Decision reliability evaluation method, device and equipment based on trusted graph neural network

Abstract

The application relates to a decision reliability evaluation method, device and equipment based on a trusted graph neural network. The method comprises the following steps: obtaining the confidence of an initial graph neural network model according to a logit vector output by the initial graph neural network model, performing decision on a decision sample by using the initial graph neural network model to obtain the accuracy rate of a decision result; calculating the expected calibration error of the initial graph neural network model according to the confidence and the accuracy rate of the decision result; constructing a graph calibration loss function of the initial graph neural network model, training the initial graph neural network model by using the graph calibration loss function to obtain a trusted graph neural network model; and performing decision evaluation according to the calibration error of the trusted graph neural network model and the trusted graph neural network model. The method can improve the decision reliability.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, computer device, and storage medium for evaluating decision reliability based on a trusted graph neural network. Background Technology

[0002] Autonomous driving technology is one of the most rapidly developing technologies today. In everyday autonomous driving scenarios, sensors are needed to detect information about the surrounding environment to automatically control the vehicle's actions. For example, at traffic intersections, during lane changes, and during acceleration and deceleration, sensors are required to detect information about the surrounding environment to control the vehicle's movements. However, when the information sensed and output by the sensors is distorted, masked, or malfunctions, errors will occur in the vehicle's automatic control. Current autonomous driving technology includes a decision center containing multiple decision models. These models are used to make decisions in some unexpected situations during autonomous driving. However, errors often occur when multiple decision models make decisions in relevant environments. Therefore, a method is needed to evaluate the reliability of decision models and measure whether a decision model can be used in real-world decision-making scenarios. Summary of the Invention

[0003] Therefore, it is necessary to provide a decision reliability assessment method, apparatus, computer equipment, and storage medium based on a trusted graph neural network that can improve decision reliability in response to the above-mentioned technical problems.

[0004] A decision reliability assessment method based on a reliable graph neural network, the method comprising:

[0005] Obtain the initial graph neural network model and test set; the graph neural network model is a decision model; the test set includes decision samples and decision results.

[0006] The confidence level of the initial graph neural network model is obtained from the logit vector output by the initial graph neural network model. The initial graph neural network model is then used to make decisions on the decision samples to obtain the accuracy of the decision results.

[0007] The expected calibration error of the initial graph neural network model is calculated based on the confidence level and the accuracy of the decision results.

[0008] Based on the expected calibration error magnitude, determine whether the initial graph neural network model needs calibration. Based on the determination result, construct the graph calibration loss function of the initial graph neural network model. Use the graph calibration loss function to train the initial graph neural network model to obtain a reliable graph neural network model.

[0009] Decision evaluation is based on the calibration error of the trusted graph neural network model and the trusted graph neural network model itself.

[0010] In one embodiment, the expected calibration error of the initial graph neural network model is calculated based on the confidence level and the accuracy of the decision result, including:

[0011] Based on the confidence level and the accuracy of the decision results, the expected calibration error of the initial graph neural network model is calculated as follows: in, Labels representing decision outcomes, y i This represents the true labels in the test set. Let represent the prediction distribution, which represents the confidence level for the network to make a prediction decision; p represents the target distribution, which represents the true probability that the network will make the correct decision.

[0012] In one embodiment, it is determined whether the initial graph neural network model needs calibration based on the expected calibration error magnitude, and a graph calibration loss function for the initial graph neural network model is constructed based on the determination result, including:

[0013] The initial graph neural network model is determined to be calibrated based on the expected calibration error. If the expected calibration error is greater than the preset error threshold, the initial graph neural network model is determined to be calibrated, and the graph calibration loss function of the initial graph neural network model is constructed.

[0014] In one embodiment, the process of constructing the graph calibration loss function for the initial graph neural network model includes:

[0015] Obtain the target distribution and the prediction distribution; the target distribution is the true probability that the network makes a correct decision; the prediction distribution is the confidence level of the initial graph neural network model in making a prediction decision;

[0016] The divergence between the target distribution and the predicted distribution is calculated to obtain the KL divergence.

