Hyperparameter optimization method and apparatus, computing device

Abstract

The embodiment of the application provides a kind of hyperparameter optimization method and device, computing device, which comprises: determining the first hyperparameter instance corresponding to the current hyperparameter of data processing system;Determine the first effect evaluation information corresponding to the first hyperparameter instance based on the first training result of the first hyperparameter instance in data processing system;Obtain a plurality of historical hyperparameter instances and a plurality of historical effect evaluation information corresponding to the historical hyperparameter instance respectively;According to the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to the plurality of historical hyperparameter instances respectively, select the target hyperparameter instance with the highest effect evaluation from the first hyperparameter instance and the plurality of historical hyperparameter instances.The embodiment of the application improves the hyperparameter optimization efficiency.

Description

Technical Field

[0001] This application relates to the field of electronic equipment technology, and in particular to a hyperparameter optimization method and apparatus, and a computing device. Background Technology

[0002] Hyperparameters are parameters set before a data processing model begins its learning process, rather than the parameter data obtained through training. Examples include the number of hidden layers in a deep network or the learning rate in a machine learning model. A data processing model's hyperparameters typically contain multiple sub-hyperparameters. Once the parameter values ​​of these sub-hyperparameters are set, they constitute a parameter instance of the hyperparameter. The selection of this parameter instance has a wide-ranging impact on the learning performance of the data processing model. Generally, selecting a set of hyperparameter instances that yields the best training results can improve the model's performance and effectiveness.

[0003] In existing technologies, hyperparameter instances are typically determined manually by the trainer based on their experience using the model before training the data processing model. After obtaining the training results, the hyperparameter instances are adjusted using these results to optimize the training effect of the data processing model. However, manually adjusting the hyperparameter instances is inefficient and lacks precision. Summary of the Invention

[0004] In view of this, embodiments of this application provide a hyperparameter optimization method, apparatus, and computing device to solve the technical problem of low efficiency in manually adjusting hyperparameters of data processing models in the prior art.

[0005] In a first aspect, embodiments of this application provide a hyperparameter optimization method, including:

[0006] Determine the first hyperparameter instance currently corresponding to the hyperparameters of the data processing system;

[0007] Based on the first training result of the first hyperparameter instance in the data processing system, determine the first effect evaluation information corresponding to the first hyperparameter instance;

[0008] Obtain multiple historical hyperparameter instances and their corresponding historical performance evaluation information.

[0009] Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances, the target hyperparameter instance with the highest effect evaluation is selected from the first hyperparameter instance and the multiple historical hyperparameter instances.

[0010] Secondly, embodiments of this application provide a hyperparameter optimization device, comprising:

[0011] The instance determination module is used to determine the first hyperparameter instance currently corresponding to the hyperparameters of the data processing system.

[0012] The effect evaluation module is used to determine the first effect evaluation information corresponding to the first hyperparameter instance based on the first training result of the first hyperparameter instance in the data processing system.

[0013] The history acquisition module is used to acquire multiple historical hyperparameter instances and the historical performance evaluation information corresponding to each of the multiple historical hyperparameter instances.

[0014] The instance selection module is used to select the target hyperparameter instance with the highest effect evaluation from the first hyperparameter instance and the multiple historical hyperparameter instances, based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to the multiple historical hyperparameter instances respectively.

[0015] Thirdly, embodiments of this application provide a computing device, including: a storage component and a processing component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked by the processing component; the processing component is used for:

[0016] Determine the first hyperparameter instance corresponding to the current hyperparameter of the data processing system; based on the first training result of the first hyperparameter instance in the data processing system, determine the first performance evaluation information corresponding to the first hyperparameter instance; obtain multiple historical hyperparameter instances and the historical performance evaluation information corresponding to the multiple historical hyperparameter instances respectively; according to the first performance evaluation information corresponding to the first hyperparameter instance and the historical performance evaluation information corresponding to the multiple historical hyperparameter instances respectively, select the target hyperparameter instance with the highest performance evaluation from the first hyperparameter instance and the multiple historical hyperparameter instances.

[0017] In this embodiment, a first hyperparameter instance corresponding to the current hyperparameters of the data processing system is determined. This hyperparameter instance is the latest hyperparameter instance of the data processing model. Based on the first training result of the first hyperparameter instance in the data processing system, first performance evaluation information corresponding to the first hyperparameter instance is determined. The first performance evaluation information can be the evaluation content generated by the first hyperparameter instance on the training effect of the data processing system. The training effect generated by the first hyperparameter instance on the data processing system can be determined through the first performance evaluation information. Then, based on the first performance evaluation information corresponding to the first hyperparameter instance and the performance evaluation information corresponding to multiple historical hyperparameter instances, a target hyperparameter instance with the highest training performance evaluation is selected from the first hyperparameter instance and multiple historical hyperparameter instances. By training the model on the latest first hyperparameter instance of the data processing system, the performance evaluation of the hyperparameter instance on the data processing system in real time can be automatically detected. By filtering with the performance evaluation information of historical hyperparameter instances, the target hyperparameter instance with the highest training performance evaluation is obtained, realizing the automated acquisition of the target hyperparameter instance with the highest performance evaluation, improving training efficiency and accuracy. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating one embodiment of a hyperparameter optimization method provided in this application;

[0020] Figure 2 A flowchart illustrating yet another embodiment of a hyperparameter optimization method provided in this application;

[0021] Figure 3a , 3b Example diagram of a causal network provided in the embodiments of this application;

[0022] Figure 4 A flowchart illustrating yet another embodiment of a hyperparameter optimization method provided in this application;

[0023] Figure 5 A flowchart illustrating yet another embodiment of a hyperparameter optimization method provided in this application;

[0024] Figure 6 This application example diagram illustrates a hyperparameter optimization method provided in this embodiment.

[0025] Figure 7 A schematic diagram of one embodiment of a hyperparameter optimization device provided in this application;

[0026] Figure 8 This is a schematic diagram of one embodiment of a computing device provided in this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the application. The singular forms “a,” “said,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. “Multiple” generally includes at least two, but does not exclude the inclusion of at least one.

[0029] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0030] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to identification.” Similarly, depending on the context, the phrases “if determination” or “if identification (of the condition or event of the statement)” can be interpreted as “when determination” or “in response to determination” or “when identification (of the condition or event of the statement)” or “in response to identification (of the condition or event of the statement).”

[0031] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a product or system comprising a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a product or system. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the product or system that includes said element.

[0032] The technical solution of this application embodiment can be applied to the hyperparameter optimization process of data processing model. By evaluating the effect of hyperparameters, the hyperparameter instance with the highest effect evaluation is selected, and the reason for the selection of the target hyperparameter instance is explained to obtain the corresponding explanation prompt information, thereby improving the selection efficiency of hyperparameters and clarifying the selection reason.

[0033] In existing technologies, hyperparameters are parameters determined before training the data processing model, such as the model depth of a machine learning model. Typically, hyperparameters are determined manually by the user of the data processing model based on their experience. If the training performance of the data processing model is poor, the hyperparameters can be manually adjusted until a satisfactory training effect is obtained. However, manually adjusting the hyperparameters of the data processing model results in low training efficiency and low accuracy.

[0034] In this embodiment, a first hyperparameter instance corresponding to the current hyperparameters of the data processing system is determined. This hyperparameter instance is the latest hyperparameter instance of the data processing model. Based on the first training result of the first hyperparameter instance in the data processing system, first performance evaluation information corresponding to the first hyperparameter instance is determined. The first performance evaluation information can be the evaluation content generated by the first hyperparameter instance on the training effect of the data processing system. The training effect generated by the first hyperparameter instance on the data processing system can be determined through the first performance evaluation information. Then, based on the first performance evaluation information corresponding to the first hyperparameter instance and the performance evaluation information corresponding to multiple historical hyperparameter instances, the target hyperparameter instance with the highest training performance evaluation is selected from the first hyperparameter instance and multiple historical hyperparameter instances. By training the model on the latest first hyperparameter instance of the data processing system, the performance evaluation of the real-time hyperparameter instance on the data processing system can be automatically detected. By filtering with the performance evaluation information of historical hyperparameter instances, the target hyperparameter instance with the highest training performance evaluation is obtained, realizing the automated acquisition of the target hyperparameter instance with the highest performance evaluation, improving training efficiency and accuracy.

[0035] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0036] like Figure 1 The diagram shown is a flowchart of an embodiment of a hyperparameter optimization method provided in this application. The method may include the following steps:

[0037] 101: Determine the first hyperparameter instance currently corresponding to the hyperparameters of the data processing system.

[0038] The hyperparameter optimization method provided in this application can be applied to computing devices, such as computers, servers, cloud servers, supercomputers, laptops, tablets, etc. This application does not impose too many limitations on the specific type of computing device.

[0039] Optionally, hyperparameters can be parameters set before learning begins, rather than parameters obtained through training; that is, they can be numerous parameters not involved in the actual training process. For example, network depth, number of iterations, and number of neurons per layer in a machine learning model can all be hyperparameters; the damage dealt to a character during an attack in a game can be a hyperparameter; the number of search terms used in a word search in e-commerce can also be a hyperparameter; and the time step and feature dimensions involved in the market in the financial field are also hyperparameters.

[0040] A data processing system may include a data processing model that requires hyperparameter optimization. Furthermore, the system may include contextual features that influence the determination of hyperparameters. For example, in the financial field, contextual features such as macroeconomic factors, microeconomic factors, and unexpected events can affect the selection of hyperparameters.

[0041] A data processing system may include machine learning algorithms or various mathematical algorithms that require data computation. The data processing model corresponding to a machine learning algorithm can be a machine learning model. The data processing model corresponding to a mathematical algorithm that requires data computation can be a mathematical computation model constructed according to that algorithm.

[0042] Machine learning algorithms, or mathematical algorithms, are mathematical concepts that can be described using mathematics and pseudocode. In machine learning, "algorithms" can run on data to create corresponding machine learning models and train their parameters. Mathematical algorithms can run on data to construct corresponding mathematical computational models and train their parameters.

[0043] A hyperparameter of a data processing system may include multiple sub-hyperparameters. That is, the hyperparameter in the embodiments of this application can actually be composed of multiple sub-hyperparameters. When the parameter data of each sub-hyperparameter is determined, it constitutes a hyperparameter instance of the hyperparameter.

[0044] The hyperparameter optimization method provided in this application can be applied to online optimization scenarios, particularly for data processing systems that require online training, such as search term recommendation scenarios or game scenarios. In online training scenarios, interactive functionality can be provided to users. Users can trigger optimization requests, thereby initiating hyperparameter optimization. At this time, the computing device can determine the first hyperparameter instance corresponding to the hyperparameters of the data processing system. The first hyperparameter instance can be the most recently obtained hyperparameter instance of the data processing model.

