Automatic hyper-parameter optimization method

Abstract

The invention relates to the technical field of data optimization, and discloses an automatic hyper-parameter optimization method, which comprises the following steps of: dynamically initializing a defined hyper-parameter search space to obtain an initial hyper-parameter of the hyper-parameter search space; performing multi-target evaluation on the initial hyper-parameter to obtain comprehensive evaluation information of the initial hyper-parameter; based on the comprehensive evaluation information, carrying out adaptive strategy acquisition on the initial hyper-parameter to obtain a candidate hyper-parameter of the initial hyper-parameter; determining a local search space corresponding to the candidate hyper-parameters according to hyper-parameter configuration in the candidate hyper-parameters; performing fine adjustment on the hyper-parameter configuration based on the local search space to obtain hyper-parameter configuration variants of the candidate hyper-parameters; screening and evaluating the hyper-parameter configuration variants to obtain optimized hyper-parameters of the candidate hyper-parameters; according to the invention, the efficiency of automatic hyper-parameter optimization can be improved.

Description

Technical Field

[0001] This invention relates to the field of data optimization technology, and in particular to an automated hyperparameter optimization method. Background Technology

[0002] The automated hyperparameter optimization method provided by this invention can accurately extract the statistical features of the hyperparameter search space and adjust its boundaries through a dynamic initialization process based on historical optimization data from a data terminal. This makes the generated initial hyperparameter data more closely match the actual optimization needs, effectively reducing invalid or redundant configurations and improving the relevance and usability of hyperparameter data from the optimization starting point, laying an efficient foundation for subsequent optimization stages. Simultaneously, this method integrates hyperparameter performance indicators, resource consumption, and structural complexity through multi-objective evaluation. After standardization and multi-criteria decision analysis, it obtains comprehensive evaluation information that fully reflects the actual applicability of hyperparameters, providing accurate basis for subsequent strategy selection and further ensuring the correctness of the optimization direction. Furthermore, based on search requirement analysis, an adaptive acquisition strategy is selected, ensuring that the acquisition process of candidate hyperparameters is highly adapted to the hyperparameter complexity level and actual optimization needs, significantly improving the quality and acquisition efficiency of candidate hyperparameters and promoting the efficient advancement of the optimization process.

[0003] This method, through local search of candidate hyperparameters, allows for fine-tuning of hyperparameter configurations, deeply exploring the optimization potential of candidate hyperparameters and effectively improving the accuracy of optimized hyperparameters. In the iterative fusion stage, a hybrid evaluation framework is constructed to achieve a dynamic balance between the performance indicators and diversity of the expanded hyperparameter set. This generates a target hyperparameter group with stronger stability and higher quality, avoiding the limitations of single-dimensional convergence and ensuring that the hyperparameter group possesses both excellent performance and reasonable expansion space. Finally, by extracting the optimal configuration from the target hyperparameter group, the generated results ensure that the output optimized hyperparameters have excellent comprehensive performance, effectively improving the efficiency and accuracy of automated hyperparameter optimization. This fully meets the data terminal's need for rapid and accurate hyperparameter optimization, providing reliable hyperparameter support for improving the performance of artificial intelligence models. Therefore, how to improve the efficiency of hyperparameter optimization has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides an automated hyperparameter optimization method to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an automated hyperparameter optimization method, comprising: S1. Dynamically initialize the defined hyperparameter search space to obtain the initial hyperparameters of the hyperparameter search space; S2. Perform multi-objective evaluation on the initial hyperparameters to obtain comprehensive evaluation information of the initial hyperparameters, wherein the comprehensive evaluation information includes: performance indicators, resource consumption and structural complexity; S3. Based on the comprehensive evaluation information, an adaptive strategy is used to collect the initial hyperparameters to obtain candidate hyperparameters for the initial hyperparameters; S4. Determine the local search space corresponding to the candidate hyperparameters based on the hyperparameter configuration in the candidate hyperparameters; S5. Based on the local search space, the hyperparameter configuration is finely adjusted to obtain the hyperparameter configuration variants of the candidate hyperparameters; S6. Screen and evaluate the hyperparameter configuration variants to obtain the optimized hyperparameters of the candidate hyperparameters.

[0006] In a preferred embodiment, the step of dynamically initializing the defined hyperparameter search space to obtain the initial hyperparameters of the hyperparameter search space includes: The configuration information of the hyperparameter search space is parsed to obtain the type definition and value range of the hyperparameters; Based on the type definition and value range, the hyperparameters are uniformly sampled to obtain the original parameter set of the hyperparameters; The original parameter set is evaluated and filtered to obtain the initial hyperparameters of the hyperparameters.

[0007] In a preferred embodiment, the step of performing multi-objective evaluation on the initial hyperparameters to obtain comprehensive evaluation information of the initial hyperparameters includes: Based on the optimization task execution process of the hyperparameters, the initial hyperparameters are dynamically analyzed to obtain the performance indicators of the initial hyperparameters. Monitor the execution time and content usage of the initial hyperparameters in the optimization task execution process, and generate the resource consumption status of the initial hyperparameters; The structural complexity of the initial hyperparameters is obtained by evaluating the data report after the completion of the optimized task execution process. By integrating the performance indicators, resource consumption, and structural complexity from multiple dimensions, a comprehensive evaluation of the initial hyperparameters is obtained.

[0008] In a preferred embodiment, the step of integrating the performance indicators, resource consumption, and structural complexity from multiple dimensions to obtain comprehensive evaluation information of the initial hyperparameters includes: The dimensional differences of the performance indicators are eliminated to obtain the standardized performance values ​​of the initial hyperparameters; The resource consumption situation is adjusted within a certain range to obtain the adjusted resource information of the initial hyperparameters; Based on the correlation between the number and type of the initial hyperparameters, the structural complexity is classified into levels to obtain the complexity level of the initial hyperparameters. A multi-criteria decision analysis is performed on the standardized performance value, the adjusted resource information, and the complexity level to obtain comprehensive evaluation information of the initial hyperparameters. The formula for calculating the comprehensive score in the comprehensive evaluation information is as follows: ; In the formula, The overall score in the comprehensive evaluation information. These are preset trade-off parameters. For, the The standardized performance values ​​of each hyperparameter configuration, For the first The adjusted resource information for each hyperparameter configuration For the first The complexity level of each hyperparameter configuration.

[0009] In a preferred embodiment, the step of adaptively acquiring candidate hyperparameters based on the comprehensive evaluation information includes: Based on the complexity level in the comprehensive evaluation information, the search requirements for the initial hyperparameters are assessed to obtain a requirement report for the initial hyperparameters. Based on the demand report, select the corresponding target acquisition strategy from the preset acquisition strategies. When it is determined that there is a high demand for exploration, the random acquisition strategy is selected first; when it is determined that it needs to be carefully selected for use, the heuristic acquisition strategy is selected. The monitoring response to the target acquisition strategy is used to make real-time corrections to the comprehensive evaluation information, thereby obtaining new hyperparameters from the initial hyperparameters. Clustering optimization is performed on the new hyperparameters to obtain candidate hyperparameters for the initial hyperparameters.

[0010] In a preferred embodiment, the step of assessing the search requirements for the initial hyperparameters based on the complexity level in the comprehensive evaluation information to obtain a requirement report for the initial hyperparameters includes: The complexity level is matched and analyzed with the task mode in the optimization task execution process to obtain the optimization task matching result of the initial hyperparameters; The resource allocation schemes for similar tasks in the optimized task execution process are estimated and analyzed to generate the resource configuration requirements of the initial hyperparameters. Based on the optimization task matching results and the complexity level, the search strategy types and effect evaluations in the optimization task execution process are non-dominatedly sorted to obtain the search direction suggestions for the initial hyperparameters. Based on the suggested search directions and resource configuration requirements, a requirement report for the initial hyperparameters is generated.

[0011] In a preferred embodiment, the step of performing a non-dominated ranking of the search strategy types and effect evaluations in the optimization task execution process based on the optimization task matching results and the complexity level to obtain search direction suggestions for the initial hyperparameters includes: The search strategy types in the optimized task execution process are evaluated and backtracked to obtain historical performance data of the optimized task execution process. The historical performance data is quantitatively reconstructed to obtain the dominance relationship matrix of the optimized task execution process; Prioritize the strategy distribution features in the dominance relationship matrix to obtain search direction suggestions for the initial hyperparameters.