[0017] A regularizer is constructed by multiplying the entropy of the target distribution and the predicted distribution. The regularization of the KL divergence and the regularizer is calculated to obtain the graph calibration loss function.

[0018] In one embodiment, the graph calibration loss function is obtained by calculating the KL divergence and the regularization of the regularizer, including:

[0019] The graph calibration loss function is obtained by calculating the KL divergence and the regularization of the regularizer. Where γ represents the regularization parameter, Represents the predicted distribution The product of the entropy of the target distribution p and the entropy of the target distribution p.

[0020] In one embodiment, decision evaluation is performed based on the calibration error of the trusted graph neural network model and the trusted graph neural network model, including:

[0021] The calibration error of the trusted graph neural network model is judged. If the calibration error is not less than the preset error threshold, the trusted graph neural network model is calibrated again until the calibration error of the trusted graph neural network model is less than the preset error threshold. Then, the trusted graph neural network model is used to make a decision.

[0022] A decision reliability assessment device based on a reliable graph neural network, the device comprising:

[0023] The data acquisition module is used to acquire the initial graph neural network model and the test set; the graph neural network model is a decision model; the test set includes decision samples and decision results.

[0024] The expected calibration error calculation module is used to obtain the confidence level of the initial graph neural network model based on the logit vector output by the initial graph neural network model, use the initial graph neural network model to make decisions on decision samples, and obtain the accuracy of the decision results; and calculate the expected calibration error of the initial graph neural network model based on the confidence level and the accuracy of the decision results.

[0025] The decision evaluation module is used to determine whether the initial graph neural network model needs to be calibrated based on the expected calibration error. Based on the judgment result, a graph calibration loss function for the initial graph neural network model is constructed. The initial graph neural network model is trained using the graph calibration loss function to obtain a reliable graph neural network model. The decision evaluation is performed based on the calibration error of the reliable graph neural network model and the reliable graph neural network model.

[0026] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0027] Obtain the initial graph neural network model and test set; the graph neural network model is a decision model; the test set includes decision samples and decision results.

[0028] The confidence level of the initial graph neural network model is obtained from the logit vector output by the initial graph neural network model. The initial graph neural network model is then used to make decisions on the decision samples to obtain the accuracy of the decision results.

[0029] The expected calibration error of the initial graph neural network model is calculated based on the confidence level and the accuracy of the decision results.

[0030] Based on the expected calibration error magnitude, determine whether the initial graph neural network model needs calibration. Based on the determination result, construct the graph calibration loss function of the initial graph neural network model. Use the graph calibration loss function to train the initial graph neural network model to obtain a reliable graph neural network model.

[0031] Decision evaluation is based on the calibration error of the trusted graph neural network model and the trusted graph neural network model itself.

[0032] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0033] Obtain the initial graph neural network model and test set; the graph neural network model is a decision model; the test set includes decision samples and decision results.

[0034] The confidence level of the initial graph neural network model is obtained from the logit vector output by the initial graph neural network model. The initial graph neural network model is then used to make decisions on the decision samples to obtain the accuracy of the decision results.

[0035] The expected calibration error of the initial graph neural network model is calculated based on the confidence level and the accuracy of the decision results.

[0036] Based on the expected calibration error magnitude, determine whether the initial graph neural network model needs calibration. Based on the determination result, construct the graph calibration loss function of the initial graph neural network model. Use the graph calibration loss function to train the initial graph neural network model to obtain a reliable graph neural network model.

[0037] Decision evaluation is based on the calibration error of the trusted graph neural network model and the trusted graph neural network model itself.

