Efficient Learning of Vehicle Controller Parameters via Multi-Fidelity Bayesian Optimization: From Simulation to Experiment

Maik Pfefferkorn; Maximilian Templer; Rolf Findeisen; Yongpeng Zhao

REVIEW 2 major objections 6 minor 34 references

Efficient Learning of Vehicle Controller Parameters via Multi-Fidelity Bayesian Optimization: From Simulation to Experiment

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A multi-fidelity Bayesian optimizer tunes a vehicle controller with a handful of real test-track runs.

desk verdict Solid simulation study; field validation is a single favorable draw that doesn't prove the multi-fidelity transfer. read the letter →

arxiv 2506.08719 v1 pith:OIT2ICR4 submitted 2025-06-10 eess.SY cs.ROcs.SY

Yongpeng Zhao , Maik Pfefferkorn , Maximilian Templer , Rolf Findeisen This is my paper · ORCID

classification eess.SYcs.ROcs.SY

keywords multi-fidelityBayesianoptimizationcontrollerparametertuningauto-regressiveGaussianprocesssimulation-to-realitytransfertrajectorytrackingcontrolexpectedimprovementvehicle

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to show that controller parameters for a vehicle trajectory-tracking system can be learned almost entirely in simulation and then confirmed with a very small number of real test-track runs. Its proposed method, AR1GP-BO, feeds a pre-collected simulation data set into an auto-regressive Gaussian process and runs Bayesian optimization on the real vehicle, so the real experiments only need to refine, not rebuild, the surrogate model. In simulation studies the method reaches a simple regret near $10^{-2}$ within ten queries, and in the field test the first real query already beat both the manually tuned parameters and the best simulation-only parameters, with true cost $0.258$ versus $0.351$ and $0.282$. A sympathetic reading takes this as evidence that the standard two-stage industrial workflow can be automated without abandoning it.

What carries the argument

The carrying object is the AR1 auto-regressive multi-fidelity Gaussian process, defined recursively by $g_h(\xi)=\rho_{h-1}g_{h-1}(\xi)+\delta_h(\xi)$, with $g_{h-1}(\xi)$ independent of the GP bias $\delta_h(\xi)$. For two fidelity levels this says the true, real-world cost is a scaled copy of the simulation cost plus a correction term. The AR1 surrogate is trained once on low-fidelity data, then used inside expected-improvement Bayesian optimization with only high-fidelity updates during field tests; this is what allows the optimizer to start from a good region and reach a competitive controller in the first real query.

What would settle it

Collect a paired data set of simulated and true costs over a broad parameter grid on a test vehicle, fit the AR1 model, and compare its held-out prediction error with a nonlinear multi-fidelity model; if the nonlinear model consistently wins and the AR1 residuals show curvature, the linearity assumption is falsified. A cheaper check is to rerun the field experiment with a deliberately poor simulation model, for example a 50% tire-force perturbation; if AR1GP-BO no longer beats single-fidelity Bayesian optimization within the first few queries, the transfer benefit depends on the linearity holding.

Watch

Extended reading notes

Core claim

The central claim is that an auto-regressive multi-fidelity Gaussian process, fed with existing low-fidelity data, lets Bayesian optimization transfer knowledge from simulation to a real vehicle without any new low-fidelity evaluations during the physical tests. The paper states the high-fidelity cost as a linear function of the low-fidelity cost plus an independent GP bias, so the real experiments can concentrate on a small number of high-value parameter queries. Across three simulated perturbed-vehicle studies, AR1GP-BO converged to a simple regret around $10^{-2}$ within 10 queries, outperforming a nonlinear multi-fidelity baseline and a single-fidelity GP baseline. On a test vehicle, 17 parameter queries were performed; the first query achieved a true cost of $0.258$, below the manually tuned cost of $0.351$ and the best simulation-only cost of $0.282$, and the method did so while preserving the two-stage simulation-then-validation workflow.

Load-bearing premise

The load-bearing premise is that the real-world cost is a linear rescaling of the simulation cost plus an independent correction; if the true simulation-to-reality map is strongly nonlinear, the surrogate can point the optimizer at the wrong parameters.

