Training method and device of high-dimensional space acceleration test model and related equipment

Abstract

The invention provides a training method and device for a high-dimensional space acceleration test model and related equipment, and relates to the technical field of high-dimensional sampling optimization, and the method comprises the steps: carrying out the clustering of a plurality of first parameter combinations, obtaining a plurality of class clusters, and enabling the number of the first parameter combinations in each class cluster to be greater than or equal to a first threshold value; searching all the first parameter combinations included in the plurality of class clusters based on an objective function to obtain a plurality of second parameter combinations, the second parameter combinations being first parameter combinations enabling the value output by the function of the objective function to be greater than or equal to a second threshold value; and constructing a training sample set according to the plurality of second parameter combinations, and training the initial model according to the training sample set to obtain a target model which is used for predicting the failure probability of the intelligent driving algorithm when the plurality of algorithm parameters are configured by inputting the parameter combinations of the target model. According to the invention, vulnerability identification of the intelligent driving algorithm can be completed efficiently and accurately.

Description

Technical Field

[0001] The present disclosure relates to the technical field of high-dimensional sampling optimization, and in particular to a training method, apparatus, and related equipment for a high-dimensional space acceleration test model. Background Art

[0002] With the rapid development of autonomous driving technology, the safety, reliability, and robustness of smart cars have become a focus of industry attention. Numerous vulnerabilities exist in smart driving algorithms, many of which have a long tail effect, meaning they are only triggered in a very few specific situations, but once triggered, they can cause serious consequences or even catastrophic losses. Furthermore, statistics show that traditional manual testing methods require extensive manual repetitive testing, and the average cost of manpower and material resources to discover a single serious vulnerability often reaches millions of dollars, making them extremely inefficient. Against this backdrop, using AI to generate simulation scenarios for testing has become a new breakthrough. However, ensuring both detection speed and coverage while maintaining sufficient computing power remains a challenge.

[0003] Previous studies have shown that efficiently sampling and searching for anomalies in high-dimensional parameter spaces is a challenge in autonomous driving testing. Researchers have proposed various approaches to identify new high-risk locations based on existing test points, including random sampling, Latin hypercube design, Bayesian optimization, reverse / adversarial testing, reinforcement learning path planning, and fuzz testing. The most straightforward methods, such as random Monte Carlo sampling or grid enumeration, suffer from the curse of dimensionality in high-dimensional environments, requiring a very large number of samples to cover rare anomalies with a high probability, resulting in significant inefficiency. To improve coverage, designs such as Latin Hypercube Sampling (LHS) have been introduced. LHS uses stratified sampling to ensure a uniform distribution of test points in high-dimensional space, making it suitable for scenarios with a large number of parameter factors. Compared to pure random sampling, LHS exhibits better uniformity in its one-dimensional projection, helping to improve the spatial coverage of sampling. However, these methods remain blindly exploratory and lack specificity, especially when anomalies have a long-tail distribution, potentially missing extreme scenarios. Some improved methods employ phased, multiple sampling to gradually approach the target region, allowing them to identify local or global anomalies with fewer samples. However, building proxy models through multiple iterations can lead to insufficient global accuracy, as each round requires solving an optimization problem to select new points, resulting in high computational overhead. In general, traditional random / uniform sampling methods struggle to balance efficiency and coverage, often necessitating a trade-off between sample size and anomaly detection rate.

[0004] Bayesian optimization uses a surrogate model combined with an acquisition function to intelligently guide test point selection, enabling the identification of high-risk areas with a limited number of tests. A typical approach involves first defining a parameter space based on real-world scenario data, then selecting an appropriate surrogate classifier (such as a Gaussian process or random forest) and acquisition function to evaluate the next scenario parameter combination most likely to trigger a risk. Bayesian optimization strikes a balance between exploration (developing new areas) and exploitation (refining the search for high-risk areas), achieving results similar to those of an exhaustive search of a large number of samples. However, its limitations lie in the fact that when the number of parameter dimensions or candidate scenario combinations is large, the computational complexity of traditional surrogate models such as Gaussian processes increases dramatically, slowing training and inference. Typical models such as Gaussian processes and TPE experience significant performance degradation when the number of scenarios reaches 100,000, necessitating the use of simpler classifiers such as nearest neighbor models to accommodate large sample sizes. Even so, Bayesian optimization requires continuous training and updating of the surrogate model, resulting in high computational complexity. Furthermore, its effectiveness relies on the accuracy of the surrogate model. If the classifier underpredicts certain hazardous areas, it may result in missed detections (incomplete coverage of abnormal scenarios). Therefore, the application of Bayesian optimization methods in high-dimensional and ultra-large spaces will face efficiency bottlenecks and the risk of missed detection.

[0005] To more effectively uncover long-tail anomalies, many studies are turning to intelligent algorithms to "reverse" search for new failure scenarios starting from existing problem samples. For example, one work proposes an edge scenario generation method based on scenario adversarial and reinforcement learning. This method can automatically generate low-probability, high-risk traffic scenarios and recreate adversarial game behavior between vehicles in simulation. This method models the scenario evolution as a closed-loop system with dynamic environmental elements. Using deep reinforcement learning to train a black-box controller for the scenario, it continuously optimizes the behavior of surrounding vehicles or pedestrians to approximate the failure conditions of the main vehicle. This type of reinforcement learning or genetic algorithm-driven method can target the failure boundary and excel in scene interaction and rare operating conditions, improving the detection rate of abnormal scenarios. Furthermore, at the perception level, the generation of adversarial examples for images and perturbation testing of autonomous driving sensor inputs also fall under the category of adversarial testing and can also reveal model vulnerabilities under specific inputs. However, these methods also have drawbacks. First, training an intelligent agent capable of generating complex scenarios often requires extensive trial-and-error simulations, resulting in significant testing overhead. Secondly, if the guidance strategy is inappropriate, purely pursuing the worst-case scenario may generate extreme scenarios that violate common sense in physics. As some studies have pointed out, some scenarios obtained through worst-case search are almost impossible to occur in reality and have little practical significance. Therefore, it is necessary to impose constraints on the generation process to ensure the rationality of the scenarios. In addition, reinforcement learning or adversarial search methods are complex to implement and are often customized for specific scenario problems, making them relatively lacking in versatility. Fuzz testing, on the other hand, uses random mutations of existing test cases to explore anomalies. Although it can discover unexpected vulnerabilities to a certain extent, due to the lack of clear guidance signals, efficiency and accuracy are difficult to guarantee in high-dimensional scenarios.

