Experimental space multi-objective optimization method based on Turbo algorithm
By using the Turbo algorithm to randomly sample and filter non-dominated points in the simulation experimental space, and combining Gaussian regression model and causal inference, the problems of high resource consumption and lack of consideration of multi-objective characteristics in traditional methods are solved, thus achieving efficient multi-objective optimization.
Patent Information
- Application Number
- CN202511891352.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-03-27
AI Technical Summary
Traditional simulation experiment space optimization methods consume high resources in complex scenarios and are difficult to handle multi-objective characteristics. They are prone to getting trapped in local optima and do not take into account the experimenters' preferences, resulting in experimental points that contradict their preferences.
The Turbo algorithm is used to randomly sample multiple points in the experimental space. Non-dominated points are screened through simulation experiments. The next round of experimental points is selected by combining Gaussian regression model and causal inference algorithm until the optimization termination condition is met, thus achieving multi-objective optimization.
In complex simulation experiments, the optimal set of experimental points can be quickly found, which improves experimental efficiency and takes into account multi-objective characteristics, avoiding local optima and preference mismatch problems.
Smart Images
Figure CN121744647A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of simulation experiment, in particular to an experiment space multi-objective optimization method based on Turbo algorithm. BACKGROUND
[0002] The simulation experiment space optimization refers to searching and optimizing in a space (i.e. experiment space) spanned by simulation experiment factors and their value ranges, to find a combination of experiment factor values (i.e. experiment point) meeting the experiment purpose, so as to assist in solving the corresponding experiment problem.
[0003] The method of traversing the simulation experiment space for optimization will obviously bring unbearable experiment resource consumption in a relatively complex experiment scenario. The traditional simulation experiment space optimization method has made many efforts around reducing the number of experiments and improving the experiment efficiency. Among them, the classic methods such as uniform experiment design and orthogonal experiment design select representative part of experiment points to replace the experiment space to carry out experiments, so as to reduce the number of experiments, but it is easy to have the problem of generalizing from a part to the whole; the evolutionary optimization method based on genetic algorithm and the like explores in the whole simulation experiment space, but it is easy to fall into local optimum.
[0004] At the same time, the traditional method generally tries to find an optimal experiment point. In actual experiment problems, it is often multi-objective, and different preferences will lead to different optimal experiment points. Before the experiment personnel give the preferences, the experiment point obtained by exploration without considering the preferences is easy to have the problem of being contrary to the preferences of the experiment personnel.
[0005] Therefore, in a relatively complex simulation experiment scenario, it is necessary to research the simulation experiment space multi-objective optimization method around improving the experiment optimization efficiency and comprehensively considering the multi-objective characteristics of the experiment problem. SUMMARY
[0006] Based on this, it is necessary to provide an experiment space multi-objective optimization method based on Turbo algorithm to solve the above technical problems, comprising: In the first round of experiments, a plurality of experiment points are randomly sampled in the experiment space; Based on the preset experiment index, simulation experiments are carried out on the plurality of experiment points to obtain a non-dominated experiment point set in the first round of experiments; Based on the Turbo algorithm and the non-dominated experiment point set, an experiment point of the next round of experiments is screened, and based on the experiment point of the next round of experiments, a simulation experiment is carried out to obtain a non-dominated experiment point set in the next round of experiments and the historical round of experiments; Until the optimization termination condition is met, a target experiment point set is obtained.
[0007] In some embodiments, the random sampling of a plurality of experiment points in the experiment space comprises: obtaining one or more experimental factors and a corresponding value range of each experimental factor; normalizing the value range of each experimental factor to obtain a normalized value range; randomly and uniformly sampling the plurality of experimental points from the normalized value range.
[0008] In some embodiments, simulation experiments are performed on the plurality of experimental points based on a preset experimental index to obtain a non-dominated experimental point set in the first round of experiments, including: For any two experimental points, simulation experiments are performed based on the preset experimental index to obtain index values corresponding to the two experimental points, respectively; Based on the relationship between the index values of the two experimental points, the non-dominated experimental point set is determined.
