{"id":"2e07cdcd-3737-4700-8326-c3a39fe22d03","arxiv_id":"2412.00419","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"AutoPQ converts point forecasts into quantile forecasts with a conditional invertible neural network and automatically tunes both the uncertainty width and the point forecaster, beating six baseline methods on six smart grid datasets while reporting electricity consumption.","lead":"AutoPQ is an automated system that turns ordinary point forecasts into quantile forecasts using a conditional invertible neural network, then tunes the uncertainty and the forecasting model automatically. It also reports the electricity cost of the tuning process, making the environmental trade-off of better forecasts explicit.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim depends on untested cINN calibration: Section 3.1's 'equivalence of uncertainty in both spaces' is asserted by citation, and sigma is tuned on validation CRPS without a test-set calibration check.","rationale":"The reader's weakest assumption identifies the cINN calibration premise, and my reading agrees: this is the single most load-bearing condition for the central claim. The paper's pipeline can produce low CRPS from an arbitrary point forecast only if the latent-space neighborhood actually represents the predictive uncertainty. The authors support this by citing prior work and by tuning sigma on validation CRPS, but they never demonstrate calibration on the test distribution. Since CRPS is a proper scoring rule, a miscalibrated but sharp forecast can still beat calibrated baselines; the reported improvements in Tables 2 and 3 therefore do not by themselves establish that the resulting quantiles are reliable for smart grid decisions. The proposed PIT/coverage check is a direct, inexpensive test that would settle this concern. If the check passes, the central claim is substantially stronger; if it fails, the performance advantage is called into question regardless of CRPS. For these reasons, the conditional-acceptance verdict is appropriate, with calibration verification as an explicit condition.","tokens_in":32594,"tokens_out":3529,"duration_ms":35677,"concrete_test":"For each of the six test sets, estimate the predictive CDF F_t from AutoPQ-advanced's quantile samples and compute PIT values p_t = F_t(y_t) at the realized observations. Assess uniformity of {p_t} via a Kolmogorov-Smirnov test and a reliability diagram; also compare the empirical coverage of the nominal 50% and 90% PIs to their nominal levels. If the KS p-value is below 0.05 or the 90% PI coverage deviates by more than 3 percentage points on any dataset, the calibration assumption fails and the CRPS comparison is insufficient.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"AutoPQ-advanced's CRPS advantage over six baselines (Table 3) rests on the assumption that the pre-trained cINN is a calibrated conditional bijection: the point forecast y_hat is mapped to z_hat = g(y_hat), samples z ~ N(z_hat, sigma^2 I) are drawn, and the backward map yields the predictive quantiles. Footnote 4 of Section 3.1 asserts 'the equivalence of uncertainty in both spaces' by citing [19], with no derivation or empirical verification in this paper. The sampling variance sigma is then tuned on validation CRPS (Eq. 5), which rewards sharpness and CRPS but does not guarantee calibration. If the cINN's learned conditional distribution in latent space is not the true conditional distribution of the data (finite sample, non-Gaussian z, bias in g(y_hat)), the pushforward through g^{-1} is not the true predictive distribution. The paper reports no PIT uniformity or empirical-coverage check on the hold-out test sets; Section 5.4 lists 'desirable probabilistic properties' as an open question, confirming calibration was not verified. A sharp but miscalibrated forecast can achieve lower CRPS than a calibrated baseline, so the reported 9.1-27.3% improvements may not reflect trustworthy quantile forecasts for smart grid decisions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces AutoPQ, an automated probabilistic forecasting pipeline that wraps an arbitrary point forecast with a conditional invertible neural network (cINN). The point forecast is mapped into a Gaussian latent space, samples are drawn around that latent representation with variance lambda_q, and the inverse map turns these samples into quantile forecasts. Two configurations are proposed: AutoPQ-default optimizes only the sampling variance for each default point forecaster, while AutoPQ-advanced jointly optimizes the point forecaster's hyperparameters and the sampling variance via a nested evolutionary/Bayesian optimization loop with successive halving. The evaluation on six smart-grid-related datasets compares CRPS against six baseline methods and