Recipe recommendation method, system, medium and device for multi-objective Bayesian optimization
Through multi-objective Bayesian optimization and Gaussian process regression agent model, the problem of inefficiency in traditional formula research and development is solved, and the rapid and accurate determination of formula experiments is achieved, reducing resource consumption.
Patent Information
- Application Number
- CN202410522622.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-28
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-04-28
AI Technical Summary
A lot of experiments are needed in the research and development of traditional formulas to explore the optimal product formula, resulting in inefficiency and waste of resources. Especially in the complex food and chemical industry, it is difficult to quickly determine the optimal formula.
Using multi-objective Bayesian optimization method, a Gaussian process regression proxy model is constructed by obtaining historical formula data for preprocessing, and iterative optimization is used to reduce the number of trial and error. The proxy model predicts performance and gradually approach the optimal formula.
The number of trial and errors from scratch is greatly reduced, the efficiency and accuracy of formula experiments are improved, resource consumption is reduced, and the rapid determination of the best formula is achieved.
Smart Images

Figure CN118349587B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of chemical formulation technology, and in particular to a multi-objective Bayesian optimization formulation recommendation method, system, medium and equipment. Background Art
[0002] With the rapid development of industry, the quality requirements of various industrial products are increasing day by day. In order to obtain products with ideal performance, formula experiments are particularly important in industrial production. Formula experiments explore the formula with the best performance through the combination of different raw materials. Especially in the field of food chemical industry, which is a highly complex and changeable industry, the design and optimization of formulas has always been a key challenge in scientific research and production.
[0003] Currently, in traditional formula research and development, researchers usually rely on experience or comprehensive experimental design to explore the optimal formula of a product. That is, a large number of experiments are needed to ensure that all possible ratios are covered, so as to find the optimal formula in the experimental data. This traditional formula experimental method requires a lot of time and resources, resulting in low efficiency in determining the optimal formula of the product. Summary of the invention
[0004] The present application provides a multi-objective Bayesian optimization recipe recommendation method, system, medium and device, which can improve the efficiency of determining the optimal recipe of a product.
[0005] In a first aspect, the present application provides a recipe recommendation method for multi-objective Bayesian optimization, the method comprising:
[0006] Obtain multiple historical formulas corresponding to the target product;
[0007] Performing data preprocessing on the data in each of the historical recipes to obtain a sample data set;
[0008] Constructing a proxy model based on the sample data set and Gaussian process regression;
[0009] Obtaining an Nth formula to be tested corresponding to the target product, and obtaining a corresponding Nth detection value according to the Nth formula and the proxy model;
[0010] Obtain the N+1th recipe, take the N+1th recipe as the Nth recipe, and execute the step of obtaining the corresponding Nth detection value according to the Nth recipe and the proxy model, until the N+Mth detection value corresponding to the N+Mth recipe meets the preset convergence condition, and take the N+Mth recipe as the recommended recipe, where N and M are both positive integers.
[0011] By adopting the above technical solution, the number of trial and error from scratch can be greatly reduced through multiple historical formulas corresponding to the target product. The historical formula data is preprocessed, and the obtained sample data set can more accurately reflect the relationship between the ratio of each raw material and the product performance. The Gaussian process regression proxy model constructed using the sample data set provides an efficient prediction model for the formula experiment. By predicting the performance of the unknown ratio experiment, the number of experiments is reduced by using the proxy model prediction, and the resource consumption is also reduced. By obtaining the Nth formula to be tested corresponding to the target product and determining the test value using the proxy model, the method realizes the continuous optimization of the experimental process. The N+1th formula is continuously iterated to obtain and its test value is evaluated until the test value of the N+Mth formula meets the preset convergence condition. This iterative process ensures that the formula experiment gradually approaches the optimal solution, rather than blindly conducting a large number of invalid experiments. This iterative optimization strategy greatly accelerates the process of determining the best formula and improves the accuracy and efficiency of the entire formula research and development.
[0012] Optionally, the historical formula includes each historical material, the material dosage corresponding to each historical material, and the detection index, and the data in each historical formula is preprocessed to obtain a sample data set, including: obtaining a standard dosage range for each material dosage; filtering out historical formulas whose material dosage exceeds the standard dosage range to obtain filtered historical formulas; encoding each historical material, each material dosage, and the detection index in the filtered historical formula to obtain a data coding set; and supplementing data with missing values in the data coding set with 0 to obtain a sample data set.
[0013] By adopting the above technical solution, by setting a standard dosage range and screening out historical formulas that exceed the range, this method ensures that the historical data considered for the model is within a certain reasonable range, excludes those formula data that may cause performance deviations due to abnormal dosage, encodes the screened historical formulas, and converts historical materials, material dosages, and test indicators into data encoding sets. This process converts unstructured or semi-structured raw data into structured data suitable for machine learning algorithm processing. For missing values in the data encoding set, a 0-filling strategy is used to improve the sample data set. This processing method is simple and efficient, and avoids the problem of data incompleteness caused by missing values, thereby ensuring the integrity of the data set and providing support for stable training of the model.
[0014] Optionally, constructing a proxy model based on the sample data set and Gaussian process regression includes: dividing the sample data set into a training set and a test set; obtaining a kernel function and hyperparameters in a pre-constructed Gaussian process regression, performing training based on the kernel function and the training set, and adjusting the hyperparameters during the training process to obtain an initial proxy model; and testing and optimizing the initial proxy model based on the test set to obtain the optimized proxy model.
[0015] By adopting the above technical solution, the Gaussian process regression algorithm is used to build a proxy model to achieve fast and efficient prediction of the performance of the recipe. The sample data set is first divided into a training set and a test set, and then the preset kernel function and hyperparameters are obtained. The model is trained based on the training set, and the hyperparameters are adjusted during the training process to obtain the initial proxy model. The initial model is then tested and optimized using an independent test set to enhance its generalization ability, and finally a proxy model with better performance is obtained.
[0016] Optionally, obtaining the corresponding Nth detection value according to the Nth recipe and the proxy model includes: predicting the Nth recipe based on the proxy model to obtain a predicted value; obtaining a real detection value of an experiment on the Nth recipe based on the predicted value, using the real detection value as the Nth detection value, and saving the Nth recipe and the Nth detection value to a database.
