Agent optimization model training method, agent optimization method and device

Abstract

The application discloses an agent optimization model training method, an agent optimization method and device, and belongs to the technical field of artificial intelligence. The agent optimization model training method comprises the following steps: acquiring agent optimization training data; the historical optimization mode information comprises a historical optimization mode and evaluation information corresponding to the historical optimization mode; an initial model for predicting optimization mode information of an agent is trained by using the agent optimization training data, so that an agent optimization model is obtained; the agent optimization model is used for receiving optimization related information of an agent, and outputting an optimization mode of the agent according to the optimization related information. The optimization mode of the agent for agent tuning output or predicted by the above agent optimization model does not need to be manually adjusted by a person, can reduce the tuning cycle of the agent, and the optimization mode of the agent for agent tuning output or predicted by the trained agent optimization model has high accuracy.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and specifically relates to a method for training an intelligent agent optimization model, an intelligent agent optimization method, and an apparatus. Background Technology

[0002] Current intelligent agent optimization techniques mainly rely on developers selecting tuning methods based on experience or using a single optimization strategy (such as global parameter tuning) for tuning.

[0003] For example, existing technologies use optimization methods such as parameter adjustment based on manual trial and error, multi-round optimization processes with a fixed order, and static optimization schemes that rely on offline evaluation to fine-tune intelligent agents.

[0004] However, the above-mentioned optimization of intelligent agents has the problem of long optimization cycle, and due to the lack of business scenario knowledge, task planning deviation or model parameter mismatch, the above-mentioned optimization of intelligent agents has the problem of low accuracy. Summary of the Invention

[0005] This application provides a method for training an intelligent agent optimization model, an intelligent agent optimization method, and an apparatus to solve the problems of long optimization cycles and low accuracy in existing intelligent agent tuning methods.

[0006] In a first aspect, embodiments of this application provide a method for training an intelligent agent optimization model, the method comprising:

[0007] Acquire agent optimization training data; wherein, the agent optimization training data includes historical optimization-related information of the agent and historical optimization method information corresponding to the historical optimization-related information; the historical optimization method information includes historical optimization methods and evaluation information corresponding to the historical optimization methods;

[0008] The initial model used to predict the optimization method information of the agent is trained using the agent optimization training data to obtain the agent optimization model;

[0009] The agent optimization model is used to receive optimization-related information of the agent and output the optimization method of the agent based on the optimization-related information.

[0010] Optionally, the initial model includes a processing layer and at least two prediction layers, wherein the at least two prediction layers are arranged sequentially.

[0011] The process includes training an initial model for predicting the agent's optimization methods using the agent's optimization training data to obtain the agent optimization model, including:

[0012] The historical optimization-related information of the agent is input into the processing layer, and the processing layer performs semantic feature analysis on the historical optimization-related information to obtain the semantic features of the historical optimization-related information.

[0013] The semantic features are input into the at least two prediction layers to obtain the agent optimization method information predicted by each prediction layer, wherein one prediction layer corresponds to one agent optimization method information;

[0014] The model parameters of the initial model are adjusted based on the predicted agent optimization method information and the historical optimization method information to obtain the agent optimization model;

[0015] The agent optimization method information predicted by the first prediction layer of the at least two prediction layers is obtained based on the semantic features.

[0016] The agent optimization method information predicted by the first prediction layer in the at least two prediction layers is obtained based on the semantic features and the agent optimization method information predicted by the second prediction layer. The first prediction layer is the prediction layer other than the first prediction layer in the at least two prediction layers, and the second prediction layer is the prediction layer before the first prediction layer in the at least two prediction layers.

[0017] Optionally, the processing layer includes a background layer, a task layer, and a connection layer;

[0018] Specifically, the historical optimization-related information of the agent is input into the processing layer, and the processing layer performs semantic feature analysis on the historical optimization-related information to obtain the semantic features of the historical optimization-related information, including:

[0019] The historical optimization-related information of the intelligent agent is subjected to a first processing to obtain a first historical optimization-related information, wherein the first processing includes at least one of the following: integrity check, logical consistency check and outlier detection;

[0020] The capability information of the intelligent agent is obtained based on the first historical optimization information;

[0021] The capability information is input into the background layer, and the first semantic feature of the capability information is extracted using the background layer.

[0022] The first semantic feature is input into the task layer, and the shape of the first semantic feature is adjusted by the task layer to obtain the second semantic feature;

[0023] The first semantic feature and the second semantic feature are input into the connection layer, and the connection layer is used to perform a second processing on the first semantic feature and the second semantic feature to obtain the semantic feature. The second processing includes at least one of the following: linking processing and flattening processing.

[0024] Optionally, the evaluation information corresponding to the historical optimization methods includes optimization accuracy value and overall optimization accuracy value; wherein, the optimization accuracy value is determined based on the historical optimization methods and the number of historical optimization methods, and the overall optimization accuracy is obtained by weighting the F1 scores corresponding to the historical optimization methods;

[0025] The process involves adjusting the model parameters of the initial model based on the predicted agent optimization method information and the historical optimization method information to obtain the agent optimization model, including:

[0026] The first loss information is obtained based on the predicted agent optimization method information and the historical optimization method information;

[0027] The model parameters of the initial model are adjusted using the first loss information to obtain the agent optimization model;

[0028] The predicted agent optimization method information includes the predicted agent optimization method, the predicted optimization accuracy value, and the predicted comprehensive optimization accuracy value.

[0029] The first loss information is obtained based on the loss information between the predicted agent optimization method and the historical optimization method, the error information between the predicted optimization accuracy value and the optimization accuracy value, and the error information between the predicted comprehensive optimization accuracy value and the comprehensive optimization accuracy value.

[0030] Optionally, the optimization method includes at least one of the following:

[0031] Adjust the parameters of the large model corresponding to the agent;

[0032] Optimize the task planning information and knowledge recommendation information of the intelligent agent;

[0033] Reinforcement learning is applied to the intelligent agent.

[0034] Optionally, the optimization-related information includes at least one of the following:

[0035] User industry classification corresponding to intelligent agents;

[0036] Information describing the application scenarios of intelligent agents;

[0037] Task description information of the intelligent agent;

[0038] Agent version information;

[0039] Information describing the capabilities of an intelligent agent.

[0040] Secondly, embodiments of this application also provide an agent optimization method, the method comprising:

[0041] Obtain optimization-related information for the intelligent agent;

[0042] The optimization-related information of the agent is input into the agent optimization model to obtain the optimization method information output by the agent optimization model; wherein, the agent optimization model is used to output the optimization method information based on the optimization-related information, and the agent optimization model is trained using agent optimization training data; the optimization method information includes the optimization method and the evaluation information corresponding to the optimization method;

[0043] The agent is optimized based on the optimization method information.

[0044] Thirdly, embodiments of this application also provide an intelligent agent optimization model training device, the device comprising:

[0045] The first acquisition module is used to acquire agent optimization training data; wherein, the agent optimization training data includes historical optimization-related information of the agent and historical optimization method information corresponding to the historical optimization-related information; the historical optimization method information includes historical optimization methods and evaluation information corresponding to the historical optimization methods;

[0046] The first processing module is used to train the initial model for predicting the optimization method information of the agent using the agent optimization training data to obtain the agent optimization model.

