acceptodds
Under review as a conference paper at ICLR 2027

Distributional Bayesian Optimal Experimental Design

Abstract

Sequential Bayesian optimal experimental design (BOED) chooses informative experiments adaptively based on the outcomes of prior experiments. The standard sequential BOED objective, expected information gain (EIG), scores an experimental design policy by the mean of its information gain (IG) distribution. However, the mean alone can obscure important characteristics of the distribution, such as its variance or lower-tail behavior. To address this limitation, we present DistBOED, a distributional learning method for sequential BOED that models the full distribution of IG induced by a design policy. DistBOED uses a quantile critic and a campaign-level actor update to learn policies for objectives beyond mean IG. We demonstrate DistBOED across three different scenarios with various risk-sensitive IG measures including: conditional value at risk, variance-penalized mean, and entropic risk measure. In each case we observe that accounting for higher-order aspects of the IG distribution yields more robust design policies than EIG alone.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian optimal experimental design, sequential experimental design, risk-sensitive decision-making, sequential decision-making, tail-risk optimization, expected information gain

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_ULK9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Distributional Bayesian Optimal Experimental Design (DistBOED), a distributional RL method for sequential BOED that models the full conditional distribution of the information gain (IG) rather than only its expectation (EIG). The authors evaluate DistBOED in three scenarios using three risk-sensitive objectives, showing that DistBOED with risk-sensitive objectives produces policies that reduce dispersion or improve lower-tail IG relative to mean-seeking policies.

Critical strengths
  1. The method is well-motivated: in sequential BOED, a policy that optimizes total EIG may favor designs with high expected information gain but also high risk. It is sensible to bias the policy towards more protection via risk-sensitive objectives over the IG distribution.
  2. Principled formulation from the risk-sensitive objectives to the score function.
Critical weaknesses
  1. The formulation assumes a finite design menu, limiting its scalability to continuous design spaces.
  2. The baseline is limited to only DAD: comparisons with baselines that could take risks into account would strengthen the claims.
Questions for authors
  1. [1] could possibly handle risk in sequential BED via explicit constraints and online planning. Could you discuss how DistBOED and the method differ and share in settings?
  2. At inference time, can the protection level of the learned policy be adjusted without retraining?

[1] Guo, Yujia, et al. "Constrained Bayesian Experimental Design via Online Planning." arXiv preprint arXiv:2605.26990 (2026).

1 comment

We are grateful to the reviewer for the positive comments on the motivation and formulation and for the constructive feedback, which we address below.

Responses to questions

Q1. Comparison to other methods

[1] could possibly handle risk in sequential BED via explicit constraints and online planning. Could you discuss how DistBOED and the method differ and share in settings?

Response. We agree that COPEx could possibly be extended to address risk in information outcomes. As presented, it optimizes expected information gain under design constraints, so a risk-sensitive extension could optimize or constrain a function of campaign information gain.

The difference is when the design optimization is performed. COPEx combines pretrained networks with online multi-step planning, while DistBOED learns a policy in advance and uses it to select designs from the observed state without solving a new design-optimization problem during the campaign.

Moving this optimization into training is particularly useful when many campaigns must be run or designs must be selected with little delay. This computational benefit is demonstrated by DAD [2] and iDAD [3], which report millisecond design selection at deployment.

COPEx would therefore offer more flexibility to accommodate changing constraints during a campaign, while DistBOED provides a reusable policy for a specified risk preference. DistBOED is particularly suited to repeated experiments in the same setting where rapid design selection matters.

Q2. Inference-time adjustments

At inference time, can the protection level of the learned policy be adjusted without retraining?

Response. Not for a single trained policy in the current implementation. As discussed in Q1 above, DistBOED trains policies for specified risk preferences. Changing the protection level requires further policy optimization or selecting another pretrained policy; this would be a disadvantage compared to an online planning method like COPEx. The learned critic can still evaluate alternative risk preferences under the existing policy without retraining.

Additional weaknesses not covered above

W1. Finite design menu

The formulation assumes a finite design menu, limiting its scalability to continuous design spaces.

