Simulation scenario optimization method and system based on interpretable machine learning algorithm

Abstract

The application discloses a simulation scene optimization method and system based on an interpretable machine learning algorithm, which comprises two parts of factor importance analysis and equivalent test scene identification, realizes analysis of evaluation factor weights from simulation results, intelligently determines factor composition for generating a scene, identifies an equivalent generated scene, and achieves the purpose of simplifying a simulation scene set for evaluation. Through the simplification of the test scene, the application improves the evaluation efficiency of the to-be-tested intelligent model algorithm in the simulation combat scene, enhances the simulation effect, improves the explainability and confidence of the evaluation result, and also improves the speed of the intelligent algorithm iteration optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent simulation, and particularly to a simulation scene optimization method and system based on an explainable machine learning algorithm. BACKGROUND

[0002] In the generation of existing intelligent simulation combat scenes, a large number of experimental factors are usually used to generate a large number of test scenes with high coverage. The intelligent model algorithm to be tested needs to complete simulation tests under all test scenes to obtain a confident intelligent model algorithm evaluation result. However, a large number of simulation test scenes require a great amount of computing resources for testing and evaluation, and also seriously affect the speed of algorithm iteration optimization. In order to solve this problem, the simulation scenes used for evaluation need to be simplified, and a simulation scene set with similar evaluation effects but smaller number is found. SUMMARY

[0003] The present application solves the technical problem of overcoming the deficiencies of the prior art and providing a simulation scene optimization method and system based on an explainable machine learning algorithm. The importance of factors is evaluated by means of the explainable machine learning algorithm, and equivalent test scenes are identified, so that optimization and simplification of test scenes are realized while maintaining the evaluation effect.

[0004] The technical solution of the present application is a simulation scene optimization method based on an explainable machine learning algorithm, which includes factor importance analysis and equivalent test scene identification.

[0005] The steps of the factor importance analysis include:

[0006] S1. Generating a simulation test scene full set S according to given experimental factors and factor levels;

[0007] S2. Simulating and deducing the algorithm to be tested under all simulation test scenes to obtain simulation deduction results corresponding to each simulation test scene, which constitute a simulation deduction result set R of the algorithm to be tested under the simulation test scene full set S;

[0008] S3. Taking <S, R> as a data set to be explained, and based on an explainable machine learning algorithm, calculating the contribution of each experimental factor to the simulation deduction results of the algorithm to be tested;

[0009] S4. Optimizing the simulation test scene full set S according to the contribution of each experimental factor to reduce the number of simulation test scenes;

[0010] The steps of the equivalent test scene identification include:

[0011] P1, obtain the original value range of a given test factor; for each test factor, a plurality of smaller value ranges are randomly generated within the original value range of the test factor, and the union of the smaller value ranges is the original value range; different value range combinations are formed by the smaller value ranges of different test factors to generate different simulation test scenarios;

[0012] P2, under the simulation scenarios generated by different value range combinations, simulation tests are respectively performed on the algorithm to be tested, and simulation deduction results are recorded;

[0013] P3, statistical analysis is performed on the simulation deduction results to screen out equivalent test factor value ranges in each smaller value range of each test factor;

[0014] P4, according to the equivalent test factor value range of each factor, the simulation test scenario is optimized, and the number of simulation test scenarios is reduced.

[0015] At the same time, two ways of factor importance analysis and equivalent test scenario identification are used, or any one way is used, to optimize the simulation scenario and reduce the number of simulation test scenarios.

[0016] Further, the S1 specifically comprises:

[0017] Determine all test factors and factor levels to be analyzed;

[0018] Digitally describe the factor levels of each test factor;

[0019] A value mode of all test factors constitutes a scenario, and each test factor takes values in its factor level range in turn to constitute a simulation test scenario set S.

[0020] Further, in the S3, for the algorithm to be tested based on the tree model, TreeExplainer is used as the SHAP interpreter of the explainable machine learning algorithm, and for other types of algorithms to be tested, KernelExplainer is used as the SHAP interpreter of the explainable machine learning algorithm.

[0021] Further, the S3 specifically comprises:

[0022] Using a random sampling method, a plurality of sub-data sets containing part of the data of the data set to be explained and consistent with the distribution of the data set to be explained are generated for <S, R>;

[0023] For each sub-data set, the SHAP value of each test factor is calculated using the selected SHAP interpreter;

[0024] The SHAP values of the same test factor of each sub-data set are averaged to obtain the global SHAP value of each test factor, that is, the contribution size of each test factor to the simulation deduction result generated by the to-be-tested algorithm.

