQC-based synthesis can induce conservatism and solution bilinearity; alternating analysis–synthesis steps or fixing multiplier basis may alleviate this (Schwenkel et al., 28 Mar 2025).
• Parameter-space visualization: For PID design, only two-dimensional gain regions can be plotted at a time; multi-dimensional objectives may compress solution space or cause intersection voids (Wang et al., 2020).
• Discrete-time advantage: Peak-norm constraints lack clean frequency-domain expressions in continuous time but admit rigorous treatment in discrete-time via time-domain LMIs and terminal-cost IQCs (Schwenkel et al., 28 Mar 2025).
• Model-free synthesis: Data requirements for Q-learning algorithms scale with quadratic state/input dimensions; unbiased exploration via trigonometric probes is required for convergence (Jiang et al., 2023).

Future research includes automated multiplier basis selection, integration with model-predictive control, and extensions to nonlinear/lipschitz uncertainties, with particular emphasis on scalable design for large interconnected systems and sampled-data/hybrid-systems continuous-time extensions (Schwenkel et al., 28 Mar 2025).

---

References

(Huang et al., 2019, Kayacan, 2021, Jiang et al., 2023, Wang et al., 2020, Schwenkel et al., 28 Mar 2025)