acceptodds
Under review as a conference paper at ICLR 2027

Bayesian optimization on amortized search space

Abstract

In real-world black-box optimization problems, a differentiable mechanistic simulator is often available that approximately predicts the behavior of the system as a function of the parameters being optimized. Here, we show how an amortized inverse map of such a simulator can be used in Bayesian optimization to restrict the search space while using a Gaussian process kernel that is aware of the original problem geometry. This allows for efficiently leveraging the low intrinsic dimensionality of the optimization problem. We derive regret bounds that highlight the benefits and limitations of this method and illustrate it with an application to drug dosing personalization in Parkinson's disease. We use, as a simulator, a pharmacokinetic/pharmacodynamic (PK/PD) model of a drug, a mechanistic model calibrated on population data with quantified parameter uncertainty. A dosing neural network is trained offline against the model, amortizing the map from patient PK/PD parameters to the regimen predicted to be optimal, and online Bayesian optimization then searches over this learned, low-dimensional representation rather than over raw dosages. Numerical simulations demonstrate the improved performance of our method in this application, as well as under simulator misspecification.

Primary area:
optimization
Keywords:
Bayesian optimization, amortization, simulator, drug dosing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_a9M2OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an approach towards improving sample efficiency of Bayesian optimization (BO) by restricting the search space to inputs that are plausibly optimal given an approximate simulator. Concretely, the search space is implicitly defined through optimal solutions of an inverse problem associated with a simulation under a particular set of task-specific parameters. To avoid repeated solution of the inverse problem, a neural net is trained to predict the solutions in its forward pass, optimized in expectation w.r.t. a prior distribution over simulation parameters. Theoretically, regret bounds are provided that depend only on the dimensionality of the resulting amortized search space, rather than the ambient dimension, along with error terms from amortization and misspecification. Results are provided on synthetic test functions, and a case study in personalized pharmacokinetics/pharmacodynamics, under various feedback modalities.

Critical strengths
  • Speeding up BO in high-dimensional domains is a relevant problem, and exploiting prior information encoded in simulators is an interesting approach. I find the considered setting quite relevant, and the PK/PD case study interesting as well.
Critical weaknesses
  • The analysis is rather simplified; in particular it assumes away the statistical problem of amortization (i.e., that the amortized network is trained only in expectation over a known prior), and rather hides it by assuming the error of amortization is uniformly bounded by \epsilon_am. This assumption seems unnecessarily strong -- in practice I would expect the approach really only requires an accurate estimate of the inverse map near the true / best approximating parameters.
  • While Appendix D shows a substantial reduction of wall-clock computation time from ~52 to <1s per patient, overall this still seems like a small gain compared to likely comparatively slow feedback ("function evaluation") times in this application... More generally, the question of when the overhead of amortization pays off seems not sufficiently explored.
Questions for authors
  • What is the computational cost of amortization in the PK/PD application? How does it relate to the per-patient optimization (cf Appendix D)? Did you explore tradeoffs regarding amortization (i.e., when in terms of number of offline samples used to train the inverse map vs sample acquisitions/evaluations of the learnt inverse map does amortization pay off within acceptable error)?
  • How is the prior over the simulation parameters chosen? What is the effect of its misspecification (i.e., placing little mass on the “true” parameters)?
  • \Delta_{model} is only specified in the appendix?
2 comments

Assumption that the error of amortization is uniformly bounded is unnecessarily strong

We agree that the assumption of uniform boundedness of the amortization error assumption is stronger than the proof requires. We revised Proposition 2 and Appendix B.3 so that the bound depends on the approximation error at a parameter whose exact simulator solution maximizes the true utility over the simulator-induced search space.

Reduction of wall-clock computation time due to amortization We now report in Table S2 of Appendix F (formerly Appendix D) the computation time of optimizing the AF when using either amortized optimization or direct optimization. Optimizing the AF when using amortized optimization is about four orders of magnitude faster.

