Multi-target task optimization method and device based on user preference

Abstract

The invention provides a multi-objective task optimization method and device based on user preferences, and the method comprises the steps: S1, constructing a multi-objective optimization task according to a search space of an input variable and a plurality of objective functions; s2, training by using the initial training data set to obtain a plurality of Gaussian process models; s3, according to the plurality of Gaussian process models, obtaining prediction performance of all points in the search space under the corresponding objective function; s4, constructing a weighted acquisition function according to user requirements, and selecting a current target input point from the search space by using the weighted acquisition function; s5, executing an experiment according to the current target input point to obtain an actual response of the current target input point; and S6, updating the initial training data set according to the current target input point and the response value thereof, and repeating the steps S2-S6 until a preset end condition is met. According to the method, the scheme meeting the multi-target requirement and meeting the user preference can be quickly found under the limited experimental budget, and the expansibility is good.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for optimizing multi-objective tasks based on user preferences. Background Technology

[0002] In the field of experimental design, especially in complex tasks involving the optimization of multiple performance indicators (such as strength, stability, yield, etc.), traditional experimental design methods often lack automation capabilities, struggle to balance conflicts between objectives, and are inefficient.

[0003] In recent years, Bayesian optimization, as an efficient strategy for handling high-cost, black-box function optimization problems, has been widely applied in experimental design fields such as materials design, process optimization, and formulation screening. Its core advantage lies in guiding experimental resources to focus on the region most likely to achieve improvement by constructing a surrogate model and acquisition function.

[0004] However, standard multi-objective Bayesian optimization methods assume that all objectives have equal weights, which fails to reflect the user's preference for different objectives. This causes the optimization process to deviate from the actual region of interest, making it difficult to meet the complex optimization needs of real-world experiments where users have clear subjective preferences and objective weights are unequal, resulting in poor scalability.

[0005] To improve scalability, some studies have attempted to incorporate user preference information into the optimization process to assist in the design of data collection strategies. However, such methods rely on users to make pairwise comparisons or explicitly score experimental schemes, increasing the interactive burden, resulting in high manual costs, low efficiency, and poor stability, making them difficult to apply to automated or high-throughput experimental scenarios. Furthermore, due to the influence of user subjectivity, they suffer from poor transferability and are difficult to reproduce.

[0006] In summary, existing technologies suffer from poor scalability. Summary of the Invention

[0007] This invention provides a multi-objective task optimization method and apparatus based on user preferences, which solves the problem of poor scalability in the prior art and achieves multi-objective task optimization based on user preferences with good scalability.

[0008] This invention provides a multi-objective task optimization method based on user preferences, comprising the following steps: S1: Construct a multi-objective optimization task based on the predefined search space of input variables and multiple pre-built objective functions; S2: Using a pre-constructed initial training dataset, the plurality of Gaussian process models are trained based on the Bayesian machine learning algorithm; wherein, the plurality of Gaussian process models correspond one-to-one with the plurality of objective functions; S3: Based on the multiple Gaussian process models, obtain the predicted performance of all points in the search space under the corresponding objective function; S4: Construct a weighted acquisition function according to user requirements, and use the weighted acquisition function to select the current target input point from the search space; S5: Perform the experiment based on the current target input point to obtain the actual response of the current target input point; S6: Update the initial training dataset based on the current target input point and its response value, and repeat steps S2-S6 until the preset termination condition is met, thus completing the scheme design for the multi-objective optimization task.

[0009] According to the multi-objective task optimization method based on user preferences provided by the present invention, multiple Gaussian process models are trained using a pre-constructed initial training dataset and a Bayesian machine learning algorithm, including: S21: Obtain the initial training dataset; S22: Normalize the response values ​​in the initial training dataset, and update the initial training dataset according to the processing result; S23: Select one of the multiple objective functions as the current objective, and train the Gaussian process model corresponding to the current objective using the initial training dataset based on the Bayesian machine learning algorithm; S24: Repeat step S23 until all objective functions are traversed, resulting in multiple Gaussian process models.

