A multi-dimensional evaluation and dynamic training method and system for optimizing an algorithm model

Abstract

This invention relates to the fields of artificial intelligence model testing and machine learning technology, and particularly to a method and system for multi-dimensional evaluation and dynamic training of optimization algorithm models. The method includes: acquiring training tasks and standardizing test items, data fusion, and grouping; initiating iterative training and collecting behavioral data streams with time labels and source identifiers; extracting state transition information sets; performing format verification, standard conversion, and multi-dimensional index storage; obtaining evaluation index data through evaluation model groups and multi-level evaluation model analysis; comparing the evaluation index data with dynamic thresholds and set evaluation thresholds, and identifying weaknesses using a statistical monitoring window; matching a strategy library and adjusting the training strategy based on the weakness identification results, and executing the next round of iterative training until the stopping condition is met. This invention achieves closed-loop optimization of training, evaluation, weakness identification, and strategy adaptation, improving the convergence efficiency, robustness, and system stability of the algorithm model.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence model testing and machine learning technology, and in particular to a method and system for multi-dimensional evaluation and dynamic training of optimization algorithm models. Background Technology

[0002] In the fields of artificial intelligence model testing and machine learning technology, existing solutions for optimizing algorithm models typically involve training execution based on the training task type, test questions, the algorithm model to be evaluated, and the training strategy. The model's operational status is reflected through fitness, network parameters, resource usage, or evaluation reports. However, these methods suffer from limitations such as a disconnect between evaluation metrics and training strategies, a lack of continuous judgment in bottleneck identification results, and reliance on manual configuration for strategy library matching. Existing methods often adjust training strategies based on changes in fitness or evaluation reports after training execution. The time labels, source identifiers, and state transition information sets in the original behavioral data stream are not fully correlated, making it difficult for evaluation metrics data to be directly incorporated into the next round of training execution.

[0003] In the training and evaluation of optimization algorithm models, if evaluation metrics are only compared using fixed thresholds, it can easily lead to unclear identification of the evaluation model corresponding to weaknesses and unclear targets for training strategy adjustments. If the statistical monitoring window, dynamic threshold mechanism, and setting evaluation thresholds are not processed in a consistent manner, convergence efficiency scores, robustness evaluation metrics, resource usage scores, and system stability metrics can easily become disconnected, making it difficult to meet the requirements for stable generation of optimization algorithm models and evaluation reports.

[0004] Regarding the joint processing of evaluation index data, dynamic threshold mechanisms, statistical monitoring windows, setting evaluation thresholds, weakness identification results, and strategy library matching, existing technologies generally suffer from insufficient coordination between evaluation data warehouse storage, evaluation model group analysis, weakness identification result generation, and training strategy adjustment. This makes it difficult to form a consistent process in the dynamic training scenario of optimization algorithm models, including raw behavioral data stream acquisition, state transition information set extraction, standardized storage data generation, evaluation index data analysis, weakness identification result determination, strategy library matching, and execution of the next round of training. Consequently, there is a lack of stable data links between training strategy adjustment results and updated training strategies. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for multi-dimensional evaluation and dynamic training of optimization algorithm models, comprising:

[0006] S100: Obtain the training task type, test questions, algorithm model to be evaluated and training strategy; after standardizing the test questions, perform data fusion, feature extraction and initial grouping to obtain the training task grouping results.

[0007] S200. Based on the training task grouping results, start iterative training, synchronously collect running records and add time tags and source identifiers to obtain the raw behavioral data stream;

[0008] S300. Based on the original behavioral data stream, extract fitness, initial optimal position, state transition related information, network parameters, and resource usage to form a state transition information set.

[0009] S400: Perform format verification, standard data conversion, and multi-dimensional index construction on the state transition information set, and store it in the evaluation data warehouse to obtain standardized storage data;

[0010] S500. Based on the standardized stored data, the evaluation index data is obtained through analysis of the evaluation model group and the multi-level evaluation model.

[0011] S600. The evaluation index data is compared with the dynamic threshold and the set evaluation threshold, and the shortcoming identification result is obtained by combining the statistical monitoring window.

[0012] S700. Based on the weakness identification results, match the strategy library and adjust the training strategy, execute the next round of iterative training until the stopping condition is met, and output the optimized algorithm model and evaluation report.

[0013] Furthermore, the process of data fusion, feature extraction, and initial grouping after standardizing the test items includes:

[0014] The standardization process includes data format unification, merging of duplicate questions, verification of source identifiers, and completion of training task types.

[0015] The data fusion refers to linking standardized multidimensional data of the same test question, training task type, and training strategy into the same multidimensional input data.

[0016] The feature extraction process involves extracting parameter space, constraints, historical running data, real-time running data, dynamic disturbances, maximum iteration count, data acquisition frequency parameters, and network parameters from the multidimensional input data to generate feature parameters.

[0017] The initial grouping is based on a similarity matrix, grouping test questions with the same training task type, similar parameter space, the same constraints, and the same training strategy into the same initial group, and generating a grouping adjustment log when test questions cross multiple initial groups or when source identifiers conflict.

[0018] Furthermore, the process of initiating iterative training, synchronously collecting runtime records, and adding time tags and source identifiers includes:

[0019] Before starting iterative training, an access verification is performed, which includes group number verification, test question number verification, training strategy reference record verification, and group adjustment log verification.

[0020] The synchronous acquisition of running records involves receiving running records from the training execution module according to the data acquisition frequency parameter, and arranging the continuously received running records to obtain a data stream to be labeled;

[0021] Adding time stamps includes writing start time stamps, writing process time stamps, and writing end time stamps;

[0022] Adding source identifiers involves binding the group number, test question number, training task type, algorithm model number to be evaluated, action network number, value network number, and network parameter version number to each run record.

[0023] Furthermore, the process of extracting fitness, initial optimal position, state transition information, network parameters, and resource usage to form a state transition information set includes:

[0024] The state transition information includes the t-th observation state, the t-th action selection, the t-th reward, and the (t+1)-th observation state;

[0025] First, the original behavioral data stream is sorted sequentially according to the time tag, and then the source is merged according to the source identifier to obtain continuous running records;

[0026] The t-th observation state, t-th action selection, t-th reward, and t+1-th observation state are extracted from the continuous running records and paired to obtain the t-th iteration features. The fitness, initial optimal position, number of iterations, maximum number of iterations, network parameters, and resource usage are extracted to obtain the training phase features.

[0027] The t-th iteration feature is associated and bound with the training phase feature.

[0028] Furthermore, the process of performing format validation, standard data conversion, and multidimensional index construction before storing the data in the evaluation data warehouse includes:

[0029] The format verification is a unified check of the record format of the t-th observation state, t-th selection action, t-th reward, t+1-th observation state, fitness, initial optimal position, network parameters, and resource usage.

[0030] The standard data conversion involves converting the aforementioned fields in the verified records into a unified record order and generating a version number.

[0031] The multidimensional index construction includes dimensions of creation time, source, training task, test question, and network parameters.

[0032] The evaluation data warehouse includes a standard data storage area, an index structure storage area, a verification result storage area, and a version number storage area.

[0033] Furthermore, the process of obtaining evaluation indicator data through the analysis of evaluation model groups and multi-level evaluation models includes:

[0034] The evaluation model set includes a convergence efficiency evaluation model, a robustness evaluation model, a resource utilization evaluation model, and a system stability evaluation model;

[0035] The multi-level evaluation model analysis includes: firstly, extracting a set of feature vectors from the standardized stored data, and then performing scoring processing through the evaluation model group to obtain convergence efficiency score, robustness evaluation index, resource utilization score, and system stability index; then, performing basic scoring processing, index merging processing, and intermediate evaluation data table generation processing through the multi-level evaluation model to finally obtain evaluation index data.

[0036] Furthermore, the process of comparing the evaluation index data with dynamic thresholds and set evaluation thresholds, and combining this with a statistical monitoring window to obtain the weakness identification results includes:

[0037] The dynamic threshold is a threshold comparison rule set according to different training task types, different test question numbers, and different version numbers;

[0038] The setting of the evaluation threshold comparison includes: first reading the verification result reference record in the evaluation index data, and then sending the verification record into the dynamic threshold mechanism for processing; the dynamic threshold mechanism processing includes threshold field reading, evaluation index data alignment and comparison status generation.

[0039] The process of obtaining the weakness identification result by combining the statistical monitoring window includes: reading the threshold comparison results under the same training task type, the same test question number, and the same version number in the statistical monitoring window according to the time tag order, and counting the continuous changes; when the comparison state corresponding to the same evaluation model is lower than the corresponding set evaluation threshold in multiple consecutive statistical monitoring windows, it is recorded as the evaluation model corresponding to the weakness; when the comparison state changes at the boundary position of the statistical monitoring window, the delayed effect mechanism is invoked, the boundary position is written to the temporary record, and reading continues in the next statistical monitoring window;

[0040] The bottleneck identification results also include: group number and feature parameters read from the training task grouping results, and the t-th observation state, t-th action selection, t-th reward, t+1-th observation state, network parameters and resource usage read from the state transition information set.

[0041] Furthermore, the process of matching the strategy library includes:

[0042] The strategy library includes convergence efficiency adjustment strategies, robustness adjustment strategies, resource usage adjustment strategies, and system stability adjustment strategies.

[0043] Based on the evaluation model corresponding to the shortcoming in the shortcoming identification results, the corresponding adjustment strategy is retrieved; when the evaluation model corresponding to the shortcoming is a convergence efficiency evaluation model, a convergence efficiency adjustment strategy is retrieved; when it is a robustness evaluation model, a robustness adjustment strategy is retrieved; when it is a resource usage evaluation model, a resource usage adjustment strategy is retrieved; and when it is a system stability evaluation model, a system stability adjustment strategy is retrieved.

[0044] Furthermore, the process of adjusting training strategies includes:

[0045] The candidate adjustment strategies are matched with the training task grouping results, network parameters, and dynamic threshold mechanism to generate training strategy adjustment results. Then, corresponding adjustment actions are executed according to the type of the candidate adjustment strategy: for convergence efficiency adjustment strategies, the network parameters and maximum number of iterations are adjusted; for robustness adjustment strategies, the dynamic perturbation records and test question groupings are adjusted; for resource usage adjustment strategies, the data acquisition frequency parameters and resource scheduling strategy are adjusted; and for system stability adjustment strategies, the dynamic threshold mechanism, statistical monitoring window, and delay-based activation mechanism are adjusted. Finally, an updated training strategy is generated based on the training strategy adjustment results.

[0046] Furthermore, a multi-dimensional evaluation and dynamic training system for optimizing algorithm models includes: a training task processing module, a training execution and data acquisition module, a state transition information processing module, a standardization storage module, an evaluation analysis module, a bottleneck identification module, a strategy library matching module, and a dynamic training module; the system is used to implement the method described in any of the above embodiments.