[0045] The hyperparameter optimization method provided in this application embodiment can be applied to offline optimization scenarios. In offline optimization scenarios, the first hyperparameter instance can be the last hyperparameter instance obtained in the offline optimization scenario.

[0046] 102: Based on the first training result of the first hyperparameter instance in the data processing system, determine the first performance evaluation information corresponding to the hyperparameter instance.

[0047] The training process of the first hyperparameter instance in the data processing system can specifically include: constructing a data processing model corresponding to the data processing system based on the first hyperparameter instance, training the model with training data to obtain the model parameters of the data processing model, and testing the model with test data to obtain the training results of the data processing model corresponding to the trained model parameters, thus obtaining the first training result. Of course, in order to improve the accuracy of training, multiple sets of training data can be used to train the model parameters of the data processing model multiple times to obtain multiple training results, and the first training result can be obtained by using the multiple training results.

[0048] The first training result can be a quantitative evaluation index of the model's training effect during the model training process. For example, it can be accuracy, precision, recall, ROC (Receiver Operating Characteristic) curve and / or AUC (Area Under Curve).

[0049] 103: Obtain multiple historical hyperparameter instances and their corresponding historical performance evaluation information.

[0050] Historical hyperparameter instances are hyperparameter instances for which performance estimation has been completed. The historical performance evaluation information for any historical hyperparameter instance can be determined from the first training result of that historical hyperparameter instance in the data processing system. The performance evaluation information for any hyperparameter instance can be obtained based on the training result of that hyperparameter instance in the data processing system, and the performance evaluation information for any two hyperparameter instances is the same.

[0051] 104: Based on the first performance evaluation information corresponding to the first hyperparameter instance and the historical performance evaluation information corresponding to multiple historical hyperparameter instances, select the target hyperparameter instance with the highest training performance evaluation from the first hyperparameter instance and multiple historical hyperparameter instances.

[0052] Optionally, the performance evaluation information can be an evaluation metric used to evaluate the training performance of hyperparameter instances in the data processing model. The performance evaluation information may actually include one or more evaluation metrics, such as performance information and / or effect information. Performance information may include data matching values, and effect information may include effect scores.

[0053] Specifically, selecting the target hyperparameter instance with the highest training effect evaluation from the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances can include: sorting the first effect evaluation information of the first hyperparameter instance and the historical effect evaluation information corresponding to the multiple historical hyperparameter instances in descending order of information content, and selecting the hyperparameter instance with the highest sorting order as the target hyperparameter instance. For example, when the effect evaluation information is determined by the effect score, the first effect score of the first hyperparameter instance and the historical effect scores corresponding to the multiple historical hyperparameter instances can be sorted in descending order, and the hyperparameter instance with the highest effect score can be selected as the target hyperparameter instance.

[0054] In this embodiment, a first hyperparameter instance corresponding to the current hyperparameters of the data processing system is determined. This hyperparameter instance is the latest hyperparameter instance of the data processing model. Based on the first training result of the first hyperparameter instance in the data processing system, first performance evaluation information corresponding to the first hyperparameter instance is determined. The first performance evaluation information can be the training effect produced by the first hyperparameter instance on the data processing system. The training effect produced by the first hyperparameter instance on the data processing system can be determined through the first performance evaluation parameter. Then, based on the first performance evaluation information corresponding to the first hyperparameter instance and the performance evaluation information corresponding to multiple historical hyperparameter instances, a target hyperparameter instance with the highest training performance evaluation is selected from the first hyperparameter instance and multiple historical hyperparameter instances. By training the model on the latest first hyperparameter instance of the data processing system, the performance evaluation of the real-time hyperparameter instance on the data processing system can be automatically detected. By filtering with the performance evaluation information of historical hyperparameter instances, the target hyperparameter instance with the highest training performance evaluation is obtained, realizing the automated acquisition of the target hyperparameter instance with the highest performance evaluation, improving training efficiency and accuracy.

[0055] The first hyperparameter instance is actually used to construct the data processing model of the data processing system; therefore, it has a significant impact on the composition of the data processing model. The training objective of the data processing model is a prerequisite for the generation of the training results. The first hyperparameter instance can be considered as the dependent variable, and the training objective as the effect variable. This allows us to measure the causal relationship between the first hyperparameter instance and the data processing model, and to determine the actual impact of the first hyperparameter instance on the construction of the data processing model through this causal relationship, thereby obtaining accurate performance rating information.

[0056] refer to Figure 2 ,and Figure 1 The difference in the illustrated embodiment is that step 102: determining the first performance evaluation information corresponding to the first hyperparameter instance based on the first training result of the first hyperparameter instance in the data processing system may include:

[0057] 201: Determine the training objective of the data processing system and the first training result generated by the first hyperparameter instance during the training of the data processing system.

[0058] The first hyperparameter instance generates the first training result during the training of the data processing system.

[0059] 202: Based on the first training result, determine the first causal relationship between the first hyperparameter instance and the training target.

[0060] The first hyperparameter instance can actually be applied to the generation process of the data processing model corresponding to the data processing system, and has a corresponding impact on the establishment of the data processing model. Since the first hyperparameter instance can have a corresponding impact on the establishment of the data processing model, when the training result of the data processing model is taken as the result, the first hyperparameter instance can be a cause of the training result, and a causal relationship is formed between the two.

[0061] Causal relationships can include: dependent variables composed of hyperparameter instances and causal variables composed of training objectives. When the dependent variable is determined, the effect variable can also be determined. That is, when the dependent variable composed of the first hyperparameter instance is determined, the data processing system can generate the first training result under the constraint of the effect variable composed of the training objective.

[0062] Causal relationships are mainly used to represent the association between dependent and effect variables. They can exist in various forms. For example, they can be expressed by clearly defining the dependent and effect variables and establishing a functional mapping between them. They can also be expressed by using the dependent and effect variables as network nodes and forming a directed acyclic causal network based on the connection relationship between the causal variables.

[0063] 203: Based on the first causal relationship, determine the first effect evaluation information corresponding to the first hyperparameter instance.

[0064] Typically, in this embodiment, the first causal relationship can be used to represent the association between the dependent variable, which is composed of the first hyperparameter instance, and the effect variable, which is composed of the training objective of the data processing system. The first causal relationship can determine the influence trend of the dependent variable on the effect variable, which is also the generated first effect evaluation information. The first effect evaluation information can evaluate the performance of the first hyperparameter instance on the first training result of the data processing system. For example, when the first hyperparameter instance is X1 and the first training result is Y1, the first effect evaluation learning can be an effect evaluation of generating Y1 from X1.

[0065] In this embodiment, the first performance evaluation information corresponding to the first hyperparameter instance can be determined through the first training result of the first hyperparameter instance in the data processing system. The first training result can determine the first causal relationship between the first hyperparameter instance and the training objective. This first causal relationship can be the association between the dependent variable formed by the first hyperparameter instance and the effect variable formed by the training objective. This first causal relationship can be used to measure the performance evaluation of the first hyperparameter instance in producing the first training result under the constraint of the training objective of the data processing system, thus obtaining the first performance evaluation information. This performance evaluation information can then be used to determine the impact of the first hyperparameter instance on the training process of the data processing system, accurately measure the training effect of each hyperparameter instance, and obtain accurate performance evaluation results.

[0066] In practical applications, the first causal relationship can be used to determine the impact of the first hyperparameter instance on the data processing model in two aspects: the degree to which the first hyperparameter instance characterizes the data processing model itself, and the training effect of the first hyperparameter instance on the data processing model.

[0067] In some embodiments, step 203: determining the first effect evaluation information corresponding to the first hyperparameter instance based on the first causal relationship may specifically include:

[0068] The first effect information is determined based on the matching degree between the first causal relationship and the first hyperparameter instance;

[0069] Based on the first causal relationship, estimate the first evaluation information corresponding to the generation of the first training result of the first hyperparameter instance;

[0070] Based on the first effect information and / or the first evaluation information, determine the first effect evaluation information corresponding to the first hyperparameter instance.

[0071] In this embodiment, by analyzing the causal relationship between the first hyperparameter instance and the data processing model, the degree to which the first hyperparameter instance characterizes the data processing model can be determined to obtain first effect information. Furthermore, the degree of influence of the first hyperparameter instance on the first training result generated by the data processing model can be determined to obtain first evaluation information. Using the first effect information and / or the first evaluation information, first effect evaluation information is determined. The influence of the first hyperparameter instance on the data processing model is quantified from two aspects to obtain accurate first effect evaluation information.

[0072] For ease of understanding, let's take a causal relationship expressed as a causal network as an example. Assume the hyperparameters of the data processing model consist of 11 sub-parameters: A, B, C, D, E, F, G, H, I, J, and K. Each sub-parameter is used as the dependent variable (301), and the training objective of the data processing system is used as the effect variable (Regret) (302). The specific causal network can be found in [reference needed]. Figure 3a Causal networks can reveal the causal relationship between dependent and effect variables. (Reference) Figure 3a It can be determined directly. Figure 3a In the causal network, dependent variables B, D, H, F, A, and G directly point to the result variable Regret302, and are therefore direct factors influencing the value of Regret. Other dependent variables C, E, I, and J do not directly point to Regret302, and are therefore indirect factors influencing Regret. Variable K is not present in the causal network, therefore it has no effect on the value of Regret. These direct, indirect, and ineffective factors influencing the result variable can provide explanatory information, explaining the various factors that produce the training results.

[0073] Taking a causal relationship expressed as a causal network as an example, let's assume that the hyperparameters of a deep learning algorithm can include two parameters: one is the batch size (the size of the dataset used for each training iteration), with a value range of [10, 1000], and the other is the learning rate n, with a value range of [0.001, 0.1]. In this embodiment, we aim to gradually adjust the combination of batch size 303 and n 304 to maximize the training objective y. Assuming that the training objective is achieved by training the model with these hyperparameters, the resulting causal network can be referenced... Figure 3b .

[0074] exist Figure 3b In this context, batchsize is a factor of n, and both batchsize and n are factors of AUC.

[0075] In the causal relationship between batchsize and n, assuming n is generated by batchsize, batchsize is the cause of n, and n is the effect of batchsize (i.e., batchsize is the dependent variable, and n is the effect variable), the fitting function between the dependent variable formed by batchsize and the effect variable formed by n is: n = f(batchsize, e), where e is a noise variable. However, assuming batchsize is generated by n, n is the cause of batchsize, and batchsize is the effect of n (i.e., n is the dependent variable, and batchsize is the dependent variable), the fitting function between the dependent variable formed by n and the dependent variable formed by batchsize is: batchsize = g(n, e). Since e and n are not independent (batchsize = f⁻¹(n) - e), there is no reverse causal relationship. The corresponding network model can be obtained through function fitting.

