AI algorithm parameter tuning method and device based on a parameter-adjusting large model, and related equipment

Abstract

This invention relates to the field of automated parameter tuning technology in machine learning, and particularly to a method, apparatus, and related equipment for AI algorithm parameter tuning based on a large-scale parameter tuning model. The method includes: acquiring parameter space and tuning task information; retrieving matching historical tuning record data from a tuning record database; inputting the tuning task information and historical tuning record data into a large-scale parameter tuning model to obtain tuning strategy information; generating candidate parameter combinations based on the parameter space and tuning strategy information and performing constraint verification; writing valid candidate parameter combinations into a target workflow node and triggering execution; collecting and analyzing the execution result data to obtain parameter tuning analysis results; storing the current round of data in the tuning record database, and outputting the target parameter combination or continuing iteration based on the analysis results. This method can improve the efficiency and reliability of AI algorithm parameter tuning.

Description

Technical Field

[0001] This invention relates to the field of automated parameter tuning technology in machine learning, and in particular to a method, apparatus and related equipment for optimizing AI algorithm parameters based on a large parameter tuning model. Background Technology

[0002] Machine learning and deep learning algorithms typically involve a large number of hyperparameters, such as tree depth, learning rate, number of iterations, sample sampling rate, and feature sampling rate for tree models, and the number of network layers, number of hidden units, regularization coefficient, learning rate, and batch size for neural network models. The values ​​of these hyperparameters directly affect model accuracy, convergence speed, training time, memory usage, and the satisfaction of business constraints. Therefore, parameter tuning is one of the most labor-intensive aspects of implementing AI algorithms.

[0003] Existing parameter tuning methods typically include manual tuning, grid search, random search, Bayesian optimization, evolutionary algorithms, and continuous halving search. While these methods can achieve parameter search to some extent, they still have shortcomings: the existing tuning process lacks a mechanism for reusing historical data. For tuning tasks with similar algorithms, similar datasets, similar evaluation metrics, or similar business constraints, the system often still starts searching from the initial search space, failing to effectively reuse existing tuning experience. Summary of the Invention

[0004] This invention provides a method, apparatus, and related equipment for AI algorithm parameter tuning based on a large-scale parameter tuning model, which solves the problems of high manual maintenance cost of parameter space, lack of parameter importance judgment in traditional search, and inability to reuse historical experience of similar tasks in the prior art.

[0005] In a first aspect, the present invention provides a method for optimizing AI algorithm parameters based on a large-scale parameter tuning model, comprising:

[0006] Obtain parameter space and optimization task information; the parameter space is constructed based on the parameter definition information of the algorithm components to be optimized, including parameter name, parameter type, parameter value range and parameter dependency; the optimization task information is input by the user, or generated based on pre-obtained user goals, evaluation indicators, resource budget, data characteristics and business constraints; Retrieve historical tuning record data that matches the tuning task information from the preset tuning record database; The optimization task information and the historical optimization record data are input into the pre-trained hyperparameter tuning model to obtain hyperparameter tuning strategy information; the hyperparameter tuning strategy information includes a subset of parameters to be searched, parameter importance scores, initial sampling configuration information, optimization algorithm, and stopping conditions; Candidate parameter combinations are generated based on the parameter space and the parameter tuning strategy information, and the candidate parameter combinations are constrained and verified to obtain valid candidate parameter combinations. Write the legal candidate parameter combination into the target workflow node corresponding to the algorithm component to be tuned, and trigger the execution of the workflow corresponding to the target workflow node; Collect the running result data corresponding to the legal candidate parameter combinations, and analyze the running result data to obtain parameter tuning analysis results; the parameter tuning analysis results are used to characterize whether the legal candidate parameter combinations meet preset requirements; The legal candidate parameter combinations, along with the corresponding runtime data and parameter tuning analysis results, are stored as new tuning record data in the tuning record database. If the parameter tuning analysis results indicate that the legitimate candidate parameter combination meets the preset requirements, then the legitimate candidate parameter combination is determined as the target parameter combination, and the target parameter combination is output; otherwise, the process returns to the following step: retrieving historical tuning record data related to the tuning task information from the preset tuning record database.

[0007] In one optional implementation, the parameter definition information includes configurable parameters and parameter constraints of the algorithm component to be tuned; wherein, the configurable parameters include at least one of continuous parameters, discrete parameters, enumerated parameters, Boolean parameters, and parameters with conditional dependencies; the parameter dependencies are used to limit the target parameter to be enabled, disabled, or have its value range restricted when the associated parameter meets a preset value condition.

[0008] In one optional implementation, obtaining historical optimization record data matching the optimization task information from a preset optimization record database includes: extracting at least one of algorithm type, evaluation index, resource budget, data characteristics, and business constraints from the optimization task information; matching the algorithm type, evaluation index, resource budget, data characteristics, and business constraints in the optimization record database to obtain historical optimization record data; wherein, the historical optimization record data includes at least one of historical candidate parameter combinations, historical running result data, historical parameter optimization analysis results, historical failure reasons, and historical target parameter combinations.

[0009] In one optional implementation, the initial sampling configuration information includes at least one of the following: the number of initial candidate parameter combinations, sampling range, sampling distribution, and number of sampling rounds; the optimization algorithm includes at least one of grid search, random search, Bayesian optimization, multi-objective evolutionary algorithm, continuous halving search, reinforcement learning search, or heuristic search; the stopping condition includes at least one of the following: reaching the maximum number of iteration rounds, the improvement of the running result data being lower than a preset improvement threshold, resource budget being exhausted, or the constraint deviation value being greater than a preset deviation threshold.

[0010] In one optional implementation, generating a candidate parameter combination based on the parameter space and the parameter tuning strategy information includes: determining parameters to be tuned from the parameter space based on the subset of parameters to be searched; determining the sampling weight of each parameter to be tuned based on the parameter importance score; generating the candidate parameter combination based on the initial sampling configuration information and the sampling weight; wherein, the sampling density is increased for parameters to be tuned with a parameter importance score higher than a preset importance threshold, and / or the sampling density is decreased for parameters to be tuned with a parameter importance score lower than the preset importance threshold.

[0011] In one optional implementation, the constraint verification of the candidate parameter combinations to obtain valid candidate parameter combinations includes: performing parameter type verification, parameter value range verification, enumeration value verification, and parameter dependency verification on the candidate parameter combinations; if a candidate parameter combination passes the parameter type verification, parameter value range verification, enumeration value verification, and parameter dependency verification, the candidate parameter combination is determined as a valid candidate parameter combination; if a candidate parameter combination fails any of the verifications, the candidate parameter combination is discarded, and the corresponding verification failure reason is recorded.

[0012] In one optional implementation, the step of writing the legal candidate parameter combination into the target workflow node and triggering the execution of the workflow corresponding to the target workflow node includes: determining the workflow node where the algorithm component to be tuned corresponding to the legal candidate parameter combination is located, and determining the workflow node as the target workflow node; writing the legal candidate parameter combination into the node parameter configuration of the target workflow node; and triggering the execution of the target workflow node and the downstream workflow nodes that take the output of the target workflow node as input.

