- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper studies acquisition when an action becomes deployable only after it has been measured, with separate measurement and global deployment budgets. A2-EDUG shortlists queries using immediate expected deployable-utility gain, evaluates a bounded two-step continuation for four candidates, executes the first query, and replans. A counterexample illustrates arbitrarily poor one-step behavior. Image-compression and speech-representation experiments report lower query-path regret under a fixed belief model and measurement budget, with matched solver-call and approximation controls.
Critical strengths
-
The queried-only deployment formulation captures a useful distinction: measuring an action also makes it available for deployment. The counterexample clearly illustrates how a global deployment budget can create complementarities that one-step acquisition misses.
-
The empirical controls are careful. The paper fixes the belief model and deployment solver and includes matched solver-call, exhaustive-planning, shortlist, and fixed-second-query comparisons. These results support the usefulness of a two-query horizon within the tested setting.
Critical weaknesses
-
The methodological contribution of the lookahead construction could be articulated more clearly. The approach combines familiar shortlisting and two-step rollout components, and practical two-step lookahead is already studied in cited work such as Wu and Frazier (2019), albeit in a different setting. The competitive fixed-second-query control suggests that much of the demonstrated gain comes from extending the horizon; it does not establish a separate benefit from the adaptive continuation machinery. Explaining the methodological insight specific to queried-only deployment would strengthen the contribution.
-
An important question is how reliably the candidate-selection rule captures the complementarity motivating the method. Ranking candidates by immediate value can exclude actions whose value emerges only jointly, while Proposition 3 uses an instance where the top-four shortlist is exhaustive. The existing shortlist-recall and exhaustive-planning analyses acknowledge this approximation, but all episodes still contain four items and 13 actions per item. It would be valuable to clarify how well this behavior is expected to extend to larger coupled decisions. Given the roughly 24.7-second planning cost versus about 10 seconds for one-step EDUG (Table 2), such evidence would also help readers identify the regimes where the extra planning is worthwhile.
Questions for authors
-
Could you clarify the methodological insight beyond standard two-step rollout, particularly in light of the competitive fixed-second-query control?
-
When does the immediate-value shortlist retain actions whose benefit is mainly complementary? Is there evidence beyond four-item allocations, or a structural condition supporting this design choice?
-
Is the equal-weight planning score in Eq. (10) intentionally a surrogate for trapezoidal AURC in Eq. (5)? For two queries and fixed initial regret, AURC weights the first and second post-query regrets 2:1. Clarifying this relationship would help assess the method’s objective.
We sincerely thank the reviewer for the careful and constructive assessment. We especially appreciate the reviewer’s recognition of both the motivation for jointly simulating two options and the breadth of our experimental controls. The comments on planning cost and the fixed-second control identify two important questions that the manuscript did not explain clearly enough. We address these questions below.
Complexity and the efficiency tradeoff. Let Uq denote the number of unseen cells at query depth q, S the number of posterior draws, k the shortlist size, Talloc the cost of one exact queried-only deployment solve, and Trank the cost of scoring one candidate under a conditional-mean continuation. In our experiments, S=256 and the primary setting is k=4.
Up to lower-order shortlist-selection and posterior-conditioning terms, the leading cost of one-step EDUG is TEDUG(q)=O(UqSTalloc). A2-EDUG first performs the same all-candidate EDUG scoring and then evaluates kS conditional continuations: TA2(q)=O(UqSTalloc+kSUqTrank+kSTalloc).
With the per-solve costs held fixed, both methods are linear in Uq when k and S are fixed. The additional cost is controlled by k and the Monte Carlo budget S. If k grows with the search space, the shortlist-dependent continuation term can become quadratic in Uq. This decomposition separates planner overhead from the exact deployment solver, whose own cost depends on the multiple-choice knapsack implementation and on N, M, and B.
The existing per-trajectory planning audit makes the practical frontier explicit. Relative to EDUG, k=4 takes 2.45× the planning time on DIV2K and 2.62× on FSDD, while reducing AURC by 11.99% and 16.25%, respectively. The cheaper k=2 setting gives 9.37% and 10.60% reductions, whereas k=8 gives 15.56% and 18.75% at roughly 4.5–4.8× the EDUG time. Thus, k=4 is an operating point on a tunable compute–quality frontier, not a claim of universal efficiency.
Fixed-second control. We agree that the fixed-second control is an important comparison. It retains the two-query horizon and the same k=4 planning setting while removing outcome-dependent re-selection of the second query. With 568.8k solver calls and essentially the same planning time, it gives a 13.33% reduction on DIV2K and a 15.44% reduction on FSDD, compared with 11.99% and 16.25% for outcome-adaptive A2-EDUG.
The conceptual advantage of outcome-adaptive continuation is available when the best second query changes with the simulated first outcome; when the best continuation is outcome-invariant, adaptive continuation reduces to a fixed continuation. In our current experiments, the difference between the two policies is small, and the paired intervals for the A2-EDUG versus fixed-second difference cross zero.
We will therefore make the claim more precise in the revision: the evidence supports the value of a bounded two-query lookahead horizon in queried-only deployment, but it does not establish a separate benefit from outcome-adaptive branching in the tested domains. We will add the precise fixed-second definition, move the compute–quality interpretation into the main discussion, and include the complexity accounting above.