Learning data creation apparatus

Abstract

The objective of this invention is to efficiently process learning data used to create a learning model. Known design variables and their response variables are read in. A Gaussian process regression model is used to learn the relationship between the design variables and the response variables. For the obtained Gaussian process regression model, an acquisition function EI is obtained, and its maximum value P is obtained. Random sampling of the design variables is performed, and the maximum value P' of the acquisition function EI based on the sampled design variables is obtained. The maximum value P and the maximum value P' are compared. If the maximum value P' is greater than the maximum value P, an acquisition function PS is obtained, which includes the product of the mean and standard deviation obtained from the Gaussian process regression model. A predetermined number of design variables with the highest ranking are extracted from the obtained acquisition function PS. If the maximum value P' is less than the maximum value P, a predetermined number of design variables with the highest ranking are extracted from the acquisition function EI.

Description

Technical Field

[0001] This invention relates to a learning data creation apparatus for creating a model that predicts the actions of a vehicle. Background Technology

[0002] Various computer-based methods have been proposed in vehicle design, including evaluation methods using various models.

[0003] Non-Patent Document 1 describes an automatic evaluation system for bench testing of driving performance, which is considered the driving performance of a vehicle. In this example, a system using Bayesian Active Learning to search the design space was constructed. Furthermore, the constructed system describes a method for efficiently searching the executable region with fewer evaluations by using the expected improvement value (EI) as an acquisition function for minimum value search and using the probability of feasibility (PoF) and entropy search for the feasible region (ESF) as acquisition functions for feasible solution search.

[0004] Non-Patent Literature 1: "Automatic Evaluation Method for Vehicle Driving Performance Using Machine Learning," Proceedings of the Japan Society of Mechanical Engineers, Vol. 88, No. 915, pp. 22-00219 Summary of the Invention In this context, various methods have been proposed to improve the accuracy of learning models when creating learning data, targeting general uniform random sampling methods such as Latin hypercube. Bayesian active learning, represented by non-patent literature 1, has been applied, but in sampling based on existing acquisition functions, it is sometimes difficult to construct models with high accuracy in learning functions with multimodal characteristics.

[0005] The learning data creation device involved in this invention is a learning data creation device for creating a model to predict the actions of a vehicle. The learning data creation device performs the following processing: reading in known design variables and their response variables; performing a Gaussian process regression model learning operation on the relationship between the design variables and the response variables; obtaining an acquisition function EI related to the amount improved through additional learning for the obtained Gaussian process regression model, and obtaining the maximum value P of the acquisition function; performing random sampling of the design variables; Obtain the maximum value P' of the acquisition function EI based on the sampled design variables; compare the maximum value P with the maximum value P'; if the maximum value P' is greater than the maximum value P, obtain the acquisition function PS which includes the product of the mean and standard deviation obtained from the Gaussian process regression model, and extract a specified number of design variables with the highest ranking from the obtained acquisition function PS; and if the maximum value P' is less than the maximum value P, extract a specified number of design variables with the highest ranking from the acquisition function EI.

[0006] The preferred function to obtain PS is a0(x) = μ *α (x)×σ *β (x), μ * σ * The mean and standard deviation are calculated and normalized according to the Gaussian process regression model, and 0 < μ. * <1, 0<σ * <1, where α and β are scaling factors, and 0 ≤ α and β.

[0007] Furthermore, the learning data creation apparatus involved in this invention is preferably as follows: the learning data creation apparatus is a learning data creation apparatus for creating a model to predict the actions of a vehicle, which performs the following processing: reading in known design variables and response variables; performing a Gaussian process regression model learning operation on the relationship between the design variables and the response variables; obtaining the acquisition function PS for the obtained Gaussian process regression model, and extracting a predetermined number of design variables with the highest ranking from the obtained acquisition function PS; the acquisition function PS is a0(x) = μ *α (x)×σ *β (x), μ * σ * The mean and standard deviation are calculated and normalized according to the Gaussian process regression model, and 0 < μ. * <1, 0<σ * <1, where α and β are scaling factors, and 0 ≤ α and β.

[0008] Invention Effects According to the learning data creation apparatus of the present invention, for learning responses with multimodal characteristics, sampling near each peak can reduce prediction error; it can search for regions where vehicle specification inputs increase; and it can identify multiple peaks that cannot be obtained through optimization alone and include them in the machine learning data. Therefore, for example, it can create high-precision AI (artificial intelligence) models for road surface inputs, enabling efficient strength and reliability studies, etc. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the processing steps of the learning data creation apparatus involved in this embodiment.

[0010] Figure 2 It is a graph that shows the magnitude of the response to a factor (design variable) together with the prediction error.

[0011] Figure 3 It is a flowchart representing the overall process of generating learning data.

[0012] Figure 4 It is a graph representing the magnitude of the evaluation function f for design variables x1 and x2.