[0047] The agent optimization model is used to receive optimization-related information of the agent and output the optimization method of the agent based on the optimization-related information.

[0048] Fourthly, embodiments of this application also provide an intelligent agent optimization device, the device comprising:

[0049] The second acquisition module is used to acquire optimization-related information of the agent;

[0050] The second processing module is used to input the optimization-related information of the agent into the agent optimization model to obtain the optimization method information output by the agent optimization model; wherein, the agent optimization model is used to output the optimization method information based on the optimization-related information, and the agent optimization model is trained using agent optimization training data; the optimization method information includes the optimization method and the evaluation information corresponding to the optimization method;

[0051] The third processing module is used to optimize the agent based on the optimization method information.

[0052] Fifthly, embodiments of this application also provide an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the steps in the agent optimization model training method as described in any one of the first aspects, or implements the steps in the agent optimization method as described in the second aspect.

[0053] Sixthly, embodiments of this application also provide a readable storage medium storing a program that, when executed by a processor, implements the steps in the agent optimization model training method as described in any one of the first aspects, or implements the steps in the agent optimization method as described in the second aspect.

[0054] In a seventh aspect, embodiments of this application also provide a computer program product, including computer instructions, which, when executed by a processor, implement the steps in the agent optimization model training method as described in any one of the first aspects, or implement the steps in the agent optimization method as described in the second aspect.

[0055] The beneficial effects of this application are:

[0056] The agent optimization model training method provided in this application acquires agent optimization model training data, which includes historical optimization-related information of the agent and historical optimization method information corresponding to the historical optimization-related information. The historical optimization method information includes historical optimization methods and corresponding evaluation information. This agent optimization training data is used to train an initial model for predicting agent optimization methods, resulting in an agent optimization model. This trained agent optimization model receives optimization-related information of the agent and outputs (or predicts) the agent's optimization methods based on this information. Using the agent optimization model trained above, the output or prediction of agent optimization methods for agent tuning eliminates the need for manual adjustments, reducing the agent tuning cycle. Furthermore, since the agent optimization model is trained based on historical optimization-related information, the output or prediction of agent optimization methods for agent tuning by the trained agent optimization model has high accuracy. Attached Figure Description

[0057] Figure 1 This is a flowchart of the intelligent agent optimization model training method provided in the embodiments of this application;

[0058] Figure 2This is a schematic diagram of the model structure of the initial model provided in the embodiments of this application;

[0059] Figure 3 This is one of the flowcharts of the intelligent agent optimization method provided in the embodiments of this application;

[0060] Figure 4 This is the second flowchart of the intelligent agent optimization method provided in the embodiments of this application;

[0061] Figure 5 This is a schematic diagram of the structure of the intelligent agent optimization model training device provided in the embodiments of this application;

[0062] Figure 6 This is a schematic diagram of the intelligent agent optimization device provided in the embodiments of this application;

[0063] Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0064] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0065] In the following description, specific details such as particular configurations and components are provided merely to aid in a comprehensive understanding of the embodiments of this application. Therefore, those skilled in the art will understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Furthermore, for clarity and brevity, descriptions of known functions and constructions have been omitted.

[0066] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0067] In the various embodiments of this application, it should be understood that the sequence number of each process described below does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0068] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and are not used to describe a specified order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, a first object can be one or more. Furthermore, in the specification and claims, "and" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0069] Furthermore, the "or" in this application indicates at least one of the connected objects. For example, "A or B" covers three scenarios: Scenario 1: includes A but excludes B; Scenario 2: includes B but excludes A; Scenario 3: includes both A and B. The character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0070] The term "instruction" in this application can be either a direct instruction (or explicit instruction) or an indirect instruction (or implicit instruction). A direct instruction can be understood as one in which the sender explicitly informs the receiver of specific information, the operation to be performed, or the requested result, etc.; an indirect instruction can be understood as one in which the receiver determines the corresponding information based on the instruction sent by the sender, or makes a judgment and determines the operation to be performed or the requested result, etc., based on the judgment result.

[0071] To address the problems of long optimization cycles, poor policy adaptability, and low relevance to business scenarios in existing agent optimization methods, this application provides an agent optimization model training method, agent optimization method, and apparatus.

[0072] like Figure 1 As shown in the figure, this application provides a method for training an agent optimization model, the method comprising:

[0073] Step 101: Obtain agent optimization training data; wherein, the agent optimization training data includes historical optimization-related information of the agent and historical optimization method information corresponding to the historical optimization-related information; the historical optimization method information includes historical optimization methods and evaluation information corresponding to the historical optimization methods.

[0074] In this step, structured historical optimization information is submitted through the standardized data collection interface provided by the intelligent agent platform. The input historical optimization information is standardized and its integrity is checked through predefined templates and verification rules to ensure data quality and provide standardized and clear input for subsequent model training.

[0075] In this step, we obtain historical optimization records, i.e., historical optimization method information, which includes historical optimization methods (or historical tuning methods) and the evaluation information corresponding to the historical optimization methods.

[0076] For example, in this embodiment, the evaluation information corresponding to historical optimization methods is represented by an optimization accuracy value. It is understood that if the performance of an agent improves after optimization based on historical optimization methods, then the historical optimization method is considered accurate. Therefore, the optimization accuracy value is obtained based on the number of accurate historical optimization methods and the total number of historical optimization methods. Specifically, if the performance of an agent improves after optimization based on historical optimization methods, then the historical optimization method is considered accurate, and the ratio of accurate historical optimization methods to the total number of historical optimization methods is used as the optimization accuracy value.

[0077] Label the historical optimization information mentioned above with the actual historical optimization method information, and construct a training sample sequence (or label data). This label data is the agent optimization training data.

[0078] That is, each sample in the intelligent agent's optimized training data includes structured input features.

[0079] Optionally, the optimization-related information includes at least one of the following (1), (2), (3), (4), (5), and (6):

[0080] (1) The user industry classification corresponding to the intelligent agent can be understood as the industry classification (or industry type) applied by the intelligent agent.

[0081] (2) Description of the application scenario of the intelligent agent.

[0082] The application scenario description information of the intelligent agent includes at least one of the following: the task background of the intelligent agent, the professional knowledge of the intelligent agent, and the precautions of the intelligent agent.

[0083] (3) Task description information of the intelligent agent.

[0084] The task description information of the intelligent agent can also be understood as detailed task description information, including at least one of the following: operation object, task objective, task output requirements, tasks performed, task requirements, and constraint information.

[0085] (4) Version information of the agent.

[0086] The version information of the agent includes the type and version information of the large model in the agent, as well as the version of the agent.

[0087] (5) Description of the agent's capabilities.

[0088] (6) Application time of the intelligent agent.

[0089] Optionally, it may also include at least one of the following: agent usage, user evaluation data, and agent evaluation and optimization records.