Response. We agree that the current implementation is limited to finite design menus. The risk-sensitive objectives can also be defined for continuous designs, but extending DistBOED would require a continuous-action policy, a critic accepting continuous designs, and an alternative to explicit summation over candidate designs in the actor update. We acknowledge the finite design menu in the manuscript’s limitations and leave the continuous extension to future work.

W2. Baselines

The baseline is limited to only DAD: comparisons with baselines that could take risks into account would strengthen the claims.

Response. A central contribution of this work is a unified framework for optimizing risk-sensitive objectives over campaign information gain. Methods such as DAD and COPEx do not directly optimize these objectives, so constructing comparisons would require extending existing methods.

We have nevertheless added a scalar actor-critic baseline that optimizes the same risk-sensitive objectives (Appendix F). This baseline achieves broadly comparable performance and provides a direct comparison with the distributional approach. We have also evaluated DAD with an explicit destruction penalty. The destruction penalty addresses a different concern, avoiding a physical event, but allows us to examine how targeting a physical source of information loss compares with targeting information risk directly. We discuss these results in more detail in the response to reviewer rPYW.

[2] Foster, A., Ivanova, D. R., Malik, I., and Rainforth, T. (2021). Deep Adaptive Design: Amortizing Sequential Bayesian Experimental Design. Proceedings of the 38th International Conference on Machine Learning, 139:3384–3395.

[3] Ivanova, D. R., Foster, A., Kleinegesse, S., Gutmann, M. U., and Rainforth, T. (2021). Implicit Deep Adaptive Design: Policy-Based Experimental Design without Likelihoods. Advances in Neural Information Processing Systems, 34:25785–25798.

Reviewer_rPYWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that sequential BOED should not target only the expected information gain (EIG), because the mean of the information gain (IG) distribution induced by a policy hides other properties that matter. The focus here is on the spread and lower-tail behaviour of the IG. It proposes DistBOED, an actor-critic method with two parts:

A quantile critic predicts the distribution of terminal IG, conditioned on a state summary and a candidate design. The actor is trained with a likelihood-ratio gradient on a transformed return fU(g)f_U(g). This transform is chosen so that its expected gradient matches the gradient of a target functional U(Fπ)U(F_\pi): the mean, CVaR, mean-variance or entropic risk.

The critic acts as an action-dependent control variate. The method is evaluated on three small problems, each with a 2D parameter and a discrete design set: a destructive probe, source localisation with a jammer, and a misspecified spring-mass system scored with generalised IG. Risk-sensitive policies reduce variance and improve lower-tail outcomes at some cost to mean IG. The mean-trained policy roughly matches DAD.