In practical applications, the constraint is often on the online amortized optimization time. Table S2 shows that relying on direct optimization implies acquisition times that can exceed 6h, against 3s when using amortization. However, we agree that in our PK/PD application, the slow feedback makes the acquisition time less relevant. Therefore, our focus is rather on the better performance in terms of regret.

Computational cost of amortization in the PK/PD application

We computed the offline training time in our PK/PD application, keeping the same training configuration, and found it to be about 2.2 hours (5.3 min per epoch). We now report this value in the paper (Appendix D, section "Training cost").

Tradeoffs regarding amortization

Appendix F compares amortized optimization with direct L-BFGS-B optimization. The mean time to produce a regimen was 0.1160.116 ms for a single network forward pass and 11.71511.715 s for direct optimization (please note that we updated these numbers to exclude the one-off JIT compilation of the ODE solver).
The time it takes to perform one acquisition varies across iterations. Table S2 reports values that are about 3 seconds when using amortized optimization (tamt_{\mathrm{am}}), against 6 hours (tdirectt_{\mathrm{direct}}) for direct optimization.
Let's denote nn the number of BO iterations. The compute costs for whole BO runs with nn iterations are:

\qquad T_{\mathrm{direct}}(n) = n\\,t_{\mathrm{direct}}.$$ where $T_{\mathrm{offline}}$ is the offline training cost. So amortization saves compute when: $$n > \frac{T_{\mathrm{offline}}} {t_{\mathrm{direct}}-t_{\mathrm{am}}}, \qquad t_{\mathrm{direct}}>t_{\mathrm{am}}.$$ In our PK/PD experiment, $T_{\mathrm{offline}} \approx 2.2 \text{ hours}$, so $T_{\mathrm{offline}} /({t_{\mathrm{direct}}-t_{\mathrm{am}}}) < 1$. This means that the extremely high cost of optimizing the acquisition function when using direct optimization of the dosing regimen implies that amortization is always beneficial. > How is the prior over the simulation parameters chosen? We construct $p(\phi)$ from the published population PK model of Senek et al. and PD model of Westin et al. Their fixed effects and random effects specify typical parameter values and inter-individual variability. We sample independently from the reported marginals because the joint distribution is unavailable, then apply the plausibility filter in Appendix E (Eq. 37) to obtain the training sample. The search space used during Bayesian optimization, $\Phi$, spans the $0.025$ to $0.975$ quantiles of the prior marginals. > What is the effect of its misspecification (i.e., placing little mass on the “true” parameters)? A prior that assigns little mass to a relevant parameter region can affect BOASS in two ways. Sparse training coverage can increase the approximation error of $\mathcal E_w$ in that region. A shifted or overly concentrated prior can also move the quantile bounds so that the online search excludes parameter values that generate good regimens. The effect on attainable utility depends on the resulting image $\mathcal M=\mathcal E_w(\Phi)$: $$\Delta_{\mathcal M} =\max_{x\in\mathcal D} U(x) - \max_{\phi\in\Phi} U\bigl(\mathcal E_w(\phi)\bigr).$$ Previously, our misspecification experiments varied the target regimen to study the effect of an optimum lying outside the learned search space. We added a new set of experiments in Appendix G (Out-of-distribution patients), which answers the reviewer's question by applying the BOASS method to a population of patients whith a PK/PD parameters distribution that differs from the prior used when training the amortized optimizer (the PK/PD model inverse map). These experiments show the robustness of our method to prior misspecification, and Appendix G provides interpretation for this robustness. > $\Delta_{model}$ is only specified in the appendix? We corrected this issue in the new version. We put the definition of $\Delta_{\mathrm{model}}$ from Appendix B.3, Eq. (19) in Proposition 2, Eq. (4).

Thank you for the clarification. I will keep my score.