[0090] It is understandable that historical optimization-related information includes at least one of the following: the user industry classification corresponding to the historical intelligent agent, the application scenario description information of the historical intelligent agent, the task description information of the historical intelligent agent, the version information of the historical intelligent agent, the capability description information of the historical intelligent agent, the application time of the historical intelligent agent, the usage of the intelligent agent, user evaluation data, and intelligent agent evaluation and tuning record data.

[0091] The optimization method refers to the specific operation of targeted optimization of the agent based on the recommendation strategy, including at least one of the following A, B, and C:

[0092] A. Adjust the parameters of the large model corresponding to the agent, i.e., adjust the parameters of the large model.

[0093] This involves adjusting the parameters of the large model on which the agent depends, such as temperature parameters and top-k parameters.

[0094] B. Optimize the agent's task planning information and knowledge recommendation information, i.e., optimize the agent's task planning and knowledge recommendation.

[0095] Specifically, this includes reconstructing and enhancing the task chain structure, task planning logic, and knowledge base retrieval strategy of the intelligent agent.

[0096] C. Perform reinforcement learning on the agent, i.e., optimization based on reinforcement learning.

[0097] Specifically, this includes utilizing the agent's interaction logs in the business environment and employing reinforcement learning algorithms to optimize the agent's strategy offline.

[0098] Specifically, the process of acquiring optimization-related information of historical intelligent agents includes:

[0099] Users submit relevant information about historical agents through the intelligent agent platform. This information is then processed in real-time using a standardized data collection interface. Specifically, this process involves extracting key information and mapping fields from unstructured text information based on a predefined industry classification system and task description template, resulting in key fields. Built-in validation rules are used to check the completeness of required fields within these key fields and to assess their logical consistency and format compliance, ensuring data conformity to preset standards. All key fields are automatically categorized and organized into unified, machine-readable structured data objects, providing standardized and high-quality input for subsequent processing stages.

[0100] Step 102: Use the agent optimization training data to train the initial model used to predict the agent's optimization method information to obtain the agent optimization model;

[0101] The agent optimization model is used to receive optimization-related information of the agent and output the optimization method of the agent based on the optimization-related information.

[0102] In this step, the agent's optimization method can also be called the agent's tuning method, or it can be understood as the optimized path recommendation result output by the agent optimization model.

[0103] Understandably, the input to the agent optimization model is the current agent optimization-related information, and the output is the predicted agent optimization method.

[0104] Using the agent optimization model obtained through the above training, the optimization method of the agent for agent tuning can be output or predicted without manual adjustment, which can reduce the tuning cycle of the agent. Moreover, the agent optimization model is trained based on historical optimization information, so that the optimization method of the agent for agent tuning output or predicted by the trained agent optimization model has high accuracy.

[0105] Optionally, the initial model is a hierarchical network architecture, which includes a processing layer and at least two prediction layers, wherein the at least two prediction layers are arranged sequentially.

[0106] Here, "arranged sequentially" can be understood as each of the at least two prediction layers being arranged independently from front to back, with the preceding prediction layer connected to the following prediction layer.

[0107] In this configuration, the first prediction layer (or the first prediction layer) of at least two prediction layers is connected to the processing layer.

[0108] In an optional embodiment, training an initial model for predicting the agent's optimization methods using the agent's optimization training data to obtain the agent optimization model includes:

[0109] The historical optimization-related information of the agent is input into the processing layer, and the processing layer performs semantic feature analysis on the historical optimization-related information to obtain the semantic features of the historical optimization-related information.

[0110] The prepared historical optimization information of the agent is input into the processing layer. The processing layer performs semantic feature analysis on the historical optimization information of the agent to obtain the semantic features of the historical optimization information.

[0111] The semantic features are input into the at least two prediction layers to obtain the agent optimization method information predicted by each prediction layer, wherein one prediction layer corresponds to one agent optimization method information;

[0112] Specifically, the agent optimization method information predicted by the first prediction layer of the at least two prediction layers is obtained based on the semantic features;

[0113] The agent optimization method information predicted by the first prediction layer in the at least two prediction layers is obtained based on the semantic features and the agent optimization method information predicted by the second prediction layer. The first prediction layer is the prediction layer other than the first prediction layer in the at least two prediction layers, and the second prediction layer is the prediction layer before the first prediction layer in the at least two prediction layers.

[0114] That is, at least two prediction layers take the semantic features output by the processing layer as input. The first prediction layer makes a prediction based on the input semantic features to obtain the agent optimization method information predicted by the first prediction layer. The input of each prediction layer other than the first prediction layer includes the semantic features output by the processing layer and the agent optimization method information predicted by each prediction layer before it. The prediction layer obtains the agent optimization method information predicted by the prediction layer based on the semantic features output by the processing layer and the agent optimization method information predicted by each prediction layer before it.

[0115] For example, the initial model performs three predictions, meaning it comprises three prediction layers: a first prediction layer, a middle prediction layer, and a final prediction layer. The first prediction layer is connected to the processing layer. The first prediction layer uses the semantic features output from the processing layer as input to obtain the agent optimization method information predicted by the first prediction layer (i.e., the agent optimization method information for the first prediction). The middle prediction layer uses the semantic features output from the processing layer and the agent optimization method information predicted by the first prediction layer as input to obtain the agent optimization method information predicted by the middle prediction layer (i.e., the agent optimization method information for the second prediction). The final prediction layer uses the semantic features output from the processing layer, the agent optimization method information predicted by the first prediction layer, and the agent optimization method information predicted by the middle prediction layers as input to obtain the agent optimization method information predicted by the final prediction layer. The agent optimization method information from each prediction layer is output in a unified format.

[0116] It should be noted that the number of prediction layers in the initial model can be set according to requirements, and is not limited in this embodiment.

[0117] The model parameters of the initial model are adjusted based on the predicted agent optimization method information and the historical optimization method information to obtain the agent optimization model.

[0118] That is, the initial model is trained and optimized based on the predicted agent optimization method information and the historical optimization method information to obtain the trained agent optimization model.

[0119] Optionally, the processing layer includes a background layer, a task layer, and a connection layer.

[0120] The background layer is used to learn user industry classification and application scenario information. The background layer includes a word embedding layer, a normalization layer, and a Text Convolutional Neural Network (TextCNN) layer.

[0121] The task layer can also be called the ALBert layer.

[0122] The connection layer includes a fusion linking layer and a fully connected layer. The semantic features output by the above processing layers are the same as the semantic features output by the fully connected layer.

[0123] In one optional implementation, the historical optimization-related information of the agent is input into the processing layer, and the processing layer performs semantic feature analysis on the historical optimization-related information to obtain the semantic features of the historical optimization-related information, including:

[0124] The input data undergoes preprocessing. Specifically, through structured mapping and standardization processes, the historical optimization information of the original agent is converted into a unified input format required by the model. This process includes data quality verification, agent capability profiling, and multimodal data format conversion, aiming to improve data standardization and consistency and provide reliable numerical feature representations for the model. The preprocessed historical optimization information of the agent is then input into the initial model.

