Multi-Fidelity Stochastic Trust Region Method with Adaptive Sampling

Juliane Mueller; Yunsoo Ha

REVIEW 3 major objections 5 minor 17 references

Multi-Fidelity Stochastic Trust Region Method with Adaptive Sampling

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A trust-region method decides when cheap simulations are safe to trust.

desk verdict Useful multi-fidelity extension with a real, possibly manageable bias in the adaptive sampling rule; deserves review. read the letter →

arxiv 2508.03901 v1 pith:ERASI7J6 submitted 2025-08-05 math.OC

Yunsoo Ha , Juliane Mueller This is my paper

classification math.OC MSC 90C5690C15

keywords multi-fidelitysimulationoptimizationadaptivesamplingtrustregionmethodMonteCarloderivative-freestochasticcorrelation-basedmodelselection

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

This paper tries to show that a stochastic trust-region optimizer can safely exploit several levels of simulation fidelity by deciding at every step where to look, which fidelity to query, and how many replications to run. The proposed method, ASTRO-MFDF, uses a multi-fidelity Monte Carlo estimator as its workhorse and only leans on cheaper lower-fidelity models when a running correlation score says they agree with the expensive model. In numerical experiments on stochastic Rosenbrock functions and continuous (s,S) inventory problems, the method reports faster convergence than the high-fidelity-only ASTRO-DF and than Nelder-Mead on most test settings, and better final solutions in several cases. The reason it matters is that simulation optimization is expensive, and the usual promise of multi-fidelity methods, cheap speedup, is undone when biased low-fidelity models drag the search away from the true optimum.

What carries the argument

The engine is the multi-fidelity adaptive sampling subroutine (MF AS) wrapped in a trust-region loop. Its core identity is the multi-fidelity Monte Carlo estimator of Eq. (2), an unbiased control-variate combination of all fidelity levels whose variance is given in Eq. (3); the adaptive sampler replaces the unknown variances and covariances with sequential estimates, solves problem (5) to minimize total query cost subject to the stochastic-error and optimality-gap balance from Eq. (4), and chooses MFMC over MC only when the predicted cost is lower. Around this, a correlation vector $\alpha_k$ is updated from the success or failure of each lower-fidelity model's candidate points, so low-fidelity models are built and trusted only when they have recently earned it.

What would settle it

Give ASTRO-MFDF a synthetic problem whose true variances and covariances are known, and compare the sample sizes and MC/MFMC choices from subproblem (5) with the cost-minimizing choices computed using the true values; if the method's choices deviate sharply while the algorithm still appears to converge, the sequential estimates are carrying the argument. Alternatively, a problem where low fidelity is strongly correlated in one region and anticorrelated in another should make $\alpha_k$ rise and then fall; if it does not, the selective-trust mechanism is not doing what the paper claims.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that the 'where, which fidelity, and how many times' question of multi-fidelity simulation optimization can be answered online by a trust-region algorithm with adaptive sampling. ASTRO-MFDF maintains separate trust regions for different fidelity levels and a correlation vector $\alpha_k$; a lower-fidelity model is used to propose the next iterate only when its recent track record indicates sufficient agreement with the high-fidelity function. At each evaluation, the multi-fidelity adaptive sampling subroutine estimates variances and covariances sequentially, solves a small cost-minimization problem to choose sample sizes and the coefficient vector $c$ for the multi-fidelity Monte Carlo estimator, and selects MFMC only when its predicted cost for meeting the accuracy target beats plain Monte Carlo. The paper reports that this selective use of low fidelity, combined with a larger preserved trust region for the high-fidelity model, yields more completed iterations and faster convergence under a fixed oracle budget in the tested stochastic Rosenbrock and inventory problems.

Load-bearing premise

Everything rests on the assumption that the variances and covariances estimated sequentially, together with the correlation score $\alpha_k$, are accurate enough that the method's choices of sample size, fidelity, and MC-versus-MFMC never systematically mislead the search.

