Multi-agent combined model automatic training method and related product

Abstract

The invention provides a multi-agent combined model automatic training method and a related product. According to one specific mode of the method, a decision-making agent corresponding to a target business scene is utilized to determine a lower-round hyper-parameter optimization execution agent for controlling a lower-round hyper-parameter optimization scheme of a target machine learning model; and based on the generated parameter adjustment record, a next-round hyper-parameter optimization execution agent is utilized to execute corresponding control operation so as to stop training of the target machine learning model, or replace a hyper-parameter search space of the target machine learning model, or determine next-round hyper-parameter configuration of the target machine learning model. And training the target machine learning model according to the next-round hyper-parameter configuration. Therefore, a traditional Bayesian hyper-parameter optimization method is intelligently transformed; the model training frequency is reduced; the model construction efficiency is improved; the model performance is improved; and self-adaptive model training is realized by automatically judging early stopping and continuing training.

Description

Technical Field

[0001] Embodiments of this disclosure relate to the field of machine learning technology, and more specifically to methods and related products for automatic training of models combining multiple agents. Background Technology

[0002] The current machine learning model training process is difficult to automate completely because some key steps in machine learning model training (such as hyperparameter optimization) are difficult to automate completely.

[0003] Although there are already automated hyperparameter search tools based on Bayesian optimization (such as Hyperopt and Optuna) that can perform hyperparameter optimization steps without human intervention, full automation is not common in actual industrial scenarios. The main reasons include: (1) the computational cost of automated search is high, especially when the sample size is large or the model structure is complex, the time consumption of a single hyperparameter experiment is significant; (2) when the search space is too large, automated hyperparameter search tools are difficult to converge efficiently within a limited time; (3) the engineer's experience judgment is still crucial - for example, according to the task type and data characteristics, the reasonable range of hyperparameters such as learning rate or regularization coefficient is manually limited, and then combined with a small amount of automated search or manual adjustment, followed by analysis of the results, further correction of the search range and repeated optimization. Therefore, the current mainstream approach is: to take human experience as the main guide, and achieve a balance between efficiency and effectiveness through the iterative cycle of "defining the hyperparameter search range → automatic search → analysis and feedback → adjustment range", supplemented by lightweight automated tools. That is, "human-machine collaboration", rather than relying on existing automated hyperparameter search tools to complete it automatically.

[0004] Therefore, a fully automated method for training machine learning models is urgently needed. Summary of the Invention

[0005] Embodiments of this disclosure provide methods, apparatus, electronic devices, storage media, and computer program products for automatic training of models incorporating multiple agents.

[0006] In a first aspect, embodiments of this disclosure provide an automatic training method for a model incorporating multiple agents, the method comprising: Based on the historical hyperparameter tuning record set and the current hyperparameter search space of the target machine learning model, the current hyperparameter configuration is generated using the current proxy model; Perform the following training and hyperparameter tuning record generation operations: train the target machine learning model according to the current hyperparameter configuration; perform performance evaluation on the target machine learning model to obtain training evaluation results; generate hyperparameter tuning records based on the current hyperparameter configuration and the training evaluation results; By utilizing the decision-making agent corresponding to the target business scenario to select the execution agent, the next hyperparameter optimization execution agent that controls the next hyperparameter optimization scheme of the target machine learning model is determined. Based on the generated parameter tuning records, the agent executes corresponding control operations using the next round of hyperparameter optimization to stop training the target machine learning model or determine the next round of hyperparameter configuration for the target machine learning model, and train the target machine learning model according to the next round of hyperparameter configuration.

[0007] In some optional implementations, the training and parameter tuning record generation operation further includes: The generated parameter tuning records are added to the global historical parameter tuning record set of the target machine learning model and the local historical parameter tuning record set of the current proxy model.

[0008] In some optional implementations, after adding the generated hyperparameter tuning records to the global historical hyperparameter tuning record set of the target machine learning model and the local historical hyperparameter tuning record set of the current proxy model, the training and hyperparameter tuning record generation operation further includes: Based on the local historical parameter tuning record set of the current proxy model and the current hyperparameter search space of the target machine learning model, the original hyperparameter configuration for the next round is generated using the current proxy model.

[0009] In some optional implementations, the step of generating the original hyperparameter configuration for the next round using the current proxy model, based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space of the target machine learning model, includes: The local historical parameter tuning records belonging to the current hyperparameter search space in the local historical parameter tuning records of the current proxy model are determined as the local historical parameter tuning records of the current space; Based on the current local historical parameter tuning records and the current hyperparameter search space, the original hyperparameter configuration for the next round is generated using the current surrogate model.

[0010] In some optional implementations, the next round of hyperparameter optimization execution agent is any of the following hyperparameter optimization execution agents: The first hyperparameter optimization execution agent is used to continue training with the original next round hyperparameter configuration; The second hyperparameter optimization execution agent is used to evaluate whether to stop training; The third hyperparameter optimization execution agent is used to instruct whether to change the hyperparameter search space and continue training for the first preset number of rounds. The fourth hyperparameter optimization execution agent is used to adopt other hyperparameter optimization schemes instead of the original next-round hyperparameter configuration during the subsequent second preset number of hyperparameter optimization rounds.

[0011] In some optional implementations, the step of utilizing the decision-making agent corresponding to the target business scenario to select the execution agent, and determining the next-round hyperparameter optimization execution agent to control the next-round hyperparameter optimization scheme of the target machine learning model, includes: Based on the preset probability or preset round frequency corresponding to each hyperparameter optimization agent, determine the next round of hyperparameter optimization agent for training the target machine learning model; or Based on the global historical parameter tuning record set, a large model is used to determine the next round of hyperparameter optimization execution agent that controls the next round of hyperparameter optimization scheme for the target machine learning model.

[0012] In some optional implementations, the step of using the next round of hyperparameter optimization to execute corresponding control operations based on the generated parameter tuning record includes: In response to the next round of hyperparameter optimization execution agent being the first next round of hyperparameter optimization execution agent, the following original hyperparameter optimization and training operations are performed: updating the current hyperparameter configuration to the original next round hyperparameter configuration, and performing the training and hyperparameter tuning record generation operation.

[0013] In some optional implementations, the step of using the next round of hyperparameter optimization to execute corresponding control operations based on the generated parameter tuning record includes: In response to the next round of hyperparameter optimization execution agent being the second hyperparameter optimization execution agent, a determination is made on whether to stop training based on the global historical hyperparameter tuning record set; In response to determining to stop training, the model parameters of the target machine learning model are persistently saved; In response to the determination not to stop training, the original hyperparameter optimization and training operations are performed.

[0014] In some optional implementations, the step of using the next round of hyperparameter optimization to execute corresponding control operations based on the generated parameter tuning record includes: In response to the next round of hyperparameter optimization execution agent being the third hyperparameter optimization execution agent, the large model is used to determine whether to change the hyperparameter search space based on the global historical hyperparameter tuning record set and the current hyperparameter search space, and a new hyperparameter search space is generated after determining to change the hyperparameter search space; In response to the successful generation of a new hyperparameter search space, the current hyperparameter search space of the target machine learning model is updated to the new hyperparameter search space. The local hyperparameter tuning record set of the current proxy model is cleared, and the local historical hyperparameter tuning records belonging to the current hyperparameter search space in the global historical hyperparameter tuning record set are added to the local hyperparameter tuning record set of the current proxy model. Based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space, the original hyperparameter configuration for the next round is generated using the current proxy model, and the original hyperparameter optimization and training operations are continuously executed for the first preset number of rounds.

[0015] In some optional implementations, the step of using the next round of hyperparameter optimization to execute corresponding control operations based on the generated parameter tuning record includes: In response to the next round of hyperparameter optimization execution agent being the fourth hyperparameter optimization execution agent, the hyperparameter optimization recommendation operation is continuously executed for the second preset number of rounds or less: determining the recommended hyperparameter configuration of the target machine learning model; updating the current hyperparameter configuration of the target machine learning model to the recommended hyperparameter configuration; and executing the training and hyperparameter tuning record generation operation.

[0016] In some optional implementations, the step of utilizing the decision-making agent corresponding to the target business scenario to select the execution agent, and determining the next-round hyperparameter optimization execution agent to control the next-round hyperparameter optimization scheme of the target machine learning model, includes: Based on the probability of changing the preset hyperparameter search space or the frequency of changing the preset hyperparameter search space rounds, the next round of hyperparameter optimization execution agent training the target machine learning model is determined as the fifth hyperparameter optimization execution agent based on the large model; and Based on the generated parameter tuning records, the agent performs corresponding control operations using the next round of hyperparameter optimization, including: In response to the next round of hyperparameter optimization execution agent being the fifth hyperparameter optimization execution agent, the following training operations after changing the hyperparameter search space are performed: using a large model based on the global historical hyperparameter tuning record set and the current hyperparameter search space to determine whether to change the hyperparameter search space, and generating a new hyperparameter search space after determining to change the hyperparameter search space; in response to successfully generating a new hyperparameter search space, updating the current hyperparameter search space of the target machine learning model to the new hyperparameter search space, clearing the local hyperparameter tuning record set of the current proxy model, adding the local historical hyperparameter tuning records belonging to the current hyperparameter search space from the global historical hyperparameter tuning record set to the local hyperparameter tuning record set of the current proxy model; based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space, generating the original next round hyperparameter configuration using the current proxy model, and performing the original hyperparameter optimization and training operations.

[0017] Secondly, embodiments of this disclosure provide an automatic model training apparatus combining multiple agents, the apparatus comprising: The current hyperparameter configuration generation module is configured to generate the current hyperparameter configuration based on the historical hyperparameter tuning record set and the current hyperparameter search space of the target machine learning model, using the current proxy model. The training and hyperparameter tuning record generation module is configured to perform the following training and hyperparameter tuning record generation operations: training the target machine learning model according to the current hyperparameter configuration; performing performance evaluation on the target machine learning model to obtain training evaluation results; and generating hyperparameter tuning records based on the current hyperparameter configuration and the training evaluation results. The agent determination module is configured to use the decision agent corresponding to the target business scenario to select the agent, and determine the next hyperparameter optimization agent to control the next hyperparameter optimization scheme of the target machine learning model. The next round hyperparameter optimization and training module is configured to, based on the generated hyperparameter tuning record, use the next round hyperparameter optimization execution agent to perform corresponding control operations, thereby stopping the training of the target machine learning model or determining the next round hyperparameter configuration of the target machine learning model, and training the target machine learning model according to the next round hyperparameter configuration.

[0018] In some optional implementations, the training and parameter tuning record generation operation further includes: The generated parameter tuning records are added to the global historical parameter tuning record set of the target machine learning model and the local historical parameter tuning record set of the current proxy model.

[0019] In some optional implementations, after adding the generated hyperparameter tuning records to the global historical hyperparameter tuning record set of the target machine learning model and the local historical hyperparameter tuning record set of the current proxy model, the training and hyperparameter tuning record generation operation further includes: Based on the local historical parameter tuning record set of the current proxy model and the current hyperparameter search space of the target machine learning model, the original hyperparameter configuration for the next round is generated using the current proxy model.