Reviewer_3ZtCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on the problem of Bayesian optimization and introduces Bayesian Optimization on Amortized Search Space (BOASS), an approach for integrating prior knowledge from differentiable mechanistic simulators, when available, to sample-efficient black-box optimization. The core idea of the paper lies on an initial training of a neural network architecture used to amortize the solution of a simulator-defined optimization problem. This is leveraged to define a more confined search space, over which the Bayesian optimization process is carried out. In addition, the paper puts forth a pullback Gaussian kernel that quantifies pairwise similarity between candidate solutions through their corresponding designs in the original search space, instead of directly utilizing simulation parameter representations. BOASS is evaluated on synthetic benchmarks including the Branin function and a GP-sample benchmark, and a personalized Levodopa-Carbidopa dosing problem for Parkinson's disease, using simulated patients, a published PK/PD model, with concentration, preference, misspecified and fully misspecified feedback. In addition to empirical evaluation, the paper provides theoretical guarantees through regret analysis where regret is decomposed into simulator misspecification, amortization error, and online optimization error.

Critical strengths

S1. The problem formulation and methodological solution are clean and it is a well-motivated problem: Judiciously exploiting mechanistic simulators, when available, to cope with high dimensionality challenges in black-box optimization problems is important particularly in settings where function evaluations are scarce. The motivation of combining amortized optimization and Bayesian optimization is clear.

S2. The integration of prior knowledge through “amortization” seems intuitive and interesting: The use of a simulator-trained neural network to construct the search space, instead of solely relying on expensive online observations to learn representations for BO, seems a reasonable choice.

S3. The appendix is comprehensive: Important components and aspects, including e.g., the PK/PD model, network, training, constraint handling, πBO calibration and Jacobian-based dimension analysis are well documented.

Critical weaknesses

W1. Theoretical assumptions and requirements described in the paper are not satisfied and/or conflict with the proposed approach. Specifically, Proposition 1 requires the learned search space (\mathcal M=E_w(\Phi)) to be a smooth, compact embedded submanifold without boundary, and the composition (g=U\circ E_w) to satisfy the derivative-tail condition of [Srinivas et al. (2010)]. However, it is not clear whether these assumptions hold for the proposed approach:

1.The learned search space is the image of a bounded parameter domain under a neural network architecture. Such an image is not guaranteed to be a smooth embedded manifold without boundary.

  1. The network in Appendix D uses ReLU activations, making (E_w) only piecewise differentiable in general. Consequently, the regularity conditions required for the pullback Gaussian process and the application of the derivative-tail assumption are not automatically satisfied.

Therefore, the eigenvalue-decay results of [Rosa et al. (2023), Rosa (2025)] , together with the GP-UCB analysis of [Srinivas et al. (2010)], cannot be directly applied to the learned search spaces without further justification. Although the theoretical results may hold under the stated idealized assumptions, the paper does not establish that these assumptions are satisfied in the experimental section.

W2. The numerical evaluation is limited and important baselines are missing. Apart from πBOASS, every online baseline ignores the simulator. Thus, it is unclear whether the gains come from the amortized search space or simply from access to a correct mechanistic model.

  • Model-based inference: In the well-specified settings φ_true ∈ Φ; thus, a reasonable alternative would be a Bayesian estimation of φ from the feedback, followed by selecting a dosing schedule based on (E_w(\hat{\phi})).. This is closely related to the model-informed precision dosing [Sheiner & Beal (1982), Darwich et al., (2017)] and Bayesian calibration [Kennedy & O’Hagan, (2001)]. Alternatively, dosing schedules could be selected through sequential Bayesian experimental design [Foster et al. (2021)].

  • More importantly well-tuned BO on D: The paper cites [Hvarfner et al. (2024)] but uses a fixed Γ(2,2) lengthscale-prior specification, rather than comparing against dimension-scaled lengthscale priors. Conventional BO with dimension-scaled priors, and SAASBO in the scalar-feedback settings, constitute non-expensive and fair baselines.

