Sample-Efficient "Clustering and Conquer" Procedures for Parallel Large-Scale Ranking and Selection

arXiv:2402.02196 · stat.ME, cs.LG · Submitted 2024-02-03 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Sample-Efficient "Clustering and Conquer" Procedures for Parallel Large-Scale Ranking and Selection".

Jane: The gist The work aims to improve sample efficiency of parallel large-scale ranking and selection (R&S) problems by leveraging correlation information.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we’re looking at this paper called "Sample-Efficient 'Clustering and Conquer' Procedures for Parallel Large-Scale Ranking and Selection." It sounds like they’re trying to make parallel ranking problems way more sample efficient by using correlation information.

Jane: Yeah, so when you think about large-scale ranking, it often feels like you have to do a ton of work just because every alternative is treated as completely independent, and this paper suggests we can use that shared information.

Lu: Exactly. The core idea here is modifying the traditional divide and conquer framework into a clustering and conquer strategy by adding this correlation-based step, which they call the Parallel Correlation Clustering and Conquer procedure.

Meng: So how does this actually help us when we’re trying to run these simulations at scale? What’s the mechanism behind that clustering?

Tom: Well, they cluster alternatives based on their correlation and then assign all alternatives within the same cluster to a single processor instead of distributing them randomly. The intuition is that if you have highly correlated alternatives grouped together, their fluctuations tend to cancel each other out in the same direction.

Jane: That cancellation effect sounds like something we see when we try to reduce variance in simulations, which is what they’re aiming for when they talk about stochastic fluctuations aligning.

Lu: They are building on ideas similar to the CRN technique, where you artificially introduce positive correlation just to reduce variance and speed up pairwise comparisons. In this paper, they focus specifically on positive correlations between alternatives because that's what helps them cancel out those negative random fluctuations.

Meng: So it’s not about finding a perfect correlation estimate, which sounds hard to get in practice, right? They say it doesn't require highly accurate estimation or precise clustering for this to work.

Tom: That’s a big part of the improvement they highlight. They show that you can integrate this with existing fixed-precision and fixed-budget R andS procedures while still achieving optimal sample complexity.

Jane: What does optimal sample complexity mean in the context of ranking? Are we talking about needing fewer total simulations to get the same quality result?

Lu: They prove that for both the fixed-precision and fixed-budget formulations, P3C is sample optimal as the number of alternatives p grows under Assumption one <ref:2402.02196#pg1,fixed-precision and fixed-budget>. Specifically, Proposition one shows that for P3C-KN and P3C-KT, the expected number of required samples NP3C−KN is O(p) and E

NP3C−KT: is O(p) as p goes to infinity in the fixed-precision formulation <ref:2402.02196#pg1>.

Title and authors: Meng: O of p? So they’re claiming that no matter how big the problem gets, you only need a number of samples that grows linearly with the number of alternatives. That sounds pretty good for deployment costs.

Tom: It’s promising because it directly addresses the sample efficiency side of things, moving away from methods that discard shared information between alternatives.

Jane: So they are tackling the fundamental assumption in large-scale R andS that all alternatives are independent, which is what allows them to leverage correlation information for improvement.

Lu: And they introduce a new parallel clustering algorithm called AC+ which is tailored for large-scale settings, allowing them to partition all alternatives into several clusters based on those initial samples from Stage zero.

Meng: What happens after you have these clusters? Does the process just stop there, or does it move into the actual ranking part?

Tom: After clustering, Stage two is where they perform R andS within each cluster to select the local best <ref:2402.02196#pg1>. Then, if there are still alternatives left in contention—if the number of remaining alternatives in contention satisfies less than or equal to pm—they perform R andS to select the global optimal alternative.

Jane: So it’s a multi-stage process where you first group things up, then you sort them locally, and only bring back what’s necessary for the final decision.

Lu: And they handle the remaining contenders by repeating Stage one and Stage two until that condition is met, which keeps the procedure robust under both fixed-precision and fixed-budget constraints <ref:2402.02196#pg1,fixed-precision and fixed-budget>.

Meng: I wonder about how this scales when we actually run it on real traffic, because I’ve seen a lot of these theoretical results that don't translate well to production environments.

Tom: That’s where the numerical experiments come in. In the fixed-precision R andS setting for drug discovery, P3C-KN and P3C-KT consistently outperform other algorithms, showing linear growth in sample size with respect to p.