[0006] It can be seen that there is an urgent need for a method that can conduct testing efficiently and discover as many vulnerabilities as possible, so as to quickly and accurately discover potential problems in intelligent driving algorithms. Summary of the Invention

[0007] The purpose of the present disclosure is to provide a training method, device and related equipment for a high-dimensional space acceleration test model, which is used to solve the technical problem that the existing technology has in the vulnerability identification of intelligent driving algorithms, and it is difficult to strike a balance between recognition accuracy and recognition efficiency.

[0008] In a first aspect, an embodiment of the present invention provides a method for training a high-dimensional space acceleration test model, the method comprising:

[0009] Clustering the multiple first parameter combinations to obtain multiple clusters, wherein the number of first parameter combinations included in each cluster is greater than or equal to a first threshold, each first parameter combination includes multiple parameter values, the multiple parameter values ​​correspond to multiple algorithm parameters of the intelligent driving algorithm in a one-to-one manner, and different first parameter combinations include different multiple parameter values;

[0010] All first parameter combinations included in the multiple clusters are searched based on an objective function to obtain multiple second parameter combinations, wherein a function input of the objective function is the first parameter combination, and a function output of the objective function represents: when the multiple algorithm parameters are configured with the corresponding first parameter combination, the risk of failure of the intelligent driving algorithm is represented, and the greater the value of the function output, the higher the risk; the second parameter combination is: a first parameter combination that makes the value of the function output of the objective function greater than or equal to a second threshold;

[0011] A training sample set is constructed according to the multiple second parameter combinations, and an initial model is trained based on the training sample set to obtain a target model, wherein the initial model is a deep neural network model, and the target model is used to predict the probability of failure of the intelligent driving algorithm when the multiple algorithm parameters are configured by inputting the parameter combination of the target model.

[0012] In a second aspect, an embodiment of the present invention further provides a training device for a high-dimensional space acceleration test model, the device comprising:

[0013] a clustering module, configured to cluster the plurality of first parameter combinations to obtain a plurality of clusters, wherein the number of first parameter combinations included in each cluster is greater than or equal to a first threshold, each first parameter combination includes a plurality of parameter values, the plurality of parameter values ​​correspond one-to-one to a plurality of algorithm parameters of the intelligent driving algorithm, and different first parameter combinations include different parameter values;

[0014] a search module configured to search all first parameter combinations included in the multiple clusters based on an objective function to obtain multiple second parameter combinations, wherein a function input of the objective function is the first parameter combination, and a function output of the objective function represents: when the multiple algorithm parameters are configured with the corresponding first parameter combination, the risk of failure of the intelligent driving algorithm is represented, and a larger value of the function output indicates a higher risk; and the second parameter combination is: a first parameter combination that makes the function output value of the objective function greater than or equal to a second threshold;

[0015] A training module is used to construct a training sample set based on the multiple second parameter combinations, and train an initial model based on the training sample set to obtain a target model, wherein the initial model is a deep neural network model, and the target model is used to predict the probability of failure of the intelligent driving algorithm when the multiple algorithm parameters are configured by inputting the parameter combination of the target model.

[0016] In a third aspect, an embodiment of the present disclosure further provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-mentioned training method for the high-dimensional space acceleration test model.

[0017] In a fourth aspect, an embodiment of the present disclosure further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the training method of the high-dimensional space acceleration test model are implemented.

[0018] In an embodiment of the present invention, suspicious areas are located in the vast parameter space corresponding to multiple first parameter combinations through clustering, that is, areas with a higher probability of causing the intelligent driving algorithm to fail are located, so as to reduce the scale of space that needs to be explored subsequently. Then, multiple second parameter combinations that will cause the intelligent driving algorithm to fail are selected from the suspicious areas, and a training sample set is constructed based on this to train the deep neural network model, thereby obtaining a target model that can quickly process high-dimensional parameter space, so as to directly predict the probability of causing the intelligent driving algorithm to fail through the input parameter combination, eliminating the process of using parameter combinations to conduct simulation experiments to determine whether the intelligent driving algorithm has failed, and further eliminating the time and computing resource overhead caused by conducting the above-mentioned simulation experiments. This can achieve higher recognition accuracy and efficiency in identifying vulnerabilities in the intelligent driving algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the description of the embodiments of the present disclosure. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0020] Figure 1 This is a flowchart of a training method for a high-dimensional space acceleration test model provided by an embodiment of the present disclosure;

[0021] Figure 2 Schematic diagram of a high-dimensional space acceleration test model training device provided by an embodiment of the present disclosure;

[0022] Figure 3 It is a structural diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present disclosure in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present disclosure without making any creative efforts shall fall within the scope of protection of the present disclosure.

[0024] The embodiment of the present invention provides a training method for a high-dimensional space acceleration test model, such as Figure 1 As shown, the method includes:

[0025] Step 101: Cluster multiple first parameter combinations to obtain multiple clusters.

[0026] Among them, the number of first parameter combinations included in each of the clusters is greater than or equal to the first threshold, each of the first parameter combinations includes multiple parameter values, the multiple parameter values ​​correspond one-to-one to multiple algorithm parameters of the intelligent driving algorithm, and different first parameter combinations include different multiple parameter values.

[0027] The multiple first parameter combinations are derived from the high-dimensional parameter space of the intelligent driving algorithm. The high-dimensional parameter space of the intelligent driving algorithm is specifically: a composite space formed by fusing the parameter value space of each parameter value of the multiple parameter values ​​of the intelligent driving algorithm.

[0028] Among them, the multiple first parameter combinations are specifically: multiple parameter combinations that have been tested among all parameter combinations supported by the high-dimensional parameter space of the intelligent driving algorithm. The multiple first parameter combinations can be obtained by sampling all the parameter combinations. Based on this setting, the amount of data for clustering operations can be reduced to further improve the resource overhead and time overhead of the solution described in the present invention in the clustering process.

[0029] The fact that different first parameter combinations include different parameter values ​​should be understood as follows: for any two different first parameter combinations in the multiple first parameter combinations, the intersection of the two is a proper subset of the union of the two.