[0038] The aforementioned decision reliability assessment method, apparatus, computer equipment, and storage medium based on a trusted graph neural network (GNN) first obtains the confidence level of the initial GNN model based on the logit vector output by the initial GNN model. Then, the initial GNN model is used to make decisions on decision samples to obtain the accuracy of the decision results. Based on the confidence level and the accuracy of the decision results, the expected calibration error of the initial GNN model is calculated. This expected calibration error is then used for subsequent trusted GNN model evaluation. Confidence level is a crucial factor in determining whether a model can be used in real-world decision-making scenarios; it means that the model should accurately reflect when it might be wrong. The initial GNN model's reliability is assessed based on the magnitude of the expected calibration error. The initial graph neural network model is assessed for the need for calibration. Based on the assessment result, a graph calibration loss function is constructed for the initial graph neural network model. This graph calibration loss function is then used to train the initial graph neural network model, resulting in a reliable graph neural network model. The reliability of decisions is then evaluated based on this reliable graph neural network model. This application minimizes the regularized KL divergence between high-confidence and low-confidence samples using the graph calibration loss function. This not only ensures the minimization of KL divergence but also reduces the Shannon entropy of the prediction distribution. This encourages the model to become more confident. The graph calibration loss function makes correct decisions more confident and incorrect decisions less reliable, leading to a better calibrated model. Therefore, the graph calibration loss function can calibrate the model while ensuring high accuracy, resulting in a more reliable reliable graph neural network model. Decision-making based on the reliable graph neural network model can reduce the confidence of incorrect decisions and increase the confidence of correct decisions. Furthermore, the accuracy of decision evaluation can be improved by using the calibration error of the reliable graph neural network model for decision evaluation. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating a decision reliability assessment method based on a trusted graph neural network in one embodiment;

[0040] Figure 2 This is a schematic diagram comparing the confidence levels of a graph neural network model and a trusted graph neural network model in one embodiment;

[0041] Figure 3 Here is a histogram showing the confidence distribution of a graph neural network model and a trusted graph neural network model in one embodiment;

[0042] Figure 4 This is a structural block diagram of a decision reliability assessment device based on a trusted graph neural network in one embodiment;

[0043] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0045] In one embodiment, such as Figure 1 As shown, a decision reliability assessment method based on a trusted graph neural network is provided, including the following steps:

[0046] Step 102: Obtain the initial graph neural network model and test set; the graph neural network model is a decision model; the test set includes decision samples and decision results.

[0047] Existing graph neural network models are poorly calibrated and have low reliability when making decisions, while reliable neural network models are necessary in safety-critical fields. Decision samples represent various real-world decision-making scenarios faced by the decision center of an autonomous vehicle, and decision results represent the correct decisions made by the validated decision-making models for various real-world scenarios.

[0048] Step 104: Obtain the confidence level of the initial graph neural network model based on the logit vector output by the initial graph neural network model; use the initial graph neural network model to make decisions on the decision samples to obtain the accuracy of the decision results; calculate the expected calibration error of the initial graph neural network model based on the confidence level and the accuracy of the decision results.

[0049] The expected calibration error of the initial graph neural network model is calculated based on the confidence level and the accuracy of the decision results. This expected calibration error is then used for subsequent model evaluation. Confidence level is a crucial factor in determining whether a model can be used in real-world decision-making scenarios; it means that the model should accurately reflect when it might be wrong.

[0050] Step 106: Determine whether the initial graph neural network model needs to be calibrated based on the expected calibration error magnitude. Construct a graph calibration loss function for the initial graph neural network model based on the determination result. Train the initial graph neural network model using the graph calibration loss function to obtain a reliable graph neural network model. Make a decision evaluation based on the calibration error of the reliable graph neural network model and the reliable graph neural network model.

[0051] The divergence between the target and predicted distributions is calculated to obtain the KL divergence. A regularizer is constructed based on the product of the entropies of the target and predicted distributions. The regularization of the KL divergence and the regularizer is then calculated to obtain the graph calibration loss function (GCL). The GCL minimizes the regularized KL divergence between high-confidence and low-confidence samples. This not only ensures the minimization of KL divergence but also reduces the Shannon entropy of the predicted distribution, encouraging the model to become more confident. GCL makes correct decisions more confident and incorrect decisions less confident. It reshapes the standard cross-entropy loss and uses a minimum entropy regularization term to regularize the KL divergence. GCL is the first end-to-end calibration method for GNNs, leading to better calibrated models through this training method. Therefore, GCL can calibrate the model while ensuring high accuracy, resulting in a more reliable and accurate trusted graph neural network model. Decision-making based on a trusted graph neural network model can reduce the confidence of incorrect decisions and increase the confidence of correct decisions. Furthermore, the trusted graph neural network model can be used to calibrate errors for decision evaluation, thereby improving the accuracy of decision evaluation.