[0076] In the causal relationship between batch size, n, and AUC, suppose we obtain the following result: when batch size is set to a large value, such as 10, then n must be set to a low level (0.001). Therefore, the setting of n changes significantly with batch size. Conversely, when n is set to a small value, such as n = 0.001, then setting batch size = 10 or batch size = 1000 will result in a high AUC. Therefore, batch size does not change significantly with the setting of n. From the above settings, we can conclude that batch size is the dependent variable, n is the effect variable, and there is a causal relationship between them. This causal relationship can form an explanation of the effect. For example, an explanation of the effect that can be displayed to the user could be: batch size is the dependent variable, n is the effect variable, and there is a causal relationship between them.

[0077] In practical applications, to improve the describability of causal relationships, they can actually exist in the form of causal networks. A causal network can use the parameters in the hyperparameters and the training objective as network nodes, and is constructed by directed edges connecting the parameters themselves and the parameters with the training objective, using the influence or connection relationships between the hyperparameters and between the parameters and the training objective.

[0078] In one possible design, the first causal relationship may include a first causal network.

[0079] Optionally, determining the first effect information based on the matching degree between the first causal relationship and the first hyperparameter instance may include:

[0080] Based on the causal network discovery algorithm, the first causal model corresponding to the first causal network is determined.

[0081] Calculate the first data matching value between the first hyperparameter instance and the first causal model;

[0082] Based on the first data matching value, obtain the first effect information.

[0083] After the selection of the target hyperparameter instance is completed, the first hyperparameter instance becomes the historical hyperparameter instance. A relationship can be established between the first causal model and the first hyperparameter instance. By establishing a relationship between the first hyperparameter instance and the first causal model, the first causal model, which has become a historical hyperparameter instance, can be used to explain the causal relationship.

[0084] In some embodiments, determining the first causal model corresponding to the first causal network based on the causal network discovery algorithm may include: inputting the first causal network into the causal network discovery algorithm to obtain the first causal model.

[0085] refer to Figure 3a The causal network shown can be identified using a causal network discovery algorithm, where the first causal model can be regret = f(B,D,H,F,A,G). This causal network model can provide the highest evaluation of the instance effect given the target hyperparameter, thus obtaining a decision explanation.

[0086] In some other embodiments, hyperparameter instances and their corresponding causal models can be obtained through matching calculations. For example, the matching relationship between hyperparameter instances and causal models can be calculated using a log-likelihood algorithm on joint probabilities, or the CFI (Comparative Fit Index) calculation method can be used to calculate the matching relationship between hyperparameter instances and causal models.

[0087] In this embodiment, a first causal model is used to quantize a first causal network to enhance its expressiveness. The quantized first causal model is then used to calculate the matching degree with a first hyperparameter instance to obtain a corresponding first data matching value. This first data matching value accurately measures the training effect of the hyperparameter instance, providing accurate first effect information.

[0088] Optionally, based on the first causal relationship, the estimated first evaluation information corresponding to the generation of the first training result for the first hyperparameter instance includes:

[0089] Based on the causal effect estimation algorithm, the first effect model corresponding to the first causal network is determined;

[0090] Input the first hyperparameter instance into the first effect model to calculate the first effect score;

[0091] Based on the first effect score, obtain the first evaluation information.

[0092] After the selection of the target hyperparameter instance is completed, the first hyperparameter instance becomes the historical hyperparameter instance. A correlation can be established between the first effect model and the first hyperparameter instance. By establishing a correlation between the first hyperparameter instance and the first effect model, the first effect model that has become the historical hyperparameter instance can be used to estimate the effect.

[0093] Similarly Figure 3a Taking the causal network shown as an example, the causal effect estimation algorithm can determine the first effect model corresponding to the causal network. This effect model can be used to estimate the evaluation information of the results produced by the hyperparameter instances. The effect score obtained through effect model estimation can be used as evaluation information to directly evaluate the training effect of the hyperparameter instances.

[0094] In some embodiments, determining the first effect model corresponding to the first causal network based on the causal effect estimation algorithm may include: inputting the first causal network into the causal effect estimation algorithm to obtain the first effect model. The first effect model can be used to estimate the effect score that hyperparameter instances can obtain during model training, thereby obtaining accurate evaluation information.

[0095] In the process of finding the target hyperparameter instance with the highest training performance, to overcome the limitation of only providing the search results, the selection of the target hyperparameter instance with the highest performance can be explained to determine the reasons for the selection and clarify the rationale for choosing that target hyperparameter instance. The explanatory information for the selection of that target hyperparameter instance can be obtained from the first hyperparameter instance and multiple historical hyperparameter instances.

[0096] In practical applications, the explanatory information for selecting target hyperparameter instances can include various types of explanatory information. For example, it can include at least one of effect explanatory information, decision explanatory information, and path explanatory information.

[0097] As one embodiment, the explanation information for selecting the target hyperparameter instance may include path explanation information, i.e., the optimized path of the target hyperparameter instance. After selecting the target hyperparameter instance with the highest performance evaluation from the first hyperparameter instance and the historical performance evaluation information corresponding to the first hyperparameter instance and the multiple historical hyperparameter instances, the method may further include:

[0098] Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances, path explanation information corresponding to the selection process of the target hyperparameter instance is generated.

[0099] The path explanation information for hyperparameters consists of a sequence of hyperparameter instances, starting from the first hyperparameter instance and ending when the optimization reaches the target hyperparameter instance. In practical applications, however, changes in hyperparameter instances are usually slow, and the training results of adjacent hyperparameter instances may be similar. To simplify the optimization path of hyperparameter instances and make the path explanation information more explicit, path explanations can be provided for the inflection points in the instance effects during the hyperparameter optimization process, thereby improving explanation efficiency.

[0100] In one possible design, the path explanation information corresponding to the selection process of the target hyperparameter instance, based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances, may include:

[0101] Obtain the training order corresponding to multiple historical hyperparameter instances, and determine that the training order of the first hyperparameter instance is the last one;

[0102] Identify the interference identifiers corresponding to multiple historical hyperparameter instances;

[0103] Based on the training order of multiple historical hyperparameter instances and the first hyperparameter instance, hyperparameter instances with adjacent training order and the same interference label in the multiple historical hyperparameter instances and the first hyperparameter instance are grouped into the same group to obtain at least one group of hyperparameter instances.

[0104] Extract the first and last hyperparameter instances from at least one set of hyperparameter instances respectively, and combine them with the remaining hyperparameter instances that were not assigned to at least one set of hyperparameter instances to obtain multiple path hyperparameter instances for explaining the path optimization process;

[0105] Based on the training order corresponding to the multiple path hyperparameter instances, generate path explanation information corresponding to the selection process of the target hyperparameter instance.

[0106] An interference label can be associated with any hyperparameter instance. This label indicates whether the hyperparameter instance meets the training conditions, classifying it as either an interference point or a non-interference point. Generally, hyperparameter instances labeled as interference points perform poorly, while those labeled as non-interference points perform better.

[0107] The path interpretation information corresponding to the selection process of generating target hyperparameter instances, based on the training order corresponding to the multiple path hyperparameter instances, may specifically include: sorting the training order corresponding to the multiple path hyperparameter instances to obtain the sorted multiple path hyperparameter instances, and determining the path interpretation information constituted by the multiple path hyperparameter instances.

[0108] When extracting the first and last hyperparameter instances from at least one set of hyperparameter instances, and combining them with the remaining hyperparameter instances not assigned to at least one set of hyperparameter instances to obtain multiple path hyperparameter instances for explaining the path optimization process, the first and last hyperparameter instances from at least one set of hyperparameter instances, and the remaining hyperparameter instances not assigned to at least one set of hyperparameter instances are used as multiple path hyperparameter instances.

[0109] Similarly Figure 3a Taking the causal network shown as an example, the generated path explanation information can be multiple obtained path hyperparameter instances:

[0110] From (A=0.01, B=0.6, C=0.1, D=5, E=8, F=8, G=0.9, H=10, I=10, J=10, K=100);

[0111] Transition to (A=0.5, B=0.6, C=10, D=10, E=1, F=0.1, G=0.7, H=1, I=8, J=10, K=0.01);

[0112] Then, transitioning to (A = ..., K = ...), it finally converges to the instance with the highest performance evaluation:

[0113] (A=0.5, B=0.5, C=1, D=10, E=1, F=0.01, G=0.07, H=1, I=1, J=1, K=0.1).

[0114] As another embodiment, the explanatory information for selecting the target hyperparameter instance can be effect explanation information, to explain why the target hyperparameter instance has the highest effect evaluation information. After selecting the target hyperparameter instance with the highest effect evaluation from the first hyperparameter instance and the historical effect evaluation information corresponding to the first hyperparameter instance and the multiple historical hyperparameter instances, the method may further include:

[0115] Determine the training objective of the data processing system and the first training result generated by the first hyperparameter instance during the training of the data processing system;

[0116] Based on the first training result, determine the first causal relationship between the first hyperparameter instance and the training objective;

[0117] Based on the first causal relationship, the implicit effect explanation information between the generated hyperparameter instances and the training objective of the data processing system is obtained.

[0118] The first causal relationship can be used to determine the association between the dependent and effect variables. Through the first causal relationship, the implicit relationship between hyperparameter instances and the training objective can be interpreted, thus obtaining information on the effect. For example, the first causal relationship can clarify the direct factors in hyperparameters that have a direct impact on the training objective, as well as the indirect factors that have an indirect impact on the training objective. Furthermore, the causal relationship can be used to determine the effect relationship between hyperparameters and the training objective.

[0119] In this embodiment of the application, the effect of hyperparameter instances on training targets can be explained through the first causal relationship, that is, the corresponding influence of hyperparameter instances on training targets, so as to improve the usability of the explanation.

[0120] Furthermore, as another embodiment, the explanatory information for selecting the target hyperparameter instance can be decision explanation information to explain the decision-making process of selecting the target hyperparameter instance as the hyperparameter instance with the highest effect evaluation. After selecting the target hyperparameter instance with the highest effect evaluation from the first hyperparameter instance and the historical effect evaluation information corresponding to the first hyperparameter instance and the multiple historical hyperparameter instances respectively, the method may further include:

[0121] Based on the causal network discovery algorithm, the first causal model corresponding to the first causal network is determined.

[0122] Based on the causal effect estimation algorithm, the first effect model corresponding to the first causal network is determined;

[0123] Input the first hyperparameter instance into the first effect model to calculate the first effect score;

[0124] Based on the first causal model and the first effect score, generate decision explanation information for selecting the target hyperparameter instance as the hyperparameter instance with the highest effect evaluation.

[0125] In this embodiment of the application, when utilizing decision explanation information, decision explanation information for selecting the target hyperparameter instance as the hyperparameter instance with the highest effect evaluation can be generated based on the first causal model and the first effect score. The first causal model can be used to explain why the first hyperparameter instance generates the corresponding training target under the constraint of the training target, thereby achieving accurate explanation.