[0125] The historical optimization-related information of the intelligent agent is subjected to a first processing to obtain a first historical optimization-related information, wherein the first processing includes at least one of the following: integrity check, logical consistency check and outlier detection.

[0126] The first process can also be called data quality verification process.

[0127] For example, the historical optimization information of the input agent is subjected to integrity checks, logical consistency checks and outlier detection to obtain the first historical optimization information. This ensures that the obtained first historical optimization information conforms to the preset specifications, providing a clean and reliable data foundation for subsequent processing.

[0128] The capability information of the intelligent agent is obtained based on the first historical optimization information.

[0129] Specifically, construct an agent capability profile: Based on the first historical optimization information verified above, systematically characterize the agent's capability information from multiple dimensions. The agent's capability information includes the capability matrix of the constructed agent model (quantitatively scored from dimensions such as language understanding and logical reasoning), execution knowledge base coverage analysis, and key features extracted from historical optimization information to obtain comprehensive agent capability information (also known as the agent's capability profile vector).

[0130] The data in the ability information of the intelligent agent mentioned above is transformed into a format. The final structured format transformation is achieved through a unified preprocessing pipeline. For example, the user industry classification is encoded, the corresponding scene information is vectorized, the task description information is deeply analyzed and the fields are aligned, and finally a standardized feature vector set (i.e., ability information) that conforms to the model is generated.

[0131] The capability information is input into the background layer, and the first semantic feature of the capability information is extracted using the background layer.

[0132] Specifically, the word embedding layer in the background layer maps capability information (such as user industry classification and application scenario information) into a distributed representation (or word embedding vector). The normalization layer in the background layer normalizes the word embedding vector to improve training efficiency. The TextCNN layer in the background layer further extracts the local and global semantic features of the text in the normalized word embedding vector to form the output of the background layer, i.e. the first semantic feature.

[0133] The first semantic feature is input into the task layer, and the shape of the first semantic feature is adjusted by the task layer to obtain the second semantic feature.

[0134] The task layer (ALBert layer) deploys a domain-fine-tuned version of the ALBert-base model. The task layer loads the domain-fine-tuned version of the ALBert-base model to parse the task semantics of the first semantic feature, extracts the feature extraction layer of the TextCNN layer, removes its output layer, and adjusts the shape of the first semantic feature output by the TextCNN layer to be consistent with the shape of ALBert, thus obtaining the second semantic feature.

[0135] The first semantic feature and the second semantic feature are input into the connection layer, and the connection layer is used to perform a second processing on the first semantic feature and the second semantic feature to obtain the semantic feature. The second processing includes at least one of the following: linking processing and flattening processing.

[0136] Specifically, the feature information (first semantic feature) generated by the background layer and the feature information (i.e., second semantic feature) generated by the task layer are fused through a fusion linking layer. The fused semantic feature is then linked and flattened according to its shape through a fully connected layer to obtain the semantic feature. The semantic feature is then input into the prediction layer for prediction.

[0137] For example, the model structure of the initial model provided in this embodiment is as follows: Figure 2 As shown.

[0138] Optionally, the evaluation information corresponding to the historical optimization methods includes optimization accuracy value and overall optimization accuracy value; wherein, the optimization accuracy value is determined based on the historical optimization methods and the number of historical optimization methods, and the overall optimization accuracy is obtained by weighting the F1 scores corresponding to the historical optimization methods.

[0139] For example, the optimized accuracy value can be represented by a score. The specific calculation process for this optimized accuracy value includes:

[0140] The agent is optimized using historical optimization methods. If the performance of the optimized agent is higher than that of the agent before optimization, the historical optimization method is considered accurate. Accurate historical optimization methods are obtained in the above manner. The number of accurate historical optimization methods is divided by the total number of historical optimization methods to obtain the optimization accuracy rate.

[0141] For example, the overall optimization accuracy value can be represented by a score (F-Score). The specific calculation of this overall optimization accuracy value includes:

[0142] The historical optimization methods are categorized (e.g., according to A, B, and C as described above), and the F1 score for each category is calculated using the following formula:

[0143]

[0144]

[0145] in, Indicates a real example, Indicates a false positive example. This indicates a false counterexample.

[0146] Specifically, a true positive is the number of accurate historical optimization methods among those that can be predicted from historical optimization information, a false positive is the number of inaccurate historical optimization methods (i.e., historical optimization methods other than accurate ones) among those that can be predicted from historical optimization information, and a false negative is the number of historical optimization methods that cannot be predicted from the agent's historical optimization information.

[0147] Next, the F1 scores corresponding to the historical optimization methods for each category are weighted (Weighted F1) to obtain the overall optimization accuracy, as shown in the following formula:

[0148]

[0149] in, This represents the F1 score corresponding to the historical optimization method in class A. This represents the F1 score corresponding to the historical optimization method of class B. This represents the F1 score corresponding to the historical optimization method of type C. express The weight, express The corresponding weights They represent The corresponding weights.

[0150] After the historical optimization method, optimization accuracy, and overall accuracy are calculated, they are revised and reviewed to ensure correctness and consistency. Then, a label data is constructed based on the historical optimization method information and related information, including the historical optimization method, optimization accuracy, and overall accuracy. This label data is the agent optimization training data.

[0151] In summary, the initial model, based on preprocessed historical optimization information, predicts historical optimization methods through a hierarchical network model. Specifically, the initial model extracts semantic features of historical optimization information from the background and task layers, fuses them, and then performs multi-stage optimization prediction through cascaded prediction layers. The final output includes information on the predicted historical optimization methods, the predicted optimization accuracy, and the overall prediction accuracy. The calculation methods for the predicted optimization accuracy and overall prediction accuracy of the initial model are explained above and will not be repeated here.

[0152] The process involves adjusting the model parameters of the initial model based on the predicted agent optimization method information and the historical optimization method information to obtain the agent optimization model, including:

[0153] First loss information is obtained based on the predicted agent optimization method information and the historical optimization method information; wherein, the predicted agent optimization method information includes the predicted agent optimization method, the predicted optimization accuracy value, and the predicted comprehensive optimization accuracy value; the first loss information is obtained based on the loss information between the predicted agent optimization method and the historical optimization method, the error information between the predicted optimization accuracy value and the optimization accuracy value, and the error information between the predicted comprehensive optimization accuracy value and the comprehensive optimization accuracy value.

[0154] In one optional implementation, the first loss information consists of loss information between the predicted agent optimization method and the historical optimization method, error information between the predicted optimization accuracy value and the optimization accuracy value, and error information between the predicted comprehensive optimization accuracy value and the comprehensive optimization accuracy value.

[0155] In another optional implementation, the first loss information is obtained by weighted summation of the loss information between the predicted agent optimization method and the historical optimization method, the error information between the predicted optimization accuracy value and the optimization accuracy value, and the error information between the predicted comprehensive optimization accuracy value and the comprehensive optimization accuracy value.