[0047] The key innovations of this invention include:

[0048] (1) The evaluation index data is not directly entered into the evaluation report, but is compared and processed by dynamic threshold mechanism, statistical monitoring window and set evaluation threshold to obtain the shortcoming identification result, so that the convergence efficiency score, robustness evaluation index, resource usage score and system stability index form a unified judgment process.

[0049] (2) The shortcoming identification results are not stored as separate records, but are matched with the strategy library, the training strategy is adjusted and the next round of training is executed to obtain the optimized algorithm model and evaluation report, so that the shortcoming identification results and the training strategy form a continuous calling relationship.

[0050] (3) The original behavioral data flows through the state transition information set, standardized storage data and evaluation index data in sequence, so that the t-th observation state, the t-th selected action, the t-th reward, the t+1-th observation state, network parameters and resource usage are entered into the shortcoming identification results and training strategy adjustment.

[0051] The following are its main beneficial effects:

[0052] (1) To address the problem of the disconnect between evaluation index data and training strategy, the evaluation index data is converted into the shortcoming identification result through dynamic threshold mechanism, statistical monitoring window and setting evaluation threshold comparison processing, so that the evaluation model, training task type and test question number corresponding to the shortcoming form a callable data object.

[0053] (2) To address the issue that strategy library matching relies on manual configuration, strategy library matching is performed based on the results of the shortcoming identification, and the training strategy is adjusted according to the candidate adjustment strategy, so that the training strategy adjustment results correspond to the convergence efficiency adjustment strategy, robustness adjustment strategy, resource usage adjustment strategy and system stability adjustment strategy.

[0054] (3) To address the problem of insufficient correlation between the original behavioral data stream and the evaluation index data, fitness, initial optimal position, network parameters and resource usage are entered into multi-level evaluation model analysis and processing through continuous processing of state transition information set, standardized storage data and evaluation model group.

[0055] (4) To address the problem of unclear positioning of the evaluation model corresponding to the shortcomings caused by fixed threshold comparison, the evaluation index data is processed by statistical monitoring window and dynamic threshold mechanism to form the shortcomings identification results under training task type, test question number and version number.

[0056] (5) To address the lack of stable data links in the next round of training, the optimization algorithm model and evaluation report are kept in relation to the previous evaluation index data through continuous processing of the shortcoming identification results, strategy library matching, training strategy adjustment and updating of the training strategy. Attached Figure Description

[0057] Figure 1 A flowchart illustrating a multi-dimensional evaluation and dynamic training method for an optimization algorithm model provided in this application embodiment;

[0058] Figure 2 This is a structural block diagram of a multi-dimensional evaluation and dynamic training system for an optimization algorithm model provided in an embodiment of this application. Detailed Implementation

[0059] Example 1: Refer to Figure 1This is a flowchart illustrating a multi-dimensional evaluation and dynamic training method for an optimization algorithm model provided in an embodiment of the present invention. The process may include at least steps S100-S700:

[0060] S100: Obtain the training task type, test questions, algorithm model to be evaluated and training strategy; after standardizing the test questions, perform data fusion, feature extraction and initial grouping to obtain the training task grouping results.

[0061] S200. Based on the training task grouping results, start iterative training, synchronously collect running records and add time tags and source identifiers to obtain the raw behavioral data stream;

[0062] S300. Based on the original behavioral data stream, extract fitness, initial optimal position, state transition related information, network parameters, and resource usage to form a state transition information set.

[0063] S400: Perform format verification, standard data conversion, and multi-dimensional index construction on the state transition information set, and store it in the evaluation data warehouse to obtain standardized storage data;

[0064] S500. Based on the standardized stored data, the evaluation index data is obtained through analysis of the evaluation model group and the multi-level evaluation model.

[0065] S600. The evaluation index data is compared with the dynamic threshold and the set evaluation threshold, and the shortcoming identification result is obtained by combining the statistical monitoring window.

[0066] S700. Based on the weakness identification results, match the strategy library and adjust the training strategy, execute the next round of iterative training until the stopping condition is met, and output the optimized algorithm model and evaluation report.

[0067] S100: Obtain the training task type, test questions, algorithm model to be evaluated and training strategy; after standardizing the test questions, perform data fusion, feature extraction and initial grouping to obtain the training task grouping results.

[0068] Specifically, S100 is triggered by the system before training execution. The training task type is a record of the category of the training task to which the test question belongs, including categories corresponding to historical running data, real-time running data, dynamic perturbations, and resource scheduling strategies. The test question is the question data used for training the optimization algorithm model to be evaluated, including multi-dimensional raw data, parameter space, constraints, and training task type. The optimization algorithm model to be evaluated is the optimization algorithm model entering training execution this time, including the particle swarm optimization algorithm model, action network, value network, and network parameters. The training strategy is the configuration record called for this training execution, including the maximum number of iterations, data acquisition frequency parameters, set evaluation threshold, and corresponding records in the strategy library. After receiving the training task type, the test question, the optimization algorithm model to be evaluated, and the training strategy, the system first checks the access records; when any input lacks a source identifier or a reserved time stamp position, the system generates a group adjustment log and places the test question with the missing record into the pending processing group.

[0069] Specifically, the system standardizes the test questions to obtain standardized multidimensional data. This standardization process includes data format unification, duplicate question merging, source identifier verification, and training task type completion. Data format unification converts raw multidimensional data from different sources into standardized multidimensional data with the same field order and data units. Duplicate question merging combines test questions with the same source identifier, training task type, and feature parameters into a single question record. Source identifier verification matches the source identifiers in the test questions with those in the training strategy. Training task type completion adds test questions lacking category records to adjacent training task types based on historical and real-time running data. After this processing, the standardized multidimensional data carries the test question number, training task type, source identifier, reserved time stamp space, and training strategy reference record.

[0070] Furthermore, the system performs data fusion and feature extraction based on the standardized multidimensional data and the training task type to obtain multidimensional input data and feature parameters. The data fusion associates the standardized multidimensional data, training task type, and training strategy for the same test item into a single multidimensional input data set. The feature extraction extracts parameter space, constraints, historical running data, real-time running data, dynamic perturbations, maximum iteration count, data acquisition frequency parameters, and network parameters from the multidimensional input data to generate feature parameters. These feature parameters serve as input fields for initializing grouping processing. Understandably, when the same test item corresponds to multiple training strategies, the system reads the training strategies according to the reserved time label position and version number order, and writes the last completed training strategy into the multidimensional input data.

[0071] Furthermore, the system performs similarity matrix calculation, initial grouping, and group adjustment log generation based on the multidimensional input data and the feature parameters to obtain the training task grouping results. Similarity matrix calculation involves matching the feature parameters of multiple test items item by item to generate a similarity matrix representing the similarity relationships between test items. Initial grouping involves grouping test items with the same training task type, similar parameter space, identical constraints, and the same training strategy into the same initial group based on the similarity matrix. Group adjustment log generation records the group before adjustment, the group after adjustment, the trigger time tag, and the trigger source identifier when test items span multiple initial groups, have conflicting source identifiers, or inconsistent training strategies. After completing the initial grouping, the system writes the group number, test item number, training task type, feature parameters, training strategy reference record, and group adjustment log into the training task grouping results.

[0072] In one engineering implementation, the training task type is resource scheduling strategy testing, the optimization algorithm model to be evaluated is a particle swarm optimization algorithm model, and the training strategy records a maximum of 200 iterations, a data acquisition frequency of once per second, and sets evaluation thresholds and network parameters. After the system receives twenty test questions, it first converts the test questions into standardized multidimensional data, and then extracts the parameter space, constraints, historical running data, and real-time running data from the standardized multidimensional data. The system calculates a similarity matrix based on the feature parameters and groups test questions with similar parameter spaces and the same training strategy into the same initial group. If the source identifier of a test question is inconsistent with other test questions in the same group, the system moves the test question to a new initial group and records the source identifier, time tag reserved position, and adjusted group in the group adjustment log. After processing, the training task grouping results are sent to S200 for S200's training execution, time tag, and source identifier addition processing.

[0073] In summary, the technical effects of this step are as follows: This step transforms the training task type, test questions, the algorithm model to be evaluated and optimized, and the training strategy into training task grouping results. These grouping results retain the source identifier, reserved time label positions, feature parameters, and grouping adjustment logs. This result provides a unified input for the S200 to collect raw behavioral data streams.

[0074] S200. Based on the training task grouping results, start iterative training, synchronously collect running records and add time tags and source identifiers to obtain the raw behavioral data stream;

[0075] Specifically, S200 is executed collaboratively by the training execution module and the data acquisition module, with the input source being the training task grouping results generated by S100. The training task grouping results include a group number, test question number, training task type, feature parameters, training strategy reference records, and group adjustment logs. After reading the training task grouping results, the training execution module retrieves the corresponding test question according to the group number and the training strategy according to the training strategy reference records. The training strategy includes the maximum number of iterations, data acquisition frequency parameters, a set evaluation threshold, and network parameters. Training execution refers to the process of driving the optimization algorithm model to be evaluated to iteratively train the test question. The data acquisition module starts synchronously when the training execution begins and receives the running records generated during the training execution process according to the data acquisition frequency parameters. These running records are organized into training phase data in this step, which includes the group number, test question number, training task type, training strategy reference records, maximum number of iterations, data acquisition frequency parameters, and network parameter call records.

[0076] Specifically, the training execution module first performs access verification on the training task grouping results. Access verification includes group number verification, test question number verification, training strategy reference record verification, and group adjustment log verification. Group number verification is used to determine if there are duplicate groups in the training task grouping results. Test question number verification is used to determine if each test question belongs to a group. Training strategy reference record verification is used to determine if each group has a corresponding training strategy. Group adjustment log verification is used to determine if the adjustment records generated by the previous initialization grouping process have been written. If a training task grouping result lacks a training strategy reference record, the training execution module pauses the training execution for that group and writes the group number into the exception record. If the training task grouping results pass the access verification, the training execution module starts training execution according to the group number order and writes the start time into the reserved timestamp location.

[0077] Furthermore, the training execution module invokes the optimization algorithm model to be evaluated during training execution. The optimization algorithm model to be evaluated includes an action network, a value network, and network parameters. The action network receives the test questions and training strategies corresponding to the training task grouping results and outputs action records during training execution. The value network receives the running records during training execution and outputs value records during training execution. The network parameters are read and their version numbers are recorded in each training stage. A training stage refers to the continuous running interval under the same group number, from the start of training execution to reaching the maximum number of iterations or before switching training strategies. At the end of each training stage, the training execution module generates training stage data and sends the training stage data to the data acquisition module. Understandably, when a test question reading failure, a training strategy reference record conflict, or an inconsistent network parameter version number occurs during training execution, the training execution module stops the training execution of the current group and writes the failure location, source identifier, and time stamp into the group adjustment log.