[0126] The training results of the first hyperparameter instance in the data processing system can be obtained through training.

[0127] As one embodiment, determining the training objective of the data processing system and the first training result generated by the first hyperparameter instance during the training of the data processing system may include:

[0128] Construct the data processing model corresponding to the data processing system using the first hyperparameter instance;

[0129] Determine the training objectives for the data processing system;

[0130] Obtain multiple training groups; where each training group includes multiple training data sets.

[0131] According to the training objective, train sequentially to obtain multiple training groups and sub-training results generated by the data processing model;

[0132] Based on the sub-training results corresponding to multiple training groups, the first training result generated by the first hyperparameter instance in the data processing system is determined.

[0133] Furthermore, in some embodiments, the data processing system may also include contextual features that influence the determination of hyperparameters. After constructing the data processing model corresponding to the data processing system using the first hyperparameter instance, it may further include determining the contextual features that influence the determination of hyperparameters.

[0134] The sub-training results obtained by training multiple training groups sequentially according to the training objective and generating sub-training results in the data processing model under the influence of contextual features can include: the sub-training results obtained by training multiple training groups sequentially according to the training objective and generating sub-training results in the data processing model under the influence of contextual features.

[0135] In this embodiment, a data processing model corresponding to the data processing system is constructed using a first hyperparameter instance. After determining the training objective of the data processing system, multiple training groups are used to train the model sequentially according to the training objective, resulting in training results generated by each training group in the data processing model. Based on the sub-training results corresponding to each of the multiple training groups, the first training result generated by the first hyperparameter instance in the data processing system is determined. By using multiple training groups to train the data processing model multiple times, and combining the sub-training results obtained from multiple training sessions to obtain the first training result, the accuracy of the training result is improved.

[0136] Since a data processing system can include any neural network model or mathematical calculation model, the specific training method used varies depending on the choice of the data processing model. Training methods can be divided into online training and offline training.

[0137] During offline training, the dataset used for training is known. Therefore, as a possible implementation, multiple training sets are obtained; wherein any training set includes multiple training data sets, including:

[0138] Read multiple training groups from the training dataset; where each training group includes multiple training data sets.

[0139] During online training, it is necessary to collect each set of training data online in real time and train the model on each set of training data. Therefore, as another embodiment, multiple training sets are obtained; wherein any training set includes multiple training data, including:

[0140] Determine the number of training sets that have been trained;

[0141] If the number of training groups exceeds the preset threshold, then multiple training groups will be obtained.

[0142] If the number of training groups is not greater than the preset threshold for the number of training groups, training data is collected online to obtain at least one online data point; if the number of data points of at least one online data point is greater than the preset threshold for the number of data points, a training group consisting of at least one online data point is determined; if the number of data points of at least one online data point is not greater than the preset threshold for the number of data points, the process returns to the step of collecting training data online to obtain at least one online data point and continues execution.

[0143] This application provides both online and offline training methods to adapt to different training scenarios and improve the training scenarios for hyperparameter optimization.

[0144] In the actual training process of hyperparameters, different parameter values ​​can be selected according to the value range of each sub-instance in the hyperparameter instance to obtain multiple hyperparameter instances for training the data processing model, thereby obtaining the highest training effect by evaluating the effect of the hyperparameter instances. In practical applications, due to the influence of hyperparameters and their selection range, the number of hyperparameter instances may be very large. If all hyperparameter instances are used to train the parameters of the data processing system, it may result in a large number of training iterations, leading to a decrease in training efficiency.

[0145] like Figure 4 The diagram shown is a flowchart of another embodiment of a hyperparameter optimization method provided in this application. The method may include:

[0146] 401: Determine the first hyperparameter instance currently corresponding to the hyperparameters of the data processing system.

[0147] 402: Determine whether the first hyperparameter instance meets the training conditions. If it does, proceed to step 403; otherwise, proceed to step 404.

[0148] 403: If satisfied, then based on the training results of the first hyperparameter instance in the data processing system, determine the first effect evaluation information corresponding to the first hyperparameter instance.

[0149] 404: If not satisfied, then randomly generate first effect evaluation information for the first hyperparameter instance whose evaluation effect is lower than the effect threshold.

[0150] 405: Obtain multiple historical hyperparameter instances and their corresponding historical performance evaluation information.

[0151] 406: Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances, select the target hyperparameter instance with the highest effect evaluation from the first hyperparameter instance and multiple historical hyperparameter instances.

[0152] In this embodiment, before actually training the hyperparameter instance, a condition judgment can be made regarding whether the training conditions of the hyperparameter instance are met. If the first hyperparameter instance meets the training conditions, it can be directly trained; if it does not meet the training conditions, a poor training result can be directly generated for it. Since the training process of hyperparameter instances in the data processing system is highly complex, judging the training conditions before training reduces the amount of training required and improves the optimization efficiency of hyperparameters.

[0153] As an example, whether the first hyperparameter instance satisfies the training conditions can be determined in the following way:

[0154] Identify the second hyperparameter instance that is associated with the first hyperparameter instance from among multiple historical hyperparameter instances;

[0155] Based on the second hyperparameter instance and the second effect evaluation information associated with the second hyperparameter instance, the first effect estimation result corresponding to the first hyperparameter instance is predicted;

[0156] If the first effect estimation result matches the preset estimation result, the first hyperparameter instance is determined to meet the training conditions.

[0157] If the first effect estimation result does not match the preset estimation result, it is determined that the first hyperparameter instance does not meet the training conditions.

[0158] The second hyperparameter instance can be a hyperparameter instance associated with the first hyperparameter instance from among multiple historical hyperparameter instances. For example, it can be the most recently obtained hyperparameter instance that precedes the first hyperparameter instance, and the causal or effect model of the second hyperparameter instance can be used to predict the effect estimation result. Alternatively, there can be multiple second hyperparameter instances, and curve fitting can be performed using multiple hyperparameter instances, and the result can be estimated based on the position of the first hyperparameter instance in the coordinate system.

[0159] In this embodiment, a second hyperparameter instance associated with the first hyperparameter instance is used to determine whether the training conditions are met. Based on the second hyperparameter instance and the second effect evaluation information associated with it, a first effect estimation result corresponding to the first hyperparameter instance can be predicted. By matching the first effect estimation result with a preset effect estimation result, it can be accurately determined whether the first hyperparameter instance meets the training conditions.

[0160] The second hyperparameter instance can be a historical hyperparameter instance most recently obtained before the first hyperparameter instance. As another embodiment, determining the second hyperparameter instance associated with the first hyperparameter instance from a plurality of historical hyperparameter instances may include:

[0161] The most recent historical hyperparameter instance that meets the training conditions among multiple historical hyperparameter instances is identified as the second hyperparameter instance.

[0162] When the second hyperparameter instance is the hyperparameter instance closest to the first hyperparameter instance. As a possible implementation, based on the second hyperparameter instance and the second effect evaluation information associated with it, the estimated first effect information corresponding to the first hyperparameter instance can be obtained by:

[0163] Obtain the second effect model associated with the second hyperparameter instance;

[0164] Input the first hyperparameter instance into the second effect model to calculate the second effect score;

[0165] The first effect estimation result corresponding to the first hyperparameter instance is determined by the second effect score.

[0166] It should be noted that the training results of the data processing model and the methods for obtaining the corresponding causal relationship, effect model and causal model are the same in any hyperparameter instance training in this application embodiment. For details, please refer to the relevant description of the first hyperparameter instance, which will not be repeated here.

[0167] Optionally, the second effect score can be directly used as the first effect estimation result. Alternatively, a ranking estimate can be performed based on the second effect score to determine the effect estimation ranking of the first hyperparameter instance, and the obtained effect estimation ranking can be used as the first effect estimation result.

[0168] In this embodiment of the application, a second effect score can be calculated by the second effect model associated with the second hyperparameter instance, and the first effect estimation result corresponding to the first hyperparameter instance can be determined by the second effect score, so as to obtain an accurate effect estimation result.

[0169] When the second hyperparameter instance is the hyperparameter instance closest to the first hyperparameter instance. As another possible implementation, based on the second hyperparameter instance and the second effect evaluation information associated with it, the estimation of the first effect estimation information corresponding to the first hyperparameter instance further includes:

[0170] Obtain the second causal model associated with the second hyperparameter instance;

[0171] Calculate the second data matching value between the first hyperparameter instance and the second causal model;

[0172] The first effect estimation result corresponding to the first hyperparameter instance is determined by matching the second data value.

[0173] Optionally, the second data matching value can be directly used as the first effect estimation result. Alternatively, a rank estimation can be performed based on the second data matching value to determine the effect estimation rank of the first hyperparameter instance, and the obtained effect estimation rank can be used as the first effect estimation result.

[0174] In this embodiment of the application, a second effect score can be calculated by the second effect model associated with the second hyperparameter instance, and the first effect estimation result corresponding to the first hyperparameter instance can be determined by the second effect score, so as to obtain an accurate effect estimation result.

[0175] The second hyperparameter instance can be one of multiple historical hyperparameter instances that are sequentially preceding the first hyperparameter instance in the training order. In some embodiments, there are multiple second hyperparameter instances; based on the second hyperparameter instances and the second effect evaluation information associated with the second hyperparameter instances, the first effect estimation information corresponding to the first hyperparameter instance is estimated by:

[0176] Obtain the third data matching values ​​associated with multiple instances of the second hyperparameter;

[0177] Data analysis and processing are performed on multiple second hyperparameter instances and the third data matching values ​​associated with each second hyperparameter instance to obtain data trend information between the matching degree of the hyperparameter instances and the causal model;

[0178] By using data trend information, the fourth data matching value corresponding to the first hyperparameter instance is estimated;

[0179] The first effect estimation information corresponding to the first hyperparameter instance is determined by matching the second data value.

[0180] By combining data trend information with the trend position of the first hyperparameter instance, we can predict the first effect estimation information of the first hyperparameter instance.

[0181] Multiple second hyperparameter instances can be consecutive historical hyperparameter instances. In one possible design, determining the second hyperparameter instance associated with the first hyperparameter instance among multiple historical hyperparameter instances may include:

[0182] Determine the number of reference instances; where the number of reference instances is greater than 1.

[0183] Sort multiple historical hyperparameter instances according to their training order;

[0184] Following the training order from back to front, select hyperparameter instances that match the number of references as the second hyperparameter instances.

[0185] When obtaining a new first hyperparameter instance, historical hyperparameter instances and their corresponding performance evaluation information can be referenced to obtain hyperparameter instances that produce better training results for the data processing model. In one possible design, determining the current first hyperparameter instance corresponding to the hyperparameters of the data processing system may include:

[0186] Determine the last obtained third hyperparameter instance among multiple historical hyperparameter instances, and the third effect evaluation information corresponding to the third hyperparameter instance;

[0187] Based on the third hyperparameter instance and the third effect evaluation information corresponding to the third hyperparameter instance, and combined with the preset parameter update algorithm, the first hyperparameter instance corresponding to the current hyperparameter of the data processing system is obtained.