[0030] These clusters can be understood as regions / subspaces corresponding to a large number of parameter combinations. When certain parameter combinations (such as environmental conditions and sensor reading configurations) form dense clusters in a high-dimensional parameter space, this may indicate a high frequency of abnormal algorithm behavior or performance degradation under these conditions. In other words, in the regions / subspaces represented by these clusters, the intelligent driving algorithm may face potential problem scenarios. Therefore, obtaining these multiple clusters can be considered a preliminary identification of parameter regions in the high-dimensional parameter space that may contain vulnerabilities. This significantly reduces the amount of data required for subsequent processing, thereby significantly improving the overall processing efficiency of the solution described herein.

[0031] Among them, the setting of the first threshold can avoid incorporating clusters with too few elements in the cluster into subsequent processing, so as to reduce the mixing of interference, ensure the accuracy of subsequent processing results, and reduce the amount of data required for subsequent processing, thereby improving subsequent processing efficiency.

[0032] Step 102: Search all first parameter combinations included in the multiple clusters based on the objective function to obtain multiple second parameter combinations.

[0033] Among them, the function input of the objective function is the first parameter combination, and the function output of the objective function represents: when the multiple algorithm parameters are configured with the corresponding first parameter combination, the risk of failure of the intelligent driving algorithm, and the larger the value of the function output, the higher the risk; the second parameter combination is: the first parameter combination that makes the value of the function output of the objective function greater than or equal to the second threshold.

[0034] Among them, when the value of the function output of the objective function is greater than or equal to the second threshold, it can be regarded that the corresponding parameter combination will cause the intelligent driving algorithm to fail.

[0035] Exemplarily, the failure of the intelligent driving algorithm includes: lane departure, incorrect obstacle identification, abnormal acceleration or deceleration, etc.

[0036] Step 103: construct a training sample set according to the plurality of second parameter combinations, and train the initial model based on the training sample set to obtain a target model.

[0037] The initial model is a deep neural network model, and the target model is used to predict the probability of failure of the intelligent driving algorithm when the multiple algorithm parameters are configured by inputting the parameter combination of the target model.

[0038] It should be noted that after obtaining the target model, all parameter combinations supported by the high-dimensional parameter space can be input into the target model in batches to obtain the probability that each parameter combination will cause the intelligent driving algorithm to fail, and the parameter combination with the corresponding probability greater than or equal to the probability threshold is determined as an abnormal parameter combination, that is, a parameter combination that causes the intelligent driving algorithm to fail / abnormal. Based on this, the vulnerability detection of the intelligent driving algorithm can be completed. Since each parameter combination is detected, the above-mentioned vulnerability detection process fully covers the high-dimensional parameter space and can discover as many vulnerabilities as possible. Moreover, since it does not involve using the corresponding parameter combination for simulation experiments, but rather performs feature analysis on the parameter combination at the data level, the detection of the corresponding parameter combination can be completed quickly, which can improve the efficiency of vulnerability identification.

[0039] In an embodiment of the present invention, suspicious areas are located in the vast parameter space corresponding to multiple first parameter combinations through clustering, that is, areas with a higher probability of causing the intelligent driving algorithm to fail are located, so as to reduce the scale of space that needs to be explored subsequently. Then, multiple second parameter combinations that will cause the intelligent driving algorithm to fail are selected from the suspicious areas, and a training sample set is constructed based on this to train the deep neural network model, thereby obtaining a target model that can quickly process high-dimensional parameter space, so as to directly predict the probability of causing the intelligent driving algorithm to fail through the input parameter combination, eliminating the process of using parameter combinations to conduct simulation experiments to determine whether the intelligent driving algorithm has failed, and further eliminating the time and computing resource overhead caused by conducting the above-mentioned simulation experiments. This can achieve higher recognition accuracy and efficiency in identifying vulnerabilities in the intelligent driving algorithm.

[0040] In one embodiment, the action of step 102 above can be implemented based on the DBSCAN algorithm.

[0041] Specifically, clustering the multiple first parameter combinations to obtain multiple clusters includes:

[0042] Performing distance calculation on the multiple first parameter combinations to obtain distance information, wherein the distance information is used to represent: an Lp norm between any two different first parameter combinations in the multiple first parameter combinations;

[0043] Analyze the distance information to determine multiple minimum neighbor distances, where the multiple minimum neighbor distances correspond to the multiple first parameter combinations in a one-to-one manner, and the minimum neighbor distance is a minimum value among multiple Lp norms of the corresponding first parameter combinations;

[0044] determining a domain radius according to an average value and a standard deviation of the plurality of minimum neighbor distances;

[0045] According to the domain radius and the distance information, the multiple first parameter combinations are clustered to obtain multiple clusters, wherein for any first parameter combination in the multiple clusters, there is at least one adjacent parameter combination in the corresponding cluster, and the adjacent parameter combination is a first parameter combination whose Lp norm with the corresponding first parameter combination is less than or equal to the domain radius.

[0046] Wherein, the Lp norm is:

[0047]

[0048] In the above formula, d(x i ,x j ) is the Lp norm between the i-th first parameter combination and the j-th first parameter combination in the plurality of first parameter combinations, x ikis the kth parameter value in the i-th first parameter combination, x jk is the kth parameter value in the jth first parameter combination, n is the number of parameters of the multiple algorithm parameters, and p is the p value of the Lp norm.

[0049] In one example, the p may be 2, in which case the Lp norm may be understood as the Euclidean distance.

[0050] The distance information is analyzed to determine multiple minimum neighbor distances, which can be expressed as:

[0051] For the sample x i , find the minimum neighbor distance Among them, the sample x i is the i-th first parameter combination.

[0052] The calculation formula of the field radius is as follows:

[0053] ε=μ ε +λ·σ ε , λ∈[0,1]

[0054] In the above formula, ε is the radius of the field, μ is ε is the average value of the multiple minimum neighbor distances, σ ε is the standard deviation of the multiple minimum neighbor distances, and λ is the robustness adjustment factor. Based on the setting of the above calculation formula, the central tendency and dispersion of the distance distribution between multiple first parameter combinations are comprehensively considered to make the calculated area radius more accurate and reliable.

[0055] The calculation formula for the average value of the multiple minimum neighbor distances is as follows:

[0056]

[0057] In the above formula, m is the number of the multiple minimum neighbor distances.