  • Simple dimensionality reduction baselines: Applying PCA to simulated regimens (E_w(\phi_s)) and subsequently performing BO within the resulting low-dimensional linear subspace would provide a useful comparison, inspired by PCA-assisted BO [Raponi et al., 2020]. Such an experiment would assess whether the nonlinear amortized representation and pullback kernel provide substantial advantages over simpler linear dimensionality reduction. Comparisons against random or adaptive embeddings [Letham et al., (2020), Papenmeier et al., (2022)] could further contextualize the merits of simulator-informed search-space construction.

References

Darwich, A. S., Ogungbenro, K., Vinks, A. A., Powell, J. R., Reny, J.-L., Marsousi, N., et al. (2017). Why has model-informed precision dosing not yet become common clinical reality? Lessons from the past and a roadmap for the future. Clinical Pharmacology & Therapeutics, 101(5), 646–656.

Foster, A., Ivanova, D. R., Malik, I., & Rainforth, T. (2021). Deep adaptive design: Amortizing sequential Bayesian experimental design. In Proceedings of the 38th International Conference on Machine Learning (Vol. 139, pp. 3384–3395). PMLR.

Hvarfner, C., Hellsten, E. O., & Nardi, L. (2024). Vanilla Bayesian optimization performs great in high dimensions. In Proceedings of the 41st International Conference on Machine Learning (Vol. 235, pp. 20793–20817). PMLR.

Kennedy, M. C., & O’Hagan, A. (2001). Bayesian calibration of computer models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 63(3), 425–464.

Letham, B., Calandra, R., Rai, A., & Bakshy, E. (2020). Re-examining linear embeddings for high-dimensional Bayesian optimization. In Advances in Neural Information Processing Systems (Vol. 33). Curran Associates, Inc.

Papenmeier, L., Nardi, L., & Poloczek, M. (2022). Increasing the scope as you learn: Adaptive Bayesian optimization in nested subspaces. In Advances in Neural Information Processing Systems (Vol. 35, pp. 11586–11601). Curran Associates, Inc.

Raponi, E., Wang, H., Bujny, M., Boria, S., & Doerr, C. (2020). High dimensional Bayesian optimization assisted by principal component analysis. In Parallel Problem Solving from Nature – PPSN XVI (Vol. 12269, pp. 169–183). Springer.

Sheiner, L. B., & Beal, S. L. (1982). Bayesian individualization of pharmacokinetics: Simple implementation and comparison with non-Bayesian methods. Journal of Pharmaceutical Sciences, 71(12), 1344–1348.

Questions for authors

Please refer to weaknesses W1 and W2 above.

2 comments

Theoretical assumptions and requirements described in the paper are not satisfied in the PK/PD experiment and/or conflict with the proposed approach

We agree that the smooth-manifold and derivative-tail assumptions of Proposition 2 remain unverified for our implementation. We have clarified the scope of the result in the Abstract, Introduction and Discussion. We used these to rely on established Gaussian process posterior contraction rate results, and think the theoretical analysis brings useful insights in understanding the behavior of the method even though these assumptions may not always hold in practice. Our experiments aim at providing empirical evidence for the benefits of the method beyond the ideal case where these assumptions can be proven. 

Baselines ignore the simulator

We updated the paper by adding the following baselines:

  • Conventional BO with a dimension-scaled prior (DSP) following Hvarfner et al. (2024a). Concretelly, we set the prior for the ARD lengthscales to the log-normal LN(2+12ln⁡14,3)\mathcal{LN}\big(\sqrt{2}+\frac{1}{2}\ln{14}, \sqrt{3} \big).
  • TurBO and SAASBO (previously only used in the synthetic experiments). The results are not substantially different from what we previously obtained for BO-on-DD.