[0025] Further, in the P2, a Monte Carlo simulation method is used for simulation test.

[0026] The application also provides a simulation scene optimization system based on an interpretable machine learning algorithm, comprising a factor importance analysis module and an equivalent test scene identification module.

[0027] The factor importance analysis module comprises:

[0028] The test scene generation unit I is used for generating a simulation test scene set S according to the given test factors and factor levels;

[0029] The simulation deduction unit is used for performing simulation deduction on the to-be-tested algorithm under all simulation test scenes, obtaining the simulation deduction result corresponding to each simulation test scene, and constituting a simulation deduction result set R of the to-be-tested algorithm under the simulation test scene set S;

[0030] The factor weight calculation unit is used for taking <S, R> as a to-be-interpreted data set, calculating the contribution size of each test factor to the simulation deduction result generated by the to-be-tested algorithm based on the interpretable machine learning algorithm;

[0031] The factor optimization unit is used for optimizing the simulation test scene set S according to the contribution size of each test factor, and reducing the number of simulation test scenes;

[0032] The equivalent test scene identification module comprises:

[0033] The test scene generation unit II is used for obtaining the original value range of the given test factor; for each test factor, a plurality of smaller value ranges are randomly generated in the original value range, and the joint of the smaller value ranges is the original value range; different value range combinations are formed by the smaller value ranges of different test factors, and different simulation test scenes are generated;

[0034] The simulation test unit is used for performing simulation test on the to-be-tested algorithm under the simulation scenes generated by different value range combinations respectively, and recording the simulation deduction results;

[0035] The scene identification unit is used for statistically analyzing the simulation deduction results, and screening out the equivalent test factor value ranges in each smaller value range of each test factor;

[0036] The scene optimization unit is used for optimizing the simulation test scene according to the equivalent test factor value range of each factor, and reducing the number of simulation test scenes.

[0037] At the same time, the simulation scene is optimized by using the factor importance analysis module and the equivalent test scene identification module, or by using any one of the modules.

[0038] Further, in the test scene generation unit I, the specific method for generating the simulation test scene set S according to the given test factors and factor levels is: determining all test factors to be analyzed and factor levels; digitizing the factor levels of each test factor; a value mode of all test factors constitutes a scene, and each test factor takes values in turn within the range of its factor levels to constitute the simulation test scene set S.

[0039] Further, in the factor weight calculation unit, an optional SHAP interpreter is provided, for the algorithm to be tested based on a tree model, the TreeExplainer is used as the SHAP interpreter of the interpretable machine learning algorithm, and for other types of algorithms to be tested, the KernelExplainer is used as the SHAP interpreter of the interpretable machine learning algorithm.

[0040] Further, in the factor weight calculation unit, the specific method for determining the contribution size of each test factor is: using a random sampling method on <S, R> to generate a plurality of sub-data sets containing part of the data of the data set to be explained and consistent with the distribution of the data set to be explained; for each sub-data set, the SHAP value of each test factor is calculated using the selected SHAP interpreter; the SHAP values of each test factor of each sub-data set are averaged to obtain the global SHAP value of each test factor, that is, the contribution size of each test factor to the simulation deduction result of the algorithm to be tested.

[0041] Further, in the simulation test unit, the Monte Carlo simulation method is used for simulation test.

[0042] Compared with the prior art, the present application has the following advantages:

[0043] (1) The present application uses interpretable machine learning algorithms to improve the interpretability and confidence of the evaluation results.

[0044] (2) The present application reduces the number of test factors and optimizes the factor value range, thereby reducing the number of test scenes and improving the efficiency of test evaluation.

[0045] (3) The present application uses fewer equivalent scenes for test evaluation, thereby improving the speed of iterative optimization of intelligent algorithms. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The present application is a method flowchart. DETAILED DESCRIPTION

[0047] In order to better understand the technical solutions of the present application, the embodiments of the present application are specifically described below in combination with the drawings.

[0048] The generation of the intelligent simulation combat scene needs to consider many factors such as weather, terrain, combat power, combat equipment, tactical strategy and the like, each factor has a different value range, resulting in a large number of generated intelligent simulation scenes, a long simulation training time, and affecting the practicability of the intelligent simulation. In order to solve the above problems, the present application provides a simulation scene optimization method based on an explainable machine learning algorithm, including factor importance analysis and equivalent test scene identification, and the specific process can be referred to Figure 1 , the two parts can be executed respectively or in sequence according to the needs of scene simplification.