[0012] In a preferred embodiment, determining the local search space corresponding to the candidate hyperparameters based on the hyperparameter configuration among the candidate hyperparameters includes: Analyze the hyperparameter configurations in the candidate hyperparameters to generate performance gradient information for the candidate hyperparameters; Based on the performance gradient information, determine the sensitive dimension of the hyperparameter configuration among the candidate hyperparameters; Based on the current distribution characteristics of the sensitive dimension, the local search space corresponding to the candidate hyperparameter is determined.

[0013] In a preferred embodiment, the step of fine-tuning the hyperparameter configuration based on the local search space to obtain hyperparameter configuration variants of the candidate hyperparameters includes: Based on the local search space, a search subspace centered on the hyperparameter configuration is constructed; The sensitivity differences of the hyperparameter configuration are serialized and parsed to obtain the coordinated adjustment sequence of the hyperparameter configuration; By extrapolating the performance gradient trend of the coordinated adjustment sequence, we obtain the hyperparameter configuration variants of the candidate hyperparameters.

[0014] In a preferred embodiment, the step of screening and evaluating the hyperparameter configuration variants to obtain the optimized hyperparameters of the candidate hyperparameters includes: Obtain the performance data of the hyperparameter configuration variant; The performance data is quantified using the super-volume index to obtain the super-volume index evaluation result. Based on the evaluation results of the hypervolume index, the optimal parameter configuration variant is selected as the optimized hyperparameter of the candidate hyperparameter.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. The automated hyperparameter optimization method provided by this invention can accurately extract the statistical features of the hyperparameter search space and adjust its boundaries through dynamic initialization operations based on historical optimization data from a data terminal. This makes the generated initial hyperparameter data more aligned with actual optimization needs, effectively reducing invalid or redundant hyperparameter configurations and improving the relevance and usability of hyperparameter data from the optimization source. Simultaneously, this method can perform multi-objective evaluation of the initial hyperparameter data, integrating performance indicators, resource consumption, and structural complexity. Through standardization and multi-criteria decision analysis, it forms comprehensive evaluation information that fully reflects the actual applicability of the hyperparameters, providing a precise basis for the selection of subsequent acquisition strategies. Furthermore, based on search needs analysis, it can select suitable target acquisition strategies and combine clustering optimization to obtain high-quality candidate hyperparameters, significantly improving the efficiency and quality of hyperparameter acquisition.

[0016] 2. This invention performs local searches on candidate hyperparameters, determines the local search space based on hyperparameter configurations, and makes fine adjustments to deeply explore the optimization potential of candidate hyperparameters, effectively improving the accuracy of optimized hyperparameters. In the iterative fusion stage, optimized hyperparameters can be integrated with the current hyperparameter group to form an expanded hyperparameter set. By constructing a hybrid evaluation framework, a dynamic balance between performance indicators and diversity is achieved, generating a more stable target hyperparameter group. Finally, the optimal configuration in the target hyperparameter group is extracted to generate the results, ensuring that the output optimized hyperparameters have excellent comprehensive performance, effectively improving the efficiency and accuracy of automated hyperparameter optimization, fully meeting the needs of data terminals for hyperparameter optimization, and providing reliable support for improving the performance of artificial intelligence models. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an automated hyperparameter optimization method according to an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides an automated hyperparameter optimization method. The execution subject of this automated hyperparameter optimization method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the automated hyperparameter optimization method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating an automated hyperparameter optimization method according to an embodiment of the present invention. In this embodiment, the automated hyperparameter optimization method includes: S1. Dynamically initialize the defined hyperparameter search space to obtain the initial hyperparameters of the hyperparameter search space; In this embodiment of the invention, the step of dynamically initializing the defined hyperparameter search space to obtain the initial hyperparameters of the hyperparameter search space includes: The configuration information of the hyperparameter search space is parsed to obtain the type definition and value range of the hyperparameters; Based on the type definition and value range, the hyperparameters are uniformly sampled to obtain the original parameter set of the hyperparameters; The original parameter set is evaluated and filtered to obtain the initial hyperparameters of the hyperparameters.

[0021] Specifically, when parsing the configuration information of the hyperparameter search space, the file or data structure storing the configuration information is first obtained. This configuration information contains a specific description of each hyperparameter. The description of each hyperparameter is identified and extracted one by one. The type definition refers to the category to which the hyperparameter belongs, such as integer, floating-point number, string category, etc. The value range refers to the possible value boundaries or specific options of the hyperparameter. For example, an integer type hyperparameter may be marked as having a value range of a certain continuous integer interval, and a string type hyperparameter may be marked as having a value range of a set of several fixed strings. In this way, the type definition and value range corresponding to each hyperparameter are clearly extracted from the configuration information.

[0022] Furthermore, when performing uniform sampling based on the obtained hyperparameter type definitions and value ranges, for integer type hyperparameters, values ​​are selected at equal intervals within their value range to ensure that the selected values ​​are evenly distributed throughout the entire value range and do not concentrate in a certain local area; for floating-point type hyperparameters, values ​​are selected evenly distributed within their value range to avoid values ​​concentrating in a certain area and to ensure coverage of all parts of the value range; for string type hyperparameters, all possible options are selected in the same proportion to ensure that each option has a corresponding number of sampled values, and there will be no situation where an option is over-selected or completely unselected. The sampled values ​​obtained by all hyperparameters in the above manner are combined to form the original parameter set of the hyperparameters.

[0023] Furthermore, when evaluating and screening the original parameter set, a clear evaluation criterion is first set. This criterion can be a quantifiable indicator such as the model's accuracy on the validation set or the convergence speed during training. Each parameter combination in the original parameter set is then applied to the corresponding model training process, and the model training and validation process is executed completely. The evaluation criterion results for each parameter combination are recorded. Then, the results of all parameter combinations are sorted according to the evaluation criterion. For example, if the evaluation criterion is accuracy, the parameter combinations are sorted from high to low accuracy. A preset number of parameter combinations are selected from the top of the sorted list. These selected parameter combinations are the initial hyperparameters of the hyperparameters.

[0024] In general, when parsing the configuration information of the hyperparameter search space, we first obtain the file or data structure that stores the configuration information, identify and extract the description of each hyperparameter one by one, clarify the category or type definition of each hyperparameter, such as integer, floating-point number, string category, etc., and determine the possible value boundaries or fixed option set of each hyperparameter, i.e., the value range. Finally, we obtain the type definition and value range of the hyperparameter from the configuration information.

[0025] In summary, when performing uniform sampling based on the type definition and value range of hyperparameters, a corresponding uniform selection method is adopted for different types of hyperparameters. For integer type hyperparameters, values ​​are selected at equal intervals within their value range; for floating-point type hyperparameters, values ​​are selected evenly distributed within their value range; and for string type hyperparameters, values ​​are selected from all their options in the same proportion. Then, the sampled values ​​obtained by the above method for all hyperparameters are combined to finally obtain the original parameter set of the hyperparameters.

[0026] In general, when evaluating and selecting the original parameter set of hyperparameters, first determine a quantifiable evaluation standard, such as the model's accuracy on the validation set or the convergence speed during training. Then, apply each parameter combination in the original parameter set to the model training and validation process, record the evaluation results corresponding to each parameter combination, and then sort the results of all parameter combinations according to the evaluation standard. Select the preset number of parameter combinations that are ranked first to finally obtain the initial hyperparameters.

[0027] S2. Perform multi-objective evaluation on the initial hyperparameters to obtain comprehensive evaluation information of the initial hyperparameters, wherein the comprehensive evaluation information includes: performance indicators, resource consumption and structural complexity; In this embodiment of the invention, the step of performing multi-objective evaluation on the initial hyperparameters to obtain comprehensive evaluation information of the initial hyperparameters includes: Based on the optimization task execution process of the hyperparameters, the initial hyperparameters are dynamically analyzed to obtain the performance indicators of the initial hyperparameters. Monitor the execution time and content usage of the initial hyperparameters in the optimization task execution process, and generate the resource consumption status of the initial hyperparameters; The structural complexity of the initial hyperparameters is obtained by evaluating the data report after the completion of the optimized task execution process. By integrating the performance indicators, resource consumption, and structural complexity from multiple dimensions, a comprehensive evaluation of the initial hyperparameters is obtained.

[0028] The process of integrating the performance indicators, resource consumption, and structural complexity from multiple dimensions to obtain comprehensive evaluation information for the initial hyperparameters includes: The dimensional differences of the performance indicators are eliminated to obtain the standardized performance values ​​of the initial hyperparameters; The resource consumption situation is adjusted within a certain range to obtain the adjusted resource information of the initial hyperparameters; Based on the correlation between the number and type of the initial hyperparameters, the structural complexity is classified into levels to obtain the complexity level of the initial hyperparameters. A multi-criteria decision analysis is performed on the standardized performance value, the adjusted resource information, and the complexity level to obtain comprehensive evaluation information of the initial hyperparameters. The formula for calculating the comprehensive score in the comprehensive evaluation information is as follows: ; In the formula, The overall score in the comprehensive evaluation information. These are preset trade-off parameters. For, the The standardized performance values ​​of each hyperparameter configuration, For the first The adjusted resource information for each hyperparameter configuration For the first The complexity level of each hyperparameter configuration.

