CN115270634B - Counterfactual interpretation generation method and system suitable for autonomous air combat field - Google Patents
Counterfactual interpretation generation method and system suitable for autonomous air combat field Download PDFInfo
- Publication number
- CN115270634B CN115270634B CN202210930992.2A CN202210930992A CN115270634B CN 115270634 B CN115270634 B CN 115270634B CN 202210930992 A CN202210930992 A CN 202210930992A CN 115270634 B CN115270634 B CN 115270634B
- Authority
- CN
- China
- Prior art keywords
- sample
- objective
- optimization
- individuals
- counterfactual
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000005457 optimization Methods 0.000 claims abstract description 145
- 238000012549 training Methods 0.000 claims abstract description 13
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 11
- 230000002068 genetic effect Effects 0.000 claims abstract description 9
- 230000006870 function Effects 0.000 claims description 29
- 230000001364 causal effect Effects 0.000 claims description 26
- 238000010276 construction Methods 0.000 claims description 6
- 238000013136 deep learning model Methods 0.000 claims description 6
- 238000010801 machine learning Methods 0.000 claims description 6
- 230000035772 mutation Effects 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000002474 experimental method Methods 0.000 description 7
- 230000008569 process Effects 0.000 description 6
- 239000004973 liquid crystal related substance Substances 0.000 description 4
- 238000002679 ablation Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 241000282326 Felis catus Species 0.000 description 1
- 235000009421 Myristica fragrans Nutrition 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000010304 firing Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 239000001115 mace Substances 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000004326 stimulated echo acquisition mode for imaging Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
The invention relates to a method and a system for generating counterfactual explanation suitable for the field of autonomous air combat, comprising the following steps: acquiring a simulated air combat data set and carrying out low-dimensional manifold representation, and constructing a multi-objective optimization function of an optimal counter fact sample according to the optimal counter fact property; constructing a black box model to be explained; training a black box model to be explained by using the low-dimensional fact sample and the corresponding decision label; according to the low-dimensional fact sample, the trained black box model to be explained and the multi-objective optimization function, a third-generation non-dominant sorting genetic algorithm is applied to obtain an optimal counter fact sample, and then an optimal counter fact sample explanation text is obtained; and obtaining a decision strategy of the autonomous air combat in the black box model to be explained based on the optimal counterfactual sample interpretation text. The optimal counterfactual sample is generated more accurately based on the low-dimensional fact sample and a more comprehensive multi-objective optimization function, and the optimal counterfactual interpretation text is further obtained.
Description
Technical Field
The invention relates to the field of counterfactual interpretation in the field of autonomous air combat, in particular to a counterfactual interpretation generation method and a counterfactual interpretation generation system suitable for the field of autonomous air combat.
Background
Autonomous Air Combat (AAC) refers to a technique in which an aircraft autonomously performs battlefield sensing, decision making, and control by means of related devices such as an onboard aircraft, to perform an air combat. Autonomous air combat based on reinforcement learning has achieved combat performance beyond that of human pilots, but its black box nature becomes a bottleneck for human-machine interaction and landing applications. By means of the inverse fact research method in the fields of economics and psychology, inverse fact interpretation becomes an important way for revealing the internal mechanism of the black box model and generating high-order semantic interpretation. So-called counterfactual, given the original sample x 0 Generating a black box model f which is as close to x as possible with the black box model f to be interpreted 0 Whereas the inverse fact sample x, from which the predictive label is different, is cf Comparison of x 0 And x cf The key characteristics of the local decision rule of the black box model can be known according to the difference of the local decision rule.At present, a local post-interpretation method based on a counterfactual sample has become an important way for revealing the internal mechanism of the black box decision model, but the existing method for generating the counterfactual sample and the counterfactual interpretation has the following disadvantages: (1) The model architecture is not uniform and the modeling method for generating the counterfactual is different, and is difficult to be commonly known, for example, part of methods directly disturb in the original data feature space, while other methods emphasize that manifold representations of data need to be learned first to improve causal feasibility and interpretability of generated samples, and modeling of the counterfactual generation problem is various, and can be regarded as searching shortest path problem, multi-objective optimization problem or Markov decision process and the like; (2) Modeling optimal counterfactual properties is incomplete, and most existing methods only involve partial optimal counterfactual sample properties; (3) What the user needs is a natural language interpretation of high-level semantics, the counterfactual sample is just an intermediate result, but existing methods mostly ignore how to span the semantic gap between the counterfactual sample and the text interpretation. In this regard, the invention provides a method and a system for generating counterfactual explanation suitable for the field of autonomous air combat.
Disclosure of Invention
The invention aims to provide a method and a system for generating a counterfactual interpretation, which are suitable for the field of autonomous air combat, wherein after low-dimensional manifold representation is carried out on air combat data, a multi-objective optimization function is constructed by considering more comprehensive optimal counterfactual sample properties, the accuracy of generating the optimal counterfactual sample is improved, and a counterfactual interpretation text is generated based on the obtained counterfactual sample.
In order to achieve the above object, the present invention provides the following solutions:
a method for generating counterfactual explanation suitable for the field of autonomous air combat comprises the following steps:
acquiring a simulated air combat data set and carrying out low-dimensional manifold representation on the simulated air combat data set to obtain low-dimensional air combat data; the low-dimensional air combat data comprise a plurality of low-dimensional fact samples and decision labels corresponding to each low-dimensional fact sample;
constructing a multi-objective optimization function of the optimal counter fact sample according to the optimal counter fact property;
constructing a black box model to be explained; the black box model to be explained is a machine learning model or a deep learning model;
training the black box model to be explained by using the low-dimensional fact sample and the corresponding decision tag to obtain a trained black box model to be explained;
Applying a third-generation non-dominant sorting genetic algorithm according to the low-dimensional fact sample, the trained black box model to be explained and the multi-objective optimization function of the optimal counter fact sample to obtain the optimal counter fact sample;
according to the optimal counterfactual sample, the characteristic value of the low-dimensional fact sample and a preset interpretation text set, interpreting the optimal counterfactual sample to obtain an optimal counterfactual sample interpretation text;
and obtaining a decision strategy of the autonomous air combat in the black box model to be explained based on the optimal counterfactual sample interpretation text.
