acceptodds
Under review as a conference paper at ICLR 2027

GUIDE-FBO: Guidance via Uncertainty Intervention and Distributional Exchange for Federated Bayesian Optimization

27 views

Abstract

Federated Bayesian Optimization (FBO) enables distributed agents to collaboratively optimize expensive black-box objectives without sharing raw local observations. However, effective knowledge transfer remains challenging under communication constraints and task heterogeneity. We propose GUIDE-FBO, in which agents exchange compact distributions over the locations of their respective optima inferred from local Gaussian process (GP) posteriors, rather than raw observations, query points, or surrogate parameters. The server merges and reweights these distributional components before returning a subset to each agent. Each agent then constructs a Federated Interventional GP (FI-GP), which preserves the local posterior mean and spatially rescales its covariance for local decision making. For the upper confidence bound (UCB) instantiation, GUIDE-UCB, we prove that any bounded FI-GP uncertainty intervention preserves the leading-order cumulative regret rate of standard GP-UCB. When the transferred distributions place greater support near an optimum than in a suboptimal region, selecting the latter requires greater local posterior uncertainty. Experiments on 12 synthetic benchmarks and three real-world optimization tasks show that GUIDE-FBO remains effective across settings ranging from homogeneous to severely heterogeneous. Ablation results highlight the importance of spatially localized uncertainty intervention, while the communication analysis shows that GUIDE-FBO exchanges only compact distributional messages.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Federated Bayesian Optimization, Bayesian Optimization, Multi-Task Optimization, Gaussian Processes, Black-Box Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mwenOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GUIDE-FBO is a Federated Bayesian Optimization framework where distributed agents optimize heterogeneous black-box tasks while keeping local data private. Agents share only compact distributions over local optima (low communication); the server merges/reweights these and sends global guidance back; each agent uses it only to rescale covariance while keeping its local posterior mean. It wins on synthetic benchmarks but gains negligibly on real-world tasks.

Critical strengths

1/Sound theory, few algebraic/logical errors. 2/Broad testing on12 synthetic + 3 real tasks with ablations. 3/Strong synthetic performance over competing FBO methods.

Critical weaknesses

1/ in the real-world tasks (Activity Recognition, FedHPO), FBO methods differ only marginally (accuracy 0.8511 vs 0.8513),gains are comparable, not superior, yet the method adds heavy overhead (DPGMM fitting, optimum extraction, server clustering) while cost not justified.

2/ in the "synthetic experiments", it shows that "federation fixed at N=16 agents", no scalability or computational-cost analysis; real federations may reach hundreds of agents.

3/ for "federated privacy", it shows that "agents upload Gaussian distribution components", the issue is that no theoretical privacy guarantee, shared components could leak an agent's objective or optimal design locations.

4/ "proofs only for GUIDE-UCB", the issue is that GUIDE-NEI and GUIDE-TS lack regret/convergence guarantees despite strong results.

5/ "GUIDE-TS on real tasks", it states "beats TS only on Landmine Detection", very weak real-world performance.

6/ "Sections 3.1 vs 3.2", it shows "m = posterior paths vs agent index", the issue is that symbol reuse conflict.

7/ in "Section 3.1 vs Appendix B.2", same symbol for scalar and matrix.

Questions for authors

see the above

2 comments
AuthorsOpenReview ↗

Response to Reviewer mwen (Part 1/2)

Thank you for your time and careful assessment. We appreciate your positive evaluation of the theory and synthetic results, as well as your comments on missing experimental evidence and ambiguous notation. We have added scalability experiments and statistical comparisons, clarified the runtime and theoretical scope, and corrected the notation. Revisions are marked in blue in the manuscript.

W1 and W5: Real-world gains, GUIDE-TS, and computational cost

We added two-sided paired t-tests over 10 runs, pairing the mean performance across agents within each run. Holm correction covers 108 Level 2 synthetic comparisons and, separately, 27 original-budget real-world comparisons. GUIDE-UCB significantly outperforms UCB and CGP-UCB on 10 of 12 synthetic benchmarks each. Under the original real-world budget, neither comparison is significant; the only significant result among the 27 tests is GUIDE-TS versus FTS-DE on Landmine Detection. We acknowledge that these original-budget results provide limited evidence of a practical advantage, and have revised the manuscript to distinguish small numerical gains from statistically significant improvements.