[0078] Furthermore, the data acquisition module configures data acquisition frequency parameters based on the training phase data. The data acquisition frequency parameter refers to the reception interval record of the original behavioral data stream. After the training phase data arrives, the data acquisition module reads the data acquisition frequency parameter and receives run records from the training execution module according to the data acquisition frequency parameter. The run record includes the training phase number, group number, test question number, training task type, action network output record, value network output record, network parameter version number, and a reserved field for resource usage. The data acquisition module arranges the continuously received run records in the receiving order to obtain a data stream to be labeled. If a run record is missing within the same training phase, the data acquisition module writes a missing record between adjacent run records and writes the receiving position corresponding to the missing record into an anomaly record. If a run record is duplicated, the data acquisition module retains the run record with the earlier timestamp and writes the duplicate record into the group adjustment log.

[0079] Further, the data acquisition module adds time tags to the data stream to be labeled. The time tag is an identifier field that records the generation order and reception time of the running records. The time tag addition process includes writing a start time tag, writing a process time tag, and writing an end time tag. The start time tag corresponds to the start position of training execution. The process time tag corresponds to the reception position of each running record. The end time tag corresponds to the end position of the training phase. After the data acquisition module writes the start time tag, process time tag, and end time tag to the data stream to be labeled, it obtains a time-tagged data stream. The time tags are subsequently used in step S300 to sequentially extract the t-th observation state, t-th action selection, t-th reward, and t+1-th observation state from the original behavioral data stream.

[0080] Furthermore, the data acquisition module adds source identifiers to the time-stamped data stream. The source identifier is an identifier field representing the source of the run record, including group number, test question number, training task type, algorithm model number to be evaluated, action network number, value network number, and network parameter version number. The source identifier addition process binds these fields to each run record, resulting in a data stream with time stamps and source identifiers. If the same run record corresponds to multiple source identifiers, the data acquisition module determines the primary source identifier according to the group number in the training task grouping result and writes the remaining source identifiers into the grouping adjustment log. If the source identifier is inconsistent with the training task grouping result, the data acquisition module writes the run record into the exception record and pauses the run record from entering the original behavior data stream.

[0081] In one engineering implementation, the training task grouping result comprises three groups, each containing several test questions. The training strategy records a maximum of 200 iterations, a data acquisition frequency of once per second, and a network parameter version number. The training execution module first reads the test questions from the first group and then calls the optimization algorithm model to be evaluated for training execution. The data acquisition module receives a running record every second and adds a process time tag and source identifier to each running record. When the training execution of the first group reaches the maximum number of iterations, the training execution module generates training phase data. The data acquisition module merges the training phase data with the continuous running records to obtain the raw behavioral data stream. Subsequently, the training execution module reads the second group and repeats the above process. After all group processing is completed, the raw behavioral data stream contains a group number, test question number, training task type, training strategy reference record, time tag, source identifier, action network output record, value network output record, and network parameter version number.

[0082] The raw behavioral data stream is the output of S200 and serves as the input of S300. Specifically, S300 extracts fitness, initial optimal position, observation state t, action selection at t, reward at t, observation state at t+1, network parameters, and resource usage based on the raw behavioral data stream to obtain a set of state transition information. The time tag in S300 is used to determine the sequential order of the running records. The source identifier in S300 is used to determine the training task grouping result and the optimization algorithm model to be evaluated for the running record. The training phase data in S300 is used to form training phase features.

[0083] In summary, the technical effects of this step are as follows: This step converts the training task grouping results into a raw behavioral data stream with time labels and source identifiers. The raw behavioral data stream retains the group number, test question number, training strategy reference record, and network parameter version number from the training execution process. This data stream provides a continuous running record for the S300 to extract the state transition information set.

[0084] S300. Based on the original behavioral data stream, extract fitness, initial optimal position, state transition related information, network parameters, and resource usage to form a state transition information set.

[0085] Specifically, S300 is triggered by the system upon receiving the raw behavioral data stream generated in S200. The raw behavioral data stream includes a group number, test question number, training task type, training strategy reference record, time label, source identifier, action network output record, value network output record, and network parameter version number. The system first organizes the raw behavioral data stream sequentially according to the time label, and then merges the sources according to the source identifier to obtain continuous running records corresponding to the same training task grouping result, the same test question, and the same optimization algorithm model to be evaluated. These continuous running records serve as the input data for this step. If a continuous running record lacks a time label, the system reads the positional relationship of adjacent running records to fill in the missing information and writes the filled position into the anomaly record. If the source identifier is inconsistent with the training task grouping result, the system temporarily stores the running record and writes the corresponding group number and test question number into the grouping adjustment log.

[0086] Specifically, the system extracts and processes fitness, initial optimal position, number of iterations, maximum number of iterations, network parameters, and resource usage based on the original behavioral data stream to obtain training phase features. Fitness is a numerical field recording the quality of solutions to test problems during training execution, derived from value network output records and training execution result records. Initial optimal position is the position record corresponding to the current fitness at the start of the same training phase, derived from the first valid running record of the optimization algorithm model to be evaluated after training execution begins. Number of iterations is the number of iteration records completed within the same training phase. Maximum number of iterations comes from training strategy reference records. Network parameters are records of parameters invoked by the action network and value network during training execution, including network parameter version numbers and parameter invocation positions. Resource usage is a record of runtime, cache invocation, and resource scheduling strategies generated during training execution. The system merges the above fields according to the training phase number to obtain training phase features, and binds the training phase features with the time tag and the source identifier.

[0087] Further, the system extracts the t-th observation state, t-th action selection, t-th reward, and t+1-th observation state based on the original behavioral data stream to obtain the t-th iteration feature. The t-th observation state is the current observation record sorted by time label within the same training phase, including the number of iterations, the maximum number of iterations, the test question number, and the network parameter version number. The t-th action selection is the action record output by the action network in the t-th observation state, originating from the action network output record. The t-th reward is the fitness change record formed after the t-th action selection is executed, originating from the fitness change between adjacent time labels. The t+1-th observation state is the next observation record after the t-th action selection is executed, originating from the adjacent running record after the t-th observation state. The system pairs these records in the order of "current observation record, action record, reward record, next observation record" to obtain the t-th iteration feature. If there is no adjacent t+1-th observation state for the same t-th observation state, the system writes the t-th observation state into the exception record and stops the running record from entering the state transition information set.

[0088] Furthermore, the system performs association processing based on the t-th iteration feature and the training stage feature to obtain a set of state transition information. This association processing involves binding the t-th iteration feature under the same source identifier, the same test question number, and the same training stage number to the corresponding training stage feature. After binding, each state transition information entry includes the t-th observation state, the t-th action selection, the t-th reward, the t+1-th observation state, fitness, initial optimal position, number of iterations, maximum number of iterations, network parameters, resource usage, time label, and source identifier. Multiple state transition information entries are arranged consecutively according to their time labels and grouped by group number to obtain a set of state transition information. Each state transition information entry in the set retains the network parameter version number, facilitating subsequent standard data format conversion and multi-dimensional index structure establishment by S400.

[0089] In one engineering implementation, the training task grouping result includes a resource scheduling strategy test group, the test questions correspond to twenty parameter space records, and the training strategy records a maximum of two hundred iterations. The training execution module has already formed the raw behavioral data stream in S200. After reading the raw behavioral data stream in S300, the system first organizes the running records according to a time label of once per second, and then divides the continuous running records according to the test question number. For a given test question, the system reads the fitness and position records from the first valid running record to generate the initial optimal position. Subsequently, the system reads the current observation record, action network output record, fitness change record, and next observation record one by one to form the t-th iteration feature. The system then binds the t-th iteration feature to the training phase features. If the action network output record of a certain running record exists, but the next observation record is missing, the system writes the running record into the exception record and does not enter the state transition information set. After completing the processing of all test questions, the system generates a state transition information set and sends the state transition information set to S400.

[0090] The state transition information set is the output of S300 and serves as the input of S400. Specifically, S400 performs standard data format conversion, multidimensional index structure establishment, and evaluation data warehouse storage processing based on the state transition information set to obtain standardized stored data. The fitness, the initial optimal position, the t-th observation state, the t-th selection action, the t-th reward, the (t+1)-th observation state, the network parameters, and the resource usage are subject to format verification and integrity checks in S400. The time stamp and the source identifier are subject to multidimensional index structure establishment processing in S400.

[0091] In summary, the technical effect of this step is as follows: This step organizes the raw behavioral data stream into a set of state transition information with time labels, source identifiers, and network parameter version numbers. This set of state transition information associates the features from the training phase with the features from the t-th iteration. This set provides continuous and traceable input for the S400 to generate standardized storage data.

[0092] S400: Perform format verification, standard data conversion, and multi-dimensional index construction on the state transition information set, and store it in the evaluation data warehouse to obtain standardized storage data;

[0093] Specifically, S400 is triggered after receiving the state transition information set generated by S300. This state transition information set includes the t-th observation state, the t-th action selection, the t-th reward, the (t+1)-th observation state, fitness, initial optimal position, number of iterations, maximum number of iterations, network parameters, resource usage, time stamp, and source identifier. The system first reads the time stamp and source identifier from the state transition information set, and then performs an access verification based on the training task grouping results, test question number, and network parameter version number. This access verification determines the order, source record, and field completeness of each state transition information. If a state transition information is missing the (t+1)-th observation state, the system writes this state transition information into an exception log. If the source identifier of a state transition information is inconsistent with the training task grouping results, the system temporarily stores this state transition information and records the corresponding test question number, time stamp, and source identifier in the grouping adjustment log.

[0094] Specifically, the system performs format verification and integrity checks based on the state transition information set to obtain verification results. The format verification performs a unified check on the record formats of the t-th observation state, t-th action selection, t-th reward, t+1-th observation state, fitness, initial optimal position, network parameters, and resource usage. The integrity check determines the continuity of state transition information corresponding to adjacent time tags under the same source identifier. The verification results include verified records, abnormal records, temporary records, and group adjustment log reference records. For verified records, the system proceeds to standard data format conversion. For abnormal records, the system retains the original record, abnormal location, and network parameter version number. For temporary records, the system waits for the source identifier verification to complete before proceeding to standard data format conversion.

[0095] Further, the system performs standard data format conversion and version number generation based on the verification results to obtain standard data. The standard data format conversion involves converting the t-th observation state, t-th action selection, t-th reward, t+1-th observation state, fitness, initial optimal position, number of iterations, maximum number of iterations, network parameters, resource usage, timestamps, and source identifiers from the verification records into a unified record order. The standard data is a record of state transition information after the standard data format conversion. The version number is a storage record number generated by the system based on the training task grouping results, test question number, network parameter version number, and conversion timestamp. Each piece of standard data is bound to a version number. If the same set of state transition information enters the standard data format conversion at different times, the system retains the newer version number and writes the earlier version number to the grouping adjustment log.