We also implemented the following methods that leverage the mechanistic simulator :

  • PCA-assisted BO (PCA-BO) (described in Appendix L).
  • Model-based inference combined with Bayesian Experimental Design based on the Expected Information Gain (EIG) (described in Appendix M). This method is inspired by Model-Informed Precision Dosing and is denoted MIPD-EIG in the updated figures. The dose is optimized based on the PK/PD model whose posterior distribution of the parameters is inferred using patient feedback or measurements. In this approach, the PK/PD model is combined with an observation model to define a likelihood function, and the prior is defined as the distribution of the parameters in the population. The next candidate dosing regimen is the one with the largest expected information gain about the parameters.

In our drug dosing experiments, we found that BOASS beats PCA-BO when considering simple regret in all settings except the fully-misspecified setting, in which piBOASS still does better. BOASS performs much better than MIPD-EIG when considering experienced regret. When considering simple regret, MIPD-EIG performs as well as BOASS with preference feedback, and slighlty better during the first 20 irerations with concentration feedback. This lack of robustness can be explained by the fact that it's aiming at maximizing information gain about the parameters of the misspecified model, rather than maximizing the patient's utility. This is confirmed by supplementary experiments (Figure S9) in which we misspecified the observation model.

I would like to thank the authors for their response. Although the applicability of the theoretical guarantees to actual implementation still remains an open concern in my point of view, I appreciate the strengthened empirical evidence and the focused responses to my comments; therefore I am increasing my rating.

Reviewer_ei3iOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors train an amortized inverse map of a differentiable simulator and use it with a Gaussian process kernel to solve the optimization problem. They apply this method to a PK/PD problem and also provide a theoretical analysis of their method.

Critical strengths
  1. The authors introduce BOASS, which uses the trained inverse map and a GP.
  2. BOASS can exploit the intrinsic lower dimension and improve sample efficiency.
  3. The authors derive an upper bound for their optimization method under several assumptions.
Critical weaknesses
  1. Training the inverse map may be computationally expensive. The reported time excludes offline training.
  2. BOASS may not find the optimal point if the learned search space excludes it. The paper studies this issue and introduces πBOASS, but it is unclear when the hard restriction should be used.
  3. BO still runs on a higher-dimensional optimization problem than the original PK/PD problem (14 vs. 13 dimensions). The computational cost of the acquisition optimization is unclear.
Questions for authors
  1. Can you find explicit low-dimensional coordinates for the search space and use them in the acquisition optimization, so that you can show the runtime and performance?
  2. Can you use the regret upper bound to design a cheaper training objective for the inverse map?
2 comments

Training the inverse map may be computationally expensive.

We computed the offline training time in our PK/PD application, keeping the same training configuration, and found it to be about 2.2 hours (5.3 min per epoch). We now report this value in Appendix D, section "Training cost". However, we emphasize that in practical applications of BO, the constraint is on the online optimization time: here, the offline training cost is paid upfront, so that later online optimization problems are solved orders of magnitude faster. The advantage on the latter is now made more explicit in the new Table S2 (Appendix F) which compares the estimated total time for optimizing the acquisition function for all iterations, which is $\sim$106s for the amortized version and $\sim$291h with direct optimization.

it is unclear when the hard restriction should be used.

Because the utility is unknown, it is not possible to provide a threshold that can be checked before optimization for selecting betwen BOASS vs π\piBOASS. We recommend BOASS when the evaluation budget is small and there is evidence that the simulator-induced search space contains high-utility designs for the task. The latter case occurs when the simulator is trusted and amortized optimization is accurate. Our PK/PD application corresponds to this scenario as 1. Appendix F (Amortization analysis) shows good performance of amortized optimization over the patients population, 2. PK/PD models are already used in model-informed precision dosing for some applications, such as oncology, which reflects the fact they can be trusted, 3. the evaluation budget is very limited due to each iteration implying giving a new regimen to the patient and gathering feedback. However, if the evaluation budget is larger or important simulator mismatch is plausible, we favor π\piBOASS.
Figure 1 illustrates both regimes. Hard BOASS performed well in the moderately misspecified setting even though the optimum lay outside M\mathcal{M}. Under stronger misspecification, methods searching the full design space achieved lower simple regret. Note that the new experiments we added on prior misspecification (Appendix G, Out-of-distribution patients) show the strong robustness of BOASS to prior misspecification, such that we did not observe benefits of using π\piBOASS compared to BOASS in this case.

