A method, system, and device for selecting a problem-solving layer of a multimodal data model

Abstract

The present invention belongs to the field of big data, and specifically relates to a method, a system, and a device for selecting a problem-solving layer of a multi-modal data model. This method is used to optimize the problem-solving layer in the partitioned order product space. It includes a preselection stage and an optimization stage. The specific steps are as follows: S1: Represent the problem-solving layer as chromosome individuals to generate a quasi-initial population. S2: Define a fitness function. S3: Iteratively update the quasi-initial population and save the optimal individual in each round of iterative update. S4: Improve the selection operator, crossover operator, and mutation operator in the classical genetic algorithm to obtain a new adaptive genetic algorithm. S5: Combine the randomly generated chromosomes and the sub-optimal population as the initial population in the optimization stage. S6: Iteratively update the initial population and select the optimal problem-solving layer after the iteration ends. The present invention solves the problems such as difficult convergence and easy entrapment in local optima when the existing classical genetic algorithm is used to handle this problem.

Description

Technical Field

[0001] The present invention belongs to the field of big data, and specifically relates to a method, a system, and a device for selecting a problem-solving layer of a multi-modal data model. Background Art

[0002] In the field of big data processing, each source or form of information can be called a modality. For example, human touch, hearing, vision, smell; information media such as speech, video, text; various sensors such as radar, infrared, accelerometers, etc. Each of the above can be called a modality. At the same time, the modality can also have a very broad definition. For example, we can regard two different languages as two modalities, and even data sets collected under two different circumstances can also be considered as two modalities.

[0003] Multi-modal data is a type of data that contains multiple different levels of information. For example, the real physical environment is a multi-modal environment, and the data used to analyze and evaluate environmental characteristics is usually multi-modal data. Multi-modal data exists widely, and effectively processing different modal information is an urgent need for big data mining. Existing methods for obtaining knowledge from multi-modal data are all based on a single level, but different levels contain different knowledge. Therefore, we hope to achieve representations at different levels of multi-modal data, and be able to mine from different perspectives, so as to better obtain knowledge.

[0004] In multi-modal data processing methods, the partition-order product space method based on granular computing provides an effective representation means. This model uses granular computing means to provide a multi-level and multi-perspective description of multi-modal information, and each perspective corresponds to a modality. The partition-order product space is a lattice structure, and each problem-solving layer in the lattice structure can provide an understanding and description of the problem from different angles and at different levels. However, due to the large solution space of the partition-order product space, selecting a suitable problem-solving layer in the model is an NP-hard problem. NP-hard problems are usually difficult mathematical problems to solve. The existing technology uses a genetic algorithm to handle NP-hard problems to obtain the best problem-solving layer. However, the classic genetic algorithm generally has problems such as slow convergence speed and easy to fall into local optimum when dealing with problem-solving layers. Summary of the Invention

[0005] In order to solve the problems of difficult convergence and easy to fall into local optimum existing in the selection process of the problem-solving layer of the multi-modal data model by the existing classic genetic algorithm, the present invention provides a method, a system, and a device for selecting a problem-solving layer of a multi-modal data model.

[0006] The present invention is implemented by adopting the following technical solutions:

[0007] A method for selecting a problem-solving layer of a multimodal data model, which is used to optimize the problem-solving layer in the partitioned order product space, and then obtain the optimal problem-solving layer. The problem-solving layer selection method includes a preselection stage and an optimization stage. The preselection stage is processed using the classical genetic algorithm to select some sub-optimal problem-solving layers. The optimization stage is processed using an improved adaptive genetic algorithm to determine the optimal problem-solving layer.

[0008] Among them, the problem-solving layer selection method includes the following processing steps:

[0009] S1: Represent the problem-solving layer in the partitioned order product space as a chromosome individual in the genetic space using real number coding, and then randomly generate a quasi-initial population containing multiple chromosomes.

[0010] S2: Define a fitness function according to the boundary constraint conditions of the classification accuracy index and the granularity index for the target problem-solving layer.

[0011] S3: Set the iteration conditions for the preselection stage, iteratively update the quasi-initial population, and save the optimal individual in each round of iterative update, and then preselect some sub-optimal individuals in the population. The preselection process is as follows:

[0012] S31: Calculate the fitness value of each individual in the quasi-initial population in the current round using the custom fitness function.

[0013] S32: Select the individual with the best fitness value as the optimal individual in the current round and save it in a preselection set.

[0014] S33: Use the classical genetic algorithm to perform selection, crossover, and mutation operations on the quasi-initial population in turn; update the quasi-initial population. The selection method for the selection operation is to use both roulette wheel selection and elitist retention strategy at the same time; the crossover method for the crossover operation is single-point crossover; the mutation method for the mutation operation is single-point mutation.

[0015] S34: Loop through steps S31 - S33 until the iteration conditions set in the preselection stage are reached, end the algorithm iteration, and output the preselection set. All individuals included in the preselection set are the preselected sub-optimal population.

[0016] S4: Improve the selection operator, crossover operator, and mutation operator in the classical genetic algorithm to obtain a new adaptive genetic algorithm. The improvement of the adaptive genetic algorithm is as follows:

[0017] S41: Design the selection operator as an adaptive operator that changes dynamically with the number of evolutionary iterations of the current population. In the early stage of iteration, both roulette wheel selection and elitist retention strategy are selected; in the later stage of iteration, only the elitist retention strategy is selected.

[0018] S42: Design the crossover operator as an adaptive crossover operator where the crossover probability varies dynamically with the current population evolution iteration count and the fitness values of the crossover individuals; select multi-point crossover as the crossover method.

[0019] S43: Design the mutation operator as an adaptive large mutation operator where the density factor varies dynamically with the current population evolution iteration count, and make the density factor in the early stage of iteration greater than that in the later stage of iteration.

[0020] S5: Randomly generate a batch of new chromosomes, and combine the new chromosomes with the sub-optimal population as the initial population in the optimization stage.

[0021] S6: Set the iteration conditions for the optimization stage, perform iterative update on the initial population, and select the optimal problem-solving layer after the iterative update ends; the detailed process is as follows:

[0022] S61: Calculate the fitness values of each individual in the initial population in the current round using a custom fitness function.

[0023] S62: Use the adaptive genetic algorithm to perform adaptive selection, adaptive crossover, and adaptive mutation operations on the initial population in sequence.

[0024] S63: Loop and execute steps S61 - S62 until the iteration conditions set in the pre-selection stage are reached, end the algorithm iteration, and output the individual with the optimal fitness value in the initial population as the required optimal problem-solving layer.

[0025] As a further improvement of the present invention, in the quasi-initial population in step S1 and the initial population in step S5, each bit in the chromosome corresponds to a perspective, and the gene value at each bit corresponds to the layer number under that perspective. For example, in the partition order product space composed of m perspectives v i constituting, 1 ≤ i ≤ m, is a problem-solving layer, 1 ≤ j i ≤ n i . Among them, n i represents the number of layers of perspective v i , represents the j-th layer of the i-th perspective.

[0026] As a further improvement of the present invention, the custom fitness function is defined as follows:

[0027]

[0028] Among them, α and β are the weights of classification accuracy and granularity respectively, satisfying 0 ≤ α ≤ 1, 0 ≤ β ≤ 1, and α + β = 1; r O (l, D) and g(l) are the classification accuracy and granularity of the problem-solving layer l respectively; r ais the classification accuracy that the problem-solving layer needs to achieve; g Thinnest is the granularity of the finest problem-solving layer in the partitioned product space of orders; abs() is a function that returns the absolute value of a given parameter.

