An automated experimental design closed-loop optimization method based on self-evolution agent
Patent Information
- Application Number
- CN202611020384.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-09
- Publication Date
- 2026-10-23
AI Technical Summary
本发明解决现有技术中启动效率低、非结构化知识无法注入优化框架、无法处理带未知约束的多目标优化以及缺乏任务间经验迁移能力的问题，实现从问题解析到实验验证的全流程自动闭环
[0046]1、通过智能体自动挖掘文献和规则并量化为先验均值与特征，初始实验的有效性相比随机采样或纯贝叶斯优化大大提升，实验轮数大大减少。
Smart Images
Figure CN122965279A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence and automated experiment optimization technology, specifically a closed-loop optimization method for automated experiment design based on a self-evolving intelligent agent. Background Technology
[0002] In existing technologies, the use of artificial intelligence to assist in experimental optimization and parameter search can be mainly divided into the following categories:
[0003] The first category is purely data-driven machine learning methods. These methods train surrogate models using a large amount of experimental data to predict the performance of unknown samples. Their drawbacks include the need for a large amount of labeled data, low prediction accuracy in sparse data scenarios, and the inability to proactively suggest the next experimental step.
[0004] The second category is traditional Bayesian optimization methods. This method uses a Gaussian process as a surrogate model, balancing exploration and utilization through a sampling function to find a relatively optimal solution within tens to hundreds of experiments. However, it has three inherent drawbacks: First, it requires manual definition of the parameter space, prior distribution, and kernel function form, making it highly dependent on domain knowledge; random sampling during cold starts leads to low experimental efficiency. Second, it cannot automatically utilize unstructured knowledge such as literature and expert rules, which are difficult to transform into inputs for Bayesian optimization. Third, it cannot handle optimization problems with unknown constraints, such as certain parameter combinations that, while superior in performance, are physically impossible to synthesize or stably exist.
[0005] The third category is agent-based methods using large language models. In recent years, there have been works utilizing large language models for experimental planning and literature review. However, most existing agents output natural language suggestions, lack a rigorous mathematical optimization framework, cannot quantify uncertainty, and cannot guarantee convergence to the optimal solution based on experimental feedback.
[0006] In summary, existing technologies fail to effectively integrate the formalization of unstructured knowledge, the sample efficiency of Bayesian optimization, and the continuous learning ability of agents, resulting in long experimental cycles, high costs, and poor transferability in experimental optimization tasks.
[0007] Therefore, to address the above problems, an automated experimental design closed-loop optimization method based on self-evolving intelligent agents is needed. Summary of the Invention
[0008] The purpose of this invention is to provide an automated closed-loop optimization method for experimental design based on self-evolving intelligent agents. This invention addresses the problems of low startup efficiency, inability to inject unstructured knowledge into the optimization framework, inability to handle multi-objective optimization with unknown constraints, and lack of experience transfer capabilities between tasks in existing technologies, achieving a fully automated closed-loop process from problem analysis to experimental verification.
[0009] This invention is implemented as follows:
[0010] This invention provides an automated closed-loop optimization method for experimental design based on self-evolving intelligent agents, comprising:
[0011] The system receives experimental optimization requirements text described by users in natural language, parses it to obtain optimization objectives, parameter types and constraints, extracts domain prior knowledge and transforms it into a structured task definition that can be recognized by Bayesian optimization.
[0012] Based on the structured task definition, a Gaussian process regression model for predicting continuous performance indicators and a Gaussian process classification model for predicting feasibility constraints are constructed. The posterior distributions and posterior means of the two models are then output through model optimization. Posterior variance And the feasibility probability of the Gaussian process classification model. ;
[0013] Based on the posterior distributions of the two models, a constrained acquisition function is constructed. A batch acquisition strategy is adopted to output multiple sets of candidate parameter vectors, which are then sent to automated experimental equipment to perform experiments, and new experimental data containing parameter vectors, performance values and feasibility labels are acquired.
[0014] The new experimental data is incorporated into the existing dataset, the kernel matrix is reconstructed, and the hyperparameters and posterior distributions of the two models are updated. The updated posterior distribution is used to conduct the next round of experimental optimization and text parameter collection.