[0188] In this embodiment, by introducing historical hyperparameter instances and the corresponding third effect evaluation information into the hyperparameter instance update process, a first hyperparameter instance that is more effective in training the data processing model can be obtained, thereby improving the training efficiency of the first hyperparameter instance and further promoting the optimization efficiency of the target hyperparameter instance.

[0189] In another possible design, determining the first hyperparameter instance corresponding to the current hyperparameter of the data processing system includes:

[0190] Detect the parameter optimization request triggered by the first user's hyperparameters in relation to the data processing system;

[0191] In response to the parameter optimization request, determine the first hyperparameter instance that currently corresponds to the hyperparameter of the data processing system.

[0192] In addition, some embodiments may include sending the target hyperparameter instance to the first user's first user equipment so that the first user equipment can output the target hyperparameter instance to the first user.

[0193] Optionally, when the technical solution of this application embodiment is directly applied to the user terminal, it can directly detect the parameter optimization request triggered by the first user for the hyperparameters of the data processing system.

[0194] Optionally, the technical solution of this application embodiment is applied to a server. A user equipment connected to the server can detect a parameter optimization request triggered by a user for the hyperparameters of the data processing system and send the parameter optimization request to the server. The server detecting the parameter optimization request triggered by the first user for the hyperparameters of the data processing system may include: receiving the parameter optimization request triggered by the hyperparameters of the data processing system sent by the first user equipment.

[0195] As another embodiment, selecting the target hyperparameter instance with the highest effect evaluation from the first hyperparameter instance and the historical effect evaluation information corresponding to the multiple historical hyperparameter instances may include:

[0196] Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances, select multiple candidate hyperparameter instances that meet the effect evaluation conditions from the first hyperparameter instance and multiple historical hyperparameter instances.

[0197] Multiple candidate hyperparameter instances are displayed to the second user so that the second user can select the target hyperparameter instance from the multiple candidate hyperparameter instances;

[0198] Obtain the target hyperparameter instance selected by the second user.

[0199] In practical applications, multiple candidate hyperparameter instances that meet certain performance evaluation criteria can be selected from the current hyperparameter instance and multiple historical hyperparameter instances. These candidate hyperparameter instances and their corresponding performance evaluation information can be viewed by the user, allowing the user to select the target hyperparameter instance from multiple candidate hyperparameter instances. By providing candidate hyperparameter instances, users can have more hyperparameter selection space, eliminating the need for users to try the training effects of different hyperparameter instances one by one, thus improving the efficiency of hyperparameter selection.

[0200] Optionally, multiple candidate hyperparameter instances that meet the effect evaluation conditions can be any of the following selection conditions: the evaluation level is higher than the predetermined standard, the evaluation parameter is higher than the predetermined value, or the evaluation words include preset keywords. The specific selection can be determined according to the actual usage requirements.

[0201] As yet another embodiment, before determining the first hyperparameter instance corresponding to the hyperparameter of the data processing system, the method may further include:

[0202] Based on the preset scenario information, extract the hyperparameters and training objectives corresponding to the data processing system.

[0203] Contextual information is related to the user's usage context of the data processing system and is relevant information abstracted from that context. In practical applications, data processing systems can be used in various contexts such as word retrieval, data querying, logistics tracking, object detection, ad click-through rate prediction, content recommendation, intelligent interaction, and autonomous driving. Different application domains have corresponding data processing systems to achieve their respective functions. Based on the different application contexts, after determining the data processing system to be used, the specific content or information relevant to that application context can be extracted. This extracted content or information is then used as contextual information, and modeling this contextual information is performed to obtain hyperparameters and training objectives.

[0204] In addition, in practical applications, besides using scene information to extract hyperparameters and training objectives of the data processing system, contextual features that have an important impact on the training process of the data processing model can also be extracted based on scene information, so as to use hyperparameters and contextual features together to train the data processing model and generate training results.

[0205] Furthermore, the data processing system can actually be any of a variety of candidate learning algorithms. These candidate algorithms can be pre-provided by the developers or provided by the user through their user devices. Additionally, users can directly provide the data processing system that requires hyperparameter optimization through their user devices.

[0206] As yet another embodiment, it may also include:

[0207] Based on preset scenario information, a data processing system that matches the scenario information is selected from multiple candidate learning algorithms.

[0208] Optionally, multiple candidate learning algorithms can be matched with preset scene information to estimate the degree of matching, and the candidate learning algorithm with the highest degree of matching can be used as the data processing system.

[0209] Optionally, the scene information can be input by the user. Before extracting the hyperparameters and training targets corresponding to the data processing system based on the preset scene information, the method may further include: detecting the scene information input by a third user.

[0210] In one possible design, the user equipment can display multiple candidate learning algorithms for the user to view, and detect the user's selection operation on any of the multiple candidate learning algorithms to obtain the data processing system that needs to be optimized for hyperparameters.

[0211] Therefore, a data processing system that selects a matching scenario information from multiple candidate learning algorithms based on preset scenario information can include:

[0212] Show multiple candidate learning algorithms to a third user;

[0213] The system detects the data processing system selected by the third user from the candidate learning algorithm that matches the scene information.

[0214] When the hyperparameter optimization method of this application embodiment is applied to a server, the user equipment can send the data processing system selected by the user for multiple candidate learning algorithms to the server so that the server can confirm the data processing system that the user has selected from multiple candidate learning algorithms that needs to be hyperparameter optimized.

[0215] In this embodiment of the application, by receiving the user's parameter optimization request and interacting with the user, the hyperparameter optimization service is provided to the user in a timely manner, thereby improving the utilization rate of the technical solution of this application.

[0216] It should be noted that the terms "first user," "second user," and "third user" in this application embodiment are merely distinguishing different users by name and do not have any actual sequential or quantitative meaning. In practical applications, "first user," "second user," etc., can refer to the same user or different users.

[0217] like Figure 5 The diagram shown is a flowchart of another embodiment of a hyperparameter optimization method provided in this application. The method may include:

[0218] 501: Determine the first hyperparameter instance currently corresponding to the hyperparameters of the data processing system.

[0219] 502: Determine whether the first hyperparameter instance meets the training conditions; if it does, proceed to step 503; if it does not, proceed to step 512.

[0220] 503: Determine the training objective of the data processing system and the first training result generated by the first hyperparameter instance during the training of the data processing system.

[0221] 504: Based on the first training result, determine the first causal network between the first hyperparameter instance and the training target.

[0222] 505: Based on the causal network discovery algorithm, determine the first causal model corresponding to the first causal network;

[0223] 506: Calculate the first data matching value between the first hyperparameter instance and the first causal model;

[0224] 507: Obtain the first effect information based on the first data matching value.

[0225] 508: Based on the causal effect estimation algorithm, determine the first effect model corresponding to the first causal network;

[0226] 509: Input the first hyperparameter instance into the first effect model and calculate the first effect score;

[0227] 510: Obtain the first evaluation information based on the first effect score.

[0228] 511: Based on the first effect information and / or the first evaluation information, determine the first effect evaluation information corresponding to the first hyperparameter instance.

[0229] 512: Randomly generate first effect evaluation information for the first hyperparameter instance whose evaluation effect is lower than the effect threshold.

[0230] 513: Obtain multiple historical hyperparameter instances and their corresponding historical performance evaluation information;

[0231] 514: Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances, select the target hyperparameter instance with the highest effect evaluation from the first hyperparameter instance and multiple historical hyperparameter instances.

[0232] 515: Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances, generate path explanation information corresponding to the selection process of the target hyperparameter instance.

[0233] 516: Based on the first causal network, generate the implicit effect explanation information between the hyperparameter instances and the training objective of the data processing system.

[0234] 517: Based on the first causal model and the first effect score, generate decision explanation information for selecting the target hyperparameter instance as the hyperparameter instance with the highest effect evaluation.

[0235] In this embodiment, after determining the first hyperparameter instance, it can be determined whether the first hyperparameter instance meets the training conditions. Then, for the first hyperparameter instance that meets the training conditions, the first training result generated under the training objective of the data processing system is determined. Subsequently, based on this first training result, a first causal network between the first hyperparameter instance and the training objective can be determined. And based on the causal network discovery algorithm, a first causal network model corresponding to the first causal network is determined. A first data matching value between the first hyperparameter instance and the first causal model is calculated, and first effect information is obtained based on this first data matching value.

[0236] Furthermore, based on a causal network effect estimation algorithm, a first effect model corresponding to the first causal network is determined. The first hyperparameter instance is input into this first effect model to calculate the first effect score. First evaluation information is then obtained based on the first effect score. Finally, based on the first effect information and / or the first evaluation information, the first effect evaluation information corresponding to the first hyperparameter instance is determined. Therefore, for first hyperparameter instances that meet the training conditions, corresponding model training and causal relationship estimation are performed to obtain accurate first effect evaluation information. For first hyperparameter instances that do not meet the training conditions, a first effect evaluation information with an evaluation effect lower than the effect threshold is directly set. This reduces the number of hyperparameter instances participating in training, lowers the space complexity during training, and improves the optimization efficiency of hyperparameter instances.

[0237] In addition, multiple historical hyperparameter instances and their corresponding historical performance evaluation information are used, combined with the first performance evaluation information of the first hyperparameter instance, to select the target hyperparameter instance with the highest performance evaluation from the first hyperparameter instance and multiple historical hyperparameter instances, so as to improve the accuracy of target hyperparameter instance acquisition.

[0238] Subsequently, path explanation information in the selection process of the target hyperparameter instance can be generated based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances. Based on the first causal network, implicit effect explanation information between the hyperparameter instance and the training objective of the data processing system can be generated. Furthermore, based on the first causal model and the first effect score, decision explanation information for selecting the target hyperparameter instance as the hyperparameter instance with the highest effect evaluation is generated. By utilizing the causal network and other content generated during training, the selection of the target hyperparameter instance can be explained from multiple perspectives and angles, clarifying the reasons for the generation of the target hyperparameter instance and further improving optimization efficiency.

[0239] To facilitate understanding, we can take the data processing system that needs to be optimized as the machine learning algorithm, the cloud server as the hyperparameter optimization method, and take the interaction with the user device as an example to introduce an application example of the embodiments of this application in detail.