reports average CRPS improvements of 9.1%–27.3% over the baselines and 5.0% over AutoPQ-default, together with an ablation study and measurements of electricity consumption and monetary cost.","tokens_in":32885,"tokens_out":8666,"duration_ms":91765,"significance":"If the empirical claims hold, AutoPQ is a practically useful contribution: it brings AutoML-style selection and hyperparameter optimization to probabilistic forecasting from point forecasts, it is accompanied by a public implementation, and its explicit reporting of electricity consumption is a welcome step for sustainable ML. The ablation study is a strength, particularly the comparison of the nested optimization structure (Algorithm 1 vs. Algorithm 3), the prior-knowledge initialization for Bayesian optimization, and the successive-halving pruning analysis. The central limitation is that the cINN-based predictive distribution is never directly checked for calibration, so the paper currently supports a claim about CRPS optimization but not yet a fully supported claim about unbiased and accurate uncertainty quantification.","major_comments":[{"comment":"The central probabilistic claim rests on an unverified distributional equivalence. The method maps a point forecast to z_hat = g(y_hat), samples z ~ N(z_hat, sigma^2 I), and maps back through g^{-1}; footnote 4 justifies this by 'the equivalence of uncertainty in both spaces' with a citation to [19], but no derivation or test-set calibration check is provided. The cINN is trained to map the realization space to a Gaussian latent space, while at inference it is evaluated at point forecasts, which are not draws from the predictive distribution of the target; this is a distribution shift that the cited equivalence would need to cover. Tuning sigma on validation CRPS (Eq. 5) does not imply calibration, since CRPS can favor a sharp but miscalibrated forecast over a calibrated but less sharp one in finite samples. The paper reports no probability integral transform histograms, empirical coverage rates, or reliability diagrams in Section 4, and Section 5.4 explicitly leaves 'desirable probabilistic properties' as an open question. Because the title, abstract, and smart-grid motivation are about producing trustworthy quantile forecasts, the authors should either verify calibration of the final predictive distributions on the test sets or explicitly reframe the contribution as CRPS optimization without the stronger claim of unbiased uncertainty quantification.","section":"Section 3.1, footnote 4; Section 5.4"},{"comment":"The headline comparison to baselines is not a controlled one. AutoPQ-advanced receives an 8-hour CASH budget (Section 3.4.1, Algorithm 1) and jointly optimizes the point forecaster's hyperparameters and the sampling variance, whereas DeepAR, QRNNs, NNQF, and the Gaussian/Empirical/Conformal PI methods are evaluated with default or standard settings; only the best base point forecast is selected for the PI benchmarks, without tuning the underlying forecaster. The reported 9.1%–27.3% average CRPS improvements therefore conflate the proposed uncertainty-quantification mechanism with the benefit of hyperparameter optimization. The authors should compare against baselines that receive a comparable tuning budget, or make AutoPQ-default (fixed default point forecasts, only sigma optimized) the primary fair comparison and explicitly quantify the additional gain attributable to HPO.","section":"Section 4.1.1, Table 3"},{"comment":"The significance claims are based on 42 one-tailed t-tests with five runs each, with no correction for multiple testing. At alpha = 0.05 one would expect roughly two false positives among 42 tests, so the statement that improvements are 'significant in 38 out of 42 tests' is not a valid family-wise or false-discovery-rate statement. The authors should report adjusted p-values (for example, Benjamini-Hochberg) or clearly state which comparisons are confirmatory versus exploratory, and they should supplement the five-run t-tests with effect sizes and confidence intervals. This is particularly relevant for the non-significant entries in Table 3, such as QRNNs on PV at 0.0% and Empirical/Conformal PIs on Mobility, which the current presentation tends to obscure.","section":"Section 4.1.2, Table 3"},{"comment":"For four of the six datasets (Load-GCP, Mobility, Price, PV), the benchmark and AutoPQ-default numbers are taken from the authors' prior papers [19, 59], while the AutoPQ-advanced numbers are new in this manuscript. This creates a reproducibility gap: the reader cannot verify that the same cINN checkpoint, normalization, train/validation/test splits, quantile post-processing, and CRPS implementation were used for the old and new numbers. The inconsistency between footnote 11 (which attributes the