[0013] In one optional implementation, the running result data includes at least one of model performance data, resource consumption data, convergence status data, and constraint deviation data; the parameter tuning analysis results include at least one of indicator improvement magnitude, resource consumption status, constraint deviation value, failure type information, whether the stopping condition is met, target parameter combination determination result, and next round search configuration information; wherein, when the parameter tuning analysis results indicate that the target parameter combination has not been determined based on the legal candidate parameter combination, the next round search configuration information includes at least one of the adjusted search range, sampling quantity, sampling distribution, random sampling ratio, and stopping condition.

[0014] Secondly, the present invention provides an AI algorithm parameter tuning device based on a large-scale parameter tuning model, comprising: an information acquisition module, a historical record acquisition module, a strategy generation module, a candidate parameter generation module, a workflow execution module, a result analysis module, a record update module, and an output iteration module.

[0015] Thirdly, the present invention provides an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement the steps of the above-described AI algorithm parameter tuning method based on a large-scale parameter tuning model.

[0016] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described AI algorithm parameter tuning method based on a large-scale parameter tuning model.

[0017] The present invention has at least the following beneficial effects: First, this invention constructs a parameter space based on the parameter definition information of the algorithm components to be tuned, enabling the system to automatically parse and utilize parameter names, parameter types, parameter value ranges, and parameter dependencies. Candidate parameter combinations are generated based on the parameter space and the parameter tuning strategy information, and the candidate parameter combinations are constrained and verified to obtain valid candidate parameter combinations. This reduces the cost of manually maintaining the search space and lowers the probability of parameter mismatch, out-of-bounds configuration, and dependency conflicts.

[0018] Second, this invention inputs optimization task information and historical optimization record data into a pre-trained large-scale parameter tuning model, which then outputs a subset of parameters to be searched, parameter importance scores, initial sampling configuration information, optimization algorithms, and stopping conditions. This allows the prior judgment of the large-scale parameter tuning model to narrow the search range and allocate more computing resources to parameters that have a greater impact on the optimization target.

[0019] Third, this invention stores historical optimization records and newly added optimization records in an optimization record database, enabling optimization tasks under similar algorithms, similar data features, similar evaluation indicators, or similar business constraints to reuse historical experience and avoid starting the search from scratch for each optimization.

[0020] Fourth, this invention writes the combination of legal candidate parameters into the target workflow node and triggers the execution of the real workflow, then collects the running result data and obtains the parameter tuning analysis results, so that the parameter tuning strategy can be updated according to the actual execution feedback, thereby forming a closed loop of parameter tuning strategy generation, candidate execution, result collection, record update and re-tuning. Attached Figure Description

[0021] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0022] Figure 1 A flowchart illustrating the AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model provided in this embodiment of the invention; Figure 2 A schematic diagram of the AI ​​algorithm parameter tuning system architecture based on a large-scale parameter tuning model provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of the AI ​​algorithm parameter tuning device based on a large-scale parameter tuning model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, those skilled in the art can make inventive decisions without making creative contributions. In this application, the "algorithm component to be tuned" can be any algorithm node, model training node, feature processing node, parameter optimization node, or model evaluation node in a componentized algorithm workflow platform. For example, the algorithm component to be tuned can be a gradient boosting tree training node, a random forest training node, a neural network training node, a support vector machine training node, a clustering algorithm node, a feature selection node, or a deep learning model training node.

[0024] In this application, "parameter definition information" refers to structured metadata used to describe the configurable parameters and parameter constraints of the algorithm component to be tuned. Parameter definition information can originate from component registration information, component configuration files, component interface definitions, component parameter description files, node metadata of the workflow platform, or parameter declaration files of the algorithm component. Parameter definition information may include at least one of the following: parameter name, parameter type, enumerated values, upper and lower limits, default values, required status, and parameter dependencies.

[0025] In this application, "parameter space" refers to a searchable set of parameters and their value constraints constructed based on parameter definition information. The parameter space can include at least one of continuous parameters, discrete parameters, enumerated parameters, Boolean parameters, and parameters with conditional dependencies. For example, a continuous parameter can be the learning rate, a discrete parameter can be the maximum tree depth, an enumerated parameter can be the optimizer type, a Boolean parameter can be whether regularization is enabled, and parameters with conditional dependencies can be optimizer internal parameters that only take effect when a particular optimizer is enabled.

[0026] In this application, "optimization task information" refers to information used to describe the objectives and constraints of this parameter tuning. Optimization task information can be user-inputted natural language parameter tuning requirements, or it can be structured optimization task information generated by the system based on user objectives, evaluation metrics, resource budget, data features, and business constraints. User objectives may include improving model accuracy, reducing validation set error, shortening training time, reducing memory usage, or balancing multiple optimization objectives; evaluation metrics may include accuracy, recall, F1 score, mean squared error, root mean square error, AUC, loss function value, etc.; resource budget may include maximum training time, maximum memory usage, maximum number of concurrent tasks, total computation time, or computational cost limits; data features may include sample size, feature dimensions, class distribution, missing rate, sparsity, and training / validation set partitioning methods, etc.; business constraints may include upper limits for model size, inference time, accuracy, training time, and resource consumption.

[0027] The "tuning record database" in this application is used to store historical tuning record data and newly added tuning record data. The tuning record database can be a relational database, key-value database, document database, vector database, time-series database, or a combination of the above. Tuning records in the tuning record database can be indexed according to algorithm type, data characteristics, evaluation metrics, resource budget, business constraints, workflow version, data version, or task identifier.

[0028] The "large-scale parameter tuning model" in this application can be a pre-trained large language model, a multimodal large model, a language model fine-tuned for algorithm parameter tuning tasks, or a parameter tuning strategy generation model that integrates a rule engine and a large language model. The large-scale parameter tuning model is used to generate parameter tuning strategy information based on tuning task information and historical tuning record data. The large-scale parameter tuning model does not necessarily directly output the final target parameter combination; it can output a subset of parameters to be searched, parameter importance scores, initial sampling configuration information, optimization algorithm, and stopping conditions, which are then further used by the optimization search process to generate candidate parameter combinations. In this way, the large-scale parameter tuning model is responsible for policy-level decision-making, while the optimization algorithm is responsible for generating candidate parameter combinations, thus separating their responsibilities.

[0029] In this application, "running result data" refers to the result data obtained after a valid combination of candidate parameters is written into the target workflow node and executed. Running result data may include at least one of the following: model performance data, resource consumption data, convergence status data, and constraint deviation data. Model performance data may include model accuracy, loss function, precision, recall, F1 score, AUC, mean squared error, or root mean square error; resource consumption data may include runtime, training time, memory usage, video memory usage, CPU utilization, GPU utilization, or concurrent task usage; convergence status data may include convergence algebra, indicator change trends, indicator fluctuation amplitude, or parameter value change amplitude; constraint deviation data may include the deviation of the running result from business constraints or resource budget.

[0030] In this application, "parameter tuning analysis results" refers to data obtained after analyzing the running results data, used to determine whether to continue the search, whether to output the target parameter combination, and how to generate the next round of search configuration information. Parameter tuning analysis results may include at least one of the following: the magnitude of indicator improvement, resource consumption status, constraint deviation value, failure type information, whether the stopping condition is met, the target parameter combination determination result, and the next round of search configuration information.

[0031] Figure 1 This is a flowchart illustrating the AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model provided in an embodiment of the present invention. Figure 1 As shown, this method can be applied to component-based algorithm workflow platforms, as well as to servers, cloud-based model training platforms, local automated machine learning platforms, or distributed algorithm training platforms. The method includes steps S101 to S110.