Jane: Linear growth sounds much better than exponential growth which you see in some of these traditional methods when the problem size gets large.

Lu: And they noticed something interesting in the KT case—the slope of KT visibly decreases after applying P3C, suggesting that P3C can further reduce its leading constant and improve finite-sample performance.

Title and authors: Meng: If it reduces the constant term, that means for a fixed budget or precision target, you actually get a better outcome with fewer total samples used in practice. That matters for deployment where resources are tight.

Tom: Moving to the fixed-budget setting, they looked at Neural Architecture Search problems and showed that P3C-OCBA and P3C-CBA improve PCStrad by more than an order of magnitude compared to FBKT.

Jane: An order of magnitude is a big jump, so if you’re dealing with huge search spaces in architecture design, that could make a real difference in how much data you need to collect.

Lu: The PCCAC+ algorithm shows that as the sample size N goes to infinity with p fixed, the lower bound of PCCAC+ converges to one minus k times one minus one over k times p sigma, which approaches one as ps increases <ref:2402.02196#pg1>.

Meng: So they’re showing a theoretical limit on how good this procedure can get in the long run for fixed sample sizes versus when you increase the number of alternatives.

Tom: It confirms that correlation-based clustering is an effective way to improve sample efficiency in large-scale R andS problems. The P3C procedure is presented as a ready-to-use tool that works without needing excessively high clustering accuracy, it only needs to identify which alternatives are highly correlated.

Jane: So the main message here is that we don’t need perfect correlation; we just need to identify those strong positive correlations between alternatives to get significant performance gains.

Lu: And they also developed a novel gradient-based framework to look at the "mean–correlation interaction" between mean and correlation information. Theorem one establishes that the PCS of an alternative tau is differentiable with respect to mu, where mu is the vector of means, and that partial derivative of PCS(tau) over mu tau is positive <ref:2402.02196#pg1>.

Tom: That gradient analysis framework is interesting because it shows how the mean and correlation information actually interact mathematically in a way that helps selection probability.

Jane: So, if an alternative has a high mean but low correlation with others, this framework suggests it might be penalized compared to one with a lower mean but higher correlated alternatives.

Title and authors: Lu: Theorem two analyzes the derivative with respect to correlation information by defining I+(tau) and I-(tau), showing that the sign of the derivative depends on whether the mean mu i is larger or smaller than mu tau <ref:2402.02196#pg1>.

Meng: That sounds like it’s a complex trade-off you have to manage, depending on whether your alternatives are generally high or low performing in terms of their means.

Tom: And Corollary one shows that when tau equals one, Dtaui is less than Iτi under Assumption 3a and 3b. This suggests that the correlation information helps push the selection probability in the right direction for a specific alternative.

Jane: It sounds like this theoretical analysis gives us confidence that what they are proposing isn't just a clever heuristic, but something grounded in how means and correlations actually affect the outcome.

Lu: So to wrap up, the paper "Sample-Efficient 'Clustering and Conquer' Procedures for Parallel Large-Scale Ranking and Selection" shows that leveraging correlation is an effective way to improve sample efficiency in large-scale R andS. The P3C procedure is a ready-to-use tool that achieves significant improvements without requiring excessively high clustering accuracy; it only needs to identify which alternatives are highly correlated.

Meng: And for the fixed-budget setting, combining P3C with allocation rules like OCBA or CBA shows that correlation-based clustering can further improve PCStrad by increasing the correlations among alternatives assigned to the same processor.

Tom: So, we have a procedure that moves us toward better sample efficiency without needing perfect knowledge upfront. It makes the distinction between good and bad alternatives more pronounced, which should accelerate the entire ranking procedure.

Jane: It’s definitely a tool that simplifies how we think about large-scale R andS by showing that grouping alternatives based on correlation is a powerful way to reduce the necessary simulation budget or precision target.

Lu: For future work, they could explore more complex clustering structures, but for now, the focus is on proving sample optimality under both fixed-precision and fixed-budget formulations as p goes to infinity.

Meng: From a deployment standpoint, I’d want to see more validation on how this performs when the correlation structure changes dynamically during a real run.

Tom: That’s definitely something worth looking into next. But for now, the P3C procedure is a ready-to-use tool achieving significant improvements based on these results in "Sample-Efficient 'Clustering and Conquer' Procedures for Parallel Large-Scale Ranking and Selection."

