Key test factor determination method and device based on Bayesian causal network

Abstract

The invention provides a key test factor determination method and device based on a Bayesian causal network, and relates to the technical field of reason tracing. The method comprises the following steps: acquiring a multi-dimensional observation parameter of a preset intelligent system, and generating an initial data set based on the multi-dimensional observation parameter; based on an improved Bayesian causal structure learning algorithm and the initial data set, determining a causal structure diagram between an input factor and a target output in the initial data set; based on the causal structure diagram, performing intervention operation on an input factor pointing to the target output to determine an average causal influence degree of the input factor on the target output; and calculating importance degree scores of the input factors according to a preset weight fusion mode and the average causal influence degree, and determining the input factors corresponding to the importance degree scores meeting a preset key factor selection condition as key test factors. In the intelligent system, the real causal-driven effect between the variables is effectively determined, and the accuracy of the key test factor is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cause tracing, for example to a method and device for determining key test factors based on a Bayesian causal network. BACKGROUND

[0002] An intelligent complex system refers to a system composed of multiple intelligent agents or subsystems that interact and depend on each other, and can exhibit overall intelligent behavior through self-organization, self-adaptation, learning or evolution. In an intelligent complex system, in order to determine whether the task execution result of an intelligent agent executing a task meets the target, key evaluation indicators need to be determined. Therefore, a test needs to be designed in advance to determine the key test factors that affect the task execution. Reasonable test design depends on comprehensive identification and effective combination of key test factors in the system, which not only directly determines the range of working conditions that can be covered by the test and the effectiveness of the data, but also affects the accuracy of subsequent model training, performance evaluation and system contribution analysis. However, with the rapid growth of the scale and complexity of artificial intelligence systems, complex equipment systems and multi-agent collaborative tasks, the dimension of system parameters has increased significantly, the dependency structure between variables is complex and opaque, and the variables are strongly coupled, resulting in unprecedented challenges for test design. Traditional test design methods are difficult to optimize the test scheme, and the combination of test schemes lacks scientific basis, making it difficult to effectively cover typical and extreme working conditions.

[0003] In related technologies, there are many factors that affect task execution. Existing feature selection or parameter screening techniques, such as filter-based methods based on statistical correlation and wrapper-based methods based on prediction performance, have improved the automation level of factor screening to some extent, but they cannot fundamentally reveal the causal dependency structure between variables. In a multi-factor strongly coupled and dynamically evolving intelligent complex system, it is difficult to effectively distinguish between the correlation and the real causal driving effect between variables, resulting in inaccurate determination of key test factors. SUMMARY

[0004] The present application aims to provide a method and device for determining key test factors based on a Bayesian causal network.

[0005] According to an aspect of the present application, a method for determining a key test factor based on a Bayesian causal network is provided, comprising: obtaining a plurality of observation parameters of a preset intelligent system, and generating an initial data set for causal inference based on the plurality of observation parameters; wherein the initial data set comprises input factors and target outputs; determining a causal structure diagram between the input factors and the target outputs in the initial data set based on an improved Bayesian causal structure learning algorithm and the initial data set; performing an intervention operation on the input factors directed to the target outputs based on the causal structure diagram to determine an average causal influence degree of the input factors on the target outputs; calculating an importance score of the input factors according to a preset weight fusion manner and the average causal influence degree, and determining the input factors corresponding to the importance score meeting a preset key factor selection condition as the key test factors.

[0006] According to an aspect of the present application, a device for determining a key test factor based on a Bayesian causal network is provided, comprising:

[0007] a data acquisition module configured to obtain a plurality of observation parameters of a preset intelligent system, and generate an initial data set for causal inference based on the plurality of observation parameters; wherein the initial data set comprises input factors and target outputs;

[0008] a causal structure diagram determination module configured to determine a causal structure diagram between the input factors and the target outputs in the initial data set based on an improved Bayesian causal structure learning algorithm and the initial data set;

[0009] an influence degree determination module configured to perform an intervention operation on the input factors directed to the target outputs based on the causal structure diagram to determine an average causal influence degree of the input factors on the target outputs;

[0010] a key test factor determination module configured to calculate an importance score of the input factors according to a preset weight fusion manner and the average causal influence degree, and determine the input factors corresponding to the importance score meeting a preset key factor selection condition as the key test factors.

[0011] According to an aspect of the present application, an electronic device is provided, comprising: a processor; a memory storing a computer program, when the computer program is executed by the processor, the computer program causes the processor to execute the method as described above.

[0012] According to an aspect of the present application, a non-transitory computer readable medium is provided, which stores readable instructions, when the instructions are executed by a processor, the instructions cause the processor to execute the method as described above.

[0013] It should be understood that the above general description and the following detailed description are only exemplary and cannot limit the present application.

[0014] Beneficial effects:

[0015] Through the above-mentioned embodiments provided by the present application, through causal structure modeling and intervention reasoning, the actual influence degree of each input factor on the target output can be quantitatively analyzed under the condition of limited samples, and the key test factors having important driving action on the performance index can be accurately screened, so as to provide an interpretable and quantifiable basis for scientific configuration of test parameters and optimization of test scheme. Compared with the method relying on expert experience or simple correlation analysis, the present application has the advantages of causal interpretability, result quantification and self-adaptive updating, and is helpful to improve the representativeness and coverage of test design, avoid invalid or redundant test settings, significantly improve the utilization efficiency of test resources and the reliability of verification results, and support the performance verification and contribution evaluation of intelligent systems. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained according to these drawings without departing from the scope of the present application.