The invention also provides a counterfactual interpretation generation system suitable for the field of autonomous air combat, which comprises the following steps:
the low-dimensional air combat data acquisition module is used for acquiring a simulated air combat data set and carrying out low-dimensional manifold representation on the simulated air combat data set to obtain low-dimensional air combat data; the low-dimensional air combat data comprise a plurality of low-dimensional fact samples and decision labels corresponding to each low-dimensional fact sample;
the optimization target determining module is used for constructing a multi-target optimization function of the optimal counter fact sample according to the optimal counter fact property;
the model construction module is used for constructing a black box model to be explained; the black box model to be explained is a machine learning model or a deep learning model;
The training module is used for training the black box model to be explained by utilizing the low-dimensional fact sample and the corresponding decision label to obtain a trained black box model to be explained;
the optimal counter fact sample acquisition module is used for applying a third generation non-dominant sorting genetic algorithm according to the low-dimensional fact sample, the trained black box model to be interpreted and the multi-objective optimization function of the optimal counter fact sample to obtain the optimal counter fact sample;
the interpretation text generation module is used for interpreting the optimal counterfacts sample according to the characteristic values of the optimal counterfacts sample and the low-dimensional fact sample and a preset interpretation text set to obtain an optimal counterfacts sample interpretation text;
and the decision strategy acquisition module is used for acquiring the decision strategy of the autonomous air combat in the black box model to be interpreted based on the optimal counter fact sample interpretation text.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention relates to a method and a system for generating counterfactual explanation suitable for the field of autonomous air combat, comprising the following steps: acquiring a simulated air combat data set and carrying out low-dimensional manifold representation, and constructing a multi-objective optimization function of an optimal counter fact sample according to the optimal counter fact property; constructing a black box model to be explained; training a black box model to be explained by using the low-dimensional fact sample and the corresponding decision label; according to the low-dimensional fact sample, the trained black box model to be explained and the multi-objective optimization function, a third-generation non-dominant sorting genetic algorithm is applied to obtain an optimal counter fact sample, and then an optimal counter fact sample explanation text is obtained; and obtaining a decision strategy of the autonomous air combat in the black box model to be explained based on the optimal counterfactual sample interpretation text. The optimal counterfactual sample is generated more accurately based on the low-dimensional fact sample and a more comprehensive multi-objective optimization function, and the optimal counterfactual interpretation text is further obtained.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a flowchart of a method for generating a counterfactual explanation applicable to the field of autonomous air combat provided in embodiment 1 of the present invention;
fig. 2 is a frame diagram of a method for generating a counterfactual explanation for the field of autonomous air combat provided in embodiment 1 of the present invention;
FIG. 3 shows the AACE performance of the DCS-AtoA data set according to example 1 of the present invention in different embedding dimensions;
FIG. 4 shows the AACE performance of the DCS-AtoG dataset according to example 1 of the present invention in different embedding dimensions;
fig. 5 is a block diagram of a system and a method for generating a counterfactual explanation suitable for the field of autonomous air combat according to embodiment 2 of the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The invention aims to provide a method and a system for generating a counterfactual interpretation, which are suitable for the field of autonomous air combat, wherein after low-dimensional manifold representation is carried out on air combat data, a multi-objective optimization function is constructed by considering more comprehensive optimal counterfactual sample properties, the accuracy of generating the optimal counterfactual sample is improved, and a counterfactual interpretation text is generated based on the obtained counterfactual sample.
In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description.
Example 1
As shown in fig. 1 and 2, the present embodiment provides a method for generating a counterfactual interpretation applicable to the field of autonomous air combat, including:
s1: acquiring a simulated air combat data set and carrying out low-dimensional manifold representation on the simulated air combat data set to obtain low-dimensional air combat data; the low-dimensional air combat data comprises a plurality of low-dimensional fact samples and decision labels corresponding to each low-dimensional fact sample.
Two air combat datasets, DCS-AtoA and DCS-AtoG, were used to validate and study the methods presented herein. DCS-AtoA is data for simulating an air-to-air dog bucket scene, decision labels are two categories (0-not fired, 1-fired), DCS-AtoG is data for simulating an air-to-ground attack scene, and decision labels are five categories (0-motionless, 1-search, 2-aiming, 3-attack, 4-break away). The two data sets are table data and are acquired based on a DCS-World air combat simulator sold by a STEAM platform.
A low-dimensional manifold representation of the data is learned using dfencoder. dfencoder is a self-encoder suitable for table data, and the core idea is to train continuous features, multi-class features and binary features into three potential variables respectively, and remove data points with larger construction errors so as to reduce abnormal deviation:
wherein L is an encoder, L -1 For decoder, X con ，X cat ,X bin Respectively continuous characteristics, multi-category characteristics and binary characteristics. argmin dist () represents a low-dimensional manifold representation representing a minimized reconstruction error, which is a loss function training the self-encoder; step S1 may learn a low-dimensional data manifold representation of the original simulated air combat data.
S2: and constructing a multi-objective optimization function of the optimal counter fact sample according to the optimal counter fact property.
The optimal counterfactual needs to have six properties of legitimacy, proximity, sparsity, diversity, causal feasibility, manifold proximity. Firstly, step S2 performs optimization in the low-dimensional manifold representation learned in step S1 to generate a counterfactual sample, which is because the proximity requires that the change of the corresponding manifold point of the counterfactual sample in the counterfactual optimizing process is as small as possible, and the small change is ignored when the counterfactual sample is generated by a decoder, so that the counterfactual sample is as similar as possible to the counterfactual sample, manifold proximity is ensured, and no occurrence occurs Outliers. This is also the value of the step S1 self-encoder. Subsequently given the label y 0 Original sample x of (2) 0 Decision strategy pi to be interpreted θ Generated anti-facts sample x cf And the expected counterfactual label y cf The counterfactual properties other than manifold proximity can be modeled as an optimization objective as follows. Specifically, step S2 includes:
s21: and determining a first optimization sub-target according to the difference between the decision label of the inverse facts sample and the decision label of the low-dimensional facts sample.
Legitimacy: legitimacy ensures that the generated anti-facts sample tags are different from the original sample tags to satisfy the anti-facts properties. Due to y 0 The decision tag predicted by the model is usually a probability, which is compared with a specific threshold value to be converted into a discrete fire control tag, so that the loss of legitimacy is modeled by using cross entropy loss. Specifically, the expression of the first optimization sub-objective is:
loss validity ＝-y cf log(f(x cf ))+(1-y cf )log(1-f(x cf ))
wherein loss is validity Representing a first optimization sub-objective; x is x cf Representing a counterfactual sample; y is cf Decision labels representing counterfactual samples; f (x) cf ) And the prediction label of the black box model to be interpreted on the counterfactual sample is represented.