[0013] Figure 5 This is a graph showing the frequency of occurrence of random sampling and sampling in this embodiment relative to the response level.

[0014] Figure 6 It is a graph that compares the accuracy of learning models based on the relationship between response level and predicted value. Detailed Implementation

[0015] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. Furthermore, the following embodiments do not limit the present invention, and structures selectively combined from multiple examples are also included in the present invention.

[0016] The learning data creation apparatus described in this embodiment is composed of a computer. By inputting prescribed input data and executing a pre-prepared program, it can perform various steps to obtain learning data.

[0017] Figure 1 This is a flowchart illustrating the processing steps of the learning data creation apparatus involved in this embodiment.

[0018] To properly evaluate the driving performance of a vehicle as shown in Non-Patent Document 1, the more evaluations based on test results, the better. However, due to limited development time, there is a desire to obtain a suitable optimal solution with as few evaluations as possible. Therefore, the selection of evaluation data is important, and this learning device can effectively perform the selection of evaluation data.

[0019] First, the pre-prepared data for the design variable x and response variable y are read in (S1). This data is obtained through vehicle design variables and driving performance, such as data on the vehicle's response (changes) to inputs from the road surface, obtained through tests such as driving.

[0020] The design variable x and the response variable y have a defined relationship, and the evaluation function is y=f(x).

[0021] Therefore, a learning operation (S2) is performed on the Gaussian process regression model for the evaluation function f.

[0022] Based on this learning, we perform an a-valued function EI for the design variable x.EI The evaluation of (x) is performed, and its maximum value P(S3) is obtained.

[0023] Here, we obtain the a of the function EI: Expected Improvement (EI). EI (x) relates to the amount improved through supplementary learning and can be represented as follows.

[0024] a EI (x) = (μ(x) - f(x)) + ))Φ(Z)+σ(x)φ(Z) Z = (μ(x) - f(x)) + )) / σ(x) Here, μ and σ are the mean and standard deviation obtained from the Gaussian process regression model, Φ and φ are the cumulative frequency distribution and probability density function of the normal distribution, and x + The design variable x is the evaluation function that maximizes P in the known response.

[0025] Next, x' (S4) is generated by random sampling within a specified range of design variables. Sampling is performed on the design variables read from S1.

[0026] Perform the retrieval of function EI for x' with respect to a' EI The evaluation of (x) is performed, and its maximum value P' (S5) is obtained.

[0027] Then, compare the two obtained maximum values ​​P and P' to determine whether P < P' (S6).

[0028] In the S6 decision, if it is "no", then a0 for the design variable x' of random sampling is calculated according to the newly constructed acquisition function PS (S7).

[0029] Here, the newly constructed acquisition function PS can be represented as follows.

[0030] a0(x) = μ *α (x)×σ *β (x) Here, μ* and σ* are the mean and standard deviation calculated and normalized according to the Gaussian process regression model, and 0 < μ. * <1, 0<σ * <1, where α and β are scaling factors, and 0 ≤ α and β.

[0031] Then, extract the first N corresponding x's of a0 based on the new acquisition function PS, and output them as design variables selected by filtering (S8).

[0032] On the other hand, if S6 determines "yes", then extract a' based on the acquisition function EI. EI The first N corresponding design variables x are output as the design variables x selected through screening (S9).

[0033] Thus, if the judgment of S6 is "yes", by obtaining the effect of function EI, the output a' with the higher probability of updating the maximum value is determined. EI If the first N corresponding design variables x are determined as “no” in S6, since the maximum value cannot be expected to be updated, the function PS is used to output the first N corresponding design variables x' located in the region of large response and large prediction error of a0.

[0034] Figure 2 This graph shows the magnitude of the response to a factor (design variable) along with the prediction error. Point B has a large prediction error, but the a0 of the acquisition function PS is relatively small. Point A has a smaller prediction error compared to point B, but the a0 of the acquisition function PS is relatively large. Therefore, by using the new acquisition function PS, it is possible to evaluate a large response and easily sample near the peak.

[0035] Thus, the learning apparatus according to this embodiment promotes sampling near the peaks, enabling sampling near each peak to reduce prediction errors for learning responses with multimodal characteristics.

[0036] Figure 3 This is a flowchart illustrating the overall process of generating learning data. In this example, sampling continues until the specified number of iterations is reached, after which the learning data is output to a learning file.

[0037] First, the search criteria file is read from the search criteria file (S11). Next, the search criteria are set (S12). Then, random sampling of the design variables is performed using the Design of Experiments (DOE) method (S13).

[0038] Next, the response variable y for the design variable x is obtained from the computational model or experimental setup (S14). S11 to S14 correspond to... Figure 1 S1 in the middle.

[0039] Thus, given the specified design variable x and response variable y, the process is carried out in conjunction with... Figure 1 The same process as S2 to S9 is applied, and as a result, a specified number of design variables x or x' are obtained.