[0049] 1. The steps of factor importance analysis specifically include:

[0050] S1. Generating a simulation test scene full set S according to the given test factors and factor levels

[0051] A preferred implementation mode is:

[0052] (1) Obtaining test factors: first, all test factors and factor levels to be analyzed need to be determined, these factors are used to generate various different intelligent simulation combat scenes.

[0053] (2) Data preprocessing: in the data preprocessing of the factors, a factor can be regarded as a feature dimension, and then the factor levels of each factor are digitally described, necessary normalization, standardization or other necessary preprocessing operations are performed, a matrix evaluation method is adopted, a value mode of all factors constitutes a scene vector, different factor levels of each factor are taken in turn, and a simulation test scene full set S is obtained, which is used for subsequent analysis and modeling.

[0054] The method for generating the simulation test scene full set S is not limited to the above mode, and other methods can be selected according to actual algorithm testing requirements to generate the simulation test scene full set S.

[0055] S2. Under all simulation test scenes, the simulation deduction of the algorithm to be tested is carried out, the simulation deduction result corresponding to each simulation test scene is obtained, and a simulation deduction result set R of the algorithm to be tested under the simulation test scene full set S is constituted

[0056] S3. Taking <S, R> as a data set to be explained, based on the explainable machine learning algorithm, the contribution size of each test factor to the simulation deduction result of the algorithm to be tested is calculated

[0057] A preferred implementation mode is:

[0058] (1) Select an interpretable model: According to the type of intelligent model algorithm to be tested and the characteristics of the data, select a suitable interpretable machine learning algorithm. For tree-based models (such as decision trees, random forests, etc.), select TreeExplainer as the SHAP interpreter; for any other model (such as complex neural network models), select KernelExplainer as the SHAP interpreter; in addition, for models that only need local interpretation, you can choose a local SHAP interpreter to perform interpretable analysis on the model.

[0059] (2) Sample generation sub-data set: Use random sampling method to generate multiple sub-data sets from the simulation test scene set S. Each sub-data set should contain part of the data of the original data set, and the distribution of the sub-data set should be consistent with the original data set.

[0060] (3) Calculate the SHAP value of the sub-data set: For each sub-data set, use the selected SHAP interpreter to calculate the SHAP value of all factors. These values represent the contribution size of each factor to the model prediction result in this sub-data set.

[0061] (4) Calculate the global SHAP value: Average the SHAP values of all sub-data sets to get the global SHAP value of each factor. These values represent the contribution size of each factor to the model prediction result in the entire data set.

[0062] S4, according to the contribution size of each test factor, optimize the simulation test scene set S, reduce the number of simulation test scenes

[0063] A preferred implementation is:

[0064] (1) Visualize the global SHAP value: Visualize the calculated global SHAP value, such as drawing factor importance bar chart, scatter plot, heat map, etc., to visually show the contribution of each factor to the prediction result.

[0065] (2) Analyze the global SHAP value: By analyzing the visualization results, you can understand which factors have the greatest impact on the prediction result in the entire data set, and how they affect the prediction result.

[0066] (3) Optimize test factors: According to the analysis results, optimize the test factors to improve the prediction performance and interpretability of the intelligent algorithm model. Specifically, you can remove unimportant (contribution less than a certain value) test factors according to the analysis results, thereby reducing the number of simulation test scenes.

[0067] 2、The steps of equivalent test scene identification include:

[0068] P1, obtain the original value range or distribution rule of a given test factor; for each test factor, a plurality of smaller value ranges are randomly generated within the original value range, and the respective smaller value ranges jointly form the original value range; the smaller value ranges of different test factors form different value range combinations, and different simulation test scenarios are generated.

[0069] A preferred implementation is:

[0070] (1) Obtain the initial value range: for each test factor generated for a test scenario, obtain its original value range. A probability model can be established to describe the value range and distribution rule, and the probability model should be able to reflect the uncertainty and variation range of the test factor.

[0071] (2) Randomly generate smaller value ranges: for each test factor, a smaller value range is randomly generated within the original value range, and the respective smaller value ranges still constitute the complete value range; the smaller value ranges of different test factors form different value range combinations, and different simulation scenarios are generated, and the number of simulation scenarios has not been reduced. The generated value range combinations should conform to the data distribution rule of the test factor.