[0058] The calculation formula for the standard deviation of the multiple minimum neighbor distances is as follows:

[0059]

[0060] The process of clustering the multiple first parameter combinations according to the domain radius and the distance information can be understood as a process of performing density cluster analysis on the multiple first parameter combinations using the domain radius as the distance threshold and the first threshold as the minimum sample number threshold MinPts. Specifically,

[0061] First, determine the ε-neighborhood set N corresponding to multiple first parameter combinations ε (xi ) = {x j | d(x i , x j ) ≤ ε}, that is, the set of all first parameter combinations whose distance from x i does not exceed ε.

[0062] If the number of elements included in a neighborhood in the neighborhood set is greater than or equal to the first threshold, that is, |N ε (x i )| ≥ MinPts, then x i is marked as a core point (Core Point), indicating that this point is inside a high-density area.

[0063] If |N ε (x i )| < MinPts, then it is considered that x i is not a core point; however, if x i is not a noise point and falls into the neighborhood of a certain core point, then x i can be marked as a border point (BorderPoint). Among them, those points (i.e., the first parameter combinations) that are neither core points nor within the neighborhood of any core point are marked as noise points. [[ID=三十二]] [[ID=三十三]]

[0064] [[ID=三十四]]By distinguishing core points, border points, and noise points from the ε-neighborhood set, a group of points that are density-reachable (Density-Reachable) and density-connected (Density-Connected) to each other among all core points are grouped into a cluster (high-density area). Specifically, if the sample points x[[ID=三十五]] i [[ID=三十六]]and x[[ID=三十七]] j [[ID=三十八]]are both core points, and x[[ID=三十九]] j [[ID=四十]]∈ N[[ID=四十一]] ε [[ID=四十二]](x[[ID=四十三]] i [[ID=四十四]]), then they are assigned to the same cluster; then this rule is iteratively applied to the neighborhood points of all core points within the cluster, continuously expanding the scope of the cluster to obtain the multiple clusters. [[ID=四十五]] [[ID=四十六]]

[0065] [[ID=四十七]]In this embodiment, by using the density clustering method, several parameter combinations that have been tested in the high-dimensional parameter space are analyzed to quickly discover suspicious high-risk areas in the high-dimensional parameter space from the past test information. Compared with the clueless global search, this method limits the search scope to near several high-density clusters (i.e., clusters) through rough screening, greatly reducing the scale of the space that needs to be explored subsequently, ensuring that the subsequent processing focuses on areas where potential anomalies are more concentrated, and improving the detection probability of long-tail scenarios. [[ID=四十八]] [[ID=四十九]]

[0066] [[ID=五十]]In one embodiment, the distance information is obtained based on the processing of multiple nodes included in the distributed cluster, and the two first parameter combinations corresponding to the Lp norms calculated by different nodes are different.

[0067] In this embodiment, multiple nodes are used to process multiple distance calculation tasks corresponding to the distance information in parallel, and each node is responsible for a part of the multiple distance calculation tasks, so as to greatly reduce the time complexity of processing multiple distance calculation tasks, thereby improving the data processing efficiency in the clustering process.

[0068] In one embodiment, the searching of all first parameter combinations included in the plurality of clusters based on the objective function to obtain a plurality of second parameter combinations includes:

[0069] Generate a loss function based on the objective function, wherein a function output of the loss function is the inverse of the function output of the objective function;

[0070] Using all first parameter combinations included in the multiple clusters as the parameter search space of the loss function, performing stochastic gradient descent on the loss function to obtain the multiple second parameter combinations, where the second parameter combinations are the first parameter combinations that make the loss function converge;

[0071] Wherein, the objective function is defined as:

[0072]

[0073] In the above formula, x is the data space composed of multiple parameter values ​​included in the target first parameter combination, and the target first parameter combination is a first parameter combination in the multiple clusters, θ k is the kth parameter value of the multiple parameter values ​​included in the target first parameter combination, K is the total number of the multiple parameter values, and the function f k (x;θ k ) represents: when the multiple algorithm parameters are configured with the target first parameter combination, the behavior performance index of the intelligent driving algorithm, the behavior performance index is an index for measuring the risk; function g k (f k (x;θ k )) is determined by the function f k (x;θ k ) is a differentiable function, Ω e (x,A e ) is used to represent multiple regular constraint items corresponding to the multiple algorithm parameters, E is the number of constraint items of the multiple regular constraint items, It is used to express the constraint term that is sensitive to the change of the input parameter value, α k is the weight corresponding to the kth parameter value, λ e is the weight corresponding to the e-th regular constraint term, and γ is The corresponding weight.

[0074] In this embodiment, by converting the objective function into a loss function and adopting the stochastic gradient descent method, multiple first parameter combinations that make the loss function converge are quickly identified from all the first parameter combinations included in the multiple clusters, so as to achieve detailed risk identification of the suspicious areas obtained by clustering, ensure the data accuracy of the identified multiple second parameter combinations, and significantly improve the efficiency of identifying multiple second parameter combinations from all the first parameter combinations included in the multiple clusters.

[0075] Specifically, the stochastic gradient descent process described above can be roughly understood as multiple "directed adversarial tests," which rapidly approximate the precise parameter values ​​(i.e., the second parameter combination) that trigger an anomaly within a continuous parameter search space, without requiring numerous random attempts like in Monte Carlo. Compared to traditional Bayesian optimization methods that require global training of proxy models each time, gradient search directly utilizes information derived from the loss function, resulting in faster convergence and more accurate identification of the extreme scenario parameters (i.e., the second parameter combination) that cause failures.

[0076] In the present invention, a differentiable function should be understood as a function for which sufficient conditions exist for total differentiation.

[0077] In application, when the function f k (x;θ k ) cannot be converted into a differentiable function, a proxy model or simulation-based gradient estimation method can be used to approximate the gradient.

[0078] In one embodiment, the plurality of second parameter combinations are obtained through distributed cluster computing, and the plurality of nodes included in the distributed cluster share a lock-free memory.

[0079] In this embodiment, multiple computing tasks corresponding to stochastic gradient descent are calculated in parallel by multiple nodes, so as to cooperate with the lock-free memory sharing mechanism to accelerate the efficiency of objective function convergence and improve the efficiency of obtaining multiple second parameter combinations.

[0080] Among them, multiple nodes sharing lock-free memory can be understood as:

[0081] For each of the multiple nodes, after independently completing the local gradient estimation in the stochastic gradient descent process, it can immediately perform non-blocking update without waiting for the completion of calculations of other nodes.