[0029] Specifically, when performing dynamic performance analysis on the initial hyperparameters based on the optimization task execution flow of the hyperparameters, the specific steps of the optimization task execution flow corresponding to the hyperparameters are first identified, including key nodes such as task startup, intermediate running stages, and task termination. At each key node, the impact of the initial hyperparameters on the task running state is tracked in real time. For example, in model training optimization tasks, the prediction bias and fitting degree of the model corresponding to the initial hyperparameters during the training process are tracked. These tracked specific situations are converted into quantifiable indicators, such as prediction accuracy and bias rate. These quantifiable indicators are the performance indicators of the initial hyperparameters.

[0030] Furthermore, when monitoring the execution time and content usage of the initial hyperparameters in the optimization task execution process, a timing operation is started from the moment the initial hyperparameters begin participating in the optimization task execution and stops timing until the optimization task execution process is completely completed. The total duration during this period is recorded, and this total duration is the execution time. At the same time, during the optimization task execution process, the usage of computing resources associated with the initial hyperparameters is tracked in real time, such as the real-time memory usage and storage space usage. These resource usage data are continuously recorded, and the recorded execution time and resource usage data are organized to form structured information containing time data and resource usage data. This structured information is the resource consumption status of the initial hyperparameters.

[0031] Furthermore, when evaluating the data report after the completion of the optimization task execution process to obtain the structural complexity of the initial hyperparameters, the complete data report generated after the completion of the optimization task execution process is first obtained. This report contains information such as the task execution result corresponding to the initial hyperparameters and the composition details of the hyperparameters themselves. Then, the content related to the structure of the initial hyperparameters in the data report is analyzed, such as the composition dimensions of the hyperparameters, the relationship between the dimensions, and whether there is a nesting level. The complexity of the initial hyperparameter structure is judged based on the analysis results. If the hyperparameters have few composition dimensions, no complex relationship between the dimensions, and no nesting level, it is judged as low complexity. If the composition dimensions are many, the dimensions are closely related, and there are multiple nesting levels, it is judged as high complexity. This judgment result is the structural complexity of the initial hyperparameters.

[0032] Furthermore, when integrating the performance indicators, resource consumption, and structural complexity from multiple dimensions to obtain the comprehensive evaluation information of the initial hyperparameters, the importance of the three dimensions in the comprehensive evaluation is first determined. For example, in efficiency-focused optimization tasks, resource consumption is more important than the other two dimensions, while in effect-focused tasks, performance indicators are the most important. Then, according to the determined importance, the information of each dimension is described in detail. For example, the specific values ​​and meanings of each item in the performance indicators are explained first, followed by the time and resource usage data in the resource consumption, and finally the basis for determining structural complexity. These descriptions are then arranged in order of importance to form a document containing complete information from the three dimensions and highlighting the key points. This document is the comprehensive evaluation information of the initial hyperparameters.

[0033] Specifically, when eliminating dimensional differences in the performance indicators, all performance indicators to be processed are first collected, and the original dimension of each indicator is determined. For example, some indicators are measured in "accuracy percentage" and others in "error value". Then, a unified processing method for eliminating dimensional differences is determined, that is, the maximum and minimum values ​​of each performance indicator are first found in the indicator data corresponding to all initial hyperparameters. Then, for the performance indicator value of each initial hyperparameter, the minimum value of the corresponding indicator is subtracted from the value, and then divided by the difference between the maximum and minimum values ​​of the indicator. Through this calculation process, performance indicators with different dimensions are transformed into dimensionless values ​​within the same range. These dimensionless values ​​are the standardized performance values ​​of the initial hyperparameters.

[0034] Furthermore, when adjusting the range of resource consumption, first clarify the specific content included in the resource consumption, which typically includes execution time and content usage. Then, set a unified target adjustment range based on the actual application scenario, for example, adjusting all resource consumption data to the range of 0 to 1. Subsequently, process the execution time and content usage separately: For execution time, find the maximum and minimum values ​​of the execution time corresponding to all initial hyperparameters. Subtract the minimum execution time from the execution time of a certain initial hyperparameter, and then divide by the difference between the maximum and minimum execution times to obtain the adjusted value of the execution time of that initial hyperparameter. For content usage, use the same processing method as for execution time. Subtract the minimum content usage from the content usage of a certain initial hyperparameter, and then divide by the difference between the maximum and minimum content usage to obtain the adjusted value of the content usage of that initial hyperparameter. Integrate the adjusted values ​​of the execution time and content usage of the same initial hyperparameter together to form the overall information, which is the adjusted resource information of the initial hyperparameter.

[0035] Furthermore, when classifying the structural complexity based on the correlation between the number and types of the initial hyperparameters, the specific number of hyperparameters contained in each initial hyperparameter is first counted, and then the correlation between these hyperparameter types is analyzed. For example, it is determined whether there is a dependency relationship between different types of hyperparameters, that is, whether the value of one hyperparameter will affect the effective range of another hyperparameter. Then, a classification standard is set. For example, the case with a small number of hyperparameters and no correlation between types is classified as low complexity level, the case with a medium number of hyperparameters and a small number of simple correlations between types is classified as medium complexity level, and the case with a large number of hyperparameters and a large number of complex correlations between types is classified as high complexity level. Then, the structural complexity of each initial hyperparameter is judged according to the standard, and the judgment result is the complexity level of the initial hyperparameter.

[0036] Furthermore, when performing multi-criteria decision analysis on the standardized performance value, the adjusted resource information, and the complexity level, the weights of the three criteria in the decision are first determined. For example, based on the core requirements of the optimization task, if the task focuses more on performance, the weight of the standardized performance value is set to the highest; if the task focuses more on resource consumption, the weight of the adjusted resource information is set to the highest. Then, the indicators under each criterion are quantified. The standardized performance value is already quantified data. The average value of the adjusted execution time and the adjusted content occupancy value in the adjusted resource information can be taken as the quantification result of the criterion. The complexity level can be assigned corresponding quantification scores according to low, medium, and high. Then, the quantification result of each criterion is multiplied by its corresponding weight to obtain the weighted score of each criterion. Then, the weighted scores of the three criteria are added together to obtain the comprehensive score of each initial hyperparameter. Finally, combining the specific situation of each criterion and the comprehensive score, a document containing the comprehensive score, the performance of each criterion, and the overall evaluation is formed. This document is the comprehensive evaluation information of the initial hyperparameter.

[0037] Furthermore, the sources of the parameters in the formula are as follows: The first one is obtained by calculating using this formula A comprehensive score for each hyperparameter configuration; These are trade-off parameters pre-set based on the actual needs or experience of the hyperparameter optimization task, used to balance the degree of influence of different parts in the formula; From the first The standardized performance values ​​obtained after eliminating dimensional differences in the performance indicators of each hyperparameter configuration; From the first The adjusted resource information is obtained by adjusting the range of resource consumption of each hyperparameter configuration; From the first The relationship between the number and type of hyperparameters in each hyperparameter configuration is used to classify the structural complexity into levels.

[0038] Furthermore, the significance of the formula lies in combining the first... A comprehensive score is calculated based on the standardized performance values ​​of each hyperparameter configuration, adjusted resource information, and complexity level. To balance the impact of the two parts, one is the minimum value of the three indicators, reflecting the weakest link in the hyperparameter configuration across the three aspects; the other is the average value of the three indicators, reflecting the overall average level of the hyperparameter configuration across the three aspects. By combining these two aspects, the comprehensive score can reflect both the shortcomings of the hyperparameter configuration and its overall average performance, thus comprehensively reflecting the [specific aspect]. A comprehensive overview of each hyperparameter configuration.

[0039] In summary, when performing dynamic performance analysis of initial hyperparameters based on the execution process of the hyperparameter optimization task, the specific steps of the optimization task execution process are combined to track the impact of initial hyperparameters on the state at each stage of task execution in real time, transform these impacts into quantifiable information, and finally obtain the performance indicators of initial hyperparameters.

[0040] In summary, when monitoring the execution time and content consumption of initial hyperparameters in the optimization task execution process, the timing starts from the moment the initial hyperparameters participate in the optimization task execution and stops when the optimization task execution process is completely finished to record the execution time. At the same time, the computing resource data consumed by the initial hyperparameters is continuously tracked to obtain the content consumption. After organizing these two types of data, the resource consumption of the initial hyperparameters is generated.