[0029] In the present invention, the definition of classification accuracy is as follows:

[0030] Suppose the equivalence classes of the universe of discourse U with respect to the decision attribute D are U / D = {X1, X2,..., X r}, then the classification accuracy of the problem-solving layer l with respect to the decision attribute D is r O (l, D), and it satisfies:

[0031]

[0032] Among them,

[0033]

[0034] In the above formula, X r represents the decision equivalence class, m represents the number of perspectives; x represents the object; "||" represents obtaining the cardinality of the set.

[0035] The calculation formula for the granularity g(l) of the problem-solving layer l is as follows:

[0036]

[0037] Among them, the larger the value of g(l), the finer the granularity of the problem-solving layer l, and vice versa, the coarser the granularity of the problem-solving layer l.

[0038] As a further improvement of the present invention, the iteration termination conditions for both the preselection stage and the optimization stage are set to reach the number of iterations; among them, the number of iterations in the optimization stage is greater than the number of iterations in the preselection stage. Among them, the partitioning methods for the early iteration stage and the late iteration stage are as follows:

[0039] Suppose the maximum number of iterations of the genetic algorithm is T, the current number of iterations is t, 0 < t < T; a threshold γ for piecewise partitioning is given: 0 < γ < 1; then when , the t-th iteration belongs to the early evolution stage, and when , the t-th iteration belongs to the late evolution stage.

[0040] As a further improvement of the present invention, in step S42, the adjustment rule for the adaptive crossover probability P c is as follows:

[0041]

[0042] In the above formula, P b is the given crossover probability boundary value, satisfying 0.5 ≤ Pb ≤ 1; f(p1) and f(p2) are the fitness values of the paternal chromosome p1 and the maternal chromosome p2; where f pavg = avg(f(p1), f(p2)), representing the average of the fitness values f(p1) and f(p2) of the paternal and maternal chromosomes.

[0043] The present invention also includes a problem-solving layer selection system for a multimodal data model. This system uses the problem-solving layer selection method of the multimodal data model as described above to optimize the problem-solving layer in the partitioned order product space. The problem-solving layer selection system includes: a population initialization module, a fitness generation module, a classical genetic algorithm processing module; a preferred sample collection module, an adaptive genetic algorithm processing module, and an optimal individual selection module.

[0044] Among them, the population initialization module is used to represent the problem-solving layer in the partitioned order product space as chromosome individuals in the genetic space by using real number coding. In the preselection stage, multiple chromosomes are randomly generated to form a quasi-initial population; and in the optimization stage, other chromosome individuals in the initial population except the sub-optimal population are generated.

[0045] The fitness generation module is used to solve the fitness values of each sample in the quasi-initial population and the initial population according to a custom fitness function in each round of iteration in the preselection stage and the optimization stage.

[0046] The classical genetic algorithm processing module is used to iteratively update the quasi-initial population in the preselection stage; the classical genetic algorithm processing module includes a selection sub-module, a crossover sub-module, and a mutation sub-module; the selection sub-module uses both roulette wheel selection and elitist retention strategy to perform selection operations on the quasi-initial population. The crossover sub-module uses single-point crossover to perform crossover operations on the quasi-initial population. The mutation sub-module uses single-point mutation to perform mutation operations on the quasi-initial population.

[0047] The preferred sample collection module is used to pick out the individual with the largest fitness value in the quasi-initial population after each round of iteration in the preselection stage to form the required sub-optimal population.

[0048] The adaptive genetic algorithm processing module is used to iteratively update the initial population containing the sub-optimal population during the optimization stage. The adaptive genetic algorithm includes a cycle recognition sub-module, an adaptive selection sub-module, an adaptive crossover sub-module, and an adaptive mutation sub-module. The cycle recognition sub-module is used to determine the type of the current iteration cycle according to a preset rule. The types of the iteration cycle are divided into the early iteration stage and the late iteration stage. The operator of the adaptive selection sub-module is an adaptive operator that dynamically changes with the number of iterations. In the early iteration stage, both roulette wheel selection and elitist retention strategy are selected; in the late iteration stage, only the elitist retention strategy is selected. The operator of the adaptive crossover sub-module is an adaptive crossover operator whose crossover probability dynamically changes with the number of evolutionary iterations of the current population and the fitness values of the crossover individuals, and the crossover operation adopts a multi-point crossover method. The operator of the adaptive mutation sub-module selects an adaptive large mutation operator whose density factor dynamically changes with the number of evolutionary iterations of the current population, and makes the density factor in the early iteration stage greater than that in the late iteration stage.

[0049] The optimal individual selection module is used to output the chromosome individual with the largest fitness value in the population after the iterative update in the optimization stage as the required optimal problem-solving layer.

[0050] The present invention further includes a problem-solving layer selection device for a multi-modal data model. The device includes a memory, a processor, and a computer program that can run on the memory and on the processor. Each functional module in the problem-solving layer selection device is deployed in the manner of the problem-solving layer selection system of the multi-modal data model as described above. When the processor executes the computer program, the steps of the problem-solving layer selection method of the multi-modal data model as described above are implemented, and then the best problem-solving layer is selected in the partitioned order product space.

[0051] The technical solution provided by the present invention has the following beneficial effects:

[0052] 1. The technical solution provided by the present invention introduces the genetic algorithm into the analysis of multi-modal data. Using the genetic algorithm to process NP-hard problems can reduce the complexity of the solution, and then solve the problem-solving layer in the partitioned order product space to obtain the best problem-solving layer that meets the requirements of multiple custom constraints. This method has a wide application prospect in the field of multi-modal big data analysis and application.

[0053] 2. The technical solution provided by the present invention adopts a special two-stage processing method. In the pre-selection stage, several sub-optimal problem-solving layers are obtained through the classical genetic algorithm to optimize the solution space. Then, in the optimization stage, an improved adaptive genetic algorithm is used to continuously iterate the solution space containing the selected sub-optimal problem-solving layers, and then the optimal problem-solving layer is obtained. This two-stage processing method can effectively narrow the range of the solution space and improve the convergence speed of the algorithm.

[0054] 3. The improved adaptive genetic algorithm of the present invention improves the selection operator, crossover operator, and mutation operator according to the state changes of the population during the update process, enabling each operator to be dynamically adjusted following the number of iterations. Furthermore, while maintaining the convergence speed of the algorithm, the diversity of the iterative population is ensured, thereby enhancing the ability of the algorithm's selection result to jump out of the local optimal solution. Specifically, considering that the diversity of the population changes with the increase in the number of population evolution iterations, an adaptive selection operator and an adaptive crossover operator that change dynamically with the current population evolution iteration number are defined. In the early stage of evolution, roulette wheel selection and elitist reservation strategy are combined and the crossover probability is increased. In the later stage of evolution, only the elitist reservation strategy is used and the crossover probability is decreased. Through this operation, the convergence speed can be accelerated while reducing the damage to the population diversity. An adaptive large mutation operator is defined. The probability of executing the large mutation operation is decreased in the early stage of evolution and increased in the later stage of evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flowchart of the steps of a method for selecting a problem-solving layer of a multi-modal data model provided in Embodiment 1 of the present invention.