[0017] By adopting the above technical solution, using the proxy model to predict first can quickly guide the experimental direction and reduce the time cost of blind trial and error; however, experimental verification must be carried out because the model prediction may be biased. Feedback and saving the measured results can not only correct the model bias, but also provide new sample data for subsequent model retraining, so that the proxy model performance can be continuously optimized.
[0018] Optionally, the method further includes: calculating the accuracy of the proxy model's prediction of the Nth recipe based on the predicted value and the actual detection value; when the accuracy is lower than an accuracy threshold, displaying a warning message to indicate that the proxy model prediction is abnormal.
[0019] By adopting the above technical solution, after obtaining the predicted value and the measured value of the new formula, the error between the two is calculated to evaluate the prediction accuracy of the proxy model. The calculated accuracy is compared with the preset threshold. If the model accuracy is lower than the set threshold, it is judged that the model prediction ability has decreased, and an early warning message will be displayed to indicate that the model is abnormal, indicating that it needs to be retrained and parameters optimized to improve its generalization performance.
[0020] Optionally, after determining the Nth detection value according to the Nth recipe and the proxy model, the method further includes: iteratively updating the proxy model based on the Nth detection value and the Nth recipe to obtain an updated proxy model; taking the N+1th recipe as the Nth recipe, and executing the step of obtaining the corresponding Nth detection value according to the Nth recipe and the proxy model, until the N+Mth detection value corresponding to the N+Mth recipe meets the preset convergence condition, and then taking the N+Mth recipe as the recommended recipe, including: taking the N+1th recipe as the Nth recipe, and executing the step of obtaining the corresponding Nth detection value according to the Nth recipe and the updated proxy model, until the N+Mth detection value corresponding to the N+Mth recipe meets the preset convergence condition, and then taking the N+Mth recipe as the recommended recipe.
[0021] By adopting the above technical solution, after obtaining the Nth recipe and the corresponding Nth detection value, the new data feedback is used to retrain the proxy model to iteratively update the model parameters to improve its prediction performance. Then, based on the updated model, the previous prediction process is repeated to continuously obtain the detection values of the new recipe.
[0022] Optionally, obtaining the N+1th formula includes: determining whether the Nth detection value meets a preset convergence condition; if the Nth detection value meets the preset convergence condition, using the Nth formula corresponding to the Nth detection value as a recommended formula; if the Nth detection value does not meet the preset convergence condition, obtaining multiple target requirements of the user for the target product, and determining an objective function based on the target requirements; within the standard usage range of each material usage, optimizing the objective function based on the acquisition function to obtain the N+1th formula.
[0023] By adopting the above technical solution, it is first determined whether the Nth detection value corresponding to the previous round of formula meets the preset convergence condition. If so, the Nth formula is directly used as the final recommended formula. If not, the user's product optimization goal is further obtained and converted into a computable objective function. Then, under the premise of meeting the constraint range of the amount of each component material, the acquisition function is used to sample and optimize the objective function to generate a new round of N+1th formula.
[0024] In a second aspect of the present application, a multi-objective Bayesian optimization recipe recommendation system is provided, the system comprising:
[0025] A historical formula acquisition module is used to acquire multiple historical formulas corresponding to a target product;
[0026] A data preprocessing module, used for preprocessing the data in each of the historical recipes to obtain a sample data set;
[0027] An agent module construction module, used to construct an agent model based on the sample data set and Gaussian process regression;
[0028] A detection value determination module, used for obtaining an Nth formula to be tested corresponding to the target product, and determining an Nth detection value according to the Nth formula and the proxy model;
[0029] A loop iteration module is provided to obtain the N+1th formula, take the N+1th formula as the Nth formula, and execute the step of obtaining the corresponding Nth detection value according to the Nth formula and the proxy model, until the N+Mth detection value corresponding to the N+Mth formula meets the preset convergence condition, and then take the N+Mth formula as the recommended formula, where N and M are both positive integers.
[0030] In a third aspect of the present application, a computer storage medium is provided, wherein the computer storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the above method steps.
[0031] In a fourth aspect of the present application, an electronic device is provided, comprising: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the above-mentioned method steps.
[0032] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:
[0033] This application can significantly reduce the number of trial and error from scratch through multiple historical formulas corresponding to the target product, preprocess the historical formula data, and the obtained sample data set can more accurately reflect the relationship between the ratio of each raw material and the product performance. The Gaussian process regression proxy model constructed using the sample data set provides an efficient prediction model for formula experiments. By predicting the performance of unknown ratio experiments, the number of experiments is reduced by using the proxy model prediction, and resource consumption is also reduced. By obtaining the Nth formula to be tested corresponding to the target product and using the proxy model to determine the test value, this method realizes continuous optimization of the experimental process. The N+1th formula is continuously iterated to obtain, and its test value is evaluated until the test value of the N+Mth formula meets the preset convergence condition. This iterative process ensures that the formula experiment gradually approaches the optimal solution, rather than blindly conducting a large number of invalid experiments. This iterative optimization strategy greatly accelerates the process of determining the best formula and improves the accuracy and efficiency of the entire formula research and development. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is a flow chart of a multi-objective Bayesian optimization recipe recommendation method provided in an embodiment of the present application;
[0035] Figure 2 It is a schematic diagram of a multi-objective Bayesian optimization recipe recommendation system module provided in an embodiment of the present application;
[0036] Figure 3 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application.
[0037] Description of reference numerals: 300, electronic device; 301, processor; 302, communication bus; 303, user interface; 304, network interface; 305, memory. DETAILED DESCRIPTION
[0038] In order to enable technicians in this field to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.
[0039] In the description of the embodiments of the present application, words such as "for example" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "for example" or "for example" is intended to present related concepts in a specific way.
[0040] In the description of the embodiments of the present application, the meaning of the term "multiple" refers to two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "include", "comprise", "have" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.
[0041] The following will provide a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.
[0042] Please refer to Figure 1, a flowchart of a multi-objective Bayesian optimization recipe recommendation method is proposed. The method can be implemented by a computer program, a single-chip microcomputer, or run on a multi-objective Bayesian optimization recipe recommendation system. The computer program can be integrated in an intelligent terminal or run as an independent tool application. Specifically, the method includes steps 10 to 50, and the steps are as follows:
[0043] Step 10: Obtain multiple historical formulas corresponding to the target product.