[0041] In general, when evaluating the data report after the completion of the optimization task to obtain the structural complexity of the initial hyperparameters, we first obtain the complete data report generated after the optimization task is completed, analyze the structural content in the report such as the composition dimensions of the initial hyperparameters, the relationship between each dimension and the nesting situation, and judge the complexity of the initial hyperparameter structure based on the analysis results, and finally obtain the structural complexity of the initial hyperparameters.

[0042] In general, when integrating performance indicators, resource consumption, and structural complexity from multiple dimensions to obtain comprehensive evaluation information of initial hyperparameters, the importance of the three dimensions in the comprehensive evaluation is first determined. The specific information of each dimension is then explained in detail according to its importance. Finally, these explanations are integrated into a complete document in order of importance, and the comprehensive evaluation information of the initial hyperparameters is obtained.

[0043] In general, when eliminating dimensional differences in performance indicators to obtain standardized performance values ​​of initial hyperparameters, all performance indicators to be processed and their original dimensions are first identified. Then, the differences caused by different dimensions are eliminated through a unified processing method, and these indicators are converted into dimensionless values, ultimately obtaining the standardized performance values ​​of the initial hyperparameters.

[0044] In general, when adjusting the range of resource consumption to obtain the adjusted resource information of the initial hyperparameter, the execution time and content usage included in the resource consumption are first identified. Then, a unified target adjustment range is set according to the actual application scenario. The execution time and content usage are adjusted in the same way. Finally, the two types of adjusted values ​​of the same initial hyperparameter are integrated to obtain the adjusted resource information of the initial hyperparameter.

[0045] In general, when classifying the structural complexity based on the relationship between the number and types of initial hyperparameters to obtain the complexity level of the initial hyperparameters, we first count the specific number of hyperparameters contained in each initial hyperparameter, then analyze the relationship between these hyperparameter types, then set clear classification criteria, and finally determine the structural complexity of each initial hyperparameter by comparing it with the criteria, thus obtaining the complexity level of the initial hyperparameters.

[0046] In general, when performing multi-criteria decision analysis on standardized performance values, adjusted resource information, and complexity levels to obtain comprehensive evaluation information of initial hyperparameters, the weights of the three criteria in the decision-making process are first determined, then the information under each criterion is quantified, then the weighted score of each criterion is calculated and summed to obtain the comprehensive score, and finally, a complete document is formed by combining the specific circumstances of each criterion with the comprehensive score, thus obtaining the comprehensive evaluation information of the initial hyperparameters.

[0047] S3. Based on the comprehensive evaluation information, an adaptive strategy is used to collect the initial hyperparameters to obtain candidate hyperparameters for the initial hyperparameters; In this embodiment of the invention, the step of adaptively acquiring candidate hyperparameters of the initial hyperparameters based on the comprehensive evaluation information includes: Based on the complexity level in the comprehensive evaluation information, the search requirements for the initial hyperparameters are assessed to obtain a requirement report for the initial hyperparameters. Based on the demand report, select the corresponding target acquisition strategy from the preset acquisition strategies. When it is determined that there is a high demand for exploration, the random acquisition strategy is selected first; when it is determined that it needs to be carefully selected for use, the heuristic acquisition strategy is selected. The monitoring response to the target acquisition strategy is used to make real-time corrections to the comprehensive evaluation information, thereby obtaining new hyperparameters from the initial hyperparameters. Clustering optimization is performed on the new hyperparameters to obtain candidate hyperparameters for the initial hyperparameters.

[0048] Based on the complexity level in the comprehensive evaluation information, the initial hyperparameters are searched for to determine the required parameters, resulting in a requirement report for the initial hyperparameters, including: The complexity level is matched and analyzed with the task mode in the optimization task execution process to obtain the optimization task matching result of the initial hyperparameters; The resource allocation schemes for similar tasks in the optimized task execution process are estimated and analyzed to generate the resource configuration requirements of the initial hyperparameters. Based on the optimization task matching results and the complexity level, the search strategy types and effect evaluations in the optimization task execution process are non-dominatedly sorted to obtain the search direction suggestions for the initial hyperparameters. Based on the suggested search directions and resource configuration requirements, a requirement report for the initial hyperparameters is generated.

[0049] Based on the optimization task matching result and the complexity level, the search strategy type and effect evaluation in the optimization task execution process are non-dominatedly ranked to obtain the search direction suggestions for the initial hyperparameters, including: The search strategy types in the optimized task execution process are evaluated and backtracked to obtain historical performance data of the optimized task execution process. The historical performance data is quantitatively reconstructed to obtain the dominance relationship matrix of the optimized task execution process; Prioritize the strategy distribution features in the dominance relationship matrix to obtain search direction suggestions for the initial hyperparameters.

[0050] Specifically, based on the complexity level in the comprehensive evaluation information, when assessing the search needs for the initial hyperparameters and obtaining the initial hyperparameter requirement report, the complexity level corresponding to each initial hyperparameter is first obtained. Simultaneously, the overall performance of the initial hyperparameters is analyzed by combining the standardized performance values ​​and adjusted resource information from the comprehensive evaluation information. If the complexity level is high, and the standardized performance values ​​of different initial hyperparameters differ significantly, and the adjusted resource information fluctuates noticeably, it indicates that a stable range of high-quality parameters has not yet been formed within the current hyperparameter search space, and the search scope needs to be expanded to explore potential high-quality parameters. If the complexity level is low or medium, and some initial hyperparameters already possess high standardized performance values ​​and reasonable adjusted resource information, it indicates that there is already a foundation of high-quality parameters, and there is no need to excessively expand the search scope; instead, it is necessary to focus on screening near existing high-quality parameters. The above complexity level analysis, the judgment results combined with other evaluation information, the search requirement conclusions of "needs exploration" or "needs selection," and suggestions for subsequent data collection directions are organized into a structured document, which is the initial hyperparameter requirement report.

[0051] Furthermore, based on the demand report, a corresponding target acquisition strategy is selected from the preset acquisition strategies. When a high exploration demand is identified, a random acquisition strategy is prioritized. When a carefully selected approach is required, a heuristic acquisition strategy is selected. First, the search demand conclusion explicitly stated in the demand report is read. If the demand report indicates "exploration required," then a high exploration demand is identified, and a random acquisition strategy is prioritized. This strategy selects new hyperparameter combinations unbiasedly from the hyperparameter search space corresponding to the initial hyperparameters, without relying on the comprehensive evaluation performance of the existing initial hyperparameters, ensuring coverage of different areas of the search space to uncover potential high-quality parameters. If the demand report indicates "careful selection required," then a carefully selected approach is selected, and a heuristic acquisition strategy is chosen. This strategy extracts the value range characteristics of high-quality parameters based on the initial hyperparameters with high comprehensive scores in the comprehensive evaluation information, and selects new hyperparameter combinations within the sub-range of the search space corresponding to these characteristics, focusing on the existing high-quality parameter base to screen for better parameters. The acquisition strategy determined in the above manner is the target acquisition strategy.

[0052] Furthermore, when monitoring the response to the target acquisition strategy and real-time correction of the comprehensive evaluation information to obtain new hyperparameters from the initial hyperparameters, real-time monitoring of the target acquisition strategy execution process is initiated. If a random acquisition strategy is used, it monitors whether the new parameter combination evenly covers each sub-region of the search space. If a sub-region is repeatedly selected or not selected, the selection direction is adjusted promptly to ensure coverage. If a heuristic acquisition strategy is used, it monitors whether the new parameter combination revolves around the initial hyperparameters with higher comprehensive scores. If it deviates too far from the range of high-quality parameters, the selection range is corrected promptly to ensure focus. The acquired new parameter combination is substituted into the hyperparameter optimization task execution process to obtain its temporary performance indicators, temporary resource consumption, and temporary structural complexity. This is compared with the evaluation data of the original initial hyperparameters in the comprehensive evaluation information. If the temporary evaluation data of the new parameters differs significantly from the original data, the evaluation description of the corresponding search region in the comprehensive evaluation information is updated, completing the real-time correction. Based on the corrected comprehensive evaluation information and the target acquisition strategy execution requirements, the final new parameter combination is determined; these new parameter combinations are the new hyperparameters of the initial hyperparameters.

