Apparatus and method for hyperparameter tuning based on gaussian processes

The Gaussian process-based hyperparameter tuning method optimizes hyperparameters to balance speed and accuracy in simulation-based optimization, addressing inefficiencies in smart manufacturing systems by reducing computational costs and enhancing production control reliability.

KR103046662B1Active Publication Date: 2026-11-11DONGGUK UNIVERSITY INDUSTRY ACADEMIC COOPERATION FOUNDATION
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
KR1020250110305
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2025-08-11
Publication Date
2026-11-11
Estimated Expiration
2045-08-11

AI Technical Summary

Technical Problem

Existing simulation-based optimization (SBO) methods in smart manufacturing systems, particularly in semiconductor manufacturing, face challenges with high computational costs and inefficiencies in real-time decision-making due to manual setting of hyperparameters like significance level (α) and sample size, leading to suboptimal simulation speed and accuracy.

Method used

A Gaussian process-based hyperparameter tuning method that includes a condition generation unit, initial value setting, acquisition function calculation, candidate selection, and simulation unit to optimize hyperparameters efficiently, balancing speed and accuracy in adaptive abstraction control simulations.

Benefits of technology

This method significantly reduces simulation execution time while maintaining accuracy within user-defined tolerances, ensuring high reliability and deriving optimal production control solutions by minimizing computational errors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112025090887339-PAT00133_ABST
    Figure 112025090887339-PAT00133_ABST
Patent Text Reader

Abstract

The present invention relates to the simulation and optimization of smart manufacturing systems, and more specifically, to an apparatus and method for efficiently tuning hyperparameters by utilizing a Gaussian process to achieve optimal performance during the simulation process. According to an embodiment of the present invention, the Gaussian process-based hyperparameter tuning apparatus and method can effectively resolve the trade-off between speed and accuracy in simulation-based optimization in complex production environments, such as semiconductor manufacturing processes. In particular, the present invention can significantly reduce simulation execution time by maximizing the speed improvement of AACS through Gaussian process-based hyperparameter optimization. Furthermore, the present invention guarantees the quality of the optimization results by maintaining simulation accuracy within a user-defined tolerance range through hyperparameter adjustment.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present invention relates to the simulation and optimization of a smart manufacturing system, and more specifically, to an apparatus and method for efficiently tuning hyperparameters by utilizing a Gaussian process to achieve optimal performance during the simulation process. Background Technology

[0003] Due to the high complexity and frequent re-entry required in semiconductor manufacturing processes, decision-making through simulation is essential. However, Simulation-Based Optimization (SBO) demands very high computational costs, and the efficiency of real-time decision-making deteriorates as simulation execution times lengthen. To address these issues, Adaptive Abstraction Control Simulation (AACS) has been proposed. AACS accelerates simulations by switching between a high-precision Discrete Event Model (DEM) and a Mean Delay Model (MDM). However, inaccurate simulations can degrade the quality of optimization results. Furthermore, to date, significant factors controlling the speed and accuracy of AACS... Hyperparameters such as the level (α) and sample size are set manually, making it difficult to find optimal values ​​depending on the production scenario. This leads to limitations in improving simulation speed and poses a problem that degrades the actual production process control performance.

[0004] The background technology of the present invention is disclosed in Korean Published Patent No. 10-2021-0190143. The problem to be solved

[0006] The present invention provides a Gaussian process-based hyperparameter tuning device and method.

[0007] The present invention provides a method for optimizing the trade-off between speed and accuracy when performing simulation-based optimization (SBO) in a smart manufacturing system. In particular, the present invention provides a method for efficiently tuning two key hyperparameters, significance level (α) and sample size, to maximize the speed improvement of Adaptive Abstraction Transition Simulation (AACS) in complex production scenarios, such as semiconductor manufacturing processes, while minimizing the loss of simulation accuracy.

[0008] The technical problems that the present invention aims to solve are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art to which the present invention belongs from the description below. means of solving the problem

[0010] According to one aspect of the present invention, a Gaussian process-based hyperparameter tuning device and method are provided.