[0056] Figure 2 It is a schematic diagram of a typical chromosome generated by using the real number coding method in Embodiment 1 of the present invention.

[0057] Figure 3 It is a probability distribution diagram of the roulette wheel selection method in Embodiment 1 of the present invention.

[0058] Figure 4 It is a schematic diagram of the chromosome state change of the single-point crossover method in Embodiment 1 of the present invention.

[0059] Figure 5 Figure 5

[0060] Figure 6 It is a schematic diagram of the chromosome state change of the multi-point crossover method in Embodiment 1 of the present invention.

[0061] Figure 7 It is a phased program flowchart of a method for selecting a problem-solving layer of a multi-modal data model provided in Embodiment 1 of the present invention.

[0062] Figure 8 It is a lattice structure schematic diagram of multi-modal data from multiple detection type perspectives of a gastric disease patient in Embodiment 1 of the present invention.

[0063] Figure 9 It is a program flowchart for formulating an examination plan for a gastric disease patient by applying the method of the present invention in Embodiment 1 of the present invention.

[0064] Figure 10 For the control experiment, on the Kr-vs-kp dataset, the granularity change curves of the selected problem-solving layers of TSAGA and GSS at different thresholds.

[0065] Figure 11 For the control experiment, on the Dermatology dataset, the granularity change curves of the selected problem-solving layers of TSAGA and GSS at different thresholds.

[0066] Figure 12 For the control experiment, on the Student-por dataset, the granularity change curves of the selected problem-solving layers of TSAGA and GSS at different thresholds.

[0067] Figure 13 For the control experiment, on the Divorce dataset, the granularity change curves of the selected problem-solving layers of TSAGA and GSS at different thresholds.

[0068] Figure 14 For the control experiment, on the Urban dataset, the granularity change curves of the selected problem-solving layers of TSAGA and GSS at different thresholds.

[0069] Figure 15 For the control experiment, on the Sonar dataset, the granularity change curves of the selected problem-solving layers of TSAGA and GSS at different thresholds.

[0070] Figure 16 For the control experiment, on the Phishing dataset, the granularity change curves of the selected problem-solving layers of TSAGA and GSS at different thresholds.

[0071] Figure 17 For the control experiment, on the Mushroom dataset, the granularity change curves of the selected problem-solving layers of TSAGA and GSS at different thresholds.

[0072] Figure 18 For the control experiment, on the SCADI dataset, the granularity change curves of the selected problem-solving layers of TSAGA and GSS at different thresholds.

[0073] Figure 19 For the ablation experiment, when using the Urban dataset, the granularity change curves of various algorithm improvements for the selected problem-solving layers at different thresholds.

[0074] Figure 20 For the ablation experiment, when using the SCADI dataset, the granularity change curves of various algorithm improvements for the selected problem-solving layers at different thresholds.

[0075] Figure 21It is a schematic diagram of modules of a problem-solving layer selection system for a multimodal data model provided in Embodiment 2 of the present invention. Detailed implementation manners

[0076] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0077] Embodiment 1

[0078] This embodiment provides a problem-solving layer selection method for a multimodal data model, which is used to optimize the problem-solving layer in the partitioned order product space, and then obtain the optimal problem-solving layer; wherein, as Figure 1 shown, the problem-solving layer selection method includes the following processing steps:

[0079] S1: Represent the problem-solving layer in the partitioned order product space as a chromosome individual in the genetic space by using a real number coding method, and then randomly generate a quasi-initial population containing multiple chromosomes. Each bit in the chromosome corresponds to a perspective, and the gene value at each bit corresponds to the layer number under that perspective.

[0080] For example, in the partitioned order product space composed of m perspectives v i , 1 ≤ i ≤ m, is a problem-solving layer, 1 ≤ j i ≤ n i . Among them, n i represents the number of layers of the perspective v i , represents the jth layer of the ith perspective. Considering that using binary coding for the problem-solving layer will result in too long chromosomes, the present invention uses real number coding to represent the problem-solving layer in the partitioned order product space as a chromosome individual in the genetic space. For example, given the problem-solving layer its corresponding chromosome is as Figure 2 shown. Figure 2 Each bit in the chromosome corresponds to a perspective, and the gene value at each bit corresponds to the layer number under that perspective. After coding, a quasi-initial population containing multiple chromosomes can be obtained.

[0081] Specifically, in this embodiment, the initialized population in the preselection stage is called a quasi-initial population, and the initialized population in the optimization stage is called an initial population. The difference between the two is that all chromosome individuals in the quasi-initial population are randomly generated, while the initial population includes both randomly generated populations and sub-optimal populations obtained through preselection. This optimizes the solution space in the optimization stage.

[0082] S2: Define a fitness function according to the boundary constraint conditions of the classification accuracy index and the granularity index for the target problem-solving layer.

[0083] The method provided in this embodiment mainly solves problems based on granular computing. In this processing method, by omitting detailed information on the coarse-grained layer to describe the problem more abstractly and concisely, problem-solving can be achieved with higher efficiency; however, the solution accuracy of the coarse-grained solution method is not high. On the fine-grained layer, the accuracy of the solution result can be increased through more specific and refined problem descriptions; however, this will also cause a decrease in the solution efficiency at the same time. In this embodiment, in order to improve the convergence rate of the algorithm processing process, the problem-solving layer selected in the partitioned ordered product space makes the granularity of the algorithm processing as coarse as possible while meeting the classification accuracy requirements.

[0084] In order to achieve a balance between the classification accuracy and granularity of the algorithm, this embodiment defines the fitness function of the genetic algorithm in two stages according to the classification accuracy and granularity of the problem-solving layer. Specifically, the custom fitness function is as follows:

[0085]

[0086] where α and β are the weights of the classification accuracy and granularity respectively, satisfying 0 ≤ α ≤ 1, 0 ≤ β ≤ 1, and α + β = 1; r O (l, D) and g(l) are the classification accuracy and granularity of the problem-solving layer l respectively; r a is the classification accuracy that the problem-solving layer needs to achieve; g Thinnest is the granularity of the finest problem-solving layer in the partitioned ordered product space; abs() is a function that returns the absolute value of the given parameter.

[0087] In this embodiment, the definition of the classification accuracy is as follows:

[0088] Suppose the equivalence classes of the universe of discourse U with respect to the decision attribute D are U / D = {X1, X2,..., X r}, then the classification accuracy of the problem-solving layer l with respect to the decision attribute D is r O (l, D), and it satisfies:

[0089]

[0090] where,

[0091]

[0092] In the above formula, X r represents the decision equivalence class, m represents the number of perspectives; x represents the object; "||" represents obtaining the cardinality of the set.

[0093] The calculation formula for the granularity g(l) of the problem-solving layer l is as follows:

[0094]

[0095] Among them, the larger the value of g(l), the finer the granularity of the problem-solving layer l; conversely, the coarser the granularity of the problem-solving layer l.

[0096] S3: Set the iteration conditions for the preselection stage, perform iterative updates on the initial population, and save the optimal individuals in each round of iterative updates, thereby preselecting some sub-optimal individuals in the population. The preselection process is as follows:

[0097] S31: Calculate the fitness value of each individual in the quasi-initial population in the current round using a custom fitness function.