These real-world tasks have only 2, 3, and 4 dimensions. With 80 evaluations per agent, final differences may be small, while 10 runs limit estimate precision. We therefore examined a smaller budget of 2d2d initial evaluations plus 5d5d BO rounds over 50 independent runs. Given the time available, this supplementary analysis focuses on Landmine Detection (d=2d=2, 4+10 evaluations).

GUIDE-UCB achieves mean final AUC 0.786961 (SE 0.000418), versus 0.783679 for UCB and 0.785256 for CGP-UCB. Both gains are significant after Holm correction across nine matched paired t-tests (p=7.21×10−7p=7.21\times10^{-7} and 5.00×10−45.00\times10^{-4}, respectively). This provides evidence of GUIDE-UCB's sample efficiency on Landmine Detection under a limited evaluation budget.

Revisions: Section 5.2; Appendix C.5 (protocol); D.2, Tables 12–13 (synthetic); D.4, Table 17 (original real-world); D.5, Tables 18–19 and Figure 8 (low-budget Landmine).

Under the original budget, GUIDE-TS improves over independent TS only on Landmine Detection, with an average rank of 9.67 across the three tasks. A possible explanation is that TS maximizes a sampled path: guidance changes the path distribution, but a particular draw can still attain its maximum outside the guided regions. This may make the effect less consistent across rounds than with UCB or NEI.

The mean runtime comparison across 12 Level 2 synthetic benchmarks is:

Acquisition ruleIndependent BOCGPGUIDE
UCB129.1 s289.4 s240.7 s
NEI254.7 s905.3 s534.1 s

GUIDE therefore requires more computation than independent BO but is faster than the corresponding CGP variants on average and on 9 of 12 benchmarks. These timings primarily measure optimizer overhead. In applications dominated by expensive objective evaluations, this overhead can be a small part of the total cost, making effective use of the evaluation budget the main consideration.

Revisions: Section 5.2 and Appendix D.4 for real-world results; Appendix D.1 discusses the TS sampling mechanism; Section 5.4 and Appendix H report runtime and its scope.

W2: Federation size and scalability

We added two studies on Level 2 Rastrigin, each with 10 runs and 30 initial evaluations plus 50 BO rounds per agent. The tables below report mean final regret; lower is better.

Varying the number of agents with d=10d=10:

NNUCBGUIDE-UCB
482.444357.4703
882.728959.6573
1680.477158.3082
3282.022759.9315
6482.821160.0240
12881.879560.9329

Varying dimension with N=8N=8:

ddUCBGUIDE-UCB
23.97863.5325
526.533725.0156
1082.728959.6573
20218.1751130.7259
40462.6876305.7158

On Rastrigin, GUIDE-UCB achieves lower mean regret at every tested federation size and dimension, demonstrating a consistent advantage from 4 to 128 agents and from 2 to 40 dimensions. Increasing the number of agents does not monotonically improve performance, but the advantage over independent UCB is maintained throughout the tested range.

We also measured runtime as NN increases. With clients executed sequentially, GUIDE-UCB takes 105.64 seconds at N=4N=4 and 3107.76 seconds at N=128N=128. After subtracting measured server time and dividing by NN, the remaining time stays between 24.15 and 28.49 seconds. Server aggregation and packet generation take 0.080–16.199 seconds per run, less than 0.6% of total runtime. These measurements show that server processing contributes only a small fraction of runtime in this implementation.

The full results, SEs, and convergence plots are in Appendix F.1; the runtime breakdown is in Appendix H. Both studies are summarized in Section 5.4.