[0017] Figure 1 The flowchart of the method for determining the key test factor based on the Bayesian causal network provided by the embodiments of the present application;

[0018] Figure 2 The flowchart of the test design optimization based on the Bayesian network provided by the embodiments of the present application;

[0019] Figure 3 The schematic diagram of the dynamic patrol task evaluation process of the unmanned vehicle provided by the embodiments of the present application;

[0020] Figure 4 The causal structure diagram of the intelligent unmanned aerial vehicle group search task provided by the embodiments of the present application;

[0021] Figure 5 The block diagram of the device for determining the key test factor based on the Bayesian causal network provided by the embodiments of the present application;

[0022] Figure 6 The structural schematic diagram of the electronic device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0023] Example embodiments now will be described more fully hereinafter with reference to the accompanying drawings. Example embodiments, however, can be implemented in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example embodiments to those skilled in the art. Like reference numerals refer to like elements throughout the several views and, thus, description of the same elements will not be repeated.

[0024] Moreover, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the application. One skilled in the relevant art will recognize, however, that the

[0025] The block diagrams in the drawings show only the functionality of the embodiments and do not imply that the embodiments will take the form discussed in connection therewith. As illustrated in the various block diagrams throughout the drawings, the illustrated components can be implemented or performed with hardware, software or both. As will be understood by those skilled in the art, the functions of the various illustrated components can be combined or divided into other components, or eliminated, without affecting the overall implementation.

[0026] The flow diagrams depicted herein are examples of sequences of operations that can be performed. Such sequences can be embodied in software or code modules executed by a processing unit, hardware logic or any combination thereof. As will be understood by those skilled in the art, the sequences of operations can be changed, or other sequences of operations can be implemented without departing from the overall method of operation. Accordingly, the sequences of operations described herein and illustrated in the diagrams are exemplary and do not limit the scope of the methods described herein.

[0027] It should be understood that although the terms first, second, third, etc. can be used herein to describe various elements, components, regions, layers and / or sections, these elements, components, regions, layers and / or sections should not be limited by these terms. These terms are only used to distinguish one element, component, region, layer or section from another element, component, region, layer or section. Thus, a first element, component, region, layer or section discussed below could be termed a second element, component, region, layer or section without departing from the teachings of the present application. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0028] The specific implementation can refer to the following embodiments.

[0029] Figure 1 A flowchart of a method for determining a key test factor based on a Bayesian causal network is provided for the embodiments of the present application. The method of the present embodiment can be applied to a test design optimization server. As shown in the figure, the method includes steps S10, S11, S12 and S13. Figure 1

[0030] ​In step S10, a preset multi-dimensional observation parameter of the intelligent system is acquired, and an initial data set for causal inference is generated based on the multi-dimensional observation parameter; wherein the initial data set includes input factors and target outputs.

[0031] In the present application, the intelligent system is used as an alias of the intelligent complex system. Simulation tests are carried out, and multi-dimensional observation data including input parameters, running states, output indicators and the like are collected from the intelligent system. The input parameters can be static parameters, such as the number of unmanned vehicles in the task of "unmanned vehicle cluster performing dynamic patrol task in the region", the cruising speed of the unmanned vehicle and the like set by the user. The output indicators can include, for example, the success rate of task execution and the like. The set of multi-dimensional observation parameters can be used as the initial data set for causal inference. There can be multiple input factors and output factors.

[0032] In step S11, based on the improved Bayesian causal structure learning algorithm and the initial data set, a causal structure graph between the input factors and the target outputs in the initial data set is determined.

[0033] In order to simplify the structure learning process and improve the practicability and computational efficiency of the method, the constraint-based PC algorithm (Peter-Clark algorithm) is used in the present application to learn the causal structure graph, and the improved Bayesian causal structure learning algorithm is obtained. The initial data set is processed by using the algorithm, and the causal structure graph between the input factors and the target outputs is obtained. In the present application, the nodes in the causal structure graph are variables (i.e. input factors and target outputs), and the edges represent the causal direction. The causal structure graph can clearly reflect the direct causal relationship between the variables.

[0034] In step S12, based on the causal structure graph, an intervention operation is performed on the input factors directed to the target outputs, so as to determine the average causal influence degree of the input factors on the target outputs.

[0035] In the present application, the intervention operation can be performed on each input factor directed to each target output. In causal inference and causal structure learning, the intervention operation refers to actively changing the level (which can be a value or a distribution) of a certain variable through external means, so as to observe its influence on other variables in the system, and thus to identify the real causal relationship. Intervention is a core means to distinguish "correlation" and "causality".

[0036] In some implementations, the influence degree of the level change of the input factor on the target output can be quantified by calculating the average causal effect of the input factor. The average causal effect is a measure of the average causal influence strength of a variable on a result variable.

[0037] In step S13, according to the preset weight fusion manner and the average causal influence degree, the importance score of the input factor is calculated, and the input factor corresponding to the importance score meeting the preset key factor selection condition is determined as the key test factor.

[0038] Based on step S12, the causal effect of the input factor on a single target output based on Bayesian causal inference can be obtained. The total importance score of the test factor can be obtained by using the weight fusion method.