[0082] Specifically, the current parameter vector x (i.e., a first parameter combination) is stored in a memory area shared by all computing nodes (e.g., using a parameter server or multi-threaded shared memory). The current global parameter vector x tThe parameters are stored in a shared memory area (such as a distributed parameter server or a thread-safe cache) that is accessible to all worker nodes. The rth node reads the current parameter copy from the shared memory at the tth iteration and calculates the local loss function for the assigned sample subset. The node then immediately performs the following update:

[0083]

[0084] Where η is the learning rate (step size), which controls the magnitude of each gradient update. Due to the asynchronous nature of this approach, nodes do not need to wait for each other to complete their computations before performing a unified update. Instead, they execute the above update formula in parallel at any time. This means that in practice, the update to x may be based on slightly outdated parameter states (because other nodes may have updated x in the meantime), but this lock-free parallel update significantly improves computational efficiency.

[0085] In terms of implementation details, lock-free updates can prevent update conflicts through atomic operations: each node's update of a specific component of the parameter (i.e., the parameter value of a specific algorithm parameter) is performed atomically on the underlying hardware to ensure that there is no serious write competition. The regularization term λΩ(x) generates an additional term in the gradient calculation (For example, for |x| 2 Regularization, whose gradient is 2λx), will synchronously affect the gradient calculation of all nodes, thereby automatically limiting the growth of parameters to extreme values ​​during the update.

[0086] Among them, the process of using distributed computing to perform stochastic gradient descent can be called distributed asynchronous stochastic gradient descent (Lock-Free Asynchronous SGD, LFA-SGD).

[0087] In one embodiment, constructing a training sample set according to the plurality of second parameter combinations, and training an initial model based on the training sample set to obtain a target model includes:

[0088] Normalizing the multiple second parameter combinations to obtain multiple standard parameter combinations, wherein each standard parameter combination includes multiple standard values, the multiple standard values ​​correspond to the multiple parameter values ​​one-to-one, and the multiple standard values ​​corresponding to the same algorithm parameter are normally distributed;

[0089] A training sample set is constructed according to the plurality of standard parameter combinations, and the initial model is trained based on the training sample set to obtain the target model.

[0090] In this embodiment, the data distribution of different groups of parameter values ​​corresponding to different algorithm parameters is adjusted to a uniform scale through normalization processing to alleviate problems such as uneven gradients and training oscillations caused by differences in feature distribution. On the basis of keeping the expected value of each dimension of input feature at 0 and the standard deviation approximately at 1 (i.e., normal distribution), some learnable transformation degrees of freedom are retained, so that the trained target model can obtain better prediction results.

[0091] It should be noted that, in this embodiment, after obtaining the target model, the parameter combination input into the target model also needs to undergo the aforementioned normalization process.

[0092] Exemplarily, the normalization process may be performed based on the following formula:

[0093]

[0094] In the above formula, x' k is the kth parameter value after normalization in the second parameter combination, x k is the kth parameter value before normalization in the second parameter combination, μ k is the mean value of the kth parameter value in the second parameter combination before normalization, σ k is the standard deviation of the kth parameter value in the second parameter combination before normalization, δ is the offset term, ∈ is the stability term, and γ k and β k are the trainable scaling parameter and offset parameter corresponding to the kth parameter value before normalization in the second parameter combination.

[0095] Among them, in the normalization process, the offset term δ, the stability term ∈ and the trainable scaling and offset parameters γ are introduced k , β k , thereby further enhancing the expressiveness and numerical robustness of the model.

[0096] In one embodiment, the initial model includes:

[0097] The input layer, hidden layer, and output layer are connected in series;

[0098] The input layer is used to extract multiple dimensional features of the input parameter combination, wherein the multiple dimensional features correspond to the multiple algorithm parameters in a one-to-one manner;

[0099] The hidden layer is used to perform setting processing on the input multiple dimensional features respectively, wherein the setting processing includes: linear calculation and nonlinear activation performed in sequence;

[0100] The output layer is used to fuse the multiple dimensional features processed by the above settings.

[0101] It should be noted that the hidden layer can be one or two layers, so that the initial model exists as a lightweight deep neural network model, thereby reducing the training overhead of the target model and improving the reasoning efficiency of the target model, thereby improving the efficiency of comprehensive vulnerability detection of intelligent driving algorithms.

[0102] After the above normalization process, the standard parameter combination can be expressed as: x'=[x'1,x'2,…,x' k ], the weight matrix and bias of the first hidden layer are W (1) and b (1) The linear transformation output z of the hidden layer (1) And the activation function output h (1) It can be expressed as:

[0103] z (1) =W (1) x'+b (1) +α (1) ⊙x'+R (1) ·ψ(x')

[0104]

[0105] In the above formula, W (1) is the weight matrix of the first hidden layer (dimension k×n, k represents the number of multiple algorithm parameters, n represents the number of samples in the training sample set), b (1) is the bias vector. Based on the conventional linear transformation, the model introduces the residual modulation term α (1) ⊙x' and high-order feature channel R (1) ·ψ(x') to enhance the responsiveness and nonlinear expression of the input dimension. The output is normalized before activation, and then adjusted by the learnable scaling γ and offset β, and then input into the activation function φ(·). The final output is also superimposed with an input residual path ξ (1) ⊙x' is used to retain low-order information. This structure integrates normalization, residual and auxiliary channel mechanisms, significantly improving feature modeling capabilities while maintaining network lightweight. The above formula is calculated element by element: For the jth neuron in the hidden layer, we have:

[0106]

[0107] That is, the linear combination of the input features is first summed through weights and biases, and then nonlinear activation is applied to obtain the output of the hidden unit. In this way, the hidden layer extracts nonlinear feature representations from the original features, providing a more informative representation for the next layer's prediction.

[0108] If the model contains multiple hidden layers, similar calculations are repeated. Specifically, the input of the second hidden layer is the output h of the first hidden layer. (1) , through the corresponding weight matrix W (2) , bias b (2) Linear transformation and activation function to obtain the new hidden representation h (2) Similarly, after several layers of nonlinear transformation, the final deep feature representation of the model will be obtained.