[0096] Furthermore, the system establishes a multi-dimensional index structure based on the standard data, the time stamps, and the source identifiers. This multi-dimensional index structure is an index record used by the evaluation data warehouse to retrieve standard data, including time, source, training task, test question, and network parameter dimensions. The time dimension is formed by time stamps. The source dimension is formed by source identifiers. The training task dimension is formed by training task type and training task grouping results. The test question dimension is formed by test question numbers. The network parameter dimension is formed by network parameter version numbers. The system writes standard data under the same time and source dimensions into the same index location and binds the index location to the version number. If the same standard data corresponds to multiple training task dimensions, the system writes it into the main index location according to the grouping number in the training task grouping results and writes the remaining index locations into the grouping adjustment log.

[0097] Further, the system performs evaluation data warehouse storage processing. The evaluation data warehouse is a storage unit that receives standard data and a multi-dimensional index structure, including a standard data storage area, an index structure storage area, a verification result storage area, and a version number storage area. The standard data storage area receives state transition information records for completed standard data format conversion. The index structure storage area receives multi-dimensional index structures corresponding to the time dimension, source dimension, training task dimension, test question dimension, and network parameter dimension. The verification result storage area receives verification success records, exception records, and temporary records. The version number storage area receives the version number corresponding to the standard data. After the system completes the writing, it generates standardized stored data. The standardized stored data includes standard data, verification results, a multi-dimensional index structure, a version number, a timestamp, and a source identifier.

[0098] In one engineering implementation, the state transition information set output by the S300 comes from the resource scheduling strategy test group. This state transition information set includes the t-th observation state, t-th action selection, t-th reward, and t+1-th observation state corresponding to two hundred training iterations. The system first reads two hundred state transition information entries according to time tags, and then verifies the test question number and network parameter version number according to the source identifier. If one state transition information entry lacks the t-th reward, the system writes this record into the anomaly record and retains its adjacent record as a pass verification record. Subsequently, the system converts the pass verification records into standard data and generates a version number. The system establishes a time dimension based on the time tag, a source dimension based on the source identifier, a training task dimension based on the training task grouping results, a test question dimension based on the test question number, and a network parameter dimension based on the network parameter version number. After these dimensions are written into the evaluation data warehouse, standardized storage data is formed.

[0099] The standardized stored data is the output of S400 and serves as the input of S500. Specifically, S500 performs feature extraction based on the standardized stored data to obtain a set of feature vectors, and then performs evaluation model group and multi-level evaluation model analysis based on the set of feature vectors. The standard data is used in the feature extraction operation in S500. The multi-dimensional index structure in S500 is used to retrieve standard data corresponding to the same training task grouping results, the same test question, and the same network parameter version number. The verification result and version number in S500 are used to distinguish between verification records, abnormal records, and historical version records.

[0100] In summary, the technical effects of this step are as follows: This step converts the state transition information set into standardized stored data with verification results, a multi-dimensional index structure, and version numbers. The standardized stored data retains time stamps, source identifiers, and network parameter version numbers. This data provides searchable input for the S500 to perform evaluation model group and multi-level evaluation model analysis and processing.

[0101] S500. Based on the standardized stored data, the evaluation index data is obtained through analysis of the evaluation model group and the multi-level evaluation model.

[0102] Specifically, S500 is executed collaboratively by the evaluation model group and the multi-level evaluation model, with the input source being the standardized stored data generated by S400. The standardized stored data includes standard data, verification results, a multi-dimensional index structure, version number, timestamp, and source identifier. The evaluation model group is an analysis unit composed of multiple evaluation models, including a convergence efficiency evaluation model, a robustness evaluation model, a resource usage evaluation model, and a system stability evaluation model. The multi-level evaluation model is a model structure that further organizes and analyzes the output results of the evaluation model group, including basic scoring processing, indicator merging processing, and intermediate evaluation data table generation processing. After receiving the standardized stored data, the system first reads the multi-dimensional index structure, and then retrieves the corresponding standard data according to the time dimension, source dimension, training task dimension, test question dimension, and network parameter dimension. If the retrieved standard data corresponds to an abnormal record, the system writes the standard data to the abnormal record reference position and stops it from entering the evaluation model group. If the retrieved standard data corresponds to a verified record, the system sends it to the feature extraction operation.

[0103] Specifically, the system performs feature extraction based on the standardized stored data to obtain a feature vector set. The feature extraction operation reads the t-th observation state, t-th action selection, t-th reward, t+1-th observation state, fitness, initial optimal position, number of iterations, maximum number of iterations, network parameters, resource usage, time stamp, and source identifier from the standard data, and groups them according to the same training task and the same test question number. The feature vector set is a structured record set formed by the above fields in the order of the training stages. Each feature vector corresponds to one state transition information. The feature vector includes fitness change records, number of iterations records, maximum number of iterations records, t-th reward record, network parameter version number, resource usage record, runtime record, and source identifier. If there are multiple version numbers for the same test question, the system reads the newest version number and writes the earlier version number into the historical version record. If adjacent time stamps are not consecutive, the system retains empty positions in the feature vector set and writes the empty positions into the verification result reference record.

[0104] Further, the system performs scoring processing on the convergence efficiency evaluation model, the robustness evaluation model, the resource usage evaluation model, and the system stability evaluation model based on the feature vector set, obtaining an intermediate evaluation data table. The convergence efficiency evaluation model receives fitness, number of iterations, maximum number of iterations, and reward at time t, and outputs a convergence efficiency score. The robustness evaluation model receives dynamic perturbation, state transition information set, and real-time performance data, and outputs a robustness evaluation index. The resource usage evaluation model receives resource usage and runtime, and outputs a resource usage score. The system stability evaluation model receives statistical monitoring window, network parameters, and state transition information set, and outputs a system stability index. The inputs to the above four evaluation models all originate from the feature vector set. The outputs of the above four evaluation models are all entered into the same intermediate evaluation data table. The intermediate evaluation data table records data according to training task type, test question number, source identifier, time tag, and version number.

[0105] Furthermore, during the scoring process, the convergence efficiency evaluation model first reads the fitness corresponding to consecutive time labels under the same test item, then reads the number of iterations and the maximum number of iterations, and arranges the reward for the t-th time label in order. The convergence efficiency evaluation model maps fitness change records and reward records for the t-th time label to the same training phase to form a convergence efficiency score. During the scoring process, the robustness evaluation model reads the test item records corresponding to dynamic perturbations and maps the set of state transition information under dynamic perturbations to real-time performance data to form a robustness evaluation index. During the scoring process, the resource usage evaluation model reads the resource usage and runtime of the same training phase to form a resource usage score. During the scoring process, the system stability evaluation model reads the network parameters and state transition information set within the statistical monitoring window and performs continuous change statistics according to version number order to form a system stability index.

[0106] Furthermore, the system performs multi-level evaluation model analysis based on the intermediate evaluation data table to obtain evaluation index data. The multi-level evaluation model first reads the convergence efficiency score, robustness evaluation index, resource usage score, and system stability index from the intermediate evaluation data table, and then merges them according to training task type and test question number. After merging, the multi-level evaluation model writes the score records under the same source identifier into the same evaluation index data. The evaluation index data includes training task type, test question number, version number, convergence efficiency score, robustness evaluation index, resource usage score, system stability index, and verification result reference record. If there is an abnormal record reference position in the intermediate evaluation data table, the multi-level evaluation model writes that position into the evaluation index data and retains the corresponding source identifier. If there are multiple test question numbers under the same training task type, the multi-level evaluation model groups and merges them according to the training task grouping results to generate corresponding evaluation index data.

[0107] In one engineering implementation, the standardized storage data output by S400 comes from the resource scheduling strategy test group. This standardized storage data contains two hundred standard data entries and their corresponding version numbers. In S500, the system reads the standard data under the same test question number according to a multi-dimensional index structure, and extracts fitness, number of iterations, maximum number of iterations, reward at time t, network parameters, and resource usage to form a feature vector set. Subsequently, the convergence efficiency evaluation model reads fitness, number of iterations, maximum number of iterations, and reward at time t to generate a convergence efficiency score. The robustness evaluation model reads dynamic perturbation, state transition information set, and real-time performance data to generate a robustness evaluation index. The resource usage evaluation model reads resource usage and runtime to generate a resource usage score. The system stability evaluation model reads the statistical monitoring window, network parameters, and state transition information set to generate a system stability index. The multi-level evaluation model writes the above outputs into an intermediate evaluation data table, and then merges them according to training task type and test question number to obtain evaluation index data.

[0108] The evaluation index data is the output of S500 and serves as the input of S600. Specifically, S600 performs dynamic thresholding, statistical monitoring, and threshold comparison processing based on the evaluation index data to obtain the bottleneck identification result. The convergence efficiency score, robustness evaluation index, resource usage score, and system stability index are processed by the dynamic thresholding mechanism in S600. The training task type, test question number, version number, and verification result reference record are used in S600 for the evaluation model corresponding to the bottleneck, training task grouping results, and feature parameter association processing.

[0109] In summary, the technical effects of this step are as follows: This step converts standardized stored data into a set of feature vectors and generates an intermediate evaluation data table through an evaluation model group. The multi-level evaluation model then merges the intermediate evaluation data into evaluation index data. This evaluation index data provides multi-dimensional input for the S600 to generate the weakness identification results.

[0110] S600. The evaluation index data is compared with the dynamic threshold and the set evaluation threshold, and the shortcoming identification result is obtained by combining the statistical monitoring window.

[0111] Specifically, S600 is executed by the bottleneck identification module, with the input source being the evaluation index data generated by S500. The evaluation index data includes training task type, test question number, version number, convergence efficiency score, robustness evaluation index, resource usage score, system stability index, and verification result reference record. The dynamic threshold mechanism is a processing mechanism that sets threshold comparison rules based on different training task types, test question numbers, and version numbers. The statistical monitoring window is a fixed observation interval that continuously reads the evaluation index data according to time labels. The set evaluation thresholds are threshold fields pre-recorded in the training strategy, corresponding to the convergence efficiency score, robustness evaluation index, resource usage score, and system stability index, respectively. After receiving the evaluation index data, the bottleneck identification module first reads the verification result reference record. If a certain evaluation index data corresponds to an abnormal record, the bottleneck identification module writes it to the abnormal record reference position and pauses the processing of that evaluation index data in the dynamic threshold mechanism. If a certain evaluation index data corresponds to a passing verification record, the bottleneck identification module continues to read the training task type, test question number, and version number.