The paper's summary: Tom: So, to wrap up what we just heard: this paper is about taking large ranking and selection problems where you have way too many alternatives and making them sample more efficiently by using correlation information instead of treating everything as totally independent.

Jane: Right, so they propose this new method called the Parallel Correlation Clustering and Conquer procedure that groups alternatives together based on how correlated they are.

Lu: What’s really interesting is how they frame the problem—they look at both fixed-precision and fixed-budget procedures, which are the two main ways you usually set up these ranking simulations.

Meng: So, if I'm running a simulation for Neural Architecture Search, this method helps me know if I can get the same quality result with fewer total samples before the budget runs out.

Tom: Exactly. They show that P3C is sample optimal under both those constraints, meaning as you have more alternatives, the number of samples needed doesn't explode; it just grows linearly with the number of alternatives, which is good for practical deployment costs.

Jane: And they show how this works mathematically by looking at a gradient framework that connects the mean of an alternative to its correlation structure.

Lu: They proved something about that interaction—that stronger correlation tends to push the selection probability in a certain direction, depending on whether the alternative's mean is high or low.

Meng: That makes sense for scaling; if we can predict which groups of alternatives are likely to be together, we can process them faster locally instead of throwing everything at every processor randomly.

Tom: The experiments back this up; in drug discovery, they saw P3C-KN and P3C-KT consistently beating other algorithms, with the KT version even showing a drop in the constant factor when you use their method.

Jane: It suggests that we don't need perfect knowledge of every single correlation to get big gains; we just need to identify which groups are highly correlated, which is much more realistic for a real-world setup.

Lu: And they built an algorithm called AC+ that helps them partition those alternatives into clusters using just the initial samples from Stage zero which is a smart way to manage the complexity upfront.

Meng: I’m curious about what this means for my team; if we can reduce the sample size needed by half with this structure, it really impacts how much data we need to collect for our next model iteration.

Tom: It’s about making the simulation process itself smarter, not just finding a better way to run the existing algorithms on more hardware.

Jane: So what they're showing us is that using correlation information is a fundamental way to improve sample efficiency in parallel large-scale ranking problems without needing excessively high clustering accuracy.

Lu: For the fixed-budget scenario, when you combine P3C with allocation rules like OCBA or CBA within those clusters, you see even bigger improvements in the performance metric PCStrad than existing methods like FBKT.

Meng: An order of magnitude improvement is significant; that’s what we need to see when we're pushing the limits of model search spaces.

Tom: It confirms that correlation-based clustering is a ready-to-use tool here, and it makes the distinction between good and bad alternatives much clearer, which should speed up the whole procedure.

Jane: So this paper gives us a concrete procedure to start using immediately, focusing on finding those strong correlations to make our ranking simulations more effective.

Lu: The theoretical analysis shows that this approach holds up under both fixed-precision and fixed-budget constraints as the number of alternatives p goes to infinity, which is a big piece of evidence for its robustness.

Meng: It’s interesting how they tie this back to the mean information too; it’s not just about grouping things together, it's about understanding how those groups behave relative to the overall performance means.

Tom: So while they prove the math works, I think we need to see more validation on how this correlation structure changes dynamically during a real run before we deploy this for production systems.

The paper's improvements: Tom: So we're talking about what actually makes this method better than just running simulations normally: they introduce a new way to look at the math, focusing on how the mean and correlation information actually interact.

Jane: What does that mean for a user trying to set up their simulation? It means if an alternative has a high average score but is totally random with all the others, this framework gives it less weight in the selection process.

Lu: They developed this gradient-based framework, Theorem one which shows that how much you change the selection probability based on the mean of an alternative is positive and linked to its correlation structure.

Meng: That sounds like a sophisticated way to handle uncertainty; instead of just looking at one number, you're seeing how that number plays out in relation to the whole group.

Tom: And then they have Theorem two where they look at the derivative with respect to correlation information, I plus tau and I minus tau, and that sign changes depending on whether the mean mu i is larger or smaller than mu tau.

Jane: So if an alternative's mean is higher than its neighbors' means, does the correlation information push it further toward being selected?