S22: a second optimization sub-objective is determined based on similarities between features of the inverse facts sample and features of the low-dimensional facts sample.
Proximity: proximity requires that the generated anti-facts sample be as close as possible to the facts sample on the basis of satisfying legitimacy to satisfy the requirement of "opposite closest".
For the continuity features, the average absolute deviation is used for normalization due to the different dimensions of the different features.
Wherein Conloss is proximity A second optimization sub-objective representing a continuous feature; m represents the number of consecutive features in the counterfactual sample,representing the ith successive feature of the counterfactual sample; x is x i Representing the ith successive feature of the fact sample; mac i Representing the average absolute deviation of the ith successive feature of the fact sample.
For discrete features, proximity only needs to compare whether the feature values are the same.
Catloss proximity A second optimization sub-objective representing a discrete feature; if it isThen I (·|·) =o, otherwise I (·|·) =1; n is the number of discrete features;
the proximity optimization objective can be modeled as a whole:
loss proximity ＝Conloss proximity +Catloss proximity
wherein loss is proximity Representing a second optimization sub-objective.
S23: and determining a third optimization sub-objective according to the changed quantity of the characteristics in the counterfactual sample.
Sparsity: sparsity means that fewer features change in the generated counterfactual sample, thereby ensuring readability of the interpretation. Experiments show that the continuous type characteristic is necessarily changed in the process of generating the counterfactual sample, so that the invention only measures the changed quantity of the discrete type characteristic and is used as an optimization target of sparsity.
The expression of the third optimization sub-objective is:
wherein loss is sparsity Representing a third optimization sub-objective; if it isThen I (·|·) =0, otherwise I (·|·) =1.
S24: determining a fourth optimization sub-objective based on differences between a plurality of said anti-facts samples generated from one of said low-dimensional facts samples;
diversity of: diversity is a measure of the difference between multiple counterfactual samples generated from one fact sample. The higher diversity means that more viable choices and more information interpretation are provided to the user. A straightforward method to calculate the diversity loss is to sum the distances of each counter pair, but in order to keep the sparsity low at the same time, the invention uses a deterministic point process (determinant point processes, DPP) to implement the diversity constraint, calculating the determinant of the distance matrix.
The expression of the fourth optimization sub-objective is:
wherein loss is diversity Representing a fourth optimization sub-objective; dist (x) cfa ，x cfb ) The distance between the two counterfactual samples is calculated.
S25: a fifth optimization sub-objective is determined based on causal constraints between features of the counterfactual sample.
Causal feasibility: causal feasibility preserves causal constraints between features. For example, the education level does not decrease with age. This causality is described by a Structural Causal Model (SCM). The structured causal model is a directed acyclic graph whose nodes represent features and edges represent causal relationships from cause to result. One node e in the SCM may be determined by its parent: e=g (pa (e), e). Where pa (e) represents the parent of e, g is a function representing causal strength, and e is Gaussian noise. The method first calculates a counterfactual sample with causal feasibility:
Wherein, the liquid crystal display device comprises a liquid crystal display device,a first exogenous variable representing a counterfactual sample; the exogenous variable is a characteristic different from the continuous and discrete characteristics in the counterfactual sample; g (-) refers to a structural causal model, and after the structural causal model G (-) and exogenous characteristic variables are given, all endogenous characteristic variables are calculated in a breadth-first traversal mode, so that a counterfactual sample with causal feasibility is obtained.
Causal feasibility is defined asAnd x cf Distance between:
namely, the expression of the fifth optimization sub-objective is:
wherein loss is causality Representing a fifth optimization sub-objective; the method comprises the steps of carrying out a first treatment on the surface of the I 2 Representing a binary norm.
S26: taking the counterfactual sample with a specified counterfactual label in the low-dimensional fact sample as a sixth optimization sub-objective; the anti-facts label refers to decision labels corresponding to the anti-facts samples.
Prototype loss: to accelerate the counterfactual generation, we have added the optimization objective of prototype loss in addition to the optimal counterfactual properties. A prototype is a representation of a sample in a fact sample that has a specified anti-fact label. Prototype proto can be calculated by the following formula:
wherein, the liquid crystal display device comprises a liquid crystal display device,for an exponential kernel defined on the distance measure D, z is a hidden spatial representation of the low-dimensional fact sample; / >Is the hidden spatial representation of the K-nearest neighbor sample with the anti-fact label of the p-th low-dimensional fact sample; k represents the number of K neighbor samples; the corner label knn represents the K-nearest neighbor algorithm.
The prototype loss, i.e. the expression of the sixth optimization sub-objective, is:
loss proto ＝||proto-z cf || 2
wherein loss is proto Representing a sixth optimization sub-objective; z cf A hidden spatial representation of the low-dimensional counterfactual sample is represented.
S27: and constructing the multi-objective optimization function according to the first optimization sub-objective, the second optimization sub-objective, the third optimization sub-objective, the fourth optimization sub-objective, the fifth optimization sub-objective and the sixth optimization sub-objective.
The final objective of the counterfactual optimization is the following multi-objective optimization task:
s3: constructing a black box model to be explained; the black box model to be explained is a machine learning model or a deep learning model.
S4: and training the black box model to be explained by using the low-dimensional fact sample and the corresponding decision label to obtain a trained black box model to be explained.
The LightGBM is used for fitting the air combat data set, the LightGBM is used for fitting the original feature space, and a 5-fold cross validation method is adopted, so that the model evaluation index AUC is ensured to reach more than 0.95.
S5: and applying a third generation non-dominant sorting genetic algorithm NSGA-III according to the low-dimensional fact sample, the trained black box model to be explained and the multi-objective optimization function of the optimal counter fact sample to obtain the optimal counter fact sample.
The step S5 specifically includes:
s51: and randomly generating an initial population obeying Gaussian distribution based on a plurality of low-dimensional fact samples.
S52: crossing and mutating individuals in the initial population to generate a middle population.
S53: and acquiring the set of the initial population and the intermediate population to obtain a set population.