[0010] According to the multi-objective task optimization method based on user preferences provided by the present invention, obtaining an initial training dataset includes: Initial sample points are sampled from the search space, and experiments are performed based on the initial sample points to obtain the response values ​​of the initial sample points; Based on the initial sample points and their response values, an initial training dataset is constructed.

[0011] According to the multi-objective task optimization method based on user preferences provided by the present invention, obtaining an initial training dataset includes: Obtain historical experimental data and construct an initial training dataset based on the historical experimental data.

[0012] According to a multi-objective task optimization method based on user preferences provided by the present invention, a weighted acquisition function is constructed according to user needs, and the current target input point is selected from the search space using the weighted acquisition function, including: The weighted acquisition function is constructed based on the pre-set weights of the users for the corresponding targets of the multiple objective functions; Candidate input points are selected from the search space, and the weighted expected hypervolume increment value of the candidate input points is calculated according to the weighted acquisition function. The candidate input point with the largest weighted expected supervolume increment is selected as the current target input point.

[0013] According to a multi-objective task optimization method based on user preferences provided by the present invention, a weighted acquisition function is constructed according to pre-set user weights for the corresponding objectives of the multiple objective functions, including: The weights of the targets corresponding to the multiple objective functions are set according to the user's level of attention to them; The weighted acquisition function is constructed based on the weights of the targets corresponding to the multiple objective functions.

[0014] The present invention also provides a multi-objective task optimization device based on user preferences, comprising the following modules: The task module is used to construct multi-objective optimization tasks based on the predefined search space of input variables and multiple pre-built objective functions; The modeling module is used to train the multiple Gaussian process models using a pre-constructed initial training dataset and based on a Bayesian machine learning algorithm; wherein the multiple Gaussian process models correspond one-to-one with the multiple objective functions; The prediction module is used to obtain the predicted performance of all points in the search space under the corresponding objective function based on the multiple Gaussian process models. The input module is used to construct a weighted acquisition function according to user requirements, and to select the current target input point from the search space using the weighted acquisition function; The experiment module is used to perform experiments based on the current target input point and obtain the actual response of the current target input point; The iterative module is used to update the initial training dataset based on the current target input point and its response value, repeating steps S2-S6 until the preset termination condition is met, thus completing the scheme design for the multi-objective optimization task.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the user preference-based multi-objective task optimization method as described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the user preference-based multi-objective task optimization method as described above.

[0017] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the user preference-based multi-objective task optimization method as described above.

[0018] The present invention provides a multi-objective task optimization method and apparatus based on user preferences. It constructs a multi-objective optimization task using a search space and multiple pre-built objective functions. Then, it models the initial training dataset to predict the performance of each point in the search space, thus obtaining the target performance under untested conditions. Next, it constructs a multi-objective acquisition function based on explicit weighted preferences according to user needs, allowing the user to select the target input point that best matches their interests. Finally, it implements the target input points on an experimental platform, feeding back new data to update the model and proceeding to the next iteration. In this way, the present invention can quickly find a solution that meets multi-objective requirements and aligns with user preferences within a limited experimental budget, exhibiting good scalability. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the multi-objective task optimization method based on user preferences provided by the present invention; Figure 2 This is an iterative round-yield diagram of an embodiment of the multi-objective task optimization method based on user preferences provided by the present invention; Figure 3 This is an iterative round-cost diagram of an embodiment of the multi-objective task optimization method based on user preferences provided by the present invention; Figure 4 This is a schematic diagram of the multi-objective task optimization device based on user preferences provided by the present invention; Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] Based on this, the present invention provides a method and apparatus for optimizing multi-objective tasks based on user preferences.

[0023] The following is combined Figures 1-3 The present invention describes a multi-objective task optimization method based on user preferences. Figure 1 This is a flowchart illustrating the multi-objective task optimization method based on user preferences provided by the present invention, as shown below. Figure 1 As shown, the method includes: Step S1: Construct a multi-objective optimization task based on the predefined search space of the input variables and multiple pre-built objective functions.