AuthorsOpenReview ↗

Response to Reviewer mwen (Part 2/2)

W4: Theoretical guarantees for GUIDE-TS and GUIDE-NEI

Our theoretical contribution is the regret and query-selection analysis of GUIDE-UCB, complemented by empirical evaluation with UCB, TS, and NEI. We do not claim convergence guarantees for all three instantiations.

Appendix B.3 establishes their common mathematical foundation: FI-GP preserves the local posterior mean and defines a positive-semidefinite covariance, hence a valid auxiliary GP. The regret argument in Lemma B.1 additionally uses the UCB selection inequality. TS and NEI use different selection rules and require separate convergence analyses.

Revisions: Opening of Section 4; Section 3.3 and Appendix B.3 for the common construction; Appendices B.4–B.5 for the UCB analysis.

W6–W7: Reused symbols

Thank you for catching these ambiguities. We now use s=1,…,Ms=1,\ldots,M for posterior-path samples, separating this index from the agent indices n,mn,m. We also use Ln,tL_{n,t} for the scalar number of active local mixture components, separating it from the KK notation used for kernel matrices.

Revisions: Sections 3.1–3.2 and Appendix I for sample indexing; Section 3.1 and Appendix B.2 for component-count and kernel-matrix notation.

W3: Privacy guarantees for the exchanged components

We agree that keeping raw observations local does not by itself provide differential privacy. The exchanged distributions can still reveal information about an agent's objective or promising design locations.

Our contribution is a communication-efficient FBO mechanism that exchanges compact distributions and uses them to guide local exploration while keeping raw observations on each agent. We do not claim that the current mechanism satisfies formal differential privacy.

We have considered four possible directions for a private extension:

  • Perturb the posterior-optimum samples on each agent before fitting the mixture.
  • Privatize intermediate statistics during mixture fitting.
  • Privatize the released component parameters on each agent.
  • Privatize aggregated components on a trusted server before redistribution.

Each route would require a guarantee for the chosen privacy unit, protection of the complete released information, and accounting across rounds. A key challenge for client-side privacy is preserving useful spatial guidance after adding noise calibrated to the sensitivity of the released information. Server-side aggregation may offer a better trade-off when individual contributions and aggregation weights are controlled, although the current clustering and weighting operations do not automatically provide a 1/N1/N sensitivity reduction. This route would also require trust in the server.

These are directions for future work rather than privacy mechanisms already established in this paper. Section 6 explicitly states the absence of a formal privacy guarantee and identifies private distribution release and server aggregation as future directions.

Manuscript location: Section 3.1 defines the exchanged information; Section 6 states the privacy limitation and future directions.

Reviewer_rTDuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GUIDE-FBO, a FL Bayesian optimization method in which clients exchange compact distributions describing promising solution locations. These summaries are aggregated by a server and used to adjust local exploration while preserving each client’s predicted objective values. The contributions include a regret analysis for the UCB variant and experiments across synthetic and application-based benchmarks.

Critical strengths
  • The experiments substantiate the central design choice. The ablations show that directing exploration toward shared promising regions is substantially more effective than increasing exploration uniformly. This appears to me as a meaningful contribution.
  • Collaboration remains useful when clients have different objectives. The synthetic results show benefits over independent optimization across varying levels of heterogeneity. This strengthens the paper’s relevance to federated settings, where identical client objectives would be a restrictive assumption.
Critical weaknesses
  • The evaluation does not establish how benefits and harms are distributed across clients. Since results are averaged across clients, this can hide negative transfer for clients whose objectives differ from the majority. Helping heterogeneous clients is central to the paper, so reporting which clients improve or worsen relative to independent optimization would be important for assessing the usefulness of collaboration.
  • The application experiments demonstrate competitiveness more clearly than a meaningful practical advantage. The strongest methods have closely matched performance, while GUIDE adds computational overhead. This is acknowledged by the authors, but leaves practical payoff of the added complexity insufficiently justified.
Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer rTDu

Thank you for the time and care you put into the review. We appreciate your recognition of the value of directing exploration toward shared promising regions. Your point about averages hiding unfavorable outcomes motivated us to examine the results before averaging across agents or runs. Revisions are marked in blue in the manuscript.