[0015] The system stores the empirical data from each round of experiments. When the accumulated data reaches a threshold, it triggers rule induction, mines effective parameter rules, and corrects the prior mean function. The optimization objective is then obtained through the analysis of the corrected prior mean function.
[0016] Furthermore, the parsing yields the optimization objective, parameter type, and constraints. A top-level agent, comprising a large language model, a memory bank, and a skill bank, receives the user's request text and identifies three types of information: optimization objective, parameter type, and constraints.
[0017] Furthermore, the top-level intelligent agent first extracts prior knowledge, retrieves historical experimental records from the memory bank, calls the large language model to parse domain literature and expert rule documents, and extracts unstructured prior knowledge, including the performance preferences of parameter combinations and the taboo thresholds of parameter values;
[0018] The unstructured prior knowledge is then transformed into mathematical objects adapted to the Bayesian optimization engine, outputting a structured task definition.
[0019] Discrete parameters are mapped to learnable embedding vectors, a legal range of values is defined for continuous parameters, a parameter space is constructed, and the parameter vector is output. Where D is the total number of parameter dimensions;
[0020] Based on the performance preference region from prior knowledge, an initial prior distribution is constructed, and a higher initial mean is set for the corresponding parameter region. The prior mean function is then output. Further clarify the definitions of the objective function and constraints, where the objective function is a continuous performance index. The constraint definition is transformed into discrete feasibility labels. ,in This indicates that the parameter combination is feasible. This indicates that it is not feasible.
[0021] Furthermore, in the Gaussian process regression model, the model input is a parameter vector. The output is a performance value. Gaussian process regression satisfies the expression:
[0022] in, Let be the performance function to be predicted. It is a Gaussian process distribution. and Let them be the parameter vectors in the parameter space. It is the prior mean function. For kernel functions;
[0023] Construct a Gaussian process classification model, with the model input being a parameter vector. The output is a feasibility label. ;
[0024] Then, hyperparameter optimization is performed by maximizing the marginal likelihood to estimate the kernel hyperparameters of the two models.
[0025] Furthermore, the acquisition function is constructed; based on the posterior distributions of the two surrogate models, a constrained improved acquisition function for the expected hypervolume is constructed, with the following expression:
[0026] in parameter vector The value of the acquisition function, To convert the parameter vector The hypervolume increment after incorporating the current Pareto front is calculated from the posterior mean and posterior variance of the Gaussian process regression. This represents the feasibility probability threshold.
[0027] The batch acquisition strategy described above uses a virtual point sequence acquisition strategy to generate a batch parameter set. Each time, the parameter vector with the largest current acquisition function value is selected, and the posterior mean of that point is used as a virtual observation value to temporarily add it to the dataset. After updating the posterior distribution of the two models, the next parameter vector is selected until the preset batch size q is reached.
[0028] The experiment execution and data feedback are described below: A batch of parameter sets is passed to the underlying experimental interface and transformed into execution instructions for automated experimental equipment; after the experiment is completed, raw measurement data is collected, and the performance value y and feasibility label z are parsed to generate new experimental data pairs. And store it in the database.
[0029] Furthermore, the reconstruction of the kernel matrix and updating of the hyperparameters and posterior distributions of the two models includes dataset augmentation, adding q pairs of new experimental data to the existing dataset. This yields the augmented dataset;
[0030] The kernel matrix reconstruction and hyperparameter optimization include recalculating the kernel matrix based on the augmented dataset, performing a marginal likelihood optimization process, and updating the kernel hyperparameters of the Gaussian process regression model and the Gaussian process classification model.
[0031] The posterior distribution update includes recalculating the posterior mean and posterior variance of the Gaussian process regression model based on the updated hyperparameters, and simultaneously updating the posterior approximation of the Gaussian process classification model based on the augmented dataset to obtain the updated feasibility probability.
[0032] Furthermore, the storage of experience data for each round of experiments includes the top-level agent storing the parameter vector for each round of experiments. The performance value y and the feasibility label z are stored as an experience record in the memory bank;
[0033] Rule induction triggers frequent pattern mining, including triggering the rule induction process when the number of new experiences added to the memory bank reaches a set threshold: This involves all... The successful experimental parameter vectors are used as the transaction set S. The frequencies of discrete parameters are directly statistically analyzed, while continuous parameters are discretized over intervals. The support of each parameter subspace R is calculated.