S54: inputting each individual in the aggregate population into the trained black box model to be explained, obtaining a decision label corresponding to each individual, and calculating a plurality of optimization target values by combining the multi-objective optimization function.
In step S54, the individuals (counterfactual samples) in the aggregate population are input into the trained black box model to be explained to obtain the decision labels corresponding to the counterfactual samples, so as to participate in the calculation of the multi-objective optimization.
S55: and performing non-dominant sorting on the aggregate population according to the optimized target values to obtain layered pareto fronts.
S56: and obtaining individuals with the same number as the individuals of the initial population from the layered pareto front edge to obtain a next generation population.
Considering that individual layered individuals of the pareto front do not necessarily just constitute the number of individuals required for the next generation population, i.e. the number of individuals of the initial population, an attempt is made to determine whether the number of individuals in the pareto front of the first h layers is equal to the number of individuals of the initial population.
Specifically, step S56 specifically includes:
judging whether the number of individuals in the pareto front edge of the first h layers is equal to the number of individuals of the initial population;
if the first-generation population is equal to the second-generation population, taking individuals in the pareto front of the first h layers as individuals in the next-generation population;
if the number of individuals in the pareto front of the first h layer is smaller than the number of individuals in the initial population, let h=h+1, and return to the step of judging whether the number of individuals in the pareto front of the first h layer is equal to the number of individuals in the initial population;
if the number of the individuals in the pareto front of the first h-1 layer is larger than the number of the individuals in the next generation population, taking the individuals in the pareto front of the first h-1 layer as part of the individuals in the next generation population, and calculating the number of the remaining individuals;
performing scalar processing on all the optimization target values corresponding to all the individuals in the aggregate population, and generating a plurality of reference points by using a simplex method according to scalar results;
calculating the occurrence times of the reference points in the niche;
selecting the reference point with the occurrence frequency smaller than the preset frequency as a target reference point;
and selecting the individuals with the number of the residual individuals from the pareto front edge of the h layer according to the similarity between the individuals and the target reference point, and adding the individuals into the next generation population to obtain a complete next generation population.
S57: and judging whether the current iteration number is equal to the maximum iteration number.
If yes, the next generation population is the optimal inverse fact sample set;
if not, the next generation population is made to be the initial population, and the step S52 is returned until the maximum iteration number is reached.
S6: and interpreting the optimal counter fact sample according to the characteristic values of the optimal counter fact sample and the low-dimensional fact sample and a preset interpretation text set to obtain an optimal counter fact sample interpretation text.
The preset interpretation text set is manually formulated according to field knowledge, such as a text describing a classical situation in a close range dog fight scene. The triggering condition is that one or several characteristic values of the fact sample and the anti-fact sample meet a certain condition, and within a certain range, specific characteristic changes in the fact sample and the anti-fact sample may trigger generation of corresponding interpretation text.
The high-dimensional space optimal inverse facts sample may also be interpreted in step S6 based on the optimal inverse facts sample of the high-dimensional space and the eigenvalues of the facts samples of the high-dimensional space.
S7: and obtaining a decision strategy of the autonomous air combat in the black box model to be explained based on the optimal counterfactual sample interpretation text.
After the optimal counter fact sample is obtained, internal rules in the black box model can be found, and then the decision strategy of autonomous air combat learned by the black box model can be known.
In this embodiment, the quality of the generated counterfactual sample obtains better performance in all six indexes, and the method is referred to as AACE. Experiments are carried out on DCS-AtoA and DCS-AtoG generated based on DCS-world simulation, and the quality of the counterfactual sample is measured by adopting the following indexes:
(1) Legal ratio (valRatio): i.e. the ratio of legal counterfactual samples to total counterfactual samples.
(2) Proximity (meanProx): i.e. the average of the distances of the generated counterfactual samples from the actual samples, is defined as the same as the proximity optimization objective.
(3) Diversity (diversity): i.e. a set of counterfactual samples generated for one fact sample, the average of the distances between two pairs.
(4) Causal feasibility ratio (causallratio): i.e. the ratio of the samples satisfying the causal constraint in the generated counterfactual samples to the total counterfactual sample
(5) IM1: the manifold closeness of the counterfactual sample is measured and defined as follows:
wherein AE cf AE for self-encoder (dimension reduction) training on a counterfactual class (not counterfactual samples) ori For a self-encoder trained on a fact class, e is a non-zero small amount.
(6) IM2: the interpretability of the counterfactual sample is measured and defined as follows:
where (5) and (6) are abbreviations of interpretable metric, two interpretability criteria, used to measure manifold proximity of the counterfactual sample. The foregoing properties lack manifold proximity and have been incorporated. The formula applies in the process of evaluating the counterfactual samples, see table 1, table 2.
Wherein AE is cf AE for self-encoder trained on counterfactual class (not counterfactual samples) full For a self-encoder trained on all classes (i.e., the entire data set), e is a non-zero small amount.
Among the six indexes, the larger the legal rate, the diversity and the causal feasibility are, the better; the smaller the proximity, the better the IM1, IM 2.
The effect of partial counterfactual generation method of nearly three years on DCS-AtoA and DCS-AtoG is compared with AACE as baseline, and the result is shown in the following table:
TABLE 1 representation of counterfactual sample generation on DCS-AtoA dataset
TABLE 2 representation of counterfactual sample generation on DCS-AtoG dataset
Analysis of the above table can result in:
(1) All methods can generate 100% legal counterfactual samples. AACE exceeded all other baselines in proximity, causal feasibility, IM1 and IM2 metrics, reaching the second highest performance in diversity.
(2) By making a heavier penalty on the infeasible counter facts, the causal feasibility of AACE is better than procae and MACE, and the latter two also add causal constraints. Indicating the rationality of considering error propagation in a structured causal model.
To explore the contribution factors to low IM1 and IM2 in AACE, further ablation experiments were performed on the method of this example, with the self-encoder and prototype losses removed, respectively, and the results are shown in tables 3 and 4 below:
TABLE 3 ablation experiments of AACE on DCS-AtoA
TABLE 4 ablation experiments of AACE on DCS-AtoA
The study found that there was a slight increase in IM1 and IM2 on both datasets without prototype guidance. However, whether or not to optimize on a dense data manifold representation has a greater impact on both metrics. Based on the above comparison, it is strongly recommended to add a self-encoder or a variant thereof as an optimization device for the counterfactual generation.