[0098] S32: Select the individual with the best fitness value as the optimal individual in the current round and save it to a preselection set.

[0099] S33: Use the classical genetic algorithm to perform selection, crossover, and mutation operations on the quasi-initial population in turn; update the quasi-initial population.

[0100] S34: Loop through steps S31 - S33 until the iteration conditions set for the preselection stage are reached, end the algorithm iteration, and output the preselection set. All the individuals included in the preselection set are the preselected sub-optimal population.

[0101] In the classical genetic algorithm of this embodiment, the selection method for the selection operation is to simultaneously adopt roulette wheel selection and elitist retention strategy; the crossover method for the crossover operation is single-point crossover; the mutation method for the mutation operation is single-point mutation. Among them, roulette wheel selection is to map the individual fitness value to a roulette wheel. The larger the individual fitness value, the larger the angle allocated to it on the roulette wheel and the greater the probability of being selected. A typical individual distribution diagram of roulette wheel selection is roughly as Figure 3 shown.

[0102] In this embodiment, the single-point crossover method is adopted for the crossover operation. Single-point crossover is to perform a cut at a randomly selected position point and exchange the part on the right side of this position, thereby obtaining two different sub-chromosomes. The specific crossover transformation process is as Figure 4 shown. The operation of the single-point mutation method is to randomly change a gene position in the chromosome from the current value to a feasible value of the threshold, thereby obtaining a new sub-chromosome. A typical transformation in the mutation process is as Figure 5 shown.

[0103] S4: Improve the selection operator, crossover operator, and mutation operator in the classical genetic algorithm to obtain a new adaptive genetic algorithm. The improvement of the adaptive genetic algorithm is as follows:

[0104] S41: Design the selection operator as an adaptive operator that changes dynamically with the number of evolutionary iterations of the current population. In the early stage of iteration, both roulette wheel selection and elitist reservation strategy are selected; in the later stage of iteration, only the elitist reservation strategy is selected.

[0105] Roulette wheel selection and elitist reservation strategy are two common selection operators. Although roulette wheel selection can increase the probability of excellent individuals being selected, it cannot guarantee that the best individual will definitely be retained in the next generation. Although the elitist reservation strategy can ensure that the best individual is retained in the next generation, it cannot control other individuals in the population. Considering that both roulette wheel selection and elitist reservation strategy obtain a faster convergence rate by sacrificing the diversity of the population, therefore, in the early stage of evolution with higher population diversity, combining the two selection operators can achieve better results. As the evolution progresses, the diversity of the population gradually decreases. If roulette wheel selection is still used at this time, it will lead to a high degree of repetition of individuals within the population and is likely to fall into a local optimal solution.

[0106] To address the above problems, this embodiment proposes an adaptive selection operator that changes dynamically with the number of evolutionary iterations of the population. First, divide the population evolution process into the early stage and the later stage by setting a threshold. When in the early stage of evolution, the population has high diversity. To make the algorithm converge as soon as possible, combine roulette wheel selection and elitist reservation strategy to accelerate the convergence rate while ensuring that the optimal individual is retained. In the later stage of evolution, the diversity of the population gradually decreases, and only the elitist reservation strategy is used to reduce the damage to the population diversity while retaining the optimal individual.

[0107] S42: Design the crossover operator as an adaptive crossover operator whose crossover probability changes dynamically with the number of evolutionary iterations of the current population and the fitness value of the crossover individuals; the crossover method selects multi-point crossover, and the specific crossover transformation process is as Figure 6 shown.

[0108] The crossover operator is based on the crossover probability. A higher crossover probability can increase the frequency of crossover operations and accelerate the search process, but it may damage chromosomes with high fitness values. A lower crossover probability can reduce the possibility of chromosomes with high fitness values being damaged, but it will slow down the evolutionary speed. At the same time, during the evolution process, individuals with smaller fitness values should obtain a larger crossover probability to optimize their genotypes, and individuals with larger fitness values should obtain a smaller crossover probability to retain their genotypes. For this reason, this embodiment proposes an adaptive crossover operator whose crossover probability changes dynamically with the number of evolutionary iterations of the current population and the fitness value of the crossover individuals; thus ensuring that the population can be continuously optimized.

[0109] Specifically, for different iteration rounds, the adjustment rule of the adaptive crossover probability P c in the crossover operator of this embodiment is as follows:

[0110]

[0111] In the above formula, P b is a given crossover probability threshold value, satisfying 0.5 ≤ P b ≤ 1; f(p1) and f(p2) are the fitness values of the paternal chromosome p1 and the maternal chromosome p2; where, f pavg = avg(f(p1), f(p2)), representing the average of the fitness values f(p1) and f(p2) of the paternal and maternal chromosomes.

[0112] In addition, it should be emphasized that in the optimization stage, the crossover operation adopts the multi-point crossover method. As Figure 6 shown, the multi-point crossover randomly sets multiple crossover points in the chromosome and then performs gene exchange. In particular, in the multi-point crossover of this embodiment, only the different gene positions in the parental chromosomes are exchanged.

[0113] S43: Design the mutation operator as an adaptive large mutation operator with the density factor changing dynamically with the number of evolutionary iterations of the current population, and make the density factor in the early stage of iteration greater than that in the later stage of iteration.

[0114] Traditional mutation operators set the mutation probability extremely small, which results in a very low possibility of the mutation operation being executed. Coupled with the fact that the mutation operation does not necessarily generate better individuals, it is very difficult to help the algorithm jump out of the local optimal solution only relying on traditional mutation operators. To address this problem, this embodiment adopts a method called large mutation operation to improve the mutation operator, that is, when the maximum fitness F max of a certain generation and the average fitness F avg satisfy:

[0115]

[0116] it is considered that the fitness values of the current population are relatively concentrated, and the mutation probability of all individuals in this generation is set to a probability P mbig more than 4 times larger than the normal mutation probability. Among them, 0.5 ≤ δ < 1 is called the density factor, which determines the probability of the large mutation operation being executed. The closer it is to 0.5, the higher the probability that the population is judged to have concentrated fitness, and the greater the probability of the large mutation operation being executed.

[0117] Based on the large mutation operation, considering that in the early stage of evolution, the population diversity is high and there are few excellent individuals in the population. At this time, it is necessary to reduce the mutation probability to prevent the destruction of excellent genes. In the late stage of evolution, the population diversity is low. At this time, it is necessary to increase the mutation probability to jump out of the local optimal solution. Therefore, this embodiment designs an adaptive large mutation operator that dynamically changes with the number of evolutionary iterations of the current population. First, the population evolution process is divided into the early stage and the late stage. In the early stage of evolution, δ is set to a relatively large value, and in the late stage of evolution, δ is set to a relatively small value. Among them, the value of δ in different periods can be manually adjusted according to specific applications to make the algorithm adapt to the actual problems to be processed.

[0118] Specifically, considering that when the mutation probability is too large, the algorithm approaches random search. To prevent the population from diverging, the method of this embodiment introduces a condition to judge whether the large mutation operation is feasible. If the fitness value of an individual after mutation is less than the fitness value of the individual before mutation, this mutation is cancelled, and the individual before mutation is still retained in the population; otherwise, we replace the original individual in the population with the individual after mutation. The mutation method is single-point mutation.