[0011] A Gaussian process-based hyperparameter tuning device according to one embodiment of the present invention may include a condition generation unit that generates a plurality of production condition candidate groups through a genetic algorithm, an initial value setting unit that sets an initial value of a Gaussian process model for each generated individual based on the previous search results of neighboring individuals, an acquisition function calculation unit that calculates an EI (Expected Improvement) acquisition function based on the initialized Gaussian process model, a candidate selection unit that selects hyperparameter candidates that maximize the EI value, and a simulation unit that performs an abstraction adjustment simulation by applying the selected hyperparameter combination.

[0012] A Gaussian process-based hyperparameter tuning method according to one embodiment of the present invention may include the steps of: generating a plurality of production condition candidate groups through a genetic algorithm; setting initial values ​​of a Gaussian process model for each generated individual based on previous search results of neighboring individuals; calculating an EI (Expected Improvement) acquisition function based on the initialized Gaussian process model; a candidate selection unit for selecting hyperparameter candidates that maximize the EI value; and performing an abstraction adjustment simulation by applying the selected hyperparameter combination. Effects of the invention

[0014] According to an embodiment of the present invention, a Gaussian process-based hyperparameter tuning device and method can effectively resolve the trade-off between speed and accuracy in Simulation-Based Optimization (SBO) in complex production environments, such as semiconductor manufacturing processes. In particular, the present invention can significantly reduce simulation execution time by maximizing the speed improvement of adaptive abstraction conversion simulations through Gaussian process-based hyperparameter optimization. Furthermore, the present invention guarantees the quality of optimization results by maintaining simulation accuracy within a user-defined tolerance range through hyperparameter adjustment. Through this, the present invention provides high reliability by minimizing computational errors occurring during the simulation process. Accordingly, the present invention can derive an optimal production control solution that reduces the average cycle time (MCT) of a process and increases throughput.

[0015] The effects of the present invention are not limited to the effects described above, and should be understood to include all effects that can be inferred from the composition of the invention described in the description or claims of the present invention. Brief explanation of the drawing

[0017] FIG. 1 is a diagram illustrating the overall configuration of a Gaussian process-based hyperparameter tuning device according to an embodiment of the present invention. FIG. 2 is a diagram illustrating an integrated simulation-based optimization process that incorporates hyperparameter adjustment according to an embodiment of the present invention. FIG. 3 is a diagram illustrating a detailed and comprehensive simulation execution process utilizing various random seeds according to an embodiment of the present invention. FIG. 4 is a diagram showing a hyperparameter optimization algorithm according to an embodiment of the present invention. FIGS. 5 to 9 are drawings illustrating experimental data and results of a Gaussian process-based hyperparameter tuning device according to an embodiment of the present invention. FIG. 10 is a flowchart of a Gaussian process-based hyperparameter tuning device according to one embodiment of the present invention. Specific details for implementing the invention

[0018] The present invention is susceptible to various modifications and may have various embodiments; therefore, specific embodiments are illustrated in the drawings and described in detail. However, this is not intended to limit the present invention to specific embodiments, and it should be understood that the invention includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention. In describing the present invention, detailed descriptions of related prior art are omitted if it is determined that such detailed descriptions would unnecessarily obscure the essence of the invention. Furthermore, singular expressions used in this specification and claims should generally be interpreted to mean "one or more" unless otherwise stated.

[0019] Throughout the specification, when it is stated that a part is "connected (connected, in contact, combined)" with another part, this includes not only cases where they are "directly connected," but also cases where they are "indirectly connected" with other members interposed between them. Furthermore, when it is stated that a part "includes" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but rather allows for the inclusion of additional components.

[0020] The terms used herein are merely for describing specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as “comprising” or “having” are intended to indicate the presence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0021] The present invention will be described below with reference to the attached drawings. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification have been given similar reference numerals.

[0023] FIG. 1 is a diagram illustrating the overall configuration of a Gaussian process-based hyperparameter tuning device according to one embodiment of the present invention.

[0024] Referring to FIG. 1, a Gaussian process-based hyperparameter tuning device (100) includes a condition generation unit (110), an initial value setting unit (130), an acquisition function calculation unit (150), a candidate selection unit (170), and a simulation unit (190).

[0025] The condition generation unit (110) generates a number of production condition candidate groups through a genetic algorithm. Specifically, the condition generation unit (110) generates many entities representing control or decision candidates during the time of Simulation-Based Optimization (SBO) execution.