[0044] The embodiments of the present application can be applied to but not limited to the fields of design and development of formula-based products. Through the proxy model and the continuous iteration process, it is ensured that the formula experiment gradually approaches the optimal solution, rather than blindly conducting a large number of invalid experiments. This iterative optimization strategy greatly accelerates the process of determining the optimal formula and improves the accuracy and efficiency of the entire formula research and development.
[0045] The target product refers to a product that needs to undergo a formula optimization experiment. The product can be a finished product or semi-finished product in the food, chemical, pharmaceutical and other industries. The specific category of the target product needs to be determined according to the actual application scenario. Historical formulas refer to multiple different formulas corresponding to the target product, that is, various different formula records used in the past R&D and production processes. These historical formulas reflect the experimental process of formula optimization of the product and contain a large amount of formula data under different component combinations.
[0046] Specifically, in order to obtain representative historical formula data, it is necessary to perform cluster analysis on all historical formulas of the target product in advance, take the proportion of each material component in the formula as the feature vector, and use clustering algorithms such as K-means to cluster all historical formulas. In this way, multiple formula cluster centers can be obtained, and each cluster center represents the component characteristics of a type of formula. Then, several formulas within each cluster are selected from each cluster as multiple historical formulas. These historical formulas can well represent the data distribution of different formula types, including the ratio information of different components. Among the multiple historical formulas obtained, each historical formula includes specific materials, that is, the names of each component, and the corresponding dosage information of each material component, that is, the weight or volume percentage of the component, which reflects the ratio of different components in the formula, and a set of standard detection indicators corresponding to each formula, which reflect the product quality or performance of the formula. By obtaining multiple historical formula data that can well represent different formula types, it not only contains rich group ratio information, but also provides corresponding product performance data in advance. This lays the foundation for the subsequent use of these data to establish a prediction model, which can more accurately characterize the intrinsic relationship between different group ratios and product performance, and use this prediction to guide subsequent formula optimization experiments.
[0047] For example, the material information in each historical formula may be: i =[M i1 ,M i2 ,…,M im ], where X i The name of the feature vector representing the i-th recipe, M ij Represents the amount of the jth material in the representation vector of the i-th recipe. For example, the first historical recipe contains the following materials: M 11 、M 12 、M 13 、M 14 、M 15 and M 16 , where material M 11 The dosage is 0.3, material M 12 The dosage is 1, material M 11 The dosage is 12.7…Material M 16 The dosage is 0.2. The test index information for each historical formula can be: i =[Q i1 ,Q i2 ,…,Q im ],Y i The name of the feature vector representing the i-th recipe, Q ij The test result of the jth test index in the representation vector of the i-th historical formula. For example, the test indexes in the first historical formula material include: Q 11 , Q 12 , Q 13 , Q 14 , Q 15 and Q 16 The detection index is Q 11 The test result is 52D and the test index is Q 12 The test result is 4 minutes and 30 seconds... The test index is Q 16 The test result is rough.
[0048] Step 20: Perform data preprocessing on the data in each historical recipe to obtain a sample data set.
[0049] Specifically, in order to build a model that can accurately predict the performance of a formula, it is necessary to ensure the integrity, consistency and quality of the historical formula data, and pre-process the data such as the raw material dosage and test indicators collected in each historical formula. The dosage of each raw material in the formula can be uniformly converted to a unit, for example, all converted to weight percentage to ensure that all sample data adopt a consistent representation format. It is also possible to set a reasonable standard range for the dosage of each raw material, that is, the maximum and minimum dosage limits, and remove abnormal formula samples that exceed the limit, which can filter out unreasonable samples with too many or too few raw materials in the formula. Afterwards, the missing values in the formula samples can be filled in by averaging or filling 0 to make the data complete. The data normalization method can also be used to parse all raw material dosages into the 0-1 interval to eliminate the dimension effect. By pre-processing the historical formula data, the quality of the sample data can be effectively improved, laying a data foundation for establishing an accurate formula performance prediction model using machine learning methods.
[0050] Based on the above embodiment, as an optional embodiment, the step of performing data preprocessing on the data in each historical recipe to obtain a sample data set may also include the following steps:
[0051] Step 201: Obtain the standard usage range of each material.
[0052] Step 202: Screen out historical formulas whose material dosage exceeds the standard dosage range to obtain screened historical formulas.
[0053] Specifically, when preprocessing historical recipe data, it is necessary to set the standard range of the dosage of each component raw material. The setting of this standard range is determined based on the professional knowledge and experience of the R&D personnel. The R&D personnel can first refer to the actual dosage range of each material in all the historical recipes obtained to determine the minimum and maximum dosage of the material, or combine the influence of the material dosage on the product quality attributes to make a judgment, such as too much or too little of a certain component may cause abnormal product taste, etc. to set the standard dosage range of each material. After obtaining the standard dosage range, it is necessary to filter the historical recipes according to the standard range, traverse each historical recipe, and check whether the dosage of each material in the historical recipe is within the corresponding pre-set standard dosage range. If all material dosages are within the standard range, the historical recipe is retained; if any raw material dosage is higher than the maximum value or lower than the minimum value, the historical recipe is removed. By filtering and removing those abnormal historical recipes with too much or too little materials, it is ensured that the dosage of each material in the filtered historical recipes is reasonable. These high-quality recipe sample data are more conducive to training a reasonable and accurate proxy model.
[0054] Step 203: Encode each historical material, each material usage, and detection index in the screened historical formula to obtain a data encoding set.
[0055] Step 204: The data with missing values in the data encoding set are supplemented with 0 to obtain a sample data set.
[0056] Specifically, after obtaining the high-quality historical recipe sample data after screening, it is necessary to further convert the data format to facilitate computer processing. First, the name of each material in the formula, the amount of each material and the detection index corresponding to the historical recipe are encoded. For example, each raw material is assigned a unique ID in turn, quantized as a floating point number, and the index is encoded as a category value, so that the text description can be converted into a digital coding representation that can be recognized by a computer, which is convenient for subsequent modeling and calculation. After completing the encoding, first determine whether there are missing values in the data encoding set, that is, whether the encoding expression is complete, whether the detection index appears in full, etc., directly using the sample data with missing values will cause the model training process to have calculation difficulties, if it is detected that there are missing values in the data encoding set, such as the lack of a certain encoding material, or the lack of the detection value of a certain indicator, it is necessary to process. In the embodiment of the present application, a strategy of using 0 value filling is adopted, and 0 is supplemented in the missing position to obtain a complete sample data set.