[0119] The key to the improved adaptive genetic algorithm provided in this embodiment is to control the selection operator, crossover operator, and mutation operator to be adjusted with the number of iteration rounds of the population. Therefore, how to divide the periods of the iteration rounds is also an important factor affecting the performance of the algorithm. Specifically, the division method of the early iteration stage and the late iteration stage in this embodiment is as follows:

[0120] Assume that the maximum number of iterations of the genetic algorithm is T, the current number of iterations is t, 0 < t < T; a threshold γ for piecewise division is given: 0 < γ < 1; then when the t-th iteration belongs to the early stage of evolution, and when the t-th iteration belongs to the late stage of evolution. Among them, the threshold γ is a variable parameter that is manually adjusted according to the actual situation and can take any value between 0 and 1 during the application process. In this embodiment, γ is taken as 0.5.

[0121] S5: Randomly generate a batch of new chromosomes, and combine the new chromosomes with the sub-optimal population as the initial population in the optimization stage.

[0122] S6: Set the iteration conditions in the optimization stage, perform iterative updates on the initial population, and select the optimal problem-solving layer after the iterative updates end; the detailed process is as follows:

[0123] S61: Calculate the fitness value of each individual in the initial population in the current round using a custom fitness function.

[0124] S62: Use the adaptive genetic algorithm to perform adaptive selection, adaptive crossover, and adaptive mutation operations on the initial population in turn.

[0125] S63: Loop and execute steps S61 - S62 until the iteration condition set in the pre - selection stage is reached, end the algorithm iteration, and output the individual with the optimal fitness value in the initial population as the required optimal problem - solving layer.

[0126] In this embodiment, the iteration termination conditions for both the pre - selection stage and the optimization stage are set to reach the number of iterations. Among them, the number of iterations in the optimization stage is greater than that in the pre - selection stage.

[0127] Combined with Figure 6 , the selection method provided in this embodiment improves the process of selecting the problem - solving layer in the partition - order product space based on the classical genetic algorithm into two stages, namely the pre - selection stage and the optimization stage. Among them, the classical genetic algorithm is still used in the pre - selection stage for processing to select some sub - optimal problem - solving layers. The optimization stage is processed using a newly designed adaptive genetic algorithm; it is used to determine the optimal problem - solving layer. Among them, some excellent problem - solving layers can be pre - selected in the pre - selection stage as part of the initial population for the subsequent optimization stage, thereby optimizing the solution space.

[0128] In the optimization stage of this embodiment, an adaptive selection operator, an adaptive crossover operator, and an adaptive large - mutation operator that are dynamically adjusted according to the number of iterations of population evolution are designed for the selection operation, crossover operation, and mutation operation of the algorithm respectively. Thus, the processing stage of the algorithm is adapted to the characteristics of the problem - solving layer. Among them, the improved adaptive selection operator not only ensures the convergence rate of the algorithm but also reduces the damage to population diversity. The improved adaptive crossover operator enables individuals with smaller fitness values to obtain a larger crossover probability to optimize their genotypes, and individuals with larger fitness values to obtain a smaller crossover probability to retain their genotypes, thereby continuously optimizing the population. In the improved adaptive large - mutation operator of the present invention, in the early stage of evolution, the population diversity is high and there are few excellent individuals in the population, so the mutation probability is appropriately reduced to prevent the destruction of excellent genes. In the later stage of evolution, the population diversity is low, and at this time, the mutation probability is increased to jump out of the local optimal solution.

[0129] In the improved selection method of this embodiment, the problems of slow convergence speed and easy entrapment in local optimal solutions of the classical genetic algorithm are effectively overcome. Therefore, the method provided in this embodiment can be applied to multi - modal data analysis and quickly select the best problem - solving layer in the partition - order product space.

[0130] Application example

[0131] The practical application of the problem-solving layer selection method for the multimodal data model provided in this embodiment is very extensive. Using this tool, the implicit information contained in complex data can be effectively analyzed, thereby giving play to the value of big data. The following takes an example in medical diagnosis to illustrate the application process of the method provided in this embodiment:

[0132] When a gastric disease patient goes to the hospital for treatment, in order to obtain accurate diagnostic results, the patient usually needs to undergo different types of medical examinations. Routine examination items usually include: whether there is abdominal pain, vomiting, bleeding, etc. Biochemical examination items usually include: blood routine, glycated hemoglobin, thyroid function, etc. Imaging examinations usually include: abdominal and urinary color Doppler ultrasound, CT, gastroscopy, etc. In addition to routine examinations, biochemical examinations, and imaging examinations, there may also be other special types of examinations.

[0133] The above different types of examinations provide the patient's physiological index information from different perspectives. Therefore, each type of examination can be regarded as a perspective, that is, a modality. In each perspective, as the specific examination items increase, the patient's information will become more and more detailed, thus forming multiple levels. There is a linear order relationship between the layers within a perspective, and multiple perspectives define multiple linear order relationships. Based on the Cartesian product of multiple linear order relationships, we can construct a multimodal data multi-granularity fusion model, that is, a partition order product space. The solution space of the partition order product space is a lattice structure composed of multiple problem-solving layers, where each problem-solving layer is composed of multiple perspectives, each perspective contains only one layer, and each problem-solving layer corresponds to a kind of examination plan.

[0134] For example, the examination plan corresponding to a certain problem-solving layer with a coarser granularity may let the patient undergo examinations for abdominal pain, vomiting, blood routine, and abdominal and urinary color Doppler ultrasound. The examination plan corresponding to a certain problem-solving layer with a finer granularity may let the patient undergo examinations for abdominal pain, vomiting, blood routine, glycated hemoglobin, abdominal and urinary color Doppler ultrasound, and CT. It can be found that the examination plan corresponding to the problem-solving layer with a coarser granularity can obtain more information, but it will increase the patient's examination cost. The examination plan corresponding to the problem-solving layer with a finer granularity can reduce the patient's examination cost, but the information obtained is less. Therefore, it is very important to find an appropriate examination plan corresponding to the problem-solving layer.

[0135] Specifically, the medical information table containing the examination contents received by different patients is a typical multimodal data, and the data structure of this multimodal data is roughly as shown in Table 1:

[0136] Table 1: Multimodal data list of medical information of gastric disease patients

[0137]

[0138] There are m types of examination types in the above medical information form, and there are n i examination items in the i-th type of examination type; when processing this multi-modal data, each type of examination can be regarded as a perspective. As the number of specific examination items increases, the patient's information will become more and more detailed, thus forming multiple levels. Based on the Cartesian product of multiple perspectives, the partition order product space is defined. The partition order product space is a lattice structure, and each node in the lattice structure is an examination plan. The lattice structure composed of m different perspectives is specifically as Figure 8 shown.

[0139] The examination plans in the partition order product space are represented as chromosome individuals in the genetic space, and the real number coding method is adopted. For example, the examination plan l=(1,2,…,2) means that the patient is to undergo ((abdominal pain), (blood routine, glycated hemoglobin),…, (abdominal urinary color Doppler ultrasound examination, CT examination)). After coding, the initial population is randomly generated.

[0140] In the process of generating examination plans in medical diagnosis, since the partition order product space is a lattice structure, the solution space is relatively large, and finding a suitable problem-solving layer in the partition order product space is a typical NP-hard problem. And using the problem-solving layer selection method provided in this embodiment can solve this NP-hard problem and help doctors give the best examination plan with the lowest cost and the best diagnosis and treatment effect.