[0032] Step S101: Obtain parameter space and tuning task information.

[0033] The parameter space is constructed based on the parameter definition information of the algorithm components to be tuned, including parameter names, parameter types, parameter value ranges, and parameter dependencies. Specifically, each algorithm component to be tuned in the componentized algorithm workflow platform can be registered according to a unified component contract. The unified component contract can include component identifier, component input, component output, parameter definition information, parameter default values, parameter upper and lower limits, parameter enumeration range, parameter mandatory status, and parameter dependencies.

[0034] The tuning system can read the parameter definition information of the algorithm components to be tuned from the component registry, workflow configuration files, or component metadata database, and construct a parameter space based on the parameter definition information. For example, for a gradient boosting tree training node, the parameter space may include the maximum tree depth, learning rate, number of iterations, sample sampling rate, and feature sampling rate; for a neural network training node, the parameter space may include the learning rate, batch size, number of network layers, number of hidden units, regularization coefficient, and optimizer type.

[0035] Parameter dependencies are used to describe the conditional constraints between different parameters. For example, when the optimizer type is Adam, the first moment estimate of the decay rate and the second moment estimate of the decay rate are searchable; when the optimizer type is SGD, the momentum coefficient is searchable; when regularization is not enabled, the regularization coefficient can be disabled or excluded from the search. Through parameter dependencies, the system can avoid generating meaningless or unexecutable parameter combinations.

[0036] Optimization task information can be directly input by the user. For example, a user can input "reduce the root mean square error of the validation set as much as possible while keeping the training time under 20 minutes." Optimization task information can also be generated by the system based on pre-acquired user objectives, evaluation metrics, resource budgets, data features, and business constraints. For example, the user objective might be to reduce validation set error, the evaluation metric might be root mean square error, the resource budget might be a maximum training time of 20 minutes and a maximum number of concurrent tasks of 4, the data features might be the number of samples, feature dimensions, and missing rate, and the business constraints might be an upper limit on model size and an upper limit on inference time.

[0037] Step S102: Obtain historical tuning record data that matches the tuning task information from a preset tuning record database.

[0038] Specifically, the system can extract algorithm type, data features, evaluation metrics, resource budget, and business constraints from optimization task information, and retrieve similar historical optimization records from the optimization record database based on this information. Matching methods may include key field matching, tag matching, similarity matching, vector retrieval, or a combination of the above methods.

[0039] In one implementation, the system can encode optimization task information into task feature vectors, encode historical optimization tasks in the optimization record database into historical task feature vectors, and calculate the similarity between the task feature vectors and the historical task feature vectors. When the similarity is higher than a preset similarity threshold, the corresponding historical optimization record is identified as historical optimization record data that matches the current optimization task information.

[0040] Historical tuning records can include historical candidate parameter combinations, historical running results data, historical parameter tuning analysis results, historical reasons for failure, historical target parameter combinations, historical data versions, historical workflow versions, historical random seeds, and the reasons for parameter selection or strategy generation in the output of historical large models.

[0041] Step S103: Input the optimization task information and the historical optimization record data into the pre-trained parameter tuning model to obtain parameter tuning strategy information.

[0042] The parameter tuning strategy information includes the subset of parameters to be searched, parameter importance scores, initial sampling configuration information, optimization algorithm, and stopping conditions.

[0043] The subset of parameters to be searched is a set of parameters selected from the parameter space for this round of search. The large-scale parameter tuning model can determine which parameters are more sensitive to the current tuning task based on parameter value changes, indicator trends, failure reasons, and historical target parameter combinations in historical tuning records, and identify the parameters with higher sensitivity as the subset of parameters to be searched.

[0044] Parameter importance scores are used to represent the degree of influence of each parameter on the evaluation metric or user objective. Parameter importance scores can be numerical values ​​between 0 and 1, or high, medium, and low level labels, or other scores that can be ranked. The large-scale parameter tuning model can output the importance score of each parameter based on the correspondence between parameter changes in historical tuning records and the running results data.

[0045] Initial sampling configuration information may include the number of initial candidate parameter combinations, sampling range, sampling distribution, and number of sampling rounds. Optimization algorithms may include grid search, random search, Bayesian optimization, multi-objective evolutionary algorithms, continuous halving search, reinforcement learning search, or heuristic search. Stopping conditions may include reaching the maximum number of iterations, the improvement in the running results data falling below a preset improvement threshold, resource budget exhaustion, constraint deviation exceeding a preset deviation threshold, or the user-defined target indicator being reached.

[0046] The large-scale parameter tuning model can also output the reasons for parameter selection, the reasons for search algorithm selection, risk warnings, and suggestions for adjusting the search scope. The reasons for parameter selection explain why a certain parameter is prioritized for search; the reasons for search algorithm selection explain why Bayesian optimization, multi-objective evolutionary algorithms, or other search algorithms are chosen; risk warnings may include risks of parameter out-of-bounds errors, resource overruns, constraint conflicts, or historical failures.

[0047] Step S104: Generate candidate parameter combinations based on the parameter space and the parameter tuning strategy information, and perform constraint verification on the candidate parameter combinations to obtain valid candidate parameter combinations.

[0048] Specifically, the system determines the parameters to be adjusted from the parameter space based on the subset of parameters to be searched, determines the sampling weight of each parameter to be adjusted based on the parameter importance score, and generates candidate parameter combinations based on the initial sampling configuration information and sampling weights. For parameters to be adjusted with an importance score higher than a preset importance threshold, the system can increase the sampling density; for parameters to be adjusted with an importance score lower than the preset importance threshold, the system can decrease the sampling density or fix it to the default value.

[0049] For example, when the importance score of the learning rate is higher than the preset importance threshold, the system can generate more candidate values ​​within the learning rate range; when the importance score of a certain low-sensitivity parameter is lower than the preset importance threshold, the system can reduce the number of candidate values ​​for that parameter, or directly adopt the value that has performed better in the past.

[0050] After candidate parameter combinations are generated, the system performs constraint validation on them. Constraint validation can include parameter type validation, parameter value range validation, enumeration value validation, and parameter dependency validation. Parameter type validation determines whether the parameter value conforms to the parameter type, such as whether an integer parameter is configured as an integer, a floating-point parameter as a floating-point number, or a Boolean parameter as a Boolean value. Parameter value range validation determines whether the parameter value is within the upper and lower limits. Enumeration value validation determines whether the parameter value belongs to a preset enumeration set. Parameter dependency validation determines whether the parameters satisfy the enable, disable, or value restriction relationship.

[0051] If a candidate parameter combination passes the above checks, it is determined to be a valid candidate parameter combination. If a candidate parameter combination fails any check, the system discards the candidate parameter combination and records the corresponding reason for the failure. The reason for the failure can be stored in the tuning record database as part of the subsequent tuning record data, so that the model can avoid similar illegal combinations in the future.

[0052] Step S105: Write the legal candidate parameter combination into the target workflow node and trigger the execution of the workflow corresponding to the target workflow node.

[0053] Specifically, the system determines the workflow node containing the algorithm component to be tuned corresponding to the valid candidate parameter combination, and designates this workflow node as the target workflow node. Subsequently, the system writes the valid candidate parameter combination into the node parameter configuration of the target workflow node and triggers the execution of the workflow corresponding to the target workflow node.