Editorial extensions

If this is right

If the claim holds, automotive teams can keep the two-stage V-model but replace most manual real-world tuning with automatic simulation-informed optimization.
The number of expensive test-track runs needed to find a good controller can drop to a handful; in the reported case a single query already improved on manual tuning.
The method inherits a clear dependence on low-fidelity accuracy: perturbing the simulation model by 10% slows convergence, though it still beats a single-fidelity baseline.
Because some queried parameters led to failed laps, the paper's own outlook is that safe Bayesian optimization must be added before the approach is deployed on systems where instability is dangerous.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper's experiments, the same transfer logic should apply to any black-box controller whose simulation model captures the dominant dynamics; the natural next test is adapting the AR1 scale and bias parameters online as real data accumulate.
A direct, testable extension is to fit both AR1 and a nonlinear multi-fidelity model to paired simulation and field costs and compare predictive residuals; if the residuals are strongly nonlinear near handling limits, a nonlinear correction would be needed for deployment at the tire friction limit.
The one-real-query result is a statistical accident of a single field test, not a guaranteed property; the more robust claim to carry forward is that simulation-pretrained AR1GP-BO reduces the number of real queries needed to reach a fixed regret, which is what the simulation studies demonstrate.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First, the punchline: you should read this paper for the simulation study, not the field experiment. The field section is a single 17-query run, one vehicle, one track, with no repeated trials or error bars. The best true cost comes from the first query, and the paper states that first query was selected using only the low-fidelity data, before any high-fidelity observation. So the real-world data does not show the multi-fidelity update doing any work; it's compatible with plain simulation-based selection getting a favorable draw.

What's genuinely new is the two-stage workflow: freeze the low-fidelity simulation data set, then run high-fidelity BO with an AR1 multi-fidelity GP surrogate without querying simulation online. That is a sensible adaptation of Kennedy and O'Hagan's AR1 model to the industry V-model process. The mathematical machinery is standard; the contribution is the coupling of the frozen low-fidelity data with the AR1 surrogate and the application to vehicle controller tuning. The simulation studies are the real strength: three different model perturbations, five trials per case, and fair baselines (SFGP-BO and NARGP-BO). AR1GP-BO converges to simple regret around 1e-2 within ten queries, noticeably faster than the baselines, and the supplementary study on low-fidelity accuracy (1%/5%/10% perturbations) shows the expected trend without breaking the method. That is honest, reproducible-in-principle work.

The soft spots: the field validation is overinterpreted. First, as noted, the first-query-win means the field data can't separate the transfer contribution from simulation-only BO. Second, the AR1 linearity assumption is untested against the sim-to-real scatter in Fig. 7; the scatter is large enough that the linear map may be questionable. But that concern is secondary: even a misspecified surrogate can lead to reasonable queries, and the field result is weak regardless of the AR1 model being correct. Third, no code or data are released, which limits independent verification.

Who this is for: people working on applied Bayesian optimization for controller calibration, and automotive engineers interested in data-efficient tuning. It deserves serious peer review. My recommendation: send it out, but require a revised framing of the field experiment as a single proof-of-concept run, not as a demonstration of multi-fidelity sample efficiency, plus either repeated field trials or an explicit statement that the transfer mechanism is only validated in simulation.

Referee Report

2 major / 6 minor

Summary. The paper proposes a two-stage multi-fidelity Bayesian optimization framework for tuning a vehicle trajectory tracking controller. In the first stage, a standard GP-based BO gathers low-fidelity (simulation) data. In the second stage, an auto-regressive AR1 GP surrogate fuses this fixed low-fidelity data set with a small number of high-fidelity (real-vehicle) evaluations, and expected-improvement BO selects new controller parameters without requiring further simulations. The approach is validated in simulation with three model-perturbation scenarios and on a test vehicle with 17 real-world queries, where the first query is reported to already achieve a lower true cost than both a manually tuned controller and the best simulation-only parameters.