[0141] Finally, the detailed process of generating examination plans using the method provided in this embodiment is roughly as Figure 9 shown: As can be seen from Figure 9 , in the first stage of the algorithm, the classical genetic algorithm with fewer iteration times is used to select examination plans from the partition order product space, and the best individual in the current population is saved in the individual set POP in each iteration. In the second stage of the algorithm, first, the individuals saved in POP are used as part of the initial population in the second stage, and the remaining chromosomes are still randomly generated according to the traditional method; thus, a better initial population is obtained in the second stage to optimize the solution space. Then, in the second stage, considering that the diversity of the population will change with the increase of the number of population evolution iterations, an adaptive selection operator and an adaptive crossover operator that change dynamically with the current population evolution iteration number are defined. Finally, in order to enhance the ability of the algorithm to jump out of the local optimal solution, an adaptive large mutation operator is defined. Thus, examination plans are further selected in the optimized solution space.

[0142] Verification experiment

[0143] In order to verify the effectiveness of a problem-solving layer selection method (two-stage adaptive genetic algorithm, TSAGA) for a multi-modal data model provided in this embodiment, the following verification experiments are specifically formulated.

[0144] 1. Dataset Preparation

[0145] In this embodiment, 9 datasets are selected for experimental verification, and the specific information of the datasets is shown in Table 2.

[0146] Table 2: Data characteristics of each test dataset used in the verification experiment

[0147]

[0148] For each dataset, first, the conditional attribute set is divided into multiple subsets AT i , 1 ≤ i ≤ n, and each attribute subset AT i is defined as a different perspective. Then, the attributes in each perspective are further divided into multiple subsets AT i k , 1 ≤ k ≤ m. By combining the attribute subsets, a nested sequence of attribute sets is generated where C j = ∪ 1≤k≤j AT i k , forming multiple levels. Finally, a partition order product space is constructed based on multiple perspectives and multiple levels.

[0149] The multi-perspective and multi-level structure of each of the above datasets is shown in Table 3.

[0150] Table 3: Multi-perspective and multi-level structure of each dataset in the verification experiment

[0151]

[0152]

[0153] 2. Experimental Content

[0154] In this embodiment, a set of control experiments is designed. The two-stage adaptive genetic algorithm provided by this experimental group is denoted as TSAGA. The existing method for solving the generalized decision satisfaction scale selection problem in the control group is denoted as GSS. Considering that in actual problem-solving, due to constraints such as time and cost, it is usually not necessary to select the optimal generalized decision scale, but only the satisfactory generalized decision scale needs to be selected. This scale only needs to keep the generalized decision of the decision-making system and the generalized decision of the finest scale consistent to a certain extent. Therefore, in the control group of this embodiment, a satisfactory generalized decision scale is selected under each perspective. Then, all the satisfactory generalized decision scales are combined to form a problem-solving layer, which is called the satisfactory problem-solving layer.

[0155] In this embodiment, the experimental group uses the classification accuracy of the satisfactory problem-solving layer selected by the control group as the standard, selects a problem-solving layer with the same classification accuracy as that selected by the two-stage adaptive genetic algorithm of this embodiment and the generalized decision-making satisfactory scale selection method of the control group, and then compares the granularities of these two problem-solving layers. Among them, the generalized decision-making optimal scale and the generalized decision-making satisfactory scale are defined as follows:

[0156] Given a generalized multi-scale decision-making system Denote the set of all scales as L. Denote as the generalized decision of object x in the decision-making system S K =(U, C K ∪{d}). Let K0∈L be the finest scale. If there is

[0157]

[0158] and then K is called the generalized decision-making optimal scale.

[0159] Given a generalized multi-scale decision-making system Denote the set of all scales as L. Let K0∈L be the finest scale. Given a threshold 0≤ζ≤1, there is

[0160]

[0161] and then K is called the generalized decision-making satisfactory scale.

[0162] In the control experiment, the initial population sizes of the two-stage genetic algorithm are both set to 60. The maximum number of iterations of the first-stage genetic algorithm is 5, the crossover probability is 0.8, and the mutation probability is 0.1. The maximum number of iterations of the second-stage genetic algorithm is 40, the threshold γ is 0.5, the crossover probability threshold P b is 0.8, the conventional mutation probability P msmall is 0.1, and the large mutation probability P mbig is 0.4.

[0163] On the basis of the control experiment, this embodiment also conducts ablation experiments on the provided TSAGA method to verify the improvement degree of the adaptive selection operator, the adaptive crossover operator, and the adaptive large mutation operator on the algorithm. In the ablation experiment, the threshold ζ is set to take 0.2, 0.4, 0.6, 0.8, and 1.0 respectively. Ten experiments are conducted under each threshold, and the average value of the results of the ten experiments is taken as the final result.

[0164] 3. Experimental Results

[0165] In the control experiment, the granularities of the problem-solving layers selected by the two methods, TSAGA and GSS, at different thresholds are shown in Table 4.

[0166] Table 4: Granularities of the problem-solving layers selected by TSAGA and GSS at different thresholds

[0167]

[0168]

[0169] To make the results more intuitive, in this embodiment, the experimental data in Table 4 are also presented in the form of a curve image. The experimental results on 9 datasets are respectively as Figures 10 - 18 shown. In Figures 10 - 18 , the abscissa is the threshold ζ, and the ordinate is the granularity of the problem-solving layer. Combining the data in Table 3 and Figures 10 - 18 it can be concluded that: as the threshold ζ increases, the granularities of the problem-solving layers selected by the two methods, TSAGA and GSS, gradually become finer on most datasets. However, the proposed two-stage adaptive genetic algorithm (TSAGA) selects problem-solving layers with coarser granularities than those selected by the GSS method on most datasets while maintaining the same classification accuracy as GSS. Through the above experimental results, the effectiveness of the proposed two-stage adaptive genetic algorithm in selecting problem-solving layers is demonstrated.

[0170] In the ablation experiment, to make the results clearer, experiments are conducted on the datasets SCADI and Urban with larger solution spaces for constructing the partition-order product space. The threshold ζ is set to take 0.2, 0.4, 0.6, 0.8, and 1.0 respectively, and ten experiments are conducted at each threshold. The average value of the ten experimental results is taken as the final result. The granularities of the problem-solving layers selected without adopting each improvement in the present invention at different thresholds are shown in Table 5.

[0171] Table 5: Granularities of the problem-solving layers of each improvement method under different threshold conditions in the ablation experiment

[0172]

[0173]

[0174] To make the results more intuitive, in this embodiment, the experimental data in Table 4 are also presented in the form of a curve image, as Figure 19 and Figure 20 shown. Figure 19 and Figure 20 the abscissa is the threshold, and the ordinate is the granularity of the problem-solving layer. Combining the data in Table 4 and Figures 19 - 20It can be seen that the improvement degree of the adaptive selection operator on the algorithm is the smallest, while the improvement degrees of the adaptive crossover operator and the adaptive large mutation operator on the algorithm are relatively large. Moreover, the method adopting the full solution provided in this embodiment performs better than other objects in the ablation experiment under various threshold conditions. This proves the effectiveness and superiority of the problem-solving layer selection method of the multimodal data model provided in this embodiment.