[0034] in, Let S be the number of parameter vectors in set S that belong to subregion R. The total number of successful experiments; sub-regions with support higher than the minimum support threshold are marked as candidate rules; the confidence score of each candidate rule is calculated:
[0035] in, This contains all experimental data in the memory bank. The number of parameter vectors belonging to subregion R in all experiments;
[0036] Next, the prior mean of rule storage is corrected, including storing the inductively derived rules in memory in both natural language and structured conditional forms. During new task initialization, if the parameter space of the new task overlaps with the existing rule coverage area, the prior mean function is corrected first, as shown in the following expression:
[0037]
[0038] in, This is the corrected prior mean function. The prior mean function before correction. The strength coefficient, The confidence level of the matching rule R; the modified prior mean function, used as the initial prior distribution for the new task.
[0039] Furthermore, this invention provides an automated experimental design and closed-loop optimization system based on a self-evolving intelligent agent:
[0040] It includes a top-level intelligent agent module, which receives user natural language requests, performs problem parsing, prior knowledge extraction and knowledge formalization, and outputs a structured task definition; it is also responsible for rule induction and prior knowledge correction, and realizes cross-task knowledge transfer.
[0041] The mid-level Bayesian optimization engine module is used to build Gaussian process regression models and Gaussian process classification models, perform hyperparameter optimization, posterior distribution calculation, construction of constrained acquisition functions and batch parameter acquisition, and output candidate experimental parameters.
[0042] The underlying experiment execution module is used to receive candidate experiment parameters, drive automated experimental equipment to complete experiments, collect and parse experimental data, and generate experimental data containing parameters, performance and feasibility labels.
[0043] The storage module stores historical experimental data and the rule knowledge base derived from induction, providing the top-level intelligent agent module with prior retrieval and experience storage capabilities.
[0044] Furthermore, the present invention provides a computer storage medium storing a computer program that can run on a processor, wherein the processor executes the computer program to implement the automated experimental design closed-loop optimization method based on a self-evolving intelligent agent as described above.
[0045] Compared with the prior art, the beneficial effects of the present invention are:
[0046] 1. By automatically mining literature and rules through intelligent agents and quantifying them into prior means and features, the effectiveness of the initial experiment is greatly improved compared with random sampling or pure Bayesian optimization, and the number of experimental rounds is greatly reduced.
[0047] It can handle multi-objective optimization with unknown constraints. The dual surrogate model models performance and feasibility separately. The acquisition function naturally avoids infeasible parameters, thus avoiding experimental waste caused by synthesis failure in traditional methods.
[0048] 2. It has the ability to continuously learn and transfer tasks. The memory bank and rule induction mechanism enable the system to share experience between different optimization tasks. New tasks do not have to start from scratch. The efficiency increases exponentially with long-term use. The decision-making process is interpretable. The high-level rules output by the intelligent agent can be presented to the R&D personnel, which helps to understand the model behavior and accumulate domain knowledge. Attached Figure Description
[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.
[0051] Figure 1 This is a flowchart of the method of the present invention.
[0052] Figure 2 This is a system structure diagram of the present invention;
[0053] Figure 3 This is a structural diagram of the storage medium of the present invention. Detailed Implementation
[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to describe selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0056] Example 1
[0057] Please see Figure 1 This invention provides an automated experimental design closed-loop optimization method based on self-evolving intelligent agents, comprising:
[0058] The system receives experimental optimization requirements text described by users in natural language, parses it to obtain optimization objectives, parameter types and constraints, extracts domain prior knowledge and transforms it into a structured task definition that can be recognized by Bayesian optimization.
[0059] Based on the structured task definition, a Gaussian process regression model for predicting continuous performance indicators and a Gaussian process classification model for predicting feasibility constraints are constructed. The posterior distributions and posterior means of the two models are then output through model optimization. Posterior variance And the feasibility probability of the Gaussian process classification model. ;
[0060] Based on the posterior distributions of the two models, a constrained acquisition function is constructed. A batch acquisition strategy is adopted to output multiple sets of candidate parameter vectors, which are then sent to automated experimental equipment to perform experiments, and new experimental data containing parameter vectors, performance values and feasibility labels are acquired.