[0057] Step 30: Construct a surrogate model based on the sample dataset and Gaussian process regression.
[0058] Specifically, for the problem of predicting continuous values of performance indicators such as recipe optimization, this application chooses to use Gaussian process regression method to establish a proxy model. Gaussian process regression is essentially a kernel method that can model nonlinear associations and has excellent fitting and generalization capabilities for continuous numerical predictions. According to the sample data set obtained after preprocessing, the encoding conversion and missing values are filled, and it can be randomly divided into a training set and a test set. Then, based on the kernel function in Gaussian process regression, such as the commonly used RBF kernel, the relevant hyperparameters are set, and the model is trained using the training set data. By maximizing the marginal likelihood function of the training data, a set of optimal hyperparameter values can be obtained, which will determine the shape of the kernel function and the prior distribution of the Gaussian process. Then, based on Bayesian estimation, the posterior distribution can be obtained as a trained Gaussian process regression model, and then the model is verified using the test set. By adjusting the hyperparameters such as the length scale and variance parameter in the kernel function, the prediction performance of the model is continuously optimized. The optimized Gaussian process regression model finally obtained can be used as a proxy model for predicting the performance of the new recipe. The proxy model establishes a probabilistic relationship between components and properties through a nonlinear Gaussian process, which can quickly predict the performance indicators of new formulas, avoid a large number of experiments, and thus guide the iterative process of formula optimization.
[0059] Based on the above embodiment, as an optional embodiment, the step of building a proxy model according to the sample data set and Gaussian process regression may further include the following steps:
[0060] Step 301: Divide the sample data set into a training set and a test set.
[0061] Specifically, the preprocessed sample data set can be randomly divided into two parts according to a certain ratio (for example, 7:3), with 70% of the data used as the training set and 30% of the data used as the test set. The purpose of dividing the training set and the test set is to train and evaluate the model. The training set is used for model training, that is, to adjust the parameters of the model so that it can obtain the best prediction effect on the training data. The test set is used to evaluate the prediction performance of the model. You can check the prediction accuracy of the model for unseen data, thereby evaluating the generalization ability of the model.
[0062] Step 302: Obtain the kernel function and hyperparameters in the pre-constructed Gaussian process regression, perform training based on the kernel function and the training set, and adjust the hyperparameters during the training process to obtain an initial proxy model.
[0063] Specifically, as a non-parametric model, Gaussian process regression has a kernel function and related hyperparameters that directly affect the prediction performance of the model. Therefore, it is necessary to pre-build a set of kernel functions and hyperparameters with good experience. Here, the RBF kernel commonly used in deep learning can be used. The kernel function contains kernel width parameters and variance parameters. Both of these are hyperparameters that need to be adjusted. After setting the kernel function and initial hyperparameters, the model is trained using this pre-built Gaussian process regression structure and the training set data. The training process needs to optimize the marginal likelihood function to determine the parameters of the posterior distribution. Subsequently, in order to further improve the prediction accuracy of the model, the hyperparameters such as the length scale and variance in the kernel function can be continuously adjusted during the training process through methods such as cross-validation, the posterior distribution can be re-established, and the training is repeated until the optimal value of the hyperparameter is found. Finally, through this process of pre-building kernel functions and hyperparameter retraining and parameter adjustment, a Gaussian process regression initial proxy model with hyperparameter tuning on a given training set can be obtained. This method of preset structure retraining not only improves the initial quality of the model, but also further improves the accuracy of the model through hyperparameter adjustment. The initial proxy model can be used for subsequent model evaluation and retraining.
[0064] For example, assuming that D = {(x i ,y i )|x i ∈R p ,y i ∈R,i∈1,2,…,n}, is a set of n historical recipes, where the vector x i Indicates that the input variable (material information) corresponds to the observed value yi (Detection value). The data is represented as D = (X, y), and the matrix X = (x1, x2, ..., x n ) T ∈R n×p and vector y=(y1,y2,…,y n ) T ∈R n Assume that y i Including mean zero, variance σ 2 Normally distributed measurement errors. Mathematically, y i =f(x i )+ε and ε~N(0,σ 2 ), where f(·) is the unknown function.
[0065] A Gaussian process can be thought of as a collection of normally distributed random variables that simulate the behavior of f(·)+ε: f~GP(m(x|θ),k(x,x'|θ)),x,x'∈R p ;
[0066] Among them, the Gaussian process regression model consists of two parts: the first part is the mean function m(x|θ); the mean function describes the average level of the process, usually 0 or some simple parameterized function. Although the mean function is usually set to zero in computer science applications (such as pattern recognition), the mean function is a natural way to incorporate physical relationships into the Gaussian process model. The second part is the covariance or kernel function k(x,x'|θ), which describes the correlation of the process under different inputs x and x'. θ is a hyperparameter that is used to define the coefficient of the mean and kernel functions. Commonly used kernel functions and hyperfunctions can be: Matern kernel function, square exponential kernel function, and rational quadratic kernel function. Training hyperparameters can provide insight into the most important dimensions of x, which is called automatic correlation determination. Or l can be regarded as a scalar, that is, the same for each dimension of x, to reduce the number of hyperparameters. Similarly (observation error variance) can be inferred from the data as a hyperparameter, or it can be fixed a priori by the modeler using experimental knowledge.
[0067] For the kernel function: θ SE =(l), where x and x' are two input vectors, and the kernel function is a function used to calculate the similarity or correlation between two points in the input space. In the Gaussian process, this correlation directly affects the degree to which the function value of one point affects the prediction of the function value of another point. The kernel function measures the similarity between two points. When the two points x and x' are close, K(x, x') is close to 1, indicating that they are very similar; when the two points are far away, K(x, x') is close to 0, indicating that they are almost dissimilar.
[0068] Given training data and values for hyperparameters and optionally assumed observation errors, the key mathematical property of a Gaussian process is that the outputs y and f follow a multivariate normal (Gaussian) distribution: Here K(X,X / θ) represents the kernel matrix, which is the kernel function k calculated element by element. For a given set of training data points X and corresponding observations Y, the goal of the Gaussian process is to predict a new point X based on these data points. * The function value f * .