[0240] refer to Figure 6In practical applications, user devices, such as mobile terminals and Internet of Things (IoT) terminals, can interact with users and communicate with servers capable of optimizing hyperparameters. Taking a mobile terminal M1 as the user device and a cloud server M2 as the server as an example: Mobile terminal M1 can detect the machine learning algorithm confirmed by the user (601). For example, it can detect the machine learning algorithm selected by the user from multiple candidate learning algorithms, and it can also detect the machine learning algorithm provided by the user. Mobile terminal M1 can send the machine learning algorithm (602) to cloud server M2. Cloud server M2 can determine the machine learning algorithm and perform hyperparameter optimization. During the optimization process, cloud server M2 can determine at any time (603) the first hyperparameter instance corresponding to the hyperparameters of the machine learning algorithm, and based on the first training result of the machine learning algorithm for that first hyperparameter instance, determine (604) the first performance evaluation information corresponding to that first hyperparameter instance. After obtaining 605 historical hyperparameter instances that have completed performance evaluations, and based on the first performance evaluation information corresponding to the first hyperparameter instance and the historical performance evaluation information corresponding to each of the historical hyperparameter instances, 606 target hyperparameter instances with the highest training performance evaluations are selected from the first hyperparameter instance and the multiple historical hyperparameter instances. By evaluating the training results of the first hyperparameter instance in the machine learning algorithm, the target hyperparameter instance with the highest training performance evaluation is obtained, thereby improving the effectiveness of training.

[0241] In addition, the cloud server M2 can also explain the reasons for selecting the target hyperparameter instance, generating 607 such as path explanation information, effect explanation information and / or decision explanation information.

[0242] Afterwards, cloud server M2 can send the target hyperparameter instance and the obtained explanation information to mobile terminal M1 (608). Upon receiving the target hyperparameter instance, mobile terminal M1 can display the target hyperparameter instance and explanation information to the user (609). When outputting the target hyperparameter instance, mobile terminal M1 can generate parameter prompt information for the target hyperparameter instance and output this parameter prompt information to the user. For example, the value of the target hyperparameter instance can be used as the prompt text to form the parameter prompt information, and this prompt text can be output to the user.

[0243] Furthermore, in practical applications, the cloud server M2 can also generate parameter prompts for the target hyperparameter instance and send these prompts to the mobile terminal M1. The output method of the target hyperparameter instance can include various forms, such as data, pages, information, or messages. This application embodiment does not impose too many limitations on the specific output method of the target hyperparameter instance.

[0244] The technical solutions of this application can be applied to various fields such as artificial intelligence interaction, data retrieval, content recommendation, click-through rate prediction, wastewater treatment monitoring, smart factories, industrial control, and facial recognition. For ease of understanding, the embodiments of this application will be described in detail using the following practical examples.

[0245] (1) Face Recognition Field. Common face recognition algorithms include Convolutional Neural Networks (CNN), Recursive Neural Networks (RNN), and Deep Neural Networks (DNN). Taking CNN as an example, the accuracy of CNN-based face recognition methods is mainly affected by three factors: training data, CNN architecture, and loss function. The CNN architecture is the machine learning model used to construct the data processing system. To improve the accuracy of face recognition, it is often necessary to accurately model the CNN architecture. Therefore, the hyperparameters used in the actual modeling process can be optimized using the technical solutions provided in this application.

[0246] Suppose that the hyperparameters of a CNN algorithm used for face recognition include the number of network layers, regularization parameter, number of learning iterations, batch size, learning rate, number of neurons in each hidden layer, and type of activation function. These parameters collectively constitute the hyperparameters, and each parameter, when given an actual value, constitutes a hyperparameter instance. For example, a hyperparameter instance can be formed by a network with 100 layers, a regularization parameter of 1, 1000 learning iterations, a batch size of 200 data points, a learning rate of 0.25, 10 neurons in each hidden layer, and a sigmoid activation function.

[0247] During the hyperparameter optimization process, the first hyperparameter instance corresponding to the above hyperparameters of the CNN algorithm can be identified. Then, based on the first training result of the machine learning algorithm for this first hyperparameter instance, the first performance evaluation information corresponding to this first hyperparameter instance is determined. Next, multiple historical hyperparameter instances and their corresponding historical performance evaluation information are obtained. Based on the first performance evaluation information corresponding to the first hyperparameter instance and the historical performance evaluation information corresponding to the multiple historical hyperparameter instances, the target hyperparameter instance with the highest performance evaluation is selected from the first hyperparameter instance and the multiple historical hyperparameter instances. Thus, the target hyperparameter instance for the CNN algorithm used for face recognition can be determined.

[0248] Subsequently, the machine learning model constructed using the target hyperparameter instance can be used to train the model parameters obtained from the CNN algorithm for face recognition. For example, a face recognition application can be constructed using the CNN model corresponding to the model parameters, and the face recognition application can be used to collect the face image of the user to be identified. Then, the CNN model corresponding to the model parameters can be used to perform face recognition to obtain the corresponding recognition results, such as authenticating the user's identity, completing bill payments or transfers, or authenticating the user's identity to open access control systems. The specific application scenarios of this face recognition algorithm in the embodiments of this application are merely illustrative and are not intended to limit the specific application scenarios.

[0249] (2) E-commerce. Search, recommendation, and computational advertising are the most common scenarios in e-commerce. This embodiment uses a search term recommendation scenario as an example for instance deployment. In this scenario, when a user clicks the search box in the APP (Application), the system will recommend some search terms (Query terms) to the user. The purpose of recommending search terms to users is to explore users' potential purchasing needs, increase user stickiness, and increase the total number of transactions. The search system uses the following architecture, combining a deep learning Encode-Decoder network to predict the recommendation of search terms. The system involves three hyperparameters (a, b, c), which are the number of retrieved search terms required by the three recall links. Previously, these were set manually based on human experience. Using this patented technology, relying on an online system for online hyperparameter optimization, the system can automatically find the instance of the hyperparameter combination with the highest performance evaluation and provide an explanation. Specifically, during the optimization of hyperparameters, the first hyperparameter instance corresponding to the three hyperparameters a, b, and c can be identified. Then, based on the first hyperparameter instance, the system can be deployed and executed, and the evaluation value obtained from the online operation can be collected. In this example, the total number of clicks is used to determine the first effect evaluation information corresponding to the first hyperparameter instance.

[0250] Then, multiple historical hyperparameter instances and their corresponding historical effect evaluation information are obtained. The causal network and causal effect estimation method corresponding to the first hyperparameter instance are determined. This causal network and causal effect estimation method are constructed based on the historical hyperparameter instances associated with the first hyperparameter instance. The actual usage effect of the first hyperparameter instance can be predicted using the first hyperparameter instance, its corresponding causal network, and causal effect.

[0251] If the first hyperparameter instance does not meet the usage conditions, the first hyperparameter instance with poor expected performance is directly filtered out, and poor performance evaluation information is set for it to form a historical hyperparameter instance.

[0252] If the first hyperparameter instance meets the usage conditions, then the causal network and causal effect estimation method corresponding to the first hyperparameter instance are constructed. After binding the causal network and causal effect estimation method corresponding to the first hyperparameter instance with the first hyperparameter instance, it is used as a historical hyperparameter instance and used for the condition judgment of subsequent hyperparameter instances. Hyperparameter instances with expectation differences are filtered out, thereby narrowing the candidate state space.

[0253] After several iterations, the target hyperparameter instance with the highest performance evaluation is selected from the latest first hyperparameter instance and multiple historical hyperparameter instances as the final result, and the reasoning and explanation information are provided.

[0254] (3) Financial Field. Stock index simulation is a crucial problem. Current technologies can solve this problem using machine learning models such as linear regression, SVM (support vector machines), and LSTM (Long Short-Term Memory). This involves numerous parameters during training, such as time_step, feature_dim, and hidden features in LSTM. Furthermore, it considers contextual features such as macroeconomic factors, microeconomic factors, and unexpected events, which influence the selection of hyperparameters. In this example, this patented technology utilizes one year's worth of historical system logs for offline hyperparameter optimization, automatically identifying the highest-performing hyperparameter combination and providing explanations for its effectiveness. Specifically, during the optimization of hyperparameters, the first hyperparameter instance corresponding to the three hyperparameters ime_step, feature_dim, and hiddenfeatrue can be identified. Then, based on this first hyperparameter instance, it can be actually deployed and executed in the simulation system, and the evaluation value obtained during the execution can be collected. In this example, it is the difference value of RMSE (Root Mean Squared Error) between the actual stock index and the first hyperparameter instance to determine the first effect evaluation information corresponding to the first hyperparameter instance.

[0255] Then, multiple historical hyperparameter instances and their corresponding historical effect evaluations and contextual feature information are obtained. Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to the multiple historical hyperparameter instances, the causal network and causal effect estimation method corresponding to the first hyperparameter instance are determined. This causal network and causal effect estimation method can be constructed from the historical hyperparameter instances associated with the first hyperparameter instance. The actual usage effect of the first hyperparameter instance can be predicted through the first hyperparameter instance and its corresponding causal network and causal effect.

[0256] If the first hyperparameter instance does not meet the usage conditions, it is directly filtered out, and a poor performance evaluation information is set for it to form a historical hyperparameter instance.

[0257] If the first hyperparameter instance meets the usage conditions, then the causal network and causal effect estimation method corresponding to the first hyperparameter instance are constructed. After binding the causal network and causal effect estimation method corresponding to the first hyperparameter instance with the first hyperparameter instance, it is used as a historical hyperparameter instance and used to judge the actual effect of the newly obtained first hyperparameter instances. Hyperparameter instances with poor expectations are directly filtered out, thereby narrowing the candidate state space.

[0258] After several iterations, the target hyperparameter instance with the highest performance evaluation is selected from the latest first hyperparameter instance and multiple historical hyperparameter instances as the final result, and an explanation is given. The explanation includes not only the causal relationship between hyperparameters but also the causal relationship between hyperparameters and contextual feature factors.

[0259] (4) Online Games. Online games typically involve many hyperparameters in their gameplay, which affect user experience and user spending rates. For example, the health loss caused by each attack of a hero in a game program can alter the game's balance, thus affecting the proportion of users choosing that hero. This patented technology can optimize these health loss values ​​and other hyperparameters to maximize the game's balance. Using this patented technology, online hyperparameter optimization can be performed on the online game system, automatically identifying the hyperparameter combination instance with the highest effect evaluation and providing an explanation. Specifically, during the hyperparameter optimization process, the first hyperparameter instance corresponding to the health loss value hyperparameter of all heroes can be identified. Then, based on this first hyperparameter instance, the game system is actually deployed and executed, and the game balance index evaluation value obtained during online operation is collected. In this example, this is the total number of clicks, determining the first effect evaluation information corresponding to the first hyperparameter instance.

[0260] Then, multiple historical hyperparameter instances and their corresponding historical effect evaluation information are obtained. Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to the multiple historical hyperparameter instances, the causal network and causal effect estimation method corresponding to the first hyperparameter instance are determined. This causal network and causal effect estimation method are constructed based on the historical hyperparameter instances associated with the first hyperparameter instance. The actual usage effect of the first hyperparameter instance can be predicted through the first hyperparameter instance and its corresponding causal network and causal effect.