[0009] In some embodiments, it is determined that the optimization termination condition is met, including: In response to the experimental round satisfying a preset upper limit of number of times, it is determined that the optimization termination condition is met; and / or, In response to the experimental round exceeding a preset number of times threshold and the non-dominated experimental point set of a continuous preset number of rounds satisfying a requirement, it is determined that the optimization termination condition is met.
[0010] In some embodiments, based on the Turbo algorithm and the non-dominated experimental point set, experimental points for the next round of experiments are screened, including: Based on the first round of experimental points and the corresponding experimental index values, a plurality of Gaussian regression models are constructed; Based on the points in the non-dominated experimental point set of the first round, a plurality of trust regions are constructed; In each trust region, the value probability of each experimental point in the next round of experiments is determined based on a causal inference algorithm, and a plurality of sampling points are randomly sampled after weighting; Based on the plurality of Gaussian regression models, the values of each experimental index in each trust region are estimated; Based on the values of the sampling points, the dominance relationship is calculated, and the experimental points for the next round of experiments are screened from each trust region.
[0011] In some embodiments, simulation experiments are performed based on the experimental points for the next round of experiments, including: Simulation experiments are performed on the experimental points for the next round of experiments based on a preset experimental index, and the causal inference algorithm and the trust region are updated.
[0012] In some embodiments, until the optimization termination condition is met, a target experimental point set is obtained, including: Until the optimization termination condition is met, the non-dominated experimental points in the last round of experiments and historical experiments are taken as the target experimental point set.
[0013] The simulation experiment space multi-objective optimization method provided in the application is a method for optimizing a simulation experiment space based on a Turbo algorithm, which is based on a Bayesian regression algorithm based on a trust region, and maximizes the experimental index value as an optimization target to perform multi-objective optimization, and can quickly find an optimal experimental point set in the simulation experiment space. BRIEF DESCRIPTION OF DRAWINGS
[0014] Figure 1 is a flowchart of the simulation experiment space multi-objective optimization method described in the application.
[0015] Figure 2 is a flowchart of the method for generating experimental points of the next round of simulation experiments based on the Turbo algorithm described in the application. DETAILED DESCRIPTION
[0016] In order to make the purpose, technical scheme and advantages of the application more clear and understandable, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.
[0017] Simulation experiment refers to a controllable experiment in a simulated scenario, such as emergency drills, business competition, etc. Simulation experiment space optimization refers to exploring and optimizing in the space formed by experimental factors and their value ranges (i.e. experimental space) to find experimental factor value combinations (i.e. experimental points) that meet the experimental purpose, to assist in solving the corresponding experimental problems. Experimental factors refer to various variables that affect experimental results. The value range is the possible numerical value or selection of these variables, and each factor has a clear value boundary. The experimental space is all possible combinations of these factors and their value ranges, forming a multi-dimensional space. For example, if there are two factors, each with three values, then the experimental space has 3x3=9 possible combinations.
[0018] The method of traversing the experimental space for optimization will obviously bring unbearable experimental resource consumption in a relatively complex experimental scenario. Traditional simulation experiment space optimization methods have made many efforts to reduce the number of experiments and improve experimental efficiency. Among them, classic methods such as uniform experimental design and orthogonal experimental design select representative experimental points to replace the experimental space to carry out experiments, thereby reducing the number of experiments, but this method is prone to the problem of generalization from partial to whole; evolutionary optimization methods based on genetic algorithms, etc. Although they are exploring in the full experimental space, they are prone to local optimization.
[0019] Meanwhile, traditional methods generally aim to find an optimal experimental point. In real-world experimental problems, there are often multiple objectives, and different preferences will lead to different optimal experimental points. Before the experimenter specifies their preferences, experimental points obtained by assigning weights based on experience do not consider preferences and can easily lead to problems that contradict the experimenter's own preferences.