[0024] Specifically, based on the search space of predefined input variables And based on multiple pre-defined objective functions to be optimized. This constitutes a multi-objective optimization task:

[0025] in, It is a discrete combination space (such as process parameters, formulation components, etc.).

[0026] In some embodiments, the search space for input variables is defined according to the specific experimental scenario or decision-making requirements.

[0027] In some embodiments, multiple objective functions are constructed based on historical data or expert experience. All objective functions are black-box in nature, and their response values ​​can only be obtained through experiments or simulations.

[0028] It should be noted that the user-preference-based multi-objective task optimization method provided in this invention is particularly directly aimed at specific experimental design and optimization tasks. Typical application scenarios include: chemical reaction optimization (e.g., with reaction yield and experimental cost as multiple objectives), material property control (e.g., with tensile strength, elongation, and stability as multiple objectives), and process parameter optimization (e.g., with cycle life and production cost as multiple objectives). In these scenarios, experimenters typically need to find the optimal solution that balances multiple objectives within a complex variable space, and the importance of different objectives varies. This invention determines the input variables according to different application scenarios, thereby defining the search space, and constructs corresponding objective functions according to the objectives to be balanced, thus clarifying the input and output variables and optimization objectives, providing the domain and sampling criteria for subsequent Bayesian modeling.

[0029] Step S2: Using a pre-constructed initial training dataset, train the multiple Gaussian process models based on the Bayesian machine learning algorithm; wherein, the multiple Gaussian process models correspond one-to-one with the multiple objective functions.

[0030] Using a pre-constructed initial training dataset as the foundation, an independent Gaussian Process Model (GPR) is trained for each objective based on the Bayesian machine learning algorithm, resulting in a Gaussian Process Model corresponding to each objective function. It should be noted that the number of Gaussian Process Models and objective functions are the same and correspond one-to-one.

[0031] Step S3: Based on the multiple Gaussian process models, obtain the predicted performance of all points in the search space under the corresponding objective function.

[0032] Based on multiple Gaussian process models, the predicted performance of all points in the entire search space under the target is obtained. The predicted performance includes the predicted mean and covariance.

[0033] in, In the objective function Below, input point The predicted value is a random variable that follows a Gaussian distribution (normal distribution). This indicates that the prediction result follows the mean. variance is The normal distribution To predict the mean, To predict covariance.

[0034] This invention utilizes a Gaussian model to model multiple objectives in order to capture the uncertainty of the objective function, which helps to achieve a balance between exploration and utilization.

[0035] Furthermore, in addition to the aforementioned predictive performance, the Gaussian model can also accept intermediate state information for visualization analysis.

[0036] Step S4: Construct a weighted acquisition function according to user requirements, and use the weighted acquisition function to select the current target input point from the search space.

[0037] A weighted sampling function is constructed based on user needs. The sampling is guided towards the user's area of ​​interest in the sampling function, thereby realizing a personalized optimization strategy and obtaining the experimental conditions recommended in the current round (i.e., the current target input point).

[0038] It is important to emphasize that the core idea of ​​this invention is to guide the Bayesian optimization process to focus on the target region most relevant to the user by constructing a multi-objective acquisition function based on explicit weighted preferences, thereby replacing the traditional method that relies on pairwise comparison feedback from the user. This invention is particularly suitable for high-throughput screening or automated decision-making scenarios in experimental design tasks, and can achieve a unification of optimization and preference guidance within a discrete design space.

[0039] Step S5: Perform the experiment based on the current target input point to obtain the actual response of the current target input point.

[0040] Step S6: Update the initial training dataset according to the current target input point and its response value, and repeat steps S2-S6 until the preset termination condition is reached to complete the scheme design of the multi-objective optimization task.

[0041] According to steps S1-S6, the present invention can automatically recommend the most informational solution.

[0042] In some embodiments, the preset termination condition includes a preset budget or convergence condition.

