A controller that is tuned to be optimal is optimal only for the disturbances its designer anticipated. For linear systems, one family of such controllers comes from regret-optimal control: they compete with a hypothetical controller that sees the future, and they are computed by reducing the design to the Nehari problem, a classical operator-theory question about how well a causal system can approximate an anti-causal one. Every member of this family carries a knob fixed before deployment. Examples are the Wasserstein ambiguity radius $R$ in distributionally robust regret control, the choice between energy-regret and pathlength-regret objectives, and, at the extremes, the plain average-case ($H_2$) and worst-case ($H_\infty$) designs. The papers that introduced these knobs each say that choosing them online is open. Changing a dynamic controller's parameters while it runs is dangerous, though, because switching between individually stable feedback laws can destabilize the loop.
We propose a specific way to make the knob adaptive without risking instability. First, precompute a small bank of Nehari-based controllers covering the knob's range. Then express each one in Youla coordinates, which describe every stabilizing controller for a plant through one stable "free" filter $Q$, with the closed loop depending affinely on $Q$. Finally, let an online learning algorithm choose convex mixing weights over the bank's $Q$ outputs at every time step. In Youla coordinates the signal feeding $Q$ does not depend on which $Q$ is used, for an exactly known plant. That gives two consequences. Any time-varying mixture stays stable. And the learner can compute exactly what the cost would have been under every alternative mixture, so it gets full counterfactual feedback without exploring. The main claim is that the loop can shift between $H_2$-like behavior under benign noise and regret-optimal or $H_\infty$-like behavior under structured noise with a stability guarantee, and that its cumulative cost comes provably close to the best mixture chosen in hindsight.
What exists and what is missing
Several papers supply the bank members. The Nehari reduction for full-information control was introduced by Regret-Optimal Estimation and Control. Its closed-form version, using one Riccati and two Lyapunov equations, comes from Regret-Optimal Full-Information Control. The measurement-feedback version, Regret-Optimal Control under Partial Observability, is itself built through Youla parameterization and spectral factorization. That is why our mixing construction fits naturally on top of it. Pathlength-optimal designs, whose regret scales with how much disturbances change over time rather than their total energy, come from Online Estimation and Control with Optimal Pathlength Regret. Distributionally robust regret controllers at a given radius $R$, solved in the frequency domain by Frank-Wolfe and then fit by a low-order rational approximation, come from Infinite-Horizon Distributionally Robust Regret-Optimal Control. We reuse all of these as fixed components.
A second group of papers motivates the gap. Kargin et al. state that adapting the nominal noise model online is open. Generalised Regret Optimal Controller Synthesis for Constrained Systems states that automated tuning of its disturbance weighting is open. The pathlength paper fixes its performance level $\gamma$ by offline bisection. Robust regret optimal control handles plant uncertainty with offline DK-iteration against a static uncertainty bound. None of these papers adapts anything at runtime.
The closest online method is Optimal Dynamic Regret in LQR Control. It reduces control with known dynamics to online linear regression over a high-dimensional policy class and keeps iterates in a safe set with a barrier function. We differ in two ways. We learn over a $K$-dimensional simplex of Nehari solutions, not a raw policy space. And stability comes from Youla structure instead of a barrier.
The adaptive-parameter literature constrains the design more than it supplies it. Adaptive balancing of gradient and update computation times uses TCP-like additive and multiplicative rules to steer a single scalar parameter. Proximal alternating penalty algorithms uses analytic penalty updates. A Trust-Region Method For Nonsmooth Nonconvex Optimization uses safeguarded radius updates. All three work inside static optimization problems, where changing a parameter has no dynamic side effects. They also report sensitivity to the meta-parameters their rules introduce. We take the TCP-style rule as a baseline and take meta-parameter sensitivity as something our method must address, through a learning rate that is tuned by a known bound or made parameter-free.
Core Claim
Fix a nominal plant with a stabilizing nominal controller. Write the performance output as $z = T_1 w + T_2\, Q\, (T_3 w)$, where $w$ is the disturbance, $T_1, T_2, T_3$ are stable maps fixed by the plant and nominal controller, and $Q$ is the stable Youla parameter. Let $Q_1,\dots,Q_K$ be the bank of Nehari-based designs, all expressed with respect to the same nominal controller. The signal $e = T_3 w$ can be computed from measurements and does not depend on $Q$. We therefore run every $Q_i$ on $e$ in parallel and apply
$$u_t = \sum_{i=1}^{K} \lambda_{t,i}\, (Q_i e)_t, \qquad \lambda_t \in \Delta_K,$$
where $\Delta_K$ is the probability simplex. A time-varying convex combination of bounded operators is bounded by the largest of their norms, so the closed loop remains stable for any weight sequence.