[0026] The initial value setting unit (130) sets the initial value of the Gaussian process model for each generated object based on the previous search results of neighboring objects. At this time, the initial value setting unit (130) has a group of candidate initial hyperparameter values. Defines, and some are composed of predicted values ​​without actual simulation.

[0027] The acquisition function calculation unit (150) calculates an EI (Expected Improvement) acquisition function based on the initialized Gaussian process model. The acquisition function calculation unit (150) calculates the predicted mean value (μ) and variance (σ) for candidate combinations of abstraction control hyperparameters and uses them to calculate the Expected Improvement (EI) acquisition function. The EI function is as shown in Equation 1 below.

[0028] [Mathematical Formula 1]

[0029]

[0030] The above The function represents the expected value that more gold will be found than the current highest gold vein, and the average estimated amount The higher the value is compared to the current highest record, the larger the result value becomes. The above The function represents a bonus score for uncertainty, and the value increases as the uncertainty of the prediction increases. The above mean estimation and variance estimation functions will be explained in more detail in Fig. 4.

[0031] The candidate selection unit (170) selects a hyperparameter candidate that maximizes the EI value. Hyperparameter candidate that maximizes the EI value It is equal to the mathematical formula 2 below.

[0032] [Mathematical Formula 2]

[0033]

[0034] The simulation unit (190) performs an abstraction adjustment simulation by applying the selected hyperparameter combination. Based on the number of events and accuracy loss information collected from the simulation, the simulation unit (190) calculates the suitability of the object by considering the speed improvement rate and accuracy constraints. The speed improvement rate is measured not by actual time, but by the ratio of the total number of events processed in the reference simulation (DOS) and the abstraction adjustment simulation (AACS).

[0036] FIG. 2 is a diagram illustrating an integrated simulation-based optimization process that incorporates hyperparameter adjustment according to an embodiment of the present invention.

[0037] Referring to FIG. 2, the present invention performs GA-based optimization and hyperparameter tuning to simultaneously perform Hyperparameter Optimization (HPO) and Simulation-Based Optimization (SBO). First, the GA-based optimizer performs optimization of control candidates specialized for the problem. The present invention generates multiple individuals of the next generation (i.e., offspring) based on the fitness of all individuals in the current population. Each individual represents a manufacturing decision candidate, and the average fitness of the individual is measured based on the results of multiple simulations using various random seeds through a user-defined objective function. Then, hyperparameter tuning is performed on the newly generated offspring population Each object It collaborates with a GA-based optimizer to perform the roles of exploration and utilization. Hyperparameter tuning explores hyperparameter points to find the optimal point in various hyperparameter spaces and allocates a budget for this purpose. Subsequently, in Equation 3, hyperparameter tuning uses the optimal hyperparameter point through adaptive abstraction transition simulation It evaluates the fitness of. In this case, hyperparameter adjustment is performed for each object Manages a Gaussian Process (GP) model for, and through this, an unknown speed enhancement function Find the optimal hyperparameter point that maximizes it.

[0038] [Mathematical Formula 3]

[0039]

[0040] In mathematical formula 3, is a hyperparameter Represents the search space of, class is existing hyperparameters These are two hyperparameters encoded in and m, respectively. The existing hyperparameters α and m are user-defined search spaces and It is defined in, that is is. The maximum α value is 2, which indicates that the candidate does not require AACS (Adaptive Abstraction Control Simulation). When supA=2, statistical becomes 0, and no abstraction transition occurs. If If the value is 2, the simulation speed improvement is always less than 1 due to unnecessary computational overhead for steady-state detection. α range It roughly defines, where each It disables ALCs to not monitor the steady-state convergence of the group.

[0041] The present invention in the above α from (or from m) Each encoding converted to [0,1] is performed considering the scale of A (or M). In addition, when the present invention applies candidates derived by GA to AACS, From to α (or Decoding is performed (from to m). The scale-based encoding and decoding procedure computes a kernel between two hyperparameters of equal importance during GP-based speed enhancement estimation, and continuous GP variables It is possible to map between and the discrete sample size m.