[0061] The new experimental data is incorporated into the existing dataset, the kernel matrix is reconstructed, and the hyperparameters and posterior distributions of the two models are updated. The updated posterior distribution is used to conduct the next round of experimental optimization and text parameter collection.
[0062] The system stores the empirical data from each round of experiments. When the accumulated data reaches a threshold, it triggers rule induction, mines effective parameter rules, and corrects the prior mean function. The optimization objective is then obtained through the analysis of the corrected prior mean function.
[0063] In this embodiment, the parsing obtains the optimization objective, parameter type, and constraints. The top-level agent, which includes a large language model, a memory bank, and a skill bank, receives the user's requirement text and identifies three types of information, including the optimization objective, parameter type, and constraints. The optimization objective is the performance index to be maximized or minimized, the parameter type includes discrete parameters and continuous parameters, and the constraints include synthesis feasibility, performance threshold, and taboo parameter combinations.
[0064] In this embodiment, the top-level agent first extracts prior knowledge, retrieves historical experimental records from the memory bank, calls the large language model to parse domain literature and expert rule documents, and extracts unstructured prior knowledge, including the performance preferences of parameter combinations and the taboo thresholds of parameter values.
[0065] The unstructured prior knowledge is then transformed into mathematical objects adapted to the Bayesian optimization engine, outputting a structured task definition.
[0066] Discrete parameters are mapped to learnable embedding vectors, a legal range of values is defined for continuous parameters, a parameter space is constructed, and the parameter vector is output. Where D is the total number of parameter dimensions;
[0067] Based on the performance preference region from prior knowledge, an initial prior distribution is constructed, and a higher initial mean is set for the corresponding parameter region. The prior mean function is then output. Further clarify the definitions of the objective function and constraints, where the objective function is a continuous performance index. The constraint definition is transformed into discrete feasibility labels. ,in This indicates that the parameter combination is feasible. This indicates that it is not feasible.
[0068] In this embodiment, the construction of the Gaussian process regression model involves a parameter vector as the model input. The output is a performance value. Gaussian process regression satisfies the expression:
[0069] in, Let be the performance function to be predicted. It is a Gaussian process distribution. and Let them be the parameter vectors in the parameter space. It is the prior mean function. For kernel functions;
[0070] The kernel function uses a Matern3 / 2 kernel with automatic correlation determination, and its expression is:
[0071]
[0072] in, For signal variance, and They are parameter vectors and The Dimensional value, For the first The length-scale hyperparameter of the dimension parameter is used to measure the weight of the corresponding parameter's impact on output performance;
[0073] The observation model is ,in, To observe the noise, For noise variance;
[0074] The aforementioned construction of a Gaussian process classification model is used to predict feasibility constraints. The model input is a parameter vector. The output is a feasibility label. Let implicit function obey: in For the mean function of the classification model, the mean of the high-quality parameter region is set to a positive value based on the domain rules. The kernel function of the classification model; parameter vector. The corresponding feasibility probability is ,in The standard normal cumulative distribution function is used; the observed value z follows a Bernoulli distribution. Posterior inference is performed using the Laplace approximation or expectation propagation method.
[0075] Next, hyperparameter optimization is performed by maximizing the marginal likelihood to estimate the kernel hyperparameters of both models. The logarithm of the marginal likelihood of the Gaussian process regression model is:
[0076]
[0077] in, For kernel hyperparameter set The vector of observed objective function values. Let be the prior mean vector. The kernel matrix after adding noise. The Element is N is the number of observed samples; the optimal hyperparameters are solved using the L-BFGS-B algorithm.