[0054] In one implementation, the system can rerun only the target workflow node and its downstream workflow nodes that take the output of the target workflow node as input, without re-executing workflow nodes that are not affected by the valid candidate parameter combination. This local execution method reduces redundant calculations and improves parameter tuning efficiency.

[0055] Step S106: Collect the running result data corresponding to the legal candidate parameter combinations, and analyze the running result data to obtain parameter tuning analysis results.

[0056] After the workflow is completed, the system collects the execution result data corresponding to the legal candidate parameter combinations. The execution result data may include at least one of the following: model performance data, resource consumption data, convergence status data, and constraint deviation data. Model performance data may include model accuracy, loss function, precision, recall, F1 score, AUC, mean squared error, or root mean square error; resource consumption data may include execution time, training time, memory usage, video memory usage, CPU utilization, or GPU utilization; convergence status data may include the convergence algebra, indicator trends, and parameter value changes; constraint deviation data may include the deviation of the execution results from the resource budget or business constraints.

[0057] The system analyzes the operational results data to obtain parameter tuning analysis results. These results include indications indicating whether the legitimate candidate parameter combinations meet preset requirements; they may also include at least one of the following: indicator improvement magnitude, resource consumption status, constraint deviation value, failure type information, whether stopping conditions are met, target parameter combination determination results, and next round search configuration information.

[0058] The improvement margin of the indicator can be determined by the difference or ratio between the evaluation indicator corresponding to the current candidate parameter combination and the historical best evaluation indicator. Resource consumption status can be used to indicate whether the runtime, memory usage, GPU memory usage, or number of concurrent tasks exceeds the resource budget. Constraint deviation value can be used to indicate the degree of deviation of the running result data from business constraints or resource budget. Failure type information can be determined based on workflow execution logs, exception information, or verification failure reasons, such as parameter dependency conflicts, insufficient memory, training non-convergence, abnormal indicator fluctuations, etc. Whether the stopping condition is met can be determined based on the stopping condition in the parameter tuning strategy information and the current running result data. The target parameter combination determination result is used to characterize whether the target parameter combination has been determined based on the legal candidate parameter combinations. The next round search configuration information is used to control the generation process of the next round of candidate parameter combinations.

[0059] When the improvement in a metric over a consecutive preset number of rounds is less than a preset improvement threshold, the system considers the metric to have stagnated and generates configuration information for the next round of search to narrow the search scope, adjust the sampling distribution, change the random sampling ratio, or adjust the stopping conditions. When resource consumption status indicators such as runtime, memory usage, or the number of concurrent tasks exceed the resource budget, the system can reduce the number of samples in the next round or trigger the stopping conditions earlier. When the constraint deviation value is greater than a preset deviation threshold, the system can increase the priority of the constraints corresponding to the business constraints or resource budget and adjust the parameter value range or constraint validation rules.

[0060] Step S107: The legal candidate parameter combination, the corresponding running result data and parameter tuning analysis results are stored as new tuning record data in the tuning record database.

[0061] New tuning record data can include the parameter names and values ​​used in this candidate execution, data version, workflow version, random seed, execution result data, failure reason, parameter tuning analysis results, and the reasons for parameter selection or strategy generation output by the large-scale tuning model. By storing the above data in a linked manner, the tuning record database can retain the complete tuning process, enabling subsequent similar tasks to reuse historical experience.

[0062] Step S108: Determine whether the parameter tuning analysis results indicate that the target parameter combination has been determined based on the legal candidate parameter combinations.

[0063] When the parameter tuning analysis results meet the stopping conditions, and the current legal candidate parameter combination is the optimal parameter combination in the current historical record, or the running result data corresponding to the current legal candidate parameter combination meets the user's goals, evaluation indicators, resource budget, and business constraints, it can be determined that the target parameter combination has been determined based on the legal candidate parameter combination.

[0064] Step S109: If the parameter tuning analysis results indicate that the target parameter combination has been determined based on the legal candidate parameter combination, then the target parameter combination is output.

[0065] When outputting the target parameter combination, the system can also output the corresponding running results data, parameter importance scores, convergence curves, failure type distributions, and next-round suggestions. The convergence curve can be generated based on the changes in evaluation metrics corresponding to multiple rounds of candidate parameter combinations. The failure type distribution can be obtained statistically from the failure reason field in the tuning record database. Next-round suggestions can include recommendations to continue searching for parameters, recommendations to fix parameters, recommendations to narrow the search scope, or recommendations to adjust the resource budget.

[0066] If the parameter tuning analysis results do not indicate that the target parameter combination is determined based on the legal candidate parameter combination, then return to step S102, that is, retrieve historical tuning record data that matches the tuning task information from the preset tuning record database.

[0067] Since the newly added tuning record data has already been stored in the tuning record database in step S107, the historical tuning record data obtained when step S102 is executed again already includes the newly added tuning record data from this round. In other words, the input to the next round of parameter tuning model not only includes the original historical tuning record data, but also the legal candidate parameter combinations, running result data, and parameter tuning analysis results from the previous round. Therefore, the parameter tuning strategy information for the next round is not simply regenerated, but dynamically updated based on the actual execution feedback from the previous round.

[0068] Through the above steps, this embodiment forms a closed-loop tuning process: "Input tuning task information and historical tuning record data into the large-scale tuning model — Output tuning strategy information — Generate candidate parameter combinations — Workflow execution — Collect running result data — Generate parameter tuning analysis results — Update tuning record database — Call the large-scale tuning model again".

[0069] In some embodiments, the parameter definition information includes configurable parameters and parameter constraints of the algorithm component to be tuned. Configurable parameters may include at least one of continuous parameters, discrete parameters, enumerated parameters, Boolean parameters, and parameters with conditional dependencies.

[0070] Continuous parameters are parameters that take values ​​within a continuous numerical range, such as learning rate, regularization coefficient, sample sampling rate, and feature sampling rate. Discrete parameters are parameters that take values ​​within a set of integers or discrete numerical values, such as maximum tree depth, number of iterations, and number of neural network layers. Enumerated parameters are parameters that take values ​​within a predefined enumeration set, such as optimizer type, loss function type, and model structure type. Boolean parameters are parameters that take values ​​of true or false, such as whether regularization is enabled or whether early stopping is enabled. Parameters with conditional dependencies are those whose participation in the search or whose value range depends on the values ​​of other parameters.

[0071] Parameter dependencies are used to limit whether a target parameter is enabled, disabled, or has its value range restricted when associated parameters meet preset value conditions. For example, when the parameter "booster" is "dart", the parameter "sample_type" is enabled; when the parameter "booster" is "gbtree", the parameter "sample_type" is disabled. As another example, when the parameter "regularization_enabled" is true, regularization coefficients participate in the search; when this parameter is false, regularization coefficients do not participate in the search.

[0072] Based on the parameter definition information above, the system can automatically construct a parameter space with type constraints, range constraints, and dependency constraints, solving the problems of high manual maintenance costs and easy errors under high-dimensional parameter spaces and complex conditional dependencies.

[0073] In some embodiments, the system extracts at least one of the following from the optimization task information: algorithm type, evaluation metric, resource budget, data characteristics, and business constraints, and matches it in the optimization record database based on the above content to obtain historical optimization record data.