W1 and the corresponding question: Which clients improve or worsen?

We added per-agent analyses on Rastrigin under both Level 2 and Level 3 heterogeneity, with 8 agents and 10 runs per setting. The two methods use the same realized tasks and initialization within each run.

First, averaging each agent index over the 10 runs, GUIDE-UCB has lower mean final regret than UCB at all eight indices under Level 2 and at six of eight under Level 3. The federation-level means decrease from 82.7289 to 59.6573 and from 96.0266 to 89.5949, respectively.

Because the tasks are regenerated across runs, these averages do not show whether every agent benefits in every run. We therefore also compare each agent separately within each run:

HeterogeneityLower regret with GUIDE-UCBHigher regret with GUIDE-UCB
Level 264/80 (80.0%)16/80 (20.0%)
Level 346/80 (57.5%)34/80 (42.5%)

GUIDE-UCB achieves lower regret in 80.0% of the comparisons at Level 2 and 57.5% at Level 3. At Level 3, it also achieves lower federation-average regret in all 10 runs, reducing the overall mean from 96.0266 to 89.5949. Thus, the gains extend to a majority of the individual comparisons, and the overall improvement is maintained even under severe heterogeneity.

Revisions: Section 5.4, “Agent-level performance,” and Appendix E.3, including per-agent means and SEs, convergence plots, the within-run comparison table, and the accompanying analysis.

W2 and the corresponding question: Practical gains and computational cost

We added two-sided paired t-tests over 10 runs, pairing the mean performance across agents within each run. Holm correction covers 108 Level 2 synthetic comparisons and, separately, 27 original-budget real-world comparisons. GUIDE-UCB significantly outperforms UCB and CGP-UCB on 10 of 12 synthetic benchmarks each. Under the original real-world budget, the only significant result among the 27 tests is GUIDE-TS versus FTS-DE on Landmine Detection. We acknowledge that these original-budget results provide limited evidence of a practical advantage, and have revised the manuscript to distinguish small numerical gains from statistically significant improvements.

These real-world tasks have only 2, 3, and 4 dimensions. With 80 evaluations per agent, final differences may be small, while 10 runs limit estimate precision. We therefore examined a smaller budget of 2d2d initial evaluations plus 5d5d BO rounds over 50 independent runs. Given the time available, this supplementary analysis focuses on Landmine Detection (d=2d=2, 4+10 evaluations).

GUIDE-UCB achieves mean final AUC 0.786961 (SE 0.000418), versus 0.783679 for UCB and 0.785256 for CGP-UCB. Both gains are significant after Holm correction across nine matched paired t-tests (p=7.21×10−7p=7.21\times10^{-7} and 5.00×10−45.00\times10^{-4}, respectively). This provides evidence of GUIDE-UCB's sample efficiency on Landmine Detection under a limited evaluation budget.

Revisions: Section 5.2; Appendix C.5 (protocol); D.2, Tables 12–13 (synthetic); D.4, Table 17 (original real-world); D.5, Tables 18–19 and Figure 8 (low-budget Landmine).

To quantify the computational trade-off, we report the following mean times per run across the 12 Level 2 synthetic benchmarks:

Acquisition ruleIndependent BOCGPGUIDE
UCB129.1 s289.4 s240.7 s
NEI254.7 s905.3 s534.1 s

GUIDE adds computation relative to independent BO, but its mean runtime is lower than that of the corresponding CGP variants. It is also faster than CGP on 9 of 12 benchmarks for each of these two acquisition rules.

The relevance of this overhead depends on the cost of evaluating the objective. These synthetic functions are inexpensive, so the timings primarily measure the optimizer itself. In the expensive black-box setting that motivates BO, an evaluation may instead require a costly simulation, physical experiment, or model-training run. When evaluation costs dominate, the additional optimizer computation can represent a small fraction of the total budget, and improved use of an evaluation budget can justify it.