Critical strengths
  1. The motivation is relevant and well made. In sequential design, the spread and lower tail of campaign IG can matter as much as the mean, and EIG cannot express that preference. The three settings are good examples: destroyed specimens, jammed sensors, and misleading inference under misspecification. The paper also gives one training procedure that accepts several functionals of the IG distribution, which is a useful framework.
  2. The experiments cover different sources of risk, and the learned policies are interpretable (Figs. 2, 5, 8). Risk-sensitive policies delay high-intensity probes, place receivers more centrally, and use weaker excitation in the spring problem.
Critical weaknesses
  1. The paper does not discuss its choice to define IG as DKL[p(θ∣hT) ∥ p(θ)]D_{KL}[p(\theta\mid h_T)\,\|\,p(\theta)] rather than the entropy reduction H[p(θ)]−H[p(θ∣hT)]H[p(\theta)] - H[p(\theta\mid h_T)] used by Lindley (1956) and Rainforth et al. (2024).
    • Both have expectation equal to EIG, but their distributions differ. The CVaR, variance and ERM of IG therefore depend on which definition is used.
    • The KL version is always non-negative. It rewards a posterior that moves far from the prior even if it stays diffuse.
    • Entropy reduction can be negative, which seems to be exactly the low-information outcome a risk-averse experimenter wants to avoid.
    • The two definitions coincide under the uniform prior in the jammer, but not under the Gaussian prior in the probe. Do the probe conclusions hold under entropy reduction?
  2. Section 3 is somewhat hard to follow because the critic is introduced in full before the reader knows why it is needed. I'd suggest reordering it: start from the objective U(Fπ)U(F_\pi) and the transformed-return policy gradient (Eqs. 15 and 18), then introduce the critic as a control variate that reduces the variance of that gradient (Eqs. 16–17), and only then describe how the critic is trained (current Sec. 3.1). This would also make clear that the critic does not change the expected gradient.
  3. Perhaps most importantly, the paper does not show that the distributional critic is needed.
    • At ϕold\phi_{old} the critic terms cancel in expectation, and the actor gradient is REINFORCE on the scalar return fU(g)f_U(g) observed from each simulated campaign (Eq. 18). The critic only reduces variance. The first ablation should therefore be standard REINFORCE with a simple baseline, to test whether a critic is needed at all. Perhaps all the benefit comes from just changing the objective from EIG to the relevant risk aware objective.
    • Even if a critic control variate helps, nothing requires it to be distributional. The actor only uses the critic through one scalar per state–design pair, 1K∑kfU(qk)≈E[fU(G)∣s,d]\frac{1}{K}\sum_k f_U(q_k) \approx E[f_U(G)\mid s,d] (Eq. 16), so a scalar critic trained directly on fU(g)f_U(g) would serve the same purpose. For the mean-variance objective, fUf_U is quadratic in gg, so a critic predicting only the first two moments of GG would suffice and is likely much easier to train than K=101K=101 quantiles. Currently, the case for the distributional machinery is not evidenced.
  4. Several trained policies lose on the objective they were trained for, which undercuts the claim that DistBOED optimises the chosen functional:
    1. In Table 2, the entropic policy has lower ERM$_1$ than the mean policy.
    2. In Table 1, the entropic policy beats the CVaR policy on CVaR$_{0.15}$.
    3. In Table 7, training at α=0.25\alpha=0.25 gives better CVaR$_{0.1}$ than training at α=0.1\alpha=0.1.
  5. The results appear to come from a single training seed with no confidence intervals. That makes these orderings, and the "similar to DAD" claim, hard to interpret.
Questions for authors
  1. In the destructive probe experiment, the risk has a direct proxy (destruction), so a natural baseline is an EIG-trained policy (e.g. DAD) with a destruction penalty or constraint. This needs none of the distributional machinery and is easier to calibrate than α\alpha or β\beta. How does DistBOED compare with this baseline? More generally, the paper should state when a functional of the IG distribution is needed rather than a constraint on an observable outcome. The spring and jammer problems seem like the stronger cases, since there the risk is defined through the IG itself.
  2. What does the distributional quantile critic give over REINFORCE with a baseline, over a scalar critic of E[fU(G)∣s,d]E[f_U(G)\mid s,d], or, for mean-variance, over a critic that predicts only E[G∣s,d]E[G\mid s,d] and E[G2∣s,d]E[G^2\mid s,d]? I suspect the last of these in particular would be easier to train than a critic with K=101K=101 outputs per design.
  3. How sensitive are the results to KK?
  4. Are the tables from a single training seed? Please report variability across seeds.

Minor

  1. "Score" is used for fUf_U and for the critic-based design values f^U,ψ(s,d)\hat f_{U,\psi}(s,d), and the gradient estimator itself is a score-function estimator. I find this repeated use of the term confusing.
  2. The Fig. 1 caption has two problems. It says "three DistBOED objectives" but four are shown. "The left tail is not as negative" does not fit a non-negative KL-based IG; perhaps say that risk-sensitive policies put less mass at low IG.
  3. Typos and small errors:
    1. Broken citation "from ?." in App. B.2.
    2. "Appendex", "to to", and "an arbitrary functions".
3 comments

We thank the reviewer for the thorough and helpful feedback and constructive comments, which we address below.

Responses to questions

Destruction penalty or constraint

In the destructive probe experiment, the risk has a direct proxy (destruction), so a natural baseline is an EIG-trained policy (e.g. DAD) with a destruction penalty or constraint. This needs none of the distributional machinery and is easier to calibrate than an α\alpha or β\beta. How does DistBOED compare with this baseline? More generally, the paper should state when a functional of the IG distribution is needed rather than a constraint on an observable outcome. The spring and jammer problems seem like the stronger cases here since the risk is defined through the IG itself.

Response. A destruction penalty is a natural choice when avoiding destruction is itself a goal. One could similarly construct task-specific utilities for other sources of risk. However, different experimental conditions can interact in their effects on information gain, making it difficult to capture those effects with separate penalties.