results to [59]) and the Table 2 caption (which says they 'originate from [19]') makes this harder to resolve. The authors should either rerun the full benchmark suite with the released code or provide an explicit compatibility statement and the exact artifacts (splits, cINN weights, evaluation code) for all six datasets.","section":"Section 4.1.1, footnotes 11 and 13; Table 2 caption"}],"minor_comments":[{"comment":"The sentence 'a point forecast can be interpreted as a sample of the random variable Y' is conceptually imprecise: a point forecast is a conditional summary of the predictive distribution, not a random draw from it. This matters because the method's validity depends on how the point forecast is embedded in the cINN's latent space.","section":"Section 3.1"},{"comment":"The caption says the benchmark and AutoPQ-default results for Load-GCP, Mobility, Price, and PV 'originate from [19]', while Section 4.1.1 footnotes 11 and 13 attribute these results to [59]. Please align the references.","section":"Section 4.1.1, Table 2 caption"},{"comment":"The log-normal prior assumption for the optimal sampling hyperparameter is justified in the text as 'valid, as demonstrated in Figure 3a' with a single dataset/model (MLP on Load-BW). This should be presented as a heuristic or supported with evidence across multiple datasets, since the prior is used to initialize the inner-loop optimizer in AutoPQ-advanced.","section":"Section 3.4.1, Figure 3"},{"comment":"The electricity consumption accounting should state its system boundary explicitly: the reported kWh cover the AutoPQ design runs, but not cINN pretraining, data preprocessing, or baseline training. This is important for interpreting the 'electricity consumption required for performance improvements' claim.","section":"Section 4.3"},{"comment":"There is a typo in the second paragraph: 'a forecasting method's configuration hat delivers' should read 'that delivers'.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is built directly on the authors' prior cINN method [19] and reuses results from their earlier paper [59] for four of six datasets. The editor may wish to confirm that the novelty increment (Automated sigma optimization, joint HPO, successive halving, and energy accounting) is sufficient for the target venue, and that the underlying cINN weights and evaluation pipeline are made fully available. The calibration issue in Section 3.1 is the main scientific risk: without a test-set calibration check, the empirical CRPS gains are not yet evidence for reliable quantile forecasts."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read for you. AutoPQ automates turning point forecasts into quantile forecasts: a pre-trained cINN maps a point forecast into a Gaussian latent space, then samples around that latent representation and maps back. The cINN idea is the authors' own prior work; what's new is the automated design: nested hyperparameter optimization that tunes the point forecaster and the sampling variance together, successive-halving method selection, and reporting electricity consumption for the whole design process. That's a useful engineering contribution, and the evaluation is more thorough than most: six datasets, nine point forecasters, five runs each, ablation of the HPO components, and a GitHub implementation.\n\nWhere I'd be careful: the paper's headline CRPS gains (9-27% over the six benchmarks, 5% over AutoPQ-default) are real as reported, but the comparison is not fully clean. Four of the six datasets' baseline and default numbers are inherited from the group's earlier papers, the benchmarks are not hyperparameter-tuned, and there's no multiple-testing correction across the 42 significance tests. None of that is fatal; it just weakens the 'state-of-the-art' framing. A missing AutoGluon-TimeSeries baseline is also worth asking about, since it's the closest AutoML competitor.\n\nThe bigger conceptual gap is calibration. The whole method rests on the claim that the cINN's latent-space neighborhood corresponds to the true predictive uncertainty; that equivalence is asserted by citation to [19], and the sampling variance is tuned on validation CRPS without any PIT or empirical-coverage check on the test sets. The authors themselves list 'desirable probabilistic properties' as an open question in Section 5.4, which is honest but confirms the gap. A sharp but miscalibrated forecast can win on CRPS, so the energy-aware optimization story is solid while the 'trustworthy quantiles for smart grid decisions' story is not yet backed up.\n\nWho gets value: practitioners who want a turnkey AutoML system for probabilistic energy forecasting, and anyone working on AutoML for time series. The energy accounting is a nice, reproducible touch — they report kWh and cloud cost, and the ablation of prior-knowledge Bayesian optimization is carefully done.