[0043] This invention provides a user-preference-based multi-objective task optimization method. It constructs a multi-objective optimization task using a search space and multiple pre-built objective functions. Then, it models the initial training dataset to predict the performance of each point in the search space, thus obtaining the target performance under untested conditions. Next, it constructs a multi-objective acquisition function based on explicit weighted preferences according to user needs, allowing users to select the target input points that best match their interests. Finally, it implements the target input points on an experimental platform, feeding back new data to update the model and proceeding to the next iteration. In this way, this invention can quickly find solutions that meet multi-objective requirements and align with user preferences within a limited experimental budget, exhibiting good scalability.

[0044] The following provides a further explanation of step S2. In some embodiments, using a pre-constructed initial training dataset, the plurality of Gaussian process models are trained based on a Bayesian machine learning algorithm, including: Step S21: Obtain the initial training dataset.

[0045] Obtain the basic data for model fitting.

[0046] Step S22: Normalize the response values ​​in the initial training dataset and update the initial training dataset based on the processing result.

[0047] All target values ​​(response values) in the initial training dataset are uniformly normalized so that each target value is mapped to a uniform numerical range (e.g., ...). This ensures that the prediction results of the Gaussian process model are comparable across different targets, and maintains consistent transformation of the normalized parameters for subsequent prediction outputs. This avoids the impact of inconsistent target scales on model training and acquisition function calculation.

[0048] Furthermore, in some embodiments, if the target value distribution is extremely uneven, z-score (standard score) can also be used for normalization.

[0049] The embodiments of the present invention employ normalization processing to eliminate differences in target scales, ensuring the stability of the acquisition function and solving the problem that existing multi-objective optimization methods, when dealing with large differences in target scales, are prone to leading to one target dominating the evaluation of the acquisition function, thus impairing the optimization effect.

[0050] Furthermore, variables in the search space can be encoded to provide vectorized representations of the model inputs, facilitating model inference.

[0051] Step S23: Select one of the multiple objective functions as the current objective, and train the Gaussian process model corresponding to the current objective using the initial training dataset based on the Bayesian machine learning algorithm.

[0052] Select the current target, use the pre-built initial training dataset as the basic data, and train a Gaussian process model (GPR) for the current target based on the Bayesian machine learning algorithm to obtain the Gaussian process model corresponding to the current target.

[0053] Step S24: Repeat step S23 until all objective functions are traversed, resulting in multiple Gaussian process models.

[0054] The following provides a further explanation of step S21. In some embodiments, obtaining the initial training dataset includes: Step S211: Sample initial sample points from the search space, perform experiments based on the initial sample points, and obtain the response values ​​of the initial sample points.

[0055] Specifically, from the search space Select the initial sample points for sampling.

[0056] In the process of selecting initial sample points, Latin hypercube sampling, CVT or random sampling can be used, and this invention does not limit the selection.

[0057] S212: Construct an initial training dataset based on the initial sample points and their response values.

[0058] The multi-objective response values ​​of the corresponding initial sample points are obtained through experiments or simulations to construct the initial training dataset.

[0059] in, For the initial training dataset, For the i-th initial sample point, Let be the response value of the Mth objective function corresponding to the i-th initial sample point. This represents the initial number of sample points.

[0060] The following provides a further explanation of step S21. In some embodiments, obtaining the initial training dataset includes: Obtain historical experimental data and construct an initial training dataset based on the historical experimental data.

[0061] Specifically, domain experts can also recommend or use known historical experimental data as the initial training dataset.

[0062] The following provides a further explanation of step S4. In some embodiments, a weighted acquisition function is constructed according to user requirements, and the current target input point is selected from the search space using the weighted acquisition function, including: Step S41: Construct the weighted acquisition function according to the pre-set weights of the users for the corresponding targets of the multiple objective functions.

[0063] Set weight vector , where each weight .

[0064] Furthermore, in some embodiments, if the user cannot specify a preference, a uniform weight can also be set.

[0065] Construct a weighted acquisition function based on the weight vector.

[0066] Step S42: Select candidate input points from the search space, and calculate the weighted expected hypervolume increment value of the candidate input points according to the weighted acquisition function.