[0052] In the aforementioned decision reliability assessment method based on trusted graph neural networks, the confidence level of the initial graph neural network model is first obtained from the logit vector output by the initial graph neural network model. The initial graph neural network model is then used to make decisions on the decision samples to obtain the accuracy of the decision results. Based on the confidence level and the accuracy of the decision results, the expected calibration error of the initial graph neural network model is calculated. This expected calibration error is then used for subsequent trusted graph neural network model evaluation. Confidence level is a crucial factor in determining whether a model can be used in real-world decision-making scenarios; it means that the model should accurately reflect when it might be wrong. The magnitude of the expected calibration error is used to evaluate the initial graph neural network model. The model is evaluated based on whether calibration is needed. A graph calibration loss function is constructed for the initial graph neural network model, and this function is used to train the initial model, resulting in a reliable graph neural network model. The reliability of decisions is assessed based on this reliable graph neural network model. This application minimizes the regularized KL divergence between high-confidence and low-confidence samples by constructing the graph calibration loss function. This not only ensures the minimization of KL divergence but also reduces the Shannon entropy of the prediction distribution, encouraging the model to become more confident. The graph calibration loss function makes correct decisions more confident and incorrect decisions less reliable, leading to a better-calibrated model. Therefore, the graph calibration loss function can calibrate the model while ensuring high accuracy, resulting in a more reliable reliable graph neural network model. Making decisions based on the reliable graph neural network model reduces the confidence of incorrect decisions and increases the confidence of correct decisions. Furthermore, the accuracy of decision evaluation is improved by using the calibration error of the reliable graph neural network model for decision assessment.

[0053] In one embodiment, the expected calibration error of the initial graph neural network model is calculated based on the confidence level and the accuracy of the decision result, including:

[0054] Based on the confidence level and the accuracy of the decision results, the expected calibration error of the initial graph neural network model is calculated as follows: in, Labels representing decision outcomes, y i This represents the true labels in the test set. Let represent the prediction distribution, which represents the confidence level for the network to make a prediction decision; p represents the target distribution, which represents the true probability that the network will make the correct decision.

[0055] In one embodiment, it is determined whether the initial graph neural network model needs calibration based on the expected calibration error magnitude, and a graph calibration loss function for the initial graph neural network model is constructed based on the determination result, including:

[0056] The initial graph neural network model is determined to be calibrated based on the expected calibration error. If the expected calibration error is greater than the preset error threshold, the initial graph neural network model is determined to be calibrated, and the graph calibration loss function of the initial graph neural network model is constructed.

[0057] In one embodiment, the process of constructing the graph calibration loss function for the initial graph neural network model includes:

[0058] Obtain the target distribution and the prediction distribution; the target distribution is the true probability that the network makes a correct decision; the prediction distribution is the confidence level of the initial graph neural network model in making a prediction decision;

[0059] The divergence between the target distribution and the predicted distribution is calculated to obtain the KL divergence.

[0060] A regularizer is constructed by multiplying the entropy of the target distribution and the predicted distribution. The regularization of the KL divergence and the regularizer is calculated to obtain the graph calibration loss function.

[0061] In one embodiment, the graph calibration loss function is obtained by calculating the KL divergence and the regularization of the regularizer, including:

[0062] The graph calibration loss function is obtained by calculating the KL divergence and the regularization of the regularizer. Where γ represents the regularization parameter, Represents the predicted distribution The product of the entropy of the target distribution p and the entropy of the target distribution p.

[0063] In one embodiment, decision evaluation is performed based on the calibration error of the trusted graph neural network model and the trusted graph neural network model, including:

[0064] The calibration error of the trusted graph neural network model is judged. If the calibration error is not less than the preset error threshold, the trusted graph neural network model is calibrated again until the calibration error of the trusted graph neural network model is less than the preset error threshold. Then, the trusted graph neural network model is used to make a decision.