The additional runtime analysis in Appendix H also shows that measured server processing accounts for less than 0.6% of total runtime from 4 to 128 agents in our serial implementation. After subtracting server time, the remaining time per agent stays between 24.15 and 28.49 seconds.

Revisions: Section 5.4 and Appendix H for the runtime results and the discussion of optimizer overhead versus objective-evaluation cost.

I thank the authors for their rebuttal and revision, which mostly address the points in my review. I read other reviewer's comments, I will interact with them during the discussion phase and update my score accordingly.

AuthorsOpenReview ↗

Response to Reviewer rTDu

Dear Reviewer,

Thank you for taking the time to read our rebuttal and revised manuscript. We’re glad that the revisions have addressed most of your concerns. Your feedback has helped us improve the paper. Thank you again for your careful review and for continuing the discussion with the other reviewers!

Best regards,
The Authors

Reviewer_ZYqJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method for federated Bayesian optimization by sharing the distributions over the local maxima of Gaussian Process posteriors. The server merges and reweights them, then sends back a subset of the hinted regions to the clients. The clients then use the respective subset to recalculate the uncertainty in the region using Federated Interventional GP. They also show that for any such bounded intervention, their method preserves the cumulative regret rate of GP-UCB. When considering a suboptimal region rather than a suggested optimal hinted region, the suboptimal region should exhibit high uncertainty. It conducts experiments on 12 synthetic benchmarks and 3 real-world optimization tasks.

Critical strengths
  • Overall, the conceptual idea of the method, identifying highly probable regions for each client, then sending those distributions to the server, the server merging and reweighting the regions, and sending back a subset of regions to the clients, where the clients upscale the uncertainty accordingly, and then running normal UCB, TS, or NEI on FI-GP is well motivated.
  • The method protects from negative transfer of information between clients through mechanisms such as only changing the uncertainty and not the guess, μ\mu, multiplying the changes by agent's own σ\sigma so the changes do nothing where the client is certain, and gradually decaying the importance with time.
  • The method is modular such that FI-GP can be combined with any acquisition rule such as UCB, TS or NEI.
  • The results across fair comparisons such as TS methods vs GUIDE-TS, UCB methods vs GUIDE-UCB and so on lead to consistent imrpovement in average rank (Table 1).
Critical weaknesses
  • The paper reports absolute values, and no standard errors are reported in the main tables, which are required for reproducibility across different seeds and for identifying signal from noise.
  • Gains in the real world optimization benchmarks (Table 2) is marginal, even at the fifth or sixth decimal place.
  • The number of clients and the dimensionality stays the same throughout the synthetic experiments, so it is difficult to know whether the method scales or generalizes well across different settings.
  • Theorem 1 only shows that the intervention does no harm, not that it helps. In the proof, the term −λtGn,t(xn∗)σn,t−1(xn∗)-\lambda_t G_{n,t}(x_n^{\ast}) \sigma_{n,t-1}(x_n^{\ast}) from Lemma B.1, which is the only term that can reflect a benefit from the guidance, is dropped. Since 1≤Sn,t(x)≤1+λt1 \le S_{n,t}(x) \le 1+\lambda_t, GUIDE-UCB is essentially GP-UCB with a bounded inflation of the confidence width, so preserving the O(TβTγn,T)\mathcal{O}(\sqrt{T\beta_T\gamma_{n,T}}) rate up to a constant is expected (as also seen in Appendix B.6). The result does not show advantage from collaboration.
  • Proposition 1 holds only when the guidance is informative, i.e. Gn,t(xn∗)≥cG_{n,t}(x_n^{\ast}) \ge c and Gn,t(xn,t)≤εGG_{n,t}(x_{n,t}) \le \varepsilon_G. However, the paper does not show when the distributions sent by the server actually satisfy these conditions, for example how they depend on the distance between the optima of different clients or on the level of task heterogeneity. So the proposition only shows that the method helps if the guidance is already good, and does not say when the guidance will be good.
Questions for authors
  • Can the authors report standard errors for Tables 1 and 2, and a significance test for the main comparisons such as GUIDE-UCB vs UCB and GUIDE-UCB vs CGP-UCB? In particular, are the real-world differences in Table 2, which appear at the fourth to sixth decimal place, statistically significant?
  • Under what conditions do the transferred distributions satisfy Gn,t(xn∗)≥cG_{n,t}(\mathbf{x}^*_n) \ge c in Proposition 1? Can this be related to the distance between the clients' optima or to the level of task heterogeneity?