[0020] In some optional implementations, the step of generating the original hyperparameter configuration for the next round using the current proxy model, based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space of the target machine learning model, includes: The local historical parameter tuning records belonging to the current hyperparameter search space in the local historical parameter tuning records of the current proxy model are determined as the local historical parameter tuning records of the current space; Based on the current local historical parameter tuning records and the current hyperparameter search space, the original hyperparameter configuration for the next round is generated using the current surrogate model.

[0021] In some optional implementations, the next round of hyperparameter optimization execution agent is any of the following hyperparameter optimization execution agents: The first hyperparameter optimization execution agent is used to continue training with the original next round hyperparameter configuration; The second hyperparameter optimization execution agent is used to evaluate whether to stop training; The third hyperparameter optimization execution agent is used to instruct whether to change the hyperparameter search space and continue training for the first preset number of rounds. The fourth hyperparameter optimization execution agent is used to adopt other hyperparameter optimization schemes instead of the original next-round hyperparameter configuration during the subsequent second preset number of hyperparameter optimization rounds.

[0022] In some optional implementations, the agent determination module is further configured to: Based on the preset probability or preset round frequency corresponding to each hyperparameter optimization agent, determine the next round of hyperparameter optimization agent for training the target machine learning model; or Based on the global historical parameter tuning record set, a large model is used to determine the next round of hyperparameter optimization execution agent that controls the next round of hyperparameter optimization scheme for the target machine learning model.

[0023] In some optional implementations, the next round of hyperparameter optimization and training module is further configured as follows: In response to the next round of hyperparameter optimization execution agent being the first next round of hyperparameter optimization execution agent, the following original hyperparameter optimization and training operations are performed: updating the current hyperparameter configuration to the original next round hyperparameter configuration, and performing the training and hyperparameter tuning record generation operation.

[0024] In some optional implementations, the next round of hyperparameter optimization and training module is further configured as follows: In response to the next round of hyperparameter optimization execution agent being the second hyperparameter optimization execution agent, a determination is made on whether to stop training based on the global historical hyperparameter tuning record set; In response to determining to stop training, the model parameters of the target machine learning model are persistently saved; In response to the determination not to stop training, the original hyperparameter optimization and training operations are performed.

[0025] In some optional implementations, the next round of hyperparameter optimization and training module is further configured as follows: In response to the next round of hyperparameter optimization execution agent being the third hyperparameter optimization execution agent, the large model is used to determine whether to change the hyperparameter search space based on the global historical hyperparameter tuning record set and the current hyperparameter search space, and a new hyperparameter search space is generated after determining to change the hyperparameter search space; In response to the successful generation of a new hyperparameter search space, the current hyperparameter search space of the target machine learning model is updated to the new hyperparameter search space. The local hyperparameter tuning record set of the current proxy model is cleared, and the local historical hyperparameter tuning records belonging to the current hyperparameter search space in the global historical hyperparameter tuning record set are added to the local hyperparameter tuning record set of the current proxy model. Based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space, the original hyperparameter configuration for the next round is generated using the current proxy model, and the original hyperparameter optimization and training operations are continuously executed for the first preset number of rounds.

[0026] In some optional implementations, the next round hyperparameter optimization and training module is further configured to: in response to the next round hyperparameter optimization execution agent being the fourth hyperparameter optimization execution agent, continuously execute the hyperparameter optimization recommendation operation for the second preset number of rounds or less: determine the recommended hyperparameter configuration of the target machine learning model; update the current hyperparameter configuration of the target machine learning model to the recommended hyperparameter configuration; and execute the training and hyperparameter tuning record generation operation.

[0027] In some optional implementations, the agent determination module is further configured to: determine the next round of hyperparameter optimization agent for training the target machine learning model as the fifth hyperparameter optimization agent based on the large model, according to the probability of changing the preset hyperparameter search space or the round frequency of changing the preset hyperparameter search space; and The next round of hyperparameter optimization and training module is further configured as follows: In response to the next round of hyperparameter optimization execution agent being the fifth hyperparameter optimization execution agent, the following training operations after changing the hyperparameter search space are performed: using a large model based on the global historical hyperparameter tuning record set and the current hyperparameter search space to determine whether to change the hyperparameter search space, and generating a new hyperparameter search space after determining to change the hyperparameter search space; in response to successfully generating a new hyperparameter search space, updating the current hyperparameter search space of the target machine learning model to the new hyperparameter search space, clearing the local hyperparameter tuning record set of the current proxy model, adding the local historical hyperparameter tuning records belonging to the current hyperparameter search space from the global historical hyperparameter tuning record set to the local hyperparameter tuning record set of the current proxy model; based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space, generating the original next round hyperparameter configuration using the current proxy model, and performing the original hyperparameter optimization and training operations.

[0028] Thirdly, embodiments of this disclosure provide a method for applying a multi-agent model, the method comprising: Obtain business data for the target business scenario; The business data is input into the target machine learning model to obtain the prediction result corresponding to the business data, wherein the target machine learning model is pre-trained using the method described in any of the implementations in the first aspect.

[0029] Fourthly, embodiments of this disclosure provide a model application apparatus incorporating multiple agents, including: The data acquisition module is configured to acquire business data for the target business scenario; The data prediction module is configured to input the business data into a target machine learning model to obtain a prediction result corresponding to the business data, wherein the target machine learning model is pre-trained using a method described in any of the implementations in the first aspect.

[0030] Fifthly, embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, which, when executed by the one or more processors, cause the one or more processors to implement the method described in any of the implementations of the first and / or third aspects.

[0031] In a sixth aspect, embodiments of this disclosure provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by one or more processors, implements the method as described in any of the implementations of the first and / or third aspects.

[0032] In a seventh aspect, embodiments of this disclosure provide a computer program product including a computer program / instructions that, when executed by a processor, implement the method as described in any of the implementations of the first and / or third aspects.

[0033] To address the challenges faced by existing Bayesian hyperparameter optimization methods in model training—including high computational costs for automated search, convergence difficulties due to excessively large search spaces, and strong reliance on human experience making full automation difficult—this disclosure provides a multi-agent model automatic training method, apparatus, electronic device, storage medium, and computer program product. First, based on a historical hyperparameter tuning record set and the current hyperparameter search space of the target machine learning model, a current hyperparameter configuration is generated using the current agent model. Then, the following training and hyperparameter tuning record generation operations are performed: training the target machine learning model according to the current hyperparameter configuration; evaluating the performance of the target machine learning model to obtain training evaluation results; and generating a hyperparameter tuning record based on the current hyperparameter configuration and training evaluation results. Next, using the decision agent corresponding to the target business scenario for selecting the execution agent, a next-round hyperparameter optimization execution agent is determined to control the next round of hyperparameter optimization scheme for the target machine learning model. Finally, based on the generated hyperparameter tuning record, the next-round hyperparameter optimization execution agent performs corresponding control operations to stop training the target machine learning model or determine the next round hyperparameter configuration of the target machine learning model, and train the target machine learning model according to the next round hyperparameter configuration. Thus, since the decision-making agent is an agent adapted to the target business scenario, it can design under what circumstances to stop training the target machine learning model and what strategy to use to determine the hyperparameter configuration of the target machine learning model in the next round, reflecting the control of the process of stopping training in advance or using what strategy to optimize hyperparameter configuration during the machine learning model training process. It utilizes the fact that instead of using the traditional Bayesian optimization method to configure hyperparameters after each round of training, training is stopped in advance or other methods are used to optimize hyperparameter configuration. This can realize the intelligent transformation of the traditional Bayesian hyperparameter optimization method, reducing the amount of computation, improving the model convergence speed, and adapting to the needs of specific business scenarios. Attached Figure Description

[0034] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is an exemplary system architecture diagram to which one embodiment of this disclosure can be applied; Figure 2 This is a flowchart of an embodiment of the multi-agent model automatic training method according to the present disclosure; Figure 3 This is a schematic diagram of the agent interaction logic according to an embodiment of steps 203 and 204 of this disclosure; Figure 4 This is a schematic diagram of a structure of an embodiment of the multi-agent model automatic training device according to the present disclosure; Figure 5 This is a flowchart of an embodiment of the multi-agent model application method according to the present disclosure; Figure 6 This is a schematic diagram of a structure of an embodiment of a multi-agent model application device according to the present disclosure; Figure 7 This is a schematic diagram of the structure of a computer system suitable for implementing embodiments of the present disclosure. Detailed Implementation

[0035] The present disclosure will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0036] It should be noted that the user data, data acquisition, and / or use involved in the embodiments of this disclosure strictly comply with the laws, regulations, and industry standards of relevant countries and regions. The collection and acquisition of data involved in the embodiments of this disclosure are all done in advance by actively prompting or prominently displaying information to inform users and obtaining authorization, or by obtaining full authorization from all parties. The processing, manipulation, forwarding, and use of data involved in the embodiments of this disclosure are all carried out on the premise that the user or relevant party is fully informed and authorized. When implementing the various embodiments of this disclosure, the types of data or information, scope of use, and usage scenarios that may be involved are informed to users or relevant parties and authorization is obtained through appropriate means. The specific methods of notification and authorization may vary according to the actual situation, and this disclosure is not limited in this regard. The processing of personal information involved in the embodiments of this disclosure is carried out under the premise of having a legal basis (such as obtaining the consent of the personal information subject or being necessary for the performance of a contract), and is only processed within the prescribed or agreed scope. Sensitive personal information such as biometric information, medical and health information, financial account information, and precise location information involved in the embodiments of this disclosure are all processed under the premise of having a specific purpose and sufficient necessity, and with the separate authorization and consent of the user or relevant party. In this embodiment of the disclosure, if a user or related party refuses to process personal information other than the information necessary for the basic functions, it will not affect the use of the basic functions of this embodiment of the disclosure.

[0037] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0038] Figure 1An exemplary system architecture 100 is shown, in which embodiments of the multi-agent model automatic training method, apparatus, electronic device and storage medium of this disclosure can be applied.

[0039] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0040] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as large language model applications, machine learning applications, natural language processing applications, speech recognition applications, short video social applications, audio and video conferencing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0041] Terminal devices 101, 102, and 103 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices with information input and output devices, including but not limited to smartphones, tablets, laptops, and desktop computers. When terminal devices 101, 102, and 103 are software, they can be installed in the terminal devices listed above. They can be implemented as multiple software programs or software modules (e.g., to provide automatic training services for models combining multiple agents) or as a single software program or software module. No specific limitations are imposed here.

[0042] In some cases, the multi-agent model automatic training method provided in this disclosure can be executed by terminal devices 101, 102, and 103, and correspondingly, the multi-agent model automatic training device can be set in terminal devices 101, 102, and 103. In this case, the system architecture 100 may not include server 105.

[0043] In some cases, the multi-agent model automatic training method provided in this disclosure can be jointly executed by terminal devices 101, 102, 103 and server 105. For example, the step of "generating the current hyperparameter configuration using the current agent model based on the historical hyperparameter tuning record set and the current hyperparameter search space of the target machine learning model" can be executed by terminal devices 101, 102, 103, and the steps of "performing the following training and hyperparameter tuning record generation operations" can be executed by server 105. This disclosure does not limit this. Correspondingly, the multi-agent model automatic training device can also be respectively set in terminal devices 101, 102, 103 and server 105.

[0044] In some cases, the multi-agent model automatic training method provided in this disclosure can be executed by server 105. Correspondingly, the multi-agent model automatic training device can also be set in server 105. In this case, the system architecture 100 may not include terminal devices 101, 102, and 103.