[0067] For all candidate input points Calculate its weighted expected hypervolume increment:

[0068] in, This represents a hyperrectangle representing the current predicted dominance region relative to the reference point; This represents a weighted collection function defined based on user weights.

[0069] In this embodiment of the invention, the weighted expected hypervolume increment value is used as the index for selecting the target input point, so that the calculation of the acquisition function is not affected by the physical dimension and has good numerical stability and transferability.

[0070] In some embodiments, the integral is estimated using multiple Monte Carlo sampling during the calculation of the weighted expected hypervolume increment.

[0071] It should be explained that traditional hypervolume increment acquisition functions (such as EHVI, qEHVI, etc.) have extremely high computational costs in high dimensions, especially in large search spaces. This invention uses Monte Carlo integration to estimate weighted hypervolume increments, significantly reducing the analytical quadrature complexity.

[0072] Furthermore, to improve design efficiency, in some embodiments, multiple candidate input points can be evaluated in parallel and sampled using multi-peak methods. This invention employs a full-scale parallel prediction and evaluation mechanism for the candidate set to avoid missing optimal solutions due to insufficient sampling; through numerical approximation and parallel strategies, it solves the problem of low efficiency in high-dimensional spaces inherent in traditional methods.

[0073] The embodiments of the present invention can accurately quantify the "gain" of candidate points on the user's preference target space.

[0074] Step S43: Select the candidate input point with the largest weighted expected supervolume increment value as the current target input point.

[0075] Select the candidate point with the largest weighted expected supervolume increment. This will serve as the next experimental point to perform experiments and update the training set:

[0076] Furthermore, in some embodiments, Top-K diversity recommendation can be set. Specifically, the top K candidate input points with the largest weighted expected hypervolume increment values ​​are selected as the current target input points, which significantly improves the diversity while ensuring the relevance of the recommendation results.

[0077] The following provides a further explanation of step S41. In some embodiments, the weighted acquisition function is constructed based on pre-set user weights for the targets corresponding to the plurality of objective functions, including: Step S411: Set the weights of the targets corresponding to the multiple objective functions based on the user's attention to the targets corresponding to the multiple objective functions.

[0078] A weight vector is set based on the user's focus on different goals. , where each weight .

[0079] Step S412: Construct the weighted acquisition function according to the weights of the targets corresponding to the multiple objective functions.

[0080] It should be noted that traditional EHVI and its extensions (such as qEHVI) default to equal weighting for each objective, lacking flexibility. This invention embeds the preference weights into the weighted acquisition function by setting preference weights, influencing the selection of points in the Pareto solution space. This enables the modeling of asymmetric optimization tendencies between objectives (such as a preference for performance A over performance B), satisfying optimization preferences under different practical needs in the acquisition strategy, and realizing user-driven multi-objective optimization.

[0081] Furthermore, to provide a more detailed explanation of the user preference-based multi-objective task optimization method provided by the present invention, a specific embodiment is given.

[0082] This embodiment uses publicly available high-throughput experimental data. This experiment investigated Pd-catalyzed CH arylation reactions, and the entire experimental variable space consisted of the following five factors: ligands (12 types), solvents (4 types), bases (4 types), reaction temperatures (3 types), and substrate concentrations (3 types), forming a total of... There are several possible experimental conditions. For each combination, the experimenter can obtain the corresponding reaction yield and reaction cost, the latter estimated based on the price of the raw materials and conditions used.

[0083] In this embodiment, the optimization objective is to find the reaction conditions that maximize the reaction yield while minimizing the total reaction cost.

[0084] To demonstrate the effectiveness of the preference-guided multi-objective Bayesian optimization method proposed in this invention, the user preference is set as follows: the pursuit of yield is significantly higher than the reduction of cost. This preference is achieved through a weighted representation of the objectives, specifically set as follows: In other words, the optimization focuses more on discovering reaction conditions with high yields.