[0039] In some implementations, the scores corresponding to different average causal influence degrees can be preset, and then the importance score is calculated.

[0040] In another implementation, since the input factor has both positive and negative driving effects on the target output, the overall importance of the input factor can be represented by taking the absolute value of the average causal effect of the input factor on each target output and then summing the weights, and then the importance score is obtained, and the formula is as follows:

[0041]

[0042] Where n represents the number of target outputs, ACE(X k →Y n ) represents the intervention effect of the input factor X k on the target output Y n , that is, the average causal influence degree, k represents the kth input factor, w n represents the weight of different target outputs, which can be preset by the user.

[0043] According to the requirement of the experiment content, the importance threshold τ is set, and the key factor selection condition is:

[0044] I k ≥τ

[0045] If the I k of a certain input factor meets the above condition, the input factor is taken as the key test factor.

[0046] The application can quantitatively analyze the actual influence degree of each input factor on the target output under the condition of limited samples through causal structure modeling and intervention reasoning, accurately screen the key test factors that have important driving effect on the performance index, and provide interpretable and quantifiable basis for scientific configuration of test parameters and optimization of test scheme. Compared with the method relying on expert experience or simple correlation analysis, the application has the advantages of causal interpretability, result quantification and adaptive updating, which helps to improve the representativeness and coverage of test design, avoid invalid or redundant test settings, significantly improve the utilization efficiency of test resources and the credibility of verification results, and support the performance verification and contribution evaluation of intelligent systems.

[0047] According to some embodiments, in the process of determining the causal structure graph based on the improved Bayesian causal structure virtual seat algorithm, the initial data set can be initialized, independence test and value assignment. Specifically, the initial data set can be initialized based on the Bayesian causal structure learning algorithm to generate a complete undirected graph; the input factors are tested for independence to generate a partial undirected graph based on the complete undirected graph; and the undirected edges of the partial undirected graph are valued using a preset structure recognition value assignment method to generate a causal structure graph.

[0048] Firstly, the application can construct a complete undirected graph, connect variables (i.e. input factors and target outputs) two by two, and assume that there is a potential dependency relationship between them. The initial data set contains multiple test processes of a task, and the input factors and target outputs of different test processes can be different. The PC algorithm is a classic structure learning method based on conditional independence test, which has the advantages of simple implementation, high efficiency and strong result interpretability. For each group of variables, i.e. each test process, the independence of the corresponding input factors is checked using the ability of the PC algorithm itself, the undirected edges in the complete undirected graph are adjusted based on the checking result, and a partial undirected graph is obtained, i.e. there is no relationship between some nodes and there is no undirected edge.

[0049] The preset structure recognition value assignment method can contain two parts: structure recognition and value assignment. The undirected edges of the partial undirected graph are valued using the preset structure recognition value assignment method, the undirected edges are further adjusted to have direction and length, and a causal structure graph is generated.

[0050] The complete undirected graph initialization of the application can avoid missing potential causal paths, ensure that the initial structure covers all possible variable relationships, and reduce the bias caused by insufficient prior knowledge. By statistical test, irrelevant edges are gradually removed to generate a partial undirected graph, effectively distinguishing real causality from false association. The undirected edges are oriented based on the preset rules (structure recognition value assignment method) to ensure that the output causal graph meets the constraints of directed acyclic graph, and the interpretability of the structure is improved.

[0051] According to some embodiments, in the independence test process, for each conditional set of each order, if the input factor is within the preset range of significance level, it is determined that the input factor is conditionally independent; if the input factor is not within the preset range of significance level, it is determined that the input factor is conditionally dependent; and the undirected edge between the input factors corresponding to the conditionally independent case is removed to generate a partial undirected graph based on the complete undirected graph.

[0052] In the present application, for each group of variables, conditional independence tests are performed under different conditional sets of orders, and the present application sets a maximum limit of the order of the conditional set. In some implementations, the maximum limit can be the total number of input factors.

[0053] Under each conditional set of order, it is determined whether the A variable and the B variable of the group are within the range of significance level. If they are within the range, it indicates that the A variable and the B variable are conditionally independent, and the undirected edge between the A variable and the B variable in the complete undirected graph is removed. If they are not within the range, it indicates that the condition is not independent, and the undirected edge is retained. Until the conditional independence test under each conditional set of order is completed, a partial undirected graph is obtained.

[0054] In some implementations, for the conditional independence test under each conditional set of order, the independence test is performed according to the gradually increasing order of the conditional set until the maximum limit.

[0055] The present application strictly distinguishes between conditional independence (retaining edges) and conditional dependence (removing edges) by presetting the range of significance level, avoiding the problem of misdeletion or misretention of edges caused by subjective judgment. The edges are preferentially removed from low order to high order, reducing the calculation amount of high-order test. The independent edges are removed immediately after each order test, avoiding repeated test of the edges that have been determined to be independent, and significantly reducing the calculation overhead in high-dimensional data. Only the statistically significant edges (conditionally dependent) are retained, and the redundant connections are removed, and the generated partial undirected graph is closer to the real causal sparsity.