Editorial extensions

If this is right

If the central claim is right, practitioners can run stochastic optimization with a simulator at multiple resolutions without pre-committing to which fidelity to trust, because the method re-decides at every iteration.
The sample-size rule makes the number of replications grow as the trust region shrinks, so early iterations are cheap and later iterations are precise, giving users a budget-aware accuracy schedule.
Because low-fidelity models are used only when the correlation score is high, problems where the cheap model is strongly biased in some regions automatically fall back to high-fidelity evaluation instead of being misled.
The cost-selection step between MC and MFMC gives a reusable criterion: switch to the variance-reduced estimator only when its predicted total cost beats standard Monte Carlo for the same accuracy.
In the reported experiments, more iterations fit under a fixed budget, which directly matters for applications where the high-fidelity simulator is the bottleneck.

Reading between the lines

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

I would expect the same selective-trust idea to transfer to surrogate-based optimization beyond trust regions: any method that builds a model from cheap evaluations could gate those evaluations on a running correlation score.
A natural stress test not reported in the paper is a problem where the low-fidelity model is correlated with the high-fidelity model only inside a small region; the $\alpha_k$ updates should rise and fall as the iterates cross that region, and failure to do so would expose a weakness in the heuristic.
The paper's cost-ratio treatment suggests a practical budgeting rule for steady-state simulation: run the shortest cheap models during exploration and shift to longer runs only as the trust region converges, since the optimal sample size grows like $\Delta^{-4}$.
If the sequential covariance estimates are the weakest link, a direct comparison against the same algorithm with oracle-known variances on a small synthetic problem would isolate how much of the benefit comes from the adaptivity versus from the MFMC estimator itself.

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

The paper extends the authors' ASTRO-BFDF to q fidelities, replacing BFMC with MFMC and adding a subproblem that picks sample sizes and chooses between MC and MFMC. That is a genuine extension, and the correlation-vector mechanism for deciding when low fidelity is worth using is sensible. The experiments on Rosenbrock and (s,S) inventory are standard but reasonably convincing: ASTRO-MFDF completes more iterations under budget and often converges faster than ASTRO-DF and Nelder-Mead. The SimOpt-based setup is reproducible, which is real credit.

The soft spots are mostly about what the paper does not address. First, the stress-test concern about data-dependent c* is correct. In MF AS, the variance and covariance estimates used in (5) come from the same sample trajectories that then produce the MFMC estimator. Equations (2) and (3) hold for fixed c. When c* is chosen from the data, the control variate term generally has nonzero expectation under CRN, so the estimator is biased and (3) is not its true variance. The stopping rule (4) relies on dVar, so if dVar is systematically low, the stochastic-error/optimality-gap balance is not enforced. The paper should acknowledge this and either give a heuristic argument or test empirically how much the actual variance exceeds dVar.

Second, there is no comparison to the authors' own ASTRO-BFDF, which is the natural baseline; without it, we cannot tell what the multi-fidelity extension adds beyond the bi-fidelity case. Third, the paper is an algorithmic proposal with no convergence theory. That is not fatal—ASTRO-DF had theory, so the gap is visible—but it should be stated plainly. Fourth, there are notational slips in Algorithm 2, and the hyperparameters and code are not included.

I'd send this to peer review. The core idea is coherent, the experiments are on external benchmarks, and the bias issue is worth a serious referee's time rather than a desk reject. For revision, I'd ask for the ASTRO-BFDF comparison, a discussion of the c* bias (ideally a numerical check with fixed vs data-dependent c), and cleaned-up notation. Whether the method works as advertised is an empirical question, and the paper gives enough evidence to make that question worth answering.

Referee Report

3 major / 5 minor

Summary. The paper proposes ASTRO-MFDF, an adaptive sampling trust-region method for multi-fidelity simulation optimization. The algorithm extends the ASTRO-DF/ASTRO-BFDF framework to multiple fidelities by using multi-fidelity Monte Carlo (MFMC) estimators, with sample sizes and MFMC coefficients chosen adaptively through a cost-minimization problem subject to a variance constraint. A dynamically updated correlation vector is used to decide when lower-fidelity models are trusted versus when the high-fidelity model is used. The method is validated on stochastic Rosenbrock functions and continuous (s,S) inventory problems using the SimOpt library, reporting faster convergence than ASTRO-DF and Nelder-Mead on most test instances.