[0020] Therefore, in relatively complex experimental scenarios, it is necessary to study multi-objective optimization methods for experimental space, focusing on improving experimental optimization efficiency and comprehensively considering the multi-objective characteristics of experimental problems.
[0021] Based on this, this embodiment of the invention focuses on a specific exploratory experiment. After selecting three experimental indicators and setting five experimental factors and their value ranges, the experimental space reached a scale of hundreds of millions due to the explosion of possible combinations of experimental factor values. Furthermore, the researchers could not clearly define the importance of the experimental indicators; therefore, multi-objective optimization of the experimental space was necessary to obtain an optimal set of experimental points that met the objectives of this experiment.
[0022] In this embodiment, an experimental space multi-objective optimization method based on the Turbo algorithm provided by the present invention is applied to improve experimental efficiency. (See appendix) Figure 1 This invention provides a multi-objective optimization method for experimental space based on the Turbo algorithm, comprising the following steps: Step S101: First round, uniformly random sampling in the simulation experimental space. One experimental site; Without loss of generality, the experimental factors chosen by the experimenters are: The corresponding range of values is The normalized range of values is ,exist The 1000 experimental points obtained by uniform random sampling are ,in, , .
[0023] Step S102: Run the simulation experiment and calculate the non-dominated experimental points in this round of experiments and previous experiments; The simulation experiment includes: first, changing the experimental points from... Mapped to Then, the experimental points are assembled into experimental samples; finally, these experimental samples are sent to the simulation system for execution.
[0024] The term "non-dominated experimental point" refers to an experimental point that is not dominated by other experimental points. The "dominance relationship" refers to the fact that if two experimental points... and After simulation testing of the corresponding experimental samples, The experimental index values of the corresponding experimental samples are better than Corresponding experimental sample, then called dominate .
[0025] Without loss of generality, let the experimental index selected by the experimenter be , which are positive indicators. and The experimental index values of the corresponding experimental samples are and Therefore, the calculation method of non-dominated points is as follows: If , then is not dominated by , denoted as .
[0026] The set composed of the experimental points of the current experiment and the previous experiments is , and the set composed of the non-dominated experimental points in the current experiment and the previous experiments is .
[0027] If , are all true, then .
[0028] The number of points in is . If , then randomly delete points.
[0029] Step S103: Initialize the radius of the trust region . Wherein, , and are hyperparameters. In the present embodiment, according to the reference, set , , .
[0030] Step S104: Determine whether the experimental space optimization termination condition is met. If not, go to step S105; otherwise, go to step S107. The experimental space optimization termination condition includes: (1) The experiment has reached the pre-set upper limit of the number of rounds . Denote the current round as . If , terminate the experiment. In the present embodiment, the upper limit of the number of rounds is set to .
[0031] (2) If the experiment has been running for more than In the third round, the non-dominated solution remained unchanged for three consecutive rounds. Let the [number] round be... The set consisting of the non-dominated experimental points of the round experiment and previous experiments is ,if and If the condition is met, the experiment will be terminated.
[0032] Step S105: Based on the experimental causal reasoning suggestions, and using the Turbo algorithm, select the experimental points for the next round of experiments; The aforementioned experimental causal reasoning suggestion refers to suggestions for the selection of experimental factors in subsequent rounds of experiments, based on data analysis using causal reasoning methods, using the experimental points and indicators of the current round and previous rounds of experiments. Here, this suggestion is simply denoted as "the probability of recommending different values for the factors in the next round of experiments." For example, experimental factors... Its normalized range Inside, there is Each of the following values is: The experimental causal reasoning suggests that the probabilities of the factors taking these values in the next round of experiments are as follows: ,and .
[0033] The method for generating experimental points for the next round of simulation experiments based on the Turbo algorithm includes the following steps: Step S501: According to Based on the values of the three corresponding experimental indicators, three independent Gaussian process regression models were constructed. Among them, Gaussian Process Regression (GPR), also known as the kriging method, is a nonparametric Bayesian regression method based on Gaussian processes. It models the unknown function using Gaussian processes and estimates the values of unobserved data points based on observed data points.