[0156] Optionally, the loss information between the predicted agent optimization method and the historical optimization method is the cross-entropy loss between the predicted agent optimization method and the historical optimization method; the error information between the predicted optimization accuracy value and the optimization accuracy value is the mean square error between the predicted optimization accuracy value and the optimization accuracy value; and the error information between the predicted comprehensive optimization accuracy value and the comprehensive optimization accuracy value is the mean square error between the predicted comprehensive optimization accuracy value and the comprehensive optimization accuracy value.

[0157] It is understandable that the first loss information can also be understood as, or referred to as, the first loss function.

[0158] The model parameters of the initial model are adjusted using the first loss information to obtain the optimized agent model.

[0159] Specifically, during model training and optimization, the model parameters are iteratively optimized by minimizing the first loss information. Minimizing the first loss information can be understood as stopping the iterative optimization of the model and completing the training of the agent optimization model when the first loss information decreases to less than a first preset threshold and the change in the first loss information is less than a second preset threshold (i.e., remains essentially unchanged).

[0160] Therefore, the entire training process is based on data, driven by loss optimization, and ends with business validation, forming a closed-loop process to ensure that the model has both predictive accuracy and business usability.

[0161] Furthermore, the method also includes training the agent optimization model using agent optimization verification data to obtain training results.

[0162] Specifically, the agent optimization verification data includes historical agent optimization information and historical agent optimization methods. It should be noted that the historical agent optimization information and historical agent optimization methods included in the agent optimization verification data are consistent with the historical agent optimization information and historical agent optimization methods included in the agent optimization training data in terms of parameter names and types, but the specific parameter values ​​are different.

[0163] Optionally, the training results include at least one of the following C and D:

[0164] C. The optimization accuracy value of the agent optimization method predicted by the agent optimization model.

[0165] The optimization accuracy value of the agent optimization method predicted by the agent optimization model is determined based on the number of accurate agent optimization methods predicted by the agent optimization model and the number of predicted agent optimization methods. The specific calculation process is described above in the optimization accuracy value, and will not be repeated here.

[0166] D. The overall optimization accuracy value of the agent optimization method predicted by the agent optimization model.

[0167] The overall optimization accuracy value predicted by the agent optimization model is determined based on the F1 score corresponding to the accurate agent optimization method predicted by the agent optimization model. The specific calculation process is described above in the context of the overall optimization accuracy value, and will not be repeated here.

[0168] If the above training results meet the first condition, the agent optimization model is determined to be up to standard or meet the requirements.

[0169] The first condition includes at least one of the following conditions: Condition 1, Condition 2, and Condition 3:

[0170] Condition 1: The optimization accuracy of the agent optimization method predicted by the agent optimization model is greater than the first preset value;

[0171] Condition 2: The mean square error of the optimization accuracy value of the agent optimization method predicted by the agent optimization model is less than the second preset value, and the mean square error of the comprehensive optimization accuracy value of the agent optimization method predicted by the agent optimization model is less than the third preset value.

[0172] Condition 3: The overall optimization accuracy of each type of agent optimization method predicted by the agent optimization model reaches the target value corresponding to the business application requirements.

[0173] Following the above process, a trained agent optimization model is obtained. This model can automatically trigger the tuning process based on current agent optimization information, i.e., generate the agent's optimization method (including at least one of large model parameter tuning, task chain and knowledge recommendation optimization, and reinforcement learning tuning). After each tuning method is executed, a version snapshot is automatically generated, and an evaluation service is called for automated testing and evaluation, generating a tuning report containing performance comparisons of each version. Finally, the optimization path and effect data are stored in the training database for model iteration.

[0174] Once the aforementioned agent optimization model has been trained, it can be put into use. When using the agent optimization model for prediction or inference, the input is structured information that is isomorphic to the agent's training data. The agent optimization model will then output the recommended optimization methods and corresponding prediction scores for the first prediction, intermediate prediction, and final prediction stages, forming a complete optimization path suggestion.

[0175] The agent optimization model training method provided in this application is applicable to solving the problem of decreased accuracy of agents in business scenarios due to knowledge gaps, task planning deviations, or unsuitable model parameters. It covers core aspects such as agent industry feature extraction, optimization strategy prediction, multi-stage automated optimization, and closed-loop performance evaluation. Through collaborative modeling of TextCNN industry feature extraction and ALBert task semantic analysis, it achieves intelligent matching between business scenarios and optimization strategies; it constructs a three-stage dynamic prediction optimization decision-making mechanism, achieving progressive optimization path recommendation and performance prediction through cascaded decisions of the first, middle, and final prediction layers; and it adopts a collaborative optimization framework of parameter tuning, task chain reconstruction, and reinforcement learning to support one-stop automated optimization of agents and establish a closed-loop evaluation and feedback mechanism.

[0176] Specifically, multimodal feature fusion and hierarchical prediction techniques are employed to intelligently identify the compatibility between the agent's business scenarios and optimization strategies, overcoming the limitations of traditional experience-driven optimization and significantly improving the accuracy and efficiency of optimization path recommendations. Through collaborative modeling of TextCNN industry feature extraction and ALBert task semantic analysis, intelligent matching of business scenarios and optimization strategies is achieved. A closed-loop system of three-stage dynamic prediction and automated optimization is constructed. A cascaded decision-making mechanism involving the first, middle, and final prediction layers enables multi-round iteration and effect prediction of optimization strategies, providing a progressive optimization solution for the agent and ensuring the reliability and stability of the optimization process. Based on a collaborative optimization framework of parameter tuning, task chain reconstruction, and reinforcement learning, a one-stop solution from problem diagnosis to optimization implementation is provided, supporting personalized optimization needs of agents in different business scenarios and significantly reducing the technical threshold and manpower costs of optimization. An automated evaluation and feedback mechanism continuously optimizes the recommendation model, forming a self-evolving capability for optimization strategies.

[0177] like Figure 3 As shown in the embodiments of this application, an agent optimization method is also provided, the method comprising:

[0178] Step 301: Obtain optimization-related information for the agent.

[0179] It is understandable that the optimization-related information of this intelligent agent can be understood as the real-time optimization-related information of the intelligent agent.

[0180] The optimization-related information includes at least one of (1), (2), (3), (4), (5), and (6) above, and optionally also includes at least one of agent usage, user evaluation data, and agent evaluation and tuning record data.

[0181] Optionally, after obtaining the optimization-related information of the agent, the obtained optimization-related information of the agent is preprocessed, such as data cleaning, according to the log recording rules and format, to form a preliminary data format.

[0182] For example, a standardized intelligent agent feature matrix can be constructed through a structured data acquisition interface (including industry classification, scenario description, and task elements).

[0183] Step 302: Input the optimization-related information of the agent into the agent optimization model to obtain the optimization method information output by the agent optimization model; wherein, the agent optimization model is used to output the optimization method information based on the optimization-related information, and the agent optimization model is trained using agent optimization training data; the optimization method information includes the optimization method and the evaluation information corresponding to the optimization method.

[0184] It is understood that the agent optimization model in this step is an agent optimization model trained using the agent optimization model training method described above or any implementation thereof.