Significance. If the algorithmic proposal is sound, ASTRO-MFDF would provide a practical and flexible extension of stochastic trust-region methods to multi-fidelity settings, directly addressing the risk of low-fidelity bias through adaptive correlation tracking and MFMC-based variance reduction. The paper is clearly written, includes complete pseudocode, and uses a standard benchmark environment with reproducible budgets and cost ratios. The numerical experiments are carefully illustrated. However, the central adaptive sampling mechanism relies on a variance formula that is valid only for fixed MFMC coefficients, and the manuscript offers no convergence or complexity analysis; these gaps materially limit the strength of the claims.

major comments (3)

[Section 2, Eq. (3) and Problem (5)] The variance expression in Eq. (3) is derived under the assumption that the coefficients c are fixed constants. In the proposed algorithm, however, the coefficients c* are obtained by solving Problem (5) using variance and covariance estimates computed from the very same sample paths that subsequently form the MFMC estimator in Eq. (2). When c is data-dependent, the identity E[c_i(avg_{n_i} - avg_{n_{i-1}})] = 0 no longer holds, so the estimator is biased and Eq. (3) is not its true variance. Consequently, the stopping condition dVar(\hat{F}^0) <= kappa^2 Delta_k^4 / lambda_k may not enforce the intended balance between stochastic error and optimality gap, because the predicted variance can systematically underestimate the actual variance. The authors should modify the procedure to estimate c on an independent pilot sample, or provide a rigorous argument showing that the bias is negligible and that the variance formula remains a valid upper bound with high probability.
[Algorithm 2, Step 7] Step 7 of Algorithm 2 contains inconsistent indices: the line reads "approximately compute X^{s,t}_k = argmin_{||X-X_k|| <= Delta^l_k} M^l_k(X)" using l instead of t, while Step 9 uses Delta^0_k in the denominator of the success ratio even though the model being evaluated is M^t_k. This ambiguity prevents the reader from determining whether the candidate is restricted to the trust region Delta^t_k associated with the t-fidelity model or to the high-fidelity region Delta^0_k. The pseudocode should be corrected so that all subscripts and superscripts correspond to the same fidelity index t throughout the algorithm.
[General (absence of convergence theory)] The paper does not provide any convergence, complexity, or local-convergence analysis for ASTRO-MFDF. Since the method is presented as an extension of ASTRO-DF, which has established convergence guarantees, the authors should at least analyze whether the adaptive sampling rule preserves the stochastic-error/optimality-gap balance required for trust-region convergence when the MFMC coefficients are data-dependent, or explicitly state the heuristic nature of the scheme and the assumptions under which the empirical behavior is expected to hold. Without this, the claim that the method "efficiently solves" the multi-fidelity simulation optimization problem rests entirely on two numerical examples.

minor comments (5)

[Algorithm 1, Step 12] The text refers to "BF AS" (bi-fidelity adaptive sampling) in several places, but the method described is multi-fidelity; this appears to be a typo for "MF AS" and should be corrected throughout.
[Section 2, Eq. (3)] The variance formula contains typographical inconsistencies in the superscripts (for example, "si" appears instead of the intended \sigma^i), which should be fixed to avoid confusion.
[Section 2, Problem (5)] The rounding of the continuous solution n* to integer sample sizes may violate the monotonicity constraint n_i <= n_{i+1} if only some components are rounded up; a rounding procedure that preserves the ordering should be specified.
[Section 3, correlation update] The threshold alpha_th and the update constants gamma_1 and gamma_2 are introduced without any guidance on their choice or a sensitivity analysis; the authors should discuss how these parameters affect the behavior of the correlation-vector heuristic.
[Section 4, Figure 10] The performance metric "fraction of problems each method successfully solved under 10% optimality gap threshold" is not precisely defined; the authors should clarify what constitutes a successful solve (e.g., within the given budget and with the final incumbents evaluated by the 200-replication simulation).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ASTRO-MFDF is a self-contained algorithmic proposal benchmarked on external SimOpt problems; self-citations are not load-bearing.