\n\nMy call: send it to peer review. The calibration check should be requested, and the inherited baseline numbers should be re-run or clearly qualified, but the system is real, the code is out, and the ablations show the design choices matter. It's a solid subfield contribution, not a paradigm shift.","headline":"AutoPQ is a solid, honest engineering extension: the cINN conversion is prior work, but the nested HPO, successive halving, and energy accounting are new; the main weakness is that calibration is asserted by citation rather than checked.","tokens_in":33395,"tokens_out":2547,"would_cite":true,"duration_ms":25018,"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":"One trained invertible network turns any point forecast into calibrated quantiles.","keywords":["probabilistic forecasting","quantile estimation","conditional invertible neural network","automated machine learning","hyperparameter optimization","smart grid","electricity consumption","uncertainty quantification"],"falsifier":"Compute the empirical coverage of the 90 percent prediction interval produced by AutoPQ on the held-out test sets of all six datasets and compare it with the nominal 90 percent (equivalently, inspect the probability integral transform histogram). If coverage deviates substantially on a dataset where the CRPS was nevertheless minimized, while a simple empirical- or conformal-residual benchmark achieves nominal coverage, the equivalence-of-uncertainty claim is falsified for that data. A second check: replace the selected point forecaster with a deliberately biased one and see whether the output quantiles move; stable quantiles would indicate the cINN captures the process's own uncertainty, while strongly shifting quantiles would show the uncertainty is an artifact of the point forecast.","tokens_in":32420,"feed_emoji":"⚡","tokens_out":17484,"duration_ms":142835,"temperature":0.7,"pith_summary":"AutoPQ claims to close a practical gap: most forecasting models used in smart-grid operations produce only point forecasts, while the decisions built on them need quantiles. The paper's solution attaches any point forecasting method to a conditional invertible neural network (cINN) that has been trained once on historical data to map between the data space and a Gaussian latent space; a point forecast is projected into that space, sampled around with a tunable width, and mapped back to produce quantile forecasts without assuming any distribution. AutoPQ automates the design choices — the sampling width, the choice of point forecaster among nine methods, and their hyperparameters — in two configurations with different compute budgets, and it reports the electricity each configuration consumes. On six datasets the advanced configuration improves the Continuous Ranked Probability Score by 9.1–30.8 percent on average over six baseline probabilistic forecasting methods and by 5.0 percent over AutoPQ-default, at a measured cost of 8.73 kWh per run.","feed_headline":"AutoPQ beats six baselines by 9.1 to 30.8 percent","feed_subtitle":"An invertible neural network maps forecasts into a Gaussian space and back, automating tuning and reporting energy use.","key_machinery":"The central object is the conditional invertible neural network (cINN), a bijection $g: \\mathcal{Y} \\to \\mathcal{Z}$ built from conditional-affine coupling blocks and trained to map the unknown distribution of the target time series into a multi-dimensional Gaussian latent space while conditioning on lag, cyclic, and exogenous features. Given a point forecast, the pipeline passes it forward through the cINN, samples the latent neighbourhood with variance $\\lambda_q = \\sigma$, and passes the samples backward to obtain quantiles; the sampling variance is the single knob that trades sharpness against coverage. Around this object AutoPQ builds an automated design loop: its default configuration optimizes only $\\sigma$ against the CRPS using Bayesian optimization with tree-structured Parzen estimators, and its advanced configuration adds a joint hyperparameter search in which an asynchronous evolutionary algorithm proposes point-forecaster configurations in an outer loop while the Bayesian inner loop, initialized with a log-normal prior over $\\sigma$ derived from previously evaluated configurations, finds the best sampling width for each trained model, with successive halving pruning whole forecasting methods. The nested structure is load-bearing because it evaluates several $\\sigma$ values per trained point model, exploiting the paper's measured asymmetry that quantile generation is orders of magnitude cheaper than model training.","core_discovery":"The paper's central claim is that one trained cINN is a universal quantile