Q-Fly Architecture in Adaptive Control & Optimization

• Q-Fly Architecture is an ambiguous label referring to both adaptive LPV control augmented by an L1 mechanism and dynamic surrogate selection in CFDNNetAdapt.
• It employs a piecewise-constant adaptation law to update uncertainty estimates in real time and iteratively selects MLP architectures based on IGD metrics.
• The frameworks ensure robust closed-loop behavior and reduced computational costs by replacing expensive simulations with reliable, dynamically optimized surrogates.

Searching arXiv for "Q-Fly Architecture" and the cited papers to ground the article in relevant sources. “Q-Fly Architecture” is not introduced as a formal method name in the available arXiv sources. Instead, the term can only be reconstructed from two distinct technical contexts: a generic $\mathcal{L}_1$ adaptive control plus LPV gain-scheduling framework validated in simulation on an F-16 short-period dynamics model (Snyder et al., 2021), and CFDNNetAdapt, a surrogate-assisted multi-objective optimization framework in which the MLP architecture is determined on the fly during optimization (Kubíčková et al., 3 Dec 2025). In this sense, the phrase denotes an interpretive label rather than a canonical architecture. The documented content therefore concerns, on the one hand, adaptive augmentation of a baseline LPV controller and, on the other, iterative surrogate selection within a CFD-driven optimization loop.

1. Terminological status and source alignment

The supplied sources do not present a single architecture formally named “Q-Fly.” One source explicitly states that the provided paper is not actually about “Learn-to-Fly (L2F)” or a UAV Q-Fly architecture, and that it does not contain a Learn-to-Fly framework, a Q-Fly architecture, switched reference models, UAV flight-test validation, or the detailed equations/figures for an L2F end-to-end pipeline (Snyder et al., 2021). The other source states that the text does not contain a formal method named “Q-Fly,” and that the actual named framework is CFDNNetAdapt (Kubíčková et al., 3 Dec 2025).

Source	Actually described	Not supported by the supplied text
(Snyder et al., 2021)	$\mathcal{L}_1$ adaptive control + LPV gain-scheduling framework	Learn-to-Fly, UAV Q-Fly architecture, switched reference models, UAV flight tests
(Kubíčková et al., 3 Dec 2025)	CFDNNetAdapt with on-the-fly MLP architecture determination	A separate named algorithmic family called Q-Fly

This suggests that “Q-Fly Architecture” is best treated as an ambiguous secondary label. A plausible implication is that, in one usage, it refers to an adaptive flight-control augmentation architecture, whereas in another it refers to the on-the-fly architecture-selection mechanism inside CFDNNetAdapt.

2. $\mathcal{L}_1$ adaptive LPV control architecture

The control architecture described in the supplied material is an adaptive augmentation layer built on top of a baseline LPV controller. Its main components are a baseline controller, a real-time aerodynamic / uncertainty modeling module, an $\mathcal{L}_1$ adaptive control augmentation, and a learning/adaptation law based on a piecewise-constant adaptation law (Snyder et al., 2021).

The baseline controller is a nominal gain-scheduled controller for the LPV plant. It provides the desired operating-point-dependent closed-loop behavior. Typical examples mentioned in the introduction are LQR, PID, and gain-scheduled controllers. The uncertainty description is not a learned aerodynamics model from flight data; instead, the plant is modeled as an LPV system with matched unknown nonlinearities, unknown input gain, and possibly time-varying and state-dependent disturbances. The adaptive component estimates these uncertainties online.

The adaptive channel follows the standard $\mathcal{L}_1$ separation principle: estimation and control are decoupled by a low-pass filter, the adaptation can therefore be very fast without destroying robustness, and the adaptive law estimates the uncertainty term or terms while the filtered estimate generates a bounded compensating control. The closed-loop structure is summarized as

$u(t)=u_b(t)+u_{ad}(t),$

where $u_b(t)$ is the baseline LPV controller output and $u_{ad}(t)$ is the adaptive augmentation.

The paper’s main novelty, as summarized in the supplied material, is the use of a piecewise-constant adaptation law for LPV systems. The uncertainty estimate is updated at discrete adaptation instants, held constant between updates, and faster update rates improve performance. The core guarantee stated in the summary is that the adaptive closed-loop input/output signals remain close to those of an ideal non-adaptive reference system, and that the bound can be made arbitrarily small by reducing the adaptation sample time (Snyder et al., 2021).