full rationale

The paper's central contribution is an algorithmic framework rather than a derived theorem: it combines the known MFMC estimator (Eq. 2) and variance formula (Eq. 3), stated for fixed sample sizes and coefficients, with an adaptive sampling subproblem (Eq. 5) and a trust-region wrapper. The variance formula is used as a design target, not as a fitted prediction of the final adaptive estimator. The claimed experimental gains are measured against external benchmarks in SimOpt, with no parameters fitted to reproduce those benchmarks, so the numerical conclusions are not forced by construction. The adaptive choice of c* from sequential variance estimates could, as a statistical matter, introduce bias and make Eq. (3) inaccurate, but that is a correctness risk rather than a definitional circularity: the paper does not define 'sufficiently accurate' to be identical to the quantity being optimized, and no equation is equivalent to its input by construction. The self-citations to the authors' prior ASTRO-BFDF and ASTRO-DF work are used as background and algorithmic lineage; the present MF AS procedure and correlation-threshold mechanism are described with their own equations and do not reduce to those prior results. Because the derivation chain is self-contained for the algorithmic proposal and no load-bearing claim reduces to a fit or to a self-citation, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No fitted parameters or invented physical entities appear. The method relies on several hand-chosen hyperparameters and heuristic assumptions about the reliability of sequential variance estimates and the trust-region framework.

free parameters (4)

kappa (adaptive sampling constant) = not reported
Appears in sample-size rule (4) and in problem (5); the paper does not specify the value used in experiments.
alpha_th (correlation threshold) = not reported
Determines when low-fidelity models are built independently; no value or sensitivity analysis is given.
zeta, eta, gamma1, gamma2, sigma_lb = not reported
Trust-region and model constants in Algorithm 1; values are not specified, so results may depend on hidden choices.
lambda_k sequence = O(log k)
The sample-size lower bound is only described asymptotically; the concrete initialization is not stated.

assumptions (4)

domain assumption Standard trust-region model accuracy: local models built on 2d+1 points are sufficiently accurate for candidate generation.
Section 3, Algorithm 1; relies on derivative-free trust-region practice from Conn, Scheinberg, and Vicente (ref [8]) without verification in this setting.
domain assumption MFMC variance formula (3) is exact and the covariance estimates are reliable with small samples.
Section 2, Eq. (3); the algorithm starts with sample sizes as low as 2 (en_i >= 2) and uses these estimates to solve problem (5).
domain assumption Common Random Numbers induce positive correlation between fidelity levels, making MFMC variance reduction beneficial.
Section 1, paragraph after Eq. (3); the paper relies on CRN but does not discuss how it affects dependence across design points.
ad hoc to paper The correlation vector alpha_k is a valid online indicator of where low fidelity is useful.
Section 3, Algorithm 1; this is a heuristic update rule with no theoretical justification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-Fidelity Stochastic Trust Region Method with Adaptive Sampling." pith.science (2026). https://pith.science/paper/ERASI7J6

@misc{pith2026250803901,
  author       = {Pith},
  title        = {Pith review of: Multi-Fidelity Stochastic Trust Region Method with Adaptive Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ERASI7J6}},
  note         = {Machine review of arXiv:2508.03901}
}

read the original abstract

Simulation optimization is often hindered by the high cost of running simulations. Multi-fidelity methods offer a promising solution by incorporating cheaper, lower-fidelity simulations to reduce computational time. However, the bias in low-fidelity models can mislead the search, potentially steering solutions away from the high-fidelity optimum. To overcome this, we propose ASTRO-MFDF, an adaptive sampling trust-region method for multi-fidelity simulation optimization. ASTRO-MFDF features two key strategies: (i) it adaptively determines the sample size and selects appropriate sampling strategies to reduce computational cost; and (ii) it selectively uses low-fidelity information only when a high correlation with the high-fidelity is anticipated, reducing the risk of bias. We validate the performance and computational efficiency of ASTRO-MFDF through numerical experiments using the SimOpt library.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 16 canonical work pages

[1]