generator: because the network is a conditional bijection between the data space and a known Gaussian latent space, the authors assert that uncertainty in one space is equivalent to uncertainty in the other, so sampling in the latent neighbourhood of a point forecast's representation and passing the samples backward yields the full predictive distribution. AutoPQ then treats the sampling variance $\\sigma$ as a task-dependent hyperparameter, optimizes it against the CRPS on validation data, selects among nine point forecasting methods from the statistical, machine-learning, and deep-learning families, and jointly tunes their hyperparameters. The claimed result of this automation is an average 5.0 percent CRPS improvement over AutoPQ-default, average improvements of 9.1–30.8 percent over six baseline probabilistic forecasting methods, significance in 38 of 42 comparisons, and a measured electricity cost that lets users trade forecast quality against energy consumption.","pith_inferences":["If the calibration claim holds across data sets, the same trained cINN should transfer to new point forecasters — and possibly to new data sets with similar distributions — without retraining, which would amortize its training energy across many deployments; this reuse claim goes beyond the paper's evaluation.","A direct calibration check (empirical coverage or PIT histogram on the test period) would test the equivalence-of-uncertainty premise more decisively than CRPS tuning alone, since CRPS rewards sharpness and calibration jointly and can mask miscalibration; this is a cheap diagnostic the paper does not run.","Because the paper's ablation shows that default-configuration performance strongly predicts which forecasters win after full hyperparameter optimization, a simpler rule — pick the best default forecaster and tune only $\\sigma$ — may capture most of the advanced configuration's advantage at the default configuration's energy cost.","The paper's own limitation discussion notes that feature selection stays manual and that CRPS is a proxy for what applications actually need; replacing the proxy with the direct decision cost (forecast value) is its stated future work, and the same nested-search argument would likely accelerate that search too."],"forward_implications":["Any existing point forecaster — from exponential smoothing to a transformer — can be upgraded to a full probabilistic forecast by one trained cINN, with no distributional assumption and no uncertainty model retrained per forecaster.","Because the sampling variance $\\sigma$ is a single tunable knob that trades sharpness against coverage, optimizing it on validation CRPS lets the same pipeline be re-targeted to different decision costs by changing one hyperparameter.","The advanced configuration's 5.0 percent average CRPS gain over the default is reported to add 2.9 kWh and about 23 US dollars per run in resource-aware settings, while a full advanced run consumes 8.73 kWh versus 0.57 kWh for the default.","The nested two-loop search locates good sampling widths early in the budget, which matters because successive halving prunes methods on early performance; the prior-knowledge initialization cuts the trials needed for the inner loop substantially at equal validation CRPS.","Dropping the three statistical methods, halving the time budget to 4 h, and using AutoPQ-default as the initial pruning round reduces the advanced configuration's energy use by about 60 percent, to 3.47 kWh."],"supporting_citations":[{"why":"Conformal prediction intervals for multi-step time series; one of the six point-forecast-based baselines AutoPQ-advanced must beat.","marker":"[17]"},{"why":"Introduces the cINN-based quantile generation mechanism AutoPQ automates, including the asserted equivalence of uncertainty between data and latent spaces, and supplies benchmark results for four of the six datasets.","marker":"[19]"},{"why":"Provides the reporting standard for electricity consumption that grounds the paper's sustainability evaluation.","marker":"[25]"},{"why":"DeepAR, a direct probabilistic forecasting benchmark that AutoPQ-advanced must outperform.","marker":"[26]"},{"why":"The conditional invertible neural network architecture for time series on which AutoPQ's bijection is built.","marker":"[28]"},{"why":"Supplies the conditional-affine coupling blocks that let the bijection condition on exogenous features.","marker":"[55]"},{"why":"Prior loss-customized hyperparameter optimization of the sampling variance; source of AutoPQ-default results for four datasets.","marker":"[59]"},{"why":"Supplies the inner-loop Bayesian optimization with tree-structured Parzen estimators used to tune the sampling variance.","marker":"[60]"},{"why":"Supplies the asynchronous evolutionary algorithm used as the