[0069] Exploiting the analytical properties of the multivariate normal distribution gives the following expected value (mean) and (co)variance of the predictions f corresponding to the new input matrix X:
[0070] μ * (X * |θ)=E(f * )=m(X * |θ)+K(X * ,X|θ)[K(X,X|θ)+σ 2 I] -1 (ym(X|θ)
[0071] ∑ * (X * |θ)=V(f * )=K(X * ,X * |θ)-K(X * ,X|θ)[K(X,X|θ)+σ 2 I] -1 K(X,X * |θ);
[0072] That is, for X * is an m×d matrix, where m is the number of prediction points and d is the dimension of each input point. In the Gaussian process, we want to predict the function value f at these points * The predicted mean vector μ * and the covariance matrix Σ * It can be calculated by the above formula. For the matrix input X, the posterior prediction mean is represented as a vector and the covariance is represented as a matrix; similarly, for the vector input x, the output is represented as a scalar μ * (x * )and The construction of the Gaussian process regression model quantifies the prediction uncertainty by calculating the variance. In addition, Gaussian process regression is a non-parametric regression model, which means that the training data is directly embedded into the model. Therefore, the Gaussian process regression model interpolates between the training data and the kernel function, describing the prediction uncertainty as a function of the distance from the training data (for a steady-state model).
[0073] Step 303: Perform test optimization on the initial proxy model based on the test set to obtain an optimized proxy model.
[0074] Specifically, although the initial proxy model obtained through training has a good effect on the training set, it is still necessary to use an independent test set to further test, evaluate and optimize the model. The materials and material usage in the divided test set samples can be used as the input of the model, and the detection values corresponding to the detection indicators can be used as the output of the model. Then, the error between the predicted output and the actual indicator in the test set is calculated. It can be evaluated with indicators such as mean square error. If the error is large, it means that the generalization ability of the initial proxy model is weak. At this time, it is necessary to continue to optimize. The kernel function hyperparameters of the initial proxy model, such as kernel width parameter, variance parameter, etc., can be adjusted to retrain the model and repeat this process. Each time, the error changes before and after training are compared until the error is reduced to an acceptable level to obtain a proxy model with excellent performance.
[0075] Step 40: Obtain the Nth formula to be tested corresponding to the target product, and obtain the corresponding Nth detection value according to the Nth formula and the proxy model.
[0076] Specifically, after the proxy model is constructed and optimized, it can be used to guide the experimental iteration process of subsequent formulas. You can first obtain the Nth formula to be experimented corresponding to the target product, where N is a positive integer, representing the Nth round of iterative experiments. The Nth formula can be designed by R&D personnel based on past experience, and the Nth formula includes various materials and the corresponding material dosage of each material. Then, based on the optimized proxy model, using the Nth formula as the model input, the detection index corresponding to the Nth formula can be predicted, that is, the predicted detection value of the formula can be obtained.
[0077] The reason why we need to use the proxy model for prediction first is that the actual formula experiment verification is very time-consuming and resource-consuming. If we directly conduct random experiments, it is very inefficient. The proxy model can be used to quickly predict the detection indicators of the new formula and guide the direction of subsequent experiments. R&D personnel can judge the rationality of the Nth formula based on the predicted value and decide whether it is necessary to actually conduct complex experiments. This prediction-first method can improve the efficiency of research and development. The real detection value of the actual experimental verification of the Nth formula based on the predicted value is obtained as the measured detection value of the Nth formula, that is, the Nth detection value. The obtained Nth formula and the real detection value, that is, the Nth detection value, are saved in the database. The database contains multiple historical formula data. The measured data is saved to train and optimize the proxy model and provide a reference for subsequent formula iterations.
[0078] Based on the above embodiment, as an optional embodiment, a process of calculating the accuracy of the proxy model is also included, and the specific process is as follows:
[0079] According to the predicted detection value of the proxy model for the Nth recipe and the true detection value obtained from the actual experiment of the Nth recipe, the prediction accuracy of the model for this recipe can be calculated. The calculation method can be calculated by using methods such as the error ratio between the predicted value and the true value.
[0080] The calculation formula of the embodiment of the present application is: e=|Y1-Y2| / |Y max -Y min |, Y1 is the predicted value, Y2 is the Nth detection value, that is, the true detection value, Y max is the maximum detection value of the detection index Y, Y min is the minimum detection value of the detection index Y.
[0081] The calculated prediction accuracy is compared with the set accuracy threshold, where the accuracy threshold can be set according to actual needs, for example, 80%. If the prediction accuracy of the proxy model is lower than the set accuracy threshold, it is considered that the prediction effect of the model for the Nth recipe is abnormal, and a warning message will be prompted, indicating that the reliability of the model prediction result is not high. The relationship between the accuracy rate and the accuracy threshold is judged because the accuracy rate can evaluate the generalization prediction ability of the proxy model. If the accuracy rate is too low, it means that the model's fitting effect is poor and the prediction of new data is unreliable. The warning can indicate that the proxy model needs to be retrained.
[0082] Based on the above embodiment, as an optional embodiment, after the Nth detection value is determined according to the Nth recipe and the agent model, the following process is also included:
[0083] Specifically, whenever a new experimental formula, such as the Nth formula, and the test value corresponding to the formula, such as the Nth test value, are obtained, the proxy model can be iteratively updated and optimized based on the Nth test value and the Nth formula to continuously improve its performance. After the prediction and real experiment of the Nth formula, the Nth formula and the Nth test value can be understood as historical formulas. The process of training the proxy model based on the Nth test value and the Nth formula is similar to the process of training the proxy model based on historical formulas. Continue to train the proxy model to obtain an updated model with improved performance. After the proxy model is updated, use the updated proxy model to predict the next new formula, such as the N+1th new formula as the Nth formula, repeat the previous prediction process, and continue to perform experimental verification to obtain the N+1th test value of the N+1th formula, ensuring that the proxy model is optimized with new experimental data each time, and predicting new formulas based on the updated model.
[0084] Step 50: Obtain the N+1th recipe, take the N+1th recipe as the Nth recipe, and execute the step of obtaining the corresponding Nth detection value according to the Nth recipe and the proxy model, until the N+Mth detection value corresponding to the N+Mth recipe meets the preset convergence condition, and take the N+Mth recipe as the recommended recipe, where N and M are both positive integers.