Our contribution is a framework for expressing preferences directly over learning outcomes themselves. DistBOED uses the same risk-sensitive objectives and policy-optimization procedure across all three settings, without adding a separate penalty for each mechanism. This lets practitioners specify what constitutes a desirable information outcome directly. A functional of the IG distribution is useful when the concern is the reliability of the information obtained, particularly when no single observable event captures that concern.

We compare a DAD policy trained with a destruction penalty against DistBOED-CVaR:

PolicyMean IG ↑CVaR15 ↑Destruction (%) ↓
DAD, no penalty3.255 [3.252, 3.258]1.734 [1.718, 1.751]52.394 [51.654, 53.135]
DAD, penalty 0.253.245 [3.241, 3.249]1.855 [1.839, 1.872]45.606 [44.890, 46.322]
DAD, penalty 0.53.207 [3.200, 3.213]1.944 [1.930, 1.958]36.044 [35.143, 36.946]
DAD, penalty 13.094 [3.089, 3.100]2.093 [2.082, 2.104]21.637 [20.987, 22.287]
DAD, penalty 22.742 [2.702, 2.782]1.974 [1.945, 2.003]0.850 [-1.067, 2.768]
DAD, penalty 42.724 [2.723, 2.726]1.961 [1.960, 1.962]0.0010 [-0.0005, 0.0025]
DistBOED-CVaR2.933 [2.902, 2.963]2.089 [2.075, 2.103]10.053 [8.250, 11.856]
(Results use 10 seeds)

These results show that a destruction penalty provides a useful tradeoff. With penalty 1, DAD achieves similar CVaR to DistBOED-CVaR (2.093 versus 2.089) and higher mean IG (3.094 versus 2.933), but has a destruction frequency of 21.637% compared with 10.053%. The tested penalties that reduce destruction below DistBOED-CVaR also yield lower mean IG and CVaR.

Necessity of a distributional critic

What does the distributional quantile critic give over REINFORCE with a baseline, over a scalar critic of E[fU(G)∣s,d]\mathbb{E}[f_U(G) \mid s, d], or, for mean-variance, over a critic that predicts only E[G∣s,d]\mathbb{E}[G \mid s, d] and E[G2∣s,d]\mathbb{E}[G^2 \mid s, d]? I suspect the last of these in particular would be easier to train than a critic with K=101K=101 outputs per design.

Response. We have added a scalar actor-critic comparison to Appendix F of the updated manuscript. It achieves broadly comparable performance, confirming that a scalar critic is a viable choice for a fixed objective in the settings studied.

We choose to use a distributional critic for its additional functionality. Practitioners can evaluate multiple utilities and characterize information risk under the trained policy without fitting separate predictors. Our general comment on the distributional critic discusses these benefits and how the choice of critic fits within the broader contribution of a practical framework for risk-sensitive sequential BOED.

Training seeds

The results appear to come from a single training seed with no confidence intervals. That makes these orderings, and the "similar to DAD" claim, hard to interpret.

Response: We have updated Tables 1--3 to use 10 training seeds per policy, with means and 95% confidence intervals, and Appendix D3 describes the evaluation and interval calculations.

These new results also address the points raised in Weakness 4. In Table 2, the entropic-trained policy now achieves a higher average ERM than the mean-trained policy. In Table 1, the CVaR-trained policy now achieves a higher average CVaR than the entropic-trained policy.

Additional weaknesses not covered above

W1. Definition of IG

The paper does not discuss its choice to define IG as DKL[p(θ∣hT)∣∣p(θ)]D_{\rm KL}[p(\theta \mid h_T)\mid\mid p(\theta)] rather than the entropy reduction H[p(θ)]−H[p(θ∣hT)]H[p(\theta)] - H[p(\theta \mid h_T)] used by Lindley (1956) and Rainforth et al. (2024).

We chose the KL formulation because it measures change in belief, including updates that do not reduce uncertainty. For example, an observation that updates a prior distribution of (0.1, 0.9) to a posterior of (0.9, 0.1) is certainly informative, but the entropy-reduction form would give this update a score of 0. Similarly, increasing posterior entropy doesn't necessarily mean an observation was uninformative, since new evidence could make previously discounted hypotheses plausible. This would leave us less certain but still provide us with meaningfully different beliefs.