[0043] FIG. 3 is a diagram illustrating a detailed and comprehensive simulation execution process utilizing various random seeds according to an embodiment of the present invention.

[0044] The present invention must specify two types of maximum simulation budgets for each object. is the total number of simulations, and is the maximum number of simulations to be used for HPO search. Here, must be an even number, but it is the number of simulations actually used in the HPO search It may vary depending on the confidence of the estimated hyperparameters. Remaining simulation budget is the optimal hyperparameter derived from the current search of the HPO It is used for multiple simulation runs to evaluate the fitness of an object. Referring to Fig. 3, the fitness of the object is It is calculated as the average of the simulation results. The number of simulations for fitness evaluation is equal to the number of required random seeds. During HPO search, for each entity, the present invention Run n simulation pairs. In each simulation pair, the same random seed is shared to give a given object Adaptive Abstraction Control Simulation (AACS) and Detailed-Only Simulation (DOS) are performed on the data. Through this, accuracy loss and speed improvement are measured, and the accuracy loss is given by Equation 4 using a random seed s and a hyperparameter x.

[0045] [Mathematical Formula 4]

[0046]

[0047] In mathematical formula 4, is AACS (Adaptive Abstraction Control Simulation), using DOS (Deterministic Optimization Simulation) This is the goodness of fit evaluated based on , x, and s.

[0048] The present invention runs multiple simulations for each hyperparameter candidate in a speed evaluation based on actual execution time. This is unavoidable due to the stochastic variability of simulation execution times, even in a single computing environment. To overcome this, for a given hyperparameter x and a random seed s, an object The AACS speed improvement is measured as the ratio of the number of events processed in DOS to the number of events processed in AACS. In other words, the speed improvement is given by Equation 5 below:

[0049] [Mathematical Formula 5]

[0050]

[0051] In mathematical formula 5, and is given Represents the number of events processed during DOS and AACS for , x, and s. Given a specific random seed, the present invention predetermines the entire set of event-based tasks for a target scenario simulation. The AACS strategy accelerates traditional discrete event simulations by omitting repetitive event-based tasks within a steady-state machine group. Consequently, the increase in the rate of improvement resulting from a specific hyperparameter candidate in AACS is associated with the number of events that can be eliminated. For each hyperparameter candidate, the total number of events processed in standard DOS for the same entity and random seed is deterministically reduced. Unknown rate of improvement function is measured using a random seed s as in Equation 6.

[0052] [Mathematical Formula 6]

[0053]

[0054] In mathematical formula 6, is a user-defined accuracy constraint. Since the speed improvement may change when the random seed s changes, the present invention Statistical variance is inevitable in measurement.

[0056] FIG. 4 is a diagram illustrating a hyperparameter optimization algorithm according to an embodiment of the present invention.

[0057] Referring to Fig. 4, hyperparameter tuning manages the GP (Gaussian Process) model of each entity as a proxy model for HPO exploration to predict the expected speed improvement at the hyperparameter point based on previous measurements before measuring the actual speed improvement. First, the entity If the GP model is not initialized for HPO, hyperparameter tuning prepares initial data called the warm-up phase. Hyperparameter tuning is a set of hyperparameter points and set the corresponding speed enhancement as initialization data. During the warm-up phase, hyperparameter adjustment is an object Using previous exploration information obtained from neighbors It attempts to predict the speed improvement at each point without simulation. If If it is predictable without simulation, is an additional set It is stored in. Otherwise, It is measured through DOS and AACS run pairs. All and the corresponding predicted or measured speed improvement are each and It is stored in. Before starting a search, hyperparameter adjustments are made to the previous search information. and Optimal hyperparameter candidates based on Derives. If it is not measured by simulation ( ), hyperparameter adjustment is accurate To calculate the value, it is re-evaluated using a simulation pair. This In a GP-based HPO approach that intensively explores the surroundings It is important to ensure speed reliability. In the HPO stage, the present invention [regarding] the budget Hyperparameter space until exhausted Explores.