The computational cost of the acquisition optimization is unclear.

We thank the reviewer for raising this point, also raised by reviewer a9M2. We now report in Table S2 of Appendix F the computation time of optimizing the AF when using either amortized optimization or direct optimization. Optimizing the AF when using amortized optimization is about four orders of magnitude faster that when using direct optimization. Again, we thank the reviewers as these comments allowed us to better emphasize the usefulness of relying on amortization.

Can you use the regret upper bound to design a cheaper training objective for the inverse map?

Yes. Following reviewer a9M2’s suggestion, we revised Proposition 2 so that the amortization term depends on the regimen distance between the learned and exact simulator-optimal designs at ϕM⋆⋆\phi^{\star}_{\mathcal{M}^{\star}}, δ!(Ew(ϕM⋆⋆),E⋆(ϕM⋆⋆))\delta\\!\left(\mathcal{E}_w(\phi^{\star}_{\mathcal{M}^{\star}}), \mathcal{E}^{\star}(\phi^{\star}_{\mathcal{M}^{\star}})\right) (ϕM⋆⋆\phi^{\star}_{\mathcal{M}^{\star}} is the simulator parameter whose exact simulator-optimal design achieves the highest true utility over the ideal search space). This suggests fitting the inverse map directly to cached approximate simulator-optimal designs E~(ϕi)\widetilde{\mathcal{E}}(\phi_i) by minimizing L(w)=1N∑i=1Nδ!(Ew(ϕi),E~(ϕi))\mathcal{L}(w)=\frac{1}{N}\sum_{i=1}^{N}\delta\\!\left(\mathcal{E}_w(\phi_i),\widetilde{\mathcal{E}}(\phi_i)\right).

This has an important implication: the training pipeline can therefore use a nondifferentiable simulator if the targets E~(ϕi)\widetilde{\mathcal E}(\phi_i) are generated with a derivative-free solver. Appendix K.1.2 already uses direct regression against the ideal map in the synthetic experiment. For the drug dosing application, any reduction in total training cost would depend on the cost of generating sufficiently accurate targets. We updated the paper to make this point. In particular, we now make explicit the fact that our proposed method does not require the simulator to be differentiable.

Can you find explicit low-dimensional coordinates for the search space and use them in the acquisition optimization, so that you can show the runtime and performance?

We implemented the PCA-BO method to explicitly restrict the search space to a 3-dimensional space. Appendix N.2 now describes the runtime measurements (reported in Table S6), and shows how explicit search space dimension reduction leads to reduced acquisition time in the BO loop. The results of PCA-BO are reported in Figure 1 alongside the other methods.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Rebuttal

We thank the reviewers for their thoughtful analysis and feedback, which has been highly valuable for understanding how to improve our paper. We are pleased that reviewers in general agreed that the problem we aimed to address is well-motivated and clearly presented (3ZtC, a9M2, ei3i), and that the proposed approach of exploiting prior information encoded in simulators to exploit their intrinsic lower dimension and improve sample efficiency is both intuitive and interesting (3ZtC, a9M2). We are also glad reviewers found that our analysis is comprehensive and well-documented (3ZtC), with the PK/PD application interesting and relevant (a9M2).

Reviewers raised thoughtful questions about the cost of training the inverse map, the overhead of amortization and the computational cost of the acquisition function (ei3i, a9M2), the assumptions of the theoretical analysis (3ZtC, a9M2), and the practical choice of piBOASS vs BOASS (ei3i). Reviewer 3ZtC also recommended adding baselines that also leverage the simulator.

In the following responses, we address each of the reviewers’ questions and concerns in detail. Additional experiments were conducted when requested, and we have uploaded a pdf containing an updated version of the paper with updated figures for these experiments and tables.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.