Lu: They showed that when you look at a specific alternative, tau equals one, Dtau i is less than I tau i under Assumption 3a and 3b, which basically confirms that the correlation data helps steer the selection probability in the right direction for a specific choice.

Meng: That sounds like it’s giving us a mathematical handle on how to filter out noise effectively without having to run millions of extra simulations just for that one step.

Tom: It means we can use this interaction between mean and correlation information to make decisions that are more informed, even when the underlying data is noisy.

Jane: This moves beyond just grouping things; it’s about using the relationship between different types of information to decide which alternatives are truly worth investigating further.

Lu: The core implication is that correlation-based clustering isn't just a trick to reduce variance; it’s a way to make the selection criteria itself more sensitive to meaningful structural relationships in the data.

Meng: For deployment, if we can build a system that uses this interaction, it means we can achieve better results with less computational overhead because we're being smarter about which alternatives get their chance to be compared closely.

Tom: It’s about making the decision-making process more efficient at the fundamental level of how the selection probability is calculated.

Jane: This gives us confidence that what they are proposing isn't just a clever heuristic; it's grounded in how means and correlations mathematically affect the final selection outcome.

Lu: They’ve shown that this approach helps accelerate the whole ranking procedure by making the distinction between a good alternative and a bad one much more pronounced.

Meng: That makes sense because if an alternative is highly correlated with others, its performance signal becomes stronger, which should speed up how fast we can identify the best one.

Tom: So we're moving past just running faster simulations; we’re changing the underlying logic of why those simulations are needed.

Jane: This research suggests that leveraging correlation is a powerful way to improve sample efficiency in large-scale ranking problems, and it gives us a ready-to-use tool for that job.

Lu: The paper's main contribution is proving that this P3C procedure achieves sample optimality under both fixed-precision and fixed-budget formulations as the number of alternatives p goes to infinity.

Meng: So even if the problem gets massive, as long as we use this structure, we still hit that O of p growth for our required samples.

Tom: It means we can scale our experiments confidently knowing that our sample requirement is being minimized in the most efficient way possible for these simulation types.

Jane: This sets a new standard for how we think about large-scale R andS, showing that grouping alternatives based on correlation is a very strong technique.

Lu: And they also developed this novel gradient framework to examine the mean–correlation interaction, which is where the real theoretical muscle of this work lies.

Meng: I'm still focused on practical deployment costs; though, I think understanding these theoretical bounds helps me justify the infrastructure we need to build for this kind of system.

Conclusion: Tom: So to wrap up, we've discussed how the Parallel Correlation Clustering and Conquer procedure improves sample efficiency in large-scale ranking and selection problems by using correlation information to group alternatives on single processors rather than random distribution.

Jane: Exactly, it's a ready-to-use tool that helps us reduce the necessary simulation budget or precision target without needing excessively high clustering accuracy.

Lu: The theoretical analysis shows this P3C procedure is sample optimal under both fixed-precision and fixed-budget formulations as the number of alternatives p goes to infinity.

Meng: So even if we're dealing with a massive search space, the required number of samples only grows linearly with the size of that space, which is good for keeping our operational costs manageable.

Lalam: I’m sure I can attest to this; the paper shows P3C is a ready-to-use tool achieving significant improvements without needing perfect knowledge upfront.

Tom: It's about making the decision-making process smarter by exploiting those underlying structural relationships between alternatives, which should really accelerate the whole ranking procedure.

Jane: This research suggests that leveraging correlation is an effective way to improve sample efficiency, and it gives us a concrete method to make large-scale R andS more practical.

Lu: They also developed a novel gradient framework to examine the mean–correlation interaction between mean and correlation information, which is where the real theoretical muscle of this work lies.

Meng: I'm still focused on practical deployment costs; though, understanding these theoretical bounds helps me justify the infrastructure we need to build for this kind of system.

Tom: And they proved that correlation-based clustering can further improve the performance metric PCStrad in fixed-budget settings by increasing correlations among alternatives assigned to the same processor.

Jane: So we've seen how grouping alternatives based on correlation is a powerful way to reduce the necessary simulation budget or precision target.

Zishi Zhang, Yijie Peng

Guanghua School of Management, Peking University · School of Management and Engineering, Nanjing University

stat.ME, cs.LG

Submitted: 2024-02-03

Updated: 2026-10-08

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: The gist The work aims to improve sample efficiency of parallel large-scale ranking and selection (R&S) problems by leveraging correlation information.