[0112] Specifically, the bottleneck identification module performs dynamic thresholding based on the evaluation metric data and the set evaluation thresholds to obtain threshold comparison results. The dynamic thresholding process includes threshold field reading, evaluation metric data alignment, and comparison state generation. Threshold field reading involves reading the set evaluation thresholds corresponding to the convergence efficiency score, robustness evaluation metric, resource usage score, and system stability metric from the training strategy. Evaluation metric data alignment involves matching the four types of metrics in the evaluation metric data with their corresponding set evaluation thresholds according to the training task type, test question number, and version number. Comparison state generation involves comparing each type of evaluation metric data with its corresponding set evaluation threshold and generating a threshold comparison result. The threshold comparison result includes the training task type, test question number, version number, evaluation model type, comparison state, time label, and source identifier. If a set evaluation threshold is missing, the bottleneck identification module reads the set evaluation threshold of the adjacent version from the training strategy reference record and writes the read position to the group adjustment log. If no set evaluation threshold exists for the adjacent version, the bottleneck identification module writes the evaluation metric data to a temporary storage record.

[0113] Furthermore, the bottleneck identification module performs continuous change statistical processing based on the threshold comparison results and the statistical monitoring window to obtain the evaluation model corresponding to the bottleneck. The continuous change statistical processing involves reading the threshold comparison results for the same training task type, the same test question number, and the same version number within the statistical monitoring window in time-labeled order, and statistically analyzing the continuous changes in convergence efficiency score, robustness evaluation index, resource usage score, and system stability index. If the comparison status corresponding to the same evaluation model is lower than the corresponding set evaluation threshold in multiple consecutive statistical monitoring windows, the bottleneck identification module records this evaluation model as the evaluation model corresponding to the bottleneck. If the comparison status changes at the boundary position of the statistical monitoring window, the bottleneck identification module invokes the delayed effect mechanism, writes the boundary position into a temporary record, and continues reading in the next statistical monitoring window. The evaluation model corresponding to the bottleneck includes one or more records of the convergence efficiency evaluation model, robustness evaluation model, resource usage evaluation model, and system stability evaluation model.

[0114] Further, the bottleneck identification module performs association processing based on the evaluation model corresponding to the bottleneck, the training task type, the training task grouping result, the state transition information set, and the feature parameters to obtain the bottleneck identification result. The association processing first reads the grouping number and feature parameters from the training task grouping result generated in S100 according to the training task type and test question number. Then, the bottleneck identification module reads the corresponding t-th observation state, t-th selected action, t-th reward, t+1-th observation state, network parameters, and resource usage from the state transition information set generated in S300 according to the source identifier and version number. The bottleneck identification module binds the above fields with the evaluation model corresponding to the bottleneck to form the bottleneck identification result. The bottleneck identification result includes the evaluation model corresponding to the bottleneck, the training task type, the training task grouping result, the test question number, the version number, the feature parameters, the state transition information set, the threshold comparison result, the statistical monitoring window, and the source identifier. If the same test question number corresponds to multiple evaluation models corresponding to bottlenecks, the bottleneck identification module writes multiple bottleneck records in time-stamped order and retains the same version number. If the state transition information set lacks a corresponding source identifier, the shortcoming identification module writes the shortcoming record into the anomaly record reference position.

[0115] In one engineering implementation, the evaluation index data output by the S500 comes from resource scheduling strategy test groups. This evaluation index data includes twenty test item numbers and version numbers corresponding to two hundred training iterations. The bottleneck identification module first reads the convergence efficiency score, robustness evaluation index, resource usage score, and system stability index for each test item number, and then reads the set evaluation threshold from the training strategy. The bottleneck identification module uses ten time labels as a statistical monitoring window, continuously reading the threshold comparison results for the same test item number. If the robustness evaluation index is lower than the corresponding set evaluation threshold in three consecutive statistical monitoring windows, the bottleneck identification module records the robustness evaluation model as the evaluation model corresponding to the bottleneck, and reads the group number and feature parameters corresponding to that test item number from the training task grouping results. Subsequently, the bottleneck identification module reads the corresponding t-th observation state, t-th choice action, t-th reward, t+1-th observation state, and network parameters from the state transition information set to form the bottleneck identification result.

[0116] The weakness identification result is the output of S600 and serves as the input of S700. Specifically, S700 performs policy library matching, training policy adjustment, and next round of training execution based on the weakness identification result. The evaluation model corresponding to the weakness is used for policy library retrieval processing in S700. The training task type, the training task grouping result, the state transition information set, and the feature parameters are used for candidate adjustment policy matching processing in S700. The threshold comparison result and the statistical monitoring window are used for writing the training policy adjustment result in S700.

[0117] In summary, the technical effects of this step are as follows: This step converts the evaluation index data into a weakness identification result with corresponding evaluation models, threshold comparison results, and statistical monitoring windows. The weakness identification result correlates the output of the evaluation model group with the training task grouping results, state transition information set, and feature parameters. This result provides direct input for the S700 to perform policy library matching and training policy adjustment.

[0118] S700. Based on the weakness identification results, match the strategy library and adjust the training strategy, execute the next round of iterative training until the stopping condition is met, and output the optimized algorithm model and evaluation report.

[0119] Specifically, S700 is executed collaboratively by the strategy library module and the dynamic training module, with the input source being the bottleneck identification results generated by S600. The bottleneck identification results include the evaluation model corresponding to the bottleneck, the training task type, the training task grouping results, the test question number, the version number, feature parameters, the state transition information set, the threshold comparison results, the statistical monitoring window, and the source identifier. The strategy library is a set of rules storing candidate adjustment strategies, including convergence efficiency adjustment strategies, robustness adjustment strategies, resource usage adjustment strategies, and system stability adjustment strategies. The dynamic training module is the execution unit that reads candidate adjustment strategies and rewrites the training strategy, connecting the strategy library module, the training execution module, and the data acquisition module. After receiving the bottleneck identification results, the system first reads the evaluation model corresponding to the bottleneck, then reads the training task type and the training task grouping results. If the bottleneck identification result lacks a source identifier, the system writes the bottleneck identification result into the exception log. If the bottleneck identification result has a version number conflict, the system reads the newer version number from S400 and writes the earlier version number into the grouping adjustment log.

[0120] Specifically, the strategy library module performs strategy library retrieval processing based on the bottleneck identification results to obtain candidate adjustment strategies. The strategy library retrieval processing involves matching the evaluation model corresponding to the bottleneck with the adjustment strategies in the strategy library. When the evaluation model corresponding to the bottleneck is a convergence efficiency evaluation model, the strategy library module retrieves convergence efficiency adjustment strategies. When the evaluation model corresponding to the bottleneck is a robustness evaluation model, the strategy library module retrieves robustness adjustment strategies. When the evaluation model corresponding to the bottleneck is a resource usage evaluation model, the strategy library module retrieves resource usage adjustment strategies. When the evaluation model corresponding to the bottleneck is a system stability evaluation model, the strategy library module retrieves system stability adjustment strategies. If the same bottleneck identification result corresponds to multiple evaluation models, the strategy library module retrieves the corresponding adjustment strategies sequentially according to the time tag order and writes the retrieval order into the group adjustment log. The candidate adjustment strategies include strategy type, applicable training task type, associated test question number, associated network parameters, associated dynamic threshold mechanism, and strategy version number.

[0121] Further, the dynamic training module performs matching processing based on the candidate adjustment strategy, the training task grouping results, the network parameters, and the dynamic threshold mechanism to obtain the training strategy adjustment result. The matching processing first reads the strategy type from the candidate adjustment strategy, and then reads the group number, test question number, and feature parameters from the training task grouping results. Subsequently, the dynamic training module reads the t-th observation state, t-th action selection, t-th reward, t+1-th observation state, network parameters, and resource usage from the state transition information set. The dynamic training module maps these fields to the candidate adjustment strategy to generate the training strategy adjustment result. The training strategy adjustment result includes the training task grouping result to be adjusted, the test question number to be adjusted, the network parameters to be adjusted, the data acquisition frequency parameter to be adjusted, the dynamic threshold mechanism to be adjusted, the statistical monitoring window to be adjusted, and the candidate adjustment strategy reference record. If the candidate adjustment strategy does not match the training task grouping result, the dynamic training module writes the candidate adjustment strategy into a temporary storage record. If the candidate adjustment strategy does not match the network parameter version number, the dynamic training module reads the adjacent version numbers under the same source identifier and writes the matching position in the group adjustment log.

[0122] Furthermore, when the candidate adjustment strategy is a convergence efficiency adjustment strategy, the dynamic training module performs training strategy adjustment processing based on the t-th selected action, the network parameters, and the maximum number of iterations. This processing reads the t-th selected action and the t-th reward under continuous time labels in the state transition information set, and reads the network parameter version number. The dynamic training module writes the network parameters corresponding to the action network into the training strategy adjustment result, and writes the maximum number of iterations into the same training strategy adjustment result. After this processing is completed, the action network call record, network parameter version number, and maximum number of iterations record in the training strategy are updated.

[0123] Furthermore, when the candidate adjustment strategy is a robust adjustment strategy, the dynamic training module performs training strategy adjustment processing based on dynamic perturbations, the test questions, and the state transition information set. This processing reads the test question number corresponding to the weakness identification result and reads the test questions under the same group number from the training task grouping result. The dynamic training module writes the dynamic perturbation record into the candidate adjustment strategy reference record and writes the t-th observation state, t-th selection action, t-th reward, and t+1-th observation state under the corresponding source identifier in the state transition information set into the training strategy adjustment result. After this processing is completed, the test question grouping record and dynamic perturbation record in the training strategy are updated.

[0124] Furthermore, when the candidate adjustment strategy is a resource usage adjustment strategy, the dynamic training module performs training strategy adjustment processing based on the data acquisition frequency parameter, the training task grouping results, and the resource scheduling strategy. This processing reads the bottleneck identification result corresponding to the resource usage score and retrieves the resource usage and runtime from the state transition information set. The dynamic training module binds the data acquisition frequency parameter, resource scheduling strategy, and training task grouping results to obtain the resource usage adjustment record in the training strategy adjustment result. After this processing is completed, the data acquisition frequency parameter, resource scheduling strategy reference record, and group number invocation order in the training strategy are updated.

[0125] Furthermore, when the candidate adjustment strategy is a system stability adjustment strategy, the dynamic training module performs training strategy adjustment processing based on the dynamic threshold mechanism, the statistical monitoring window, and the delayed activation mechanism. This processing reads the threshold comparison results corresponding to the system stability index and reads the network parameter change records and state transition information set within the statistical monitoring window. The dynamic training module writes the set evaluation threshold, statistical monitoring window, and delayed activation mechanism from the dynamic threshold mechanism into the training strategy adjustment results. If the threshold comparison result changes at the boundary position of the statistical monitoring window, the dynamic training module writes this boundary position into a temporary record and reads the corresponding record of the delayed activation mechanism in the next round of training execution.