[0053] Furthermore, when performing cluster optimization on the new hyperparameters to obtain candidate hyperparameters for the initial hyperparameters, the clustering criteria are first determined: based on the comprehensive evaluation information of the new hyperparameters, new hyperparameters with standardized performance values ​​in the same range, adjusted resource information differences within a preset small range, and the same complexity level are grouped into the same cluster. Specifically, the standardized performance values ​​of all new hyperparameters are first divided into several fixed ranges according to their value, and the adjusted resource information is then divided into ranges according to a reasonable range. The standardized performance value range, adjusted resource information range, and complexity level of each new hyperparameter are then assessed one by one, and new hyperparameters that match all three are grouped into the same cluster group. After clustering, each cluster group is evaluated: the average comprehensive score of all new hyperparameters within the group is calculated, and the average comprehensive score of all cluster groups is compared. The top few cluster groups with the highest average comprehensive score are selected. Within the selected cluster groups, the comprehensive score of each new hyperparameter is compared one by one and sorted from high to low, and the new hyperparameters with the highest ranking are selected. These new hyperparameters, filtered by clustering, are organized into a set, which is the candidate hyperparameters of the initial hyperparameters.

[0054] Specifically, when matching the complexity level with the task modes in the optimization task execution process to obtain the optimization task matching result of the initial hyperparameters, the task modes included in the optimization task execution process are first identified. Common task modes include fast verification mode, high-precision optimization mode, and resource-constrained mode. Then, the appropriate complexity level corresponding to different task modes is determined—for example, fast verification mode is adapted to low complexity level, high-precision optimization mode is adapted to high complexity level, and resource-constrained mode is adapted to medium-low complexity level. Subsequently, the complexity level of the initial hyperparameters is compared with the task mode adopted by the current optimization task execution process to determine whether the complexity level meets the adaptation requirements of the task mode. If it meets the requirements, "matching successful" and the adaptation basis are recorded; if it does not meet the requirements, "matching needs adjustment" and the adjustment direction are recorded. These judgment results and basis are organized into structured information, which is the optimization task matching result of the initial hyperparameters.

[0055] Furthermore, when estimating and analyzing the resource allocation schemes of similar tasks in the optimized task execution process and generating the resource configuration requirements of the initial hyperparameters, firstly, similar tasks with the same type and consistent goals as the current task are collected from the historical records of the optimized task execution process. The hyperparameter complexity levels and resource allocation schemes corresponding to these similar tasks are extracted, and the correspondence between complexity levels and resource allocation in similar tasks is analyzed—for example, in similar tasks, low-complexity hyperparameters have a short execution time quota and a low memory usage limit, while high-complexity hyperparameters have a long execution time quota and a high memory usage limit. Then, combined with the complexity level of the current initial hyperparameters and referring to the resource allocation standards of hyperparameters of the same complexity level in similar tasks, the resource allocation details are adjusted to suit the current task. The adjusted resource allocation requirements for execution time, memory, storage, etc., are compiled into a clear document, which is the resource configuration requirement of the initial hyperparameters.

[0056] Furthermore, based on the optimized task matching result and the complexity level, a non-dominated ranking of the search strategy types and effect evaluations in the optimized task execution process is performed to obtain the search direction suggestions for the initial hyperparameters. First, the available search strategy types in the optimized task execution process are listed, including random search strategy, heuristic search strategy, and grid search strategy. Simultaneously, two core dimensions of effect evaluation are determined: search efficiency and search accuracy. Then, a ranking rule is set based on the optimized task matching result and complexity level—if the optimized task matching result is "successfully matched" and the complexity level is low, search efficiency is prioritized; if the matching result is "successfully matched" and the complexity level is high, search accuracy is prioritized; if the matching result is "matching needs adjustment," efficiency and accuracy need to be balanced to adapt to the adjusted task requirements. Subsequently, each search strategy type was scored according to its effectiveness evaluation dimensions. A strategy that was no less efficient and accurate than another strategy, and had at least one dimension higher than the other strategy, was judged as "dominant". The strategies were then ranked according to the non-dominant rules. Based on the ranking results, the preferred strategy type and the focus of effect optimization were determined. This content was compiled into a guiding document, which is the search direction suggestion for the initial hyperparameters.

[0057] Furthermore, when generating the initial hyperparameter requirement report by integrating the search direction suggestions and resource configuration requirements, the core content of the search direction suggestions is first extracted, including the preferred search strategy type, the focus of performance evaluation, and the adjustment direction for adapting to the optimization task. Then, the core content of the resource configuration requirements is extracted, including execution time quotas, memory usage limits, storage space requirements, and their corresponding basis. Subsequently, the two core contents are logically integrated: first, the background of the optimization task is explained; then, the specific terms of the search direction suggestions and the specific indicators of the resource configuration requirements are presented sequentially; finally, overall execution suggestions are added. All the integrated content forms a complete structured report, which is the initial hyperparameter requirement report.

[0058] Specifically, when performing task evaluation backtracking on the search strategy types in the optimization task execution process to obtain historical performance data of the optimization task execution process, firstly, identify all search strategy types included in the current optimization task execution process. Then, from the history of hyperparameter optimization tasks, select historical tasks that are consistent with the current task mode and have similar initial hyperparameter complexity levels. Next, extract specific execution data corresponding to each search strategy type from these selected historical tasks—including the time consumed to complete the entire search process, the comprehensive score corresponding to the hyperparameters found during the search process, and whether the resource consumption during the execution process meets the resource configuration requirements at that time. Then, classify and organize these data according to search strategy type. Each strategy type corresponds to a set of specific records containing efficiency, accuracy, and resource adaptability. These classified record sets are the historical performance data of the optimization task execution process.

[0059] Furthermore, when quantifying and reconstructing the historical performance data to obtain the dominance relationship matrix of the optimized task execution process, two core dimensions of quantification and reconstruction are first determined: search efficiency and search accuracy. For the efficiency data of each search strategy type in the historical performance data, fixed levels are divided according to the length of time to complete the search, and a corresponding score is assigned to each level. For the accuracy data, fixed levels are divided according to the comprehensive score of the found hyperparameters, and a corresponding score is assigned to each level. Then, all search strategy types are compared in pairs. If the efficiency score of strategy X is not lower than that of strategy Y, and the accuracy score of strategy X is not lower than that of strategy Y, and at least one dimension's score is higher than that of strategy Y, then strategy X is determined to dominate strategy Y. Then, a matrix with search strategy types as rows and search strategy types as columns is constructed. In the matrix, the element positions corresponding to each row strategy and column strategy are marked as "dominant" if the row strategy dominates the column strategy, and marked as "non-dominant" if they do not dominate. This matrix with completed dominance relationship marking is the dominance relationship matrix of the optimized task execution process.

[0060] Furthermore, when prioritizing the strategy distribution features in the dominance relationship matrix to obtain the search direction suggestions for the initial hyperparameters, the number of times each search strategy type in the dominance relationship matrix is ​​dominated by other strategies is first counted. The fewer times a strategy is dominated, the more stable and excellent its overall performance in terms of efficiency and accuracy is. All strategies are initially prioritized according to the number of times they are dominated, from least to most. Then, the initial ranking is adjusted based on the optimization task matching result and complexity level: if the optimization task matching result is "matched successfully" and the complexity level is low, the strategy with the higher efficiency score is prioritized in the initial ranking; if the optimization task matching result is "matched successfully" and the complexity level is high, the strategy with the higher accuracy score is prioritized; if the optimization task matching result is "matching needs adjustment", the strategy with a relatively balanced efficiency and accuracy score is prioritized. Finally, the adjusted priority ranking result, the core advantages of each strategy, and the specific reasons for adapting to the current task are compiled into a structured guidance document, which is the search direction suggestion for the initial hyperparameters.

[0061] In general, when assessing the search needs for initial hyperparameters based on the complexity level in the comprehensive evaluation information to obtain a requirement report, the standardized performance values ​​and adjusted resource information in the comprehensive evaluation information are combined to analyze the overall performance of the initial hyperparameters. This determines whether it is necessary to expand the search scope to explore potential high-quality parameters or focus on screening near existing high-quality parameters. The complexity level analysis results, search requirement conclusions, and suggestions for subsequent collection directions are then compiled into a structured document, which is the requirement report for initial hyperparameters.

[0062] In general, when selecting a target acquisition strategy from the preset acquisition strategies based on the requirements report, the search requirements conclusions in the requirements report are read first. If it is determined that there is a high demand for exploration, the random acquisition strategy is selected first. If it is determined that it needs to be carefully selected, the heuristic acquisition strategy is selected. The acquisition strategy determined in this way is the target acquisition strategy of the initial hyperparameters.