[0065] In one embodiment, it is theoretically proven that the graph calibration loss function can calibrate a graph neural network model, making its decision-making more reliable.

[0066] In the following equation, the strength of γ balances two regularization terms: the entropy regularization term is used to improve the reliability of the model, and the KL-divergence term is used to maintain classification accuracy.

[0067]

[0068] in, It is the cross-entropy loss, and K represents the number of classes. As is well known,

[0069]

[0070] Therefore, the relationship between GCL and KL divergence is as follows:

[0071]

[0072] This indicates that the graph calibration loss is a lower bound on the regularization of the KL divergence and prediction entropy. The GCL can be considered as adding a minimum entropy regularizer to the same implicit minimization as the cross-entropy loss. At the target probability p... y In the case of ∈{0,1}, the prediction confidence equals the true probability if and only if the prediction confidence is equal to the true probability. This will be minimized. Therefore, it is encouraged... Reducing the KL divergence helps narrow the gap between predicted confidence and true probability, thus improving calibration. Specifically, when GCL is minimized, KL divergence can be minimized, while reducing the predicted distribution during training. The entropy. It can be seen that minimizing the GCL leads to minimizing the KL divergence, which maintains the model's accuracy and the predicted distribution entropy. Minimizing the graph calibration loss function calibrates the graph neural network model.

[0073] In a specific embodiment, using Figure 2The reliability plots illustrate the results of GCN on different datasets. The GCN in the left plot uses cross-entropy loss, while the GCN in the right plot uses GCL. The x-axis represents the confidence score of samples in 20 bins, and the y-axis represents the average precision of samples in each bin. The left plot shows a histogram of accuracy as a function of confidence on the three datasets before calibration, while the right plot shows a histogram of accuracy as a function of confidence on the three datasets after calibration using GCL. Figure 2 As can be seen, the average accuracy in the left-hand plot is much higher than the average confidence, which corresponds to the low confidence of the GNN. By comparing the reliability plots before and after calibration, we found that in the right-hand plot, the gap between confidence and accuracy is significantly reduced, and the histogram fits the diagonal better, indicating that the GCL calibration GNN in this application is more confident.

[0074] like Figure 3 The image shows a visualization of the confidence distribution for correct and incorrect decision samples, where the x-axis represents the confidence level of samples in each of the 50 bins, and the y-axis represents the percentage of samples in each bin. Before calibration, we can observe that a large number of correctly classified samples are distributed in the low confidence range, while some incorrectly classified samples are in the high confidence range. After using GCL, the situation of a large number of correctly classified samples having low confidence levels is significantly improved. Figure 3 As shown in the right figure, the distribution of samples with incorrect decisions varies little, while samples with correct decisions are encouraged to have higher confidence and tend to follow a long-tailed distribution, with the majority in the head class. According to... Figure 3 Observations show that after GCL calibration, correct decision samples tend to exhibit greater confidence, while incorrect decision samples show little change, which is beneficial for model self-training. Therefore, using a GCL-based self-trained model can improve the accuracy and reliability of decision evaluation.

[0075] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but may be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0076] In one embodiment, such as Figure 4As shown, a decision reliability assessment device based on a trusted graph neural network is provided, comprising: a data acquisition module 402, an expected calibration error calculation module 404, and a model training and decision assessment module 406, wherein:

[0077] Data acquisition module 402 is used to acquire the initial graph neural network model and test set; the graph neural network model is a decision model; the test set includes decision samples and decision results;

[0078] The expected calibration error calculation module 404 is used to obtain the confidence level of the initial graph neural network model based on the logit vector output by the initial graph neural network model, use the initial graph neural network model to make decisions on decision samples, and obtain the accuracy of the decision results; and calculate the expected calibration error of the initial graph neural network model based on the confidence level and the accuracy of the decision results.

[0079] The decision evaluation module 406 is used to determine whether the initial graph neural network model needs to be calibrated based on the expected calibration error. Based on the judgment result, a graph calibration loss function of the initial graph neural network model is constructed. The initial graph neural network model is trained using the graph calibration loss function to obtain a reliable graph neural network model. The decision evaluation is performed based on the calibration error of the reliable graph neural network model and the reliable graph neural network model.