[0056] According to some embodiments, the structure identification assignment mode includes V-structure identification method and Meek rule. In the process of assigning values to the undirected edges of the partial undirected graph by using the structure identification assignment mode, the direction of the undirected edge can be determined according to the V-structure identification method; the correlation degree of the input factor is determined according to the Meek rule, so as to set the edge length according to the correlation degree; and the partial undirected graph is adjusted according to the direction and the edge length to generate a causal structure graph.

[0057] V-structure Identification (V-structure Identification) is a core step in PC algorithm and other constraint-based causal structure learning processes, which is a method for determining the causal direction relationship between variables in a directed acyclic graph in causal structure learning. Meek's Orientation Rules is a directed edge direction inference rule, which is used to further determine the direction of undirected edges to ensure the directed acyclic nature of the graph after determining the partial causal structure.

[0058] The present application can process the partial undirected graph by V-structure Identification method, determine the direction of undirected edges, use Meek's Orientation Rules to judge the size of the correlation between different nodes, and then adjust the edge length of undirected edges. The partial undirected graph is adjusted according to the direction and edge length, and a complete causal structure graph is obtained.

[0059] The present application identifies collision nodes through V-structure (A→C←B), converts undirected edges into directed edges, and avoids the problem of ambiguous direction in traditional methods. Based on the objective results of conditional independence test, the error of causal inference caused by subjective hypothesis is reduced. The Meek's Orientation Rules are applied to globally check the directionality of edges to ensure that the final causal structure graph meets the constraints of directed acyclic graph.

[0060] According to some embodiments, in the process of determining the average influence degree based on the causal structure graph, a plurality of pairs of level values of the input factor in the causal structure graph can be obtained; for each pair of level values in the plurality of pairs of level values, the output expectation value corresponding to each pair of level values is calculated according to a preset Bayesian network, and the causal effect difference value corresponding to each pair of level values is determined according to the output expectation value; and the causal effect difference values corresponding to the plurality of pairs of level values are averaged to determine the average causal influence degree of the input factor on the target output.

[0061] In the present application, the average causal influence degree is an important indicator in causal inference, which measures the average influence of a variable on a result by calculating the expected change of a factor on a result. In the specific calculation process, the process is as follows:

[0062] Let the value set of the input factor X k be

[0063] For each pair of level values of the single input factor X k , the output expectation value corresponding to each pair of level values is calculated according to a preset Bayesian network, and the causal effect difference value corresponding to each pair of level values is determined according to the output expectation value. Since each task contains multiple experiment processes, each experiment process is not unique for the value (horizontal value) of the input factor, in some implementations, for the current experiment process, the input factor is the number of unmanned vehicles, which can be selected as 3, 5 and 8, so the input factor corresponds to multiple pairs of values, such as (3, 5), (3, 8) and (5, 8).

[0064] The output expected value under different pairs of values of a single input factor is calculated by the Bayesian network, and the formula is as follows:

[0065]

[0066] Where, Y n represents the target output currently concerned, such as the success rate of the task, X k represents the kth input factor being intervened, such as the number of unmanned vehicles. E[Y n |do(X k =x)] represents the expected value of the target output Y k after the intervention of X n to a certain value x (such as the number of unmanned vehicles being 5). do(X k =x) is the “intervention” operation in causal inference, which means that the value of X k is forcibly set to x, which is equivalent to cutting off all incoming edges of X k in the causal graph, and then fixing X k to x, that is, if X k is actively changed, what effect will it have on Y n .

[0067] Z represents the set of all other related factors except X k , which is usually the non-descendant variable of X k . ∑Z represents the sum of all possible combinations of Z values, and P(Z) is the marginal probability distribution of Z in the original (unintervened) Bayesian network. Since the do(X k =x) operation only affects X k and its descendant factors, the probability distribution of the factors in the set Z does not change when X k is intervened. The intervention operation essentially simulates a “random experiment”, which only controls the value of X k and does not affect the intrinsic probability of other variables. E[Y n |X k =x,Z] represents the expected value of Y n in the intervened Bayesian network given X k =x and Z.

[0068] After calculating the output expectation value, the causal effect difference value can be calculated for each set of level values of each input factor to measure the influence degree of the input factor on the target output when the input factor changes, and the calculation formula is as follows:

[0069]

[0070] Then, the average (Avg) of the causal effect difference values corresponding to all pairs of combinations is taken, and then the average causal influence degree of a single input factor X k to the target output Y n is:

[0071]

[0072] The above steps are repeated, and the causal inference operation is performed on all input factors pointing to the target output in the causal structure diagram to obtain the average causal influence degree of all input factors on the output indicator.

[0073] The present application covers different intervention scenarios by obtaining multiple pairs of level values of input factors, avoiding the deviation caused by a single value. Based on the preset Bayesian network, the output expectation value corresponding to each pair of values is calculated to ensure that the causal effect reflects the real intervention effect. The average of the causal effect difference values of each pair of level values weakens the influence of extreme values and improves the stability of the results.

[0074] According to some embodiments, after obtaining the multi-dimensional observation parameters, data processing can be performed to obtain an initial data set. Specifically, the multi-dimensional observation parameters are obtained; and encoding, missing value processing, and discretization processing operations are performed on the multi-dimensional observation parameters according to a preset causal inference requirement to generate the initial data set.

[0075] In the present application, the causal inference requirement can be preset, that is, in order to make the parameters available for causal structure learning, the parameters need to be encoded, missing value processed, and discretized, etc.