Significance. The proposed workflow is practically relevant because it preserves the conventional two-stage V-model development process and avoids the continuous low-fidelity evaluations required by many multi-fidelity BO methods. The simulation study is a genuine strength: it uses three perturbation cases, five independent trials, and meaningful baselines (NARGP-BO and SFGP-BO), and the AR1GP-BO convergence advantage is clearly demonstrated in Fig. 4. The real-world experiment on a production test vehicle is also a valuable contribution. However, as detailed in the major comments, the field evidence does not currently establish the paper's headline claim about real-world sample efficiency, and the AR1 linearity assumption is not validated against the collected data.

major comments (2)

[Section V, Fig. 8, Table III] The real-world validation rests entirely on a single sequence of 17 queries on one vehicle and one track, with no repeated trials, no confidence intervals, and no comparison run using a baseline method in the field. The paper states in Section V that "the first tested parameter is determined based on only the low-fidelity data set since high-fidelity data are not available in the first run"; that first query already yields the best true cost in Table III (0.258), and the subsequent 16 queries do not improve on it. Consequently, the field experiment cannot separate the contribution of the AR1 multi-fidelity update from a simulation-only acquisition, and the abstract's claim of achieving high-quality performance with "very few real-world experiments" is not independently supported by the field data. The simulation studies in Fig. 4 support the mechanism, but the field claim needs either repeated runs, measurement-noise characterization, or a matched real-world comparison against SFGP-BO or NARGP-BO.
[Section III-B, Eq. (7); Fig. 7] The AR1 model assumes g_h(x) = rho * g_{h-1}(x) + delta_h(x), i.e., a linear relationship between the simulation cost and the real-world cost. This assumption is never tested against the collected field data. Fig. 7 shows substantial scatter between simulated and true costs, and the paper itself notes the "fidelity difference between simulations and field tests." If the sim-to-real map is strongly nonlinear, the AR1 surrogate could place queries in misleading regions. A concrete way to address this is to compare leave-one-out or cross-validated predictive performance of AR1 versus NARGP on the 17 field-test points, or to run both surrogates in repeated real-world trials.

minor comments (6)

[Section V, paragraph 1] The text says "we adopt the same implementation as described in IV-B," but the implementation details (Matérn-5/2 kernel, Gamma prior, normalization) appear in Section IV-A before the "Results" subsection; the cross-reference should be corrected.
[Table I, row 3] The entry "D×103%" is ambiguous; it should be written as "D multiplied by 1.03" or "D increased by 3%" to avoid confusion with a factor of 103 percent.
[Section IV-A and Section V] The number of low-fidelity evaluations used for the real-world experiment is not stated; specify how many simulation queries form the low-fidelity data set in Section V, because this number is important for interpreting the cost of the overall workflow.
[Table III and Section V] The manually tuned controller is described as suboptimal, but no information is given about the effort, number of iterations, or expertise behind the manual tuning baseline; reporting this context would make the comparison more meaningful.
[Fig. 7 and Fig. 8] Each point in these figures appears to be a single measurement with no error bars or repeated runs; the captions should state this explicitly to avoid over-interpreting point estimates.
[Reference [22]] Reference [22] lists "M. Temper" while the author list of the paper uses "Maximilian Templer"; the spelling should be corrected.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular reasoning: AR1GP-BO is benchmarked against independent baselines and an external test-track experiment; self-citations are not load-bearing.

full rationale

The paper's central claim is that multi-fidelity Bayesian optimization with an AR1 Gaussian process surrogate reduces the number of real-world experiments needed for vehicle controller tuning. The AR1 model in Eq. (7) is a standard Kennedy-O'Hagan construction, not a restatement of the target result. Its hyperparameters, including the scaling factor rho and the bias GP, are inferred via evidence maximization from data, not fixed to force the observed outcome. The simulation studies compare AR1GP-BO against NARGP-BO and single-fidelity SFGP-BO on perturbed-vehicle test problems, giving independent baselines; the low-fidelity data are generated from different perturbed models and are not constructed to match the high-fidelity optimum. The real-world experiment is an external benchmark: low-fidelity data come from nominal-model simulations, high-fidelity data from the actual test vehicle, and the first query is explicitly selected with no high-fidelity data, so its success cannot be a fitted high-fidelity prediction. The paper also honestly states that it cannot be confirmed whether the best observed cost is the global minimum. The only self-citations are background references or the description of the control system and trajectory planner, none of which supply a load-bearing uniqueness theorem or ansatz. The field validation being a single unreplicated run is a legitimate evidence-strength concern, but it is not circularity: no equation or fitted parameter reduces by construction to the claimed result.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method adds no new physical entities or theory. Its effectiveness rests on the standard Bayesian machinery, the linear AR1 transfer assumption, and several hand-chosen constants (cost weights, failure penalties, noise bounds) that would need sensitivity analysis to establish robustness.

free parameters (3)