[0080] In one embodiment, the expected calibration error calculation module 404 is further configured to calculate, based on the confidence level and the accuracy of the decision result, the expected calibration error of the initial graph neural network model, including:

[0081] Based on the confidence level and the accuracy of the decision results, the expected calibration error of the initial graph neural network model is calculated as follows: in, Labels representing decision outcomes, y i This represents the true labels in the test set. Let represent the prediction distribution, which represents the confidence level for the network to make a prediction decision; p represents the target distribution, which represents the true probability that the network will make the correct decision.

[0082] In one embodiment, the decision evaluation module 406 is further configured to determine whether the initial graph neural network model needs calibration based on the expected calibration error magnitude, and to construct a graph calibration loss function for the initial graph neural network model based on the determination result, including:

[0083] The initial graph neural network model is determined to be calibrated based on the expected calibration error. If the expected calibration error is greater than the preset error threshold, the initial graph neural network model is determined to be calibrated, and the graph calibration loss function of the initial graph neural network model is constructed.

[0084] In one embodiment, the process of constructing the graph calibration loss function for the initial graph neural network model includes:

[0085] Obtain the target distribution and the prediction distribution; the target distribution is the true probability that the network makes a correct decision; the prediction distribution is the confidence level of the initial graph neural network model in making a prediction decision;

[0086] The divergence between the target distribution and the predicted distribution is calculated to obtain the KL divergence.

[0087] A regularizer is constructed by multiplying the entropy of the target distribution and the predicted distribution. The regularization of the KL divergence and the regularizer is calculated to obtain the graph calibration loss function.

[0088] In one embodiment, the decision evaluation module 406 is further configured to calculate the regularization of the KL divergence and the regularizer to obtain a graph calibration loss function, including:

[0089] The graph calibration loss function is obtained by calculating the KL divergence and the regularization of the regularizer. Where γ represents the regularization parameter, Represents the predicted distribution The product of the entropy of the target distribution p and the entropy of the target distribution p.

[0090] In one embodiment, the decision evaluation module 406 is further configured to perform a decision evaluation based on the calibration error of the trusted graph neural network model and the trusted graph neural network model, including:

[0091] The calibration error of the trusted graph neural network model is judged. If the calibration error is not less than the preset error threshold, the trusted graph neural network model is calibrated again until the calibration error of the trusted graph neural network model is less than the preset error threshold. Then, the trusted graph neural network model is used to make a decision.

[0092] Specific limitations regarding the decision reliability assessment device based on trusted graph neural networks can be found in the limitations of the decision reliability assessment method based on trusted graph neural networks mentioned above, and will not be repeated here. Each module in the aforementioned decision reliability assessment device based on trusted graph neural networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0093] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a decision reliability assessment method based on a trusted graph neural network. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0094] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0095] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0096] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0097] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A decision reliability assessment method based on a reliable graph neural network, characterized in that, The method includes: Obtain an initial graph neural network model and a test set; the graph neural network model is a decision model; the test set includes decision samples and decision results; the decision samples represent various actual decision scenarios faced by the decision center of the autonomous vehicle, and the decision results represent the correct decision results made by the decision model in various actual decision scenarios that have been verified; The confidence level of the initial graph neural network model is obtained from the logit vector output by the initial graph neural network model. The initial graph neural network model is then used to make decisions on the decision samples to obtain the accuracy of the decision results. The expected calibration error of the initial graph neural network model is calculated based on the confidence level and the accuracy of the decision result. Based on the expected calibration error magnitude, determine whether the initial graph neural network model needs calibration. Based on the determination result, construct a graph calibration loss function for the initial graph neural network model. Use the graph calibration loss function to train the initial graph neural network model to obtain a reliable graph neural network model. Decision evaluation is performed based on the calibration error of the trusted graph neural network model and the trusted graph neural network model itself. The expected calibration error of the initial graph neural network model is calculated based on the confidence level and the accuracy of the decision result, including: Based on the confidence level and the accuracy of the decision result, the expected calibration error of the initial graph neural network model is calculated as follows: ,in, Labels indicating decision outcomes This represents the true labels in the test set. This represents the predicted distribution and the confidence level for the network to make predictive decisions. This represents the target distribution, which is the true probability that the network will make the correct decision. Based on the expected calibration error magnitude, determine whether the initial graph neural network model needs calibration, and construct a graph calibration loss function for the initial graph neural network model based on the determination result, including: Based on the expected calibration error, it is determined whether the initial graph neural network model needs to be calibrated. If the expected calibration error is greater than a preset error threshold, it is determined that the initial graph neural network model needs to be calibrated, and a graph calibration loss function for the initial graph neural network model is constructed. The process of constructing the graph calibration loss function of the initial graph neural network model includes: Obtain the target distribution and the prediction distribution; the target distribution is the true probability that the network makes a correct decision; the prediction distribution is the confidence level of the initial graph neural network model in making a prediction decision; The KL divergence is obtained by calculating the divergence between the target distribution and the predicted distribution; A regularizer is constructed based on the product of the entropy of the target distribution and the predicted distribution. The regularization of the KL divergence and the regularization of the regularizer are calculated to obtain the graph calibration loss function.