To verify the robustness of the self-encoder, the performance of AACE at different embedding dimensions was explored, see fig. 3 and 4.
When the embedding size is from 4 to 7, other indexes than IM2 on DCS-AtoA are stable. For the AtoG dataset, all metrics were stable at different embedding scales. The above experiments demonstrate the overall stability of the self-encoder used by AACE.
Finally, taking a non-firing label of the air-to-air task DCS-AtoA as an example to generate an explanation text. According to expert knowledge, the air-to-air dog buckets most easily occur within a distance of 500-1000 meters, which is an effective attack range for most aircraft cannons. In addition, if the pilot can enter the range of the tail of the enemy aircraft, better attack gestures can be archived for the enemy aircraft, and the departure angle of the aircraft at the moment is generally within 30 DEG . Accordingly, an interpretation set is developedThe trigger II is (AOA and dist are features in DCS-AtoA and respectively represent the departure angle and the distance of a friend or foe) that:
for the following example, the explanation of the counterfactual is: if the distance is within 1000 meters, the aircraft will fire.
TABLE 5 fact-counterfactual sample pair for generating interpretations in DCS-AtoA task
The quality of the counterfactual generation and the validity of the text interpretation generation are fully verified step by step through a plurality of experiments. From the experimental results, the solution proposed by the invention is novel, reliable and effective.
Example 2
As shown in fig. 5, the present embodiment provides a counterfactual interpretation generating system applicable to the field of autonomous air combat, including:
the low-dimensional air combat data acquisition module T1 is used for acquiring a simulated air combat data set and carrying out low-dimensional manifold representation on the simulated air combat data set to obtain low-dimensional air combat data; the low-dimensional air combat data comprise a plurality of low-dimensional fact samples and decision labels corresponding to each low-dimensional fact sample;
the optimization target determining module T2 is used for constructing a multi-target optimization function of the optimal counter fact sample according to the optimal counter fact property;
The optimization target determining module T2 specifically includes:
a first optimization target determining sub-module, configured to determine a first optimization sub-target according to a difference between the decision tag of the inverse facts sample and the decision tag of the low-dimensional facts sample;
a second optimization objective determination submodule for determining a second optimization objective according to the similarity between the features of the inverse fact sample and the features of the low-dimensional fact sample;
a third optimization objective determination submodule for determining a third optimization objective according to the number of features changed in the counterfactual sample;
a fourth optimization objective determination sub-module for determining a fourth optimization sub-objective based on differences between a plurality of said counterfactual samples generated from one of said low-dimensional fact samples;
a fifth optimization objective determination submodule for determining a fifth optimization sub-objective according to causal constraints between features of the anti-facts sample;
a first optimization objective determination sub-module configured to take the counterfactual sample with a specified counterfactual tag in the low-dimensional fact sample as a sixth optimization sub-objective; the anti-facts label refers to decision labels corresponding to the anti-facts samples;
and the optimization target determination sub-module is used for constructing the multi-target optimization function according to the first optimization sub-target, the second optimization sub-target, the third optimization sub-target, the fourth optimization sub-target, the fifth optimization sub-target and the sixth optimization sub-target.
The model construction module T3 is used for constructing a black box model to be explained; the black box model to be explained is a machine learning model or a deep learning model;
the training module T4 is used for training the black box model to be explained by using the low-dimensional fact sample and the corresponding decision tag to obtain a trained black box model to be explained;
the optimal counter fact sample acquisition module T5 is used for applying a third-generation non-dominant sorting genetic algorithm according to the low-dimensional fact sample, the trained black box model to be explained and the multi-objective optimization function of the optimal counter fact sample to obtain the optimal counter fact sample;
the optimal counterfactual sample acquiring module T5 specifically includes:
the initial population construction submodule is used for randomly generating an initial population obeying Gaussian distribution based on a plurality of low-dimensional fact samples;
the cross mutation sub-module is used for carrying out cross and mutation on individuals in the initial population to generate a middle population;
the population merging sub-module is used for acquiring the set of the initial population and the intermediate population to obtain a set population;
the optimization target value calculation sub-module is used for inputting each individual in the aggregate population into the trained black box model to be explained, obtaining a decision tag corresponding to each individual and calculating a plurality of optimization target values by combining the multi-target optimization function;
The non-dominant ranking sub-module is used for performing non-dominant ranking on the aggregate population according to a plurality of optimized target values to obtain layered pareto fronts;
the next generation population generation sub-module is used for acquiring individuals with the same number as the individuals of the initial population from the layered pareto front edge to obtain a next generation population;
the next generation population generation submodule specifically comprises:
the judging unit is used for judging whether the number of individuals in the pareto front edge of the first h layer is equal to the number of individuals of the initial population;
if the first-generation population is equal to the second-generation population, taking individuals in the pareto front of the first h layers as individuals in the next-generation population;
if the number of individuals in the pareto front of the first h layer is smaller than the number of individuals in the initial population, let h=h+1, and return to the step of judging whether the number of individuals in the pareto front of the first h layer is equal to the number of individuals in the initial population;
if the number of the individuals in the pareto front of the first h-1 layer is larger than the number of the individuals in the next generation population, taking the individuals in the pareto front of the first h-1 layer as part of the individuals in the next generation population, and calculating the number of the remaining individuals;
the reference point generation unit is used for performing scalar processing on all the optimization target values corresponding to all the individuals in the aggregate population and generating a plurality of reference points by using a simplex method according to scalar results;
The target reference point determining unit is used for calculating the occurrence times of the reference points in the niche and selecting the reference points with the occurrence times smaller than the preset times as target reference points;
the next generation population generation unit is used for selecting the individuals with the residual individual numbers from the pareto front edge of the h layer according to the similarity between the individuals and the target reference point, and adding the individuals into the next generation population to obtain a complete next generation population.
The judging sub-module is used for judging whether the current iteration number is equal to the maximum iteration number;
if yes, the next generation population is the optimal inverse fact sample set;
if not, the next generation population is made to be the initial population, and the step of 'crossing and mutating individuals in the initial population to generate an intermediate population' is returned until the maximum iteration number is reached.