Cost function weights [w1, w2, w3] = [1, 3, 0.03]
Chosen by hand in Eq. (15) to balance lateral error, heading error, and steering smoothness. They define the objective that determines the 'optimal' parameters, and no sensitivity analysis is provided.
Heuristic failure cost = 0.5 (simulation), 1.0 (experiment)
Assigned to laps that deviate more than 3 m from the reference path. This hand-chosen penalty is higher than typical successful-lap costs and affects the surrogate model and acquisition function.
Noise variance box constraint = [1e-5, 2e-4]
Hand-chosen bounds on sigma_n^2 in the real-world experiments (Section V). This constrains the GP noise estimate and can influence acquisition behavior.

assumptions (5)

standard math Gaussian process regression formulas (Eq. 6) and the AR1 recursion (Eq. 7) are valid and applicable to the cost function J(theta).
Used without proof from Rasmussen and Williams (2006) and Kennedy and O'Hagan (2000), and standard in the BO literature.
domain assumption The single-track vehicle model with magic formula tires (Eqs. 2-3) captures the trajectory tracking dynamics well enough that simulation cost correlates with real-world cost.
The paper relies on this to make low-fidelity data informative; model mismatch is acknowledged but not quantified.
ad hoc to paper The cost function (Eq. 15) with weights [1, 3, 0.03] is an adequate scalarization of controller performance.
Weights are chosen by the authors and no sensitivity analysis is provided; the 'optimal' parameters are optimal only with respect to this subjective cost.
ad hoc to paper Failed laps receive a fixed penalty cost (0.5 in simulation, 1.0 in experiments) instead of a measured value.
This heuristic injects a bias that depends on the arbitrary penalty magnitude and could distort the surrogate around unstable regions.
standard math GP hyperparameters inferred by evidence maximization generalize to unseen data.
Standard practice in BO; no guarantee is provided for the small real-world dataset of 17 points.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Efficient Learning of Vehicle Controller Parameters via Multi-Fidelity Bayesian Optimization: From Simulation to Experiment." pith.science (2026). https://pith.science/paper/OIT2ICR4

@misc{pith2026250608719,
  author       = {Pith},
  title        = {Pith review of: Efficient Learning of Vehicle Controller Parameters via Multi-Fidelity Bayesian Optimization: From Simulation to Experiment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OIT2ICR4}},
  note         = {Machine review of arXiv:2506.08719}
}

read the original abstract

Parameter tuning for vehicle controllers remains a costly and time-intensive challenge in automotive development. Traditional approaches rely on extensive real-world testing, making the process inefficient. We propose a multi-fidelity Bayesian optimization approach that efficiently learns optimal controller parameters by leveraging both low-fidelity simulation data and a very limited number of real-world experiments. Our approach significantly reduces the need for manual tuning and expensive field testing while maintaining the standard two-stage development workflow used in industry. The core contribution is the integration of an auto-regressive multi-fidelity Gaussian process model into Bayesian optimization, enabling knowledge transfer between different fidelity levels without requiring additional low-fidelity evaluations during real-world testing. We validate our approach through both simulation studies and realworld experiments. The results demonstrate that our method achieves high-quality controller performance with only very few real-world experiments, highlighting its potential as a practical and scalable solution for intelligent vehicle control tuning in industrial applications.

Figures

Figures reproduced from arXiv: 2506.08719 by the authors.

**Figure 2.** Reference path on the oval test track [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Trajectory samples of cost terms during the learning process. Darker [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Comparison of simple regret over the course of the optimization process for our proposed method AR1GP-BO (blue), NARGP-BO (orange), and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 7.** Figure 7: Simulated and true costs of the sampled controller parameters. The [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 6.** Figure 6: The test vehicle (Volkswagen Golf VII GTI) with a trajectory [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 32 canonical work pages

[1]

Impact of digital trans- formation on the automotive industry,

C. Llopis-Albert, F. Rubio, and F. Valero, “Impact of digital trans- formation on the automotive industry,”Technological Forecasting and Social Change, vol. 162, p. 120343, 2021

work page 2021
[2]

Intelligent Vehicle Sys- tems: Applications and New Trends,

O. Gusikhin, D. Filev, and N. Rychtyckyj, “Intelligent Vehicle Sys- tems: Applications and New Trends,” inInformatics in Control Automation and Robotics: Selected Papers from the International Conference on Informatics in Control Automation and Robotics 2006, 2008, pp. 3–14

work page 2006
[3]