[0109] The output layer of the model is used to give the confidence or probability that the input parameter combination belongs to the "problem area". Assume that the activation output of the last hidden layer is h (L) , the output layer uses the weight vector w (o) (with h (L) The dimensions remain consistent) and bias b (o) Perform a linear transformation on it and obtain the output value through the activation function (the Sigmoid function σ is usually used for binary classification problems)

[0110] z (o) =W (o) ·(γ⊙h (L) +δ·ψ(h (L) ))+b (o)

[0111] Here (o) It is the linear weighted sum (scalar) of the output layer, which is composed of multiple sub-channels. Its value is mapped to the final output of the model after being processed by the activation function.

[0112] The activation function uses the Sigmoid form: This function compresses any real value into the interval [0,1], so that the output result has probabilistic meaning. The output value is used to express the probability that the model predicts that the current input parameter combination belongs to the abnormal problem area (that is, the probability that the intelligent driving algorithm fails). For example, when When the output value is close to 1, it means that the model highly believes that the scene corresponding to the input may cause intelligent driving abnormalities; on the contrary, if If it is close to 0, it means that the scene is more likely to be normal or safe. Set the probability threshold (such as 0.5) to convert the probability into a binary classification label, that is, to determine whether it belongs to the problem scenario.

[0113] In one example, the target model of the present invention can be called a flexible lightweight neural network index (FLEX) model, which is intended to serve as a learning index in a high-dimensional parameter space to quickly predict whether a given parameter combination belongs to a potential problem area.

[0114] Exemplarily, constructing a training sample set according to the plurality of second parameter combinations includes:

[0115] The multiple second parameter combinations are used as negative samples corresponding to high-risk abnormal scenarios, and several parameter combinations corresponding to normal scenarios are used as positive samples, and labels y are set for positive samples and negative samples respectively (y=1 indicates that the parameter combination is confirmed as a problem scenario, and y=0 indicates a normal scenario).

[0116] The goal of initial model training is to minimize the difference between the predicted output and the true label. The cross entropy loss function J(θ) can be used as the optimization target, where θ represents the set of all trainable parameters (weights and biases) of the model. For binary cross entropy, there is:

[0117]

[0118] Where N is the total number of training samples in the training sample set, y (i) is the true label of the i-th sample, is the probability that the model predicts that the i-th sample is an abnormal scenario. By minimizing the cross entropy, the model's ability to distinguish between positive and negative categories can be improved. In applications, the adaptive optimization algorithm Adam can be used as an upgraded version of gradient descent to train network parameters. The Adam optimizer dynamically adjusts the learning rate based on the first-order and second-order moment estimates in each iteration, thereby accelerating convergence and improving stability. The Adam optimizer calculates the first-order momentum m of the gradient t and the second-order momentum v t To correct the learning step size, the update rule can be summarized as:

[0119] First-order moment (momentum) update: m t =β1,m t-1 +(1-β1),g t

[0120] Second-order moment (mean square) update:

[0121] Parameter update:

[0122] where g tis the gradient of the current batch, β1,β2 are momentum decay coefficients (e.g. 0.9 and 0.999), α is the global learning rate, and ∈ is a small constant to avoid division by zero.

[0123] By using the Adam algorithm, the model parameters of the initial model can be adjusted at an appropriate pace during the training process to quickly approach the optimal solution.

[0124] In this embodiment, a lightweight deep neural network model is trained to serve as a learning index for the high-dimensional parameter space, which is equivalent to a "guide index" that accelerates the search over the high-dimensional parameter space, thereby significantly reducing invalid tests. By instantly generalizing existing test experience and providing risk estimates for unseen scenarios, the vulnerability detection efficiency of the intelligent driving algorithm can be further improved.

[0125] In some embodiments, the training sample set can be randomly divided into a training set and a validation set. The training set is used to update model parameters, and the validation set is used to evaluate the model's performance on unseen data, thereby preventing overfitting. By monitoring the model's key evaluation metrics on the validation set, such as accuracy, recall, and F1 score, it is possible to ensure that the model has good generalization ability.

[0126] The specific calculation method is as follows:

[0127] Accuracy measures the proportion of correct predictions overall: Among them, TP, TN

[0128] They are the number of samples that the model predicts as positive and negative (true positive and true negative), respectively. FP and FN are the number of incorrect predictions (false positive and false negative).

[0129] Recall focuses on the proportion of true positive classes that are successfully identified by the model: It indicates how many of all actual abnormal scenarios are recognized by the model. The higher the recall rate, the fewer missed detections.

[0130] The F1 score is the harmonic average of precision and recall, and is used to comprehensively evaluate the performance of the model in positive class detection: in

[0131] The higher the F1 value, the stronger the model's ability to detect abnormal scenarios while controlling false positives.

[0132] After multiple rounds of iterative training, when the validation set metrics stabilize and meet expected requirements, the model training is considered converged. This enables the converged target model to quickly evaluate new parameter combinations with high accuracy and generalization, predicting whether they represent potential problem areas. During deployment, this lightweight target model serves as an index for simulation testing: For large, high-dimensional parameter spaces, it is no longer necessary to exhaustively run a full simulation of all possible combinations. Instead, the target model is used to predict and screen high-risk candidate areas, and simulation verification is then focused on these areas, greatly accelerating the vulnerability detection process for intelligent driving algorithms.

[0133] See also Figure 2 , Figure 2 This is a training device for a high-dimensional space acceleration test model provided by an embodiment of the present disclosure, such as Figure 2 As shown, the training device 200 of the high-dimensional space acceleration test model includes:

[0134] a clustering module 201 for clustering the plurality of first parameter combinations to obtain a plurality of clusters, wherein the number of first parameter combinations included in each cluster is greater than or equal to a first threshold, each first parameter combination includes a plurality of parameter values, the plurality of parameter values ​​correspond one-to-one to a plurality of algorithm parameters of the intelligent driving algorithm, and different first parameter combinations include different parameter values;

[0135] A search module 202 is configured to search all first parameter combinations included in the multiple clusters based on an objective function to obtain multiple second parameter combinations, wherein the function input of the objective function is the first parameter combination, and the function output of the objective function represents: when the multiple algorithm parameters are configured with the corresponding first parameter combination, the risk of failure of the intelligent driving algorithm is represented, and the greater the value of the function output, the higher the risk; the second parameter combination is: a first parameter combination that makes the function output value of the objective function greater than or equal to a second threshold;

[0136] The training module 203 is used to construct a training sample set based on the multiple second parameter combinations, and train the initial model based on the training sample set to obtain a target model, wherein the initial model is a deep neural network model, and the target model is used to predict the probability of failure of the intelligent driving algorithm when the multiple algorithm parameters are configured by inputting the parameter combination of the target model.