[0074] For example, the algorithm type for the current optimization task is gradient boosting tree, the evaluation metric is root mean square error, the resource budget is a training time of no more than 20 minutes, the data characteristics are sample size, feature dimension, and missing rate, and the business constraint is that the model inference time is lower than a preset threshold. The system can retrieve historical optimization records in the optimization record database that have the same algorithm type, similar data characteristics, the same or similar evaluation metrics, and similar resource budgets.

[0075] Historical tuning records may include at least one of the following: historical candidate parameter combinations, historical execution result data, historical parameter tuning analysis results, historical failure reasons, and historical target parameter combinations. Historical candidate parameter combinations characterize the values ​​of parameters searched in the past; historical execution result data characterizes the model performance and resource consumption of historical candidate parameter combinations; historical parameter tuning analysis results characterize the magnitude of indicator improvement, constraint deviation, and satisfaction of stopping conditions during the historical search process; historical failure reasons indicate which parameter combinations are prone to execution failure; and historical target parameter combinations provide a reference for optimal parameters under similar tasks.

[0076] Through the aforementioned historical record matching mechanism, the system can reuse existing optimization experience in similar tasks, avoiding starting the search from scratch for each optimization.

[0077] In some embodiments, the initial sampling configuration information includes at least one of the following: the number of initial candidate parameter combinations, the sampling range, the sampling distribution, and the number of sampling rounds.

[0078] The initial number of candidate parameter combinations limits the number of candidate parameter combinations generated in the first round or the current round. The sampling range limits the upper and lower limits of the values ​​for each parameter to be searched, or enumerates the set of values. The sampling distribution can include a uniform distribution, a normal distribution, a log-uniform distribution, a custom probability distribution, or a bias distribution determined based on historical optimal parameters. The number of sampling rounds limits the number of candidate generation rounds under the current parameter tuning strategy.

[0079] The optimization algorithms include at least one of the following: grid search, random search, Bayesian optimization, multi-objective evolutionary algorithms, continuous halving search, reinforcement learning search, or heuristic search. Grid search is suitable for low-dimensional, small-range parameter spaces; random search is suitable for rapid exploration of larger parameter spaces; Bayesian optimization is suitable for continuous or mixed parameter spaces with high evaluation costs; multi-objective evolutionary algorithms are suitable for multi-objective scenarios that simultaneously optimize model performance and resource consumption; continuous halving search is suitable for gradually eliminating low-performance candidate combinations; reinforcement learning search is suitable for scenarios where the tuning strategy can be learned iteratively through states and rewards; and heuristic search is suitable for parameter tuning scenarios that incorporate domain rules or historical experience.

[0080] The stopping conditions include at least one of the following: reaching the maximum number of iterations, the improvement in the running results data being less than a preset improvement threshold, resource budget being exhausted, or constraint deviation exceeding a preset deviation threshold. By using these stopping conditions, the system can avoid invalid or over-budget searches.

[0081] In some embodiments, the system determines the parameters to be adjusted from the parameter space based on a subset of parameters to be searched; determines the sampling weight of each parameter to be adjusted based on the parameter importance score; and generates candidate parameter combinations based on the initial sampling configuration information and the sampling weights.

[0082] Specifically, if a parameter has a high importance score, it indicates that the parameter has a significant impact on the current evaluation metric or user goal. The system can increase the sampling density of this parameter during the candidate parameter combination generation process, generating more candidate values ​​within a unit value range. If a parameter has a low importance score, it indicates that the parameter has a relatively small impact on the current evaluation metric or user goal. The system can reduce the sampling density of this parameter, decreasing the number of its candidate values, or fix it to a default value or a historical preferred value.

[0083] By using parameter importance scoring to drive sampling weights, the system can concentrate limited computing power on highly sensitive parameters, reduce invalid sampling on low-sensitivity parameters, and improve search efficiency.

[0084] In some embodiments, the system performs parameter type verification, parameter value range verification, enumeration value verification, and parameter dependency verification on candidate parameter combinations.

[0085] Parameter type validation is used to confirm that the values ​​of each parameter in the candidate parameter combination are consistent with the parameter types in the parameter definition information. For example, integer parameters cannot be configured as strings, and boolean parameters cannot be configured as non-boolean values.

[0086] Parameter value range validation is used to confirm that the values ​​of each parameter in the candidate parameter combination are within the upper and lower limits specified in the parameter definition information. For example, the learning rate should be between 0.001 and 0.3, and the maximum tree depth should be between 3 and 12.

[0087] Enumeration value validation is used to confirm that the value of an enumeration parameter belongs to the enumeration set in the parameter definition information. For example, the optimizer type can only be one of the preset Adam, SGD, RMSprop, etc.

[0088] Parameter dependency validation is used to confirm that candidate parameter combinations satisfy enabled, disabled, or value restriction relationships between parameters. For example, when regularization is not enabled, regularization coefficients are not included in the search; when a certain model structure is not selected, the internal parameters associated with that model structure are not included in the search.

[0089] If a candidate parameter combination passes all validations, it is determined as a valid candidate parameter combination; if a candidate parameter combination fails any validation, it is discarded, and the corresponding validation failure reason is recorded. The validation failure reason can be stored in the tuning record database to reduce the probability of generating similar combinations when generating tuning strategies for subsequent large-scale models.

[0090] In some embodiments, the system determines the workflow node where the algorithm component to be tuned is located corresponding to the legal candidate parameter combination, and identifies the workflow node as the target workflow node.

[0091] For example, in a component-based algorithm workflow platform, a complete workflow may include a data reading node, a data preprocessing node, a feature engineering node, a model training node, a model evaluation node, and a model export node. If the algorithm component to be tuned corresponding to a valid candidate parameter combination is a model training node, then that model training node is the target workflow node.

[0092] The system writes the valid candidate parameter combinations into the node parameter configuration of the target workflow node. The writing methods can include updating the node configuration file, calling the workflow platform interface, modifying the node parameter table, updating the task runtime configuration, or generating a new runtime configuration version.

[0093] After the write operation is complete, the system triggers the execution of the target workflow node and its downstream workflow nodes that use the target workflow node's output as input. For example, if the target workflow node is a model training node, the system can execute the model training node and the model evaluation node without re-executing the data reading node and the data preprocessing node. In this way, the system leverages the platform's local execution capabilities to reduce redundant computations and improve parameter tuning efficiency.

[0094] In some embodiments, the runtime result data includes at least one of model performance data, resource consumption data, convergence status data, and constraint deviation data.

[0095] Model performance data reflects the effectiveness of the model corresponding to the candidate parameter combination, such as model accuracy, loss function, root mean square error, and AUC. Resource consumption data reflects the resource usage during the execution of the candidate parameter combination, such as runtime, memory usage, GPU memory usage, and concurrent task usage. Convergence status data reflects the convergence status during model training, such as the convergence algebra, the decreasing trend of the loss function, and the magnitude of metric fluctuations. Constraint deviation data reflects the degree of deviation of the running results from business constraints or resource budgets, such as the time difference between training time exceeding the budget, the proportion of memory usage exceeding the threshold, or the proportion of model size exceeding the upper limit.

[0096] The parameter tuning analysis results include at least one of the following: the magnitude of indicator improvement, resource consumption status, constraint deviation value, failure type information, whether the stopping condition is met, the result of the target parameter combination judgment, and the configuration information for the next round of search.