[0126] Furthermore, the dynamic training module updates the training strategy based on the adjustment results, resulting in an updated training strategy. This update process involves writing the adjustment results into the current training strategy and generating a new version number. The updated training strategy includes the training task type, training task grouping results, test question number, maximum iteration count, data acquisition frequency parameter, set evaluation threshold, network parameters, dynamic threshold mechanism, statistical monitoring window, delayed activation mechanism, candidate adjustment strategy reference record, and version number. If the adjustment results correspond to multiple candidate adjustment strategies, the dynamic training module writes them into the updated training strategy according to the time label order of the evaluation model corresponding to the weakest link. If the updated training strategy and the current training strategy have different records for the same field, the dynamic training module retains the record corresponding to the newer version number and writes the earlier record into the grouping adjustment log.

[0127] Further, the training execution module performs the next round of training execution based on the updated training strategy, obtaining the optimized algorithm model and evaluation report. The next round of training execution reads the updated training strategy and training task grouping results, and re-drives the optimized algorithm model to be evaluated to train the test questions. During training execution, raw behavioral data streams continue to be generated, and these raw behavioral data streams continue to enter S300. If the next round of training execution reaches the maximum number of iterations, the training execution module reads the optimized algorithm model under the current version number. If the convergence efficiency score, robustness evaluation index, resource usage score, and system stability index all meet the set evaluation thresholds within the corresponding statistical monitoring windows, the dynamic training module stops the training strategy update process and outputs the optimized algorithm model and evaluation report. The evaluation report includes the training task type, training task grouping results, bottleneck identification results, candidate adjustment strategies, training strategy adjustment results, updated training strategy, version number, and source identifier.

[0128] In one engineering implementation, the bottleneck identification result output by S600 shows the bottleneck corresponding to the robustness evaluation model in the resource scheduling strategy test group. After reading the bottleneck identification result, the strategy library module retrieves robust adjustment strategies and obtains candidate adjustment strategies. The dynamic training module reads the corresponding test question number and state transition information set, writes the dynamic disturbance record into the training strategy adjustment result, and rewrites the corresponding test question into the training task group result. Subsequently, the dynamic training module generates an updated training strategy. After reading the updated training strategy, the training execution module re-drives the optimization algorithm model to be evaluated for the next round of training execution. The data acquisition module continues to collect raw behavioral data streams, and subsequently, S300, S400, S500, and S600 continue to run according to the aforementioned process. When the stopping condition is reached, the system outputs the optimization algorithm model and evaluation report.

[0129] The optimization algorithm model and the evaluation report are the outputs of S700. The optimization algorithm model retains the network parameters, action network, and value network corresponding to the updated training strategy. The evaluation report retains the weakness identification results, candidate adjustment strategies, training strategy adjustment results, updated training strategy, and version number. If the system enters the next round of training execution, the updated training strategy flows back to S200 for processing and invocation of training execution, time label addition, and source identifier addition in S200. The raw behavioral data stream generated by the next round of training execution continues to enter S300, forming a closed-loop operation process.

[0130] In summary, the technical effects of this step are as follows: This step converts the weakness identification results into candidate adjustment strategies and training strategy adjustment results. The updated training strategy is then fed back into the next round of training. This process connects the weakness records corresponding to the evaluation index data with the training execution of the optimization algorithm model in a closed loop.

[0131] In one specific embodiment:

[0132] In S700, based on the weakness identification results, a strategy library is matched and the training strategy is adjusted. The next round of iterative training is executed until the stopping condition is met, and the optimized algorithm model and evaluation report are output.

[0133] The S700 is executed collaboratively by the policy library module and the dynamic training module. The input source is the bottleneck identification result generated by the S600, which includes the evaluation model corresponding to the bottleneck, the training task type, the training task grouping result, the test item number, the version number, feature parameters, the state transition information set, the threshold comparison result, the statistical monitoring window, and the source identifier. The policy library is a set of rules storing candidate adjustment strategies, including convergence efficiency adjustment strategies, robustness adjustment strategies, resource usage adjustment strategies, and system stability adjustment strategies. The dynamic training module is the execution unit that reads candidate adjustment strategies and rewrites the training strategy, connecting the policy library module, the training execution module, and the data acquisition module. After receiving the bottleneck identification result, the system first reads the evaluation model corresponding to the bottleneck, then reads the training task type and the training task grouping result. If the bottleneck identification result lacks a source identifier, the system writes the bottleneck identification result to the exception log; if there is a version number conflict, the system reads the newer version number from the S400 and writes the earlier version number to the grouping adjustment log.

[0134] The strategy library module performs strategy library retrieval processing based on the bottleneck identification results to obtain candidate adjustment strategies. Strategy library retrieval processing involves matching the evaluation model corresponding to the bottleneck with the adjustment strategies in the strategy library by field: if the evaluation model corresponding to the bottleneck is a convergence efficiency evaluation model, it retrieves convergence efficiency adjustment strategies; if it is a robustness evaluation model, it retrieves robustness adjustment strategies; if it is a resource usage evaluation model, it retrieves resource usage adjustment strategies; and if it is a system stability evaluation model, it retrieves system stability adjustment strategies. If the same bottleneck identification result corresponds to multiple evaluation models, the strategy library module retrieves the corresponding adjustment strategies sequentially according to the time tag order and writes the retrieval order to the group adjustment log. Candidate adjustment strategies include strategy type, applicable training task type, associated test item number, associated network parameters, associated dynamic threshold mechanism, and strategy version number.

[0135] To quantify the severity of shortcomings and drive policy matching, the dynamic training module first transforms the threshold comparison results within multiple consecutive statistical monitoring windows in the shortcomings identification results into a comprehensive "shortcoming strength index". Formula ① uses a focus loss weighting method to nonlinearly amplify deviations that are continuously below the set evaluation threshold, so that persistent and severe shortcomings receive higher adjustment priority.

[0136] Formula ①:

[0137]

[0138] in, The weakest link strength index indicates that the larger the value, the more serious and persistent the weakness.

[0139] : Summation index, indicating the index within the statistical monitoring window. Several windows, with values ​​ranging from 1 to... Integers;

[0140] The number of windows within the statistical monitoring window that are continuously below the set evaluation threshold is derived from the statistical results of the continuous changes in the statistical monitoring window.

[0141] : No. The actual evaluation metric values ​​(such as convergence efficiency score, robustness evaluation metric, etc.) within each window are derived from the evaluation metric data output by S500.

[0142] The corresponding evaluation threshold is derived from the training strategy;

[0143] : Focus loss adjustment factor, which takes the value of a positive real number and is preset by the policy library according to the training task type;

[0144] The natural logarithm function acts on... .

[0145] Simple numerical example: Suppose that for a certain test item, the robustness evaluation index is below the set evaluation threshold for three consecutive windows. Three windows The evaluation thresholds were set to 0.75, 0.70, and 0.65 respectively. Regulatory factors Then calculate the deviation ratio for each window. The focus weights are 0.882, 0.824, and 0.765, respectively. , , After taking the logarithm and summing them up, we get This metric, once output, will be substituted into the cost function of subsequent policy matching to determine whether to trigger higher-order adjustment actions.

[0146] The dynamic training module then performs matching processing based on candidate adjustment strategies, training task grouping results, network parameters, and a dynamic threshold mechanism to obtain the training strategy adjustment result. The matching process first reads the strategy type from the candidate adjustment strategies, then reads the group number, test item number, and feature parameters from the training task grouping results, and subsequently reads the t-th observation state, t-th action selection, t-th reward, t+1-th observation state, network parameters, and resource usage from the state transition information set. The dynamic training module maps these fields to the candidate adjustment strategies to generate the training strategy adjustment result. The training strategy adjustment result includes the training task grouping result to be adjusted, the test item number to be adjusted, the network parameters to be adjusted, the data acquisition frequency parameter to be adjusted, the dynamic threshold mechanism to be adjusted, the statistical monitoring window to be adjusted, and the candidate adjustment strategy reference record. If the candidate adjustment strategy does not match the training task grouping result, the dynamic training module writes the candidate adjustment strategy to a temporary record; if it does not match the network parameter version number, it reads the adjacent version number under the same source identifier and writes it to the group adjustment log.

[0147] For different candidate adjustment strategies, the dynamic training module performs differentiated adjustment processing. When the candidate adjustment strategy is a convergence efficiency adjustment strategy, the dynamic training module adjusts the training strategy based on the t-th selected action, network parameters, and maximum number of iterations. This processing reads the t-th selected action and t-th reward under continuous time labels in the state transition information set, reads the network parameter version number, writes the network parameters corresponding to the action network into the training strategy adjustment result, and writes the maximum number of iterations into the same result. When the candidate adjustment strategy is a robust adjustment strategy, the module adjusts based on dynamic perturbation, test questions, and the state transition information set. It reads the test question number corresponding to the weakness identification result, reads the test questions under the same group number from the training task grouping result, writes the dynamic perturbation record into the candidate adjustment strategy reference record, and writes the t-th observation state, t-th selected action, t-th reward, and t+1-th observation state under the corresponding source identifier in the state transition information set into the training strategy adjustment result. When the candidate adjustment strategy is a resource usage adjustment strategy, the module adjusts based on the data acquisition frequency parameter, training task grouping results, and resource scheduling strategy. It reads the bottleneck identification results corresponding to the resource usage score, retrieves resource usage and runtime from the state transition information set, and binds the data acquisition frequency parameter, resource scheduling strategy, and training task grouping results to obtain a resource usage adjustment record. When the candidate adjustment strategy is a system stability adjustment strategy, the module adjusts based on a dynamic threshold mechanism, statistical monitoring window, and delayed effectiveness mechanism. It reads the threshold comparison results corresponding to the system stability index, reads the network parameter change records and state transition information set within the statistical monitoring window, and writes the set evaluation threshold, statistical monitoring window, and delayed effectiveness mechanism from the dynamic threshold mechanism into the training strategy adjustment result.

[0148] To determine the optimal adjustment combination when multiple adjustment strategies are triggered simultaneously, the dynamic training module introduces a cost function based on a Lyapunov function. This function measures the deviation between the current training strategy and the ideal stable state, and selects the best-fitting adjustment action by minimizing this deviation. Equation ② defines a first-order difference Lyapunov candidate function, and Equation ③ constructs the adjustment cost based on this function.

[0149] Formula ②:

[0150]

[0151] in, : The energy deviation between the current training strategy and the reference steady state;

[0152] The change in network parameters to be adjusted (the difference between the action network parameters in the candidate adjustment strategy and the current network parameters) comes from the network parameter field in the state transition information set;

[0153] The change in the data acquisition frequency parameter (the difference between the suggested frequency in the candidate adjustment strategy and the current frequency) originates from the acquisition frequency parameter in the training strategy.

[0154] The change in the dynamic disturbance record (the intensity increment of the dynamic disturbance in the candidate adjustment strategy) comes from the dynamic disturbance field in the state transition information set.

[0155] Simple numerical example: Assume the change in network parameters Changes in data acquisition frequency Hz, dynamic disturbance change ,but .Should The smaller the value, the closer the current adjustment strategy is to a stable state.