[0175] Embodiment 2

[0176] This embodiment further includes a problem-solving layer selection system for a multimodal data model. This system adopts the problem-solving layer selection method of the multimodal data model in Embodiment 1, uses the input multimodal data set, takes a custom fitness function as a restrictive condition, and optimally selects the problem-solving layer in the partitioned order product space to solve the corresponding big data analysis problem.

[0177] As Figure 21 shown, the problem-solving layer selection system provided in this embodiment includes: a population initialization module, a fitness generation module, a classical genetic algorithm processing module; a preferred sample collection module, an adaptive genetic algorithm processing module, and an optimal individual selection module.

[0178] Among them, the population initialization module is used to represent the problem-solving layer in the partitioned order product space as chromosome individuals in the genetic space by using real number coding. In the preselection stage, multiple chromosomes are randomly generated to form a quasi-initial population; and in the optimization stage, other chromosome individuals in the initial population except the sub-optimal population are generated.

[0179] The fitness generation module is used to solve the fitness values of each sample in the quasi-initial population and the initial population according to a custom fitness function in each round of iteration in the preselection stage and the optimization stage.

[0180] The classical genetic algorithm processing module is used to iteratively update the quasi-initial population in the preselection stage; the classical genetic algorithm processing module includes a selection sub-module, a crossover sub-module, and a mutation sub-module; the selection sub-module simultaneously performs selection operations on the quasi-initial population by using the roulette wheel selection and the elite retention strategy. The crossover sub-module performs crossover operations on the quasi-initial population by using single-point crossover. The mutation sub-module performs mutation operations on the quasi-initial population by using single-point mutation.

[0181] The preferred sample collection module is used to pick out the individual with the largest fitness value in the quasi-initial population after each round of iteration in the preselection stage to form the required sub-optimal population.

[0182] The adaptive genetic algorithm processing module is used to iteratively update the initial population containing the sub-optimal population during the optimization phase. The adaptive genetic algorithm includes a cycle recognition sub-module, an adaptive selection sub-module, an adaptive crossover sub-module, and an adaptive mutation sub-module. The cycle recognition sub-module is used to judge the type of the current iteration cycle according to the preset rules. The types of the iteration cycle are divided into the early iteration stage and the late iteration stage. The operator of the adaptive selection sub-module is an adaptive operator that dynamically changes with the number of iterations. In the early iteration stage, both roulette wheel selection and elitist retention strategy are selected; in the late iteration stage, only the elitist retention strategy is selected. The operator of the adaptive crossover sub-module is an adaptive crossover operator whose crossover probability dynamically changes with the number of evolutionary iterations of the current population and the fitness values of the crossover individuals, and the crossover operation adopts the multi-point crossover method. The operator of the adaptive mutation sub-module selects an adaptive large mutation operator whose density factor dynamically changes with the number of evolutionary iterations of the current population, and makes the density factor in the early iteration stage greater than that in the late iteration stage.

[0183] The optimal individual selection module is used to output the chromosome individual with the largest fitness value in the population after the iterative update in the optimization phase, as the required optimal problem-solving layer.

[0184] In this embodiment, the adaptive crossover probability P c in the adaptive crossover sub-module is adjusted according to the following rules:

[0185]

[0186] In the above formula, t represents the current number of iterations; T represents the maximum number of iterations in the current iteration process; γ represents the threshold for segmenting the iteration cycle; represents that the t-th iteration belongs to the early stage of evolution, represents that the t-th iteration belongs to the late stage of evolution; P b is the given crossover probability boundary value, satisfying 0.5 ≤ P b ≤ 1; f(p1) and f(p2) are the fitness values of the paternal chromosome p1 and the maternal chromosome p2; where, f pavg = avg(f(p1), f(p2)), representing the average value of the fitness values f(p1) and f(p2) of the paternal and maternal chromosomes.

[0187] Example 3

[0188] This embodiment further includes a problem-solving layer selection device for a multimodal data model. The device includes a memory, a processor, and a computer program that runs on the memory and can be executed by the processor. Each functional module in the problem-solving layer selection device is deployed in the same manner as the problem-solving layer selection system of the multimodal data model described above. When the processor executes the computer program, it implements the steps of the problem-solving layer selection method of the multimodal data model described above, and further selects the best problem-solving layer in the partitioned order product space.

[0189] The computer device can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a rack server, a blade server, a tower server, or a cabinet server (including an independent server or a server cluster composed of multiple servers) that can execute programs. The computer device of this embodiment at least includes, but is not limited to, a memory and a processor that can communicate with each other through a system bus.

[0190] In this embodiment, the memory (i.e., the readable storage medium) includes flash memory, a hard disk, a multimedia card, a card-type memory (such as an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory can be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc., equipped on the computer device. Of course, the memory can also include both the internal storage unit and the external storage device of the computer device. In this embodiment, the memory is usually used to store the operating system and various application software installed on the computer device. In addition, the memory can also be used to temporarily store various data that have been output or will be output.

[0191] The processor may be a Central Processing Unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor is generally used to control the overall operation of a computer device. In this embodiment, the processor is used to run the program code stored in the memory or process data to implement the processing process of the problem-solving layer selection method of the multi-modal data model in the foregoing Embodiment 1, and then, according to the various metadata in the given multi-modal data model, obtain the best problem-solving layer according to the requirements of the custom fitness function to solve the actual big data analysis and processing problems.

[0192] The foregoing are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for selecting a problem-solving layer of a multimodal data model, which is used to optimize the problem-solving layer in the partitioned order product space, and then obtain the optimal problem-solving layer; characterized in that, The problem-solving layer selection method includes a preselection stage and an optimization stage. The preselection stage is processed using the classical genetic algorithm to select some sub-optimal problem-solving layers. The optimization stage is processed using an improved adaptive genetic algorithm to determine the optimal problem-solving layer. Among them, the problem-solving layer selection method includes the following processing steps: S1: Represent the problem-solving layers in the partitioned product space as chromosome individuals in the genetic space using real number coding, and then randomly generate a quasi-initial population containing multiple chromosomes. S2: Define a fitness function according to the boundary constraint conditions of the classification accuracy index and the granularity index for the target problem-solving layer. S3: Set the iteration conditions for the preselection stage, iteratively update the quasi-initial population, and save the optimal individual in each round of iterative update, and then preselect some sub-optimal individuals in the population. The preselection process is as follows: S31: Calculate the fitness value of each individual in the quasi-initial population in the current round using a custom fitness function. S32: Select the individual with the best fitness value as the optimal individual in the current round and save it in a preselection set. S33: Use the classical genetic algorithm to perform selection, crossover, and mutation operations on the quasi-initial population in turn; update the quasi-initial population. S34: Loop through steps S31 - S33 until the iteration conditions set in the preselection stage are reached, end the algorithm iteration, and output the preselection set; all individuals included in the preselection set are the preselected sub-optimal population. S4: Improve the selection operator, crossover operator, and mutation operator in the classical genetic algorithm to obtain a new adaptive genetic algorithm. The improvement method of the adaptive genetic algorithm is as follows: S41: Design the selection operator as an adaptive operator that changes dynamically with the number of evolutionary iterations of the current population; use both roulette wheel selection and elitist retention strategy in the early stage of iteration; only use the elitist retention strategy in the later stage of iteration. S42: Design the crossover operator as an adaptive crossover operator whose crossover probability changes dynamically with the number of evolutionary iterations of the current population and the fitness values of the crossover individuals; select multi-point crossover as the crossover method. S43: Design the mutation operator as an adaptive large mutation operator whose density factor changes dynamically with the number of evolutionary iterations of the current population, so that the density factor in the early stage of iteration is greater than that in the later stage of iteration. S5: Randomly generate a batch of new chromosomes, and merge the new chromosomes and the sub-optimal population as the initial population for the optimization stage. S6: Set the iteration conditions for the optimization stage, iteratively update the initial population; and select the optimal problem-solving layer after the iterative update ends. The detailed process is as follows: S61: Calculate the fitness value of each individual in the initial population in the current round using a custom fitness function. S62: Use the adaptive genetic algorithm to perform adaptive selection, adaptive crossover, and adaptive mutation operations on the quasi-initial population in turn. S63: Loop through steps S61 - S62 until the iteration conditions set in the preselection stage are reached, end the algorithm iteration, and output the individual with the best fitness value in the initial population as the required optimal problem-solving layer.