[0078] In this embodiment, the acquisition function is constructed as follows: Based on the posterior distributions of the two surrogate models, a constrained improved acquisition function for the desired hypervolume is constructed, with the expression:
[0079]
[0080] in parameter vector The value of the acquisition function, To convert the parameter vector The hypervolume increment after incorporating the current Pareto front is calculated from the posterior mean and posterior variance of the Gaussian process regression. The feasibility probability threshold, the indicator function exist The value is 1 if the condition is met, and 0 otherwise. For the currently available experimental dataset, The expectation operator with respect to the posterior distribution is approximated by Monte Carlo integration;
[0081] The batch acquisition strategy employs a virtual point sequence acquisition strategy to generate a batch parameter set. Each time, the parameter vector with the largest current acquisition function value is selected, and the posterior mean of that point is temporarily added to the dataset as a virtual observation. After updating the posterior distributions of the two models, the next parameter vector is selected, continuing until the preset batch size q is reached. The optimization objective of batch acquisition is:
[0082]
[0083] in, The selected q parameter vectors form a batch set. For parameter space, For excess volume indicators, For the current dataset The corresponding Pareto front;
[0084] The experiment execution and data feedback are described below: A batch of parameter sets is passed to the underlying experimental interface and transformed into execution instructions for automated experimental equipment; after the experiment is completed, raw measurement data is collected, and the performance value y and feasibility label z are parsed to generate new experimental data pairs. And store it in the database.
[0085] In this embodiment, the reconstruction of the kernel matrix and updating the hyperparameters and posterior distributions of the two models includes dataset augmentation, which involves adding q pairs of new experimental data to the existing dataset. The augmented dataset is obtained as follows:
[0086] ;
[0087] The kernel matrix reconstruction and hyperparameter optimization include recalculating the kernel matrix based on the augmented dataset, performing a marginal likelihood optimization process, and updating the kernel hyperparameters of the Gaussian process regression model and the Gaussian process classification model.
[0088] The posterior distribution update includes recalculating the posterior mean and posterior variance of the Gaussian process regression model based on the updated hyperparameters, expressed as:
[0089]
[0090] in, Let be the vector of parameters to be measured. For the prior mean function in The value at that location, The parameter vector to be measured The kernel function value vector is the vector of all observed parameter vectors. The inverse of the augmented kernel matrix. The augmented objective function value vector. The prior mean vector is used; the posterior approximation of the Gaussian process classification model is updated simultaneously based on the augmented dataset to obtain the updated feasibility probability.
[0091] In this embodiment, the experience storage includes the top-level agent storing the parameter vectors of each round of experiments. The performance value y and the feasibility label z are stored as an experience record in the memory bank;
[0092] Rule induction triggers frequent pattern mining, including triggering the rule induction process when the number of new experiences added to the memory bank reaches a set threshold: This involves all... The successful experimental parameter vectors are used as the transaction set S. For discrete parameters, frequencies are directly counted; for continuous parameters, interval discretization is performed. The support of each parameter subspace R is calculated.
[0093] in, Let S be the number of parameter vectors in set S that belong to subregion R. The total number of successful experiments; sub-regions with support higher than the minimum support threshold are marked as candidate rules; the confidence score of each candidate rule is calculated:
[0094]
[0095] in, This contains all experimental data in the memory bank. The number of parameter vectors belonging to subregion R in all experiments;
[0096] In this embodiment, prior mean correction for rule storage is performed, including storing the inductively derived rules in memory in both natural language and structured condition formats. During new task initialization, if the new task parameter space overlaps with the existing rule coverage area, the prior mean function is corrected first, as shown in the following expression:
[0097]
[0098] in, This is the corrected prior mean function. The prior mean function before correction. The strength coefficient, The confidence level of the matching rule R; the modified prior mean function, used as the initial prior distribution for the new task.
[0099] Example 2
[0100] In this embodiment, the present invention provides an automated experimental design and closed-loop optimization system based on a self-evolving intelligent agent:
[0101] It includes a top-level intelligent agent module, which receives user natural language requests, performs problem parsing, prior knowledge extraction and knowledge formalization, and outputs a structured task definition; it is also responsible for rule induction and prior knowledge correction, and realizes cross-task knowledge transfer.
[0102] The mid-level Bayesian optimization engine module is used to build Gaussian process regression models and Gaussian process classification models, perform hyperparameter optimization, posterior distribution calculation, construction of constrained acquisition functions and batch parameter acquisition, and output candidate experimental parameters.
[0103] The underlying experiment execution module is used to receive candidate experiment parameters, drive automated experimental equipment to complete experiments, collect and parse experimental data, and generate experimental data containing parameters, performance and feasibility labels.