[0076] The operation contents uniformly performed are extracted from the causal inference requirement, and then the corresponding operations are performed on the collected multi-dimensional observation parameters to obtain the initial data set.

[0077] The present application standardizes the encoding of heterogeneous data to ensure that the causal algorithm can be directly processed.

[0078] According to some embodiments, after determining the importance score and the key test factor, a test factor combination scheme can be generated. Specifically, according to the average causal influence degree, the importance score, the actual test resource, the working condition requirement, and a preset combination optimization algorithm, a test factor combination scheme containing the key test factor is generated, so that the corresponding test observation data is collected based on the test factor combination scheme, and the mechanism and parameters of the Bayesian causal structure learning algorithm are updated.

[0079] In this application, the combination optimization algorithm can be pre-set to process the average causal impact degree, importance score, actual test resource, and working condition demand to generate an optimized test factor combination, which contains key test factors. The mechanism and parameters of the Bayesian causal structure learning algorithm can be adaptively optimized subsequently.

[0080] On this basis, the flowchart of test design optimization based on Bayesian network can be referred to as shown in Figure 2 For the current test, the data can be collected, the causal diagram can be constructed, the intervention operation can be performed, and the causal importance can be quantified to obtain the importance score according to the above-mentioned flowchart. Then, the efficient test factor combination (test factor combination scheme) can be generated in combination with the actual scene and task demand, the corresponding next test can be performed under the combination, the same flowchart of the current test can be executed, and the Bayesian causal network can be updated based on the obtained result feedback to realize adaptive optimization.

[0081] The application combines the average causal impact degree, importance score, test resource, and working condition demand to generate a test factor combination scheme through a combination optimization algorithm, which can preferentially select a factor combination with high output impact, high engineering importance, and implementability under a limited budget, avoid redundant tests caused by the traditional trial-and-error method, and significantly reduce the data acquisition cost. Based on the newly collected test observation data, the structure and parameters of the Bayesian network are updated in real time to improve the fitting accuracy of the model to the real causal relationship. Through the test design of working condition adaptation, the generalization ability of the model in complex scenes is enhanced.

[0082] According to some embodiments, a specific task instance is provided for the overall flowchart shown in Figure 2 For simplicity of illustration, the key evaluation index (target output) is only "patrol area coverage integrity", which reflects the completeness of the target area covered by the unmanned vehicle within a specified time window. The input factors include: group capacity (X1), single vehicle cruising speed (X2), and perception distance (X3). At this time, the evaluation of the importance of the test factors will be simplified to importance evaluation without multi-index weight fusion. The test collects large-scale data on different combination samples through a simulation platform, constructs a Bayesian causal diagram, calculates the average causal effect of each input factor, and evaluates the importance of the experimental factors, as shown in Figure 3 .

[0083] Step 1: Test data acquisition and preprocessing:

[0084] Based on expert experience, the range of values ​​for swarm size and decision speed were determined. Swarm size X1 ∈ {5, 10, 20, 30, 40}, patrol speed X2 ∈ {10, 20, 30, 40, 50} (km / h), and perception distance X3 ∈ {3m, 10m, 20m, 50m}. One hundred combined experiments were constructed, collecting a total of 2500 samples. All variables underwent preprocessing such as discretization, encoding, and missing value completion to ensure the feasibility of causal structure learning.

[0085] Step 2: Learning the cause-effect graph structure:

[0086] A constraint-based PC algorithm is used to construct a causal graph and learn the causal path between input factors and output variables. The results show that swarm size, cruising speed and sensing distance are all directly causally related to decision accuracy.

[0087] Step 3: Intervention Reasoning and Causal Effect Calculation:

[0088] To quantify the impact of individual input factors, a single-factor intervention was performed for each input factor, i.e., a value of a certain factor was artificially fixed, and the output variable Y was evaluated under the intervention. n The expected change. Taking the swarm size as an example, with a value set of {5, 10, 20, 30, 40}, the expected success rate is calculated for each intervention value:

[0089] E[Y n |do(capacity = x) i )],forx i ∈{5,10,20,30,40}

[0090] The following table shows the values ​​of the swarm size X1 and the expected output after intervention.

[0091] X1 5 10 20 30 40 E(Y) 0.58 0.66 0.75 0.78 0.76

[0092] Next, calculate the average causal effect:

[0093]

[0094] ACE(X) can be calculated 容量 →Y n Similarly, by conducting intervention analysis on the input factor "cruising speed", the average causal effect ACE(X) can be calculated. 速度 →Y n ).

[0095] Step 4: Factor Importance Assessment

[0096] Since the only key assessment indicator is "patrol area coverage integrity," the factor importance is simply the corresponding average causal effect, i.e.:

[0097] I 容量 = 0.074

[0098] Similarly, the patrol speed can be calculated to evaluate the importance of the index I 巡航速度 and the importance of the index I 感知距离 .

[0099] Step 5, factor set optimization and combination scheme generation

[0100] According to the above steps, the average causal effect (ACE) of the unmanned vehicle cluster patrol task input factor set cluster size X1, patrol speed X2 has been obtained, and the causal importance degree relative to the key evaluation index has been calculated. The factor set optimization scheme covers the key input factor combination with larger absolute value of ACE priority, determines the adjustable range of each factor, and uses appropriate optimization algorithm to construct the optimal combination scheme.