[0085] Without assuming any prior historical experimental data, this embodiment first employs a Latin hypercube sampling (LHS) strategy to randomly select five response conditions from the entire search space as initial training data. Subsequently, the actual yields and costs corresponding to these five combinations are obtained through simulation or table lookup. All target data are uniformly normalized before entering the modeling process to avoid scale inconsistencies between targets causing biases in the acquisition function and to ensure the stability of model predictions.

[0086] Next, Gaussian process models were trained for the two objectives of yield and cost, respectively, and the reaction cost was negative, so that both objectives were transformed into maximization problems.

[0087] Based on this, predictions were made for all unevaluated points (i.e., the remaining 1723 combinations excluding the points already tested), and the weighted expected hypervolume increment for each combination point at the current Pareto front was calculated. Under weighted guidance, this acquisition function can effectively prioritize the exploration of high-yield regions while appropriately considering costs.

[0088] Based on the calculated weighted expected hypervolume increment, five values ​​are selected in each round. The point with the highest value is recommended for the next round of experiments. After each round of recommendation experiments, the new experimental results are added to the training set, and the model is retrained to achieve iterative updates of Bayesian optimization.

[0089] In this embodiment, after only 5 iterations (25 experiments in total), the system recommends multiple reaction conditions that simultaneously satisfy high yield and low cost. The yield and cost in the optimization process are as follows: Figure 2 , Figure 3 As shown. In the third round, a combination with a yield of 99.98% and a cost of 0.1114 was successfully discovered; and in the fifth round, an experimental point with a yield of 100% and a cost of 0.0989 was further recommended.

[0090] The results far exceeded the experimenter's expected target (yield). With a success rate of 95% and a relatively low cost, the method provided by this invention demonstrates its ability to achieve rapid convergence and intelligent recommendation under preference-driven conditions.

[0091] From a technical perspective, the weighted hypervolume metric gradually improves with each iteration during the optimization process, and the frontier solution continuously converges towards the user-preferred region. This indicates that the acquisition function successfully guides the model to explore high-value regions.

[0092] In summary, this embodiment demonstrates that the user-preference-based multi-objective task optimization method provided by this invention can fully utilize user objective preferences and quickly find an optimized solution that meets multiple objective requirements within a very small number of experimental rounds, without requiring a large amount of historical data for cold start. This efficient and flexible optimization capability has great practical value and potential for widespread application in tasks such as new reaction development, process screening, or material combination optimization.

[0093] The following describes the multi-objective task optimization device based on user preferences provided by the present invention. The multi-objective task optimization device based on user preferences described below and the multi-objective task optimization method based on user preferences described above can be referred to in correspondence. Figure 4 This is a schematic diagram of the multi-objective task optimization device based on user preferences provided by the present invention, as shown below. Figure 4 As shown, the device includes the following modules: Task module 410 is used to construct a multi-objective optimization task based on a predefined search space of input variables and multiple pre-built objective functions.

[0094] The modeling module 420 is used to train the plurality of Gaussian process models using a pre-constructed initial training dataset and based on a Bayesian machine learning algorithm; wherein the plurality of Gaussian process models correspond one-to-one with the plurality of objective functions.

[0095] The prediction module 430 is used to obtain the predicted performance of all points in the search space under the corresponding objective function based on the multiple Gaussian process models.

[0096] The input module 440 is used to construct a weighted acquisition function according to user requirements and use the weighted acquisition function to select the current target input point from the search space.

[0097] Experiment module 450 is used to perform an experiment based on the current target input point and obtain the actual response of the current target input point.

[0098] The iteration module 460 is used to update the initial training dataset according to the current target input point and its response value, repeat steps S2-S6 until the preset termination condition is reached, and complete the scheme design of the multi-objective optimization task.

[0099] According to the present invention, a multi-objective task optimization device based on user preferences, using a pre-constructed initial training dataset, trains multiple Gaussian process models based on a Bayesian machine learning algorithm, including: S21: Obtain the initial training dataset; S22: Normalize the response values ​​in the initial training dataset, and update the initial training dataset according to the processing result; S23: Select one of the multiple objective functions as the current objective, and train the Gaussian process model corresponding to the current objective using the initial training dataset based on the Bayesian machine learning algorithm; S24: Repeat step S23 until all objective functions are traversed, resulting in multiple Gaussian process models.