[0156] Formula ③:

[0157]

[0158] in, Adjust the total cost to compare the merits of different candidate adjustment strategies;

[0159] Same as the definition in formula ②;

[0160] Balance factor, preset by the policy library according to the type of training task;

[0161] Same as the definition in formula ①.

[0162] Simple numerical example: Continuing from the previous example, , , ,but The dynamic training module will select such that... The smallest candidate adjustment strategy is used as the actual strategy. After this process, the dynamic training module updates the training strategy based on the adjustment results, resulting in an updated training strategy. The training strategy update process involves writing the adjustment results into the current training strategy and generating a new version number. The updated training strategy includes the training task type, training task grouping results, test question number, maximum number of iterations, data acquisition frequency parameters, set evaluation threshold, network parameters, dynamic threshold mechanism, statistical monitoring window, delayed activation mechanism, candidate adjustment strategy reference record, and version number. If the adjustment results correspond to multiple candidate adjustment strategies, the dynamic training module writes them into the updated training strategy according to the time label order of the evaluation model corresponding to the shortest component. If there are different records with the same field as the current training strategy, the record corresponding to the newer version number is retained, and the earlier record is written to the grouping adjustment log. The output of this section is the updated training strategy, which will be added and processed by the S200's training execution, time label, and source identifier.

[0163] The training execution module performs the next round of training execution based on the updated training strategy, obtaining the optimized algorithm model and evaluation report. The next round of training execution reads the updated training strategy and training task grouping results, and re-drives the optimized algorithm model to be evaluated to train the test questions. During training execution, raw behavioral data streams continue to be generated, and these raw behavioral data streams continue to enter S300. To dynamically determine whether to stop adjustment, the system calculates a comprehensive convergence index after each iteration. This index is obtained by weighting the scores of the four evaluation models and uses a complementary filtering mechanism to smooth fluctuations and avoid false stops caused by early oscillations. Formula ④ gives the comprehensive score of the fusion convergence efficiency score, robustness evaluation index, resource utilization score, and system stability index.

[0164] Formula ④:

[0165]

[0166] in, The overall score, ranging from 0 to 1, represents the overall performance of the model.

[0167] The weighting coefficients of each evaluation indicator are preset by the policy library according to the training task type, satisfying the following requirements: ;

[0168] Convergence efficiency score, derived from the output of the convergence efficiency evaluation model in S500;

[0169] : Robustness evaluation metrics, derived from the output of the robustness evaluation model in S500;

[0170] Resource usage score, derived from the output of the resource usage assessment model in S500;

[0171] System stability index, derived from the output of the system stability assessment model in S500.

[0172] Simple numerical example: Assume a convergence efficiency score of 0.92, a robustness evaluation index of 0.78, a resource utilization score of 0.85, and a system stability index of 0.88, with weighting coefficients of 0.3, 0.3, 0.2, and 0.2 respectively. .

[0173] Formula ⑤ uses complementary filtering to smooth the combined score of two consecutive iterations, avoiding the impact of a single abnormal fluctuation on the stopping condition determination:

[0174]

[0175] in, : No. The smoothed overall score after rounds of iteration;

[0176] : Filter coefficients (complementary filter coefficients), with values ​​ranging from 1 to 2. Preset by the dynamic training module;

[0177] The first one obtained from formula ④ The overall score of each iteration;

[0178] : No. The smoothed comprehensive score of the round of iterations is initially set to 0;

[0179] : Iteration round index, a positive integer.

[0180] Simple numerical example: Let's assume the previous round's smooth score. Current round's overall score Filter coefficients ,but If the next training round reaches the maximum number of iterations, and all individual metrics meet the set evaluation threshold within the corresponding statistical monitoring window (i.e., ... , , , ), while smoothing the overall score If the preset stopping threshold (e.g., 0.85) is exceeded, the dynamic training module stops processing the training strategy update and outputs the optimized algorithm model and evaluation report. The evaluation report includes the training task type, training task grouping results, weakness identification results, candidate adjustment strategies, training strategy adjustment results, updated training strategy, version number, and source identifier. The optimized algorithm model retains the network parameters, action network, and value network corresponding to the updated training strategy. If the system enters the next round of training execution, the updated training strategy flows back to S200 for processing calls in S200's training execution, time label, and source identifier addition; the raw behavioral data stream generated in the next round of training execution continues to enter S300, forming a closed-loop operation. The output products of this section are the optimized algorithm model and evaluation report, which are the final outputs of this invention.

[0181] This section summarizes the technical effects: This step converts the weakness identification results into candidate adjustment strategies and training strategy adjustment results. It achieves multi-strategy optimization by using a weakness strength index weighted by focus loss and a Lyapunov cost function, and achieves stable stopping determination by smoothing the comprehensive score through complementary filtering, forming a closed-loop self-evolutionary architecture of evaluation-diagnosis-adjustment-retraining.

[0182] Example 2: Figure 2 This diagram illustrates the structural block diagram of a multi-dimensional evaluation and dynamic training system for an optimization algorithm model according to an embodiment of the present invention. Figure 2 As shown, the structure may include:

[0183] The training task processing module 01 is used to acquire the training task type, test questions, the optimization algorithm model to be evaluated, and the training strategy; perform data fusion, feature extraction, and initial grouping processing to obtain the training task grouping result; and send the training task grouping result to the training execution and data acquisition module. Specifically, the training task processing module receives the externally input training task type, test questions, optimization algorithm model to be evaluated, and training strategy. The training task processing module sets up a data entry point, a feature extraction interface, and an initial grouping interface. The data entry point performs field verification on the training task type, the test questions, the optimization algorithm model to be evaluated, and the training strategy. The feature extraction interface extracts feature parameters from the test questions and associates the feature parameters with the training task type. The initial grouping interface performs similarity matrix calculation and initial grouping processing based on the feature parameters and records the grouping adjustment log. The training task processing module combines the group number, test question number, training task type, feature parameters, training strategy reference record, and grouping adjustment log into the training task grouping result and sends it to the training execution and data acquisition module through the module connection interface.

[0184] The training execution and data acquisition module 02 is used to perform training execution, add time tags and source identifiers based on the training task grouping results to obtain a raw behavior data stream, and send the raw behavior data stream to the state transition information processing module. Specifically, the training execution and data acquisition module receives the training task grouping results from the training task processing module. The training execution and data acquisition module sets up a training execution interface, a data acquisition interface, a time tag interface, and a source identifier interface. The training execution interface reads the group number, test question number, and training strategy reference record from the training task grouping results and drives the optimization algorithm model to be evaluated to perform training execution. The data acquisition interface receives the running records formed by the training execution according to the data acquisition frequency parameter. The time tag interface writes a time tag to each running record. The source identifier interface writes a source identifier to each running record. The training execution and data acquisition module assembles the group number, test question number, training task type, training strategy reference record, time tag, source identifier, action network output record, value network output record, and network parameter version number into a raw behavior data stream, and sends the raw behavior data stream to the state transition information processing module.

[0185] The state transition information processing module 03 is used to extract and process fitness, initial optimal position, t-th observation state, t-th action selection, t-th reward, t+1-th observation state, network parameters, and resource usage based on the original behavioral data stream, to obtain a state transition information set, and then send the state transition information set to the standardized storage module. Specifically, the state transition information processing module receives the original behavioral data stream from the training execution and data acquisition module. The state transition information processing module is equipped with a sequence sorting interface, a training phase feature extraction interface, a t-th iteration feature extraction interface, and a correlation processing interface. The sequence sorting interface arranges the original behavioral data stream according to time tags and merges the running records under the same test question number according to the source identifier. The training phase feature extraction interface reads fitness, initial optimal position, number of iterations, maximum number of iterations, network parameters, and resource usage from the original behavioral data stream. The t-th iteration feature extraction interface reads the t-th observation state, t-th action selection, t-th reward, and t+1-th observation state from the original behavioral data stream. The association processing interface binds the training phase features and the t-th iteration features according to the same time label, source identifier, and network parameter version number to obtain a state transition information set. The state transition information processing module sends the state transition information set to the standardized storage module.

[0186] The standardized storage module 04 is used to perform standard data format conversion, multidimensional index structure establishment, and evaluation data warehouse storage processing based on the state transition information set, obtaining standardized storage data, and sending the standardized storage data to the evaluation and analysis module. Specifically, the standardized storage module receives the state transition information set from the state transition information processing module. The standardized storage module is equipped with a format verification interface, an integrity check interface, a standard data format conversion interface, a multidimensional index structure establishment interface, and an evaluation data warehouse storage interface. The format verification interface checks the record format of the t-th observation state, t-th selection action, t-th reward, t+1-th observation state, fitness, initial optimal position, network parameters, and resource usage. The integrity check interface checks the continuity of adjacent state transition information according to time tags. The standard data format conversion interface converts the verified records into standard data and generates a version number. The multidimensional index structure establishment interface establishes a multidimensional index structure based on time tags, source identifiers, training task grouping results, test question numbers, and network parameter version numbers. The evaluation data warehouse storage interface writes standard data, verification results, multidimensional index structure, version number, timestamp, and source identifier into the evaluation data warehouse to obtain standardized storage data, and then sends the standardized storage data to the evaluation analysis module.

[0187] Evaluation and analysis module 05 is used to perform evaluation model group and multi-level evaluation model analysis and processing based on the standardized stored data to obtain evaluation index data, and then send the evaluation index data to the bottleneck identification module. Specifically, the evaluation and analysis module receives standardized stored data from the standardized storage module. The evaluation and analysis module sets up a feature extraction interface, an evaluation model group interface, and a multi-level evaluation model interface. The feature extraction interface reads the standard data according to the multi-dimensional index structure and extracts fitness, number of iterations, maximum number of iterations, reward at time t, network parameters, resource usage, running time, time label, and source identifier from the standard data to obtain a feature vector set. The evaluation model group interface calls the convergence efficiency evaluation model, robustness evaluation model, resource usage evaluation model, and system stability evaluation model to score the feature vector set and obtain an intermediate evaluation data table. The multi-level evaluation model interface reads the intermediate evaluation data table and merges it according to training task type, test question number, version number, and source identifier to obtain evaluation index data. The evaluation and analysis module sends the evaluation index data to the bottleneck identification module.

[0188] The bottleneck identification module 06 is used to perform dynamic thresholding, statistical monitoring, and set evaluation threshold comparison processing based on the evaluation index data to obtain the bottleneck identification result, and then send the bottleneck identification result to the strategy library matching module. Specifically, the bottleneck identification module receives evaluation index data from the evaluation analysis module. The bottleneck identification module sets up a dynamic thresholding mechanism interface, a statistical monitoring window interface, a set evaluation threshold comparison interface, and an association processing interface. The dynamic thresholding mechanism interface reads the set evaluation thresholds in the training strategy and aligns the evaluation index data according to the training task type, test question number, and version number. The set evaluation threshold comparison interface compares the convergence efficiency score, robustness evaluation index, resource usage score, and system stability index with the corresponding set evaluation thresholds to obtain threshold comparison results. The statistical monitoring window interface continuously reads the threshold comparison results according to time tags and counts the comparison status of the same evaluation model in multiple statistical monitoring windows. The association processing interface associates the training task type, training task grouping results, state transition information set, and feature parameters according to the evaluation model corresponding to the bottleneck to obtain the bottleneck identification result. The shortcoming identification module sends the shortcoming identification result to the strategy library matching module.