[0185] In this step, the preprocessed optimization-related information of the agent is input into the agent optimization model. The agent optimization model outputs optimization method information based on the optimization-related information of the agent, that is, the agent optimization model recommends the agent's optimization method.

[0186] Specifically, this involves using a preprocessing pipeline to perform deep semantic analysis on multi-source information, and on this basis, generating personalized optimization solutions through a hierarchical prediction model (based on TextCNN industry feature extraction, ALBert task semantic analysis, and a three-stage optimization strategy prediction).

[0187] Optionally, the optimization method includes at least one of A, B, and C above.

[0188] Step 303: Optimize the agent according to the optimization method information.

[0189] In this step, the agent is automatically optimized based on the optimization method in the optimization method information.

[0190] Specifically, based on the optimization methods recommended by the model, the corresponding tuning engine components are automatically invoked to perform optimization operations. If the recommended optimization methods include "reinforcement learning tuning," the platform first collects a large amount of evaluation and production logs (including task information, task chain planning, execution process, interaction records, and result data) of relevant intelligent agents in the industry scenario, and then uses this log data to perform reinforcement learning training on the intelligent agents and the large models they depend on.

[0191] Optionally, an automated tuning engine is then employed, combining techniques such as parameter optimization, task chain reconstruction, and reinforcement learning to achieve multi-dimensional optimization of the agent. For each optimized version, dynamic test cases are executed and weighted scores are calculated through a closed-loop evaluation framework, ultimately generating a tuning report that includes performance comparisons. The entire process achieves end-to-end intelligentization from problem diagnosis to optimization implementation through a collaborative mechanism of feature fusion prediction and automated tuning, significantly improving tuning efficiency, strategy adaptability, and business relevance compared to traditional experience-driven methods.

[0192] Furthermore, the method also includes: automatically generating a version snapshot after each optimization method is executed, and calling the intelligent agent evaluation service to conduct multi-dimensional business evaluation of the optimized version (including indicators such as accuracy and response time) based on a preset test case set, calculating a weighted evaluation score (fscore) in combination with usage data, and generating an optimization report containing performance comparison.

[0193] Optionally, the complete flowchart for agent optimization is as follows: Figure 4 As shown, the complete process includes: acquiring optimization-related information of the agent; preprocessing the optimization-related information of the agent to obtain preprocessed information; inputting the preprocessed information into the agent optimization model, and the agent optimization model outputs optimization method information based on the preprocessed information; and the agent automatically optimizing itself based on the optimization method information.

[0194] The agent optimization method provided in this application automates, personalizes, and enables multi-strategy collaboration in the agent tuning process, significantly reducing manual tuning costs, shortening the optimization cycle, and improving the accuracy and stability of agents in real-world scenarios. This technology is applicable to various industries, such as intelligent customer service, smart office, content generation, and industrial quality inspection, demonstrating broad applicability and replicability. Given the context of enterprises promoting AI application implementation, lowering the barriers to model maintenance, and improving agent response quality, this technology possesses strong commercialization potential and can be integrated into existing AI platforms as an independent service or core module, providing continuous optimization support for enterprises and developers.

[0195] like Figure 5 As shown in the figure, this application embodiment also provides an intelligent agent optimization model training device, the device comprising:

[0196] The first acquisition module 501 is used to acquire agent optimization training data; wherein, the agent optimization training data includes historical optimization-related information of the agent and historical optimization method information corresponding to the historical optimization-related information; the historical optimization method information includes historical optimization methods and evaluation information corresponding to the historical optimization methods;

[0197] The first processing module 502 is used to train the initial model for predicting the optimization method information of the agent using the agent optimization training data to obtain the agent optimization model.

[0198] The agent optimization model is used to receive optimization-related information of the agent and output the optimization method of the agent based on the optimization-related information.

[0199] Optionally, the initial model includes a processing layer and at least two prediction layers, wherein the at least two prediction layers are arranged sequentially.

[0200] The first processing module 502 includes:

[0201] The first processing unit is used to input the historical optimization-related information of the agent into the processing layer, and use the processing layer to perform semantic feature analysis on the historical optimization-related information to obtain the semantic features of the historical optimization-related information.

[0202] The second processing unit is used to input the semantic features into the at least two prediction layers to obtain agent optimization information predicted by each prediction layer, wherein one prediction layer corresponds to one agent optimization information.

[0203] The third processing unit is used to adjust the model parameters of the initial model according to the predicted agent optimization method information and the historical optimization method information to obtain the agent optimization model.

[0204] The agent optimization method information predicted by the first prediction layer of the at least two prediction layers is obtained based on the semantic features.

[0205] The agent optimization method information predicted by the first prediction layer in the at least two prediction layers is obtained based on the semantic features and the agent optimization method information predicted by the second prediction layer. The first prediction layer is the prediction layer other than the first prediction layer in the at least two prediction layers, and the second prediction layer is the prediction layer before the first prediction layer in the at least two prediction layers.

[0206] Optionally, the processing layer includes a background layer, a task layer, and a connection layer;

[0207] The first processing unit is specifically used for:

[0208] The historical optimization-related information of the intelligent agent is subjected to a first processing to obtain a first historical optimization-related information, wherein the first processing includes at least one of the following: integrity check, logical consistency check and outlier detection;

[0209] The capability information of the intelligent agent is obtained based on the first historical optimization information;

[0210] The capability information is input into the background layer, and the first semantic feature of the capability information is extracted using the background layer.

[0211] The first semantic feature is input into the task layer, and the shape of the first semantic feature is adjusted by the task layer to obtain the second semantic feature;

[0212] The first semantic feature and the second semantic feature are input into the connection layer, and the connection layer is used to perform a second processing on the first semantic feature and the second semantic feature to obtain the semantic feature. The second processing includes at least one of the following: linking processing and flattening processing.

[0213] Optionally, the evaluation information corresponding to the historical optimization methods includes optimization accuracy value and overall optimization accuracy value; wherein, the optimization accuracy value is determined based on the historical optimization methods and the number of historical optimization methods, and the overall optimization accuracy is obtained by weighting the F1 scores corresponding to the historical optimization methods;

[0214] The third processing unit is specifically used for:

[0215] The first loss information is obtained based on the predicted agent optimization method information and the historical optimization method information;

[0216] The model parameters of the initial model are adjusted using the first loss information to obtain the agent optimization model;

[0217] The predicted agent optimization method information includes the predicted agent optimization method, the predicted optimization accuracy value, and the predicted comprehensive optimization accuracy value.

[0218] The first loss information is obtained based on the loss information between the predicted agent optimization method and the historical optimization method, the error information between the predicted optimization accuracy value and the optimization accuracy value, and the error information between the predicted comprehensive optimization accuracy value and the comprehensive optimization accuracy value.

[0219] Optionally, the optimization method includes at least one of the following:

[0220] Adjust the parameters of the large model corresponding to the agent;

[0221] Optimize the task planning information and knowledge recommendation information of the intelligent agent;

[0222] Reinforcement learning is applied to the intelligent agent.