[0261] If the first hyperparameter instance does not meet the usage conditions, the first hyperparameter instance with poor expected performance is directly filtered out, and poor performance evaluation information is set for it to form a historical hyperparameter instance.

[0262] If the first hyperparameter instance meets the usage conditions, then the causal network and causal effect estimation method corresponding to the first hyperparameter instance are constructed. After binding the causal network and causal effect estimation method corresponding to the first hyperparameter instance with the first hyperparameter instance, it is used as a historical hyperparameter instance and used for the condition judgment of subsequent hyperparameter instances. Hyperparameter instances with expectation differences are filtered out, thereby narrowing the candidate state space.

[0263] After several iterations, the target hyperparameter instance with the highest performance evaluation is selected from the latest first hyperparameter instance and multiple historical hyperparameter instances as the final result, and an explanation is given.

[0264] like Figure 7 The diagram shown is a structural schematic of another embodiment of a hyperparameter optimization device provided in this application. The device may include:

[0265] Instance determination module 701: Used to determine the first hyperparameter instance currently corresponding to the hyperparameters of the data processing system.

[0266] Effect evaluation module 702: used to determine the first effect evaluation information corresponding to the first hyperparameter instance based on the first training result of the first hyperparameter instance in the data processing system.

[0267] Historical Acquisition Module 703: Used to acquire multiple historical hyperparameter instances and the historical performance evaluation information corresponding to each of the multiple historical hyperparameter instances;

[0268] Instance selection module 704: is used to select the target hyperparameter instance with the highest effect evaluation from the first hyperparameter instance and the historical effect evaluation information corresponding to the first effect evaluation information of the first hyperparameter instance and the historical effect evaluation information corresponding to the multiple historical hyperparameter instances.

[0269] As one example, the effect evaluation module may include:

[0270] The result determination unit is used to determine the training objective of the data processing system and the first training result generated by the first hyperparameter instance during the training of the data processing system.

[0271] A relation building unit is used to determine the first causal relationship between the first hyperparameter instance and the training target based on the first training result;

[0272] The effect evaluation unit is used to determine the first effect evaluation information corresponding to the first hyperparameter instance based on the first causal relationship.

[0273] In some embodiments, the effect evaluation unit may include:

[0274] The relationship matching subunit is used to determine the first effect information based on the matching degree between the first causal relationship and the first hyperparameter instance;

[0275] The evaluation estimation subunit is used to estimate the first evaluation information corresponding to the generation of the first training result by the first hyperparameter instance based on the first causal relationship.

[0276] The effect determination subunit is used to determine the first effect evaluation information corresponding to the first hyperparameter instance based on the first effect information and / or the first evaluation information.

[0277] As one possible implementation, the relation matching subunit may include:

[0278] The first model module is used to determine the first causal model corresponding to the first causal network based on the causal network discovery algorithm.

[0279] The matching calculation module is used to calculate the first data matching value between the first hyperparameter instance and the first causal model.

[0280] The first acquisition module is used to obtain the first effect information based on the first data matching value.

[0281] As one possible implementation, the first causal relationship includes a first causal network, and the evaluation estimation subunit may include:

[0282] The second model module is used to determine the first effect model corresponding to the first causal network based on the causal effect estimation algorithm.

[0283] The effect estimation module is used to input the first hyperparameter instance into the first effect model and calculate the first effect score.

[0284] The second acquisition module is used to acquire the first evaluation information based on the first effect score.

[0285] As yet another example, Figure 7 The apparatus shown may also include:

[0286] The first explanation module is used to generate path explanation information corresponding to the selection process of the target hyperparameter instance based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances.

[0287] In some embodiments, the device may further include:

[0288] The result determination module is used to determine the training objective of the data processing system and the first training result generated by the first hyperparameter instance during the training of the data processing system.

[0289] The relationship building module is used to determine the first causal relationship between the first hyperparameter instance and the training target based on the first training result;

[0290] The second explanation module is used to generate implicit effect explanation information between hyperparameter instances and the training objective of the data processing system based on the first causal relationship.

[0291] The result building module and the result building unit process the same content, and the relation building module and the relation building unit process the same content.

[0292] In some embodiments, the first causal relationship includes a first causal network; the apparatus may further include:

[0293] The first model module is used to determine the first causal model corresponding to the first causal network based on the causal network discovery algorithm.

[0294] The second model module is used to determine the first effect model corresponding to the first causal network based on the causal effect estimation algorithm.

[0295] The effect estimation module is used to input the first hyperparameter instance into the first effect model and calculate the first effect score.

[0296] The third explanation module is used to generate decision explanation information for selecting the target hyperparameter instance as the hyperparameter instance with the highest effect evaluation, based on the first causal model and the first effect score.

[0297] Furthermore, optionally, the first interpretation module can be specifically used for:

[0298] Obtain the training order corresponding to multiple historical hyperparameter instances, and determine that the training order of the first hyperparameter instance is the last one;

[0299] Identify the interference identifiers corresponding to multiple historical hyperparameter instances;

[0300] Based on the training order of multiple historical hyperparameter instances and the first hyperparameter instance, hyperparameter instances with adjacent training order and the same interference label in the multiple historical hyperparameter instances and the first hyperparameter instance are grouped into the same group to obtain at least one group of hyperparameter instances.

[0301] Extract the first and last hyperparameter instances from at least one set of hyperparameter instances respectively, and combine them with the remaining hyperparameter instances that were not assigned to at least one set of hyperparameter instances to obtain multiple path hyperparameter instances for explaining the path optimization process;

[0302] Based on the training order corresponding to the multiple path hyperparameter instances, generate path explanation information corresponding to the selection process of the target hyperparameter instance.

[0303] As yet another embodiment, the effect evaluation module may also include:

[0304] The first evaluation unit is used to determine the first effect evaluation information corresponding to the first hyperparameter instance based on the training result of the first hyperparameter instance in the data processing system if the first hyperparameter instance meets the training conditions.

[0305] The performance evaluation module may also include:

[0306] The second evaluation unit is used to randomly generate first effect evaluation information for the first hyperparameter instance if the first hyperparameter instance does not meet the training conditions, and the evaluation effect is lower than the effect threshold.

[0307] As one possible implementation, the first evaluation unit and the second evaluation unit may specifically include:

[0308] The associated acquisition subunit is used to determine the second hyperparameter instance associated with the first hyperparameter instance among multiple historical hyperparameter instances;

[0309] The associated evaluation subunit is used to estimate the first effect estimation information corresponding to the first hyperparameter instance based on the second hyperparameter instance and the second effect evaluation information associated with the second hyperparameter instance.

[0310] The first judgment subunit determines that the first hyperparameter instance satisfies the training conditions if the first effect estimation result matches the preset estimation result.

[0311] The second judgment subunit is used to determine that the first hyperparameter instance does not meet the training conditions if the first effect estimation result does not match the preset estimation result.

[0312] In some embodiments, the association evaluation subunit may specifically include:

[0313] The first association module is used to obtain the second effect model associated with the second hyperparameter instance;

[0314] The first calculation module is used to input the first hyperparameter instance into the second effect model and calculate the second effect score.

[0315] The first estimation module is used to determine the first effect estimation result corresponding to the first hyperparameter instance through the second effect score.

[0316] In some embodiments, the association evaluation subunit may further include:

[0317] The second association module is used to obtain the second causal model associated with the second hyperparameter instance;

[0318] The second calculation module is used to calculate the second data matching value between the first hyperparameter instance and the second causal model;

[0319] The second estimation module is used to determine the first effect estimation result corresponding to the first hyperparameter instance through the second data matching value.

[0320] Furthermore, optionally, the associated acquisition of sub-units can be specifically used for:

[0321] The most recent historical hyperparameter instance that meets the training conditions among multiple historical hyperparameter instances is identified as the second hyperparameter instance.

[0322] As one embodiment, the second hyperparameter instance includes multiple instances; the correlation evaluation subunit may also include:

[0323] The data acquisition module is used to acquire the second data matching values ​​associated with multiple second hyperparameter instances respectively;

[0324] The data analysis module is used to perform data analysis and processing on multiple second hyperparameter instances and the second data matching values ​​associated with each second hyperparameter instance, and to obtain data trend information between the matching degree of the hyperparameter instances and the causal model.

[0325] The data prediction module is used to predict the first effect estimation information corresponding to the first hyperparameter instance by using data trend information.

[0326] As one possible implementation, the association-based acquisition of sub-units can also be used for:

[0327] Determine the number of reference instances; where the number of reference instances is greater than 1.

[0328] Sort multiple historical hyperparameter instances according to their training order;

[0329] Following the training order from back to front, select hyperparameter instances that match the number of references as the second hyperparameter instances.

[0330] In some embodiments, the result determination unit may include:

[0331] The model building subunit is used to construct the data processing model corresponding to the data processing system using the first hyperparameter instance;

[0332] The target determination subunit is used to determine the training target of the data processing system;

[0333] The training data sub-unit is used to acquire multiple training groups; each training group includes multiple training data sets.

[0334] The sub-result acquisition sub-unit is used to sequentially train according to the training objective to obtain the sub-training results generated by the data processing model for multiple training groups.

[0335] The result acquisition sub-unit is used to determine the first training result generated by the first hyperparameter instance in the data processing system based on the sub-training results corresponding to multiple training groups.

[0336] As yet another example, the training data subunit can specifically be used for:

[0337] Read multiple training groups from the training dataset; where each training group includes multiple training data sets.

[0338] As yet another example, the training data subunit can specifically be used for:

[0339] Determine the number of training sets that have been trained;

[0340] If the number of training groups exceeds the preset threshold, then multiple training groups will be obtained.

[0341] If the number of training groups is not greater than the preset threshold for the number of training groups, training data is collected online to obtain at least one online data point; if the number of data points of at least one online data point is greater than the preset threshold for the number of data points, a training group consisting of at least one online data point is determined; if the number of data points of at least one online data point is not greater than the preset threshold for the number of data points, the process returns to the step of collecting training data online to obtain at least one online data point and continues execution.

[0342] As one embodiment, the instance determination module may include:

[0343] The most recent instance unit is used to determine the last obtained third hyperparameter instance among multiple historical hyperparameter instances and the third effect evaluation information corresponding to the third hyperparameter instance.

[0344] The instance generation unit is used to obtain the first hyperparameter instance corresponding to the hyperparameter of the data processing system based on the third hyperparameter instance and the third effect evaluation information corresponding to the third hyperparameter instance, and in combination with the preset parameter update algorithm.

[0345] As yet another embodiment, the instance determination module may further include:

[0346] The request detection unit is used to detect parameter optimization requests triggered by the first user for the hyperparameters of the data processing system;

[0347] The request-response unit is used to respond to parameter optimization requests and determine the first hyperparameter instance that currently corresponds to the hyperparameter of the data processing system.