[0101] Step 6, test scheme execution and adaptive iterative optimization

[0102] According to the generated optimization combination scheme, the test is executed, new test observation data is collected, and the new data is continuously fed back to update the Bayesian causal network structure and parameters, dynamically correct the causal dependence relationship and factor effect, and iteratively optimize the test factor set and combination scheme, realize the adaptive closed-loop improvement of test design, and improve the scientificity and resource utilization efficiency of test design.

[0103] Another specific task example is provided. Taking the search task of multiple autonomous unmanned vehicles in complex terrain as an example, multiple autonomous unmanned vehicles cooperatively cover the specified area, and exchange position, sensing and decision-making information in real time. The key performance indicators (target outputs) are selected as: (1) task completion time (Y1), that is, the total time required from task start to successful positioning and rescue of the target, the task completion time weight is 0.6, and the resource consumption weight is 0.4; (2) cluster resource consumption (Y2), which refers to the power consumption of the entire unmanned vehicle cluster in the task, directly determines the system sustainability and subsequent task capability. The input factors include: formation size (X1), communication frequency (X2), detection radius (X3), and obstacle density (X4). The test collects large-scale data on different combination samples through the simulation platform, constructs a Bayesian causal graph, calculates the average causal effect of each input factor, and evaluates the importance of the experimental factor.

[0104] Step 1: test data collection and preprocessing:

[0105] Based on the simulation platform, the factor level combination is determined based on expert experience: formation scale X1∈{5, 10, 20, 30}, communication frequency X2∈{1, 5, 10 Hz}, detection radius X3∈{10 m, 20 m, 30 m}, and obstacle density X3∈{5, 10, 20}. Through the design of orthogonal combination and random disturbance experiment, 72 typical simulation experiments are carried out, generating more than 5000 effective samples, covering multiple scenarios and working conditions, recording task completion time and resource consumption. All data are completed, outliers are removed and discretized, to ensure the accuracy and stability of causal structure learning.

[0106] Step 2: Causal structure diagram learning:

[0107] The PC algorithm based on constraints is used to perform conditional independence test on sample data, and the conditional independent edges are gradually removed to determine the directed causal relationship between variables. The result is shown in the following table: Figure 4 As shown in the table, the formation scale X1 directly affects the task completion time Y1 and the resource consumption Y2; the detection radius X3 and the obstacle density X4 have a direct impact on the task completion time Y1; in addition, the obstacle density X4 also has an impact on the resource consumption Y2.

[0108] Step 3: Intervention reasoning and causal effect calculation:

[0109] All input factors are intervened one by one. Taking the formation scale as an example, the expected difference in task completion time is compared between the scales of 5, 10, 20, and 30, respectively, for a total of 6 groups; similarly, the expected difference in resource consumption is calculated between the communication frequencies of 1 Hz, 5 Hz, and 10 Hz, respectively, for a total of 3 groups.

[0110] The complete average causal effect (ACE) matrix is calculated for all factor-index pairs, quantifying the impact of each input factor on different key indicators.

[0111] Step 4: Factor importance score

[0112] The ACE of all factors on the two indicators is weighted and summed according to the expert-determined weights (e.g., task completion time 0.6, resource consumption 0.4). Taking the formation scale X1 as an example:

[0113]

[0114] Similarly, the comprehensive causal importance measure of input factors X2, X3, and X4 on target outputs Y1 and Y2 is calculated, and the experimental factors are sorted according to the results to find the experimental factor set with the highest importance and most significant impact on the experimental indicators.

[0115] Step 5: Factor set optimization and combination scheme generation

[0116] After completing the large-sample simulation test of the multi-agent UAV cluster complex terrain search task and the Bayesian causal structure learning, the causal importance degree of the input factors (formation size X1, communication frequency X2, detection radius X3, and obstacle density X4) to the key performance indicators "task completion time (Y1)" and "cluster resource consumption (Y2)" has been obtained. The factor set optimization scheme preferentially covers the key input factor combination with a larger absolute value of ACE, determines the adjustable range of each factor, and uses a suitable optimization algorithm to construct the optimal combination scheme.

[0117] Step 6, test scheme execution and adaptive iterative optimization

[0118] The test is performed according to the generated optimization combination scheme, new test observation data is collected, and the new data is continuously fed back to update the Bayesian causal network structure and parameters, dynamically correct the causal dependence relationship and factor effect, iteratively optimize the test factor set and combination scheme, realize adaptive closed-loop improvement of the test design, and improve the scientificity and resource utilization efficiency of the test design.

[0119] The device embodiments of the present application are described below, which can be used to execute the method embodiments of the present application. For details not disclosed in the device embodiments of the present application, reference can be made to the method embodiments of the present application.

[0120] Figure 5 A block diagram of a device for determining key test factors based on a Bayesian causal network is provided for the embodiments of the present application. As shown in Figure 5 The device for determining key test factors based on a Bayesian causal network 500 includes a data acquisition module 501, a causal structure diagram determination module 502, an influence degree determination module 503, and a key test factor determination module 504.

[0121] The data acquisition module 501 is configured to acquire a plurality of observation parameters of a preset intelligent system, and generate an initial data set for causal inference based on the plurality of observation parameters; wherein the initial data set includes input factors and target outputs.

