- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper seeks a Bayesian optimization setting in which the reward-maximizing policy can be computed exactly rather than approximated by myopic acquisition functions. The unknown function is linear in the stereographic projection of the input, , so each function has one peak and one valley and inference is Bayesian linear regression; queries are priced as with a known expensive direction c; and the learner is scored on one final query by the squared posterior-mean gap between its query and a random point drawn on the boundary of the affordable region. Under an assumption on the posterior mean (A3), the terminal value is with , which annihilates c. The authors then show that the Bellman equation has the closed-form solution , so the optimal schedule is non-adaptive; that it reduces to an exact design problem over information matrices characterized via Schur–Horn; and, when c is an eigenvector of the prior precision, that the optimal design is water-filling on the prior spectrum with no weight on c, computable in . Synthetic experiments verify the value function against Monte Carlo and compare with cost-blind A-optimal design, uncertainty sampling and random queries for n = 3 to 128.
Critical strengths
(1) The goal of identifying a nontrivial sequential design problem where the Bayes-optimal policy can be written down, and using it to explain when adaptivity has value, is a worthwhile and clearly argued research direction. (2) The derivations are clean and, under the stated assumptions, correct. (3) The paper is reasonably well written, with good intuition for each step.
Critical weaknesses
My main concern is the practicality of the problem. It seems the problem setting is engineered so that the main conclusions hold, and the link to Bayesian optimization is weak. The random baseline on the boundary slice is chosen so that its mean exactly cancels the expensive-direction component of the final query, which is what makes information along c worthless. Also, the squared reward with adversarial sign rewards peaks and valleys alike and even rewards a query orthogonal to on the degenerate branch. None of this corresponds to a standard BO objective such as the expected value of f at the recommended point or simple regret. Once the stereographic projection drops out (Section 6), it seems that the problem is Bayesian L-optimal linear design with linear costs, for which non-adaptivity under a quadratic terminal reward is well known (posterior covariance independent of responses, law of total variance), so the conceptual novelty is limited.
Questions for authors
So my questions are: Can the authors motivate the scoring rule (squared posterior-mean gap against a random point on the affordable boundary, with adversarial sign) from a BO perspective, and say which conclusions survive under a standard objective such as the expected value of f at the final query or simple regret? How do your results differ from classical Bayesian L-optimal design for linear-Gaussian models, where open-loop optimality under quadratic terminal reward is known?
Thanks for the detailed response, and I do not have further questions. Part of my concerns are solved after reading the "What is classical and what is new" section, but I still feel that the square score is considered only because it makes the Bellman equation solvable. So the practicality of the work is a little weak, and I will keep my score unchanged.