An interpretation text generation module T6, configured to interpret the optimal counterfacts sample according to the feature values of the optimal counterfacts sample and the low-dimensional fact sample and a preset interpretation text set, and obtain an optimal counterfacts sample interpretation text;
and the decision strategy acquisition module T7 is used for acquiring the decision strategy of the autonomous air combat in the black box model to be interpreted based on the optimal counter fact sample interpretation text.
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.
The principles and embodiments of the present invention have been described herein with reference to specific examples, the description of which is intended only to assist in understanding the methods of the present invention and the core ideas thereof; also, it is within the scope of the present invention to be modified by those of ordinary skill in the art in light of the present teachings. In view of the foregoing, this description should not be construed as limiting the invention.
Claims (6)
1. A method for generating counterfactual explanation suitable for the field of autonomous air combat is characterized by comprising the following steps:
acquiring a simulated air combat data set and carrying out low-dimensional manifold representation on the simulated air combat data set to obtain low-dimensional air combat data; the low-dimensional air combat data comprise a plurality of low-dimensional fact samples and decision labels corresponding to each low-dimensional fact sample;
Constructing a multi-objective optimization function of the optimal counter fact sample according to the optimal counter fact property;
constructing a black box model to be explained; the black box model to be explained is a machine learning model or a deep learning model;
training the black box model to be explained by using the low-dimensional fact sample and the corresponding decision tag to obtain a trained black box model to be explained;
applying a third-generation non-dominant sorting genetic algorithm according to the low-dimensional fact sample, the trained black box model to be explained and the multi-objective optimization function of the optimal counter fact sample to obtain the optimal counter fact sample;
according to the optimal counterfactual sample, the characteristic value of the low-dimensional fact sample and a preset interpretation text set, interpreting the optimal counterfactual sample to obtain an optimal counterfactual sample interpretation text;
obtaining a decision strategy of autonomous air combat in the black box model to be explained based on the optimal counterfactual sample interpretation text;
the multi-objective optimization function for constructing the optimal counterfactual sample according to the optimal counterfactual property specifically comprises the following steps:
determining a first optimization sub-objective according to the fact-sample decision tag and the low-dimensional fact-sample decision tag;
Determining a second optimization sub-objective according to the similarity between the features of the counterfactual sample and the features of the low-dimensional fact sample;
determining a third optimization sub-objective according to the number of features changed in the counterfactual sample;
determining a fourth optimization sub-objective based on differences between a plurality of said anti-facts samples generated from one of said low-dimensional facts samples;
determining a fifth optimization sub-objective based on causal constraints between features of the counterfactual sample;
taking the counterfactual sample with a specified counterfactual label in the low-dimensional fact sample as a sixth optimization sub-objective; the anti-facts label refers to a decision label corresponding to the anti-facts sample;
constructing the multi-objective optimization function according to the first optimization sub-objective, the second optimization sub-objective, the third optimization sub-objective, the fourth optimization sub-objective, the fifth optimization sub-objective and the sixth optimization sub-objective;
the multi-objective optimization function according to the low-dimensional fact sample, the trained black box model to be explained and the optimal counter fact sample applies a third generation non-dominant sorting genetic algorithm to obtain the optimal counter fact sample, and the method specifically comprises the following steps:
Randomly generating an initial population obeying Gaussian distribution based on a plurality of low-dimensional fact samples;
crossing and mutating individuals in the initial population to generate a middle population;
acquiring a set of the initial population and the intermediate population to obtain a set population;
inputting each individual in the aggregate population into the trained black box model to be explained to obtain a decision tag corresponding to each individual and calculating a plurality of optimization target values by combining the multi-objective optimization function;
non-dominant ordering is carried out on the aggregate population according to the plurality of optimized target values, and a layered pareto front is obtained;
obtaining individuals with the same number as the individuals of the initial population from the layered pareto front edge to obtain a next generation population;
judging whether the current iteration number is equal to the maximum iteration number or not;
if yes, the next generation population is the optimal inverse fact sample set;
if not, the next generation population is made to be the initial population, and the step of 'crossing and mutating individuals in the initial population to generate an intermediate population' is returned until the maximum iteration number is reached.
2. The method of claim 1, wherein the expression of the first optimization sub-objective is:
loss validity ＝-y cf log(f(x cf ))+(1-y cf )log(1-f(x cf ))
Wherein loss is validity Representing a first optimization sub-objective; x is x cf Representing a counterfactual sample; y is cf Decision labels representing counterfactual samples; f (x) cf ) A prediction label of a black box model to be explained on a counterfactual sample is represented;
the expression of the second optimization sub-objective is:
loss proximity ＝Conloss proximity +Catloss proximity
wherein loss is proximity Representing a second optimization sub-objective; conloss proximity A second optimization sub-objective representing a continuous feature; catloss proximity A second optimization sub-objective representing a discrete feature;m represents the number of consecutive features in the counterfactual sample, < >>Representing the ith successive feature of the counterfactual sample; x is x i Representing the ith successive feature of the fact sample; mac i Representing the average absolute deviation of the ith successive feature of the fact sample;if->Then I (·|·) =0, otherwise I (·|·) =1; n is the number of discrete features;
the expression of the third optimization sub-objective is:
wherein loss is sparsity Representing a third optimization sub-objective; if it isThen I (·|·) =0, otherwise I (·|·) =1;
the expression of the fourth optimization sub-objective is:
wherein loss is diversity Representing a fourth optimization sub-objective; dist (x) cfa ，x cfb ) The distance between the two counterfactual samples is calculated.
3. The method of claim 1, wherein the expression of the fifth optimization sub-objective is:
Wherein loss is causality Representing a fifth optimization sub-objective; g (-) refers to the structural causal model; e represents Gaussian noise; i 2 Representing a binary norm; />A first exogenous variable representing a counterfactual sample; the exogenous variable is a characteristic different from the continuous and discrete characteristics in the counterfactual sample;
the expression of the sixth optimization sub-objective is:
loss proto ＝||proto-z cf || 2
wherein loss is proto Representing a sixth optimization sub-objective;for an exponential kernel defined on the distance measure D, z is a hidden spatial representation of the low-dimensional fact sample;is the hidden spatial representation of the K-nearest neighbor sample with the anti-fact label of the p-th low-dimensional fact sample; k represents the number of K neighbor samples; the corner label knn represents the K-nearest neighbor algorithm.