3. Reference-system analysis, predictor role, and validation limits

A central analytical device is the introduction of a non-adaptive reference system that assumes perfect knowledge of uncertainties. It represents the ideal closed-loop behavior if the controller knew the true plant mismatch. The actual adaptive system is then compared against this ideal reference system, and the stated guarantee concerns a bounded transient mismatch between adaptive closed-loop signals and reference-system signals (Snyder et al., 2021).

The state predictor is described as simulating the nominal LPV dynamics in real time, incorporating the adaptive estimate of the uncertainty, comparing predicted state to measured state, and using prediction error to update the adaptive parameter estimate. The prediction error is written in the supplied material as

$\tilde{x}(t)=\hat{x}(t)-x(t),$

or equivalently $x(t)-\hat{x}(t)$, depending on convention.

The performance discussion emphasizes guaranteed transient performance, robustness under fast adaptation, and reduced sensitivity to initial uncertainty mismatch. The qualitative result given in the summary is that large initial error does not destabilize the loop, the system tracks the reference model with a bounded transient gap, and smaller adaptation sample time reduces that gap.

The validation described in the supplied material is simulation on the short-period dynamics of an F-16 aircraft. The same source explicitly states that, contrary to the query’s mention of UAV flight tests, the paper does not document UAV flight-test validation in the supplied text (Snyder et al., 2021). It also states that no switched reference-model architecture is presented there; instead, the paper uses a single LPV reference system and compares the adaptive closed loop against a non-adaptive ideal reference system.

4. CFDNNetAdapt and on-the-fly architecture determination

Within the second source, the relevant architecture is CFDNNetAdapt, a surrogate-assisted multi-objective optimization framework in which the surrogate is an MLP and the MLP architecture is determined on the fly during optimization rather than fixed in advance (Kubíčková et al., 3 Dec 2025). If “Q-Fly” is being used to refer to that on-the-fly architecture-selection mechanism, then its role is exactly adaptive architecture search within CFDNNetAdapt.

The overall workflow is described as a cycle between expensive simulation-based global search and surrogate training. The steps are: Run NSGA-II with the true cost function $\mathcal{L}_1$0, which in the real problem is a CFD model in OpenFOAM; collect evaluated data points $\mathcal{L}_1$1 in a data bank; train multiple MLP surrogates with different hidden-layer sizes; use each trained MLP as a surrogate inside another NSGA-II run to predict a Pareto set/front; select the best surrogate architecture using a comparison of its predicted Pareto set against the current CFD-based Pareto set; verify the selected surrogate’s Pareto set with a few CFD evaluations; and, if verification is good enough, stop using CFD for the main search and let the MLP replace simulation for the remainder of the global search, with CFD then used only for verification and correction (Kubíčková et al., 3 Dec 2025).

The candidate surrogates are MLPs with a fixed number of hidden layers and variable hidden-layer widths. The input layer size equals the number of design parameters $\mathcal{L}_1$2, the output layer size equals the number of objectives $\mathcal{L}_1$3, the number of hidden layers $\mathcal{L}_1$4 is fixed for all candidates, and the transfer function is always $\mathcal{L}_1$5. Only the widths of the hidden layers vary. The generic size rule is

$\mathcal{L}_1$6

followed by clipping to

$\mathcal{L}_1$7

In the first iteration, the averages $\mathcal{L}_1$8 are preset. In later iterations, the averages are inherited from the best MLP found in the previous iteration (Kubíčková et al., 3 Dec 2025).

Architecture suitability is measured by the inverted generational distance (IGD):

$\mathcal{L}_1$9

The paper chooses the MLP with the lowest IGD. Architecture search is performed once per CFDNNetAdapt iteration: generate data, train $\mathcal{L}_1$0 different MLP architectures, select the best one, verify it by CFD, and either terminate or carry its architecture forward.

5. Verification gate, stopping rule, and surrogate deployment

The surrogate does not immediately replace CFD. The paper uses a validation gate: train candidates on the current accumulated dataset; pick the best architecture by IGD; randomly select $\mathcal{L}_1$1 points from the surrogate’s predicted Pareto set; re-evaluate those points with CFD; compute mean verification error $\mathcal{L}_1$2; and, if $\mathcal{L}_1$3, accept the surrogate and terminate the iteration (Kubíčková et al., 3 Dec 2025).