Suggestions:

  • Experiments with a varying number of clients NN and higher dimensions dd would show whether the method scales, since all synthetic experiments use N=16N=16 and d=10d=10, and the real-world tasks have only 2–3 hyperparameters.
2 comments
AuthorsOpenReview ↗

Response to Reviewer ZYqJ (Part 1/2)

Thank you for your time and careful review, for recognizing the merits of our method, and for pointing out where supporting evidence was missing. We have added uncertainty estimates, statistical comparisons, and experiments beyond the original federation size and dimension. Revisions are marked in blue in the manuscript. We address the empirical concerns below and the theoretical questions in Part 2.

W1–W2 and Q1: Standard errors, statistical significance, and real-world gains

We have added one standard error (SE) for the performance estimates underlying Tables 1 and 2. The synthetic regret means and their SEs are reported in Appendix D.1–D.3, and the real-world SEs are in Appendix D.4, Table 16. Both main table captions point to these results. Since Table 1 reports average ranks, we provide SEs for the underlying regret estimates rather than the ranks themselves.

We also report SEs for the ablations and agent-level results in Appendix E, and for the hyperparameter, federation size, and dimension studies in Appendix F. The convergence plots include bands showing the mean ±1 SE.

We added two-sided paired t-tests over 10 runs, pairing the mean performance across agents within each run. Holm correction covers 108 Level 2 synthetic comparisons and, separately, 27 original-budget real-world comparisons. GUIDE-UCB significantly outperforms UCB and CGP-UCB on 10 of 12 synthetic benchmarks each. Under the original real-world budget, the only significant result among the 27 tests is GUIDE-TS versus FTS-DE on Landmine detection. We acknowledge that these original-budget results provide limited evidence of a practical advantage, and have revised the manuscript to distinguish small numerical gains from statistically significant improvements.

These real-world tasks have only 2, 3, and 4 dimensions. With 80 evaluations per agent, final differences may be small, while 10 runs limit estimate precision. We therefore examined a smaller budget of 2d2d initial evaluations plus 5d5d BO rounds over 50 independent runs. Given the time available, this supplementary analysis focuses on Landmine detection (d=2d=2, 4+10 evaluations).

GUIDE-UCB achieves mean final AUC 0.786961 (SE 0.000418), versus 0.783679 for UCB and 0.785256 for CGP-UCB. Both gains are significant after Holm correction across nine matched paired t-tests (p=7.21×10−7p=7.21\times10^{-7} and 5.00×10−45.00\times10^{-4}, respectively). This provides evidence of GUIDE-UCB's sample efficiency on Landmine detection under a limited evaluation budget.

Revisions: Section 5.2; Appendix C.5 (protocol); D.2, Tables 12–13 (synthetic); D.4, Table 17 (original real-world); D.5, Tables 18–19 and Figure 8 (low-budget Landmine detection).

W3 and the scalability suggestion: Varying the number of agents and dimension

We added two studies on Rastrigin under Level 2 heterogeneity, varying the number of agents from 4 to 128 and the dimension from 2 to 40. Each configuration uses 10 runs, 30 initial evaluations, and 50 BO rounds per agent. The tables below report mean final regret; lower is better.

Varying the number of agents with d=10d=10:

NNUCBGUIDE-UCB
482.444357.4703
882.728959.6573
1680.477158.3082
3282.022759.9315
6482.821160.0240
12881.879560.9329

Varying dimension with N=8N=8:

ddUCBGUIDE-UCB
23.97863.5325
526.533725.0156
1082.728959.6573
20218.1751130.7259
40462.6876305.7158

On Rastrigin, GUIDE-UCB achieves lower mean regret at every tested federation size and dimension, demonstrating a consistent advantage from 4 to 128 agents and from 2 to 40 dimensions. Increasing the number of agents does not monotonically improve performance, but the advantage over independent UCB is maintained throughout the tested range.