[0045] It should be noted that server 105 can be either hardware or software. When server 105 is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When server 105 is software, it can be implemented as multiple software programs or software modules (for example, used to provide distributed services), or as a single software program or software module. No specific limitations are made here.

[0046] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0047] Continue to refer to Figure 2 The diagram illustrates a flow 200 of an embodiment of an automatic training method for a multi-agent model according to the present disclosure, which includes the following steps: Step 201: Based on the historical hyperparameter tuning record set and the current hyperparameter search space of the target machine learning model, generate the current hyperparameter configuration using the current surrogate model.

[0048] In this embodiment, the execution subject of the multi-agent model automatic training method (e.g.) Figure 1 The server shown can generate the current hyperparameter configuration based on the historical hyperparameter tuning record set and the current hyperparameter search space of the target machine learning model, using the current proxy model. In some embodiments, the historical hyperparameter tuning record set can be a local historical hyperparameter tuning record set of the current proxy model, where the current proxy model refers to the proxy model in the current round. The input is given to the current surrogate model (e.g., a Gaussian process). The current surrogate model can provide a complete probability distribution, such as a Gaussian distribution, for any unknown point x in the hyperparameter space, including: Mean μ(x): The most likely performance score predicted by the current agent model.

[0049] Variance σ²(x): The uncertainty of the prediction.

[0050] From a technical classification perspective, the target machine learning model can be any type of machine learning model, and this disclosure does not impose any specific limitations on it. For example, it may include, but is not limited to, decision tree models (such as random forest, XGBoost, LightGBM, CatBoost), neural network models (such as multilayer perceptron MLP, convolutional neural network CNN, sequence neural network RNN / Transformer), linear models (such as logistic regression models), and other machine learning models.

[0051] In terms of application areas, the target machine learning model can also be one of the various machine learning models shown in Table 1 below: Table 1. Classification of application areas for target machine learning models In some alternative implementations, the target machine learning model can be a risk control model. Here, "risk control model" is a broad concept, referring to a collection of mathematical models and algorithms used to quantify, assess, and predict risk. In terms of model type, a risk control model is not a single, specific model, but rather a system that integrates multiple models with prediction and classification at its core. Common specific model types include: Logistic regression model: A classic and highly interpretable model often used in credit scoring to predict the probability of events (such as the probability of default).

[0052] Decision trees / random forests / gradient boosting trees: very powerful ensemble learning models that can handle complex nonlinear relationships and are widely used in anti-fraud and credit assessment.

[0053] Neural network / deep learning models are used to process extremely complex and high-dimensional data, such as images, text, and sequences of network behaviors, especially in the field of fraud prevention to identify hidden patterns.

[0054] Key characteristics: Risk control models are typically based on supervised or semi-supervised learning paradigms because they require training on a large number of "good and bad" samples with known outcomes (such as "normal users" and "fraudulent users") to learn to predict the risk of new samples.

[0055] The input data for risk control models is extremely extensive and complex, aiming to depict a profile of an entity (such as an individual, enterprise, or transaction) from multiple dimensions. The data can generally be categorized into the following main types: 1. Basic Information of the Main Entity Individual: age, gender, education level, occupation, income level, marital status, and residential stability.

[0056] Company: Industry, registered capital, years of establishment, business address, shareholder structure.

[0057] 2. Historical application data on resource borrowing and behavior Resource borrowing and return records: number of historical resource borrowing applications, resource return records, number of overdue resource returns, current unreturned resource rate, and approved resource amount.

[0058] Consumption and resource transfer behavior: historical resource transfer value, resource transfer frequency, resource transfer time preference, and frequently visited merchant types.

[0059] 3. Real-time data on application / resource transfer Application Information: Information about the digital resources applied for, such as amount, duration, and purpose.

[0060] Resource transfer information includes: resource transfer value, resource transfer target object information, resource transfer channel, resource transfer time, and resource transfer address. In some embodiments, the resource transfer value may be the transaction amount, the resource transfer target object information may be the payee information, the resource transfer channel may be the transaction channel (such as APP, webpage), the resource transfer time may be the transaction time (whether it is during abnormal periods such as late at night), and the resource transfer address may be the transaction location / IP address (whether it does not match the commonly used address).

[0061] 4. External Data Official credit data: authoritative credit history records.

[0062] Third-party resource application data service providers: provide information related to resource borrowing applications from multiple parties (e.g., data on users applying for resource borrowing on multiple platforms), blacklist information of accounts applying for resource borrowing, judicial enforcement information, consumption capacity scores, etc.

[0063] The output of a risk control model is not a simple "yes" or "no," but a quantitative risk assessment result, primarily in the following forms: 1. Account security This is the most common output format, which outputs a score (e.g., 300-1000 points). The higher the score, the higher the account security.

[0064] 2. Account Abnormality Probability The system directly outputs a probability value between 0 and 1 to indicate that the account is abnormal, such as "The probability that this user's transaction is fraudulent is 0.92 (92%)". Business personnel can set a threshold (such as 0.8), and trigger an alarm or reject the transaction if the threshold is exceeded.

[0065] 3. Account risk level / label Subjects are classified into different risk levels based on scores or probabilities.

[0066] For example: "low risk", "medium risk", "high risk".

[0067] Alternatively: "Premium Customer", "Normal Customer", "Restricted Customer", "Rejected Customer".

[0068] The learning objective of a risk control model is to identify patterns or regularities from massive amounts of historical data that can effectively distinguish between "high-risk" and "low-risk" behaviors.

[0069] Specifically, it learns: High-risk entities' characteristic combination: Learning what kind of personal information, behavioral data, and transaction characteristics, when combined, can significantly increase the probability of an account being an abnormal account.

[0070] The complex relationship between variables and risk: Learn the nonlinear relationships and interactions between each input variable (feature) and the final risk outcome.

[0071] For example, the model learns that "high income" is inherently a low-risk characteristic, but when combined with "extremely high debt ratio," the risk increases dramatically. "Nighttime transactions" alone may not seem risky, but when combined with "new equipment" and "large sums," the risk becomes extremely high.

[0072] Evolving fraud methods: Risk, especially fraud risk, is dynamic. Fraudsters constantly change their strategies. Therefore, risk control models need to be iterated rapidly and frequently, continuously learning from new data to adapt to and identify emerging attack patterns and maintain their predictive capabilities.

[0073] Here, the historical parameter tuning record set can refer to the global historical parameter tuning record set of the target machine learning model, or more preferably, it can refer to the local historical parameter tuning record set of the current proxy model.

[0074] Here, the global history hyperparameter tuning record set of the target machine learning model stores all the hyperparameter configurations and corresponding training and evaluation results that have been used for the target machine learning model.

[0075] The local history hyperparameter tuning record set of the current surrogate model stores: the hyperparameter configurations and corresponding training and evaluation results of all previously used target machine learning models for the current surrogate model, and / or the historical hyperparameter tuning records of the target machine learning model's global history hyperparameter tuning record set belonging to the current hyperparameter search space. Using the current surrogate model and the acquisition function, the next round of hyperparameter configurations can be generated based on the local history hyperparameter tuning record set of the current surrogate model. Specifically, the current surrogate model refers to a probabilistic surrogate model, which can simulate and predict the performance y of any point x in the current hyperparameter search space of the target machine learning model based on the historical tuning record set (e.g., the local history hyperparameter tuning record set of the current surrogate model).

[0076] As an example, step 201 can be performed as follows: First, the local historical parameter tuning records of the current proxy model are collected. The input is given to the current surrogate model (e.g., a Gaussian process). Then, for any unknown point x in the hyperparameter space, the current surrogate model will provide a complete probability distribution, such as a Gaussian distribution, including: Mean μ(x): The most likely performance score predicted by the current agent model.

[0077] Variance σ²(x): The uncertainty of the prediction.

[0078] Then, the acquisition function is used to comprehensively analyze the mean μ and variance σ², and output a set of hyperparameter configurations.

[0079] In simple terms: if a region has a high μ (good prediction performance), the acquisition function will tend to select it (utilizing known good regions). If a region has a large σ² (high uncertainty), the acquisition function will also tend to select it (exploring unknown regions, where surprises may be hidden).

[0080] The most common acquisition function is Expected Improvement (EI), which calculates how much the training evaluation result (i.e., the optimization target) can be expected to improve after trying a new set of parameter configurations compared to the current historical best training evaluation result (i.e., the optimization target). It automatically balances "exploitation" and "exploration," selecting the point with the largest EI value (i.e., the hyperparameter configuration) as the target for the next evaluation (i.e., the current hyperparameter configuration).

[0081] As an example, Table 2 provides some examples of machine learning models involved in hyperparameter optimization using surrogate models, including model types, specific model examples, typical hyperparameters that can be optimized using Bayesian methods, and optimization objectives (the Y-value of the surrogate model).

[0082] Table 2 Typical hyperparameters and optimization objectives of machine learning models Step 202: Perform the training and parameter tuning record generation operation.

[0083] Here, the training and parameter tuning record generation operation may include the following steps 2021 to 2023: Step 2021: Train the target machine learning model according to the current hyperparameter configuration.

[0084] Here, the target machine learning model can be trained according to its current hyperparameter configuration. Specifically, the target machine learning model can be trained according to its specific model category, without any specific limitations in this disclosure.

[0085] Step 2022: Perform performance evaluation on the target machine learning model to obtain training evaluation results.

[0086] Here, the corresponding performance evaluation index (i.e., the Y value of the optimization target and the surrogate model in Table 2) can be used to evaluate the performance of the target machine learning model and obtain the training evaluation results.

[0087] Step 2023: Generate hyperparameter tuning records based on the current hyperparameter configuration and training evaluation results.

[0088] Here, hyperparameter tuning records can be generated using the current hyperparameter configuration of the target machine learning model and the corresponding training evaluation results.

[0089] Optionally, the training and parameter tuning record generation operation may also include the following steps (2024): Step 2024: Add the parameter tuning record to the global historical parameter tuning record set of the target machine learning model and the local historical parameter tuning record set of the current proxy model.

[0090] Here, the parameter tuning records generated in step 2023 can be added to the global historical parameter tuning record set of the target machine learning model and the local historical parameter tuning record set of the current proxy model.

[0091] Here, the global history hyperparameter tuning record set of the target machine learning model stores all the hyperparameter configurations and corresponding training and evaluation results that have been used for the target machine learning model.

[0092] The local history hyperparameter tuning record set of the current proxy model stores the hyperparameter configurations and corresponding training evaluation results of all the target machine learning models that have been used for the current proxy model.

[0093] It should be noted that different surrogate models differ in their modeling assumptions and structures. For example, the differences among Gaussian process surrogate models can be reflected in the kernel function type (e.g., RBF vs. Matérn), kernel hyperparameters (e.g., length scale), mean function, and noise level settings; while tree-based surrogate models (e.g., random forests) differ through splitting criteria and ensemble strategies; and TPE-type surrogate models rely on prior and likelihood probability density estimation methods. Furthermore, different surrogate models also vary in their adaptability to the hyperparameter search space (e.g., support for continuous, discrete, or conditional variables).