[0348] In some embodiments, it also includes:

[0349] The instance sending module is used to send the target hyperparameter instance to the first user's first user device, so that the first user device can output the target hyperparameter instance to the first user.

[0350] As yet another embodiment, the instance selection module may include:

[0351] The candidate selection unit is used to select multiple candidate hyperparameter instances that meet the effect evaluation conditions from the first hyperparameter instance and the multiple historical hyperparameter instances, based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to the multiple historical hyperparameter instances respectively.

[0352] The candidate display unit is used to present multiple candidate hyperparameter instances to a second user, so that the second user can select the target hyperparameter instance from the multiple candidate hyperparameter instances.

[0353] The instance acquisition unit is used to acquire the target hyperparameter instance selected by the second user.

[0354] As yet another embodiment, the device may further include:

[0355] The scene processing module is used to extract the hyperparameters and training objectives corresponding to the data processing system based on preset scene information.

[0356] In some embodiments, it also includes:

[0357] The algorithm selection module is used to select a data processing system that matches the scenario information from multiple candidate learning algorithms based on preset scenario information.

[0358] As one possible implementation, it also includes:

[0359] The scene detection module is used to detect scene information input by a third user.

[0360] In some embodiments, the algorithm selection module can be used to:

[0361] A data processing system is used to present multiple candidate learning algorithms to a third user and to detect which algorithm the third user selects from the candidate learning algorithms that matches the scene information.

[0362] Figure 7 The hyperparameter optimization device can perform Figure 1 The implementation principle and technical effects of the hyperparameter optimization method in the illustrated embodiments will not be elaborated further. The specific implementation methods of the various modules, units, and sub-units executed by the processing components in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0363] In practical applications, Figure 7 The hyperparameter optimization device shown can be configured as a computing device, reference Figure 8This is a schematic diagram of the structure of one embodiment of a computing device provided in this application. The device may include: a storage component 801 and a processing component 802; the storage component 801 is used to store one or more computer instructions; one or more computer instructions are invoked by the processing component 802 to execute. Figure 1 The hyperparameter optimization method is shown in the examples.

[0364] The processing component 802 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0365] Storage component 801 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0366] Of course, computing devices may also include other components, such as input / output interfaces and communication components. Input / output interfaces provide an interface between processing components and peripheral interface modules, which can be output devices, input devices, etc. Communication components are configured to facilitate wired or wireless communication between the computing device and other devices.

[0367] Furthermore, embodiments of this application also provide a computer-readable storage medium that can store one or more computer instructions, which, when executed, are used to implement any of the hyperparameter optimization methods in embodiments of this application.

[0368] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0369] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of a necessary general-purpose hardware platform, or by a combination of hardware and software. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a computer product. This application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0370] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A hyperparameter optimization method, characterized in that, include: Determine the first hyperparameter instance currently corresponding to the hyperparameters of the data processing system; Based on the first training result of the first hyperparameter instance in the data processing system, determine the first effect evaluation information corresponding to the first hyperparameter instance; The first performance evaluation information is used to evaluate the training performance of the first hyperparameter instance in the data processing model, which is constructed based on the first hyperparameter instance; the data processed by the data processing model includes text data or image data. Obtain multiple historical hyperparameter instances and their corresponding historical performance evaluation information. Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances, select the target hyperparameter instance with the highest effect evaluation from the first hyperparameter instance and the multiple historical hyperparameter instances. The determination of the first performance evaluation information corresponding to the first hyperparameter instance based on the first training result of the first hyperparameter instance in the data processing system includes: Determine the training objective of the data processing system and the first training result generated by the first hyperparameter instance during the training of the data processing system; Based on the first training result, a first causal relationship is determined between the first hyperparameter instance and the training target; Based on the first causal relationship, determine the first effect evaluation information corresponding to the first hyperparameter instance.

2. The method according to claim 1, characterized in that, The step of determining the first effect evaluation information corresponding to the first hyperparameter instance based on the first causal relationship includes: The first effect information is determined based on the matching degree between the first causal relationship and the first hyperparameter instance; Based on the first causal relationship, estimate the first evaluation information corresponding to the generation of the first training result by the first hyperparameter instance; Based on the first effect information and / or the first evaluation information, determine the first effect evaluation information corresponding to the first hyperparameter instance.

3. The method according to claim 2, characterized in that, The first causal relationship includes a first causal network; determining the first effect information based on the matching degree between the first causal relationship and the first hyperparameter instance includes: Based on the causal network discovery algorithm, the first causal model corresponding to the first causal network is determined. Calculate the first data matching value between the first hyperparameter instance and the first causal model; Based on the first data matching value, obtain the first effect information.

4. The method according to claim 2, characterized in that, The first causal relationship includes a first causal network; the step of estimating the first evaluation information corresponding to the generation of the first training result based on the first causal relationship includes: Based on the causal effect estimation algorithm, the first effect model corresponding to the first causal network is determined; Input the first hyperparameter instance into the first effect model to calculate the first effect score; Based on the first effect score, obtain the first evaluation information.

5. The method according to claim 1, characterized in that, Also includes: Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to multiple historical hyperparameter instances, path explanation information corresponding to the selection process of the target hyperparameter instance is generated.

6. The method according to claim 5, characterized in that, Also includes: Determine the training objective of the data processing system and the first training result generated by the first hyperparameter instance during the training of the data processing system; Based on the first training result, a first causal relationship is determined between the first hyperparameter instance and the training target; Based on the first causal relationship, implicit effect explanation information is generated between the hyperparameter instance and the training objective of the data processing system.

7. The method according to claim 6, characterized in that, The first causal relationship includes a first causal network; it also includes: Based on the causal network discovery algorithm, the first causal model corresponding to the first causal network is determined. Based on the causal effect estimation algorithm, the first effect model corresponding to the first causal network is determined; Input the first hyperparameter instance into the first effect model to calculate the first effect score; Based on the first causal model and the first effect score, generate decision explanation information for selecting the target hyperparameter instance as the hyperparameter instance with the highest effect evaluation.

8. The method according to claim 1, characterized in that, The determination of the first performance evaluation information corresponding to the first hyperparameter instance based on the training results of the first hyperparameter instance in the data processing system includes: If the first hyperparameter instance meets the training conditions, then based on the training results of the first hyperparameter instance in the data processing system, the first effect evaluation information corresponding to the first hyperparameter instance is determined.

9. The method according to claim 8, characterized in that, Also includes: If the first hyperparameter instance does not meet the training conditions, then a first effect evaluation message with an evaluation effect lower than the effect threshold is randomly generated for the first hyperparameter instance.

10. The method according to claim 1, characterized in that, Determining the training objective of the data processing system and the first training result generated by the first hyperparameter instance during the training of the data processing system includes: The data processing model corresponding to the data processing system is constructed using the first hyperparameter instance; Determine the training objective of the data processing system; Obtain multiple training groups; each training group includes multiple training data sets. According to the training objective, the multiple training groups are trained sequentially to obtain the sub-training results generated by the data processing model, respectively. Based on the sub-training results corresponding to the multiple training groups, the first training result generated by the first hyperparameter instance in the data processing system is determined.

11. The method according to claim 1, characterized in that, The first hyperparameter instance corresponding to the current hyperparameter of the determined data processing system includes: Determine the last obtained third hyperparameter instance among the plurality of historical hyperparameter instances and the third effect evaluation information corresponding to the third hyperparameter instance; Based on the third hyperparameter instance and the third effect evaluation information corresponding to the third hyperparameter instance, and combined with the preset parameter update algorithm, the first hyperparameter instance corresponding to the current hyperparameter of the data processing system is obtained.

12. The method according to claim 1, characterized in that, The first hyperparameter instance corresponding to the current hyperparameter of the determined data processing system includes: Detect the parameter optimization request triggered by the first user for the hyperparameters of the data processing system; In response to the parameter optimization request, determine the first hyperparameter instance that currently corresponds to the hyperparameter of the data processing system.

13. The method according to claim 12, characterized in that, Also includes: The target hyperparameter instance is sent to the first user's first user equipment so that the first user equipment can output the target hyperparameter instance to the first user.

14. The method according to claim 1, characterized in that, The step of selecting the target hyperparameter instance with the highest performance evaluation from the first hyperparameter instance and the multiple historical hyperparameter instances, based on the first performance evaluation information corresponding to the first hyperparameter instance and the historical performance evaluation information corresponding to the multiple historical hyperparameter instances, includes: Based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to the multiple historical hyperparameter instances, select multiple candidate hyperparameter instances that meet the effect evaluation conditions from the first hyperparameter instance and the multiple historical hyperparameter instances. Multiple candidate hyperparameter instances are presented to the second user so that the second user can select the target hyperparameter instance from the multiple candidate hyperparameter instances; Obtain the target hyperparameter instance selected by the second user.

15. The method according to claim 1, characterized in that, Before determining the first hyperparameter instance corresponding to the hyperparameter of the data processing system, the method further includes: Based on the preset scenario information, the hyperparameters and training objectives corresponding to the data processing system are extracted.

16. The method according to claim 15, characterized in that, Also includes: Based on preset scenario information, a data processing system that matches the scenario information is selected from multiple candidate learning algorithms.

17. The method according to claim 15, characterized in that, Also includes: Detect scene information input by a third user.

18. The method according to claim 16, characterized in that, The data processing system that selects the data matching the scene information from the plurality of candidate learning algorithms based on preset scene information includes: The multiple candidate learning algorithms are then presented to a third user. The data processing system selected by the third user from the candidate learning algorithm that matches the scene information is detected.

19. A hyperparameter optimization device, characterized in that, include: The instance determination module is used to determine the first hyperparameter instance currently corresponding to the hyperparameter of the data processing system. The performance evaluation module is used to determine the training objective of the data processing system and the first training result generated by the first hyperparameter instance during the training of the data processing system. Based on the first training result, a first causal relationship between the first hyperparameter instance and the training target is determined; based on the first causal relationship, first performance evaluation information corresponding to the first hyperparameter instance is determined. The first performance evaluation information is used to evaluate the training performance of the first hyperparameter instance in the data processing model, which is constructed based on the first hyperparameter instance; the data processed by the data processing model includes text data or image data. The history acquisition module is used to acquire multiple historical hyperparameter instances and the historical performance evaluation information corresponding to each of the multiple historical hyperparameter instances. The instance selection module is used to select the target hyperparameter instance with the highest effect evaluation from the first hyperparameter instance and the multiple historical hyperparameter instances, based on the first effect evaluation information corresponding to the first hyperparameter instance and the historical effect evaluation information corresponding to the multiple historical hyperparameter instances respectively.

20. A computing device, characterized in that, It includes: a storage component and a processing component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked by the processing component to execute the hyperparameter optimization method according to any one of claims 1 to 18.

Images