The new V-Model of VDI 2206 and its validation,

I. Graessler and J. Hentze, “The new V-Model of VDI 2206 and its validation,”at - Automatisierungstechnik, vol. 68, no. 5, pp. 312–324, 2020

work page 2020
[4]

Effects of Model Complexity on the Performance of Automated Vehicle Steering Controllers: Model Development, Validation and Comparison,

D. E. Smith and J. M. Starkey, “Effects of Model Complexity on the Performance of Automated Vehicle Steering Controllers: Model Development, Validation and Comparison,”Vehicle System Dynamics, vol. 24, no. 2, pp. 163–181, 1995

work page 1995
[5]

Advanced driver assistance systems: Objective and subjective perfor- mance evaluation,

R. A. Auckland, W. J. Manning, O. M. Carsten, and A. H. Jamson, “Advanced driver assistance systems: Objective and subjective perfor- mance evaluation,”Vehicle System Dynamics, vol. 46, pp. 883–897, 2008

work page 2008
[6]

Optimum Settings for Automatic Controllers,

J. G. Ziegler and N. B. Nichols, “Optimum Settings for Automatic Controllers,”Transactions of the American Society of Mechanical Engineers, vol. 64, no. 8, pp. 759–765, 1942

work page 1942
[7]

Iterative feedback tuning: theory and applications,

H. Hjalmarsson, M. Gevers, S. Gunnarsson, and O. Lequin, “Iterative feedback tuning: theory and applications,”IEEE Control Systems Magazine, vol. 18, no. 4, pp. 26–41, 1998

1998
[8]

PID tuning using extremum seeking: online, model-free performance optimization,

N. Killingsworth and M. Krstic, “PID tuning using extremum seeking: online, model-free performance optimization,”IEEE Control Systems Magazine, vol. 26, no. 1, pp. 70–79, 2006

work page 2006

Show all 34 references

[9]

Genetic tuning of digital PID controllers,

B. Porter and A. H. Jones, “Genetic tuning of digital PID controllers,” Electronics Letters, vol. 28, no. 9, pp. 843–844, 1992

1992
[10]

A particle swarm optimization approach for optimum design of PID controller in A VR system,

Z.-L. Gaing, “A particle swarm optimization approach for optimum design of PID controller in A VR system,”IEEE Transactions on Energy Conversion, vol. 19, no. 2, pp. 384–391, 2004

2004
[11]

Data-Efficient Autotuning With Bayesian Optimization: An Industrial Control Study,

M. Neumann-Brosig, A. Marco, D. Schwarzmann, and S. Trimpe, “Data-Efficient Autotuning With Bayesian Optimization: An Industrial Control Study,”IEEE Transactions on Control Systems Technology, vol. 28, no. 3, pp. 730–740, 2020

2020
[12]

Learning Energy-Efficient Trajectory Planning for Robotic Manipulators Using Bayesian Optimization,

P. Holzmann, M. Pfefferkorn, J. Peters, and R. Findeisen, “Learning Energy-Efficient Trajectory Planning for Robotic Manipulators Using Bayesian Optimization,” inEuropean Control Conference, 2024, pp. 1374–1379

2024
[13]

Stability- informed Bayesian Optimization for MPC Cost Function Learning,

S. Hirt, M. Pfefferkorn, A. Mesbah, and R. Findeisen, “Stability- informed Bayesian Optimization for MPC Cost Function Learning,” in Conference on Nonlinear Model Predictive Control, 2024, to appear

2024
[14]

Time-Series-Informed Closed-loop Learning for Sequential Decision Making and Control,

S. Hirt, L. Theiner, and R. Findeisen, “Time-Series-Informed Closed-loop Learning for Sequential Decision Making and Control,” arXiv:2412.02423, 2024

2024
[15]

Safe Learning-Based Optimization of Model Predictive Control: Application to Battery Fast-Charging,

S. Hirt, A. H ¨ohl, J. Pohlodek, J. Schaeffer, M. Pfefferkorn, R. D. Braatz, and R. Findeisen, “Safe Learning-Based Optimization of Model Predictive Control: Application to Battery Fast-Charging,” in American Control Conference, 2025

2025
[16]