[0058] The present invention relates to already collected hyperparameters and the corresponding speed enhancement , The previous GP is updated based on. The GP model integrates prior knowledge through the kernel and measures the uncertainty of the prediction. Subsequently, the present invention provides optimal hyperparameter candidates Acquires. New hyperparameters The current optimal value If it is sufficiently close to , the present invention considers the hyperparameter x to have converged. Hyperparameter convergence is defined as a termination condition to prevent further searching, and Is Converges to, Satisfies the conditions. represents a user-defined threshold pair for a 2D hyperparameter. If If it does not converge, the present invention for Evaluate the new speed improvement value Perform additional optimization by comparing with. Discovery point and the corresponding speed improvement values ​​are respectively and It is saved to. After hyperparameter search is completed, hyperparameter adjustment is the remaining simulation budget It evaluates fit values ​​via AACS using . Subsequently, hyperparameter tuning returns the average of the fit values ​​to the GA-based optimizer. The newly created objects within If it was evaluated in the previous generation but not selected, hyperparameter adjustment is , , Reuses the previous search and evaluation results.

[0059] Meanwhile, Gaussian Process (GP)-based hyperparameter tuning determines optimal hyperparameters by analyzing the production dynamics of individuals to maximize AACS speed enhancement. During the warm-up phase, hyperparameter tuning determines an initial hyperparameter set based on the past search results of neighboring individuals. Creates. is a fixed set of points, and It consists of randomly sampled points. Hyperparameter The speed improvement in is estimated as shown in Equation 7 based on the results of neighbors.

[0060] [Mathematical Formula 7]

[0061]

[0062] Here, is a neighbor hyperparameters It is the speed enhancement gradient. The above gradient is equal to the following mathematical formula 8.

[0063] [Mathematical Formula 8]

[0064]

[0065] Here, is the distance vector between neighbors, represents the difference in speed improvement values. In the HPO (Hyperparameter Optimization) stage, the set of hyperparameter points and corresponding speed enhancement value Update the GP model using, and use Bayes's rule for the given prior-based posterior GP distribution It is equal to mathematical formula 9.

[0066] [Mathematical Formula 9]

[0067]

[0068] Here, the posterior mean estimation function is equal to Equation 10, and the variance estimation function It is equal to mathematical formula 11.

[0069] [Mathematical Formula 10]

[0070]

[0071] [Mathematical Formula 11]

[0072]

[0073] The new hyperparameter point is defined as the Expected Improvement (EI) function and is as shown in Equation 12 below.

[0074] [Mathematical Formula 12]

[0075]

[0076] Here, class are the GP posterior predicted mean and predicted standard deviation at point x, respectively, and represents the standard normal distribution function. And the optimal point is as shown in Equation 2 above.

[0077] [Mathematical Formula 2]

[0078]

[0079] Hyperparameter adjustment is two points and The search is terminated when the termination condition is satisfied, otherwise the search is continued by calculating the fitness through AACS and DOS execution. The present invention performs optimization after calculating the first derivative using a generally well-functioning technique such as L-BFGS-B (Limited-memory Broyden-Fletcher-Goldfarb-Shanno with Box constraints).

[0081] FIGS. 5 to 9 are drawings illustrating experimental data and results of a Gaussian process-based hyperparameter tuning device according to an embodiment of the present invention.

[0082] Referring to FIG. 5, the present invention designs a case to evaluate the performance of the proposed SBO and optimizes hyperparameter values ​​to improve simulation speed within a given budget. The present invention consists of four types of machine groups: deposition, patterning, etching, and chemical-mechanical polishing, and the detailed sub-processes within each machine may vary depending on the product recipe of the lot. There are a total of five product types, each requiring different mask layers (e.g., 4, 8, 12, 16, 20). In this experiment, the number of reentrances increases as the number of required layers increases (e.g., deposition → patterning → etching → chemical-mechanical polishing). Therefore, each wafer lot revisits the same machine group multiple times for various process steps, and the number of reentrances varies depending on the product type of the target lot. Each lot consists of 25 wafers (raw materials). The present invention aims to produce 100 wafer lots per product type, meaning that a total of 500 wafer lots for five product types must be produced using different machines according to their respective processing recipes. Overall performance, such as Mean Cycle Time (MCT) and Throughput, can vary significantly depending on the lot release rate from the inventory of each product type. Generally, as the lot release interval increases, the MCT lengthens and throughput decreases due to reduced machine utilization. The performance indicators for minimizing MCT and maximizing throughput are given by Equation 13.