[0097] When the parameter tuning analysis results indicate that the target parameter combination cannot be determined based on the legal candidate parameter combinations, the system can generate the next round of search configuration information. The next round of search configuration information includes at least one of the following: adjusted search range, number of samples, sampling distribution, random sampling ratio, and stopping conditions.

[0098] For example, when the improvement in metrics over multiple consecutive rounds is less than a preset improvement threshold, the next round's search configuration information may include narrowing the learning rate search range, increasing the sampling density in the region near the tree's maximum depth, or changing the sampling distribution variance. When resource consumption exceeds the resource budget, the next round's search configuration information may include reducing the number of samples, reducing the number of training rounds, or triggering the stopping condition earlier. When the constraint deviation value is greater than a preset deviation threshold, the next round's search configuration information may include adjusting the parameter value range, increasing constraint priority, or updating constraint validation rules.

[0099] The following description, in conjunction with specific application examples, illustrates the following: In one specific embodiment, the algorithm component to be tuned is a gradient boosting tree training node. The parameter space includes the maximum tree depth, learning rate, number of iterations, sample sampling rate, and feature sampling rate.

[0100] The user's objective is to reduce the root mean square error of the validation set while keeping the training time under 20 minutes. The evaluation metric is root mean square error, the resource budget is a training time of no more than 20 minutes, and the business constraint is that the model inference time does not exceed a preset threshold.

[0101] The system retrieves historical tuning records from the tuning record database that match gradient boosting trees, root mean square error (RMSE), training time budget, and data characteristics. The historical tuning record data shows that, on similar datasets, the learning rate and maximum tree depth have a significant impact on the RMSE, and excessively large maximum tree depths can lead to training time exceeding limits.

[0102] The system inputs the optimization task information and historical optimization records into the large-scale hyperparameter tuning model. The large-scale hyperparameter tuning model outputs the hyperparameter tuning strategy information: the subset of parameters to be searched includes the learning rate and the maximum tree depth; the parameter importance score for the learning rate is high, the parameter importance score for the maximum tree depth is high, and the importance scores for the sample sampling rate and feature sampling rate are medium; the initial sampling configuration information includes generating several candidate parameter combinations; the optimization algorithm selected is Bayesian optimization; the stopping conditions include reaching the maximum number of iterations, the root mean square error improvement being less than a threshold for multiple consecutive iterations, or the training time exceeding 20 minutes.

[0103] The optimization search process generates candidate parameter combinations based on the aforementioned parameter tuning strategy information. Before writing these combinations to the target workflow node, it performs parameter type validation, upper and lower bound validation, enumeration value validation, and parameter dependency validation. Once the validation passes, the system writes the valid candidate parameter combinations to the gradient boosting tree training node, triggering execution of the training node and its downstream model evaluation nodes.

[0104] After the workflow is completed, the system collects runtime data such as root mean square error (RMSE), training time, and memory usage. If the RMSE improvement is less than the preset improvement threshold for several consecutive rounds, the parameter tuning analysis results can indicate insufficient improvement. The system then generates search configuration information for the next round, such as narrowing the search range for the learning rate and increasing the sampling density for the maximum tree depth. If the training time exceeds 20 minutes, the system generates search configuration information for the next round, such as lowering the upper limit of the maximum tree depth or reducing the number of iteration rounds.

[0105] Each candidate execution's valid parameter combinations, execution results, parameter tuning analysis results, data version, workflow version, random seed, failure reason, and the parameter selection rationale output by the large-scale model are associated and stored in the tuning record database. At the end of tuning, the system outputs the target parameter combination, parameter importance score, convergence curve, failure type distribution, and suggestions for the next round. These tuning records can also be reused in subsequent similar datasets or similar gradient boosting tree tasks.

[0106] like Figure 2 As shown, Figure 2 This is a schematic diagram of the AI ​​algorithm parameter tuning system architecture based on a large-scale parameter tuning model provided in this application embodiment. The system includes a core engine, a tuning record database, an input unit, and an output unit. The core engine is communicatively connected to the input unit, the output unit, and the tuning record database to form a closed-loop tuning process from tuning task input, parameter tuning strategy generation, candidate parameter execution, running result collection, to tuning record updating.

[0107] The input unit provides the core engine with the algorithm component to be tuned, the parameter space, and tuning task information. The algorithm component to be tuned is the component that requires parameter tuning and can be a model training component, feature processing component, data processing component, or model evaluation component in a componentized algorithm workflow. The parameter space defines the selectable values ​​of each parameter in the algorithm component and the constraints between parameters, and can include at least one of the following: parameter name, parameter type, parameter value range, default value, enumerated value, and parameter dependency. The tuning task information describes the goals and constraints of this parameter tuning and can include at least one of the following: user goals, evaluation metrics, resource budget, data features, and business constraints.

[0108] The core engine includes a parameter space resolver, a large model tuning planner, a candidate parameter generator, a workflow candidate executor, an indicator collector, a tuning memory and attribution unit, and an engine control / data bus. The engine control / data bus connects to each of these functional modules, transmitting parameter space, tuning task information, historical tuning record data, tuning strategy information, candidate parameter combinations, workflow execution status, running result data, and parameter tuning analysis results, and coordinating the execution order of each functional module.

[0109] The parameter space resolver is used to parse the parameter space corresponding to the algorithm component to be tuned, and to determine the parameters to be tuned participating in the search, as well as their type constraints, value range constraints, and parameter dependencies. In one implementation, the parameter space resolver is also used to convert the parameter definition formats used by different algorithm components into a unified structured parameter space, so as to facilitate unified processing by subsequent modules.

[0110] The large model tuning planner is used to acquire tuning task information, the parsed parameter space, and historical tuning record data provided by the tuning record database. This information is then input into a pre-trained large-scale model to obtain tuning strategy information. This strategy information may include at least one of the following: a subset of parameters to be searched, parameter importance scores, initial sampling configuration information, optimization algorithms, and stopping conditions. The large model tuning planner is primarily used to plan the tuning strategy layer, such as determining the parameters to be searched first in this round, the search priority of each parameter, and the applicable optimization algorithms.

[0111] The candidate parameter generator is used to generate at least one candidate parameter combination based on the parameter space and tuning strategy information. The candidate parameter generator can determine the sampling weight of each parameter to be tuned based on the parameter importance score, so that parameters with higher parameter importance scores have higher sampling density, and parameters with lower parameter importance scores have lower sampling density or remain at the default value.

[0112] The candidate parameter generator is also used to perform constraint validation on candidate parameter combinations. This constraint validation may include parameter type validation, parameter value range validation, enumeration value validation, and parameter dependency validation. If a candidate parameter combination passes the constraint validation, it is determined as a valid candidate parameter combination; if it fails the constraint validation, it is discarded, and the corresponding reason for the validation failure is recorded.

[0113] The workflow candidate executor is used to write valid candidate parameter combinations into the target workflow node corresponding to the algorithm component to be tuned, and trigger the execution of the workflow corresponding to the target workflow node. In one implementation, the workflow candidate executor only re-executes the target workflow node and the downstream workflow nodes that take the output of the target workflow node as input, and does not re-execute the upstream workflow nodes that are not affected by the valid candidate parameter combinations, so as to reduce redundant calculations.