2. The method according to claim 1, characterized in that, The graph calibration loss function is obtained by calculating the KL divergence and the regularization of the regularizer, including: The graph calibration loss function is obtained by calculating the KL divergence and the regularization of the regularizer. ,in, Represents the regularization parameter. Represents the predicted distribution and target distribution p The product of entropy.

3. The method according to claim 1, characterized in that, Decision evaluation based on the calibration error of the trusted graph neural network model and the trusted graph neural network model includes: The calibration error of the trusted graph neural network model is judged. If the calibration error is not less than a preset error threshold, the trusted graph neural network model is calibrated again until the calibration error of the trusted graph neural network model is less than the preset error threshold. Then, the trusted graph neural network model is used to make a decision.

4. A decision reliability assessment device based on a reliable graph neural network, characterized in that, The device includes: The data acquisition module is used to acquire an initial graph neural network model and a test set; the graph neural network model is a decision model; the test set includes decision samples and decision results. The expected calibration error calculation module is used to obtain the confidence level of the initial graph neural network model based on the logit vector output by the initial graph neural network model, use the initial graph neural network model to make decisions on decision samples, and obtain the accuracy of the decision results; and calculate the expected calibration error of the initial graph neural network model based on the confidence level and the accuracy of the decision results, including: Based on the confidence level and the accuracy of the decision result, the expected calibration error of the initial graph neural network model is calculated as follows: ,in, Labels indicating decision outcomes This represents the true labels in the test set. This represents the predicted distribution and the confidence level for the network to make predictive decisions. This represents the target distribution, which is the true probability that the network will make the correct decision. The decision evaluation module is used to determine whether the initial graph neural network model needs calibration based on the expected calibration error magnitude, construct a graph calibration loss function for the initial graph neural network model based on the determination result, train the initial graph neural network model using the graph calibration loss function to obtain a reliable graph neural network model, and perform a decision evaluation based on the calibration error of the reliable graph neural network model and the reliable graph neural network model, including: Based on the expected calibration error, it is determined whether the initial graph neural network model needs to be calibrated. If the expected calibration error is greater than a preset error threshold, it is determined that the initial graph neural network model needs to be calibrated, and a graph calibration loss function for the initial graph neural network model is constructed. The process of constructing the graph calibration loss function for the initial graph neural network model includes: Obtain the target distribution and the prediction distribution; the target distribution is the true probability that the network makes a correct decision; the prediction distribution is the confidence level of the initial graph neural network model in making a prediction decision; The KL divergence is obtained by calculating the divergence between the target distribution and the predicted distribution; A regularizer is constructed based on the product of the entropy of the target distribution and the predicted distribution. The regularization of the KL divergence and the regularization of the regularizer are calculated to obtain the graph calibration loss function.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 3.

Images