[0137] In one embodiment, the clustering module 201 is specifically configured to:

[0138] Performing distance calculation on the multiple first parameter combinations to obtain distance information, wherein the distance information is used to represent: an Lp norm between any two different first parameter combinations in the multiple first parameter combinations;

[0139] Analyze the distance information to determine multiple minimum neighbor distances, where the multiple minimum neighbor distances correspond to the multiple first parameter combinations in a one-to-one manner, and the minimum neighbor distance is a minimum value among multiple Lp norms of the corresponding first parameter combinations;

[0140] determining a domain radius according to an average value and a standard deviation of the plurality of minimum neighbor distances;

[0141] According to the domain radius and the distance information, the multiple first parameter combinations are clustered to obtain multiple clusters, wherein for any first parameter combination in the multiple clusters, there is at least one adjacent parameter combination in the corresponding cluster, and the adjacent parameter combination is a first parameter combination whose Lp norm with the corresponding first parameter combination is less than or equal to the domain radius.

[0142] In one embodiment, the distance information is obtained based on processing of multiple nodes included in the distributed cluster, and the two first parameter combinations corresponding to the Lp norms calculated by different nodes are different.

[0143] In one embodiment, the searching of all first parameter combinations included in the plurality of clusters based on the objective function to obtain a plurality of second parameter combinations includes:

[0144] Generate a loss function based on the objective function, wherein a function output of the loss function is the inverse of the function output of the objective function;

[0145] Using all first parameter combinations included in the multiple clusters as the parameter search space of the loss function, performing stochastic gradient descent on the loss function to obtain the multiple second parameter combinations, where the second parameter combinations are the first parameter combinations that make the loss function converge;

[0146] Wherein, the objective function is defined as:

[0147]

[0148] In the above formula, x is the data space composed of multiple parameter values ​​included in the target first parameter combination, and the target first parameter combination is a first parameter combination in the multiple clusters, θ k is the kth parameter value of the multiple parameter values ​​included in the target first parameter combination, K is the total number of the multiple parameter values, and the function f k (x;θ k ) represents: when the multiple algorithm parameters are configured with the target first parameter combination, the behavior performance index of the intelligent driving algorithm, the behavior performance index is an index for measuring the risk; function g k (fk (x;θ k )) is determined by the function f k (x;θ k ) is a differentiable function, Ω e (x,A e ) is used to represent multiple regular constraint items corresponding to the multiple algorithm parameters, E is the number of constraint items of the multiple regular constraint items, It is used to express the constraint term that is sensitive to the change of the input parameter value, α k is the weight corresponding to the kth parameter value, λ e is the weight corresponding to the e-th regular constraint term, γ is The corresponding weight.

[0149] In one embodiment, the plurality of second parameter combinations are obtained through distributed cluster computing, and the plurality of nodes included in the distributed cluster share a lock-free memory.

[0150] In one embodiment, the search module 202 is specifically configured to:

[0151] Normalizing the multiple second parameter combinations to obtain multiple standard parameter combinations, wherein each standard parameter combination includes multiple standard values, the multiple standard values ​​correspond to the multiple parameter values ​​one-to-one, and the multiple standard values ​​corresponding to the same algorithm parameter are normally distributed;

[0152] A training sample set is constructed according to the plurality of standard parameter combinations, and the initial model is trained based on the training sample set to obtain the target model.

[0153] In one embodiment, the training module 203 is specifically configured to:

[0154] Normalizing the multiple second parameter combinations to obtain multiple standard parameter combinations, wherein each standard parameter combination includes multiple standard values, the multiple standard values ​​correspond to the multiple parameter values ​​one-to-one, and the multiple standard values ​​corresponding to the same algorithm parameter are normally distributed;

[0155] A training sample set is constructed according to the plurality of standard parameter combinations, and the initial model is trained based on the training sample set to obtain the target model.

[0156] In one embodiment, the initial model includes:

[0157] The input layer, hidden layer, and output layer are connected in series;

[0158] The input layer is used to extract multiple dimensional features of the input parameter combination, wherein the multiple dimensional features correspond to the multiple algorithm parameters in a one-to-one manner;

[0159] The hidden layer is used to perform setting processing on the input multiple dimensional features respectively, wherein the setting processing includes: linear calculation and nonlinear activation performed in sequence;

[0160] The output layer is used to fuse the multiple dimensional features processed by the above settings.

[0161] The training device 200 for the high-dimensional space acceleration test model provided in the embodiment of the present disclosure can implement each process in the above method embodiment. To avoid repetition, it will not be described here.

[0162] According to an embodiment of the present disclosure, the present disclosure further provides an electronic device and a readable storage medium.

[0163] Figure 3 A schematic block diagram of an example electronic device 300 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are provided as examples only and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0164] like Figure 3 As shown, the device 300 includes a computing unit 301, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 302 or a computer program loaded from a storage unit 308 into a random access memory (RAM) 303. Various programs and data required for the operation of the device 300 can also be stored in the RAM 303. The computing unit 301, the ROM 302, and the RAM 303 are connected to each other via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0165] Various components in device 300 are connected to I / O interface 305, including: an input unit 306, such as a keyboard, mouse, etc.; an output unit 307, such as various types of displays, speakers, etc.; a storage unit 308, such as a magnetic disk, optical disk, etc.; and a communication unit 309, such as a network card, modem, wireless communication transceiver, etc. The communication unit 309 allows device 300 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0166] The computing unit 301 can be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 301 performs the various methods and processes described above, such as the training method of the high-dimensional space acceleration test model. For example, in some embodiments, the training method of the high-dimensional space acceleration test model can be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 308. In some embodiments, part or all of the computer program can be loaded and / or installed on the device 300 via the ROM 302 and / or the communication unit 309. When the computer program is loaded into the RAM 303 and executed by the computing unit 301, one or more steps of the training method of the high-dimensional space acceleration test model described above can be performed. Alternatively, in other embodiments, the computing unit 301 may be configured to execute the training method of the high-dimensional space acceleration test model in any other appropriate manner (for example, by means of firmware).