[0114] The index collector is used to collect the running result data corresponding to the legal candidate parameter combinations. The running result data may include at least one of model performance data, resource consumption data, convergence status data, and constraint deviation data. For example, model performance data may include model accuracy, loss function value, mean square error, or root mean square error; resource consumption data may include running time, memory usage, or video memory usage; convergence status data may include convergence algebra, index change trend, and index improvement magnitude.

[0115] The tuning memory and attribution unit is used to perform correlation analysis on legitimate candidate parameter combinations, execution result data, workflow execution status, and failure reasons to obtain parameter tuning analysis results. These results may include at least one of the following: parameter importance, resource consumption status, constraint deviation value, failure type, whether stopping conditions are met, target parameter combination determination results, and next-round search configuration information. For example, when the improvement in an indicator for a consecutive preset number of rounds is less than a preset improvement threshold, it can be determined that the tuning process has stalled, and next-round search configuration information can be generated to narrow the search scope, adjust the sampling distribution, or change the optimization algorithm; when resource consumption exceeds the resource budget, the number of samples can be reduced or the stopping conditions adjusted.

[0116] The tuning record database is used to store the result data, log data, and experimental reproduction data generated during parameter tuning. The tuning record database includes a parameter-metric memory, an experimental version and random seed storage area, and a convergence / resource / failure log storage area. The parameter-metric memory is used to associate and store candidate parameter combinations and their corresponding execution result data; the experimental version and random seed storage area is used to store data versions, algorithm component versions, workflow versions, and random seeds; the convergence / resource / failure log storage area is used to store metric convergence status, resource usage, and reasons for candidate parameter combination execution failures.

[0117] Historical tuning records in the tuning record database are fed back to the core engine and used by the large model tuning planner to generate the next round of parameter tuning strategy information. Therefore, the next round of parameter tuning does not start from the initial state, but is adjusted based on the previous round of candidate parameter combinations and their actual running results.

[0118] The output unit is used to output recommended parameter combinations, parameter importance, and a parameter tuning explanation report. Recommended parameter combinations are target parameter combinations that satisfy the optimization objectives, evaluation metrics, resource budget, and business constraints. Parameter importance characterizes the degree of influence of each parameter to be tuned on the evaluation metrics. The parameter tuning explanation report may include the reasons for parameter selection, the reasons for optimization algorithm selection, the metric convergence process, resource consumption, reasons for failure, and the basis for determining the recommended parameter combinations.

[0119] When the parameter tuning analysis results indicate that a candidate parameter combination meets the preset requirements, the candidate parameter combination is output as the recommended parameter combination. When the parameter tuning analysis results do not indicate that a candidate parameter combination meets the preset requirements, the historical tuning record data containing the newly added tuning record data in this round is retrieved from the tuning record database, and the large model tuning planner regenerates the parameter tuning strategy information to continue executing the next round of parameter tuning.

[0120] Figure 3 This is a schematic diagram of the AI ​​algorithm parameter tuning device based on a large-scale parameter tuning model provided in an embodiment of the present invention. Figure 3 As shown, the device includes an information acquisition module 31, a history record acquisition module 33, a strategy generation module 33, a candidate parameter generation module 34, a workflow execution module 35, a result analysis module 36, a record update module 37, and an output iteration module 38.

[0121] The information acquisition module 31 is used to acquire parameter space and tuning task information. The parameter space is constructed based on the parameter definition information of the algorithm components to be tuned, including parameter names, parameter types, parameter value ranges, and parameter dependencies. The tuning task information is either user input or generated based on user objectives, evaluation metrics, resource budgets, data characteristics, and business constraints.

[0122] The historical record acquisition module 32 is used to retrieve historical tuning record data that matches the tuning task information from a preset tuning record database.

[0123] The strategy generation module 33 is used to input the tuning task information and historical tuning record data into the pre-trained hyperparameter tuning model to obtain hyperparameter tuning strategy information. The hyperparameter tuning strategy information includes the subset of parameters to be searched, parameter importance scores, initial sampling configuration information, optimization algorithm, and stopping conditions.

[0124] The candidate parameter generation module 34 is used to generate candidate parameter combinations based on parameter space and parameter tuning strategy information, and to perform constraint verification on the candidate parameter combinations to obtain valid candidate parameter combinations.

[0125] The workflow execution module 35 is used to write the combination of valid candidate parameters into the target workflow node and trigger the workflow execution corresponding to the target workflow node.

[0126] The results analysis module 36 is used to collect the running result data corresponding to the legal candidate parameter combinations, and analyze the running result data to obtain the parameter tuning analysis results.

[0127] The record update module 37 is used to store the legal candidate parameter combinations, the corresponding running result data and parameter tuning analysis results as new tuning record data in the tuning record database.

[0128] The output iteration module 38 is used to output the target parameter combination if the parameter tuning analysis result characterization has determined the target parameter combination based on the legal candidate parameter combination; otherwise, it triggers the history record acquisition module to retrieve the historical tuning record data that matches the tuning task information from the preset tuning record database.

[0129] It should be noted that each module in the above device embodiments can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor of the electronic device in hardware form, or stored in the memory of the electronic device in software form, so that the processor can call and execute the operations corresponding to each module. The above device embodiments and method embodiments belong to the same inventive concept. Specific limitations on each module in the device can be found in the foregoing method embodiments, and will not be repeated here.

[0130] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Figure 4 As shown, the electronic device may include a processor 41, a communication interface, a memory 42, and a communication bus, wherein the processor 41, the communication interface, and the memory 42 communicate with each other through the communication bus.

[0131] The memory 42 is used to store computer programs. The processor is used to execute the computer programs stored in the memory to implement the steps of the AI ​​algorithm parameter tuning method based on the large-scale parameter tuning model in any of the above embodiments.

[0132] Specifically, when the processor executes the computer program, it can perform the following steps: acquire parameter space and tuning task information; retrieve historical tuning record data matching the tuning task information from a preset tuning record database; input the tuning task information and historical tuning record data into a pre-trained large-scale tuning model to obtain tuning strategy information; generate candidate parameter combinations based on the parameter space and tuning strategy information, and perform constraint verification on the candidate parameter combinations to obtain legal candidate parameter combinations; write the legal candidate parameter combinations into the target workflow node and trigger the execution of the workflow corresponding to the target workflow node; collect the running result data corresponding to the legal candidate parameter combinations, and analyze the running result data to obtain parameter tuning analysis results; store the legal candidate parameter combinations, running result data, and parameter tuning analysis results as newly added tuning record data in the tuning record database; if the parameter tuning analysis results indicate that the target parameter combination has been determined based on the legal candidate parameter combinations, then output the target parameter combination; otherwise, return to the step of acquiring historical tuning record data.

[0133] The electronic device in this embodiment can be a server, personal computer, cloud computing node, model training platform device, edge computing device, or other computing device with data processing capabilities. This embodiment does not limit the specific implementation of the electronic device.

[0134] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program implements the steps of the AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model in any of the above embodiments.

[0135] The non-transitory computer-readable storage medium may include read-only memory, random access memory, disk, optical disk, flash memory, portable hard disk, USB flash drive, or other media capable of storing program code.

[0136] Those skilled in the art will understand that all or part of the processes of the above method embodiments can be implemented by computer program instructions to related hardware. The computer program can be stored in a non-transitory computer-readable storage medium, and when executed, it can include the processes of the above method embodiments.