[0100] According to the present invention, a multi-objective task optimization device based on user preferences acquires an initial training dataset, including: Initial sample points are sampled from the search space, and experiments are performed based on the initial sample points to obtain the response values ​​of the initial sample points; Based on the initial sample points and their response values, an initial training dataset is constructed.

[0101] According to the present invention, a multi-objective task optimization device based on user preferences acquires an initial training dataset, including: Obtain historical experimental data and construct an initial training dataset based on the historical experimental data.

[0102] According to the present invention, a multi-objective task optimization device based on user preferences constructs a weighted acquisition function according to user needs and uses the weighted acquisition function to select the current target input point from the search space, comprising: The weighted acquisition function is constructed based on the pre-set weights of the users for the corresponding targets of the multiple objective functions; Candidate input points are selected from the search space, and the weighted expected hypervolume increment value of the candidate input points is calculated according to the weighted acquisition function. The candidate input point with the largest weighted expected supervolume increment is selected as the current target input point.

[0103] According to a user preference-based multi-objective task optimization device provided by the present invention, the weighted acquisition function is constructed according to the pre-set weights of users on the corresponding objectives of the multiple objective functions, including: The weights of the targets corresponding to the multiple objective functions are set according to the user's level of attention to them; The weighted acquisition function is constructed based on the weights of the targets corresponding to the multiple objective functions.

[0104] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a user-preference-based multi-objective task optimization method. This method includes: S1: constructing a multi-objective optimization task based on a predefined search space of input variables and multiple pre-built objective functions; S2: training multiple Gaussian process models using a pre-built initial training dataset based on a Bayesian machine learning algorithm; wherein each Gaussian process model corresponds one-to-one with each objective function; S3: obtaining the predicted performance of all points in the search space under the corresponding objective function based on the multiple Gaussian process models; S4: constructing a weighted acquisition function according to user requirements and selecting the current target input point from the search space using the weighted acquisition function; S5: performing an experiment based on the current target input point to obtain the actual response of the current target input point; S6: updating the initial training dataset based on the current target input point and its response value, repeating steps S2-S6 until a preset termination condition is met, thus completing the design of the multi-objective optimization task.

[0105] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the user-preference-based multi-objective task optimization method provided by the above methods. This method includes: S1: constructing a multi-objective optimization task based on a predefined search space of input variables and multiple pre-built objective functions; S2: training the multiple Gaussian process models using a pre-built initial training dataset based on a Bayesian machine learning algorithm; wherein the multiple... A Gaussian process model corresponds one-to-one with the multiple objective functions; S3: Based on the multiple Gaussian process models, obtain the predicted performance of all points in the search space under the corresponding objective functions; S4: Construct a weighted acquisition function according to user requirements, and use the weighted acquisition function to select the current target input point from the search space; S5: Perform experiments based on the current target input point to obtain the actual response of the current target input point; S6: Update the initial training dataset based on the current target input point and its response value, and repeat steps S2-S6 until the preset termination condition is reached to complete the scheme design of the multi-objective optimization task.

[0107] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the user preference-based multi-objective task optimization method provided by the above methods. The method includes: S1: constructing a multi-objective optimization task based on a predefined search space of input variables and multiple pre-constructed objective functions; S2: training the multiple Gaussian process models using a pre-constructed initial training dataset based on a Bayesian machine learning algorithm; wherein the multiple Gaussian process models correspond one-to-one with the multiple objective functions; S3: obtaining the predicted performance of all points in the search space under the corresponding objective functions based on the multiple Gaussian process models; S4: constructing a weighted acquisition function according to user requirements, and selecting the current target input point from the search space using the weighted acquisition function; S5: performing an experiment based on the current target input point to obtain the actual response of the current target input point; S6: updating the initial training dataset based on the current target input point and its response value, repeating steps S2-S6 until a preset termination condition is reached, thus completing the scheme design of the multi-objective optimization task.