2. The problem-solving layer selection method for the multimodal data model according to claim 1, characterized in that: In the quasi-initial population in step S1 and the initial population in step S5, each bit in the chromosome corresponds to a perspective, and the gene value at each bit corresponds to the level number under that perspective.

3. The problem-solving layer selection method of the multimodal data model according to claim 1, wherein: The custom fitness function is defined as follows: Among them, α and β are the weights of classification accuracy and granularity respectively, satisfying 0 ≤ α ≤ 1, 0 ≤ β ≤ 1, and α + β = 1; r O (l, D) and g(l) are the classification accuracy and granularity of the problem-solving layer l respectively; r a is the classification accuracy that the problem-solving layer needs to achieve; g Thinnest is the granularity of the finest problem-solving layer in the partition order product space; abs() is a function that returns the absolute value of a given parameter.

4. The problem-solving layer selection method for the multi-modal data model according to claim 1, characterized in that: In the classical genetic algorithm in step S33, the selection method for the selection operation is to simultaneously adopt roulette wheel selection and elitist retention strategy; the crossover method for the crossover operation is single-point crossover; the mutation method for the mutation operation is single-point mutation.

5. The problem-solving layer selection method of the multimodal data model according to claim 1, characterized in that: The iteration termination conditions for both the preselection stage and the optimization stage are set to reach the number of iterations; among them, the number of iterations in the optimization stage is greater than the number of iterations in the preselection stage.

6. The problem-solving layer selection method of the multimodal data model according to claim 1, characterized in that: In step S41, the division method for the early iteration stage and the late iteration stage is as follows: Suppose the maximum number of iterations of the genetic algorithm is T, the current number of iterations is t, where 0 < t < T; a threshold γ for piecewise partitioning is given: 0 < γ < 1; then when the t-th iteration belongs to the early stage of evolution, and when the t-th iteration belongs to the late stage of evolution.

7. The problem-solving layer selection method of the multimodal data model according to claim 6, characterized in that: In step S42, the adjustment rule of the adaptive crossover probability P c is as follows: In the above formula, P b is the given crossover probability threshold value, satisfying 0.5 ≤ P b ≤ 1; f(p1) and f(p2) are the fitness values of the paternal chromosome p1 and the maternal chromosome p2; where, f pavg = avg(f(p1), f(p2)), representing the average of the fitness values f(p1) and f(p2) of the paternal and maternal chromosomes.

8. A problem-solving layer selection system for a multi-modal data model, characterized in that: It adopts the problem-solving layer selection method of the multimodal data model described in any one of claims 1-7, and optimizes the problem-solving layer in the partitioned order product space; the problem-solving layer selection system includes: A population initialization module, which is used to represent the problem-solving layer in the partitioned order product space as chromosome individuals in the genetic space by using real number coding; randomly generate multiple chromosomes in the preselection stage to form a quasi-initial population; and generate other chromosome individuals in the initial population except the sub-optimal population in the optimization stage; A fitness generation module, which is used to solve the fitness values of each sample in the quasi-initial population and the initial population according to a custom fitness function in each round of iteration in the preselection stage and the optimization stage; A classical genetic algorithm processing module, which is used to iteratively update the quasi-initial population in the preselection stage; the classical genetic algorithm processing module includes a selection sub-module, a crossover sub-module, and a mutation sub-module; the selection sub-module performs a selection operation on the quasi-initial population by simultaneously adopting roulette wheel selection and elitist retention strategy; the crossover sub-module performs a crossover operation on the quasi-initial population by using single-point crossover; the mutation sub-module performs a mutation operation on the quasi-initial population by using single-point mutation; An optimal sample collection module, which is used to select the individual with the largest fitness value in the quasi-initial population after each round of iteration in the preselection stage to form the required sub-optimal population; An adaptive genetic algorithm processing module, which is used to iteratively update the initial population containing the sub-optimal population in the optimization stage; the adaptive genetic algorithm includes a cycle recognition sub-module, an adaptive selection sub-module, an adaptive crossover sub-module, and an adaptive mutation sub-module; wherein the cycle recognition sub-module is used to judge the type of the current iteration cycle according to a preset rule, and the types of the iteration cycle are divided into the early iteration stage and the late iteration stage; the operator of the adaptive selection sub-module is an adaptive operator that changes dynamically with the number of iterations. In the early iteration stage, both roulette wheel selection and elitist retention strategy are selected; in the late iteration stage, only the elitist retention strategy is selected; the operator of the adaptive crossover sub-module is an adaptive crossover operator whose crossover probability changes dynamically with the number of evolutionary iterations of the current population and the fitness values of the crossover individuals, and the crossover operation adopts a multi-point crossover method; the operator of the adaptive mutation sub-module selects an adaptive large mutation operator whose density factor changes dynamically with the number of evolutionary iterations of the current population, and makes the density factor in the early iteration stage greater than that in the late iteration stage; and An optimal individual selection module, which is used to output the chromosome individual with the largest fitness value in the population after the iterative update in the optimization stage, as the required optimal problem-solving layer.

9. The problem-solving layer selection system of the multimodal data model according to claim 8, wherein: The adjustment rule of the adaptive crossover probability P in the adaptive crossover sub-module is as follows: c ​ In the above formula, t represents the current number of iterations; Let \(T\) denote the maximum number of iterations in the current iteration process; let \(\gamma\) denote the threshold for segmenting the iteration cycle. Indicates that the \(t\)-th iteration belongs to the early stage of evolution. Indicates that the \(t\)-th iteration belongs to the late stage of evolution; \(P\) b is the given crossover probability boundary value, satisfying \(0.5\leq P\) b \(\leq1\); \(f(p1)\) and \(f(p2)\) are the fitness values of the paternal chromosome \(p1\) and the maternal chromosome \(p2\); where, \(f\) pavg \(=\text{avg}(f(p1),f(p2))\), representing the average of the fitness values \(f(p1)\) and \(f(p2)\) of the paternal and maternal chromosomes.

10. An apparatus for selecting a problem-solving layer of a multimodal data model, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that: Each functional module in the problem-solving layer selection device is deployed in the manner of the problem-solving layer selection system of the multi-modal data model as described in claim 8. When the processor executes the computer program, the steps of the problem-solving layer selection method of the multi-modal data model as described in any one of claims 1-7 are implemented, so as to select the best problem-solving layer in the partitioned ordered product space.

Images