Our choice therefore reflects a preference for valuing belief revision, but the entropy reduction form is certainly legitimate -- it is appropriate when the practitioner specifically values uncertainty reduction. We note that nothing about the DistBOED framework prevents using the entropy reduction formulation; replacing the terminal reward with the entropy reduction would allow the same method to optimize risk-sensitive objectives under that definition.

We agree that this point needs better clarification in the manuscript, and we have updated Section 2.1 in the manuscript to reflect this discussion.

W2. Section 3 clarity

Section 3 is somewhat hard to follow because the critic is introduced in full before the reader knows why it is needed.

Response. Thank you to the reviewer for this very helpful suggestion. We've reordered Section 3 in the updated manuscript to reflect the suggested exposition and hope that it reads more clearly now.

Minor issues

We thank the reviewer for pointing these out and have made these corrections in the revised manuscript. In particular, the use of "score" has been changed to “return transformation” for ff, “transformed return” for f(g)f(g), and “design value” for the critic prediction.

Thank you to the authors for their detailed responses and substantial additional work, which I genuinely appreciate. I believe there is a valuable paper here, but I think it requires a clearer framing of its contribution. The experiments convince me of the value of risk-sensitive BOED, but not of the need for distributional BOED. The distinguishing distributional component is the distributional critic that explicitly models the IG distribution, yet the new experiments suggest that this is not necessary to achieve comparable performance from a risk-sensitive perspective. Indeed, just as standard BOED optimises the expectation of IG, one can optimise other functionals of IG without explicitly learning its distribution. I think a paper framed around risk-sensitive BOED could be competitive, but the current distributional framing is not sufficiently justified. Alternatively, if the main advantage of the distributional critic is amortising a learned representation that can be reused to evaluate different risk preferences, the experiments should demonstrate this benefit directly, including its computational advantages and predictive accuracy, rather than primarily showing the benefits of risk-sensitive objectives. For these reasons, I believe a more substantial revision is needed and will maintain my current assessment.

Reviewer_5jDrOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

While typical sequential Bayesian optimal experimental design (BOED) scores the efficacy of a potential experiment based on the "mean" of the information gain (IG), this paper investigates how this may obscure important characteristics regarding how the IG is distributed. For example, the lower tail behavior of sequential BOED might be of interest, which may get lost when using the expected IG alonw, and this paper exactly tackles this issue in this work. In order to address this limitation, the authors propose DistBOED, a "distributional" learning method for sequential BOED, where the entire IG distribution is considered for designing experiments, rather than the mean IG alone. The study shows how DistBOED can lead to more robust experimental design policies compared to standard sequential BOED methods, across three scenarios using different "risk-sensitive" IG measures.

Critical strengths
  1. Overall, the manuscript gives a concise yet clear overview of standard sequential BOED approaches and motivates the proposed distribution approach towards sequential BOED. The manuscript is well-written overall, with good technical and literary presentations of DistBOED.

  2. When robustness and reliability of experimental design policies are of critical importance, it makes sense to consider the entire distribution of IG instead of EIG alonw in BOED. DistBOED tackles this important problem and based on various risk-sensitive IG measure - such as conditional value at risk (CVaR), variance-penalized mean, and entropic risk measure - the study demonstrates potential advantages of DistBOED in enhancing the robustness of experimental design.

  3. The experimental results presented in the manuscript point to the potential trade-off between different strategies, what is gained (or might be lost) by taking a distributional approach (i.e., DistBOED), and also provide interesting and often intuitive insights.

Critical weaknesses
  1. Probably one of the major weaknesses of the current study is that the proposed DistBOED is somewhat limited in terms of novelty. While taking the entire distribution of IG is a well-motivated and interesting idea, which might have been also underexplored, the actual strategy for implementing more robust and risk-sensitive sequential BOED is not necessarily novel.

  2. Having an ablation study that removes/limites the role of the distribution critic might be helpful in more clearly understanding the role and merit of the proposed critic. For example, the authors may want to consider how the performance may be affected when no distribution critic is used or if the critic only utilizes the mean IG instead of its distribution.