4. The method according to claim 1, wherein the obtaining the same number of individuals as the initial population from the layered pareto front results in a next generation population, in particular comprising:
judging whether the number of individuals in the pareto front edge of the first h layers is equal to the number of individuals of the initial population;
if the first-generation population is equal to the second-generation population, taking individuals in the pareto front of the first h layers as individuals in the next-generation population;
if the number of individuals in the pareto front of the first h layer is smaller than the number of individuals in the initial population, let h=h+1, and return to the step of judging whether the number of individuals in the pareto front of the first h layer is equal to the number of individuals in the initial population;
If the number of the individuals in the pareto front of the first h-1 layer is larger than the number of the individuals in the next generation population, taking the individuals in the pareto front of the first h-1 layer as part of the individuals in the next generation population, and calculating the number of the remaining individuals;
performing scalar processing on all the optimization target values corresponding to all the individuals in the aggregate population, and generating a plurality of reference points by using a simplex method according to scalar results;
calculating the occurrence times of the reference points in the niche;
selecting the reference point with the occurrence frequency smaller than the preset frequency as a target reference point;
and selecting the individuals with the number of the residual individuals from the pareto front edge of the h layer according to the similarity between the individuals and the target reference point, and adding the individuals into the next generation population to obtain a complete next generation population.
5. A counterfactual interpretation generation system applicable to the field of autonomous air combat based on the method of any of claims 1 to 4, comprising:
the low-dimensional air combat data acquisition module is used for acquiring a simulated air combat data set and carrying out low-dimensional manifold representation on the simulated air combat data set to obtain low-dimensional air combat data; the low-dimensional air combat data comprise a plurality of low-dimensional fact samples and decision labels corresponding to each low-dimensional fact sample;
The optimization target determining module is used for constructing a multi-target optimization function of the optimal counter fact sample according to the optimal counter fact property;
the model construction module is used for constructing a black box model to be explained; the black box model to be explained is a machine learning model or a deep learning model;
the training module is used for training the black box model to be explained by utilizing the low-dimensional fact sample and the corresponding decision label to obtain a trained black box model to be explained;
the optimal counter fact sample acquisition module is used for applying a third generation non-dominant sorting genetic algorithm according to the low-dimensional fact sample, the trained black box model to be interpreted and the multi-objective optimization function of the optimal counter fact sample to obtain the optimal counter fact sample;
the interpretation text generation module is used for interpreting the optimal counterfacts sample according to the characteristic values of the optimal counterfacts sample and the low-dimensional fact sample and a preset interpretation text set to obtain an optimal counterfacts sample interpretation text;
the decision strategy acquisition module is used for acquiring a decision strategy of the autonomous air combat in the black box model to be interpreted based on the optimal counter fact sample interpretation text;
the optimization target determining module specifically comprises:
The first optimization target determining sub-module is used for determining a first optimization sub-target according to the fact that the decision label of the anti-fact sample is different from the decision label of the low-dimensional fact sample;
a second optimization objective determination submodule for determining a second optimization objective according to the similarity between the features of the inverse fact sample and the features of the low-dimensional fact sample;
a third optimization objective determination submodule for determining a third optimization objective according to the number of features changed in the counterfactual sample;
a fourth optimization objective determination sub-module for determining a fourth optimization sub-objective based on differences between a plurality of said counterfactual samples generated from one of said low-dimensional fact samples;
a fifth optimization objective determination submodule for determining a fifth optimization sub-objective according to causal constraints between features of the anti-facts sample;
a first optimization objective determination sub-module configured to take the counterfactual sample with a specified counterfactual tag in the low-dimensional fact sample as a sixth optimization sub-objective; the anti-facts label refers to decision labels corresponding to the anti-facts samples;
an optimization objective determination sub-module configured to construct the multi-objective optimization function according to the first optimization sub-objective, the second optimization sub-objective, the third optimization sub-objective, the fourth optimization sub-objective, the fifth optimization sub-objective, and the sixth optimization sub-objective;
The optimal counterfactual sample acquisition module specifically comprises:
the initial population construction submodule is used for randomly generating an initial population obeying Gaussian distribution based on a plurality of low-dimensional fact samples;
the cross mutation sub-module is used for carrying out cross and mutation on individuals in the initial population to generate a middle population;
the population merging sub-module is used for acquiring the set of the initial population and the intermediate population to obtain a set population;
the optimization target value calculation sub-module is used for inputting each individual in the aggregate population into the trained black box model to be explained, obtaining a decision tag corresponding to each individual and calculating a plurality of optimization target values by combining the multi-target optimization function;
the non-dominant ranking sub-module is used for performing non-dominant ranking on the aggregate population according to a plurality of optimized target values to obtain layered pareto fronts;
the next generation population generation sub-module is used for acquiring individuals with the same number as the individuals of the initial population from the layered pareto front edge to obtain a next generation population;
the judging sub-module is used for judging whether the current iteration number is equal to the maximum iteration number;
If yes, the next generation population is the optimal inverse fact sample set;
if not, the next generation population is made to be the initial population, and the step of 'crossing and mutating individuals in the initial population to generate an intermediate population' is returned until the maximum iteration number is reached.