[0083] [Mathematical Formula 13]

[0084]

[0085] In mathematical formula 13, means an entity representing a combination of lot release intervals for five product types, and is the lot release interval for the i-th product type. The above w1 and w2 are user-defined weights, and through simulation, objects Estimated average cycle time (MCT) for, represents the average number of lots produced per hour (throughput).

[0087] Referring to Fig. 6, the first experiment measures the estimation difference and speed improvement between AACS and DOS using a simple lot release scenario. The present invention analyzes the results by releasing 25 wafer lots (5 lots per product type) for all product types at a fixed rate and varying the release interval for each product type from 100 minutes to 180 minutes. As the significance level α decreases, AACS allows for more statistical invariance between two consecutive samples, resulting in more frequent switching from Discrete-Event Models (DEM) to Mean Delay Models (MDM). As MDM is activated more frequently, speed is improved but accuracy is reduced. As the sample size decreases, the Adaptive Level Controller (ALC) detects steady-state convergence or divergence more quickly, increasing the activation period of MDM and guaranteeing speed improvement; however, due to the short sampling period, there is a higher possibility that the observed mean value will represent the behavior in a biased manner within a specific simulation time interval, which may lead to a decrease in accuracy. The activation rate of MDM decreases as release intervals shorten, as production becomes unstable; consequently, the speed improvement of AACS is minimal. If the release interval becomes extremely short, there is a possibility that ALC may accidentally activate MDM, but this is subsequently corrected through steady-state deviation detection.

[0088] On the other hand, as the release interval increases, MDM is activated more frequently, leading to increased speed improvement, and in stable production scenarios, AACS accuracy improves with increasing intervals. However, if the release interval falls within the intermediate range (approximately 140 minutes), frequent switching between DEM and MDM can negatively impact both accuracy and speed improvement. As such, speed improvement and accuracy vary depending on the hyperparameter combination and production scenario, confirming the importance of carefully determining the hyperparameter point that maximizes speed improvement under accuracy constraints. In particular, this invention requires a more cautious approach when searching for candidates in areas with insufficient accuracy, and as AACS inaccuracy increases, there is a high probability of a discrepancy occurring between the actual optimal point derived through DOS-based SBO (Simulation-Based Optimization) and the AACS.

[0090] Referring to Fig. 7, the second experiment of the present invention aims to explore the optimal release intervals for five product types. To this end, the present invention sets weights for all possible release interval combinations by considering the importance of Mean Cycle Time (MCT) and throughput, prioritizing throughput over cycle time. This setting induces the GA-based optimizer to explore decision candidates that cause relatively unstable production, and the selection of appropriate hyperparameters is essential to minimize the loss of accuracy in AACS. The GA-based optimizer was developed using integer encoding and the Boltzmann selection method, and is set with a population size of 15, a maximum number of generations of 30, a crossover rate of 0.7, a swap rate of 0.3, and a reversal rate of 0.2. Additionally, considering stochastic simulation, the total simulation budget is set to 12 and the maximum number of search simulations to 8; for gradient-based rate-up estimation, the minimum number of neighbors is set to 5, the maximum number of neighbors to 8, and the maximum neighborhood radius to 7. The accuracy constraint of the HPO is set to 2.5%, and the hyperparameter space is It consists of 16 evenly distributed points within a defined fixed set of points. The experiment is performed in an environment equipped with an Intel Xeon E5-1650 3.6 GHz CPU and 32GB of memory. In the initial phase, the tuner estimates speed improvements without simulation by utilizing the previous search results of neighbors at the fixed points and calculates fitness by consuming the entire search budget. As generations progress in the evolutionary process, HPO budget usage decreases, allowing more budget to be allocated to fitness evaluations that account for stochastic characteristics. This indicates that as individuals survive longer or new ones are created, the likelihood of incorrect fitness predictions due to stochastic variation decreases.

[0092] Referring to Figure 8, the distribution shape in Figure 8 represents the hyperparameter points used to evaluate the goodness of fit of the entire population via AACS during the SBO execution. Therefore, the graph results are at the significance level This indicates that it has a significant influence on sample size determination. Low It requires a larger sample size to ensure statistical reliability, but analysis costs may increase due to the large sample size. On the other hand, high It can increase speed by allowing a smaller sample size, but this may affect the accuracy of the analysis.