outer-loop trial generator for point-forecaster hyperparameters.","marker":"[61]"},{"why":"Defines the calibration-sharpness rationale and the CRPS metric used for both tuning and evaluation.","marker":"[67]"}],"fun_headline_variants":["AutoPQ automates probabilistic forecasts with a single invertible network","One cINN to quantile them all: AutoPQ automates smart grid forecasting","AutoPQ: greener probabilistic forecasting with automated tuning","AutoPQ cuts forecasting energy use while boosting accuracy","AutoPQ: 9-31% better probabilistic forecasts, automatically"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the cINN trained on historical data is a calibrated conditional bijection, so that sampling around a point forecast's latent representation in Gaussian space and mapping back produces quantiles that are genuinely calibrated for the test period; the paper relies on the asserted 'equivalence of uncertainty in both spaces' and tunes the sampling width on validation CRPS rather than checking calibration directly.","fun_headline_variants_meta":{"raw":{"variants":["AutoPQ automates probabilistic forecasts with a single invertible network","One cINN to quantile them all: AutoPQ automates smart grid forecasting","AutoPQ: greener probabilistic forecasting with automated tuning","AutoPQ cuts forecasting energy use while boosting accuracy","AutoPQ: 9-31% better probabilistic forecasts, automatically"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000662,"raw_usage":{"total_tokens":3027,"prompt_tokens":952,"completion_tokens":2075,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":1999}},"tokens_in":568,"tokens_out":2075,"duration_ms":14362,"temperature":1.0,"reasoning_tokens":1999,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:24:38.531957+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the empirical coverage of the 90 percent prediction interval produced by AutoPQ on the held-out test sets of all six datasets and compare it with the nominal 90 percent (equivalently, inspect the probability integral transform histogram). If coverage deviates substantially on a dataset where the CRPS was nevertheless minimized, while a simple empirical- or conformal-residual benchmark achieves nominal coverage, the equivalence-of-uncertainty claim is falsified for that data. A second check: replace the selected point forecaster with a deliberately biased one and see whether the output quantiles move; stable quantiles would indicate the cINN captures the process's own uncertainty, while strongly shifting quantiles would show the uncertainty is an artifact of the point forecast.","supporting_citations":[{"cited_title":"Conformal time series forecasting,","cited_arxiv_id":null,"evidence_quote":"Conformal prediction intervals for multi-step time series; one of the six point-forecast-based baselines AutoPQ-advanced must beat."},{"cited_title":"Generating probabilis- tic forecasts from arbitrary point forecasts using a conditional invertible neural network,","cited_arxiv_id":null,"evidence_quote":"Introduces the cINN-based quantile generation mechanism AutoPQ automates, including the asserted equivalence of uncertainty between data and latent spaces, and supplies benchmark results for four of the six datasets."},{"cited_title":"Reporting electricity consumption is essen- tial for sustainable AI,","cited_arxiv_id":null,"evidence_quote":"Provides the reporting standard for electricity consumption that grounds the paper's sustainability evaluation."},{"cited_title":"Controlling non-stationarity and periodicities in time series generation using condi- tional invertible neural networks,","cited_arxiv_id":null,"evidence_quote":"The conditional invertible neural network architecture for time series on which AutoPQ's bijection is built."},{"cited_title":"Ardizzone, C","cited_arxiv_id":null,"evidence_quote":"Supplies the conditional-affine coupling blocks that let the bijection condition on exogenous features."},{"cited_title":"Loss- customised probabilistic energy time series fore- casts using automated hyperparameter optimisation,","cited_arxiv_id":null,"evidence_quote":"Prior loss-customized hyperparameter optimization of the sampling variance; source of AutoPQ-default results for four datasets."},{"cited_title":"Making a sci- ence of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures,","cited_arxiv_id":null,"evidence_quote":"Supplies the inner-loop Bayesian optimization with tree-structured Parzen estimators used to tune the sampling variance."},{"cited_title":"Massively parallel genetic op- timization through asynchronous propagation of populations,","cited_arxiv_id":null,"evidence_quote":"Supplies the asynchronous evolutionary algorithm used as the outer-loop trial generator for point-forecaster hyperparameters."}],"review_version":1}