Because $z$ is affine in the mixed output, the learner can run $T_2$ on each $Q_i e$ in parallel. This lets it evaluate the counterfactual running cost $\ell_t(\lambda) = \|z_t(\lambda)\|^2$ for every fixed $\lambda$, and each such cost is a known convex quadratic. An online-convex-optimization learner with memory handling, using exponentiated gradient with fixed-share for drifting regimes, then satisfies
$$\sum_{t=1}^{T} \ell_t(\lambda_t) - \min_{\lambda \in \Delta_K} \sum_{t=1}^{T} \ell_t(\lambda) \le O\!\left(\sqrt{T \log K}\right) + C_{\text{mem}},$$
where $T$ is the horizon. The $\sqrt{T\log K}$ term is the standard exponentiated-gradient rate for bounded convex losses on a simplex. The term $C_{\text{mem}}$ accounts for the lag between changing weights and their effect on $z$, and it is controlled by the decay rate of $T_2$ and a bound on how fast the weights move. The mixture inherits every per-member worst-case guarantee in the weaker form set by the norm bound, and its average-case cost tracks the best member for the current regime.
Research Program
The first component is theory. We will prove stability and the meta-regret bound above for full-information and measurement-feedback plants. We will also state a switching-comparator version, in which the learner competes with the best sequence of mixtures that changes at most $S$ times. We will characterize what breaks under model mismatch: when the true plant differs from the nominal one, $e$ picks up a component that depends on $Q$, and the small-gain argument then needs a margin. We will bound the tolerable mismatch in terms of $\max_i \|Q_i\|_\infty$.
The second component is construction. We will build banks on standard benchmarks. The core bank has $K=5$ members: $H_2$, full-information regret-optimal, pathlength-optimal, $H_\infty$, and one distributionally robust controller at a mid-range $R$. A larger bank adds a grid of radii $R$. We will check that the per-step cost of evaluating counterfactuals, which amounts to $K$ parallel filter banks, stays within real-time limits.
The third component is empirical. We will test whether mixing beats every fixed member and every naive adaptation rule under non-stationary disturbances, and whether it loses little under stationary ones.
Baselines
The naive reference points are the individual bank members deployed alone, and the most important of these is the single full-information regret-optimal controller. It is already a compromise design. On the AC15 aircraft model, Regret-Optimal Full-Information Control reports a Frobenius (average-case) norm of $1.88 \times 10^4$ against $1.61 \times 10^4$ for $H_2$, and an operator (worst-case) norm of $2.28 \times 10^6$ against $2.72 \times 10^6$ for $H_2$. Beating it means cumulative cost below it on non-stationary sequences while never exceeding its worst-case operator norm by more than the mixture bound allows. On measurement-feedback problems, the reference is the partial-observability regret-optimal controller, which reports regret 3.1322 on the unstable synthetic System 1, against 6.9475 for $H_2$ and 3.6365 for $H_\infty$ (Regret-Optimal Control under Partial Observability). We will reproduce these figures before running anything new. For distributionally robust regret control, we found no reported numerical result on Lacuna, since results there are shown only in plots. We will therefore recompute that controller ourselves on REA4 across $R$.
The strongest competitor that adapts is periodic re-estimation followed by hard switching. At fixed intervals, estimate the disturbance spectrum from a sliding window, pick $R$ or the best bank member by its estimated cost, and switch the controller. This is what a practitioner would build in response to Kargin et al.'s open problem. It gets the same bank, the same window budget, and a grid search over window length and interval on held-out sequences. We will implement it in two ways: switching in the plant's control-input space, which is what practitioners actually do, and switching in Youla space. Separating the two shows whether our gain comes from learning or merely from coordinates.
A TCP-style rule adapted from Karimireddy et al. will steer a single scalar knob multiplicatively from the observed cost trend. The strongest high-dimensional online competitor is the method of Baby and Wang, run with the same known plant and horizon.