[0094] Referring to FIG. 9, the present invention provides results of a comparative evaluation of SBO performance across three simulation paradigms: DOS-based optimization, AACS-based optimization using fixed hyperparameters, and AACS-based optimization where hyperparameters are adaptively adjusted via HPO. Each evaluation focuses on the total execution time and the quality of the calculated optimal manufacturing solution (minimization of mean goodness of fit). In AACS-based optimization using fixed hyperparameters, the α value is set to 0.01, 0.05, 0.1, and 0.5. FIG. 9(a) sets a pair of α=0.01, 0.05, 0.1, and 0.5 with a sample size of 20, and the total execution time (SBO) decreases as the α value decreases for the fixed sample size. When α=0.01, the execution time is reduced by approximately 345 seconds (22.8%) compared to DOS-based optimization, but the goodness of fit is approximately 2.4% higher than that of DOS-based optimization, resulting in somewhat lower performance. On the other hand, HPO-based SBO sets different hyperparameter points for each object to maximize speed improvement under accuracy constraints. Although HPO-based SBO achieves approximately 20.5% lower speed improvement than AACS-based SBO using a minimum α, the fitness of the calculated optimal object is closer to the actual optimal value of DOS-based SBO. Furthermore, compared to fixed hyperparameter methods using α=0.05, 0.1, and 0.5, HPO-based SBO provides optimal objects of similar quality while further reducing execution time.

[0096] FIG. 10 is a flowchart of a Gaussian process-based hyperparameter tuning device according to one embodiment of the present invention.

[0097] Referring to FIG. 10, in step S1010, the Gaussian process-based hyperparameter tuning device generates a plurality of production condition candidate sets through a genetic algorithm. The Gaussian process-based hyperparameter tuning device generates the candidate sets using a genetic algorithm (GA)-based optimization program and sets optimal hyperparameters for abstraction adjustment for each individual.

[0098] In step S1020, the Gaussian process-based hyperparameter tuning device sets the initial value of the Gaussian process model for each generated object based on the previous search result of neighbor objects.

[0099] In step S1030, the Gaussian process-based hyperparameter tuning device calculates an Expected Improvement (EI) acquisition function based on the initialized Gaussian process model. At this time, the Gaussian process-based hyperparameter tuning device calculates the predicted mean value (μ) and variance (σ) for candidate combinations of abstraction control hyperparameters and uses them to calculate the Expected Improvement (EI) acquisition function.

[0100] In step S1040, the Gaussian process-based hyperparameter tuning device selects hyperparameter candidates that maximize the EI value.

[0101] In step S1050, the Gaussian process-based hyperparameter tuning device performs an abstraction tuning simulation by applying the selected hyperparameter combination. Based on the number of events and accuracy loss information collected from the simulation, the Gaussian process-based hyperparameter tuning device calculates the fitness of the object by considering the speed improvement rate and accuracy constraints.

[0103] The Gaussian process-based hyperparameter tuning method described above can be implemented as computer-readable code on a computer-readable medium. The computer-readable recording medium may be, for example, a removable recording medium (CD, DVD, Blu-ray disc, USB storage device, removable hard disk) or a fixed recording medium (ROM, RAM, computer-equipped hard disk). The computer program recorded on the computer-readable recording medium can be transmitted to another computing device via a network such as the Internet and installed on the other computing device, thereby allowing it to be used on the other computing device.

[0104] Although it has been described above that all components constituting an embodiment of the present invention are combined or operate as a single unit, the present invention is not necessarily limited to such an embodiment. That is, within the scope of the purpose of the present invention, all components may be selectively combined in one or more ways to operate.

[0105] Although operations are depicted in a specific order in the drawings, it should not be understood that the operations must be executed in the specific order depicted or in a sequential order, or that all depicted operations must be executed to obtain the desired result. In certain situations, multitasking and parallel processing may be advantageous. Furthermore, the separation of various configurations in the embodiments described above should not be understood as a necessary separation, and it should be understood that the described program components and systems can generally be integrated together into a single software product or packaged into multiple software products.