[0072] P2, under different value range combinations generated by simulation scenarios, respectively simulate the test algorithm, and record the simulation deduction results

[0073] A preferred implementation is:

[0074] Monte Carlo simulation test: test and evaluate the tested intelligent algorithm on the simulation scenarios generated by different value range combinations, record the simulation deduction results, and record the changes and results of various parameters in the process.

[0075] Specifically, for the test and evaluation of the missile weapon system intelligent algorithm to be tested, this step specifically includes:

[0076] (1) Select the algorithm to be tested: first select the object to be tested, face different algorithm types such as perception, navigation, decision, and strike, algorithm functions such as target recognition, situation fusion, fire distribution, and Zida strike, select the scene that can test the algorithm and the scene change dimension and level that verify the intelligence of the algorithm. Load the intelligent algorithm into the simulation environment, and configure the initial parameters according to the algorithm requirements to ensure that the algorithm can operate normally.

[0077] (2) Select the simulation scene elements: include scene background, regional dimensions such as sea, land, air, and sky, environment such as weather, atmosphere, rain, snow, and trees, red and blue equipment such as ships, missiles, and vehicles, complete the basic element selection of the simulation scene under specific algorithm types and functions.

[0078] (3) Constructing simulation test question bank: based on the selected scene elements, combined with variable test factor selection and factor range selection, matrix construction of different difficulty simulation test scene question bank of full set.

[0079] (4) Constructing index system: based on the type and function of the algorithm to be tested, combined with the simulation test question bank, the parameters of the algorithm are qualitatively and quantitatively evaluated during the test process, the key parameter changes in the algorithm execution process are recorded, including algorithm performance parameters, function parameters, algorithm instruction output influenced equipment parameters such as missile trajectory, target tracking state, and simulation battlefield process change and result parameters. As a parameter measurement of evaluating scene equivalence.

[0080] (5) Obtain simulation deduction result: through the visual interface, evaluate the influence of different parameters in different scenes on the indicators and other parameters.

[0081] P3, statistical analysis of test evaluation results, screening out equivalent test factor value ranges in each smaller value range of each test factor

[0082] A preferred implementation is:

[0083] Statistical analysis of test results: statistical analysis of test evaluation test results, calculating various statistics such as mean, variance, and standard deviation. By analyzing these statistics, equivalent test factor value ranges can be screened out.

[0084] Specifically, for testing and evaluating the intelligent algorithm of the missile weapon system, this step specifically includes:

[0085] (1) Simulation deduction data collection and arrangement: based on the constructed index system and the record of key parameter changes in the algorithm execution process, including the values of different test factors (such as algorithm parameters, environmental conditions, target characteristics, etc.) and the corresponding functional performance indicators (such as hit rate, response time, resource consumption, etc.), and the selection of each corresponding test factor and factor parameter in each simulation scene, check and clean up errors, missing values or outliers in the data, and ensure the integrity and accuracy of the data.

[0086] (2) Statistical analysis: calculate the mean, median, mode, variance, and standard deviation of each test factor and its corresponding performance indicators to provide basic information about the data distribution, and analyze the correlation between test factors and performance indicators. Correlation coefficients (such as Pearson correlation coefficient, Spearman rank correlation coefficient) can be used for quantification.

[0087] (3) Screening equivalent value range:

[0088] a) Grouping and comparison: Based on the statistical results, the values of the test factors are grouped. Cluster analysis (K-means clustering) or statistical difference-based methods (t-test, ANOVA) can be used to determine which values are statistically in the same group or "equivalent" group.

[0089] b) Set threshold: Based on the simulation scenario test characteristics and the tolerance range of algorithm function requirements, set one or more threshold values for performance indicators, and any test factor value within the threshold range can be considered equivalent.

[0090] c) Determine equivalent range: Based on the grouping results and threshold setting, determine the equivalent value range of each test factor.

[0091] (4) Cross-validation and adjustment: Use the test data set that does not participate in model training to verify the effectiveness of the equivalent value range selected by cross-validation, and further analyze the impact of small changes in test factors within the equivalent value range on simulation process parameters and simulation results.

[0092] P4, according to the equivalent test factor value range of each factor, optimize the simulation test scene, reduce the number of simulation test scenes

[0093] A preferred implementation is:

[0094] Optimize test factor value range: According to the analysis results, the equivalent test factor value range can be optimized. This process can include adjusting the upper and lower limits of the value range, changing the distribution of the value range, etc., so that the equivalent test factor value range is closer to the actual test conditions, thereby improving the accuracy and reliability of the Monte Carlo method.