6. The system of claim 5, wherein the next generation population generation submodule specifically comprises:
the judging unit is used for judging whether the number of individuals in the pareto front edge of the first h layer is equal to the number of individuals of the initial population;
if the first-generation population is equal to the second-generation population, taking individuals in the pareto front of the first h layers as individuals in the next-generation population;
if the number of individuals in the pareto front of the first h layer is smaller than the number of individuals in the initial population, let h=h+1, and return to the step of judging whether the number of individuals in the pareto front of the first h layer is equal to the number of individuals in the initial population;
if the number of the individuals in the pareto front of the first h-1 layer is larger than the number of the individuals in the next generation population, taking the individuals in the pareto front of the first h-1 layer as part of the individuals in the next generation population, and calculating the number of the remaining individuals;
the reference point generation unit is used for performing scalar processing on all the optimization target values corresponding to all the individuals in the aggregate population and generating a plurality of reference points by using a simplex method according to scalar results;
The target reference point determining unit is used for calculating the occurrence times of the reference points in the niche and selecting the reference points with the occurrence times smaller than the preset times as target reference points;
the next generation population generation unit is used for selecting the individuals with the residual individual numbers from the pareto front edge of the h layer according to the similarity between the individuals and the target reference point, and adding the individuals into the next generation population to obtain a complete next generation population.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210930992.2A CN115270634B (en) | 2022-08-04 | 2022-08-04 | Counterfactual interpretation generation method and system suitable for autonomous air combat field |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210930992.2A CN115270634B (en) | 2022-08-04 | 2022-08-04 | Counterfactual interpretation generation method and system suitable for autonomous air combat field |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115270634A CN115270634A (en) | 2022-11-01 |
| CN115270634B true CN115270634B (en) | 2023-10-20 |
Family
ID=83749219
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210930992.2A Active CN115270634B (en) | 2022-08-04 | 2022-08-04 | Counterfactual interpretation generation method and system suitable for autonomous air combat field |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115270634B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116383731B (en) * | 2023-03-06 | 2023-11-14 | 南京航空航天大学 | Tactical maneuver identification method, tactical maneuver identification system, electronic equipment and storage medium |
| CN116994687B (en) * | 2023-09-28 | 2024-02-06 | 之江实验室 | Clinical decision support model interpretation system based on inverse fact comparison |
| CN117237096B (en) * | 2023-09-28 | 2024-02-23 | 大连理工大学 | Interpretable machine learning driven credit wind control system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111240353A (en) * | 2020-01-07 | 2020-06-05 | 南京航空航天大学 | Unmanned aerial vehicle collaborative air combat decision method based on genetic fuzzy tree |
| CN112085252A (en) * | 2020-08-03 | 2020-12-15 | 清华大学 | Counterfactual prediction method about set type decision effect |
| CN112541520A (en) * | 2019-09-20 | 2021-03-23 | 罗伯特·博世有限公司 | Apparatus and method for generating counterfactual data samples for neural networks |
| CN113269310A (en) * | 2021-06-03 | 2021-08-17 | 北京邮电大学 | Graph neural network interpretable method based on counterfactual |
| US11403538B1 (en) * | 2020-11-05 | 2022-08-02 | Arthur AI, Inc. | Methods and apparatus for generating fast counterfactual explanations for black-box models using reinforcement learning |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200293918A1 (en) * | 2019-03-15 | 2020-09-17 | Cognitive Scale, Inc. | Augmented Intelligence System Assurance Engine |
| US20210174253A1 (en) * | 2019-12-10 | 2021-06-10 | Triad National Security, Llc | Analysis of system log data using machine learning |
| US20220188645A1 (en) * | 2020-12-16 | 2022-06-16 | Oracle International Corporation | Using generative adversarial networks to construct realistic counterfactual explanations for machine learning models |
-
2022
- 2022-08-04 CN CN202210930992.2A patent/CN115270634B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112541520A (en) * | 2019-09-20 | 2021-03-23 | 罗伯特·博世有限公司 | Apparatus and method for generating counterfactual data samples for neural networks |
| EP3796228A1 (en) * | 2019-09-20 | 2021-03-24 | Robert Bosch GmbH | Device and method for generating a counterfactual data sample for a neural network |
| CN111240353A (en) * | 2020-01-07 | 2020-06-05 | 南京航空航天大学 | Unmanned aerial vehicle collaborative air combat decision method based on genetic fuzzy tree |
| CN112085252A (en) * | 2020-08-03 | 2020-12-15 | 清华大学 | Counterfactual prediction method about set type decision effect |
| US11403538B1 (en) * | 2020-11-05 | 2022-08-02 | Arthur AI, Inc. | Methods and apparatus for generating fast counterfactual explanations for black-box models using reinforcement learning |
| CN113269310A (en) * | 2021-06-03 | 2021-08-17 | 北京邮电大学 | Graph neural network interpretable method based on counterfactual |
Non-Patent Citations (1)
| Title |
|---|
| 无人机集群编队及其避障控制关键技术研究;赵伟伟;中国博士学位论文数据库 工程科技Ⅱ辑;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115270634A (en) | 2022-11-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115270634B (en) | Counterfactual interpretation generation method and system suitable for autonomous air combat field | |
| Li et al. | Machine learning in aerodynamic shape optimization | |
| Bagheri et al. | Financial forecasting using ANFIS networks with quantum-behaved particle swarm optimization | |
| CN110197282B (en) | Threat estimation and situation assessment method based on genetic fuzzy logic tree | |
| Stepchenko et al. | Nonlinear, non-stationary and seasonal time series forecasting using different methods coupled with data preprocessing | |
| CN113255366A (en) | Aspect-level text emotion analysis method based on heterogeneous graph neural network | |
| Karny et al. | Dealing with complexity: a neural networks approach | |
| Yadong et al. | A fault diagnosis method for complex chemical process based on multi-model fusion | |
| Fan et al. | A two-layer Wang-Mendel fuzzy approach for predicting the residuary resistance of sailing yachts | |
| Kubalík et al. | Toward physically plausible data-driven models: A novel neural network approach to symbolic regression | |
| Zhang et al. | Closeness degree-based hesitant trapezoidal fuzzy multicriteria decision making method for evaluating green suppliers with qualitative information | |
| Han et al. | A Hybrid BPNN-GARF-SVR PredictionModel Based on EEMD for Ship Motion. | |
| Zhou et al. | Online recommendation based on incremental-input self-organizing map | |
| Muhamediyeva | Aproaches to the Construction of Fuzzy Models of Intellectual Analysis of the State of the Low-Formalized Processes | |
| CN107480381B (en) | The method of response surface model is built based on simulated annealing and applies its system | |
| CN116208399A (en) | Network malicious behavior detection method and device based on metagraph | |
| CN111610514B (en) | Inversion method and device for propagation characteristics of evaporation waveguide | |
| CN115357715A (en) | Short text clustering method based on singular value decomposition and field pre-training | |
| Jang et al. | Learning to boost training by periodic nowcasting near future weights | |
| Wu | Fuzzy robust ν-support vector machine with penalizing hybrid noises on symmetric triangular fuzzy number space | |
| Kenesei et al. | Interpretability of computational intelligence-based regression models | |
| Dominguez-Olmedo et al. | On data manifolds entailed by structural causal models | |
| Iqbal | Learning of geometric-based probabilistic self-awareness model for autonomous agents | |
| Tamayo et al. | Integrating machine learning with QFD for selecting functional requirements in construction automation | |
| Eslami et al. | Learning a Unified Latent Space for NAS: Toward Leveraging Structural and Symbolic Information |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |