{"id":"1d208674-f432-4354-88bb-963993a3c147","arxiv_id":"2411.16267","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"VDP-GIBO, a local Bayesian optimization method with virtual data points for crashed evaluations, tunes 2 to 8 parameter controllers on a coupled tank system with tens of evaluations.","lead":"This paper combines two existing Bayesian optimization ideas, local gradient-based search and virtual penalties for crashed trials, into one controller tuning algorithm. The combination tunes PI, LQI, and MPC controllers on a coupled tank system, improving hardware tracking performance by about 33% to 50%.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The crash-handling heuristic in Eq. (4) is the load-bearing assumption; no analysis or ablation demonstrates that the virtual penalty reliably makes the GP gradient point away from infeasible regions.","rationale":"The reader's weakest_assumption already identifies Eq. (4) as the soft spot, and my reading agrees. The paper's reported experiments are internally consistent and do support the narrow claim that VDP-GIBO outperforms random search on the higher-dimensional simulated cases and improves the tested hardware controllers. The load-bearing concern is not that the algorithm fails in those specific settings, but that the mechanism for handling crashes is a heuristic whose adequacy is only illustrated, not established. A focused ablation varying beta and comparing against a fixed-penalty or constrained-BO baseline would settle whether the virtual penalty is actually doing the claimed work, or whether the results depend on the particular objective, initialization, and crash thresholds. Since this is an addressable empirical question rather than a demonstrated flaw, the reader's CONDITIONAL verdict remains appropriate; I do not see grounds to accept outright or to reject.","tokens_in":12720,"tokens_out":5741,"duration_ms":61059,"concrete_test":"Construct a crash-constrained test problem where f(x) = ||x - x*||^2 and the feasible region is {x : x_1 >= 0} with x* = (0, a), so the optimum lies on the crash boundary and the objective gradient at feasible points points toward the infeasible set. Run VDP-GIBO from the same feasible initialization with beta in {0.5, 1, 3, 5, 10}, record the best feasible function value and the direction of nabla mu_Dhat(x*) after the first crash. If beta=3 does not consistently rotate the gradient away from the boundary, or if the best feasible value is substantially worse than what a fixed-large-penalty or classifier-constrained BO baseline achieves, then Eq. (4) is not a reliable crash handler and the central data-efficiency claim is limited to the reported testbed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that VDP-GIBO is data-efficient for crash-constrained controller tuning depends on crashed evaluations being converted into useful gradient information. The only mechanism for this is the virtual penalty in Eq. (4): yhat_i = max(mu_D(xhat_i), mu_D(x*)) + beta*sqrt(k_D(xhat_i)), with beta=3. The paper justifies this by a lower bound (the virtual value exceeds the posterior mean at x*) and by the single illustration in Fig. 2; there is no derivation, no sensitivity analysis, and no comparison against alternative crash-constraint mechanisms. The lower bound only ensures a positive secant slope between x* and each crash location; it does not ensure that the posterior gradient nabla mu_Dhat(x*) points away from the crash region. Moreover, after the virtual points are inserted and GP hyperparameters are re-optimized (Algorithm 1, line 14), even this secant-slope property can change. If the true objective has a steep descent toward the crash boundary and the constrained optimum lies just on the feasible side, a too-small penalty can leave the learned gradient pointing into the infeasible region; the algorithm then falls back to resetting to the best feasible evaluation (Algorithm 1, line 12) and progress can stall. The reported experiments show one favorable instance, but the data-efficiency claim is intended to generalize beyond it. This is not an internal inconsistency, but it is the least secure load-bearing condition for the paper's central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VDP-GIBO, a local Bayesian optimization method for controller tuning under crash constraints, where failed evaluations do not return a function value. The method extends the authors' earlier GIBO algorithm by adding virtual data points at crash locations with an adaptive penalty (Eq. 4), by evaluating batches that minimize the total variance of the GP posterior gradient, and by resetting to a known feasible point when an update becomes infeasible (Algorithm 1). The method is evaluated on a simulated coupled-tank system with PI, cascaded PI, MPC+EKF, and LQI controllers (2, 5, 6, and 8 tuning parameters, respectively) against a random-search baseline, and on a hardware testbed for PI and cascaded PI control. The paper reports that VDP-GIBO solves the higher-dimensional problems within tens to a few hundred evaluations and improves hardware PI tracking by about 33% and cascaded PI by about 50%.","tokens_in":13059,"tokens_out":6226,"duration_ms":59513,"significance":"If the reported results hold, VDP-GIBO would be a practically useful tool for controller tuning in settings where global Bayesian optimization is too sample-hungry and where crashed experiments are common. The paper has several strengths: simulation results are averaged over ten random seeds with randomized initializations; the same hyperparameters are used across all problems; and the hardware experiments demonstrate the crash-constraint machinery on a real system. The proposed integration of virtual data points into a local BO framework is clean and easy to combine with different acquisition procedures. However, the novelty is incremental relative to prior work by the same group (GIBO and VDP-BO), and the load-bearing crash-handling heuristic in Eq. (4) is not supported by a derivation or sensitivity analysis. The empirical evaluation compares only against random search, and the hardware results are single-run point estimates. These gaps limit the strength of the data-efficiency claim in its current form.","major_comments":[{"comment":"The virtual penalty in Eq. (4) is the only mechanism that converts crashed evaluations into gradient information, yet the paper does not establish that the posterior gradient ∇μ_𝒟hat(x*) reliably points away from the crash region. The lower bound max(μ_D(x*), μ_D(x̂_i)) only ensures a non-negative secant slope between x* and each crash location; it does not control the direction of the GP gradient, and re-optimizing hyperparameters in Algorithm 1 (line 14) can alter the posterior after the virtual points are inserted. Please add a sensitivity analysis over β (e.g., 1, 3, 10), an ablation replacing Eq. (4) with a fixed penalty, and a diagnostic showing the direction of the gradient update after a crash. Without such evidence, the data-efficiency claim for crash-constrained problems rests on a single illustrative example (Fig. 2) and on the specific β=3 value used in the experiments.","section":"4.1, Eq. (4)"},{"comment":"The hardware results are based on one run per configuration. The reported 33% and 50% improvements are point estimates from single trajectories; Fig. 7 shows no error bars or repeated experiments. Because these results are the main evidence for the practical value of the method, please report multiple runs with variability, or clearly label the results as single-session illustrations and adjust the strength of the conclusions accordingly.","section":"6, Fig. 7"},{"comment":"The paper states that the hyperparameters (β=3, L=0.25I, σ_f=0.5, μ=1) were chosen manually from initial experimentation with the PI controller and then concludes that VDP-GIBO is \"not very sensitive\" to the choice of hyperparameters. This conclusion is not supported by any sensitivity study. Please provide such a study (e.g., varying β and L over a range and reporting performance on at least one problem) or replace the claim with the more limited statement that the same hyperparameters produced good results on the tested problems.","section":"5.3, Tab. 2"},{"comment":"The only baseline is random search. While this supports the specific claim that VDP-GIBO outperforms random search on the higher-dimensional problems, it does not support the broader data-efficiency claim relative to established Bayesian optimization variants with crash handling. Please add comparisons to at least one alternative, such as global BO with a fixed penalty (as in [5]) or a constrained BO method (as in [7]), or clearly scope the contribution as \"better than random search\" rather than \"data-efficient\" in a general sense.","section":"5.3, Fig. 4"}],"minor_comments":[{"comment":"The gradient update writes \"∇μ_𝒟(x_k)\" but the loop variable is x*; please use consistent notation (e.g., ∇μ_𝒟hat(x*)).","section":"Algorithm 1, line 9"},{"comment":"The caption says \"4-(cascaded PI)\" but Tab. 1 and Section 5.2 identify the cascaded PI case as having 5 parameters; please correct the inconsistency.","section":"Fig. 4 caption"},{"comment":"The text says \"eight iterations, corresponding to 33 evaluations\" for a 2-D problem with batch size b=d+1=3; a reader would expect 8 iterations to involve at least 25 evaluations. Please clarify how the total of 33 is counted (e.g., including the initial evaluation and the per-update evaluation).","section":"Section 6, first experiment"}],"recommendation":"major_revision","confidential_remarks":"The paper is a straightforward integration of two prior works by the same group (GIBO [8] and VDP-BO [11]), with batch evaluation and a reset rule added. The empirical study is new and not fitted to the results, which is a positive feature. However, the load-bearing heuristic in Eq. (4) is not analyzed, and the experimental evidence for the headline claims is thinner than the text suggests. These issues are fixable by additional experiments and a more carefully scoped abstract, and the method itself is plausible and relevant to the journal's applied audience. I would support inviting a revised version rather than rejecting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this is a clean, incremental engineering paper that combines two existing ideas—GIBO's local gradient search and virtual data points for crashed evaluations—and tests the result on a coupled tank system both in simulation and on hardware. The specific combination is new, and the hardware results are real: roughly 33% improvement for a PI controller and 50% for a cascaded PI, within tens of evaluations. The paper is honest about its limits: it only compares against random search, and it reports the crashes that happened on hardware. If you work on BO for controller tuning, this is worth a look.\n\nWhat's genuinely good: the method is simple to describe, the same hyperparameters are used across all simulation and hardware runs, and the authors don't overclaim theory. The simulation covers 2, 4, 5, 6, and 8 parameters, and the advantage over random search is convincing in the higher-dimensional cases. The hardware section is a credible sanity check, including a run with three crashes where tuning still succeeded. The fact that GIBO and VDP are prior work from the same group is not a problem here—the new empirical evaluation stands on its own.\n\nThe soft spots are mostly about scope. Random search is the only baseline; no comparison to other crash-constraint BO methods or even a fixed-penalty variant. Hardware runs are single-shot, so the improvement percentages have no error bars. The hyperparameters were hand-chosen, and the claim that the method is insensitive to them is asserted without a sensitivity study. No code or data is released, which hurts reproducibility.\n\nThe stress-test concern about Eq. (4) is fair but not fatal. The virtual penalty is a heuristic; the lower bound only ensures a positive secant slope, not a gradient that points away from the crash region. But the paper doesn't promise a guarantee—it presents an empirical demonstration, and the experiments back that up. A revision should include a sensitivity analysis on β and maybe an ablation without virtual points.\n\nBottom line: this is a solid, useful contribution that deserves a serious referee. It needs a revision with stronger baselines and released code/data, but the central claim is supported. I'd accept it conditionally and would cite it in my own work if I were active in this area.\n\nRecommendation: send to peer review, with a request for additional baselines and reproducibility artifacts.","headline":"Useful, incremental controller-tuning paper with real hardware evidence; the crash-penalty heuristic is the main soft spot but not a reason to reject.","tokens_in":13553,"tokens_out":2653,"would_cite":true,"duration_ms":23880,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A local Bayesian optimizer that adds virtual points at crash locations tunes 2- to 8-parameter controllers in a few hundred evaluations, and improves hardware PI tracking by roughly 33% and cascaded PI by 50%.","keywords":["Bayesian optimization","controller tuning","crash constraints","local optimization","Gaussian process","virtual data points","gradient information","coupled tank system"],"falsifier":"On a one-dimensional synthetic problem with feasible set [0, 0.5] and an objective that decreases steeply toward a minimum just inside the boundary at 0.5, run VDP-GIBO with β=3 from any feasible start. If the algorithm ever proposes a second evaluation at a crash location (x > 0.5) without the virtual point's gradient pointing outward — or if the reset rule triggers at the same iterate more than once — the heuristic is insufficient in that regime. More directly, compute the angle between ∇μ_hatD(x̂_i) and the outward normal of the feasible set after inserting the virtual point; the method requires that angle to exceed 90° at every crash location.","tokens_in":87,"feed_emoji":"⚙️","tokens_out":11732,"duration_ms":301875,"temperature":0.7,"pith_summary":"Controller tuning often fails outright for some parameter settings: the closed-loop system crashes and no performance value is returned. This paper proposes VDP-GIBO, a local Bayesian optimization method that handles these crash constraints by inserting a virtual data point with an adaptive penalty at each crash location into the Gaussian process model. The virtual points steer the learned gradient away from infeasible regions, so the standard GIBO gradient-update loop can keep improving the controller. In a simulated coupled tank system the method tunes 2-, 5-, 6- and 8-parameter PI, cascaded PI, LQI and MPC controllers within tens to a few hundred evaluations, and on real hardware it improves PI tracking by about 33% and cascaded PI by about 50%.","feed_headline":"Crash-aware local Bayesian tuning cuts control error by up to half","feed_subtitle":"Virtual points at crash sites keep the search feasible, tuning 2-8 parameter controllers in a few hundred evaluations.","key_machinery":"The central object is the augmented Gaussian process posterior built from both real and virtual observations. When an evaluation at x̂_i crashes, the algorithm adds the point (x̂_i, ŷ_i) with ŷ_i = max(μ_D(x̂_i), μ_D(x*)) + β√k_D(x̂_i) and β=3, which lies above the posterior mean at the current iterate and at the crash site; this creates a ridge that tilts the GP gradient away from the infeasible region. The gradient update x* ← x* − η_k ∇μ_hatD(x_k) uses the posterior mean of the derivative, and the step size η_k is normalized by the GP lengthscale via Eq. (5). The next batch of evaluations minimizes the total trace variance of the gradient posterior (Eq. (2)–(3)), concentrating queries where the gradient estimate is most uncertain, and a reset rule returns the iterate to the best feasible point seen so far if an update lands in the crash region.","core_discovery":"On its own terms, the paper establishes that crash constraints can be handled at the GP modeling level rather than at the acquisition-function level: virtual observations with value ŷ_i = max(μ_D(x̂_i), μ_D(x*)) + β√k_D(x̂_i) modify the posterior so that the GP gradient at the current iterate points away from observed crash locations, allowing the GIBO update rule to keep making descent steps. The authors demonstrate this on five simulated test cases with deterministic crashes, showing that VDP-GIBO matches or beats random search on the easy 2D problems and clearly outperforms it on the 5-, 6-, and 8-dimensional problems within the same evaluation budget. They further validate the approach on a physical coupled tank, where tuning runs complete in 33 to 54 evaluations and improve closed-loop tracking error by roughly one-third (PI) to one-half (cascaded PI).","pith_inferences":["The virtual-penalty heuristic could be stress-tested on synthetic functions with known steep gradients toward the feasible boundary; one would predict that β=3 sometimes fails to turn the GP gradient outward, which would indicate when a larger β or a different penalty shape is needed.","The paper assumes deterministic crashes; if real systems can crash stochastically, the reset rule no longer guarantees a feasible iterate, and the algorithm would likely need a probabilistic feasibility model to remain safe.","The same virtual-data strategy could be combined with global acquisition functions like expected improvement to handle crashes in non-local search; the paper only demonstrates it inside the local GIBO loop, so the transfer remains untested."],"forward_implications":["Because crash handling is confined to the GP update, the virtual-data mechanism can in principle be dropped into other BO variants that use a Gaussian process, without redesigning the acquisition function.","Controller tuning with many parameters becomes feasible: the paper demonstrates 8-parameter LQI tuning in about 200 evaluations, a regime where random search makes little progress.","Since the search is local and the penalty discourages revisiting crashes, tuning sessions on physical systems will rarely trigger emergency stops; in the hardware experiments almost all evaluations stayed feasible.","The fixed hyperparameter set (including