The key ablation mixes the $K$ controllers' outputs directly in control-input space, each controller running its own internal state from measurements, instead of mixing in Youla coordinates. This isolates whether Youla mixing is what supplies stability. We expect, and will try to exhibit, sequences of weights under which this ablation destabilizes an unstable plant.
A second ablation feeds the learner only the realized loss rather than counterfactual losses. This tests whether full-information feedback matters.
Tuning is matched across methods. Every adaptive method gets the same hyperparameter budget on a disjoint set of training disturbance sequences, and it is scored on held-out sequences. The "best fixed mixture in hindsight" is reported as an oracle, not a competitor.
Why not just…?
Why not just deploy the single regret-optimal controller, since it already balances average and worst case? This is the most serious threat. The AC15 numbers show it costs only about 17% more than $H_2$ on average while cutting the worst-case peak. If disturbance regimes change rarely, mixing may win by only a few percent. The experiments settle this directly by reporting the gap to the regret-optimal controller on stationary Gaussian, stationary structured, and switching sequences. We state in advance that the proposal is only worthwhile if the gain on switching sequences clearly exceeds that 17% headroom, or if pathlength-type regimes appear. In those regimes the pathlength paper reports orders-of-magnitude gaps under constant offsets, which no single energy-regret design closes.
Why not just use an adaptive Kalman or $H_2$ controller with noise covariances rescaled from the observed residuals? Rescaling covariances changes how large the design assumes the noise is, but not its whiteness assumption. Regret-Optimal Estimation and Control reports roughly 50% lower squared error than the extended Kalman filter under sinusoidal noise, and the pathlength paper reports orders-of-magnitude gains under constant offsets. Scale adaptation cannot fix either failure, because both come from temporal structure in the noise. We include covariance-matched $H_2$ control as a baseline on exactly those sequences.
Why not just re-estimate the ambiguity radius periodically and re-solve? This is the switching baseline above. It has two weaknesses that our experiments measure. Hard switching of dynamic controllers can destabilize the loop, which we test in the control-input-space ablation. And its decisions rest on noisy spectral estimates instead of exact counterfactual costs. If Youla-space switching with good estimates matches our method, the honest conclusion is that the coordinates, not the learner, matter. We would then report that.
Why not just run the online regression method of Baby and Wang? It learns over a much larger policy class and needs barrier projections. It carries no frequency-domain certificate, and it does not use the structure of the Nehari solutions. We expect it to adapt more slowly on short horizons. The same-horizon comparison will show whether this holds.
Proposed Experiments
We will use AC15 for full-information multi-input control, with directly comparable published numbers. We will use the unstable synthetic System 1 and AC5 for measurement feedback, REA4 for the single-input distributionally robust bank, and the Boeing 747 model under multiplicative uncertainty from Liu and Seiler to test mismatch. Disturbance sequences will switch among white Gaussian, sinusoidal, sawtooth, constant-offset, and each member's worst-case input, with regime lengths drawn from a geometric distribution.
We report four metrics: cumulative quadratic cost, regret against the clairvoyant controller computed offline, meta-regret against the best fixed and best $S$-switch mixtures, and the empirical induced gain over the worst-case sequences.
Three checks decide whether the idea works. First, on switching sequences the mixture must beat every fixed member and the re-estimate-and-switch baseline in cumulative cost. Second, its worst-case induced gain must stay within the bound set by the largest member norm, while the control-input-space ablation is shown to lose stability on at least one unstable plant. Third, meta-regret must grow sublinearly in $T$ and track the $\log K$ dependence as the radius grid grows. Under plant mismatch we will map the perturbation size at which the stability margin predicted by our analysis is exhausted and compare it with the observed onset of instability.
References
- Regret-Optimal Estimation and Control
- Regret-Optimal Full-Information Control
- Regret-Optimal Control under Partial Observability
- Online Estimation and Control with Optimal Pathlength Regret
- Infinite-Horizon Distributionally Robust Regret-Optimal Control
- Generalised Regret Optimal Controller Synthesis for Constrained Systems
- Robust regret optimal control
- Optimal Dynamic Regret in LQR Control
- Adaptive balancing of gradient and update computation times using global geometry and approximate subproblems
- Proximal alternating penalty algorithms for nonsmooth constrained convex optimization
- A Trust-Region Method For Nonsmooth Nonconvex Optimization