Two-stage estimation and variance modeling for latency- constrained variational quantum algorithms,

Y. Ha, S. Shashaani, and M. Menickelly, “Two-stage estimation and variance modeling for latency- constrained variational quantum algorithms,” INFORMS Journal on Computing, vol. 37, no. 1, pp. 125– 145, 2025

work page 2025
[2]

Stochastic simulation- optimization framework for the design and assessment of renewable energy systems under uncertainty,

G. K. Sakki, I. Tsoukalas, P. Kossieris, C. Makropoulos, and A. Efstratiadis, “Stochastic simulation- optimization framework for the design and assessment of renewable energy systems under uncertainty,” Renewable and Sustainable Energy Reviews , vol. 168, p. 112886, 2022

work page 2022
[3]

Improved multi-fidelity simulation-based optimisa- tion: application in a digital twin shop floor,

Z. Zhang, Z. Guan, Y. Gong, D. Luo, and L. Yue, “Improved multi-fidelity simulation-based optimisa- tion: application in a digital twin shop floor,” International Journal of Production Research , vol. 60, no. 3, pp. 1016–1035, 2022

work page 2022
[4]

Stochastic co-kriging for steady-state simulation metamodeling,

X. Chen, S. Hemmati, and F. Yang, “Stochastic co-kriging for steady-state simulation metamodeling,” in 2017 Winter Simulation Conference (WSC) (W. K. V. Chan, A. D’Ambrogio, G. Zacharewicz, N. Mustafee, G. Wainer, and E. Page, eds.), pp. 1750–1761, IEEE, 2017

work page 2017
[5]

Multi-fidelity bayesian optimization: A review,

B. Do and R. Zhang, “Multi-fidelity bayesian optimization: A review,” arXiv preprint arXiv:2311.13050, 2024

arXiv 2024
[6]

On the Effects of Heterogeneous Errors on Multi-fidelity Bayesian Optimization

Z. Z. Foumani, A. Yousefpour, M. Shishehbor, and R. Bostanabad, “On the effects of heterogeneous errors on multi-fidelity bayesian optimization,” arXiv preprint arXiv:2309.02771 , 2023

work page Pith review arXiv 2023
[7]

ASTRO-DF: A class of adaptive sampling trust-region algorithms for derivative-free stochastic optimization,

S. Shashaani, F. S Hashemi, and R. Pasupathy, “ASTRO-DF: A class of adaptive sampling trust-region algorithms for derivative-free stochastic optimization,” SIAM Journal on Optimization , vol. 28, no. 4, pp. 3145–3176, 2018

work page 2018
[8]

A. R. Conn, K. Scheinberg, and L. N. Vicente, Introduction to derivative-free optimization. Society for Industrial and Applied Mathematics, 1st ed., 2009

work page 2009

Show all 17 references

[9]

Adaptive sampling-based bi-fidelity stochastic trust region method for derivative- free stochastic optimization,

Y. Ha and J. Mueller, “Adaptive sampling-based bi-fidelity stochastic trust region method for derivative- free stochastic optimization,” arXiv preprint arXiv:2408.04625 , 2024

2024 arXiv
[10]

Optimal model management for multifidelity monte carlo estimation,

B. Peherstorfer, K. Willcox, and M. Gunzburger, “Optimal model management for multifidelity monte carlo estimation,” SIAM Journal on Scientific Computing , vol. 38, no. 5, pp. A3163–A3194, 2016

2016
[11]

Complexity of zeroth-and first-order stochastic trust-region algorithms,

Y. Ha, S. Shashaani, and R. Pasupathy, “Complexity of zeroth-and first-order stochastic trust-region algorithms,” arXiv preprint arXiv:2405.20116 , 2024

2024 arXiv
[12]

Stochastic trust-region response-surface method (strong)—a new response-surface framework for simulation optimization,

K.-H. Chang, L. J. Hong, and H. Wan, “Stochastic trust-region response-surface method (strong)—a new response-surface framework for simulation optimization,” INFORMS Journal on Computing , vol. 25, no. 2, pp. 230–243, 2013

2013
[13]