We also examined runtime as the federation grows. In our serial implementation, server processing accounts for less than 0.6% of total runtime throughout the tested range. The full timing breakdown is reported in Appendix H.

Revisions: Section 5.4; Appendix F.1 for optimization results and Appendix H for the runtime breakdown.

AuthorsOpenReview ↗

Response to Reviewer ZYqJ (Part 2/2)

W4–W5 and Q2: What the theory establishes, when guidance is informative, and the contribution of federation

Your distinction between preserving a regret rate and demonstrating a benefit from collaboration is well taken. We address it through the scope of Theorem 1, the new geometric conditions, and the self-only guidance ablation.

The role of Theorem 1. The theorem establishes that bounded guidance, including misleading guidance, preserves the leading-order GP-UCB cumulative-regret rate. It does not establish a smaller regret than independent UCB. Dropping the nonpositive term in Lemma B.1 is appropriate for this worst-case statement; the term is retained in Proposition 1 to analyze informative guidance. The revised discussion makes this distinction explicit.

When transferred components support the recipient's optimum. Appendix B.8 now connects the guidance condition to cross task optimum proximity and the actual server operations. Consider a component from sender jj that is merged and delivered to recipient nn. Let anja_{nj} be the distance between their optima, ej,te_{j,t} the distance from the uploaded center to the sender's optimum, and δmerge\delta_{\mathrm{merge}} the complete-linkage threshold. If the delivered component has weight at least w>0w>0 and covariance with minimum eigenvalue at least ν2>0\nu^2>0, then

w \exp\left[ -\frac{(a_{nj}+e_{j,t}+\sqrt{d}\delta_{\mathrm{merge}})^2}{2\nu^2} \right].$$ The merging contribution follows because complete linkage bounds the displacement of the merged center by $\sqrt{d}\delta_{\mathrm{merge}}$. Thus, for fixed retained weight and covariance scale, nearby task optima and accurate local components give a larger lower bound on guidance at the recipient's optimum. Appendix B.8 also bounds guidance at a suboptimal point: if its Mahalanobis distance from every received component is at least $R$, then $G_{n,t}(x)\leq\exp(-R^2/2)$. Together, these bounds provide explicit values of $c$ and $\epsilon_G$ for Proposition 1. When they satisfy Eq. (4), the uncertainty threshold for selecting that suboptimal point exceeds the standard GP-UCB threshold. *Revisions: Section 4, especially the discussion of Theorem 1 and Proposition 1 and Eq. (4); Appendices B.7–B.8, especially Eqs. (6)–(8).* **Does federated information add value beyond local guidance?** We evaluated a new Self-only guidance ablation under both Level 2 and Level 3 heterogeneity. It retains posterior-optimum sampling, dominant component extraction, the local FI-GP intervention, and the intervention schedule, but constructs guidance only from the agent's own component, without server aggregation or inter-agent exchange. The following table reports the Level 2 results, the Level 3 results are provided in Appendix E.2. | Benchmark | Self-only guidance | Full GUIDE-UCB | |---|---:|---:| | Ackley | 8.7819 | 8.4785 | | Rastrigin | 60.1082 | 58.3082 | | Zakharov | 22.7750 | 22.1717 | | Michalewicz | 5.9382 | 5.8629 | Full GUIDE-UCB achieves lower mean regret than Self-only guidance on all four benchmarks at Level 2 and on three of four benchmarks at Level 3. Self-only guidance is slightly better on Michalewicz at Level 3. These comparisons support the value of sharing information across agents in addition to using locally constructed guidance. *Revisions: Section 5.3, Table 3, and Appendix E, including the Self-only guidance definition and the Level 2–3 results in E.1–E.2.*

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.