[0094] Step 2024 allows the parameter tuning records generated in step 2023 to be added to the global historical parameter tuning record set of the target machine learning model and the local historical parameter tuning record set of the current proxy model.

[0095] Adding the hyperparameter tuning records generated in step 2023 to the global historical hyperparameter tuning record set of the target machine learning model allows for the selection of historical hyperparameter tuning records from the global historical hyperparameter tuning record set that belong to the hyperparameter search space of the new current surrogate model after the surrogate model changes and is reconstructed into a new current surrogate model. The selected historical hyperparameter tuning records are then used as the local historical hyperparameter tuning record set of the new current surrogate model. Subsequently, based on the local historical hyperparameter tuning record set of the current surrogate model and the current hyperparameter search space of the target machine learning model, the current hyperparameter configuration is generated using the current surrogate model.

[0096] The local historical parameter tuning record set of the current surrogate model is the training data of the current surrogate model and is the basis for the model to fit the optimization objective. The goal of the surrogate model is to continuously approximate the true optimization objective (such as validation loss), and each new parameter tuning record (i.e., observation data or training data) is a "sampling" of the optimization objective. Adding the parameter tuning record generated in step 2023 to the local historical parameter tuning record set of the current surrogate model can achieve technical effects including but not limited to the following: First, improve the accuracy of the current surrogate model. Specifically, each newly generated parameter tuning record, as observation data, provides the Y value of the true optimization target at a certain point x, allowing the current surrogate model to more accurately approximate the global or local structure of the objective function in the hyperparameter space, especially in regions with excellent performance (such as the neighborhood of the currently known optimal configuration), thereby enhancing its modeling ability and prediction accuracy for high-potential regions.

[0097] Second, it guides the next round of exploration / utilization towards a better balance, reducing the duplication of sampling of the same or similar points. Specifically, since the calculation of the acquisition function depends on the predicted mean and uncertainty of the surrogate model output, adding new observation data can reduce the prediction uncertainty of the already evaluated areas, thereby guiding the acquisition function to shift its search focus to more promising, underexplored areas, avoiding invalid sampling in low-performance or converged areas, and thus avoiding wasting subsequent computational resources.

[0098] Third, it achieves "sequential decision making." Specifically, Bayesian optimization is essentially a sequential experimental design process: each step makes the optimal choice for the next step based on existing knowledge, and then updates the knowledge with the new results. If no new observation data is added, it's equivalent to the surrogate model not learning or optimizing, leading to blind random search.

[0099] Step 203: Using the decision agent corresponding to the target business scenario to select the execution agent, determine the next round hyperparameter optimization execution agent to control the next round hyperparameter optimization scheme of the target machine learning model.

[0100] Here, the decision agent is used to determine the agent identifier of the next round of hyperparameter optimization execution agent according to the needs of specific business scenarios. It can be, for example, a rule-based agent, a basic heuristic algorithm-based agent, a reinforcement learning-based agent, or a large language model-based agent.

[0101] In some embodiments, the decision-making agent is a rule-based decision-making agent, which refers to a decision-making agent whose next round of hyperparameter optimization execution agent is determined in advance by professional technicians according to the needs of the specific target business scenario. Based on the aforementioned next-round hyperparameter optimization execution agent determination rules, a rule-based decision-making agent is generated. Then, in step 203, the next-round hyperparameter optimization execution agent, which controls the next-round hyperparameter optimization scheme of the target machine learning model, can be determined by calling the aforementioned rule-based decision-making agent. Here, the next-round hyperparameter optimization execution agent determination rules corresponding to the target business scenario are used to specify under what conditions what type of next-round hyperparameter optimization execution agent is triggered.

[0102] In other embodiments, the decision agent is a large-model-based decision agent, meaning a large-model-based decision agent designed for a specific business scenario, or a general large-model-based decision agent can be designed. The large-model-based decision agent is used to determine the next round of hyperparameter optimization execution agent based on the local historical parameter tuning records of the current agent model or the global parameter tuning records of the target machine learning model. Furthermore, in step 203, the aforementioned large-model-based decision agent can be invoked to determine the next round of hyperparameter optimization execution agent that controls the next round of hyperparameter optimization scheme for the target machine learning model.

[0103] The target business scenario is only a business scenario for the sake of convenience, and is not limited to a specific business scenario.

[0104] It should be noted that, as a software entity, the decision-making agent is not specifically limited in its final form or invocation method in this disclosure.

[0105] Simply put, an intelligent agent is a "brain" or "engine" that needs to interact with users and the world through different "bodies" or "interfaces".

[0106] As an example, deployment can be achieved through application programming interfaces (APIs), web services, standalone applications, or background processes / services.

[0107] In some alternative implementations, step 203 may be performed as follows: Based on the preset probability or preset round frequency corresponding to each hyperparameter optimization agent, determine the next round of hyperparameter optimization agent to train the target machine learning model.

[0108] In other words, N types of hyperparameter optimization execution agents can be pre-designed based on the needs of the target business scenario, where N is a positive integer. Then, a corresponding preset probability or preset round frequency can be assigned to each type of hyperparameter optimization execution agent.

[0109] As an example, suppose there are four hyperparameter optimization agents: AG1, AG2, AG3, and AG4. Preset probabilities Pro1, Pro2, Pro3, and Pro4 can be assigned to AG1, AG2, AG3, and AG4, respectively. Then, step 2031 can be to determine the next round of hyperparameter optimization agents for training the target machine learning model as AG1, AG2, AG3, and AG4, respectively, according to probabilities Pro1, Pro2, Pro3, and Pro4.

[0110] As an example, suppose there are four hyperparameter optimization agents: AG1, AG2, AG3, and AG4. Preset round frequencies can be assigned to AG1, AG2, AG3, and AG4, respectively: AG1 is executed M1 times every R1 round, AG2 is executed M2 times every R2 round, AG3 is executed M3 times every R3 round, and AG4 is executed M4 times every R4 round. Then, step 2031 could be: every R1 round, determine the next hyperparameter optimization agent for training the target machine learning model as AG1; every R2 round, determine AG2; every R3 round, determine AG3; and every R4 round, determine AG4.

[0111] By employing the above optional implementation methods, it is possible to customize the probability or frequency of each hyperparameter optimization agent controlling the next round of hyperparameter optimization for specific business scenarios. Because it is adapted to specific business scenarios, it can improve model training time efficiency and reduce the number of model training rounds in the corresponding business scenarios.

[0112] In some alternative implementations, step 203 may also be performed as follows: Based on a global historical hyperparameter tuning record set, and utilizing a decision-making agent based on a large model, a hyperparameter optimization execution agent is determined to control the next round of hyperparameter optimization scheme for the target machine learning model.

[0113] Here, the decision agent based on the large model can first generate prompt words based on the global historical hyperparameter tuning record set. The prompt words are used to instruct the large model to analyze the training evaluation result trend information corresponding to the hyperparameter configuration based on the global historical hyperparameter tuning record, with a focus on the most recently trained global historical hyperparameter tuning record. Then, it selects the hyperparameter optimization execution agent for the next round from the preset N hyperparameter optimization execution agents and outputs the agent identifier of the selected hyperparameter optimization execution agent.

[0114] The optional implementation of step 203 can reflect the trend of training evaluation results corresponding to historical hyperparameter configurations analyzed by the large model based on prompt words, and then select the agent to perform hyperparameter optimization in the next round. It utilizes the intelligent decision-making ability of the large model, and can realize the real-time and intelligent determination of the agent to perform hyperparameter optimization in the next round based on historical parameter tuning records, thereby improving the model training efficiency and avoiding problems such as overfitting or slow model convergence speed.

[0115] Step 204: Based on the generated hyperparameter tuning record, the agent performs corresponding control operations using the next round of hyperparameter optimization to stop training the target machine learning model or determine the next round of hyperparameter configuration for the target machine learning model, and train the target machine learning model according to the next round of hyperparameter configuration.

[0116] In existing Bayesian hyperparameter optimization methods, after each round of training and performance evaluation using a particular hyperparameter configuration, a Bayesian optimization method (surrogate model and acquisition function) is used to generate the next round of hyperparameter configuration based on historical tuning records. This process requires manual input of the hyperparameter search space based on experience, necessitating human-machine collaboration. However, by employing steps 203 and 204, the appropriate next round of hyperparameter configuration can be selected based on the specific business scenario. Since this method is customized for the specific business scenario, it achieves fully automated searching without human intervention.

[0117] In some alternative implementations, the training and parameter tuning record generation operation may further include the following steps 2025: Step 2025: Based on the local historical hyperparameter tuning record set of the current surrogate model and the current hyperparameter search space of the target machine learning model, generate the original hyperparameter configuration for the next round using the current surrogate model.

[0118] Here, it can refer to using the hyperparameter configuration of the target machine learning model as the original hyperparameter configuration for the next round, based on the local historical parameter tuning record set of the current surrogate model and the acquisition function, while keeping the hyperparameter search space, kernel function type, kernel function parameters, mean function and noise level assumptions or estimates of the current surrogate model unchanged. That is, using the hyperparameter configuration generated by the current surrogate model for the next round.

[0119] Here, the specific steps for generating the hyperparameter configuration of the target machine learning model using the current proxy model and the acquisition function can be found in the relevant record of step 201, and will not be repeated here.

[0120] By adopting the optional implementation of step 2025 above, it is possible to generate the original hyperparameter configuration for the next round in real time using the current agent model, so that the agent executing the next round hyperparameter optimization can use the original hyperparameter configuration for the next round.

[0121] In some alternative implementations, step 2025 may be performed as follows: First, the local historical parameter tuning records belonging to the current hyperparameter search space in the local historical parameter tuning records of the current proxy model are identified as the local historical parameter tuning records of the current space.

[0122] Finally, based on the local historical parameter tuning records of each current space and the current hyperparameter search space, the original hyperparameter configuration for the next round is generated using the current surrogate model.

[0123] By adopting the optional implementation of step 2025 above, when reusing historical parameter tuning records to construct a probabilistic surrogate model (or surrogate model), the historical parameter tuning records and the output hyperparameter configurations share the same hyperparameter search space, thereby improving the generation efficiency of generating the original next round hyperparameter configurations.

[0124] In some optional implementations, the hyperparameter optimization execution agent determined in step 203 can be any one of the following: the first hyperparameter optimization execution agent, the second hyperparameter optimization execution agent, the third hyperparameter optimization execution agent, and the fourth hyperparameter optimization execution agent. Wherein: The first hyperparameter optimization execution agent is used to continue training with the original hyperparameter configuration for the next round.

[0125] Here, the first hyperparameter optimization execution agent can be a rule-based agent.

[0126] The second hyperparameter optimizer is used to evaluate whether to stop training.

[0127] Here, the second hyperparameter optimization agent can be a rule-based agent or a large model-based agent. The third hyperparameter optimization agent is used to instruct whether to change the hyperparameter search space and continue training for the first preset number of rounds.

[0128] Here, the third hyperparameter optimization agent can be an agent based on a large model.

[0129] The fourth hyperparameter optimization execution agent is used to adopt other hyperparameter optimization schemes instead of the original hyperparameter configuration in the subsequent second preset number of hyperparameter optimization rounds.

[0130] Here, the fourth hyperparameter optimization agent can be an agent based on a large model.