Stochastic optimization using a trust-region method and random models,

R. Chen, M. Menickelly, and K. Scheinberg, “Stochastic optimization using a trust-region method and random models,” Mathematical Programming, vol. 169, no. 2, pp. 447–487, 2018

2018
[14]

Nocedal and S

J. Nocedal and S. J. Wright, Numerical Optimization. Springer New York, 2nd ed., 2006

2006
[15]

Simopt: A testbed for simulation-optimization experiments,

D. J. Eckman, S. G. Henderson, and S. Shashaani, “Simopt: A testbed for simulation-optimization experiments,” INFORMS Journal on Computing , vol. 35, no. 2, pp. 495–508, 2023

2023
[16]

Analytical benchmark problems for multifidelity optimization meth- ods,

L. Mainini, A. Serani, M. P. Rumpfkeil, E. Minisci, D. Quagliarella, H. Pehlivan, S. Yildiz, S. Ficini, R. Pellegrini, F. Di Fiore, et al. , “Analytical benchmark problems for multifidelity optimization meth- ods,” arXiv preprint arXiv:2204.07867 , 2022

2022 arXiv
[17]

D. J. Eckman, S. G. Henderson, S. Shashaani, and R. Pasupathy, “SimOpt.” https://github.com/ simopt-admin/simopt, 2023. 13

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Two-stage estimation and variance modeling for latency- constrained variational quantum algorithms,

Y. Ha, S. Shashaani, and M. Menickelly, “Two-stage estimation and variance modeling for latency- constrained variational quantum algorithms,” INFORMS Journal on Computing, vol. 37, no. 1, pp. 125– 145, 2025

work page 2025

[2] [2]

Stochastic simulation- optimization framework for the design and assessment of renewable energy systems under uncertainty,

G. K. Sakki, I. Tsoukalas, P. Kossieris, C. Makropoulos, and A. Efstratiadis, “Stochastic simulation- optimization framework for the design and assessment of renewable energy systems under uncertainty,” Renewable and Sustainable Energy Reviews , vol. 168, p. 112886, 2022

work page 2022

[3] [3]

Improved multi-fidelity simulation-based optimisa- tion: application in a digital twin shop floor,

Z. Zhang, Z. Guan, Y. Gong, D. Luo, and L. Yue, “Improved multi-fidelity simulation-based optimisa- tion: application in a digital twin shop floor,” International Journal of Production Research , vol. 60, no. 3, pp. 1016–1035, 2022

work page 2022

[4] [4]

Stochastic co-kriging for steady-state simulation metamodeling,

X. Chen, S. Hemmati, and F. Yang, “Stochastic co-kriging for steady-state simulation metamodeling,” in 2017 Winter Simulation Conference (WSC) (W. K. V. Chan, A. D’Ambrogio, G. Zacharewicz, N. Mustafee, G. Wainer, and E. Page, eds.), pp. 1750–1761, IEEE, 2017

work page 2017

[5] [5]

Multi-fidelity bayesian optimization: A review,

B. Do and R. Zhang, “Multi-fidelity bayesian optimization: A review,” arXiv preprint arXiv:2311.13050, 2024

arXiv 2024

[6] [6]

On the Effects of Heterogeneous Errors on Multi-fidelity Bayesian Optimization

Z. Z. Foumani, A. Yousefpour, M. Shishehbor, and R. Bostanabad, “On the effects of heterogeneous errors on multi-fidelity bayesian optimization,” arXiv preprint arXiv:2309.02771 , 2023

work page Pith review arXiv 2023

[7] [7]

ASTRO-DF: A class of adaptive sampling trust-region algorithms for derivative-free stochastic optimization,

S. Shashaani, F. S Hashemi, and R. Pasupathy, “ASTRO-DF: A class of adaptive sampling trust-region algorithms for derivative-free stochastic optimization,” SIAM Journal on Optimization , vol. 28, no. 4, pp. 3145–3176, 2018

work page 2018

[8] [8]

A. R. Conn, K. Scheinberg, and L. N. Vicente, Introduction to derivative-free optimization. Society for Industrial and Applied Mathematics, 1st ed., 2009

work page 2009