[0085] Specifically, in the process of formula optimization design, in order to obtain the best formula combination, it is necessary to set the termination condition of iterative optimization, that is, when the detection index of the formula meets the preset convergence condition, the training iteration process is carried out and the current formula is used as the recommended formula. The next formula of the Nth formula, that is, the N+1th formula, is obtained, and the N+1th formula is used as the Nth formula, and the step of obtaining the corresponding Nth detection value according to the Nth formula and the proxy model is performed. At this time, the proxy model is an updated proxy model. Then the N+1th detection value is obtained, and it is determined whether the N+1th detection value meets the preset convergence condition. If it does not meet the preset convergence condition, the next formula is obtained until the N+Mth detection value corresponding to the N+Mth formula meets the preset convergence condition, and the N+Mth formula is used as the recommended formula, where N and M are both positive integers. In the embodiment of the present application, the preset convergence condition can be the change rate of the detection value of the previous formula and the detection value of the next formula, and the convergence condition is when the change rate is less than 1%.
[0086] Based on the above embodiment, as an optional embodiment, the step of obtaining the N+1th formula may further include the following steps:
[0087] Step 501: Determine whether the Nth detection value meets a preset convergence condition.
[0088] Step 502: If the Nth detection value meets the preset convergence condition, the Nth recipe corresponding to the Nth detection value is used as the recommended recipe.
[0089] Specifically, during the iterative optimization process, a convergence condition needs to be set to determine when to end the iteration and output the final result. For each round of new recipe optimization, after obtaining the predicted test value of the new recipe, it is necessary to make a judgment and compare the predicted test value of the new recipe with the preset convergence condition to determine whether the test value meets the convergence condition. For example, before obtaining the N+1th recipe, it is necessary to determine whether the Nth test value meets the preset convergence condition. If the Nth test value meets the preset convergence condition, the Nth recipe corresponding to the Nth test value is used as the recommended recipe, that is, the step of obtaining the N+1th recipe is not continued.
[0090] Step 503: If the Nth detection value does not meet the preset convergence condition, multiple target requirements of the user for the target product are obtained, and the target function is determined according to the target requirements.
[0091] Specifically, when it is judged that the predicted test value of the new formula does not meet the preset convergence conditions, it means that the performance indicators of the current formula are not optimized and stable enough, and the next round of iterative optimization is required. At this time, it is necessary to obtain the user's specific needs for multiple optimization goals of the target product, such as minimizing costs, maximizing quality indicators, improving stability, and other aspects. After obtaining the user's multi-dimensional optimization goals, it is necessary to integrate and convert them into a computable objective function, weight each detection indicator to indicate its priority, and set the weight of each detection indicator. The weight of important detection indicators is set to be larger, and the weight of unimportant detection indicators is set to be smaller. The objective function comprehensively reflects the needs of the user.
[0092] For example, for the objective function Z, the weight values of each detection index can be: i =1 / m, where q i is the weight of the i-th detection index in a certain recipe, and m is the number of detection indexes in the recipe. For different detection indexes in different recipes, in the embodiment of the present application, the user can set different orientations for different detection indexes. For example, the favorable detection index can be set as a positive index. Then the positive index value: Z1=(y i -y min ) / (y max -y min ); Set the unfavorable detection index as a negative index, then the negative index value: Z2 = (y max -y i ) / (y max -y min ); Set the neutral detection index as the intermediate index, then the intermediate index value: Z3 = |y i -y min | / (y max -y min) , where y max and min are the maximum and minimum detection values in the formula, respectively, i is the detection value of the current detection index. In the embodiment of the present application, there are 16 detection indexes for each formula. Assuming that the first detection index is a positive index, the second detection index is a negative detection index, the third detection index is an intermediate detection index, the fourth detection index is a negative detection index... the 16th detection index is a positive detection index, then the objective function is: Z = Z1*q i +Z2*q2+Z3*q3+Z2*q4…+Z1q 16 .
[0093] Among them, the preset convergence condition can be: |Z N -Z N+1 | / |Z N+1|<epsilon, epsilon is 0.01, where Z N represents the objective function value of the Nth recipe, Z N+1 It represents the objective function value of the N+1th recipe. The objective function value Z is calculated according to the above calculation method. For example, when the objective function value Z corresponding to the 15th recipe is 15 The objective function value Z corresponding to the 16th recipe 16 When the rate of change is less than 0.01, the 16th recipe is taken as the recommended recipe.
[0094] Step 504: Within the standard dosage range of each material dosage, the objective function is optimized according to the acquisition function to obtain the N+1th formula.
[0095] Specifically, when it is judged that the current formula has not reached the convergence condition, it is necessary to design a new round of formulas for further optimization. In order to ensure the feasibility of the formula, it is necessary to consider the dosage range constraints of each component material. That is to say, the dosage of each component of the new formula must be within the standard dosage range of the component, meeting the minimum dosage requirements and not exceeding the maximum dosage limit. This is a prerequisite for ensuring the rationality of the formula. On the premise of meeting the dosage range constraints, an acquisition function can be constructed to sample and optimize the current objective function. The objective function comprehensively reflects the user's multiple optimization goals, and the acquisition function can efficiently search the entire domain to find the optimal solution for the optimization objective function. After sampling and iteration of the acquisition function, a material dosage combination can be obtained so that the objective function reaches the optimal value under the current conditions, and this combination is used as the N+1th formula of the new round.
[0096] See also Figure 2 , is a schematic diagram of a multi-objective Bayesian optimization recipe recommendation system module provided in an embodiment of the present application. The multi-objective Bayesian optimization recipe recommendation system may include: a historical recipe acquisition module, a data preprocessing module, an agent module construction module, a detection value determination module and a loop iteration module, wherein:
[0097] A historical formula acquisition module is used to acquire multiple historical formulas corresponding to a target product;
[0098] A data preprocessing module, used for preprocessing the data in each of the historical recipes to obtain a sample data set;
[0099] An agent module construction module, used to construct an agent model based on the sample data set and Gaussian process regression;
[0100] A detection value determination module, used for obtaining an Nth formula to be tested corresponding to the target product, and determining an Nth detection value according to the Nth formula and the proxy model;
[0101] A loop iteration module is provided to obtain the N+1th formula, take the N+1th formula as the Nth formula, and execute the step of obtaining the corresponding Nth detection value according to the Nth formula and the proxy model, until the N+Mth detection value corresponding to the N+Mth formula meets the preset convergence condition, and then take the N+Mth formula as the recommended formula, where N and M are both positive integers.