[0223] Optionally, the optimization-related information includes at least one of the following:

[0224] User industry classification corresponding to intelligent agents;

[0225] Information describing the application scenarios of intelligent agents;

[0226] Task description information of the intelligent agent;

[0227] Agent version information;

[0228] Information describing the capabilities of an intelligent agent.

[0229] It should be noted that the agent optimization model training device provided in this application embodiment is a device capable of executing the above-described agent optimization model training method. Therefore, all embodiments of the above-described agent optimization model training method are applicable to this device and can achieve the same or similar technical effects.

[0230] like Figure 6 As shown in the figure, this application embodiment also provides an intelligent agent optimization device, the device comprising:

[0231] The second acquisition module 601 is used to acquire optimization-related information of the intelligent agent;

[0232] The second processing module 602 is used to input the optimization-related information of the agent into the agent optimization model to obtain the optimization method information output by the agent optimization model; wherein, the agent optimization model is used to output the optimization method information based on the optimization-related information, and the agent optimization model is trained using agent optimization training data; the optimization method information includes the optimization method and the evaluation information corresponding to the optimization method;

[0233] The third processing module 603 is used to optimize the intelligent agent according to the optimization method information.

[0234] It should be noted that the intelligent agent optimization device provided in this application embodiment is a device capable of executing the above-described intelligent agent optimization method. Therefore, all embodiments of the above-described intelligent agent optimization method are applicable to this device and can achieve the same or similar technical effects.

[0235] like Figure 7 As shown in the figure, this application embodiment also provides an electronic device, including: a processor 701; and a memory 703 connected to the processor 701 via a bus interface 702, the memory 703 being used to store programs and data used by the processor 701 when performing operations, and the processor 701 calling and executing the programs and data stored in the memory 703.

[0236] The transceiver 704 is connected to the bus interface 702 and is used to receive and send data under the control of the processor 701. Specifically, the processor 701 is used to read the program in the memory 703 and to execute the following processes:

[0237] Acquire agent optimization training data; wherein, the agent optimization training data includes historical optimization-related information of the agent and historical optimization method information corresponding to the historical optimization-related information; the historical optimization method information includes historical optimization methods and evaluation information corresponding to the historical optimization methods;

[0238] The initial model used to predict the optimization method information of the agent is trained using the agent optimization training data to obtain the agent optimization model;

[0239] The agent optimization model is used to receive optimization-related information of the agent and output the optimization method of the agent based on the optimization-related information.

[0240] Optionally, the initial model includes a processing layer and at least two prediction layers, wherein the at least two prediction layers are arranged sequentially.

[0241] The processor 701 is used for:

[0242] The historical optimization-related information of the agent is input into the processing layer, and the processing layer performs semantic feature analysis on the historical optimization-related information to obtain the semantic features of the historical optimization-related information.

[0243] The semantic features are input into the at least two prediction layers to obtain the agent optimization method information predicted by each prediction layer, wherein one prediction layer corresponds to one agent optimization method information;

[0244] The model parameters of the initial model are adjusted based on the predicted agent optimization method information and the historical optimization method information to obtain the agent optimization model;

[0245] The agent optimization method information predicted by the first prediction layer of the at least two prediction layers is obtained based on the semantic features.

[0246] The agent optimization method information predicted by the first prediction layer in the at least two prediction layers is obtained based on the semantic features and the agent optimization method information predicted by the second prediction layer. The first prediction layer is the prediction layer other than the first prediction layer in the at least two prediction layers, and the second prediction layer is the prediction layer before the first prediction layer in the at least two prediction layers.

[0247] Optionally, the processing layer includes a background layer, a task layer, and a connection layer;

[0248] Specifically, the processor 701 is used for:

[0249] The historical optimization-related information of the intelligent agent is subjected to a first processing to obtain a first historical optimization-related information, wherein the first processing includes at least one of the following: integrity check, logical consistency check and outlier detection;

[0250] The capability information of the intelligent agent is obtained based on the first historical optimization information;

[0251] The capability information is input into the background layer, and the first semantic feature of the capability information is extracted using the background layer.

[0252] The first semantic feature is input into the task layer, and the shape of the first semantic feature is adjusted by the task layer to obtain the second semantic feature;

[0253] The first semantic feature and the second semantic feature are input into the connection layer, and the connection layer is used to perform a second processing on the first semantic feature and the second semantic feature to obtain the semantic feature. The second processing includes at least one of the following: linking processing and flattening processing.

[0254] Optionally, the evaluation information corresponding to the historical optimization methods includes optimization accuracy value and overall optimization accuracy value; wherein, the optimization accuracy value is determined based on the historical optimization methods and the number of historical optimization methods, and the overall optimization accuracy is obtained by weighting the F1 scores corresponding to the historical optimization methods;

[0255] Specifically, the processor 701 is used for:

[0256] The first loss information is obtained based on the predicted agent optimization method information and the historical optimization method information;

[0257] The model parameters of the initial model are adjusted using the first loss information to obtain the agent optimization model;

[0258] The predicted agent optimization method information includes the predicted agent optimization method, the predicted optimization accuracy value, and the predicted comprehensive optimization accuracy value.

[0259] The first loss information is obtained based on the loss information between the predicted agent optimization method and the historical optimization method, the error information between the predicted optimization accuracy value and the optimization accuracy value, and the error information between the predicted comprehensive optimization accuracy value and the comprehensive optimization accuracy value.

[0260] Optionally, the optimization method includes at least one of the following:

[0261] Adjust the parameters of the large model corresponding to the agent;

[0262] Optimize the task planning information and knowledge recommendation information of the intelligent agent;

[0263] Reinforcement learning is applied to the intelligent agent.

[0264] Optionally, the optimization-related information includes at least one of the following:

[0265] User industry classification corresponding to intelligent agents;

[0266] Information describing the application scenarios of intelligent agents;

[0267] Task description information of the intelligent agent;

[0268] Agent version information;

[0269] Information describing the capabilities of an intelligent agent.

[0270] Among them, Figure 7 In this context, the bus architecture may include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 701) and memory (memory 703). The bus architecture may also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides a user interface 705. A transceiver 704 may be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium. Processor 701 is responsible for managing the bus architecture and general processing, and memory 703 may store data used by processor 701 during operation.

[0271] This application also provides an electronic device, including: a processor; and a memory connected to the processor via a bus interface, the memory being used to store programs and data used by the processor when performing operations, and the processor calling and executing the programs and data stored in the memory.

[0272] The transceiver is connected to the bus interface and is used to receive and send data under the control of the processor.

[0273] The processor is used to read programs from the memory, and the processor is used to execute the following processes:

[0274] Obtain optimization-related information for the intelligent agent;

[0275] The optimization-related information of the agent is input into the agent optimization model to obtain the optimization method information output by the agent optimization model; wherein, the agent optimization model is used to output the optimization method information based on the optimization-related information, and the agent optimization model is trained using agent optimization training data; the optimization method information includes the optimization method and the evaluation information corresponding to the optimization method;

[0276] The agent is optimized based on the optimization method information.