[0131] In the above optional embodiments, the original hyperparameter optimization configuration for the next round is not used for training every time. Instead, different hyperparameter optimization execution agents are used to control the hyperparameter optimization scheme for the next round according to different situations. For example, the second hyperparameter optimization execution agent can stop training, thereby reducing the number of times the model is trained using the original hyperparameter configuration generated by the current surrogate model. Another example is that the third hyperparameter optimization execution agent can replace the hyperparameter search space with a more reasonable one, which can reduce the search space for the original hyperparameter configuration generated by the surrogate model in the next round and reduce the amount of computation. Yet another example is that the fourth hyperparameter optimization execution agent can use other hyperparameter optimization schemes instead of the original hyperparameter configuration in the subsequent second preset number of hyperparameter optimization rounds, thereby reducing the amount of computation of traditional Bayesian optimization.

[0132] Accordingly, if the hyperparameter optimization agent determined in step 203 is the first hyperparameter optimization agent, then step 204 can be executed as follows: perform the original hyperparameter optimization and training operations. Specifically, this can be achieved by first executing step 2041, and then proceeding to step 202 to perform the original hyperparameter optimization and training operations. Step 2041: Update the current hyperparameter configuration to the original hyperparameter configuration for the next round.

[0133] After completing step 2041, proceed to step 202.

[0134] Specifically, this refers to updating the current hyperparameter configuration to the original next-round hyperparameter configuration generated by the current surrogate model, based on the local historical hyperparameter tuning record set of the current surrogate model and the current hyperparameter search space of the target machine learning model, as described in step 2025. Then, step 202 is executed again, followed by steps 203 and 204.

[0135] In other words, it can be understood that the first hyperparameter optimization execution agent is equivalent to determining the hyperparameter configuration for the next round according to the existing Bayesian optimization method, without controlling the existing Bayesian optimization method in any way. Optionally, when step 2031 determines the hyperparameter optimization execution agent for training the target machine learning model according to the preset probabilities Pro1, Pro2, Pro3, and Pro4 corresponding to the first, second, third, and fourth hyperparameter optimization execution agents, Pro1 can be the maximum value in the above. As an example, Pro1 can be 95%, that is, the existing Bayesian optimization method is used to determine the hyperparameter configuration for the next round under the probability Pro1, without controlling the existing Bayesian optimization method.

[0136] By adopting the above optional implementation methods, it is possible to determine the hyperparameter configuration for the next round using the traditional Bayesian optimization method in some cases (for example, in most cases), thereby achieving hyperparameter optimization and adjustment.

[0137] In some optional implementations, if the hyperparameter optimization execution agent determined in step 203 is the second hyperparameter optimization execution agent, then step 204 may include the following steps 2042 and 2043: Step 2042: Determine whether to stop training based on the global historical parameter tuning record set.

[0138] Here, various implementation methods can be used to determine whether to stop training based on the global historical parameter tuning record set.

[0139] As an example, a training stopping method based on simple rules can be used.

[0140] Specifically, different types of rules can be used to determine whether to stop training. For example, the following types of rules can be used to determine whether to stop training: 1. Performance convergence judgment rule: If the improvement of the best value of the validation set performance in the most recent M iterations is less than a threshold ε, then training can be stopped.

[0141] 2. Resource Exhaustion Determination Rules: Training will stop if at least one of the following resource exhaustion conditions is met. The resource exhaustion conditions include: reaching the maximum number of tuning rounds (max_trials), reaching the maximum total training time (max_wall_time), and reaching the maximum computational budget (e.g., CPU hours).

[0142] 3. Performance Target Determination: Training can stop once a model that meets business requirements has been found. Specifically, training stops if the performance evaluation on the validation set in any round exceeds the preset target threshold T.

[0143] As another example, a large model can also determine whether to stop training based on a global historical parameter tuning record set. Specifically, a prompt word can be generated based on the global historical parameter tuning record set and input into the large model. The large model analyzes the recent training evaluation results trend based on the global historical parameter tuning record set, focusing on recent historical parameter tuning records, and combines this with the training time already taken for the target machine learning model to determine whether to stop training, and outputs an indication message indicating whether to stop training.

[0144] If it is determined to stop training in step 2042, proceed to step 2043.

[0145] Step 2043: Persistently save the model parameters of the target machine learning model.

[0146] That is, if it is determined in step 2042 to stop training the target machine learning model, the model parameters of the target machine learning model can be persistently saved.

[0147] If it is determined in step 2042 that training should not be stopped, the original hyperparameter optimization and training operations can be performed. That is, the process can proceed to step 2041.

[0148] Specifically, you can first execute step 2041, and then proceed to execute steps 202, 203 and 204.

[0149] Specifically, this refers to updating the current hyperparameter configuration to the original next-round hyperparameter configuration generated by the current surrogate model, based on the local historical hyperparameter tuning record set of the current surrogate model and the current hyperparameter search space of the target machine learning model, as described in step 2025. Then, step 202 is executed again, followed by steps 203 and 204.

[0150] By adopting the above optional implementation methods, model training can be stopped early, reducing unnecessary model training times and improving model training efficiency.

[0151] In some optional implementations, if the hyperparameter optimization execution agent determined in step 203 is the third hyperparameter optimization execution agent, then step 204 may include the following steps 2044 to 2048: Step 2044: Using the large model, determine whether to change the hyperparameter search space based on the global historical hyperparameter tuning record set and the current hyperparameter search space, and generate a new hyperparameter search space after determining to change the hyperparameter search space.

[0152] Here, prompt words can be generated and input into the large model, so that the large model can determine whether to change the hyperparameter search space based on the global historical hyperparameter tuning record set of the target machine learning model and the current hyperparameter search space, and generate a new hyperparameter search space after determining to change the hyperparameter search space.

[0153] Specifically, by using pre-set prompts, large models can be guided to analyze and reason about global historical hyperparameter tuning records based on thought chains, thereby dynamically and adaptively adjusting the search space of hyperparameters.

[0154] As an example, the following is a specific implementation of step 2044, including the following steps: S11: Construct prompt word templates to guide large models in search space optimization.

[0155] This template is crucial for the technical implementation, and its structure may include, for example: (1) Role definition section: clearly instruct the large model to play the role of a "machine learning hyperparameter optimization expert".

[0156] (2) Task context and input data section: Clearly describe the background of the hyperparameter optimization task to the large model and provide structured input data.

[0157] The input data includes at least a global historical hyperparameter tuning record set, the current hyperparameter search space, and a hyperparameter tuning experience guide. The hyperparameter tuning experience guide is optional text information that contains domain-related hyperparameter tuning experience, prior knowledge, or constraints (e.g., "learning rate is usually searched in logarithmic space" or "increasing model complexity parameters may exacerbate overfitting").

[0158] (3) Reasoning Instructions for the Thinking Chain: Clear instructions require the large model to use a thinking chain-based approach, performing reasoning and analysis step by step. Specific instructions may include, for example: A. Performance Range Analysis: "First, carefully analyze the provided historical hyperparameter tuning records. Identify which hyperparameter configurations or hyperparameter search ranges tend to produce the best model performance (e.g., high accuracy, low loss). At the same time, identify which configurations or ranges lead to model overfitting (validation set performance is much worse than training set performance) or underfitting (both training and validation set performance is poor)." B. Search Space Rationality Assessment: "Next, based on the above analysis, assess whether the current search space definition is reasonable. Determine whether the current range of each hyperparameter is too wide, too narrow, or fails to cover the potential optimal region. Please provide a list of hyperparameters that need to have their range narrowed, expanded, or remain unchanged, along with the reasons." C. Output Format Specification: Large models are strictly required to output their inference conclusions and new search space definitions in a specified, structured format. For example, the output should be a JSON object containing fields such as "analysis_summary" (analysis summary in text form), "reasoning_for_adjustment" (reasons for each hyperparameter adjustment), and "new_search_space" (the newly defined hyperparameter search space).

[0159] S12: The complete prompt words, populated with specific data (global historical parameter tuning record set, current search space, etc.) based on the template in S1, are input into a pre-trained large language model. This large language model needs to possess the following characteristics: Powerful natural language understanding and reasoning capabilities: capable of understanding complex task instructions and structured data contexts. Preferably, general-purpose large language models with hundreds of billions of parameters are used, such as the GPT-4 series, Claude series, and Qwen series.

[0160] Ability to follow instructions and output in a structured manner: Able to strictly follow the steps in the prompts and output results in the required format.

[0161] A certain level of numerical analysis and logical reasoning ability: able to identify trends and patterns from numerical historical records.

[0162] S13: The large model receives and processes cue words, internally executing a thought-chain-based reasoning process. It first parses the historical hyperparameter tuning record set to identify high-performance and problematic hyperparameter search ranges; then, combining tuning experience guidelines, it evaluates the rationality of the current hyperparameter search space; finally, it generates a structured output containing a detailed analysis report and a new, optimized definition of the hyperparameter search space.

[0163] S14: Analyze the output of the large model and extract the structured "new hyperparameter search space definition". Apply this new hyperparameter search space to the next round or stage of the hyperparameter optimization process (e.g., as a new hyperparameter configuration boundary for the Bayesian optimization algorithm).

[0164] S15: Optionally, this process can be iterative. After several rounds of hyperparameter evaluation in the new search space, new historical parameter tuning records are collected, and the process from S1 to S4 is triggered again to achieve continuous dynamic optimization of the search space.

[0165] The optional implementation of step 2044 above can achieve, but is not limited to, the following technical effects: Intelligentization and Automation: The powerful reasoning capabilities of large models are introduced into the hyperparameter optimization loop, realizing the automation of hyperparameter search space adjustment and reducing the reliance on the manual experience of domain experts.

[0166] Dynamic adaptability: It can dynamically adjust the hyperparameter search space based on feedback from historical optimization, making the optimization process more focused on potential regions, thereby accelerating the convergence speed and increasing the probability of finding a better configuration.

[0167] Explainability: The thought chain-based analysis process makes the search space adjustment decisions more explainable, and users can understand the reasons for the adjustment through the analysis summary generated by the model.

[0168] Flexibility: By modifying the "parameter tuning experience guide" in the prompt word template, prior knowledge from different fields or specific tasks can be easily incorporated, making the method highly flexible.

[0169] Step 2045: Determine whether the new hyperparameter search space has been successfully generated.

[0170] If the new hyperparameter search space is successfully generated, then proceed to steps 2046 and 2047.

[0171] Step 2046: Update the current hyperparameter search space of the target machine learning model to the new hyperparameter search space, clear the local hyperparameter tuning record set of the current proxy model, and add the local historical hyperparameter tuning records belonging to the current hyperparameter search space from the global historical hyperparameter tuning record set to the local hyperparameter tuning record set of the current proxy model.

[0172] Here, by clearing the local parameter tuning record set of the current proxy model and adding the local historical parameter tuning records belonging to the current hyperparameter search space from the global historical parameter tuning record set to the local parameter tuning record set of the current proxy model, the local parameter tuning record set of the current proxy model is made up of local parameter tuning records belonging to the current hyperparameter search space. This can reduce the search range of subsequent proxy models and improve computational efficiency.

[0173] Step 2047: Based on the local historical parameter tuning record set of the current surrogate model and the current hyperparameter search space, generate the original hyperparameter configuration for the next round using the current surrogate model.