[0040] Then, having extracted the design variables x and x', the response variable y' for the extracted x' is obtained from the calculation model or experimental setup (S15). Furthermore, since the response variable y for the design variable x was input in S14, it is sufficient to maintain the response variable y output in S9.

[0041] Next, it is determined whether the number of times S2 to S9 has been performed is more than the predetermined number (S16).

[0042] If the decision in S15 is "No", then the design variable x and response variable y, which are the processing objects in S2, are updated to the update performed in S15 (S17), and the process returns to S2.

[0043] Then, if the decision in S15 is "yes", the design variable x and response variable y obtained in S15 are output as a learning data file (S18). The process ends.

[0044] Thus, in this embodiment, the acquisition functions EI and PS are adaptively used to sample the learning data.

[0045] Figure 4 This is a graph representing the magnitude of the evaluation function f for design variables x1 and x2. Thus, by capturing the effect of function EI, we can capture the maximum value at the center; by capturing the effect of function PS, we can capture the region with the largest response.

[0046] Figure 5 This is a graph showing the frequency of occurrence of random sampling and sampling according to this embodiment relative to the response level. White represents the overlapping area of ​​the two, the diagonal line represents the area where random sampling is dominant, and gray represents the area where this embodiment is dominant. Thus, based on the sampling of this embodiment, it can be seen that the proportion of samples with large responses is increased.

[0047] Figure 6 This is a graph comparing the accuracy of the learning model based on the relationship between the response level and the predicted value. Thus, through this implementation, the accuracy of the learning model (AI model) is improved compared to random sampling in regions with large response levels.

[0048] As described above, the learning data creation apparatus according to this embodiment can achieve the following effects.

[0049] (i) In the study of strength reliability in platform development, it is required to quickly predict inputs from the road surface for various vehicle specifications. Therefore, simulations modeled by agents through machine learning are used, but there are issues with prediction accuracy. This is because the inputs from the road surface vary non-linearly with respect to vehicle specifications, and it is necessary to include the sometimes large inputs in the machine learning data.

[0050] According to this embodiment, sampling can be performed to create a proxy model capable of predicting road surface inputs with high accuracy.

[0051] In particular, it can search for regions where vehicle specification inputs increase, and it can identify multiple peaks that cannot be obtained through optimization alone and include them in the machine learning data.

[0052] (ii) In the development of performance such as strength and reliability, a high-precision AI model for road surface input is created, which enables efficient strength and reliability research.

[0053] (iii) It can perform sampling near each peak to reduce prediction error for learning responses with multimodal characteristics, can search for regions that increase with vehicle specification input, and can identify multiple peaks that cannot be obtained through optimization alone and include them in the machine learning data. Therefore, it can improve the accuracy of the learning model for the region of interest.

[0054] (iV) can help improve the performance of the developed vehicle and shorten the development period.

Claims

1. A learning data creation apparatus for creating a model to predict the actions of a vehicle, the learning data creation apparatus being characterized by performing the following processing: Read in the known design variables and their response variables; The Gaussian process regression model was used to learn and calculate the relationship between the design variables and the response variables. For the obtained Gaussian process regression model, the acquisition function EI related to the amount improved through additional learning is obtained, and the maximum value P of the acquisition function is obtained; Perform random sampling of design variables; Obtain the maximum value P' of the acquisition function EI based on the sampled design variables; Compare the maximum value P with the maximum value P'; When the maximum value P' is greater than the maximum value P, obtain the acquisition function PS, which contains the product of the mean and standard deviation obtained from the Gaussian process regression model, and extract a specified number of design variables with the highest ranking from the obtained acquisition function PS; and If the maximum value P' is less than the maximum value P, extract a specified number of design variables with the highest ranking from the acquisition function EI.

2. The learning data creation apparatus according to claim 1, characterized in that, The function to obtain PS is as follows: a0(x)=μ *α (x)×σ *β (x), μ * σ * The mean and standard deviation are calculated and normalized according to the Gaussian process regression model, and 0 < μ. * <1, 0<σ * <1, where α and β are scaling factors, and 0 ≤ α and β.

3. A learning data creation apparatus for creating a model to predict the actions of a vehicle, the learning data creation apparatus being characterized by performing the following processing: Read in the known design variables and response variables; Learning operations were performed using a Gaussian process regression model to investigate the relationship between design and response variables; and For the obtained Gaussian process regression model, the acquisition function PS is obtained, and a specified number of design variables with the highest ranking are extracted from the acquired function PS. The function to obtain PS is as follows: a0(x)=μ *α (x)×σ *β (x), μ * σ * The mean and standard deviation are calculated and normalized according to the Gaussian process regression model, and 0 < μ. * <1, 0<σ * <1, where α and β are scaling factors, and 0 ≤ α and β.

Images