  3. Currently, the manuscript doesn't report the statistical significance of the results summarized in the figures and tables. (e.g., confidence intervals). To make sure that the presented results are representative and not anecdotal, statistical significance needs to be reported.

Questions for authors

Please see the comments in the weaknesses section. Provide clarifications and additional results to address these concerns.

2 comments

We are very grateful to the reviewer for the feedback and for their positive comments on the paper's clarity and importance. We appreciate the constructive feedback, which we address below.

Responses to Weaknesses

W1. Novelty

Probably one of the major weaknesses of the current study is that the proposed DistBOED is somewhat limited in terms of novelty. While taking the entire distribution of IG is a well-motivated and interesting idea, which might have been also underexplored, the actual strategy for implementing more robust and risk-sensitive sequential BOED is not necessarily novel.

Response.

To our knowledge, no prior work provides a unified framework for learning sequential BOED policies under a range of risk-sensitive functions of campaign information gain.

The novel contribution of this paper is that DistBOED provides this capability, giving practitioners a practical way to express preferences over the information gained from an experimental campaign. The experiments demonstrate that these preferences produce meaningfully different policies and learning outcomes. We believe that providing this capability within a flexible framework is a useful contribution to BOED, even though the underlying optimization tools are established.

W2. Distributional critic

Having an ablation study that removes/limites the role of the distribution critic might be helpful in more clearly understanding the role and merit of the proposed critic. For example, the authors may want to consider how the performance may be affected when no distribution critic is used or if the critic only utilizes the mean IG instead of its distribution.

Response. We have added a scalar actor-critic ablation to Appendix F. The comparison retains the same risk-sensitive objectives and actor update, while replacing the distributional critic with a scalar predictor of the expected transformed return.

The scalar critic achieves broadly comparable performance, showing that a distributional critic is not essential for optimizing a fixed objective in the settings studied. However, the distributional critic provides additional functionality beyond just training. Practitioners can estimate multiple risk measures and probabilities of low-information outcomes under the trained policy without fitting separate predictors. This flexibility motivates our choice. We discuss these uses of the learned distribution further in our general response.

W3. Statistical significance

Currently, the manuscript doesn't report the statistical significance of the results summarized in the figures and tables. (e.g., confidence intervals). To make sure that the presented results are representative and not anecdotal, statistical significance needs to be reported.

Response. We apologize for the oversight and have added 95% confidence intervals across 10 independent training seeds to Tables 1--3.

Thank you for the additional information and results. W1. Considering that there exist studies on risk-sensitive/risk-aware BOED to robustly achieve specific objectives/goals, and more broadly, on robust and risk-sensitive/risk-aware sequential decision-making, it would be helpful if the authors provide a more comprehensive literature review on these topics to provide a clearer context for the current work, novelty, and significance. W2 Thanks for adding this result. This is helpful. W3. Thank you for adding these statistical significance analyses. This is helpful for understanding the presented results. Overall, I think the original rating is still suitable, and I would like to maintain the score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to general concerns

We first would like to thank all the reviewers for their time and thoughtful, constructive reviews, which have strengthened the paper. Several reviewers asked about the role of the distributional critic and requested ablations to assess its contribution to optimizing the chosen objective. We have added a comparison to a scalar actor-critic baseline in Appendix F of the updated manuscript. It achieves broadly comparable performance, so a scalar critic is a viable choice for a fixed objective in the settings we study. We chose a distributional critic for the additional functionality it provides.

In distributional RL, the choice of a distributional critic is motivated by potential benefits of richer predictions for function approximation and learning. In our setting, this broader representation has a further practical use: it supports questions about experimental outcomes that go beyond the objective used to train the policy.

For any given experimental history, the learned critic can be used to estimate EIG, multiple risk measures, the probability of an outcome below a given information threshold, and other utilities of IG under the current policy. Practitioners can inspect these quantities together and assess alternative preferences without fitting a separate predictor for each one. A scalar critic trained for one utility does not provide this information.

We see our main contribution as a practical framework for risk-sensitive sequential BOED, together with evidence that different preferences over information gain produce different policies and learning outcomes. The choice of critic is one component of that framework. Given broadly comparable optimization performance, the ability to evaluate multiple preferences and characterize information risk from one learned representation motivates our choice.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.