[0174] Here, the specific operation of step 2047 can be referred to the relevant record of step 2025 above, and will not be repeated here.

[0175] Step 2048: Continuously execute the first preset number of rounds of original hyperparameter optimization and training operations.

[0176] That is, the original hyperparameter optimization and training operations are continuously executed for a first preset number of rounds. Each original hyperparameter optimization and training operation is executed as follows: first execute step 2041, then execute step 202. After executing step 202 in each round, the process does not proceed to step 203, but directly executes the next round's process of first executing step 2041, then executing step 202, until the first preset number of such operations are completed.

[0177] In other words, after changing the hyperparameter search space, the existing Bayesian optimization method needs to be used to continuously perform the first preset number of rounds (e.g., 200 rounds) of hyperparameter optimization, and then proceed to step 203 to determine the agent to perform the next round of hyperparameter optimization.

[0178] In some optional implementations, if the hyperparameter optimization execution agent determined in step 203 is the fourth hyperparameter optimization execution agent, then step 204 may include the following step 2049: Step 2049: Execute the second preset number of rounds of hyperparameter optimization recommendation operation continuously.

[0179] Here, the recommended hyperparameter optimization operation may include the following steps 20491 and 20492: Step 20491: Determine the recommended hyperparameter configuration for the target machine learning model.

[0180] Here, various implementation methods can be used to determine the recommended hyperparameter configuration for the target machine learning model.

[0181] As an example, one can first identify similar machine learning models that are similar to the type of task of the target machine learning model.

[0182] Then, based on similarity recall and historical performance evaluation using similar machine learning models, the hyperparameter configuration is obtained.

[0183] Finally, the obtained hyperparameter configuration is determined as the recommended hyperparameter configuration.

[0184] In other words, other hyperparameter configurations that are similar to the target machine learning model and have good performance verified by historical hyperparameter tuning records are used as recommended hyperparameter configurations.

[0185] As another example, recommended hyperparameter configurations can also be generated using a large model based on a global historical hyperparameter tuning record set.

[0186] Specifically, prompt words can be generated first and then input into a large model. The large model can then use this prompt word to generate recommended hyperparameter configurations based on a global historical hyperparameter tuning record set. In other words, through a carefully designed prompt word project, the large model is guided to perform in-depth analysis based on historical hyperparameter tuning records and directly output recommended hyperparameter configurations.

[0187] As an example, the following is a specific implementation method, including the following steps: The system guides a large language model to perform a structured analysis of historical hyperparameter tuning records using pre-defined prompt templates. Based on thought chain reasoning, it generates optimization suggestions and recommended hyperparameter configurations for the next round. The template defines the model's role, input data format, analysis instructions, and structured output specifications. The large model then outputs a response containing analysis conclusions and recommended hyperparameter configurations in JSON format.

[0188] Step 20492: Update the current hyperparameter configuration of the target machine learning model to the recommended hyperparameter configuration.

[0189] That is, after executing step 20492, proceed to step 202 to perform training and parameter tuning record saving operations, and after executing step 202, continue to execute step 203 to determine the next round of hyperparameter optimization execution agent.

[0190] By executing the fourth hyperparameter optimization agent, a round of training can be performed according to the recommended hyperparameter configuration, and the hyperparameter tuning record can be recorded before determining the hyperparameter optimization agent for the next round.

[0191] In some optional implementations, step 203 may be performed as follows: by changing the probability or changing the round frequency according to the preset hyperparameter search space, the next round hyperparameter optimization execution agent for training the target machine learning model is determined as the fifth hyperparameter optimization execution agent based on the large model.

[0192] Here, we assume the preset hyperparameter search space change probability is Pasp, where Pasp can be a pre-set value greater than 0 and less than 1 based on the needs of a specific business scenario. That is, during the execution of step 203 after step 202, a probability of Pasp will determine the next round of hyperparameter optimization execution agent training the target machine learning model as the fifth hyperparameter optimization execution agent based on the large model. This is to determine whether to change the hyperparameter search space and, if so, generate a new hyperparameter search space. Optionally, the remaining probability (1-Pasp) can determine the next round of hyperparameter optimization execution agent training the target machine learning model as the first hyperparameter optimization execution agent mentioned above. That is, the traditional Bayesian optimization method is still used to determine the next round of hyperparameter configuration.

[0193] Here, the hyperparameter search space is pre-defined to change every Rasp epochs consecutively every Masp epochs. The agent responsible for hyperparameter optimization in the next epoch of training the target machine learning model is designated as the fifth hyperparameter optimization agent. For example, the agent responsible for hyperparameter optimization in the next epoch of training the target machine learning model is designated as the fifth hyperparameter optimization agent every 200 epochs consecutively. That is, after changing the hyperparameter search space once, the same hyperparameter search space is used for the subsequent 200 training epochs. The Bayesian optimization method is used to generate the hyperparameter configuration for the next epoch, and training is performed using the generated hyperparameter configuration. The hyperparameter configuration and training evaluation results are then recorded.

[0194] Accordingly, step 204 can be performed as follows: in response to the next round of hyperparameter optimization execution agent being the fifth hyperparameter optimization execution agent, perform the training operation after changing the hyperparameter search space.

[0195] Here, the training operation after changing the hyperparameter search space can include the following steps one through four: The first step is to use a large model to determine whether to change the hyperparameter search space based on the global historical hyperparameter tuning record set and the current hyperparameter search space, and to generate a new hyperparameter search space after determining to change the hyperparameter search space.

[0196] Here, the specific implementation method and technical effects of the first step are basically the same as those of the specific implementation method and technical effects of the second step 2044, and will not be repeated here.

[0197] The second step is to update the current hyperparameter search space of the target machine learning model to the new hyperparameter search space in response to the successful generation of the new hyperparameter search space, clear the local hyperparameter tuning record set of the current proxy model, and add the local historical hyperparameter tuning records belonging to the current hyperparameter search space from the global historical hyperparameter tuning record set to the local hyperparameter tuning record set of the current proxy model.

[0198] Here, the specific implementation method and technical effects of the first step are basically the same as those of steps 2045 and 2046, and will not be repeated here.

[0199] The third step is to generate the original hyperparameter configuration for the next round based on the local historical parameter tuning record set of the current surrogate model and the current hyperparameter search space.

[0200] Here, the specific implementation method and technical effects of the third step are basically the same as those of the specific implementation method and technical effects of step 2025, and will not be repeated here.

[0201] After completing the third step, proceed to the fourth step.

[0202] The fourth step is to continuously perform the original hyperparameter optimization and training operations a fourth preset number of times.

[0203] By employing the above optional implementation method, it is possible to determine whether to change the hyperparameter search space based on the probability of changing the preset hyperparameter search space or the frequency of changing the preset hyperparameter search space rounds. If a change is determined, a new hyperparameter search space is established, and the next round's hyperparameter configuration is generated using a Bayesian optimization method based on the new hyperparameter search space. Compared to the traditional Bayesian optimization method that does not change the hyperparameter search space in each round, this method can reduce the search space range, thereby improving the computational speed of Bayesian optimization.

[0204] The multi-agent model automatic training method provided in the above embodiments of this disclosure utilizes a decision agent corresponding to the target business scenario to determine the next-round hyperparameter optimization execution agent to control the next-round hyperparameter optimization scheme of the target machine learning model; then, based on the generated parameter tuning records, the next-round hyperparameter optimization execution agent performs corresponding control operations to stop training the target machine learning model or determine the next-round hyperparameter configuration of the target machine learning model, and train the target machine learning model according to the next-round hyperparameter configuration. This intelligent transformation of traditional model training methods based on Bayesian hyperparameter optimization methods can achieve, but is not limited to, the following technical effects: First, reduce the number of model training epochs. The applicant's experience has shown that this can reduce the number of model training epochs by 35%.

[0205] Second, adaptive model training is achieved by automatically determining whether to stop training prematurely, changing the hyperparameter space, and generating hyperparameter configurations through multiple executing agents. This improves both model performance and model construction time efficiency. The applicant's practical experience has demonstrated that this can achieve a 45% improvement in time efficiency.

[0206] Specifically, the applicant conducted a systematic verification based on eight different internal private risk prediction datasets (each dataset contains 100,000 to 120,000 data entries, with 200 to 400 feature dimensions).

[0207] The experiment adopted a controlled testing method: two engineers with more than two years of modeling experience used traditional manual model training methods (LightGBM algorithm combined with Hyperopt hyperparameter optimization tool) and the average KS index of the out-of-time test set (OOT) was used as the evaluation standard.

[0208] Experimental results show that traditional methods require an average of about 1500 rounds of hyperparameter optimization on 8 datasets to achieve a deliverable model performance. The average time for model training is 6.3 hours, and the average time for manual analysis and code adjustment is about 2 hours, for a total average of 8.3 hours.

[0209] The intelligent training system using this solution automatically terminates the optimization process after an average of 900 training rounds on eight datasets. Model training takes an average of 3.8 hours, while agent decision-making and large model inference take an average of 0.6 hours, for a total average of 4.4 hours. In terms of model training performance, compared to traditional methods, the KS index difference is within ±1% on five of the eight datasets, and the average KS index improvement exceeds 3% on three of the datasets.

[0210] Optionally, the multi-agent model automatic training method provided in the above embodiments of this disclosure can be applied to MLOPs (Machine Learning Operations). MLOps is a term for automated machine learning operations, referring to a standardized, automated framework for efficiently managing the development, deployment, monitoring, and continuous iteration of machine learning models. Its essence is the deep integration of data engineering, model development, and operation processes, ensuring the reliable, pipeline-like delivery of AI systems. Through the above-described multi-agent model automatic training method, the technical problem of "how to ensure the continuous availability of online machine learning services without human intervention" can be solved, achieving, but is not limited to, the following technical effects: First, it is unattended and reduces reliance on manual modeling: the next round of hyperparameter optimization execution agent determines the rules or calls the next round of hyperparameter optimization execution agent to determine the next round of hyperparameter optimization execution agent to control the hyperparameter optimization scheme of the target machine learning model. The hyperparameter optimization process does not require human intervention, thus achieving unattended operation and reducing reliance on manual modeling.

[0211] Second, optionally, by using the fourth hyperparameter optimization execution agent, the hyperparameter optimization process in the subsequent second preset number of rounds is achieved by using the agent to continuously control the generation of multiple rounds of hyperparameter configuration instead of the original next round hyperparameter configuration, thereby improving the reliability of the model and ensuring the model construction effect. Third, optionally, the execution agent can be optimized through the third hyperparameter, thereby enabling the agent to modify the hyperparameter space, improving the reliability of the model and ensuring the model construction effect. Fourth, by optimizing the execution agent through the second hyperparameter, training can be automatically stopped or continued, rather than simply stopping early. Instead, the agent makes autonomous decisions on whether to terminate the current training task based on a joint evaluation of training dynamics, resource consumption, and business metrics. This avoids ineffective computation from consuming online system resources and improves the timeliness of model updates.

[0212] Fifth, model training is more stable and autonomous: Specifically, by using the next round of hyperparameter optimization to determine the agent rules or calling the next round of hyperparameter optimization to determine the agent, rather than relying on the human experience of technicians, stable and autonomous model training can be achieved.