The verification error is defined as

$\mathcal{L}_1$4

If the surrogate passes verification, it is used as the cost-function evaluator in the remainder of the global search, while CFD is retained only for verification, correction, and stopping criteria. The source explicitly describes this as CFD being replaced by the DNN in the optimization loop, not as a one-off fit.

The optimization problem is posed as

$\mathcal{L}_1$5

For multiple objectives, Pareto dominance is used: one solution $\mathcal{L}_1$6 dominates another $\mathcal{L}_1$7 if

$\mathcal{L}_1$8

and

$\mathcal{L}_1$9

This formulation shows that the architecture is not merely a neural-network training routine. It is an iterative optimization stack in which architecture choice, surrogate trust, and expensive simulation are coupled by explicit selection and verification criteria.

6. Benchmarks, ejector case, and practical significance

The method was tested on ZDT1, ZDT2, ZDT3, ZDT4, and ZDT6, and was compared against plain NSGA-II, SOCEMO using RBF surrogates, and a CFDNNetAdapt variant with GPR surrogates (Kubíčková et al., 3 Dec 2025). On ZDT1 and ZDT2, surrogate-assisted methods converged faster than NSGA-II. On ZDT3, which has a discontinuous Pareto front, the DNN-based CFDNNetAdapt was described as the most flexible, while GPR and RBF surrogates struggled. On ZDT4, all surrogate methods failed because the parameter space is too topologically difficult. On ZDT6, CFDNNetAdapt achieved the lowest indicator values and the least distorted front.

The real problem is a 14-parameter constrained shape optimization of a single-phase ejector, focusing on the converging part and diffuser. The objectives are to maximize energy efficiency $\mathcal{L}_1$0 and minimize total length $\mathcal{L}_1$1, with

$\mathcal{L}_1$2

The cost vector is defined as

$\mathcal{L}_1$3

For this ejector case, CFDNNetAdapt needed 10 iterations / 6,600 samples to reach $\mathcal{L}_1$4. The best MLP found had 5 fully connected layers with widths $\mathcal{L}_1$5. Compared with continuing the baseline CFD-NSGA-II search, the reported computational cost was about 6,912 core-hours for baseline continuation, about 152 core-hours for the CFDNNetAdapt architecture search, and about 480 core-hours for re-evaluation of 500 surrogate-generated designs (Kubíčková et al., 3 Dec 2025). The surrogate-guided search found new ejector designs expanding the Pareto front, and four optimized shapes were 3D printed and experimentally validated, with experimental trends agreeing with the optimized and CFD-predicted trends.

These results suggest a specific practical significance for the on-the-fly architecture mechanism: it is not only a model-selection heuristic, but a device for deciding when a learned surrogate is sufficiently reliable to replace an expensive simulator in the main search loop.

7. Misconceptions, interpretive use, and relation to “Q-Fly”

Two misconceptions are explicitly contradicted by the supplied material. First, the control source does not provide a Learn-to-Fly pipeline, a UAV Q-Fly architecture, switched reference models, or UAV flight-test validation; it provides a gain-scheduled LPV baseline controller augmented by an $\mathcal{L}_1$6 adaptive loop with piecewise-constant estimation and low-pass filtered compensation, designed to keep the adaptive closed loop close to an ideal uncertainty-known reference system (Snyder et al., 2021). Second, the optimization source does not define a formal method named “Q-Fly”; the named framework is CFDNNetAdapt (Kubíčková et al., 3 Dec 2025).

Accordingly, “Q-Fly Architecture” is best understood as an informal label applied to one of two documented ideas. In the control interpretation, it denotes the architecture composed of a baseline gain-scheduled controller, a state predictor, a piecewise-constant adaptation law, and low-pass filtered adaptive compensation. In the optimization interpretation, it denotes the mechanism that dynamically generates candidate MLP architectures, evaluates them using surrogate-based NSGA-II, picks the architecture with minimum IGD against CFD-derived Pareto data, verifies it using CFD, and, if acceptable, deploys it as the simulation replacement.

A plausible implication is that the phrase “Q-Fly” compresses the notion of adaptation performed “on the fly,” but the sources do not support a single canonical definition. The technically documented architectures are instead those of $\mathcal{L}_1$7-augmented LPV control and CFDNNetAdapt’s iterative surrogate selection.