Automatic Gait Optimization with Gaussian Process Regression,

D. J. Lizotte, T. Wang, M. H. Bowling, D. Schuurmanset al., “Automatic Gait Optimization with Gaussian Process Regression,” in International Joint Conference on Artificial Intelligence, vol. 7, 2007, pp. 944–949

2007
[17]

Gaussian Process Bandit Optimisation with Multi-Fidelity Evalua- tions,

K. Kandasamy, G. Dasarathy, J. B. Oliva, J. Schneider, and B. P ´oczos, “Gaussian Process Bandit Optimisation with Multi-Fidelity Evalua- tions,” inAdvances in Neural Information Processing Systems, vol. 29, 2016

2016
[18]

Virtual vs. real: Trading off simulations and physical experiments in reinforcement learning with Bayesian optimization,

A. Marco, F. Berkenkamp, P. Hennig, A. P. Schoellig, A. Krause, S. Schaal, and S. Trimpe, “Virtual vs. real: Trading off simulations and physical experiments in reinforcement learning with Bayesian optimization,” inIEEE International Conference on Robotics and Automation (ICRA...

2017
[19]

Recent advances in Bayesian optimization,

X. Wang, Y . Jin, S. Schmitt, and M. Olhofer, “Recent advances in Bayesian optimization,”ACM Computing Surveys, vol. 55, no. 13s, pp. 1–36, 2023

2023
[20]

Pacejka,Tire and Vehicle Dynamics, 2nd ed

H. Pacejka,Tire and Vehicle Dynamics, 2nd ed. Elsevier, 2005

2005
[21]

Zeitoptimale Trajektorienplanung f ¨ur automatisiertes Fahren bis in den fahrdynamischen Grenzbereich,

I. Gundlach, “Zeitoptimale Trajektorienplanung f ¨ur automatisiertes Fahren bis in den fahrdynamischen Grenzbereich,” Ph.D. dissertation, Technische Universit¨at Darmstadt, 2020

2020
[22]

Learning Delta Policies For Automated Driving Via Reinforcement Learning,

M. Temper, J. Kaste, P. Hochrein, and B. Mennenga, “Learning Delta Policies For Automated Driving Via Reinforcement Learning,” in Proceedings of the 30th Aachen Colloquium Sustainable Mobility, 2017

2017
[23]

C. E. Rasmussen and C. K. I. Williams,Gaussian Processes for Machine Learning, 2nd ed. The MIT press, 2006

2006
[24]

Predicting the Output from a Complex Computer Code When Fast Approximations Are Available,

M. C. Kennedy and A. O’Hagan, “Predicting the Output from a Complex Computer Code When Fast Approximations Are Available,” Biometrika, vol. 87, no. 1, pp. 1–13, 2000

2000
[25]

Recursive co-kriging model for Design of Computer experiments with multiple levels of fidelity,

L. Le Gratiet and J. Garnier, “Recursive co-kriging model for Design of Computer experiments with multiple levels of fidelity,”Interna- tional Journal for Uncertainty Quantification, vol. 4, no. 5, pp. 365– 386, 2014

2014
[26]

Nonlinear information fusion algorithms for data-efficient multi-fidelity modelling,

P. Perdikaris, M. Raissi, A. Damianou, N. D. Lawrence, and G. E. Kar- niadakis, “Nonlinear information fusion algorithms for data-efficient multi-fidelity modelling,”Proceedings of the Royal Society A: Math- ematical, Physical and Engineering Sciences, vol. 473, no. 2198, p. 2...

2017
[27]

Garnett,Bayesian Optimization

R. Garnett,Bayesian Optimization. Cambridge University Press, 2023

2023
[28]

Gaussian Pro- cess Regression: Active Data Selection and Test Point Rejection,

S. Sambu, M. Wallat, T. Graepel, and K. Obermayer, “Gaussian Pro- cess Regression: Active Data Selection and Test Point Rejection,” in Proceedings of the IEEE-INNS-ENNS International Joint Conference on Neural Networks, vol. 3, 2000, pp. 241–246

2000
[29]

BoTorch: A Framework for Efficient Monte- Carlo Bayesian Optimization,

M. Balandat, B. Karrer, D. R. Jiang, S. Daulton, B. Letham, A. G. Wilson, and E. Bakshy, “BoTorch: A Framework for Efficient Monte- Carlo Bayesian Optimization,” inAdvances in Neural Information Processing Systems 33, vol. 33, 2020, pp. 21 524–21 538