β=3, the kernel lengthscale and the step-size schedule) worked across all test cases, indicating the method is not strongly sensitive to these choices.","As long as crashes are deterministic, the reset rule guarantees a feasible parameterization after every update, so the tuning process never ends in an infeasible state."],"supporting_citations":[{"why":"Provides the base GIBO algorithm: learning the objective gradient from the GP derivative posterior and choosing experiments to minimize total variance.","marker":"[8]"},{"why":"Introduces virtual data points with adaptive penalties for crashed evaluations; VDP-GIBO adopts Eq. (4) from this work.","marker":"[11]"},{"why":"Gives the Gaussian-process-with-failures formulation and convergence proof that defines the crash-constrained optimization problem.","marker":"[4]"},{"why":"Presents an alternative combined-GP approach for robot learning with crash constraints, the main prior method the paper contrasts with.","marker":"[7]"},{"why":"Shows minimizing total gradient variance is equivalent to minimizing worst-case gradient estimation error, supporting the batch selection rule.","marker":"[10]"},{"why":"Supplies the coupled tank model and linearization used for all simulation test cases and the hardware testbed.","marker":"[31]"}],"fun_headline_variants":["Local BO tames crashes to halve control error","Crash-aware Bayesian optimizer halves control error","GP-level crash handling boosts controller tuning","Virtual crash points steer Bayesian tuning to success"],"cache_read_input_tokens":15616,"weakest_assumption_plain":"The whole crash-handling scheme rests on the heuristic that a virtual point at the crash location with value max(μ_D(x̂_i), μ_D(x*)) + 3√k_D(x̂_i) is large enough to make the learned gradient point away from the infeasible region; if the real objective's slope toward the crash region is steeper than this penalty, the algorithm can keep proposing infeasible points and stall.","fun_headline_variants_meta":{"raw":{"variants":["Local BO tames crashes to halve control error","Crash-aware Bayesian optimizer halves control error","GP-level crash handling boosts controller tuning","Virtual crash points steer Bayesian tuning to success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000127,"raw_usage":{"total_tokens":1047,"prompt_tokens":811,"completion_tokens":236,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":180}},"tokens_in":427,"tokens_out":236,"duration_ms":10747,"temperature":1.0,"reasoning_tokens":180,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:18:31.998969+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a one-dimensional synthetic problem with feasible set [0, 0.5] and an objective that decreases steeply toward a minimum just inside the boundary at 0.5, run VDP-GIBO with β=3 from any feasible start. If the algorithm ever proposes a second evaluation at a crash location (x > 0.5) without the virtual point's gradient pointing outward — or if the reset rule triggers at the same iterate more than once — the heuristic is insufficient in that regime. More directly, compute the angle between ∇μ_hatD(x̂_i) and the outward normal of the feasible set after inserting the virtual point; the method requires that angle to exceed 90° at every crash location.","supporting_citations":[{"cited_title":"Local policy search with Bayesian optimization","cited_arxiv_id":null,"evidence_quote":"Provides the base GIBO algorithm: learning the objective gradient from the GP derivative posterior and choosing experiments to minimize total variance."},{"cited_title":"Gaus- sian process optimization with failures: classification and convergence proof","cited_arxiv_id":null,"evidence_quote":"Gives the Gaussian-process-with-failures formulation and convergence proof that defines the crash-constrained optimization problem."},{"cited_title":"Robot learn- ing with crash constraints.IEEE Robotics and Automation Letters, 6(2):1439–1446, 2021","cited_arxiv_id":null,"evidence_quote":"Presents an alternative combined-GP approach for robot learning with crash constraints, the main prior method the paper contrasts with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows minimizing total gradient variance is equivalent to minimizing worst-case gradient estimation error, supporting the batch selection rule."},{"cited_title":"Evaluation of data enhanced model predictive control for a coupled tank system","cited_arxiv_id":null,"evidence_quote":"Supplies the coupled tank model and linearization used for all simulation test cases and the hardware testbed."}],"review_version":1}