[0063] In summary, when monitoring the target acquisition strategy and correcting the comprehensive evaluation information in real time to obtain new hyperparameters, the execution process of the target acquisition strategy is monitored in real time. The acquired new parameter combinations are substituted into the optimization task execution process to obtain temporary evaluation data. The relevant descriptions are corrected by comparing with the original comprehensive evaluation information. Then, the final new parameter combinations are determined based on the corrected evaluation information. These new parameter combinations are the new hyperparameters of the initial hyperparameters.

[0064] In general, when clustering new hyperparameters to obtain candidate hyperparameters, the comprehensive evaluation information of the new hyperparameters is used as the basis. New hyperparameters with the same standardized performance value range, adjusted resource information range, and complexity level are grouped into the same cluster group. The average comprehensive score of each cluster group is calculated, and the group with the highest score is selected. Then, within the selected group, the new hyperparameters with the highest comprehensive scores are selected. These screened new hyperparameters are organized into a set, which is the candidate hyperparameters for the initial hyperparameters.

[0065] In general, when matching complexity levels with task patterns in the optimized task execution process to obtain the optimization task matching results of the initial hyperparameters, we first identify the task patterns included in the optimized task execution process and the complexity levels adapted to each pattern. Then, we compare the complexity levels of the initial hyperparameters with the current task patterns to determine whether they meet the adaptation requirements. The judgment results and adjustment directions are organized into structured information, which is the optimization task matching result of the initial hyperparameters.

[0066] In general, when estimating and analyzing the resource allocation schemes of similar tasks in the task execution process to generate the resource configuration requirements of the initial hyperparameters, similar tasks that are consistent with the current task type and goal are collected, the correspondence between the hyperparameter complexity level and the resource allocation scheme in these tasks is extracted, the resource allocation details are adjusted in combination with the complexity level of the current initial hyperparameters, and the adjusted resource requirements are organized into a clear document, which is the resource configuration requirements of the initial hyperparameters.

[0067] In summary, when performing non-dominated sorting of search strategy types and effect evaluations in the optimization task execution process based on the optimization task matching results and complexity level to obtain initial hyperparameter search direction suggestions, the available search strategy types and effect evaluation dimensions are listed. Sorting rules are set according to matching results and complexity level, strategies are sorted according to non-dominated rules, and the sorting results are adjusted in combination with task requirements. The adjusted sorting, strategy advantages, and adaptation reasons are compiled into a guidance document, which serves as the initial hyperparameter search direction suggestion.

[0068] In summary, when generating the initial hyperparameter requirement report by combining the search direction suggestions and resource allocation requirements, the strategy selection, effect focus and adjustment direction in the search direction suggestions and the resource indicators and basis in the resource allocation requirements are extracted. These contents are logically integrated to supplement the overall execution suggestions and form a complete structured report, which is the initial hyperparameter requirement report.

[0069] In general, when evaluating and backtracking the search strategy types in the optimization task execution process to obtain historical performance data, we first identify all the search strategy types included in the current process, then select historical tasks that are consistent with the current task mode and have similar initial hyperparameter complexity levels from the historical records, extract the search efficiency, search accuracy, and resource adaptability data corresponding to each strategy type, and then classify and organize these data according to strategy type. The resulting set of classified records is the historical performance data of the optimization task execution process.

[0070] In general, when quantifying and reconstructing historical performance data to obtain the dominance relationship matrix, search efficiency and search accuracy are determined as the core quantification dimensions. Efficiency data is graded and assigned scores according to the length of time, and accuracy data is graded and assigned scores according to the overall score. Then, all strategies are compared pairwise to determine the dominance relationship. A matrix is ​​then constructed with strategy type as the row and column, and "dominant" or "non-dominant" is marked at the corresponding position. The resulting marked matrix is ​​the dominance relationship matrix for optimizing the task execution process.

[0071] In general, when prioritizing the strategy distribution features in the dominance relationship matrix to obtain search direction suggestions, the number of times each strategy is dominated by other strategies is counted, and the strategies are initially sorted from least to most frequent. Then, the sorting is adjusted by combining the optimization task matching results and complexity level. Finally, the adjusted sorting results, the core advantages of the strategies, and the reasons for their suitability are compiled into a guidance document, which is the initial hyperparameter search direction suggestion.

[0072] S4. Determine the local search space corresponding to the candidate hyperparameters based on the hyperparameter configuration in the candidate hyperparameters; In this embodiment of the invention, determining the local search space corresponding to the candidate hyperparameters based on the hyperparameter configurations among the candidate hyperparameters includes: Analyze the hyperparameter configurations in the candidate hyperparameters to generate performance gradient information for the candidate hyperparameters; Based on the performance gradient information, determine the sensitive dimension of the hyperparameter configuration among the candidate hyperparameters; Based on the current distribution characteristics of the sensitive dimension, the local search space corresponding to the candidate hyperparameter is determined.

[0073] Specifically, when analyzing the hyperparameter configurations among the candidate hyperparameters and generating the performance gradient information of the candidate hyperparameters, the specific values ​​of each candidate hyperparameter configuration and its corresponding comprehensive score in the comprehensive evaluation information are first extracted. Then, all candidate hyperparameter configurations are sorted according to the similarity of their hyperparameter values. The difference in the comprehensive score between two adjacent configurations is compared with the difference in their hyperparameter values, and the difference in their comprehensive scores is calculated. At the same time, the difference in the value of each hyperparameter dimension is recorded. Through this comparison, the influence trend of the value change of each hyperparameter dimension on the change of the comprehensive score is clarified. These influence trends are classified and organized according to the hyperparameter dimensions to form a document containing the correlation between the value change of each dimension and the performance change. This document is the performance gradient information of the candidate hyperparameters.

[0074] Furthermore, when determining the sensitive dimensions of hyperparameter configuration among the candidate hyperparameters based on the performance gradient information, first list all hyperparameter dimensions included in the candidate hyperparameters, and then analyze the performance of each dimension in the performance gradient information one by one. For each dimension, the magnitude of the change in the comprehensive score when the value of the dimension changes is statistically analyzed. If a small change in the value of a certain dimension leads to a large increase or decrease in the comprehensive score, it indicates that the dimension has a high degree of influence on the performance of the candidate hyperparameters. If a large change in the value of a certain dimension does not cause significant fluctuations in the comprehensive score, it indicates that the dimension has a low degree of influence on the performance. All hyperparameter dimensions with high influence are selected, and these selected dimensions are the sensitive dimensions of hyperparameter configuration among the candidate hyperparameters.

[0075] Furthermore, when determining the local search space corresponding to the candidate hyperparameter based on the current distribution characteristics of the sensitive dimension, the specific values ​​of all candidate hyperparameters on the sensitive dimension are first collected, and the distribution characteristics of these values ​​are analyzed—including the concentrated range of all values ​​on the sensitive dimension, the maximum and minimum values, and the density of values. Then, the range of the local search space is set according to the distribution characteristics: taking the concentrated range of the values ​​of the sensitive dimension as the core, the range boundary is extended to the outside of the maximum and minimum values ​​by a certain margin. For non-sensitive dimensions, their value range is kept consistent with the value range of that dimension in the candidate hyperparameters. The expanded range of all sensitive dimensions is integrated with the original range of non-sensitive dimensions to form the complete value space, which is the local search space corresponding to the candidate hyperparameter.

[0076] In general, when analyzing the hyperparameter configurations in the candidate hyperparameters to generate performance gradient information for the candidate hyperparameters, the specific values ​​of each candidate hyperparameter configuration and its corresponding comprehensive score are first extracted. All candidate hyperparameter configurations are sorted according to the similarity of their values, and the differences in comprehensive scores and hyperparameter values ​​of adjacent configurations are compared one by one to clarify the influence trend of changes in the value of each hyperparameter dimension on the change in the comprehensive score. These influence trends are then classified and organized by dimension to form a document containing the correlation between the value of each dimension and performance. This document is the performance gradient information of the candidate hyperparameters.

[0077] In general, when determining the sensitive dimensions of hyperparameter configuration among candidate hyperparameters based on performance gradient information, we first list all the hyperparameter dimensions included in the candidate hyperparameters and analyze the performance of each dimension in the performance gradient information one by one—statistically calculating the change in the comprehensive score when the value of the dimension changes. If a small change in the dimension value leads to a large fluctuation in the comprehensive score, it indicates that it has a high degree of impact on performance. All such hyperparameter dimensions with a high degree of impact are selected, and these selected dimensions are the sensitive dimensions of hyperparameter configuration among the candidate hyperparameters.