[0277] It should be noted that the terminal device provided in this application embodiment is similar to... Figure 7The structures of the electronic devices shown are basically the same, so they will not be described in detail here.

[0278] In addition, specific embodiments of this application also provide a readable storage medium storing a computer program thereon, wherein when the program is executed by a processor, it implements the steps in the agent optimization model training method as described above, or implements the steps in the agent optimization method as described above.

[0279] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0280] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can be physically included separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0281] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions that cause a computer device (which may be a personal computer, server, or network device, etc.) to execute partial steps of the resource selection method described in the various embodiments of this application, or to execute partial steps of the information transmission method described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0282] A specific embodiment of this application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above-described functionality. Figure 1 or Figure 3 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0283] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for training an intelligent agent optimization model, characterized in that, The method includes: Acquire agent optimization training data; wherein, the agent optimization training data includes historical optimization-related information of the agent and historical optimization method information corresponding to the historical optimization-related information; the historical optimization method information includes historical optimization methods and evaluation information corresponding to the historical optimization methods; The initial model used to predict the optimization method information of the agent is trained using the agent optimization training data to obtain the agent optimization model; The agent optimization model is used to receive optimization-related information of the agent and output the optimization method of the agent based on the optimization-related information.

2. The method according to claim 1, characterized in that, The initial model includes a processing layer and at least two prediction layers, wherein the at least two prediction layers are arranged sequentially. The process involves training an initial model for predicting the agent's optimization methods using the agent's optimization training data to obtain an agent optimization model, including: The historical optimization-related information of the agent is input into the processing layer, and the processing layer performs semantic feature analysis on the historical optimization-related information to obtain the semantic features of the historical optimization-related information. The semantic features are input into the at least two prediction layers to obtain the agent optimization method information predicted by each prediction layer, wherein one prediction layer corresponds to one agent optimization method information; The model parameters of the initial model are adjusted based on the predicted agent optimization method information and the historical optimization method information to obtain the agent optimization model; The agent optimization method information predicted by the first prediction layer of the at least two prediction layers is obtained based on the semantic features. The agent optimization method information predicted by the first prediction layer in the at least two prediction layers is obtained based on the semantic features and the agent optimization method information predicted by the second prediction layer. The first prediction layer is the prediction layer other than the first prediction layer in the at least two prediction layers, and the second prediction layer is the prediction layer before the first prediction layer in the at least two prediction layers.

3. The method according to claim 2, characterized in that, The processing layer includes a background layer, a task layer, and a connection layer; Specifically, the historical optimization-related information of the agent is input into the processing layer, and the processing layer performs semantic feature analysis on the historical optimization-related information to obtain the semantic features of the historical optimization-related information, including: The historical optimization-related information of the intelligent agent is subjected to a first processing to obtain a first historical optimization-related information, wherein the first processing includes at least one of the following: integrity check, logical consistency check and outlier detection; The capability information of the intelligent agent is obtained based on the first historical optimization information; The capability information is input into the background layer, and the first semantic feature of the capability information is extracted using the background layer. The first semantic feature is input into the task layer, and the shape of the first semantic feature is adjusted by the task layer to obtain the second semantic feature; The first semantic feature and the second semantic feature are input into the connection layer, and the connection layer is used to perform a second processing on the first semantic feature and the second semantic feature to obtain the semantic feature. The second processing includes at least one of the following: linking processing and flattening processing.

4. The method according to claim 2, characterized in that, The evaluation information corresponding to the historical optimization methods includes optimization accuracy value and overall optimization accuracy value; wherein, the optimization accuracy value is determined based on the historical optimization methods and the number of historical optimization methods, and the overall optimization accuracy value is obtained by weighting the F1 scores corresponding to the historical optimization methods; The process involves adjusting the model parameters of the initial model based on the predicted agent optimization method information and the historical optimization method information to obtain the agent optimization model, including: The first loss information is obtained based on the predicted agent optimization method information and the historical optimization method information; The model parameters of the initial model are adjusted using the first loss information to obtain the agent optimization model; The predicted agent optimization method information includes the predicted agent optimization method, the predicted optimization accuracy value, and the predicted comprehensive optimization accuracy value. The first loss information is obtained based on the loss information between the predicted agent optimization method and the historical optimization method, the error information between the predicted optimization accuracy value and the optimization accuracy value, and the error information between the predicted comprehensive optimization accuracy value and the comprehensive optimization accuracy value.

5. The method according to any one of claims 1 to 4, characterized in that, The optimization method includes at least one of the following: Adjust the parameters of the large model corresponding to the agent; Optimize the task planning information and knowledge recommendation information of the intelligent agent; Reinforcement learning is applied to the intelligent agent.

6. The method according to any one of claims 1 to 4, characterized in that, The optimization-related information includes at least one of the following: User industry classification corresponding to intelligent agents; Information describing the application scenarios of intelligent agents; Task description information of the intelligent agent; Agent version information; Information describing the capabilities of an intelligent agent.

7. A method for optimizing an intelligent agent, characterized in that, The method includes: Obtain optimization-related information for the intelligent agent; The optimization-related information of the agent is input into the agent optimization model to obtain the optimization method information output by the agent optimization model; wherein, the agent optimization model is used to output the optimization method information based on the optimization-related information, and the agent optimization model is trained using agent optimization training data; the optimization method information includes the optimization method and the evaluation information corresponding to the optimization method; The agent is optimized based on the optimization method information.

8. A training device for an intelligent agent optimization model, characterized in that, The device includes: The first acquisition module is used to acquire agent optimization training data; wherein, the agent optimization training data includes historical optimization-related information of the agent and historical optimization method information corresponding to the historical optimization-related information; the historical optimization method information includes historical optimization methods and evaluation information corresponding to the historical optimization methods; The first processing module is used to train the initial model for predicting the optimization method information of the agent using the agent optimization training data to obtain the agent optimization model. The agent optimization model is used to receive optimization-related information of the agent and output the optimization method of the agent based on the optimization-related information.

9. An intelligent agent optimization device, characterized in that, The device includes: The second acquisition module is used to acquire optimization-related information of the agent; The second processing module is used to input the optimization-related information of the agent into the agent optimization model to obtain the optimization method information output by the agent optimization model; wherein, the agent optimization model is used to output the optimization method information based on the optimization-related information, and the agent optimization model is trained using agent optimization training data; the optimization method information includes the optimization method and the evaluation information corresponding to the optimization method; The third processing module is used to optimize the agent based on the optimization method information.

10. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps in the agent optimization model training method as described in any one of claims 1 to 6, or implements the steps in the agent optimization method as described in claim 7.

11. A readable storage medium, characterized in that, The readable storage medium stores a program that, when executed by a processor, implements the steps in the agent optimization model training method as described in any one of claims 1 to 6, or implements the steps in the agent optimization method as described in claim 7.

12. A computer program product, characterized in that, The method includes computer instructions that, when executed by a processor, implement the steps in the agent optimization model training method as described in any one of claims 1 to 6, or implement the steps in the agent optimization method as described in claim 7.

Images