[0034] Step S502: with Using point M as the center, construct a side with length M of Each hypercube serves as a corresponding trust region. Indicates the trust region radius. express The number of midpoints.
[0035] The The upper bound of the value of each dimension of the trust region is: The lower bound is .in, This represents the coordinates of the center point of the trust region.
[0036] Step S503: Within each trust region, determine the probability of the value based on causal inference. Weighted uniform random sampling One point.
[0037] Step S504: Utilize the previously constructed Gaussian process regression model To estimate (predict) the values of each experimental index at each sampling point in each trust region.
[0038] Step S505: Within each trust region, calculate The dominance relationship between the sampling points is used to filter out the samples based on the dominance relationship. Each point is selected as a candidate point in the trust region, and the points selected from each trust region are output as the experimental points for the next round of simulation experiments. This indicates rounding up to the nearest integer.
[0039] Step S106: Call step S102, update the experimental causal inference suggestion based on it, update the trust region radius, and go to step S104.
[0040] The method for updating the trust region radius is as follows: If, in step S105, a non-dominant point is updated or added in each trust domain, the exploration is considered successful; otherwise, the exploration is considered unsuccessful. The cumulative number of successful explorations in each trust domain is denoted as . The cumulative number of failed explorations was .if Then update the radius of the trust region to be At the same time, ;if Then update the radius of the trust region to be At the same time, In this embodiment, hyperparameters , .
[0041] It is worth noting that for points that are updated or added in each trust domain, the radius of the trust domain will be the updated radius when the trust domain is constructed in the next round; if it is not updated, the original trust domain radius will be inherited.
[0042] Step S107: Output the non-dominated experimental points of the current round of experiments and previous experiments as the preferred experimental point set for this experiment.
[0043] In one embodiment, an experimental space multi-objective optimization apparatus based on the Turbo algorithm is provided. The apparatus includes: The sampling module is used to randomly sample multiple experimental points in the experimental space during the first round of experiments; The simulation module is used to conduct simulation experiments on the multiple experimental points based on preset experimental indicators to obtain the set of non-dominated experimental points in the first round of experiments. The filtering module is used to filter the experimental points for the next round of experiments based on the Turbo algorithm and the set of non-dominated experimental points, and to conduct simulation experiments based on the experimental points of the next round of experiments to obtain the set of non-dominated experimental points in the next round of experiments and the historical rounds of experiments. The optimization module is used to obtain the target set of experimental points until the optimization termination condition is met.
[0044] It should be noted that the experimental space multi-objective optimization device based on the Turbo algorithm provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the experimental space multi-objective optimization device based on the Turbo algorithm provided in the above embodiments and the experimental space multi-objective optimization method embodiments based on the Turbo algorithm belong to the same concept. The specific implementation process is detailed in the method embodiments and will not be repeated here.
[0045] According to one aspect of this application, embodiments of the present invention also provide a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication component. When the computer program is executed by a processor, it performs the experimental space multi-objective optimization method based on the Turbo algorithm provided in embodiments of this application.
[0046] Furthermore, embodiments of the present invention also provide a computer device, which includes a processor and a memory. The memory stores a computer program, and the processor is capable of executing the computer program stored in the memory. When the computer program is executed by the processor, it can implement the experimental space multi-objective optimization method based on the Turbo algorithm provided in any of the above embodiments.
[0047] The above embodiments are only used to illustrate the technical solutions of the present invention, but the present invention is not limited thereto. Any changes in the art that do not reflect inventiveness, as well as improvements made in accordance with the principles and spirit of the present invention, should be covered within the protection scope of the present invention.