[0104] The storage module stores historical experimental data and the rule knowledge base derived from induction, providing the top-level intelligent agent module with prior retrieval and experience storage capabilities.
[0105] Example 3
[0106] In this embodiment, the present invention provides a computer storage medium storing a computer program that can run on a processor. When the processor executes the computer program, it implements the automated experimental design closed-loop optimization method based on a self-evolving intelligent agent as described above.
[0107] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations will be apparent to those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An automated experimental design closed-loop optimization method based on self-evolving intelligent agents, characterized in that: include: The system receives experimental optimization requirements text described by users in natural language, parses it to obtain optimization objectives, parameter types and constraints, extracts domain prior knowledge and transforms it into a structured task definition that can be recognized by Bayesian optimization. Based on the structured task definition, a Gaussian process regression model for predicting continuous performance indicators and a Gaussian process classification model for predicting feasibility constraints are constructed. The posterior distributions and posterior means of the two models are then output through model optimization. Posterior variance And the feasibility probability of the Gaussian process classification model. ; Based on the posterior distribution of the two models, a constrained acquisition function is constructed. A batch acquisition strategy is adopted to output multiple sets of candidate parameter vectors, which are then sent to automated experimental equipment to perform experiments, and new experimental data containing parameter vectors, performance values and feasibility labels are acquired. The new experimental data is incorporated into the existing dataset, the kernel matrix is reconstructed, and the hyperparameters and posterior distributions of the two models are updated. The updated posterior distribution is used to conduct the next round of experimental optimization and text parameter collection. The system stores the empirical data from each round of experiments. When the accumulated data reaches a threshold, it triggers rule induction, mines effective parameter rules, and corrects the prior mean function. The optimization objective is then obtained through the analysis of the corrected prior mean function.
2. The automated experimental design closed-loop optimization method based on self-evolving intelligent agents according to claim 1, characterized in that, The aforementioned parsing yields the optimization objective, parameter type, and constraints. A top-level agent, comprising a large language model, a memory bank, and a skill bank, receives the user's request text and identifies three types of information: optimization objective, parameter type, and constraints.
3. The automated experimental design closed-loop optimization method based on self-evolving intelligent agents according to claim 2, characterized in that, The top-level intelligent agent first extracts prior knowledge, retrieves historical experimental records from the memory bank, calls the large language model to parse domain literature and expert rule documents, and extracts unstructured prior knowledge, including the performance preferences of parameter combinations and the taboo thresholds of parameter values; The unstructured prior knowledge is then transformed into mathematical objects adapted to the Bayesian optimization engine, outputting a structured task definition. Discrete parameters are mapped to learnable embedding vectors, valid ranges of values are defined for continuous parameters, a parameter space is constructed, and parameter vectors are output. Where D is the total number of parameter dimensions; Based on the performance preference region from prior knowledge, an initial prior distribution is constructed, and a higher initial mean is set for the corresponding parameter region. The prior mean function is then output. Further clarify the definitions of the objective function and constraints, where the objective function is a continuous performance index. The constraint definition is transformed into discrete feasibility labels. ,in This indicates that the parameter combination is feasible. This indicates that it is not feasible.
4. The automated experimental design closed-loop optimization method based on self-evolving intelligent agents according to claim 1, characterized in that, The Gaussian process regression model is described above, where the model input is a parameter vector. The output is a performance value. Gaussian process regression satisfies the expression: in, Let be the performance function to be predicted. It is a Gaussian process distribution. and Let them be the parameter vectors in the parameter space. It is the prior mean function. For kernel functions; Construct a Gaussian process classification model, with the model input being a parameter vector. The output is a feasibility label. ; Then, hyperparameter optimization is performed by maximizing the marginal likelihood to estimate the kernel hyperparameters of the two models.