[0213] To better understand the technical logic of the key steps 203 and 204 in the above-mentioned automatic training method for multi-agent models, the following references are provided. Figure 3 , Figure 3 A schematic diagram of the agent interaction logic of one embodiment of steps 203 and 204 above is shown.

[0214] like Figure 3 As shown, by utilizing a decision-making agent, step 203 is executed to determine the agent for hyperparameter optimization in the next round.

[0215] If the next round of hyperparameter optimization execution agent is determined to be the first next round of hyperparameter optimization execution agent, switch to the first next round of hyperparameter optimization execution agent (or non-intervention control agent) 302, and perform the action: determine the original next round hyperparameter configuration as the final next round hyperparameter configuration, that is, execute the above step 2041.

[0216] If it is determined that the hyperparameter optimization execution agent for the next round is the second hyperparameter optimization execution agent, then proceed to the second hyperparameter optimization execution agent (or the stop training evaluation agent) 303, and execute the action: determine whether to stop training based on the global historical parameter tuning record set, that is, execute the above steps 2042 and 2043.

[0217] If the hyperparameter optimization execution agent for the next round is determined to be the third hyperparameter optimization execution agent, then proceed to the third hyperparameter optimization execution agent (or hyperparameter search space analysis agent) 304, and perform the following actions: using the large model based on the global historical parameter tuning record set and the current hyperparameter search space to determine whether to change the hyperparameter search space, and generating a new hyperparameter search space after determining to change the hyperparameter search space, which corresponds to executing the above step 2044, and after executing step 2044, continue to execute steps 2045 to 2048.

[0218] If it is determined that the hyperparameter optimization execution agent for the next round is the fourth hyperparameter optimization execution agent, proceed to the fourth hyperparameter optimization execution agent (or hyperparameter configuration generation agent) 305, and perform the action: continuously execute the second preset number of rounds of hyperparameter optimization recommendation operation, that is, execute the above step 2049.

[0219] Further reference Figure 4 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of an automatic model training device combining multiple agents, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0220] like Figure 4As shown, the multi-agent model automatic training device 400 of this embodiment includes: a current hyperparameter configuration generation module 401, a training and parameter tuning record generation module 402, an execution agent determination module 403, and a next round hyperparameter optimization and training module 404. The current hyperparameter configuration generation module 401 is configured to generate the current hyperparameter configuration based on the historical hyperparameter tuning record set and the current hyperparameter search space of the target machine learning model, using the current agent model; the training and hyperparameter tuning record generation module 402 is configured to perform the following training and hyperparameter tuning record generation operations: training the target machine learning model according to the current hyperparameter configuration; performing performance evaluation on the target machine learning model to obtain training evaluation results; generating hyperparameter tuning records based on the current hyperparameter configuration and the training evaluation results; the agent determination module 403 is configured to use the decision agent corresponding to the target business scenario to select the agent, and determine the next round hyperparameter optimization agent to control the next round hyperparameter optimization scheme of the target machine learning model; the next round hyperparameter optimization and training module 404 is configured to, based on the generated hyperparameter tuning records, use the next round hyperparameter optimization agent to perform corresponding control operations to stop training the target machine learning model or determine the next round hyperparameter configuration of the target machine learning model, and train the target machine learning model according to the next round hyperparameter configuration.

[0221] In this embodiment, the specific processing and technical effects of the current hyperparameter configuration generation module 401, training and parameter tuning record generation module 402, execution agent determination module 403, and next round hyperparameter optimization and training module 404 of the multi-agent model automatic training device 400 can be referred to respectively. Figure 2 The relevant descriptions of steps 201, 202, 203 and 204 in the corresponding embodiments will not be repeated here.

[0222] In some optional implementations, the training and parameter tuning record generation operation may further include: The generated parameter tuning records are added to the global historical parameter tuning record set of the target machine learning model and the local historical parameter tuning record set of the current proxy model.

[0223] In some optional implementations, after adding the generated hyperparameter tuning records to the global historical hyperparameter tuning record set of the target machine learning model and the local historical hyperparameter tuning record set of the current proxy model, the training and hyperparameter tuning record generation operation may further include: Based on the local historical parameter tuning record set of the current proxy model and the current hyperparameter search space of the target machine learning model, the original hyperparameter configuration for the next round is generated using the current proxy model.

[0224] In some optional implementations, the step of generating the original hyperparameter configuration for the next round using the current proxy model, based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space of the target machine learning model, may include: The local historical parameter tuning records belonging to the current hyperparameter search space in the local historical parameter tuning records of the current proxy model are determined as the local historical parameter tuning records of the current space; Based on the current local historical parameter tuning records and the current hyperparameter search space, the original hyperparameter configuration for the next round is generated using the current surrogate model.

[0225] In some optional implementations, the next round of hyperparameter optimization execution agent is any of the following hyperparameter optimization execution agents: The first hyperparameter optimization execution agent is used to continue training with the original next round hyperparameter configuration; The second hyperparameter optimization execution agent is used to evaluate whether to stop training; The third hyperparameter optimization execution agent is used to instruct whether to change the hyperparameter search space and continue training for the first preset number of rounds. The fourth hyperparameter optimization execution agent is used to adopt other hyperparameter optimization schemes instead of the original next-round hyperparameter configuration during the subsequent second preset number of hyperparameter optimization rounds.

[0226] In some alternative implementations, the agent determination module 403 may be further configured as follows: Based on the preset probability or preset round frequency corresponding to each hyperparameter optimization agent, determine the next round of hyperparameter optimization agent for training the target machine learning model; or Based on the global historical parameter tuning record set, a large model is used to determine the next round of hyperparameter optimization execution agent that controls the next round of hyperparameter optimization scheme for the target machine learning model.

[0227] In some optional implementations, the next round of hyperparameter optimization and training module 404 can be further configured as follows: In response to the next round of hyperparameter optimization execution agent being the first next round of hyperparameter optimization execution agent, the following original hyperparameter optimization and training operations are performed: updating the current hyperparameter configuration to the original next round hyperparameter configuration, and performing the training and hyperparameter tuning record generation operation.

[0228] In some optional implementations, the next round of hyperparameter optimization and training module 404 can be further configured as follows: In response to the next round of hyperparameter optimization execution agent being the second hyperparameter optimization execution agent, a determination is made on whether to stop training based on the global historical hyperparameter tuning record set; In response to determining to stop training, the model parameters of the target machine learning model are persistently saved; In response to the determination not to stop training, the original hyperparameter optimization and training operations are performed.

[0229] In some optional implementations, the next round of hyperparameter optimization and training module 404 can be further configured as follows: In response to the next round of hyperparameter optimization execution agent being the third hyperparameter optimization execution agent, the large model is used to determine whether to change the hyperparameter search space based on the global historical hyperparameter tuning record set and the current hyperparameter search space, and a new hyperparameter search space is generated after determining to change the hyperparameter search space; In response to the successful generation of a new hyperparameter search space, the current hyperparameter search space of the target machine learning model is updated to the new hyperparameter search space. The local hyperparameter tuning record set of the current proxy model is cleared, and the local historical hyperparameter tuning records belonging to the current hyperparameter search space in the global historical hyperparameter tuning record set are added to the local hyperparameter tuning record set of the current proxy model. Based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space, the original hyperparameter configuration for the next round is generated using the current proxy model, and the original hyperparameter optimization and training operations are continuously executed for the first preset number of rounds.

[0230] In some optional implementations, the next round hyperparameter optimization and training module 404 may be further configured to: in response to the next round hyperparameter optimization execution agent being the fourth hyperparameter optimization execution agent, continuously execute the hyperparameter optimization recommendation operation for the second preset number of rounds or less: determine the recommended hyperparameter configuration of the target machine learning model; update the current hyperparameter configuration of the target machine learning model to the recommended hyperparameter configuration; and execute the training and hyperparameter tuning record generation operation.

[0231] In some alternative implementations, the next round of hyperparameter optimization execution agent can be a fifth hyperparameter optimization execution agent based on a large model; and The next round hyperparameter optimization and training module 404 can be further configured as follows: In response to the next round hyperparameter optimization execution agent being the fifth hyperparameter optimization execution agent, the module continuously performs a training operation after changing the hyperparameter search space up to a third preset number of times: using a large model based on the global historical hyperparameter tuning record set and the current hyperparameter search space to determine whether to change the hyperparameter search space and generating a new hyperparameter search space after determining to change the hyperparameter search space; in response to successfully generating a new hyperparameter search space, updating the current hyperparameter search space of the target machine learning model to the new hyperparameter search space, clearing the local hyperparameter tuning record set of the current proxy model, adding the local historical hyperparameter tuning records belonging to the current hyperparameter search space in the global historical hyperparameter tuning record set to the local hyperparameter tuning record set of the current proxy model; based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space, generating the original next round hyperparameter configuration using the current proxy model, and performing the original hyperparameter optimization and training operation.

[0232] It should be noted that the implementation details and technical effects of each module in the multi-agent model automatic training device provided in the embodiments of this disclosure can be referred to the descriptions of other embodiments in this disclosure, and will not be repeated here.

[0233] Continue to refer to Figure 5 The diagram illustrates a flow 500 of an embodiment of a multi-agent model application method according to the present disclosure, which includes the following steps: Step 501: Obtain business data for the target business scenario.

[0234] Here, the target business scenario is the business scenario corresponding to the target machine learning model. The target machine learning model employs methods such as... Figure 2 The method described in the implementation shown or its alternative implementations is pre-trained.

[0235] As an example, Table 3 shows the business data types corresponding to machine learning models in various application areas. User information data.

[0236] Table 3. Business data types corresponding to machine learning models in various application areas Step 502: Input the business data into the target machine learning model to obtain the prediction results corresponding to the business data.

[0237] As an example, Table 4 shows the prediction results of machine learning models for various application areas, that is, the data format of the output data, and the business actions that may occur in specific business areas based on the prediction results.

[0238] Table 4. Prediction Result Data Formats and Business Actions Corresponding to Machine Learning Models in Various Application Areas The multi-agent model application method provided in the above embodiments of this disclosure obtains prediction results corresponding to the business data by inputting business data of the target business scenario into a target machine learning model. The target machine learning model employs methods such as... Figure 2 The methods described in any of the illustrated embodiments and their optional implementations are pre-trained. During the training process of the target machine learning model, by utilizing the decision-making agent corresponding to the target business scenario, a next-round hyperparameter optimization execution agent is determined to control the next round of hyperparameter optimization scheme for the target machine learning model; then, based on the generated parameter tuning records, the next-round hyperparameter optimization execution agent executes corresponding control operations to stop training the target machine learning model or determine the next-round hyperparameter configuration of the target machine learning model, and train the target machine learning model according to the next-round hyperparameter configuration. Thus, the intelligent transformation of traditional model training methods based on Bayesian hyperparameter optimization methods can achieve, but is not limited to, the following technical effects: First, reduce the number of model training epochs. The applicant's experience has shown that this can reduce the number of model training epochs by 35%.

[0239] Second, adaptive model training is achieved by automatically determining whether to stop training prematurely, changing the hyperparameter space, and generating hyperparameter configurations through multiple executing agents. This improves both model performance and model construction time efficiency. The applicant's practical experience has demonstrated that this can achieve a 45% improvement in time efficiency.

[0240] Furthermore, the aforementioned target machine learning model can be rapidly iterated according to the needs of specific business scenarios, thereby improving the model's predictive performance.