[0106] The present invention has been described above with reference to its embodiments. Those skilled in the art will understand that the present invention may be embodied in modified forms without departing from the essential characteristics of the invention. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the invention is defined by the claims, not by the foregoing description, and all variations within the scope of the claims should be interpreted as being included in the invention. Explanation of the symbols

[0108] 100: Gaussian process-based hyperparameter tuning device 110: Condition generation section 130: Initial value generation section 150: Acquisition function calculation unit 170: Candidate Selection Department 190: Simulation Section

Claims

Claim 1 A Gaussian process-based hyperparameter tuning device comprising: a condition generation unit that generates a plurality of production condition candidate groups through a genetic algorithm; an initial value setting unit that sets an initial value of a Gaussian process model for each generated individual based on the previous search results of neighboring individuals; an acquisition function calculation unit that calculates an Expected Improvement (EI) acquisition function based on the initialized Gaussian process model; a candidate selection unit that selects hyperparameter candidates that maximize the EI value; and a simulation unit that performs an abstraction adjustment simulation by applying the selected hyperparameter combination. Claim 2 In claim 1, the condition generation unit is a Gaussian process-based hyperparameter tuning device that generates the candidate group using a genetic algorithm (GA)-based optimization program. Claim 3 A Gaussian process-based hyperparameter tuning device according to claim 1, wherein the acquisition function calculation unit calculates the predicted mean value (μ) and variance (σ) for candidate combinations of abstraction control hyperparameters and calculates the Expected Improvement (EI) acquisition function using the above. Claim 4 In claim 1, the simulation unit is a Gaussian process-based hyperparameter tuning device that calculates the fitness of an object by considering the speed improvement rate and accuracy constraints based on the number of events and accuracy loss information collected from the simulation. Claim 5 A Gaussian process-based hyperparameter tuning device that, in paragraph 4, measures the speed improvement rate not as actual time, but as the ratio of the total number of events processed in the reference simulation (DOS) and the abstraction adjustment simulation (AACS). Claim 6 A Gaussian process-based hyperparameter tuning method comprising: a step of generating a plurality of production condition candidate groups through a genetic algorithm; a step of setting initial values ​​of a Gaussian process model for each generated individual based on previous search results of neighboring individuals; a step of calculating an EI (Expected Improvement) acquisition function based on the initialized Gaussian process model; a candidate selection unit for selecting hyperparameter candidates that maximize the EI value; and a step of performing an abstraction adjustment simulation by applying the selected hyperparameter combination. Claim 7 In claim 6, the step of generating a plurality of production condition candidate groups through the genetic algorithm is a Gaussian process-based hyperparameter tuning method that generates the candidate groups using a genetic algorithm (GA)-based optimization program. Claim 8 In claim 6, the step of calculating an Expected Improvement (EI) acquisition function based on the initialized Gaussian process model comprises calculating the predicted mean value (μ) and variance (σ) for candidate combinations of abstract control hyperparameters and using them to calculate an Expected Improvement (EI) acquisition function, a Gaussian process-based hyperparameter tuning method. Claim 9 In claim 6, the step of performing an abstraction tuning simulation by applying the selected hyperparameter combination is a Gaussian process-based hyperparameter tuning method that calculates the fitness of an object by considering the speed improvement rate and accuracy constraints based on the number of events and accuracy loss information collected from the simulation. Claim 10 A Gaussian process-based hyperparameter tuning method according to claim 9, wherein the speed improvement rate is measured not by actual time, but by the ratio of the total number of events processed in the reference simulation (DOS) and the abstraction adjustment simulation (AACS). Claim 11 A computer program recorded on a computer-readable recording medium that executes any one of the Gaussian process-based hyperparameter tuning methods of paragraphs 6 through 10.

Citation Information

Patent Citations

  • System and method for generating mixed variable type multivariate temporal synthetic data

    EP4209968A1

  • Meta-automated machine learning with improved multi-armed bandit algorithm for selecting and tuning a machine learning algorithm

    US20210224585A1

  • Machine learning based tuning of radio frequency apparatuses

    US20240044961A1

  • Function optimization method and apparatus

    WO2023287239A1