[0189] The strategy library matching module 07 is used to perform strategy library retrieval processing based on the shortcoming identification results, obtain candidate adjustment strategies, and send the candidate adjustment strategies to the dynamic training module. Specifically, the strategy library matching module receives the shortcoming identification results from the shortcoming identification module. The strategy library matching module sets up a strategy library interface, a retrieval interface, and a strategy version number recording interface. The strategy library interface stores convergence efficiency adjustment strategies, robustness adjustment strategies, resource usage adjustment strategies, and system stability adjustment strategies. The retrieval interface reads the evaluation model corresponding to the shortcoming in the shortcoming identification results and matches the evaluation model corresponding to the shortcoming with the adjustment strategies in the strategy library. When the evaluation model corresponding to the shortcoming is a convergence efficiency evaluation model, a convergence efficiency adjustment strategy is retrieved. When the evaluation model corresponding to the shortcoming is a robustness evaluation model, a robustness adjustment strategy is retrieved. When the evaluation model corresponding to the shortcoming is a resource usage evaluation model, a resource usage adjustment strategy is retrieved. When the evaluation model corresponding to the shortcoming is a system stability evaluation model, a system stability adjustment strategy is retrieved. The strategy version number recording interface writes the strategy type, training task type, test question number, network parameters, dynamic threshold mechanism, and strategy version number into the candidate adjustment strategy, and sends the candidate adjustment strategy to the dynamic training module.

[0190] The dynamic training module 08 is used to perform matching processing based on the candidate adjustment strategies, the training task grouping results, the network parameters, and the dynamic threshold mechanism to obtain training strategy adjustment results. Based on the training strategy adjustment results, it performs training strategy update processing to obtain an updated training strategy. Based on the updated training strategy, it performs the next round of training execution processing to obtain an optimized algorithm model and an evaluation report. Specifically, the dynamic training module receives candidate adjustment strategies from the strategy library matching module and calls the training task grouping results in the training task processing module, the network parameters in the state transition information processing module, and the dynamic threshold mechanism in the bottleneck identification module. The dynamic training module sets up a matching processing interface, a training strategy update interface, and a next round of training execution interface. The matching processing interface reads the strategy type from the candidate adjustment strategies and matches the strategy type with the training task grouping results, network parameters, and dynamic threshold mechanism to obtain the training strategy adjustment results. The training strategy update interface writes the training strategy adjustment results into the current training strategy and generates an updated training strategy. The next round of training execution interface sends the updated training strategy to the training execution and data acquisition module to drive the next round of training execution. After the next round of training is completed, the dynamic training module reads the optimization algorithm model and generates an evaluation report based on the weakness identification results, candidate adjustment strategies, training strategy adjustment results, updated training strategy, version number, and source identifier.

Claims

1. A method for multi-dimensional evaluation and dynamic training of an optimization algorithm model, characterized in that, include: S100: Obtain the training task type, test questions, algorithm model to be evaluated and training strategy; after standardizing the test questions, perform data fusion, feature extraction and initial grouping to obtain the training task grouping results. S200. Based on the training task grouping results, start iterative training, synchronously collect running records and add time tags and source identifiers to obtain the raw behavioral data stream; S300. Based on the original behavioral data stream, extract fitness, initial optimal position, state transition related information, network parameters, and resource usage to form a state transition information set. S400: Perform format verification, standard data conversion, and multi-dimensional index construction on the state transition information set, and store it in the evaluation data warehouse to obtain standardized storage data; S500. Based on the standardized stored data, the evaluation index data is obtained through analysis of the evaluation model group and the multi-level evaluation model. S600. The evaluation index data is compared with the dynamic threshold and the set evaluation threshold, and the shortcoming identification result is obtained by combining the statistical monitoring window. S700. Based on the weakness identification results, match the strategy library and adjust the training strategy, execute the next round of iterative training until the stopping condition is met, and output the optimized algorithm model and evaluation report.

2. The method according to claim 1, characterized in that, The process of standardizing test questions and then performing data fusion, feature extraction, and initial grouping includes: The standardization process includes data format unification, merging of duplicate questions, verification of source identifiers, and completion of training task types. The data fusion refers to linking standardized multidimensional data of the same test question, training task type, and training strategy into the same multidimensional input data. The feature extraction process involves extracting parameter space, constraints, historical running data, real-time running data, dynamic disturbances, maximum iteration count, data acquisition frequency parameters, and network parameters from the multidimensional input data to generate feature parameters. The initial grouping is based on a similarity matrix, grouping test questions with the same training task type, similar parameter space, the same constraints, and the same training strategy into the same initial group, and generating a grouping adjustment log when test questions cross multiple initial groups or when source identifiers conflict.

3. The method according to claim 2, characterized in that, The process of initiating iterative training, synchronously collecting running records, and adding time tags and source identifiers includes: Before starting iterative training, an access verification is performed, which includes group number verification, test question number verification, training strategy reference record verification, and group adjustment log verification. The synchronous acquisition of running records involves receiving running records from the training execution module according to the data acquisition frequency parameter, and arranging the continuously received running records to obtain a data stream to be labeled; Adding time stamps includes writing start time stamps, writing process time stamps, and writing end time stamps; Adding source identifiers involves binding the group number, test question number, training task type, algorithm model number to be evaluated, action network number, value network number, and network parameter version number to each run record.

4. The method according to claim 3, characterized in that, The process of extracting fitness, initial optimal position, state transition information, network parameters, and resource usage to form a state transition information set includes: The state transition information includes the t-th observation state, the t-th action selection, the t-th reward, and the (t+1)-th observation state; First, the original behavioral data stream is sorted sequentially according to the time tag, and then the source is merged according to the source identifier to obtain continuous running records; The t-th observation state, t-th action selection, t-th reward, and t+1-th observation state are extracted from the continuous running records and paired to obtain the t-th iteration features. The fitness, initial optimal position, number of iterations, maximum number of iterations, network parameters, and resource usage are extracted to obtain the training phase features. The t-th iteration feature is associated and bound with the training phase feature.

5. The method according to claim 4, characterized in that, The process of performing format validation, standard data conversion, and multidimensional index construction before storing the data in the evaluation data warehouse includes: The format verification is a unified check of the record format of the t-th observation state, t-th selection action, t-th reward, t+1-th observation state, fitness, initial optimal position, network parameters, and resource usage. The standard data conversion involves converting the aforementioned fields in the verified records into a unified record order and generating a version number. The multidimensional index construction includes dimensions of creation time, source, training task, test question, and network parameters. The evaluation data warehouse includes a standard data storage area, an index structure storage area, a verification result storage area, and a version number storage area.

6. The method according to claim 1, characterized in that, The process of obtaining evaluation indicator data through evaluation model group and multi-level evaluation model analysis includes: The evaluation model set includes a convergence efficiency evaluation model, a robustness evaluation model, a resource utilization evaluation model, and a system stability evaluation model; The multi-level evaluation model analysis includes: firstly, extracting a set of feature vectors from the standardized stored data, and then performing scoring processing through the evaluation model group to obtain convergence efficiency score, robustness evaluation index, resource utilization score, and system stability index; then, performing basic scoring processing, index merging processing, and intermediate evaluation data table generation processing through the multi-level evaluation model to finally obtain evaluation index data.

7. The method according to claim 6, characterized in that, The process of comparing the evaluation index data with dynamic thresholds and set evaluation thresholds, and combining this with a statistical monitoring window to obtain the weakness identification results includes: The dynamic threshold is a threshold comparison rule set according to different training task types, different test question numbers, and different version numbers; The setting of the evaluation threshold comparison includes: first reading the verification result reference record in the evaluation index data, and then sending the verification record into the dynamic threshold mechanism for processing; the dynamic threshold mechanism processing includes threshold field reading, evaluation index data alignment and comparison status generation. The process of obtaining the weakness identification result by combining the statistical monitoring window includes: reading the threshold comparison results under the same training task type, the same test question number, and the same version number in the statistical monitoring window according to the time tag order, and counting the continuous changes; when the comparison state corresponding to the same evaluation model is lower than the corresponding set evaluation threshold in multiple consecutive statistical monitoring windows, it is recorded as the evaluation model corresponding to the weakness; when the comparison state changes at the boundary position of the statistical monitoring window, the delayed effect mechanism is invoked, the boundary position is written to the temporary record, and reading continues in the next statistical monitoring window; The bottleneck identification results also include: group number and feature parameters read from the training task grouping results, and the t-th observation state, t-th action selection, t-th reward, t+1-th observation state, network parameters and resource usage read from the state transition information set.

8. The method according to claim 7, characterized in that, The process of matching the strategy library includes: The strategy library includes convergence efficiency adjustment strategies, robustness adjustment strategies, resource usage adjustment strategies, and system stability adjustment strategies. Based on the evaluation model corresponding to the shortcoming in the shortcoming identification results, the corresponding adjustment strategy is retrieved; when the evaluation model corresponding to the shortcoming is a convergence efficiency evaluation model, a convergence efficiency adjustment strategy is retrieved; when it is a robustness evaluation model, a robustness adjustment strategy is retrieved; when it is a resource usage evaluation model, a resource usage adjustment strategy is retrieved; and when it is a system stability evaluation model, a system stability adjustment strategy is retrieved.

9. The method according to claim 8, characterized in that, The process of adjusting training strategies includes: The candidate adjustment strategies are matched with the training task grouping results, network parameters, and dynamic threshold mechanism to generate training strategy adjustment results. Then, corresponding adjustment actions are executed according to the type of the candidate adjustment strategy: for convergence efficiency adjustment strategies, the network parameters and maximum number of iterations are adjusted; for robustness adjustment strategies, the dynamic perturbation records and test question groupings are adjusted; for resource usage adjustment strategies, the data acquisition frequency parameters and resource scheduling strategy are adjusted; and for system stability adjustment strategies, the dynamic threshold mechanism, statistical monitoring window, and delay-based activation mechanism are adjusted. Finally, an updated training strategy is generated based on the training strategy adjustment results.

10. A multi-dimensional evaluation and dynamic training system for an optimization algorithm model, characterized in that, include: The system comprises a training task processing module, a training execution and data acquisition module, a state transition information processing module, a standardized storage module, an evaluation and analysis module, a bottleneck identification module, a strategy library matching module, and a dynamic training module; the system is used to implement the method described in any one of claims 1-9.

Images