[0102] Optionally, the data preprocessing module is also used to obtain a standard usage range for each material usage; filter out historical formulas whose material usage exceeds the standard usage range to obtain filtered historical formulas; encode each historical material, each material usage and detection index in the filtered historical formulas to obtain a data coding set; and fill in missing values in the data coding set with 0 to obtain a sample data set.
[0103] Optionally, the proxy module construction module is further used to divide the sample data set into a training set and a test set; obtain a kernel function and hyperparameters in a pre-constructed Gaussian process regression, perform training based on the kernel function and the training set, and adjust the hyperparameters during the training process to obtain an initial proxy model;
[0104] The initial proxy model is tested and optimized based on the test set to obtain the optimized proxy model.
[0105] Optionally, the detection value determination module is also used to predict the Nth formula based on the proxy model to obtain a predicted value; obtain a real detection value of an experiment on the Nth formula based on the predicted value, use the real detection value as the Nth detection value, and save the Nth formula and the Nth detection value to a database.
[0106] Optionally, a multi-objective Bayesian optimization recipe recommendation system also includes an accuracy calculation module, which is used to calculate the accuracy of the proxy model's prediction of the Nth recipe based on the predicted value and the true detection value; when the accuracy is lower than the accuracy threshold, a warning message is displayed to indicate that the proxy model prediction is abnormal.
[0107] Optionally, a multi-objective Bayesian optimization recipe recommendation system also includes a model updating module, which is used to iteratively update the proxy model based on the Nth detection value and the Nth recipe to obtain an updated proxy model; taking the N+1th recipe as the Nth recipe, and executing the step of obtaining the corresponding Nth detection value according to the Nth recipe and the proxy model, until the N+Mth detection value corresponding to the N+Mth recipe meets the preset convergence condition, and then taking the N+Mth recipe as the recommended recipe, includes: taking the N+1th recipe as the Nth recipe, and executing the step of obtaining the corresponding Nth detection value according to the Nth recipe and the updated proxy model, until the N+Mth detection value corresponding to the N+Mth recipe meets the preset convergence condition, and then taking the N+Mth recipe as the recommended recipe.
[0108] Optionally, the loop iteration module is also used to determine whether the Nth detection value meets the preset convergence conditions; if the Nth detection value meets the preset convergence conditions, the Nth formula corresponding to the Nth detection value is used as the recommended formula; if the Nth detection value does not meet the preset convergence conditions, multiple target requirements of the user for the target product are obtained, and the objective function is determined according to the target requirements; within the standard usage range of each material usage, the objective function is optimized according to the acquisition function to obtain the N+1th formula.
[0109] It should be noted that: when the system provided in the above embodiment realizes its functions, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiment belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.
[0110] An embodiment of the present application also provides a computer storage medium, which can store multiple instructions. The instructions are suitable for being loaded by a processor and executing a multi-objective Bayesian optimization recipe recommendation method of the above embodiment. The specific execution process can be found in the specific description of the above embodiment, which will not be repeated here.
[0111] Please refer to Figure 3 The application also discloses an electronic device. Figure 3 The electronic device 300 may include: at least one processor 301 , at least one network interface 304 , a user interface 303 , a memory 305 , and at least one communication bus 302 .
[0112] The communication bus 302 is used to realize the connection and communication between these components.
[0113] The user interface 303 may include a display screen (Display) and a camera (Camera), and the optional user interface 303 may also include a standard wired interface and a wireless interface.
[0114] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).
[0115] Among them, the processor 301 may include one or more processing cores. The processor 301 uses various interfaces and lines to connect various parts in the entire server, and executes various functions of the server and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 305, and calling data stored in the memory 305. Optionally, the processor 301 can be implemented in at least one hardware form of digital signal processing (Digital Signal Processing, DSP), field programmable gate array (Field-Programmable Gate Array, FPGA), and programmable logic array (Programmable Logic Array, PLA). The processor 301 can integrate one or a combination of a central processing unit (Central Processing Unit, CPU), a graphics processing unit (Graphics Processing Unit, GPU) and a modem. Among them, the CPU mainly processes the operating system, user interface and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 301, and it can be implemented separately through a chip.
[0116] Among them, the memory 305 may include a random access memory (Random Access Memory, RAM) and may also include a read-only memory (Read-Only Memory). Optionally, the memory 305 includes a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 305 may optionally also be at least one storage device located away from the aforementioned processor 301. Refer to Figure 3 , the memory 305 as a computer storage medium may include an operating system, a network communication module, a user interface module and an application program of a multi-objective Bayesian optimization recipe recommendation method.
[0117] exist Figure 3 In the electronic device 300 shown, the user interface 303 is mainly used to provide an input interface for the user and obtain the data input by the user; and the processor 301 can be used to call the application program storing a multi-objective Bayesian optimization recipe recommendation method in the memory 305, and when executed by one or more processors 301, the electronic device 300 executes one or more of the methods described in the above embodiments. It should be noted that for the aforementioned method embodiments, for the sake of simple description, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for the present application.
[0118] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0119] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are only schematic, such as the division of units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interfaces, and the indirect coupling or communication connection of devices or units can be electrical or other forms.
[0120] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0121] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0122] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a memory and includes several instructions for a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the various embodiments of the present application. The aforementioned memory includes: various media that can store program codes, such as USB flash drives, mobile hard drives, magnetic disks or optical disks.
[0123] The above is only an exemplary embodiment of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.
[0124] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.