[0095] After the factor importance analysis and equivalent test scene identification, it can be verified whether the simplified scene is equivalent to the original scene, which includes: obtaining the optimized test factor and its optimized value range, that is, the matrix test factor is simplified from the row and column dimensions. Re-perform simulation test evaluation test of intelligent model algorithm on the simplified scene set, and verify whether the simplified scene can achieve equivalent test evaluation function according to the evaluation results and the change of process parameters.

[0096] It is to be understood that the application is described in one preferred embodiment, and that modifications or equivalents can occur to those skilled in the art without departing from the spirit and scope of the application. Furthermore, it is to be understood that the features and embodiments described herein are capable of modification or equivalent substitution without departing from the spirit and scope of the application. Accordingly, the application is not to be limited by the specific embodiments disclosed herein, but rather by the scope of the claims appended hereto.

[0097] The details of the application are set forth in the accompanying description. Other features, objects and advantages of the application will be apparent to those skilled in the art from the following detailed description, taken in conjunction with the accompanying drawings.

Claims

1. An explainable machine learning algorithm based simulation scenario optimization method, characterized in that: The method comprises two parts of factor importance analysis and equivalent test scenario identification. The steps of the factor importance analysis comprise: S1, generating a simulation test scenario set S according to given test factors and factor levels; S2, performing simulation deduction on the to-be-tested algorithm under all simulation test scenarios to obtain a simulation deduction result corresponding to each simulation test scenario, thereby forming a simulation deduction result set R of the to-be-tested algorithm under the simulation test scenario set S; S3, taking <S, R> as a to-be-interpreted data set, and calculating a contribution size of each test factor to the simulation deduction result of the to-be-tested algorithm based on an interpretable machine learning algorithm; S4, optimizing the simulation test scenario set S according to the contribution size of each test factor, and reducing the number of simulation test scenarios. The steps of the equivalent test scenario identification comprise: P1, obtaining an original value range of a given test factor; for each test factor, a plurality of smaller value ranges are randomly generated in the original value range, and the joint of the smaller value ranges is the original value range; different test factor smaller value ranges form different value range combinations to generate different simulation test scenarios; P2, performing simulation test on the to-be-tested algorithm under the simulation scenarios generated by different value range combinations, and recording simulation deduction results; P3, performing statistical analysis on the simulation deduction results to screen equivalent test factor value ranges in each smaller value range of each test factor; the statistical analysis comprises: calculating statistics of each test factor and a corresponding performance index to provide basic information of data distribution, and analyzing the correlation between test factors and between test factors and performance indexes; the screening mode specifically comprises: a) determining which values belong to an equivalent group based on the statistical analysis results of each test factor and the corresponding performance index, and grouping the values of the test factors; b) setting one or more threshold values of performance index changes based on the simulation scenario test characteristics and the tolerance range of the function requirements of the to-be-tested algorithm, and considering the test factor values changing within the threshold range as equivalent; c) determining the equivalent test factor value range of each test factor based on the grouping results and the threshold setting; P4, optimizing the simulation test scenarios according to the equivalent test factor value range of each test factor, and reducing the number of simulation test scenarios. The simulation scenarios are optimized by using both the factor importance analysis and the equivalent test scenario identification, or by using any one of them, to reduce the number of simulation test scenarios. 2.The simulation scenario optimization method based on an interpretable machine learning algorithm according to claim 1, characterized in that: The S1 specifically comprises: determining all test factors and factor levels to be analyzed; digitally describing the factor levels of each test factor; one value mode of all test factors constitutes a scenario, and each test factor takes values in the range of the factor level to constitute the simulation test scenario set S. 3.The simulation scenario optimization method based on explainable machine learning algorithm according to claim 1, characterized in that: In the S3, for a to-be-tested algorithm based on a tree model, TreeExplainer is used as a SHAP interpreter of the interpretable machine learning algorithm, and for other types of to-be-tested algorithms, KernelExplainer is used as a SHAP interpreter of the interpretable machine learning algorithm. 4.The simulation scenario optimization method based on an interpretable machine learning algorithm according to claim 1, characterized in that: The S3 specifically comprises: using a random sampling method, generating a plurality of sub-datasets containing part of the data of the dataset to be explained and consistent with the distribution of the dataset to be explained; for each sub-dataset, using a selected SHAP interpreter to calculate the SHAP value of each test factor; averaging the SHAP values of the same test factor of each sub-dataset to obtain the global SHAP value of each test factor, i.e. to obtain the contribution size of each test factor to the simulation deduction result generated by the algorithm to be tested. 5.The simulation scenario optimization method based on explainable machine learning algorithm according to claim 1, characterized in that: In the P2, a Monte Carlo simulation method is used for simulation test.