5. The automated experimental design closed-loop optimization method based on self-evolving intelligent agents according to claim 1, characterized in that, Acquisition function construction: Based on the posterior distributions of the two surrogate models, a constrained improved acquisition function for the expected hypervolume is constructed, with the following expression: in For parameter vectors The value of the acquisition function, To convert the parameter vector The hypervolume increment after incorporating the current Pareto front is calculated from the posterior mean and posterior variance of the Gaussian process regression. This represents the feasibility probability threshold. The batch acquisition strategy described above uses a virtual point sequence acquisition strategy to generate a batch parameter set. Each time, the parameter vector with the largest current acquisition function value is selected, and the posterior mean of that point is used as a virtual observation value to temporarily add it to the dataset. After updating the posterior distribution of the two models, the next parameter vector is selected until the preset batch size q is reached. The experiment execution and data feedback are described below: A batch of parameter sets is passed to the underlying experimental interface and transformed into execution instructions for automated experimental equipment; after the experiment is completed, raw measurement data is collected, and the performance value y and feasibility label z are parsed to generate new experimental data pairs. And store it in the database.
6. The automated experimental design closed-loop optimization method based on self-evolving intelligent agents according to claim 1, characterized in that, The process of reconstructing the kernel matrix and updating the hyperparameters and posterior distributions of the two models includes dataset augmentation, which involves adding q pairs of new experimental data to the existing dataset. This yields the augmented dataset; The kernel matrix reconstruction and hyperparameter optimization include recalculating the kernel matrix based on the augmented dataset, performing a marginal likelihood optimization process, and updating the kernel hyperparameters of the Gaussian process regression model and the Gaussian process classification model. The posterior distribution update includes recalculating the posterior mean and posterior variance of the Gaussian process regression model based on the updated hyperparameters, and simultaneously updating the posterior approximation of the Gaussian process classification model based on the augmented dataset to obtain the updated feasibility probability.
7. The automated experimental design closed-loop optimization method based on self-evolving intelligent agents according to claim 1, characterized in that: The aforementioned storage of experience data for each round of experiments includes the top-level agent storing the parameter vectors for each round of experiments. The performance value y and the feasibility label z are stored as an experience record in the memory bank; Rule induction triggers frequent pattern mining, including triggering the rule induction process when the number of new experiences added to the memory bank reaches a set threshold: This involves all... The successful experimental parameter vectors are used as the transaction set S. The frequencies of discrete parameters are directly statistically analyzed, while continuous parameters are discretized over intervals. The support of each parameter subspace R is calculated. in, Let S be the number of parameter vectors in set S that belong to subregion R. The total number of successful experiments; sub-regions with support higher than the minimum support threshold are marked as candidate rules; the confidence score of each candidate rule is calculated: in, This contains all experimental data in the memory bank. This represents the number of parameter vectors belonging to subregion R in all experiments.
8. The automated experimental design closed-loop optimization method based on self-evolving intelligent agents according to claim 7, characterized in that: Next, the prior mean of rule storage is corrected, including storing the inductively derived rules in memory in both natural language and structured conditional forms. During new task initialization, if the parameter space of the new task overlaps with the coverage area of existing rules, the prior mean function is corrected first, as shown in the following expression: in, This is the corrected prior mean function. The prior mean function before correction. The strength coefficient, The confidence score for matching rule R; the modified prior mean function, used as the initial prior distribution for the new task.
9. An automated experimental design and closed-loop optimization system based on a self-evolving intelligent agent, characterized in that: It includes a top-level intelligent agent module, which receives user natural language requests, performs problem parsing, prior knowledge extraction and knowledge formalization, and outputs a structured task definition; it is also responsible for rule induction and prior knowledge correction, and realizes cross-task knowledge transfer. The mid-level Bayesian optimization engine module is used to build Gaussian process regression models and Gaussian process classification models, perform hyperparameter optimization, posterior distribution calculation, construction of constrained acquisition functions and batch parameter acquisition, and output candidate experimental parameters. The underlying experiment execution module is used to receive candidate experiment parameters, drive automated experimental equipment to complete experiments, collect and parse experimental data, and generate experimental data containing parameters, performance and feasibility labels. The storage module stores historical experimental data and the rule knowledge base derived from induction, providing the top-level intelligent agent module with prior retrieval and experience storage capabilities.
10. A computer storage medium storing a computer program executable on a processor, characterized in that, When the processor executes the computer program, it implements the automated experimental design closed-loop optimization method based on a self-evolving intelligent agent as described in any one of claims 1-8.