[0078] In general, when determining the local search space corresponding to candidate hyperparameters based on the current distribution characteristics of the sensitive dimension, we first collect the specific values ​​of all candidate hyperparameters on the sensitive dimension, analyze the concentration range, maximum value, minimum value and density of these values, expand the value range of the sensitive dimension to both sides with the concentration range as the core, and keep the original value range of the non-sensitive dimension. The expanded range of the sensitive dimension is integrated with the original range of the non-sensitive dimension to form the complete value space, which is the local search space corresponding to the candidate hyperparameter.

[0079] S5. Based on the local search space, the hyperparameter configuration is finely adjusted to obtain the hyperparameter configuration variants of the candidate hyperparameters; In this embodiment of the invention, the step of finely adjusting the hyperparameter configuration based on the local search space to obtain hyperparameter configuration variants of the candidate hyperparameters includes: Based on the local search space, a search subspace centered on the hyperparameter configuration is constructed; The sensitivity differences of the hyperparameter configuration are serialized and parsed to obtain the coordinated adjustment sequence of the hyperparameter configuration; By extrapolating the performance gradient trend of the coordinated adjustment sequence, we obtain the hyperparameter configuration variants of the candidate hyperparameters.

[0080] Specifically, when constructing a search subspace centered on the hyperparameter configuration based on the local search space, the complete value range of each hyperparameter dimension in the local search space is first extracted. Then, the specific values ​​of the current hyperparameter configuration in each dimension are determined, and these specific values ​​are used as the central benchmark for constructing the subspace. For each hyperparameter dimension, with the central benchmark value as the midpoint, a smaller value interval is defined within the local search space value range of that dimension. The boundary of the interval must be controlled within the local search space range, and the values ​​within the interval must cover the region near the central benchmark where there may be better performance. This avoids the interval being too large, which would lead to a decrease in search accuracy, and also avoids the interval being too small, which would miss potential optimal values. All the small value intervals of all hyperparameter dimensions are integrated to form a complete value space centered on the original hyperparameter configuration and with a focused range. This space is the search subspace centered on the hyperparameter configuration.

[0081] Furthermore, when serializing and parsing the sensitivity differences of the hyperparameter configuration to obtain the collaborative adjustment sequence of the hyperparameter configuration, the previously determined hyperparameter configuration sensitivity dimensions and their sensitivity levels are reviewed to clarify the sensitivity ranking of each dimension. Then, the correlation between different sensitivity dimensions is analyzed. For example, adjusting a high-sensitivity dimension may change the optimal value range of another sensitivity dimension; dimensions with such dependencies should be marked as correlation groups. Following the principle of "high-sensitivity dimensions take precedence over low-sensitivity dimensions, and high-sensitivity dimensions within correlation groups drive low-sensitivity dimensions," all hyperparameter dimensions or dimension correlation groups are arranged from highest to lowest adjustment priority, forming a step list containing a clear adjustment order and a description of the correlation relationships. This list constitutes the collaborative adjustment sequence of the hyperparameter configuration.

[0082] Furthermore, when extrapolating the performance gradient trend of the collaborative adjustment sequence to obtain the hyperparameter configuration variants of the candidate hyperparameters, the performance gradient information corresponding to each hyperparameter dimension in the collaborative adjustment sequence is first retrieved to clarify the performance gradient trend of each dimension—for example, when the value of a certain dimension increases, the comprehensive score of the hyperparameter configuration continues to rise; when the value of a certain dimension decreases to a specific range, the comprehensive score reaches a peak and then begins to decline. According to the priority order of the collaborative adjustment sequence, the values ​​of each dimension are fine-tuned: for dimensions where increasing the value increases performance, the value is appropriately increased along the upward trend within the value range of that dimension in the search subspace; for dimensions with a peak range, the value is adjusted to a position closer to the peak range within the value range of that dimension in the search subspace. After each dimension is adjusted, the adjusted hyperparameter combination is substituted into the optimization task for a brief verification, and its comprehensive score is recorded. If the score is better than the original hyperparameter configuration, the adjusted value is retained. After all dimensions have been adjusted sequentially, all adjusted value combinations that have achieved a better overall score than the original hyperparameter configuration in the validation are collected. These combinations are the hyperparameter configuration variants of the candidate hyperparameters.

[0083] In general, when constructing a search subspace centered on hyperparameter configuration based on the local search space, the complete value range of each hyperparameter dimension in the local search space is first extracted, the specific value of the current hyperparameter configuration in each dimension is determined and used as the central benchmark, and a smaller value interval is defined for each dimension within the value range of the local search space. Then, the small value intervals of all hyperparameter dimensions are integrated to form a focused complete value space, which is the search subspace centered on hyperparameter configuration.

[0084] In general, when serializing and parsing the sensitivity differences of hyperparameter configurations to obtain the collaborative adjustment sequence, we first review the previously determined hyperparameter configuration sensitivity dimensions and their sensitivity levels, analyze the correlation between different sensitivity dimensions and mark the dependent correlation groups, and arrange all hyperparameter dimensions or dimension correlation groups in descending order of adjustment priority according to the principle of "high sensitivity dimensions take precedence over low sensitivity dimensions, and high sensitivity dimensions drive low sensitivity dimensions within correlation groups". This forms a list of steps containing a clear adjustment order and a description of the correlation relationship. This list is the collaborative adjustment sequence of hyperparameter configurations.

[0085] In summary, when extrapolating the performance gradient trend of the coordinated adjustment sequence to obtain the hyperparameter configuration variants of the candidate hyperparameters, the performance gradient information corresponding to each hyperparameter dimension in the coordinated adjustment sequence is retrieved first. The values ​​of each dimension are fine-tuned according to the priority order of the coordinated adjustment sequence. After each adjustment, the new parameter combination is substituted into the optimization task for brief verification and the comprehensive score is recorded. The adjusted values ​​with scores better than the original configuration are retained. After all dimensions are adjusted, all adjusted value combinations that are better than the original hyperparameter configuration are collected. These combinations are the hyperparameter configuration variants of the candidate hyperparameters.

[0086] S6. Screen and evaluate the hyperparameter configuration variants to obtain the optimized hyperparameters of the candidate hyperparameters.

[0087] In this embodiment of the invention, the step of screening and evaluating the hyperparameter configuration variants to obtain the optimized hyperparameters of the candidate hyperparameters includes: Obtain the performance data of the hyperparameter configuration variant; The performance data is quantified using the super-volume index to obtain the super-volume index evaluation result. Based on the evaluation results of the hypervolume index, the optimal parameter configuration variant is selected as the optimized hyperparameter of the candidate hyperparameter.

[0088] Specifically, when obtaining the performance data of the hyperparameter configuration variants, each hyperparameter configuration variant is substituted into the corresponding hyperparameter optimization task execution process, and the entire training and validation process of the task is executed. During the execution, the key performance indicators corresponding to each variant are recorded in real time. These indicators include the prediction accuracy of the model on the validation set, the convergence speed during training, and the resource utilization efficiency during task execution. All key performance indicators of each variant are organized into a structured data table in a unified format, and this table is the performance data of the hyperparameter configuration variant.

[0089] Furthermore, when quantifying the performance data using the hypervolume index to obtain the hypervolume index evaluation result, the reference point for calculating the hypervolume index is first determined. The reference point is determined as follows: in the performance data of all hyperparameter configuration variants, the worst value of each key performance indicator is found, and these worst values ​​are combined to form a multi-dimensional reference point. Then, the dimensions of the performance space are defined, with each key performance indicator corresponding to one dimension of the performance space. The performance data of all hyperparameter configuration variants are mapped to this multi-dimensional performance space, with each variant corresponding to a data point in the performance space. Next, the volume of the multi-dimensional region formed between each variant data point and the reference point is calculated. The calculation method is: for each performance dimension, the performance indicator value of the variant is subtracted from the value of the corresponding dimension of the reference point to obtain the length of each dimension. Then, the lengths of all dimensions are multiplied to obtain the hypervolume value corresponding to the variant. Finally, the hypervolume values ​​of all hyperparameter configuration variants are organized into a list containing variant identifiers and corresponding hypervolume values. This list is the hypervolume index evaluation result.

[0090] Furthermore, based on the hypervolume index evaluation results, when selecting the optimal parameter configuration variant as the optimized hyperparameter of the candidate hyperparameter, the hypervolume value corresponding to each hyperparameter configuration variant in the hypervolume index evaluation results is first read. The larger the hypervolume value, the wider the high-quality performance region covered by the variant in the multidimensional performance space, and the better its performance is compared with the reference point and other variants with smaller hypervolume values. The hypervolume values ​​of all variants are compared, and the hyperparameter configuration variant with the largest hypervolume value is selected. If there are multiple variants with the same and largest hypervolume values, the key performance index details of these variants are further compared—the variant that performs better in the core performance index of the optimization task is given priority. The hyperparameter configuration variant determined in the above manner is the optimized hyperparameter of the candidate hyperparameter.