2020
[30]

Practical Bayesian Op- timization of Machine Learning Algorithms,

J. Snoek, H. Larochelle, and R. P. Adams, “Practical Bayesian Op- timization of Machine Learning Algorithms,” inAdvances in Neural Information Processing Systems, vol. 25, 2012

2012
[31]

Emukit: A Python toolkit for decision making under uncertainty,

A. Paleyes, M. Mahsereci, and N. D. Lawrence, “Emukit: A Python toolkit for decision making under uncertainty,” inPython in Science Conference, 2023, pp. 68–75

2023
[32]

Safe learning of regions of attraction for uncertain, nonlinear systems with Gaussian processes,

F. Berkenkamp, R. Moriconi, A. P. Schoellig, and A. Krause, “Safe learning of regions of attraction for uncertain, nonlinear systems with Gaussian processes,” inIEEE 55th Conference on Decision and Control (CDC), 2016, pp. 4661–4666

2016
[33]

Safe Bayesian Optimization Using Interior-Point Methods—Applied to Personalized Insulin Dose Guidance,

D. Krishnamoorthy and F. J. Doyle, “Safe Bayesian Optimization Using Interior-Point Methods—Applied to Personalized Insulin Dose Guidance,”IEEE Control Systems Letters, vol. 6, pp. 2834–2839, 2022

2022
[34]

Safe and Stable Closed- Loop Learning for Neural-Network-Supported Model Predictive Con- trol,

S. Hirt, M. Pfefferkorn, and R. Findeisen, “Safe and Stable Closed- Loop Learning for Neural-Network-Supported Model Predictive Con- trol,” inConference on Decision and Control, 2024

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Impact of digital trans- formation on the automotive industry,

C. Llopis-Albert, F. Rubio, and F. Valero, “Impact of digital trans- formation on the automotive industry,”Technological Forecasting and Social Change, vol. 162, p. 120343, 2021

work page 2021

[2] [2]

Intelligent Vehicle Sys- tems: Applications and New Trends,

O. Gusikhin, D. Filev, and N. Rychtyckyj, “Intelligent Vehicle Sys- tems: Applications and New Trends,” inInformatics in Control Automation and Robotics: Selected Papers from the International Conference on Informatics in Control Automation and Robotics 2006, 2008, pp. 3–14

work page 2006

[3] [3]

The new V-Model of VDI 2206 and its validation,

I. Graessler and J. Hentze, “The new V-Model of VDI 2206 and its validation,”at - Automatisierungstechnik, vol. 68, no. 5, pp. 312–324, 2020

work page 2020

[4] [4]

Effects of Model Complexity on the Performance of Automated Vehicle Steering Controllers: Model Development, Validation and Comparison,

D. E. Smith and J. M. Starkey, “Effects of Model Complexity on the Performance of Automated Vehicle Steering Controllers: Model Development, Validation and Comparison,”Vehicle System Dynamics, vol. 24, no. 2, pp. 163–181, 1995

work page 1995

[5] [5]

Advanced driver assistance systems: Objective and subjective perfor- mance evaluation,

R. A. Auckland, W. J. Manning, O. M. Carsten, and A. H. Jamson, “Advanced driver assistance systems: Objective and subjective perfor- mance evaluation,”Vehicle System Dynamics, vol. 46, pp. 883–897, 2008

work page 2008

[6] [6]

Optimum Settings for Automatic Controllers,

J. G. Ziegler and N. B. Nichols, “Optimum Settings for Automatic Controllers,”Transactions of the American Society of Mechanical Engineers, vol. 64, no. 8, pp. 759–765, 1942

work page 1942

[7] [7]

Iterative feedback tuning: theory and applications,

H. Hjalmarsson, M. Gevers, S. Gunnarsson, and O. Lequin, “Iterative feedback tuning: theory and applications,”IEEE Control Systems Magazine, vol. 18, no. 4, pp. 26–41, 1998

1998

[8] [8]

PID tuning using extremum seeking: online, model-free performance optimization,

N. Killingsworth and M. Krstic, “PID tuning using extremum seeking: online, model-free performance optimization,”IEEE Control Systems Magazine, vol. 26, no. 1, pp. 70–79, 2006

work page 2006