[0167] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-chip (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special-purpose or general-purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0168] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0169] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0170] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0171] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0172] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0173] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0174] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A training method for a high-dimensional space acceleration test model, characterized in that: The method comprises: Clustering the multiple first parameter combinations to obtain multiple clusters, wherein the number of first parameter combinations included in each cluster is greater than or equal to a first threshold, each first parameter combination includes multiple parameter values, the multiple parameter values ​​correspond to multiple algorithm parameters of the intelligent driving algorithm in a one-to-one manner, and different first parameter combinations include different multiple parameter values; All first parameter combinations included in the multiple clusters are searched based on an objective function to obtain multiple second parameter combinations, wherein a function input of the objective function is the first parameter combination, and a function output of the objective function represents: when the multiple algorithm parameters are configured with the corresponding first parameter combination, the risk of failure of the intelligent driving algorithm is represented, and the greater the value of the function output, the higher the risk; the second parameter combination is: a first parameter combination that makes the value of the function output of the objective function greater than or equal to a second threshold; A training sample set is constructed according to the multiple second parameter combinations, and an initial model is trained based on the training sample set to obtain a target model, wherein the initial model is a deep neural network model, and the target model is used to predict the probability of failure of the intelligent driving algorithm when the multiple algorithm parameters are configured by inputting the parameter combination of the target model.

2. The training method according to claim 1, characterized in that The clustering of the plurality of first parameter combinations to obtain a plurality of clusters includes: Performing distance calculation on the multiple first parameter combinations to obtain distance information, wherein the distance information is used to represent: an Lp norm between any two different first parameter combinations in the multiple first parameter combinations; Analyze the distance information to determine multiple minimum neighbor distances, where the multiple minimum neighbor distances correspond to the multiple first parameter combinations in a one-to-one manner, and the minimum neighbor distance is a minimum value among multiple Lp norms of the corresponding first parameter combinations; determining a domain radius according to an average value and a standard deviation of the plurality of minimum neighbor distances; According to the domain radius and the distance information, the multiple first parameter combinations are clustered to obtain multiple clusters, wherein for any first parameter combination in the multiple clusters, there is at least one adjacent parameter combination in the corresponding cluster, and the adjacent parameter combination is a first parameter combination whose Lp norm with the corresponding first parameter combination is less than or equal to the domain radius.

3. The method according to claim 2, characterized in that The distance information is obtained based on processing of multiple nodes included in the distributed cluster, and the two first parameter combinations corresponding to the Lp norms calculated by different nodes are different.

4. The method according to claim 1, wherein The step of searching all first parameter combinations included in the plurality of clusters based on the objective function to obtain a plurality of second parameter combinations includes: Generate a loss function based on the objective function, wherein a function output of the loss function is the inverse of the function output of the objective function; Using all first parameter combinations included in the multiple clusters as the parameter search space of the loss function, performing stochastic gradient descent on the loss function to obtain the multiple second parameter combinations, where the second parameter combinations are the first parameter combinations that make the loss function converge; Wherein, the objective function is defined as: In the above formula, x is the data space composed of multiple parameter values ​​included in the target first parameter combination, and the target first parameter combination is a first parameter combination in the multiple clusters, θ k is the kth parameter value of the multiple parameter values ​​included in the target first parameter combination, K is the total number of the multiple parameter values, and the function f k (x;θ k ) represents: when the multiple algorithm parameters are configured with the target first parameter combination, the behavior performance index of the intelligent driving algorithm, the behavior performance index is an index for measuring the risk; function g k (f k (x;θ k )) is determined by the function f k (x;θ k ) is a differentiable function, Ω e (x,A e ) is used to represent multiple regular constraint items corresponding to the multiple algorithm parameters, E is the number of constraint items of the multiple regular constraint items, It is used to express the constraint term that is sensitive to the change of the input parameter value, α k is the weight corresponding to the kth parameter value, λ e is the weight corresponding to the e-th regular constraint term, and γ is The corresponding weight.

5. The method according to claim 4, characterized in that The multiple second parameter combinations are obtained through distributed cluster calculation, and the multiple nodes included in the distributed cluster share a lock-free memory.

6. The method according to claim 1, characterized in that The step of constructing a training sample set according to the plurality of second parameter combinations and training an initial model based on the training sample set to obtain a target model includes: Normalizing the multiple second parameter combinations to obtain multiple standard parameter combinations, wherein each standard parameter combination includes multiple standard values, the multiple standard values ​​correspond to the multiple parameter values ​​one-to-one, and the multiple standard values ​​corresponding to the same algorithm parameter are normally distributed; A training sample set is constructed according to the plurality of standard parameter combinations, and the initial model is trained based on the training sample set to obtain the target model.

7. The method according to claim 1, characterized in that The initial model includes: The input layer, hidden layer, and output layer are connected in series; The input layer is used to extract multiple dimensional features of the input parameter combination, wherein the multiple dimensional features correspond to the multiple algorithm parameters in a one-to-one manner; The hidden layer is used to perform setting processing on the input multiple dimensional features respectively, wherein the setting processing includes: linear calculation and nonlinear activation performed in sequence; The output layer is used to fuse the multiple dimensional features processed by the above settings.

8. A training device for a high-dimensional space acceleration test model, characterized in that: The device comprises: a clustering module, configured to cluster the plurality of first parameter combinations to obtain a plurality of clusters, wherein the number of first parameter combinations included in each cluster is greater than or equal to a first threshold, each first parameter combination includes a plurality of parameter values, the plurality of parameter values ​​correspond one-to-one to a plurality of algorithm parameters of the intelligent driving algorithm, and different first parameter combinations include different parameter values; a search module configured to search all first parameter combinations included in the multiple clusters based on an objective function to obtain multiple second parameter combinations, wherein a function input of the objective function is the first parameter combination, and a function output of the objective function represents: when the multiple algorithm parameters are configured with the corresponding first parameter combination, the risk of failure of the intelligent driving algorithm is represented, and a larger value of the function output indicates a higher risk; and the second parameter combination is: a first parameter combination that makes the function output value of the objective function greater than or equal to a second threshold; A training module is used to construct a training sample set based on the multiple second parameter combinations, and train an initial model based on the training sample set to obtain a target model, wherein the initial model is a deep neural network model, and the target model is used to predict the probability of failure of the intelligent driving algorithm when the multiple algorithm parameters are configured by inputting the parameter combination of the target model.

9. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the method according to any one of claims 1 to 7 when executed by the processor.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which implements the steps of the method according to any one of claims 1 to 7 when executed by a processor.

Images