[0122] The causal structure diagram determination module 502 is configured to determine a causal structure diagram between the input factors and the target outputs in the initial data set based on an improved Bayesian causal structure learning algorithm and the initial data set.

[0123] The influence degree determination module 503 is configured to perform an intervention operation on the input factors directed to the target outputs based on the causal structure diagram, so as to determine the average causal influence degree of the input factors on the target outputs.

[0124] The key test factor determination module 504 is configured to calculate the importance score of the input factor according to the preset weight fusion manner and the average causal influence degree, and determine the input factor corresponding to the importance score meeting the preset key factor selection condition as the key test factor.

[0125] Optionally, the causal structure diagram determination module 502 is specifically configured to:

[0126] Based on the Bayesian causal structure learning algorithm, the initial data set is initialized to generate a complete undirected graph.

[0127] The input factor is subjected to independence test to generate a partial undirected graph based on the complete undirected graph.

[0128] The preset structure recognition assignment manner is used to assign the undirected edges of the partial undirected graph to generate the causal structure diagram.

[0129] Optionally, the causal structure diagram determination module 502 is specifically configured to:

[0130] For each order condition set, if the input factor is within the preset significance level range, the input factor is determined to be conditionally independent.

[0131] If the input factor is not within the preset significance level range, the input factor is determined to be conditionally independent.

[0132] The undirected edges between the input factors corresponding to the conditionally independent case are removed to generate a partial undirected graph based on the complete undirected graph.

[0133] Optionally, the structure recognition assignment manner includes V-structure recognition method and Meek rule; and the causal structure diagram determination module 502 is specifically configured to:

[0134] According to the V-structure recognition method, the direction of the undirected edge is determined.

[0135] According to the Meek rule, the correlation degree of the input factor is determined to set the edge length according to the correlation degree.

[0136] The partial undirected graph is adjusted according to the direction and the edge length to generate the causal structure diagram.

[0137] Optionally, the influence degree determination module 503 is specifically configured to:

[0138] Obtain multiple pairs of horizontal values of the input factor in the causal structure diagram.

[0139] For each pair of horizontal values, the output expectation value corresponding to each pair of horizontal values is calculated according to the preset Bayesian network, and the causal effect difference value corresponding to each pair of horizontal values is determined according to the output expectation value;

[0140] The causal effect difference values corresponding to the multiple pairs of horizontal values are averaged to determine the average causal influence degree of the input factor on the target output.

[0141] Optionally, the data acquisition module 501 is specifically configured to:

[0142] Obtain the multi-dimensional observation parameters;

[0143] According to the preset causal inference requirement, the multi-dimensional observation parameters are encoded, missing value processed and discretized to generate an initial data set.

[0144] Optionally, the key test factor determination apparatus 500 based on the Bayesian causal network further includes a combination optimization module 505 configured to:

[0145] According to the average causal influence degree, the importance score, the actual test resource, the working condition requirement and the preset combination optimization algorithm, a test factor combination scheme containing the key test factor is generated, so that the corresponding test observation data is collected based on the test factor combination scheme, and the mechanism and the parameters of the Bayesian causal structure learning algorithm are updated.

[0146] The apparatus performs similar functions to the method provided above, and other functions can be referred to the foregoing description, which will not be described here.

[0147] Figure 6 The structural schematic diagram of the electronic device provided in the embodiment of the present application is shown in FIG. 6. Figure 6 As shown in FIG. 6, the electronic device 600 of the embodiment can include a memory 601 and a processor 602.

[0148] The computer program is stored on the memory 601, and when the computer program is executed by the processor 602, the aforementioned processor 602 executes the method in the above embodiment.

[0149] The processor 602 and the memory 601 are connected, such as through a bus.

[0150] Optionally, the electronic device 600 can further include a transceiver. It should be noted that the transceiver in actual application is not limited to one, and the structure of the electronic device 600 does not constitute a limitation on the embodiments of the present application.

[0151] The processor 602 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It can implement or execute the various exemplary logical blocks, modules and circuits described in connection with the disclosure. The processor 602 can also be a combination of computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

[0152] The bus can include a path that carries information between the aforementioned components. The bus can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, or the like. The bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only one thick line is shown in the figure, but it does not mean that there is only one bus or only one type of bus.

[0153] The memory 601 can be a ROM (Read Only Memory) or other type of static storage device that can store static information and instructions, a RAM (Random Access Memory) or other type of dynamic storage device that can store information and instructions, an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto.

[0154] The memory 601 is used to store application program code for implementing the scheme of the present application, and is controlled by the processor 602 for execution. The processor 602 is used to execute the application program code stored in the memory 601 to realize the content shown in the foregoing method embodiments.

[0155] The electronic device includes, but is not limited to, a mobile terminal such as a mobile phone, a notebook computer, a digital broadcast receiver, a PDA (Personal Digital Assistant), a PAD (Tablet Personal Computer), a PMP (Portable Multimedia Player), a car terminal (for example, a car navigation terminal), and the like, and a stationary terminal such as a digital TV, a desktop computer, and the like. It can also be a server or the like. Figure 6 The illustrated electronic device is merely an example and should not impose any limitation on the function and use range of the embodiments of the present application.

[0156] The electronic device of the embodiments can be used to execute the method of any of the above embodiments, and has similar implementation principles and technical effects, which will not be described here.