6. An interpretable machine learning algorithm based simulation scenario optimization system, characterized in that: It comprises a factor importance analysis module and an equivalent test scenario identification module. The factor importance analysis module comprises: a test scenario generation unit I, configured to generate a simulation test scenario set S according to the given test factors and factor levels; a simulation deduction unit, configured to perform simulation deduction on the algorithm to be tested under all simulation test scenarios, to obtain the simulation deduction result corresponding to each simulation test scenario, and to form a simulation deduction result set R of the algorithm to be tested under the simulation test scenario set S; a factor weight calculation unit, configured to take <S, R> as the dataset to be explained, and based on the interpretable machine learning algorithm, to calculate the contribution size of each test factor to the simulation deduction result generated by the algorithm to be tested; a factor optimization unit, configured to optimize the simulation test scenario set S according to the contribution size of each test factor, to reduce the number of simulation test scenarios; The equivalent test scenario identification module comprises: a test scenario generation unit II, configured to obtain the original value range of the given test factors; for each test factor, a plurality of smaller value ranges are randomly generated within the original value range, and the joint of each smaller value range is the original value range; different test factor smaller value ranges form different value range combinations to generate different simulation test scenarios; a simulation test unit, configured to perform simulation test on the algorithm to be tested under the simulation scenarios generated by different value range combinations, and record the simulation deduction results; a scenario identification unit, configured to statistically analyze the simulation deduction results, and screen out the equivalent test factor value ranges in each smaller value range of each test factor; the statistical analysis comprises: calculating the statistics of each test factor and its corresponding performance indicators to provide basic information of data distribution, and analyzing the correlation between test factors and between test factors and performance indicators; the screening mode specifically comprises: a) based on the statistical analysis results of each test factor and its corresponding performance indicators, determining which values are statistically equivalent, and grouping the values of the test factors; b) based on the simulation scenario test characteristics and the tolerance range of the function requirements of the algorithm to be tested, setting one or more threshold values of performance indicators, and considering the test factor values changing within the threshold range as equivalent; c) based on the grouping results and threshold setting, determining the equivalent test factor value range of each test factor; a scenario optimization unit, configured to optimize the simulation test scenarios according to the equivalent test factor value range of each test factor, to reduce the number of simulation test scenarios; At the same time, the simulation scene is optimized by using the factor importance analysis module and the equivalent test scene identification module, or by using any one of the modules, so as to reduce the number of simulation test scenes.

7. The interpretable machine learning algorithm-based simulation scenario optimization system of claim 6, wherein: In the test scene generation unit I, the specific method for generating the simulation test scene set S according to the given test factors and factor levels is as follows: all test factors and factor levels to be analyzed are determined; the factor levels of each test factor are digitally described; one value mode of all test factors constitutes a scene, and each test factor is sequentially valued within the range of the factor level, thereby constituting the simulation test scene set S.

8. The interpretable machine learning algorithm based simulation scenario optimization system of claim 6, wherein: In the factor weight calculation unit, an optional SHAP interpreter is provided, for the algorithm to be tested based on a tree model, the TreeExplainer is used as the SHAP interpreter of the interpretable machine learning algorithm, and for other types of algorithms to be tested, the KernelExplainer is used as the SHAP interpreter of the interpretable machine learning algorithm.

9. The interpretable machine learning algorithm based simulation scenario optimization system of claim 6, wherein: In the factor weight calculation unit, the specific method for determining the contribution size of each test factor is as follows: a random sampling method is used on <S, R> to generate a plurality of sub-data sets containing part of the data set to be explained and consistent with the distribution of the data set to be explained; for each sub-data set, the SHAP value of each test factor is calculated by using the selected SHAP interpreter; the SHAP values of each test factor of each sub-data set are averaged to obtain the global SHAP value of each test factor, that is, the contribution size of each test factor to the simulation deduction result of the algorithm to be tested is obtained.

10. The interpretable machine learning algorithm based simulation scenario optimization system of claim 6, wherein: In the simulation test unit, the Monte Carlo simulation method is used for simulation test.

Images