[0108] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0109] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-objective task optimization method based on user preferences, characterized in that, include: S1: Construct a multi-objective optimization task based on the predefined search space of input variables and multiple pre-built objective functions; S2: Using a pre-constructed initial training dataset, the plurality of Gaussian process models are trained based on the Bayesian machine learning algorithm; wherein, the plurality of Gaussian process models correspond one-to-one with the plurality of objective functions; S3: Based on the multiple Gaussian process models, obtain the predicted performance of all points in the search space under the corresponding objective function; S4: Construct a weighted acquisition function according to user requirements, and use the weighted acquisition function to select the current target input point from the search space; S5: Perform the experiment based on the current target input point to obtain the actual response of the current target input point; S6: Update the initial training dataset based on the current target input point and its response value, and repeat steps S2-S6 until the preset termination condition is met, thus completing the scheme design for the multi-objective optimization task.

2. The multi-objective task optimization method based on user preferences according to claim 1, characterized in that, Using a pre-constructed initial training dataset, and based on a Bayesian machine learning algorithm, the multiple Gaussian process models are trained, including: S21: Obtain the initial training dataset; S22: Normalize the response values ​​in the initial training dataset, and update the initial training dataset according to the processing result; S23: Select one of the multiple objective functions as the current objective, and train the Gaussian process model corresponding to the current objective using the initial training dataset based on the Bayesian machine learning algorithm; S24: Repeat step S23 until all objective functions are traversed, resulting in multiple Gaussian process models.

3. The multi-objective task optimization method based on user preferences according to claim 2, characterized in that, Obtain the initial training dataset, including: Initial sample points are sampled from the search space, and experiments are performed based on the initial sample points to obtain the response values ​​of the initial sample points; Based on the initial sample points and their response values, an initial training dataset is constructed.

4. The multi-objective task optimization method based on user preferences according to claim 2, characterized in that, Obtain the initial training dataset, including: Obtain historical experimental data and construct an initial training dataset based on the historical experimental data.

5. The multi-objective task optimization method based on user preferences according to claim 1, characterized in that, Constructing a weighted acquisition function based on user requirements, and using the weighted acquisition function to select the current target input point from the search space, including: The weighted acquisition function is constructed based on the pre-set weights of the users for the corresponding targets of the multiple objective functions; Candidate input points are selected from the search space, and the weighted expected hypervolume increment value of the candidate input points is calculated according to the weighted acquisition function. The candidate input point with the largest weighted expected supervolume increment is selected as the current target input point.

6. The multi-objective task optimization method based on user preferences according to claim 5, characterized in that, Based on pre-set user weights for the corresponding objectives of the multiple objective functions, the weighted acquisition function is constructed, including: The weights of the targets corresponding to the multiple objective functions are set according to the user's level of attention to them; The weighted acquisition function is constructed based on the weights of the targets corresponding to the multiple objective functions.

7. A multi-objective task optimization device based on user preferences, characterized in that, include: The task module is used to construct multi-objective optimization tasks based on the predefined search space of input variables and multiple pre-built objective functions; The modeling module is used to train the multiple Gaussian process models using a pre-constructed initial training dataset and based on a Bayesian machine learning algorithm; wherein the multiple Gaussian process models correspond one-to-one with the multiple objective functions; The prediction module is used to obtain the predicted performance of all points in the search space under the corresponding objective function based on the multiple Gaussian process models. The input module is used to construct a weighted acquisition function according to user requirements, and to select the current target input point from the search space using the weighted acquisition function; The experiment module is used to perform experiments based on the current target input point and obtain the actual response of the current target input point; The iterative module is used to update the initial training dataset based on the current target input point and its response value, repeating steps S2-S6 until the preset termination condition is met, thus completing the scheme design for the multi-objective optimization task.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the user preference-based multi-objective task optimization method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the user preference-based multi-objective task optimization method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the user preference-based multi-objective task optimization method as described in any one of claims 1 to 6.

Images