[0157] The present application also provides a non-transitory computer-readable storage medium having computer-readable instructions stored thereon, and when the instructions are executed by a processor, the processor executes the method in the above embodiments.

[0158] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instructions related to hardware. The foregoing program can be stored in a non-transitory computer-readable storage medium. The program, when executed, performs steps including the above-mentioned method embodiments; and the foregoing storage medium includes ROM, RAM, magnetic or optical disk, and various media that can store program codes.

[0159] The embodiments of the present application are described in detail above, and the specific examples are applied to the principles and implementation modes of the present application. The above description of the embodiments is only used to help understand the method of the present application and its core idea. Meanwhile, the changes or deformations made by the skilled in the art according to the idea of the present application, based on the specific implementation mode and application range of the present application, all belong to the scope of protection of the present application. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for determining a key trial factor based on a Bayesian causal network, characterized by, The method comprises the following steps: acquiring a preset multi-dimensional observation parameter of an intelligent system, and generating an initial data set for causal inference based on the multi-dimensional observation parameter; wherein the initial data set comprises input factors and target outputs; determining a causal structure graph between the input factors and the target outputs in the initial data set based on an improved Bayesian causal structure learning algorithm and the initial data set; performing an intervention operation on the input factors directed to the target outputs based on the causal structure graph to determine an average causal influence degree of the input factors on the target outputs; calculating an importance score of the input factors according to a preset weight fusion manner and the average causal influence degree, and determining the input factors corresponding to the importance score that meets a preset key factor selection condition as key test factors.

2. The method of claim 1, wherein, The method of determining the causal structure graph between the input factors and the target outputs in the initial data set based on the improved Bayesian causal structure learning algorithm and the initial data set comprises: initializing the initial data set based on the Bayesian causal structure learning algorithm to generate a complete undirected graph; performing an independence test on the input factors to generate a partial undirected graph based on the complete undirected graph; assigning values to undirected edges of the partial undirected graph using a preset structure identification assignment manner to generate the causal structure graph.

3. The method of claim 2, wherein, The method of performing an independence test on the input factors to generate a partial undirected graph based on the complete undirected graph comprises: for each order of a conditional set, if the input factors are within a preset significance level range, determining that the input factors are conditionally independent; if the input factors are not within the preset significance level range, determining that the input factors are conditionally dependent; removing undirected edges between the input factors corresponding to the conditionally independent case to generate the partial undirected graph based on the complete undirected graph.

4. The method of claim 2, wherein, The structure identification assignment manner comprises a V-structure identification method and a Meek rule; wherein the method of assigning values to the undirected edges of the partial undirected graph using the preset structure identification assignment manner to generate the causal structure graph comprises: determining a direction of the undirected edges according to the V-structure identification method; determining a correlation degree of the input factors according to the Meek rule to set an edge length according to the correlation degree; adjusting the partial undirected graph according to the direction and the edge length to generate the causal structure graph.

5. The method of claim 1, wherein, The method of performing an intervention operation on the input factors directed to the target outputs based on the causal structure graph to determine an average causal influence degree of the input factors on the target outputs comprises: acquiring multiple pairs of level values of the input factors in the causal structure graph; for each pair of level values in the multiple pairs of level values, calculating an output expectation value corresponding to each pair of level values according to a preset Bayesian network, and determining a causal effect difference value corresponding to each pair of level values according to the output expectation value; The average causal effect difference corresponding to the multiple pairs of horizontal value is averaged to determine the average causal influence degree of the input factor on the target output.

6. The method of claim 1, wherein, The multi-dimensional observation parameters of the preset intelligent system are acquired, and an initial data set for causal inference is generated based on the multi-dimensional observation parameters, including: The multi-dimensional observation parameters are acquired; According to the preset causal inference requirement, the multi-dimensional observation parameters are encoded, missing value processed and discretized to generate the initial data set.

7. The method according to any one of claims 1 to 6, characterized in that, Also includes: According to the average causal influence degree, the importance score, the actual test resource, the working condition requirement and the preset combination optimization algorithm, a test factor combination scheme containing the key test factor is generated, test observation data corresponding to the test factor combination scheme is collected based on the test factor combination scheme, and the mechanism and parameters of the Bayesian causal structure learning algorithm are updated.

8. A device for determining key experimental factors based on Bayesian causal networks, characterized in that, Including: The data acquisition module is used for acquiring the multi-dimensional observation parameters of the preset intelligent system, and generating an initial data set for causal inference based on the multi-dimensional observation parameters; wherein the initial data set includes input factors and target outputs; The causal structure graph determination module is used for determining the causal structure graph between the input factors and the target outputs in the initial data set based on the improved Bayesian causal structure learning algorithm and the initial data set; The influence degree determination module is used for performing intervention operation on the input factor pointing to the target output based on the causal structure graph, to determine the average causal influence degree of the input factor on the target output; The key test factor determination module is used for calculating the importance score of the input factor according to the preset weight fusion mode and the average causal influence degree, and determining the input factor corresponding to the importance score meeting the preset key factor selection condition as the key test factor.

9. An electronic device, comprising: Including: A processor; A memory storing a computer program, when the computer program is executed by the processor, the processor executes the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium, comprising: A computer readable instruction is stored thereon, when the instruction is executed by a processor, the processor executes the method of any one of claims 1-7.

Images