[0241] Further reference Figure 6 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a model application device combining multiple agents, which is similar to... Figure 5 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0242] like Figure 6 As shown, the multi-agent model application device 600 of this embodiment includes: a data acquisition module 601 and a data prediction module 602. The data acquisition module 601 is configured to acquire business data of a target business scenario; the data prediction module 602 is configured to input the business data into a target machine learning model to obtain a prediction result corresponding to the business data, wherein the target machine learning model employs, for example... Figure 2 The method described in any of the illustrated embodiments and its alternative implementations is pre-trained.

[0243] In this embodiment, the specific processing of the data acquisition module 601 and the data prediction module 602 of the multi-agent model application device 600 and the resulting technical effects can be referred to respectively. Figure 5 The relevant descriptions of steps 501 and 502 in the corresponding embodiments will not be repeated here.

[0244] It should be noted that the implementation details and technical effects of each module in the multi-agent model application device provided in the embodiments of this disclosure can be referred to the descriptions of other embodiments in this disclosure, and will not be repeated here.

[0245] The following is for reference. Figure 7 It shows a schematic diagram of the structure of a computer system 700 suitable for implementing the electronic device of the present disclosure. Figure 7 The computer system 700 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0246] like Figure 7 As shown, the computer system 700 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage device 708 into a random access memory (RAM) 703. The RAM 703 also stores various programs and data required for the operation of the computer system 700. The processing device 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0247] Typically, the following devices can be connected to I / O interface 705: input devices 706 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, etc.; output devices 707 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 708 including, for example, magnetic tapes, hard disks, etc.; and communication devices 709. Communication device 709 allows computer system 700 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 7 A computer system 700 with various electronic devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0248] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 709, or installed from storage device 708, or installed from ROM 702. When the computer program is executed by processing device 701, it performs the functions defined in the methods of embodiments of this disclosure.

[0249] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0250] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0251] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the following functions: Figure 2The embodiments shown and their alternative implementations illustrate an automatic training method for models incorporating multiple agents and / or, as described above. Figure 5 The method for applying a multi-agent model is shown.

[0252] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, C++, Python, and C#, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0253] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0254] The modules described in the embodiments of this disclosure can be implemented in software or hardware. The names of modules do not necessarily limit the module itself; for example, the current hyperparameter configuration generation module can also be described as "a module that generates the current hyperparameter configuration using the current surrogate model based on the historical hyperparameter tuning record set and the current hyperparameter search space of the target machine learning model."

[0255] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

Claims

1. An automatic training method for a multi-agent model, comprising: Based on the historical hyperparameter tuning record set and the current hyperparameter search space of the target machine learning model, the current hyperparameter configuration is generated using the current surrogate model; Perform the following training and hyperparameter tuning record generation operations: train the target machine learning model according to the current hyperparameter configuration; perform performance evaluation on the target machine learning model to obtain training evaluation results; generate hyperparameter tuning records based on the current hyperparameter configuration and the training evaluation results; By utilizing the decision-making agent corresponding to the target business scenario to select the execution agent, the next hyperparameter optimization execution agent that controls the next hyperparameter optimization scheme of the target machine learning model is determined. Based on the generated parameter tuning records, the agent executes corresponding control operations using the next round of hyperparameter optimization to stop training the target machine learning model or determine the next round of hyperparameter configuration for the target machine learning model, and train the target machine learning model according to the next round of hyperparameter configuration.

2. The method according to claim 1, wherein, The training and parameter tuning record generation operation also includes: The generated parameter tuning records are added to the global historical parameter tuning record set of the target machine learning model and the local historical parameter tuning record set of the current proxy model.

3. The method according to claim 2, wherein, After adding the generated hyperparameter tuning records to the global historical hyperparameter tuning record set of the target machine learning model and the local historical hyperparameter tuning record set of the current proxy model, the training and hyperparameter tuning record generation operation further includes: Based on the local historical parameter tuning record set of the current proxy model and the current hyperparameter search space of the target machine learning model, the original hyperparameter configuration for the next round is generated using the current proxy model.

4. The method according to claim 3, wherein, The process of generating the original hyperparameter configuration for the next round using the current proxy model, based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space of the target machine learning model, includes: The local historical parameter tuning records belonging to the current hyperparameter search space in the local historical parameter tuning records of the current proxy model are determined as the local historical parameter tuning records of the current space; Based on the current local historical parameter tuning records and the current hyperparameter search space, the original hyperparameter configuration for the next round is generated using the current surrogate model.

5. The method according to claim 3 or 4, wherein, The hyperparameter optimization execution agent in the next round is any one of the following hyperparameter optimization execution agents: The first hyperparameter optimization execution agent is used to continue training with the original next round hyperparameter configuration; The second hyperparameter optimization execution agent is used to evaluate whether to stop training; The third hyperparameter optimization execution agent is used to instruct whether to change the hyperparameter search space and continue training for the first preset number of rounds. The fourth hyperparameter optimization execution agent is used to adopt other hyperparameter optimization schemes instead of the original next-round hyperparameter configuration during the subsequent second preset number of hyperparameter optimization rounds.

6. The method according to claim 2, wherein, The step of determining the next-round hyperparameter optimization execution agent, which controls the next-round hyperparameter optimization scheme of the target machine learning model, by utilizing the decision-making agent corresponding to the target business scenario for selecting the execution agent, includes: Based on the preset probability or preset round frequency corresponding to each hyperparameter optimization agent, determine the next round of hyperparameter optimization agent for training the target machine learning model; or Based on the global historical parameter tuning record set, a large model is used to determine the next round of hyperparameter optimization execution agent that controls the next round of hyperparameter optimization scheme for the target machine learning model.

7. The method according to claim 5, wherein, Based on the generated parameter tuning records, the agent performs corresponding control operations using the next round of hyperparameter optimization, including: In response to the next round of hyperparameter optimization execution agent being the first next round of hyperparameter optimization execution agent, the following original hyperparameter optimization and training operations are performed: updating the current hyperparameter configuration to the original next round hyperparameter configuration, and performing the training and hyperparameter tuning record generation operation.

8. The method according to claim 5, wherein, Based on the generated parameter tuning records, the agent performs corresponding control operations using the next round of hyperparameter optimization, including: In response to the next round of hyperparameter optimization execution agent being the second hyperparameter optimization execution agent, determine whether to stop training based on the global historical hyperparameter tuning record set; In response to determining to stop training, the model parameters of the target machine learning model are persistently saved; In response to the determination not to stop training, the original hyperparameter optimization and training operations are performed.

9. The method according to claim 5, wherein, The process of using the generated parameter tuning record and the next round of hyperparameter optimization to execute corresponding control operations by the agent includes: In response to the next round of hyperparameter optimization execution agent being the third hyperparameter optimization execution agent, the large model is used to determine whether to change the hyperparameter search space based on the global historical hyperparameter tuning record set and the current hyperparameter search space, and a new hyperparameter search space is generated after determining to change the hyperparameter search space; In response to the successful generation of a new hyperparameter search space, the current hyperparameter search space of the target machine learning model is updated to the new hyperparameter search space. The local hyperparameter tuning record set of the current proxy model is cleared, and the local historical hyperparameter tuning records belonging to the current hyperparameter search space in the global historical hyperparameter tuning record set are added to the local hyperparameter tuning record set of the current proxy model. Based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space, the original hyperparameter configuration for the next round is generated using the current proxy model, and the original hyperparameter optimization and training operations are continuously executed for the first preset number of rounds.

10. The method according to claim 5, wherein, The process of using the generated parameter tuning record and the next round of hyperparameter optimization to execute corresponding control operations by the agent includes: In response to the next round of hyperparameter optimization execution agent being the fourth hyperparameter optimization execution agent, the hyperparameter optimization recommendation operation is continuously executed for the second preset number of rounds or less: determining the recommended hyperparameter configuration of the target machine learning model; updating the current hyperparameter configuration of the target machine learning model to the recommended hyperparameter configuration; and executing the training and hyperparameter tuning record generation operation.

11. The method according to claim 7, wherein, The step of determining the next-round hyperparameter optimization execution agent, which controls the next-round hyperparameter optimization scheme of the target machine learning model, by utilizing the decision-making agent corresponding to the target business scenario for selecting the execution agent, includes: Based on the probability of changing the preset hyperparameter search space or the frequency of changing the preset hyperparameter search space rounds, the next round of hyperparameter optimization execution agent training the target machine learning model is determined as the fifth hyperparameter optimization execution agent based on the large model; and The process of using the generated parameter tuning record and the next round of hyperparameter optimization to execute corresponding control operations by the agent includes: In response to the next round of hyperparameter optimization execution agent being the fifth hyperparameter optimization execution agent, the following training operations after changing the hyperparameter search space are performed: using a large model based on the global historical hyperparameter tuning record set and the current hyperparameter search space to determine whether to change the hyperparameter search space, and generating a new hyperparameter search space after determining to change the hyperparameter search space; in response to successfully generating a new hyperparameter search space, updating the current hyperparameter search space of the target machine learning model to the new hyperparameter search space, clearing the local hyperparameter tuning record set of the current proxy model, adding the local historical hyperparameter tuning records belonging to the current hyperparameter search space from the global historical hyperparameter tuning record set to the local hyperparameter tuning record set of the current proxy model; based on the local historical hyperparameter tuning record set of the current proxy model and the current hyperparameter search space, generating the original next round hyperparameter configuration using the current proxy model, and performing the original hyperparameter optimization and training operations.

12. An automatic training device for a multi-agent model, comprising: The current hyperparameter configuration generation module is configured with a local historical hyperparameter tuning record set and the current hyperparameter search space of the target machine learning model, and generates the current hyperparameter configuration using the current proxy model; The training and hyperparameter tuning record generation module is configured to perform the following training and hyperparameter tuning record generation operations: training the target machine learning model according to the current hyperparameter configuration; performing performance evaluation on the target machine learning model to obtain training evaluation results; and generating hyperparameter tuning records based on the current hyperparameter configuration and the training evaluation results. The agent determination module is configured to use the decision agent corresponding to the target business scenario to select the agent, and determine the next hyperparameter optimization agent to control the next hyperparameter optimization scheme of the target machine learning model. The next round hyperparameter optimization and training module is configured to, based on the generated hyperparameter tuning record, use the next round hyperparameter optimization execution agent to perform corresponding control operations, thereby stopping the training of the target machine learning model or determining the next round hyperparameter configuration of the target machine learning model, and training the target machine learning model according to the next round hyperparameter configuration.

13. A method for applying a multi-agent model, comprising: Obtain business data for the target business scenario; The business data is input into the target machine learning model to obtain the prediction result corresponding to the business data, wherein the target machine learning model is pre-trained using the method described in any one of claims 1-11.

14. A model application device combining multiple agents, comprising: The data acquisition module is configured to acquire business data for the target business scenario; The data prediction module is configured to input the business data into a target machine learning model to obtain a prediction result corresponding to the business data, wherein the target machine learning model is pre-trained using the method described in any one of claims 1-11.

15. An electronic device comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method as described in any one of claims 1-11 and / or claim 13.

16. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed by one or more processors, it implements the method as described in any of claims 1-11 and / or claim 13.

17. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the method as described in any one of claims 1-11 and / or claim 13.

Images