Claims
1. A multi-objective Bayesian optimization recipe recommendation method, characterized in that: The method comprises: Obtain multiple historical formulas corresponding to the target product; Performing data preprocessing on the data in each of the historical recipes to obtain a sample data set; Constructing a proxy model based on the sample data set and Gaussian process regression; Obtaining an Nth formula to be tested corresponding to the target product, and obtaining a corresponding Nth detection value according to the Nth formula and the proxy model; Determining whether the Nth detection value meets a preset convergence condition; If the Nth detection value meets the preset convergence condition, the Nth formula corresponding to the Nth detection value is used as the recommended formula; If the Nth detection value does not meet the preset convergence condition, multiple target requirements of the user for the target product are obtained, and the objective function is determined according to the target requirements. The objective function is: Z = Z1*q1+Z2*q2+Z3*q3+ Z2*q4…+ Z1qi, where Z1 is the positive indicator value, Z2 is the negative indicator value, and Z3 is the intermediate indicator value; Z1=(y i -y min ) / (y max -y min ); Z2=(y max -y i ) / (y max -y min ), Z3=|y i -y min | / (y max -y min) ,y max and min are the maximum and minimum detection values in the formula, respectively, i is the detection value of the current detection index; q i =1 / m, where q i is the weight of the i-th detection index in the recipe, and m is the number of detection indexes in the recipe; Within the standard dosage range of each material dosage, the objective function is optimized according to the acquisition function to obtain the N+1th formula; wherein the preset convergence condition is: |Z N -Z N+1 | / |Z N+1 |<epsilon, epsilon is 0.01, where Z N represents the objective function value of the Nth recipe, Z N+1 Represents the objective function value of the N+1th recipe; The N+1th recipe is used as the Nth recipe, and the step of obtaining the corresponding Nth detection value according to the Nth recipe and the proxy model is performed, until the N+Mth detection value corresponding to the N+Mth recipe meets the preset convergence condition, the N+Mth recipe is used as the recommended recipe, and N and M are both positive integers.
2. The multi-objective Bayesian optimization recipe recommendation method according to claim 1, characterized in that: The historical formula includes each historical material, the material usage corresponding to each historical material, and the detection index. The data in each historical formula is preprocessed to obtain a sample data set, including: Obtaining the standard dosage range of each material dosage; Screen out historical formulas where the material dosage exceeds the standard dosage range to obtain screened historical formulas; Encoding each historical material, each material usage, and a detection index in the screened historical formula to obtain a data encoding set; The data with missing values in the data encoding set are supplemented with 0 to obtain a sample data set.
3. The multi-objective Bayesian optimization recipe recommendation method according to claim 1, characterized in that: The constructing of the proxy model according to the sample data set and Gaussian process regression includes: Dividing the sample data set into a training set and a test set; Obtaining a kernel function and hyperparameters in a pre-constructed Gaussian process regression, performing training based on the kernel function and the training set, and adjusting the hyperparameters during the training process to obtain an initial proxy model; The initial proxy model is tested and optimized based on the test set to obtain the optimized proxy model.
4. The multi-objective Bayesian optimization recipe recommendation method according to claim 1, characterized in that: The obtaining a corresponding Nth detection value according to the Nth recipe and the proxy model includes: Predicting the Nth recipe based on the proxy model to obtain a predicted value; A true detection value of an experiment on the Nth recipe based on the predicted value is obtained, the true detection value is used as the Nth detection value, and the Nth recipe and the Nth detection value are saved in a database.
5. The multi-objective Bayesian optimization recipe recommendation method according to claim 4, characterized in that: The method further comprises: Calculating the accuracy of the proxy model's prediction of the Nth recipe based on the predicted value and the true detection value; When the accuracy rate is lower than the accuracy rate threshold, a warning message is displayed to prompt that the proxy model prediction is abnormal.
6. The multi-objective Bayesian optimization recipe recommendation method according to claim 4, characterized in that: After determining the Nth detection value according to the Nth recipe and the proxy model, the method further includes: Iteratively updating the proxy model based on the Nth detection value and the Nth recipe to obtain an updated proxy model; The step of taking the N+1th recipe as the Nth recipe and performing the step of obtaining the corresponding Nth detection value according to the Nth recipe and the proxy model until the N+Mth detection value corresponding to the N+Mth recipe meets the preset convergence condition, and taking the N+Mth recipe as the recommended recipe includes: The N+1th recipe is used as the Nth recipe, and the step of obtaining the corresponding Nth detection value according to the Nth recipe and the updated proxy model is performed, until the N+Mth detection value corresponding to the N+Mth recipe meets the preset convergence condition, and the N+Mth recipe is used as the recommended recipe.
7. A multi-objective Bayesian optimization recipe recommendation system, characterized in that: The system comprises: A historical formula acquisition module is used to acquire multiple historical formulas corresponding to a target product; A data preprocessing module, used for preprocessing the data in each of the historical recipes to obtain a sample data set; An agent module construction module, used to construct an agent model based on the sample data set and Gaussian process regression; A detection value determination module, used for obtaining an Nth formula to be tested corresponding to the target product, and determining an Nth detection value according to the Nth formula and the proxy model; The loop iteration module is used to determine whether the Nth detection value meets the preset convergence condition; if the Nth detection value meets the preset convergence condition, the Nth formula corresponding to the Nth detection value is used as the recommended formula; if the Nth detection value does not meet the preset convergence condition, multiple target requirements of the user for the target product are obtained, and the objective function is determined according to the target requirements. The objective function is: Z= Z1*q1+Z2*q2+Z3*q3+ Z2*q4…+ Z1qi, where Z1 is the positive indicator value, Z2 is the negative indicator value, and Z3 is the intermediate indicator value; Z1=(y i -y min ) / (y max -y min ); Z2=(y max -y i ) / (y max -y min ), Z3=|y i -y min | / (y max -y min) ,y max and min are the maximum and minimum detection values in the formula, respectively, i is the detection value of the current detection index; q i =1 / m, where q i is the weight of the i-th detection index in the formula, and m is the number of detection indexes in the formula; within the standard dosage range of each material dosage, the objective function is optimized according to the acquisition function to obtain the N+1-th formula; wherein the preset convergence condition is: |Z N -Z N+1 | / |Z N+1 |<epsilon, epsilon is 0.01, where Z N represents the objective function value of the Nth recipe, Z N+1 Represents the objective function value of the N+1th recipe; taking the N+1th recipe as the Nth recipe, and executing the step of obtaining the corresponding Nth detection value according to the Nth recipe and the proxy model, until the N+Mth detection value corresponding to the N+Mth recipe meets the preset convergence condition, taking the N+Mth recipe as the recommended recipe, where N and M are both positive integers.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the method according to any one of claims 1 to 6.
9. An electronic device, characterized in that: It includes a processor, a memory, a user interface and a network interface, the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the method as described in any one of claims 1-6.
Citation Information
Patent Citations
Automatic synthesis test system and method for realizing rapid optimization of formula
CN116046764A
Chemical experiment multi-index optimization method and system, storage medium and electronic equipment
CN116959629A
Loess landslide slip prediction method and system based on meta-learning and Bayesian optimization
CN117349751A
Model training method, device and equipment for predicting charge quality and medium
CN117575412A
Liquid crystal material formula analysis method and device based on data mining
CN117649898A