Key concepts

Ranking and Selection (R&S): This is a problem where you need to identify the best alternative among many options. The goal is to find the unique true best alternative under constraints like a fixed budget or fixed precision, often in parallel.
Correlation Information: Instead of assuming alternatives are independent, this method uses information about how alternatives relate to each other. Alternatives with high positive correlation are grouped together because they tend to fluctuate in the same direction, which helps stabilize the selection process.
Parallel Correlation Clustering and Conquer (P3C): This is the proposed procedure that modifies traditional methods by first clustering alternatives based on their correlation. It assigns alternatives within the same cluster to one processor, allowing that processor to handle them together, which is more efficient than random distribution.

Terminology

Clustering — 5× in this paper · explained in 2 episodes
Sample Complexity — 2× in this paper · explained in 20 episodes
Sample Efficiency — 2× in this paper · explained in 7 episodes
Benchmark — 1× in this paper · explained in 2 episodes

Summary

The gist The work aims to improve sample efficiency of parallel large-scale ranking and selection (R&S) problems by leveraging correlation information.

Problem Formulation

The paper frames Ranking and Selection (R&S) into two main formulations: fixed-precision and fixed-budget procedures. Fixed-precision R&S procedures terminate the simulation once a pre-specified level of precision is achieved, while fixed-budget R&S procedures stop when a predetermined total simulation budget is exhausted to optimize precision. The literature on large-scale R&S can be divided into two branches: one focusing on parallel computing implementation challenges and another focusing on improving sample efficiency. Existing R&S algorithms typically assume independence across alternatives, thereby discarding valuable shared information. Leveraging such shared information offers a promising direction for improving sample efficiency, yet this remains largely unexplored in the large-scale R&S literature. The objective of R&S is to identify the unique true best alternative [1] under either a fixed-budget constraint or a fixed-precision constraint. The precision metric used is the traditional PCS, defined as Pxbar[1] > xbarj, j!= [1]. This generalizes to the individual PCS of τ as PCStrad = P(xbarτ > xbarj, j!= τ), which serves as a probabilistic criterion for assessing the performance of alternative τ.

How it works

The proposed method modifies the traditional divide and conquer framework by adding a correlation-based clustering step to create the Parallel Correlation Clustering and Conquer (P3C) procedure. This modification clusters alternatives based on their correlation and assigns alternatives from the same cluster to a single processor, rather than distributing them randomly. The intuition behind P3C is that clustering strongly positively correlated alternatives together can effectively cancel out stochastic fluctuations in the same direction. This concept is similar to the CRN technique, which introduces positive correlation artificially to reduce variance and expedite pairwise comparisons.

The P3C procedure involves several stages as outlined in Algorithm 1:

Stage 0 (Initialization): Simulate each alternative at least N0 ≥ 3 times and estimate the covariance matrix Σ.
Stage 1 (Clustering): Continue sampling and apply the correlation-based clustering algorithm AC+ using the initialization samples collected in Stage 0 to partition all alternatives into several clusters. Assign all alternatives within the same cluster to a single processor.
Stage 2 (Conquer): Perform R&S within each cluster to select the local best.
Stage 3 (Final Comparison): If the number of remaining alternatives in contention satisfies ≤ pm, perform R&S to select the global optimal alternative; otherwise, repeat Stage 1 and Stage 2 for the remaining alternatives.

Theoretical Analysis

The theoretical analysis is conducted from two perspectives: establishing sample optimality and examining the benefit of correlation-based clustering. P3C is shown to be sample optimal under both the fixed-precision and fixed-budget formulations as the number of alternatives p grows under Assumption 1. Specifically, Proposition 1 establishes that for P3C-KN and P3C-KT, E[NP3C−KN] = O(p) and E[NP3C−KT] = O(p) as p → ∞ in the fixed-precision formulation. For the fixed-budget formulation with equal allocation (EA), Proposition 2 shows that P3C-EA has RMCS Θ(p) and is sample optimal under the condition N > pη1.

The analysis also develops a novel gradient-based framework to examine the mean–correlation interaction between mean and correlation information. Theorem 1 establishes that PCS(τ) is differentiable with respect to µ = (µ1,·····, µp) and ∂PCS(τ)/∂µτ > 0. Theorem 2 analyzes the derivative with respect to correlation information by defining I+(τ) and I-(τ), showing that the sign of the derivative depends on whether the mean µi is larger or smaller than µτ. Corollary 1 demonstrates that when τ = [1], Dτ[i] < Iτ[i] under Assumption 3(a) and 3(b).