[0137] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing AI algorithm parameters based on a large-scale parameter tuning model, characterized in that, include: Obtain parameter space and tuning task information; The parameter space is constructed based on the parameter definition information of the algorithm components to be tuned, including parameter name, parameter type, parameter value range, and parameter dependency; the optimization task information is input by the user, or generated based on pre-acquired user goals, evaluation indicators, resource budget, data characteristics, and business constraints. Retrieve historical tuning record data that matches the tuning task information from the preset tuning record database; The optimization task information and historical optimization record data are input into the pre-trained hyperparameter tuning model to obtain hyperparameter tuning strategy information; the hyperparameter tuning strategy information includes a subset of parameters to be searched, parameter importance scores, initial sampling configuration information, optimization algorithm, and stopping conditions; Candidate parameter combinations are generated based on the parameter space and the parameter tuning strategy information, and the candidate parameter combinations are constrained and verified to obtain valid candidate parameter combinations. Write the legal candidate parameter combination into the target workflow node corresponding to the algorithm component to be tuned, and trigger the execution of the workflow corresponding to the target workflow node; Collect the execution result data corresponding to the legal candidate parameter combinations; The operational results data are then analyzed to obtain parameter tuning analysis results; The parameter tuning analysis results are used to characterize whether the legal candidate parameter combinations meet the preset requirements; The legal candidate parameter combinations, along with the corresponding runtime data and parameter tuning analysis results, are stored as new tuning record data in the tuning record database. If the parameter tuning analysis results indicate that the legal candidate parameter combination meets the preset requirements, then the legal candidate parameter combination is determined as the target parameter combination, and the target parameter combination is output. Otherwise, return to the following step: retrieve historical tuning record data related to the tuning task information from the preset tuning record database.

2. The AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model according to claim 1, characterized in that, The parameter definition information includes the configurable parameters and parameter constraint relationships of the algorithm component to be tuned; The configurable parameters include at least one of continuous parameters, discrete parameters, enumerated parameters, Boolean parameters, and parameters with conditional dependencies. The parameter dependency relationship is used to limit whether the target parameter is enabled, disabled, or has its value range restricted when the associated parameter meets the preset value conditions.

3. The AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model according to claim 1, characterized in that, The step of retrieving historical tuning record data matching the tuning task information from a preset tuning record database includes: Extract at least one of the following from the optimization task information: algorithm type, evaluation metric, resource budget, data characteristics, and business constraints, as a matching item; Based on the matching item, a matching is performed in the optimization record database to obtain historical optimization record data; The historical optimization record data includes at least one of the following: historical candidate parameter combinations, historical operation result data, historical parameter optimization analysis results, historical failure reasons, and historical target parameter combinations.

4. The AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model according to claim 1, characterized in that, The initial sampling configuration information includes at least one of the following: the number of initial candidate parameter combinations, the sampling range, the sampling distribution, and the number of sampling rounds; The optimization algorithm includes at least one of grid search, random search, Bayesian optimization, multi-objective evolutionary algorithm, continuous halving search, reinforcement learning search, or heuristic search; The stopping conditions include at least one of the following: reaching the maximum number of iterations, the improvement of the running result data being less than a preset improvement threshold, resource budget being exhausted, or constraint deviation being greater than a preset deviation threshold.

5. The AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model according to claim 1, characterized in that, The step of generating candidate parameter combinations based on the parameter space and the parameter tuning strategy information includes: Based on the subset of parameters to be searched, determine the parameters to be adjusted that will participate in the search from the parameter space; The sampling weight of each parameter to be adjusted is determined based on the importance score of the parameter. Based on the initial sampling configuration information and the sampling weights, the candidate parameter combination is generated; Specifically, the sampling density is increased for parameters whose importance scores are higher than a preset importance threshold, and / or the sampling density is decreased for parameters whose importance scores are lower than the preset importance threshold.

6. The AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model according to claim 1, characterized in that, The constraint verification of the candidate parameter combinations to obtain valid candidate parameter combinations includes: The candidate parameter combinations are validated for parameter type, parameter value range, enumeration value, and parameter dependency. If the candidate parameter combination passes the parameter type verification, parameter value range verification, enumeration value verification, and parameter dependency verification, the candidate parameter combination is determined as the legal candidate parameter combination. If any candidate parameter combination fails any validation, the candidate parameter combination is discarded, and the corresponding validation failure reason is recorded.

7. The AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model according to claim 1, characterized in that, The step of writing the legal candidate parameter combination into the target workflow node of the algorithm component to be tuned, and triggering the execution of the workflow corresponding to the target workflow node, includes: Determine the workflow node where the algorithm component to be tuned corresponds to the legal candidate parameter combination, and determine the workflow node as the target workflow node; Write the combination of the legal candidate parameters into the node parameter configuration of the target workflow node; Trigger the execution of the workflow corresponding to the target workflow node.

8. The AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model according to claim 1, characterized in that, The running result data includes at least one of the following: model performance data, resource consumption data, convergence status data, and constraint deviation data. The parameter tuning analysis results include at least one of the following: indicator improvement magnitude, resource consumption status, constraint deviation value, failure type information, whether the stopping condition is met, target parameter combination judgment result, and next round search configuration information; Wherein, if the parameter tuning analysis results indicate that the target parameter combination has not been determined based on the legal candidate parameter combination, the next round of search configuration information includes at least one of the following: adjusted search range, number of samples, sampling distribution, random sampling ratio, and stopping condition.

9. A parameter tuning device for AI algorithms based on a large-scale parameter tuning model, characterized in that, include: The information acquisition module is used to acquire parameter space and optimization task information; The parameter space is constructed based on the parameter definition information of the algorithm components to be tuned, including parameter name, parameter type, parameter value range, and parameter dependency; the optimization task information is input by the user, or generated based on user goals, evaluation indicators, resource budget, data characteristics, and business constraints. The historical record acquisition module is used to acquire historical optimization record data that matches the optimization task information from a preset optimization record database. The strategy generation module is used to input the optimization task information and the historical optimization record data into a pre-trained hyperparameter tuning model to obtain hyperparameter tuning strategy information; the hyperparameter tuning strategy information includes a subset of parameters to be searched, parameter importance scores, initial sampling configuration information, optimization algorithm, and stopping conditions; The candidate parameter generation module is used to generate candidate parameter combinations based on the parameter space and the parameter tuning strategy information, and to perform constraint verification on the candidate parameter combinations to obtain valid candidate parameter combinations. The workflow execution module is used to write the combination of the legal candidate parameters into the target workflow node and trigger the execution of the workflow corresponding to the target workflow node; The results analysis module is used to collect the running result data corresponding to the legal candidate parameter combinations, and analyze the running result data to obtain parameter tuning analysis results. The record update module is used to store the legal candidate parameter combination, the corresponding running result data and parameter tuning analysis results as new tuning record data into the tuning record database. The output iteration module is used to output the target parameter combination if the parameter tuning analysis result characterization has determined the target parameter combination based on the legal candidate parameter combination; otherwise, it triggers the history acquisition module to retrieve historical tuning record data that matches the tuning task information from the preset tuning record database.

10. An electronic device comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the AI ​​algorithm parameter tuning method based on a large-scale parameter tuning model as described in any one of claims 1 to 8.

Images