[0048] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A multi-objective optimization method for experimental space based on the Turbo algorithm, characterized in that, include: In the first round of experiments, multiple experimental points were randomly sampled in the experimental space; Simulation experiments were conducted on the multiple experimental points based on preset experimental indicators to obtain the set of non-dominated experimental points in the first round of experiments; where a non-dominated experimental point is one that is not dominated by other experimental points. Based on the Turbo algorithm and the set of non-dominated experimental points, the experimental points for the next round of experiments are selected, and simulation experiments are conducted based on the experimental points of the next round of experiments to obtain the set of non-dominated experimental points in the next round of experiments and the historical rounds of experiments. The process continues until the optimization termination condition is met, at which point the target set of experimental points is obtained.
2. The method according to claim 1, characterized in that, The random sampling of multiple experimental points in the experimental space includes: Obtain one or more experimental factors, and the range of values for each experimental factor; The range of values for each experimental factor is normalized to obtain the normalized range of values. The multiple experimental points are obtained by randomly and uniformly sampling from the normalized value range.
3. The method according to claim 2, characterized in that, Simulation experiments were conducted on the multiple experimental points based on preset experimental indicators to obtain the set of non-dominated experimental points in the first round of experiments, including: For any two experimental points, a simulation experiment is performed based on the preset experimental index to obtain the index values corresponding to the two experimental points respectively. The set of non-dominated experimental points is determined based on the relationship between the index values of the two experimental points.
4. The method according to claim 1, characterized in that, Determining that the optimization termination condition is met includes: In response to the experimental rounds meeting the preset maximum number of times, it is determined that the optimization termination condition is met; And / or, In response to the number of experimental rounds exceeding a preset threshold and the set of non-dominated experimental points for a consecutive preset number of rounds meeting the requirements, the optimization termination condition is determined to be satisfied.
5. The method according to claim 1, characterized in that, Based on the Turbo algorithm and the aforementioned set of non-dominated experimental points, the experimental points for the next round of experiments are selected, including: Based on the first round of experimental points and the corresponding experimental index values, multiple Gaussian regression models were constructed. Multiple trust domains are constructed centered on the points in the first round of non-dominant speech points; In each trust region, the probability of each experimental point in the next round of experiments is determined based on a causal inference algorithm, and multiple sampling points are randomly sampled after weighting. Predict the values of each experimental index in each trust region based on multiple Gaussian backtracking models; The dominance relationship is calculated based on the values of the sampling points, and the experimental points for the next round of experiments are selected from each trust region.
6. The method according to claim 5, characterized in that, Simulation experiments will be conducted based on the experimental points of the next round of experiments, including: Based on preset experimental indicators, simulation experiments are conducted on the experimental points of the next round of experiments, and the causal reasoning algorithm and the trust region are updated.
7. The method according to claim 1, characterized in that, Until the optimization termination condition is met, the target experimental point set is obtained, including: Until the optimization termination condition is met, the non-dominated experimental points in the last round of experiments and the historical experiments are taken as the target experimental point set.
8. A multi-objective optimization device for experimental space based on the Turbo algorithm, characterized in that, include: The sampling module is used to randomly sample multiple experimental points in the experimental space during the first round of experiments; The simulation module is used to conduct simulation experiments on the multiple experimental points based on preset experimental indicators to obtain the set of non-dominated experimental points in the first round of experiments. The filtering module is used to filter the experimental points for the next round of experiments based on the Turbo algorithm and the set of non-dominated experimental points, and to conduct simulation experiments based on the experimental points of the next round of experiments to obtain the set of non-dominated experimental points in the next round of experiments and the historical rounds of experiments. The optimization module is used to obtain the target set of experimental points until the optimization termination condition is met.
9. The apparatus according to claim 8, characterized in that, The sampling module is specifically used for: Obtain one or more experimental factors, and the range of values for each experimental factor; The range of values for each experimental factor is normalized to obtain the normalized range of values. The multiple experimental points are obtained by randomly and uniformly sampling from the normalized value range.
10. The apparatus according to claim 9, characterized in that, The simulation module is specifically used for: For any two experimental points, a simulation experiment is performed based on the preset experimental index to obtain the index values corresponding to the two experimental points respectively. The set of non-dominated experimental points is determined based on the relationship between the index values of the two experimental points.