Performance Evaluation

Numerical experiments demonstrate the effectiveness of P3C in both fixed-precision and fixed-budget settings. In the fixed-precision R&S setting for drug discovery, P3C-KN and P3C-KT consistently outperform other algorithms, exhibiting linear growth in sample size with respect to p. Notably, the slope of KT decreases visibly after applying P3C, suggesting that P3C can further reduce its leading constant and improve finite-sample performance. In the fixed-budget R&S setting for Neural Architecture Search (NAS), P3C-OCBA and P3C-CBA improve PCStrad by more than an order of magnitude compared to FBKT, demonstrating its effectiveness in large-scale fixed-budget R&S. The PCCAC+ algorithm shows that as the sample size N → ∞ (with p fixed), the lower bound of PCCAC+ converges to 1 − k(1 − 1/k)ps, which approaches 1 as ps increases.

Conclusion

The paper concludes that leveraging correlation is an effective way to improve sample efficiency in large-scale R&S. The P3C procedure is a ready-to-use tool that achieves significant improvements without requiring excessively high clustering accuracy; it only needs to identify which alternatives are highly correlated. The analysis confirms that correlation-based clustering can further improve PCS by increasing the correlations among alternatives assigned to the same processor. This makes the distinction between good and bad alternatives more pronounced, thereby accelerating the R&S procedure. The P3C procedure is a ready-to-use tool that achieves significant improvements without requiring excessively high clustering accuracy; it only needs to identify which alternatives are highly correlated.

References

Al-Kandari NM, Jolliffe IT (2001) Variable selection and interpretation of covariance principal components.

Bechhofer RE (1954) A single-sample multiple decision procedure for ranking means of normal populations with known variances.

Chen CH, Chick SE, Lee LH, Pujowidianto NA (2015) Ranking and selection: Efficient simulation budget allocation.

Chen CH, Lin J, Y¨ucesan E, Chick SE (2000) Simulation budget allocation for further enhancing the efficiency of ordinal optimization.

Chen X, Ankenman BE, Nelson BL (2012) The effects of common random numbers on stochastic kriging metamodels.<ref:2402.

Improvements for AI systems

Bold Header: Parallel Correlation Clustering and Conquer (P3C) Procedure Implementation

This procedure transforms traditional divide and conquer into a clustering and conquer strategy by adding a correlation-based step, allowing alternatives to be assigned to a single processor based on their estimated correlations, which is shown to "substantially improve the performance of large-scale R&S procedures."

Bold Header: Improved Sample Efficiency in Fixed-Precision R&S

The modified approach leverages a novel gradient analysis framework that reveals a mean–correlation interaction, showing that stronger correlation tends to increase the selection probabilities of high-mean alternatives while decreasing those of low-mean alternatives, leading to an improvement in the probability of correct selection (PCS).

Bold Header: Enhanced Performance in Fixed-Budget R&S

For fixed-budget problems, P3C can be combined with various budget allocation rules like OCBA or CBA within clusters, and it is shown that correlation-based clustering can further improve PCS by increasing the correlations among alternatives assigned to the same processor.

Bold Header: Parallel Few-Shot Alternative Clustering (AC+)

The proposed AC+ algorithm enables efficient parallelization by splitting alternatives into a support set and a query set, allowing for the selection of prototypes and parallel matching across processors without requiring an estimation of the entire correlation matrix.

Bold Header: Sample Optimality Under Large-Scale Constraints

The P3C procedure is shown to be sample optimal under both the fixed-precision and fixed-budget formulations, achieving an asymptotic order of O(p) for required sample size as p approaches infinity, confirming that it can "achieve sample optimality when combined with classic fixed-budget and fixed-precision R&S algorithms."

Bold Header: Application in Neural Architecture Search (NAS)

In large-scale NAS problems, P3C-CBA and P3C-OCBA demonstrate that they improve by more than an order of magnitude over existing strategies, significantly surpassing FBKT performance and leading to a higher PCStrad under fixed budgets.

Sources

Analyzing and provably improving fixed budget ranking and selection algorithms