[0091] In general, when obtaining performance data of hyperparameter configuration variants, each hyperparameter configuration variant is substituted into the corresponding hyperparameter optimization task execution process, and the training and validation phases are executed completely. The key performance indicators of each variant are recorded in real time, and then all key performance indicators are organized into a structured data table in a unified format. This table is the performance data of the hyperparameter configuration variants.

[0092] In general, when quantifying performance data using hypervolume metrics to obtain hypervolume metric evaluation results, the worst combination of key performance metrics for all variants is first used as a multidimensional reference point. The performance data of each variant is then mapped to a multidimensional performance space composed of key metrics to form data points. Next, the volume of the multidimensional region formed by each data point and the reference point is calculated. Finally, the hypervolume values ​​of all variants are compiled into a labeled list, which is the hypervolume metric evaluation result.

[0093] In summary, when selecting the optimal parameter configuration variant as the optimized hyperparameter of the candidate hyperparameter based on the hypervolume index evaluation results, the hypervolume values ​​of all variants are compared, and the variant with the largest hypervolume value is selected first. If there are multiple variants with the largest values, their performance on the core performance index of the optimization task is further compared, and the optimal variant is finally determined as the optimized hyperparameter of the candidate hyperparameter.

[0094] In the several embodiments provided by this invention, it should be understood that the disclosed method can be implemented in other ways.

[0095] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0096] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, and technology that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An automated hyperparameter optimization method, characterized in that, The method includes: S1. Dynamically initialize the defined hyperparameter search space to obtain the initial hyperparameters of the hyperparameter search space; S2. Perform multi-objective evaluation on the initial hyperparameters to obtain comprehensive evaluation information of the initial hyperparameters, wherein the comprehensive evaluation information includes: performance indicators, resource consumption and structural complexity; S3. Based on the comprehensive evaluation information, an adaptive strategy is used to collect the initial hyperparameters to obtain candidate hyperparameters for the initial hyperparameters; S4. Determine the local search space corresponding to the candidate hyperparameters based on the hyperparameter configuration in the candidate hyperparameters; S5. Based on the local search space, the hyperparameter configuration is finely adjusted to obtain the hyperparameter configuration variants of the candidate hyperparameters; S6. Screen and evaluate the hyperparameter configuration variants to obtain the optimized hyperparameters of the candidate hyperparameters.

2. The automated hyperparameter optimization method as described in claim 1, characterized in that, The dynamic initialization of the defined hyperparameter search space to obtain the initial hyperparameters of the hyperparameter search space includes: The configuration information of the hyperparameter search space is parsed to obtain the type definition and value range of the hyperparameters; Based on the type definition and value range, the hyperparameters are uniformly sampled to obtain the original parameter set of the hyperparameters; The original parameter set is evaluated and filtered to obtain the initial hyperparameters of the hyperparameters.

3. The automated hyperparameter optimization method as described in claim 1, characterized in that, The multi-objective evaluation of the initial hyperparameters to obtain comprehensive evaluation information of the initial hyperparameters includes: Based on the optimization task execution process of the hyperparameters, the initial hyperparameters are dynamically analyzed to obtain the performance indicators of the initial hyperparameters. Monitor the execution time and content usage of the initial hyperparameters in the optimization task execution process, and generate the resource consumption status of the initial hyperparameters; The structural complexity of the initial hyperparameters is obtained by evaluating the data report after the completion of the optimized task execution process. By integrating the performance indicators, resource consumption, and structural complexity from multiple dimensions, a comprehensive evaluation of the initial hyperparameters is obtained.

4. The automated hyperparameter optimization method as described in claim 3, characterized in that, The process of integrating the performance indicators, resource consumption, and structural complexity from multiple dimensions to obtain comprehensive evaluation information for the initial hyperparameters includes: The dimensional differences of the performance indicators are eliminated to obtain the standardized performance values ​​of the initial hyperparameters; The resource consumption situation is adjusted within a certain range to obtain the adjusted resource information of the initial hyperparameters; Based on the correlation between the number and type of the initial hyperparameters, the structural complexity is classified into levels to obtain the complexity level of the initial hyperparameters. A multi-criteria decision analysis is performed on the standardized performance value, the adjusted resource information, and the complexity level to obtain comprehensive evaluation information of the initial hyperparameters. The formula for calculating the comprehensive score in the comprehensive evaluation information is as follows: ； In the formula, The overall score in the comprehensive evaluation information. These are preset trade-off parameters. For, the The standardized performance values ​​of each hyperparameter configuration, For the first The adjusted resource information for each hyperparameter configuration For the first The complexity level of each hyperparameter configuration.

5. The automated hyperparameter optimization method as described in claim 1, characterized in that, Based on the comprehensive evaluation information, an adaptive strategy is used to collect candidate hyperparameters from the initial hyperparameters, including: Based on the complexity level in the comprehensive evaluation information, the search requirements for the initial hyperparameters are assessed to obtain a requirement report for the initial hyperparameters. Based on the demand report, select the corresponding target acquisition strategy from the preset acquisition strategies. When it is determined that there is a high demand for exploration, the random acquisition strategy is selected first; when it is determined that it needs to be carefully selected for use, the heuristic acquisition strategy is selected. The monitoring response to the target acquisition strategy is used to make real-time corrections to the comprehensive evaluation information, thereby obtaining new hyperparameters from the initial hyperparameters. Clustering optimization is performed on the new hyperparameters to obtain candidate hyperparameters for the initial hyperparameters.

6. The automated hyperparameter optimization method as described in claim 5, characterized in that, Based on the complexity level in the comprehensive evaluation information, the initial hyperparameters are searched for to determine the required parameters, resulting in a requirement report for the initial hyperparameters, including: The complexity level is matched and analyzed with the task mode in the optimization task execution process to obtain the optimization task matching result of the initial hyperparameters; The resource allocation schemes for similar tasks in the optimized task execution process are estimated and analyzed to generate the resource configuration requirements of the initial hyperparameters. Based on the optimization task matching results and the complexity level, the search strategy types and effect evaluations in the optimization task execution process are non-dominatedly sorted to obtain the search direction suggestions for the initial hyperparameters. Based on the suggested search directions and resource configuration requirements, a requirement report for the initial hyperparameters is generated.

7. The automated hyperparameter optimization method as described in claim 6, characterized in that, Based on the optimization task matching result and the complexity level, the search strategy type and effect evaluation in the optimization task execution process are non-dominatedly ranked to obtain the search direction suggestions for the initial hyperparameters, including: The search strategy types in the optimized task execution process are evaluated and backtracked to obtain historical performance data of the optimized task execution process. The historical performance data is quantitatively reconstructed to obtain the dominance relationship matrix of the optimized task execution process; Prioritize the strategy distribution features in the dominance relationship matrix to obtain search direction suggestions for the initial hyperparameters.

8. The automated hyperparameter optimization method as described in claim 1, characterized in that, The step of determining the local search space corresponding to the candidate hyperparameters based on the hyperparameter configurations in the candidate hyperparameters includes: Analyze the hyperparameter configurations in the candidate hyperparameters to generate performance gradient information for the candidate hyperparameters; Based on the performance gradient information, determine the sensitive dimension of the hyperparameter configuration among the candidate hyperparameters; Based on the current distribution characteristics of the sensitive dimension, the local search space corresponding to the candidate hyperparameter is determined.

9. The automated hyperparameter optimization method as described in claim 1, characterized in that, The process of fine-tuning the hyperparameter configuration based on the local search space to obtain hyperparameter configuration variants for the candidate hyperparameters includes: Based on the local search space, a search subspace centered on the hyperparameter configuration is constructed; The sensitivity differences of the hyperparameter configuration are serialized and parsed to obtain the coordinated adjustment sequence of the hyperparameter configuration; By extrapolating the performance gradient trend of the coordinated adjustment sequence, we obtain the hyperparameter configuration variants of the candidate hyperparameters.

10. The automated hyperparameter optimization method as described in claim 1, characterized in that, The process of screening and evaluating the hyperparameter configuration variants to obtain the optimized hyperparameters from the candidate hyperparameters includes: Obtain the performance data of the hyperparameter configuration variant; The performance data is quantified using the super-volume index to obtain the evaluation result of the super-volume index. Based on the evaluation results of the hypervolume index, the optimal parameter configuration variant is selected as the optimized hyperparameter of the candidate hyperparameter.

Images