Genetic breeding strategy optimization method based on evolutionary algorithm

Abstract

The invention discloses a genetic breeding strategy optimization method based on an evolutionary algorithm, and the method comprises the steps: S1, collecting and outputting a breeding multi-source data set, and obtaining the genotype and phenotype data of a parent group; s2, constructing a mathematical model of genetic breeding, and designing a chromosome coding scheme; s3, initializing an initial population, and outputting a fitness score of a breeding strategy by adopting an improved DeepFM model; s4, executing evolutionary iteration operation, and generating a filial generation population by adopting a CMA-ES algorithm; s5, judging whether a termination condition is met or not, if not, repeating iteration, and if yes, outputting an optimal solution breeding strategy; and S6, decoding the optimal solution breeding strategy, and outputting a specific breeding scheme. According to the method, the strategy potential is accurately evaluated through the improved DeepFM model, efficient optimization is carried out in combination with the CMA-ES algorithm, intelligence and automation of breeding decision making are achieved, and the breeding efficiency and the success rate of good variety breeding are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and genetic breeding, and in particular to a genetic breeding strategy optimization method based on evolutionary algorithms. Background Technology

[0002] Evolutionary algorithms for optimizing genetic breeding strategies have been widely applied in recent years to crop improvement, livestock breeding, and other fields due to their global search capabilities in high-dimensional solution spaces, becoming an important development direction for achieving precision breeding. However, in practical applications, modern breeding problems face many challenges, such as high dimensionality of decision variables, nonlinear genotype-phenotype relationships, and diversified breeding objectives. The deployment effectiveness of traditional evolutionary algorithms is still constrained by many factors.

[0003] Most current breeding optimization methods rely on simplified mathematical models or empirical formulas to evaluate breeding strategies, making it difficult to fully explore the complex linkage and interaction effects in parental genotypes. This results in a lack of accuracy in assessing the fitness of breeding strategies. Some systems only use fixed crossover and mutation operators to evolve the population, ignoring the mixed encoding characteristics of continuous and discrete variables, which limits the search efficiency and convergence performance of the algorithm in complex strategy spaces. Furthermore, the optimization process lacks an adaptive adjustment mechanism for the search direction, making it difficult to dynamically guide population evolution based on historical success experiences, thus affecting the algorithm's ability and stability to find the global optimum.

[0004] Furthermore, most existing evolutionary algorithms employ static search mechanisms in breeding optimization, failing to dynamically adjust the search step size and direction based on changes in population distribution and fitness landscape. This leads to the algorithms easily getting stuck in local optima or converging slowly in complex terrains, making it difficult to adapt to the continuous adjustment and optimization of breeding objectives. Consequently, the practical value and breeding efficiency of the models in real breeding projects are severely affected.

[0005] Therefore, how to provide a genetic breeding strategy optimization method based on evolutionary algorithms is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a genetic breeding strategy optimization method based on evolutionary algorithms. This invention fully integrates key steps such as multi-source breeding data construction, hybrid coding chromosome design, improved DeepFM model evaluation, CMA-ES algorithm evolutionary iteration, and optimal strategy decoding output. It constructs an improved DeepFM model, introducing graph neural network technology into breeding strategy evaluation. This breaks through the traditional model's assumption of genotype data independence, enabling in-depth mining of linkage and interaction effects between genes, thereby achieving accurate and biologically interpretable evaluation of breeding strategy potential. Using the CMA-ES algorithm as the optimization engine, it achieves efficient and robust global optimization in a complex hybrid coding space through adaptive adjustment of the search distribution. The synergistic innovation of these two methods constructs a data-driven, self-evolving intelligent breeding decision-making closed loop, significantly improving the scientific rigor and efficiency of breeding strategy optimization. It provides a novel technical path to solve core bottleneck problems in traditional breeding, such as reliance on experience, biased evaluation, and blind optimization.

[0007] A genetic breeding strategy optimization method based on evolutionary algorithms according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect breeding multi-source data, output breeding multi-source dataset, and simultaneously obtain genotype and phenotypic data of parental populations;

[0009] S2. Based on the multi-source dataset of breeding, construct a mathematical model of genetic breeding, design a chromosome coding scheme, map decision variables to genes on chromosomes, and output the chromosome coding scheme.

[0010] S3. Using the chromosome coding scheme, the initial population is randomly initialized, and the improved DeepFM model is trained using genotype and phenotypic data based on the parent population to output the fitness score of the breeding strategy.

[0011] S4. Perform evolutionary iteration operation, input the current population into the improved DeepFM model, obtain the fitness score of each breeding strategy, and use the CMA-ES algorithm to efficiently recombine and mutate the breeding strategy population represented in continuous variable form to generate offspring population.

[0012] S5. Determine whether the evolutionary iteration operation meets the preset termination condition. If it does, terminate the iteration and select the breeding strategy with the highest fitness score from the current offspring population as the optimal solution breeding strategy output. If it does not meet the condition, take the offspring population as the new generation current population and repeat the evolutionary iteration operation.

[0013] S6. Use the chromosome encoding scheme to decode the optimal breeding strategy, reverse analyze the chromosome encoding into a specific breeding strategy scheme and output it.

[0014] Optionally, S1 specifically includes:

[0015] S11. Obtain the breeding project plan, and extract from the breeding project plan the trait indicators used to define the breeding objectives, the optional parent populations as decision variables, and the constraints of the breeding process.

[0016] S12. For each parent in the selectable parent population, collect leaf tissue samples, perform genotyping and sequencing on the leaf tissue samples, and obtain genotypic data of the parent population.

[0017] S13. Record multiple key growth stages of the selectable parent population, use measuring instruments to record the values ​​of plant height, ear length, thousand-grain weight and yield-related traits of each parent, and associate the recorded values ​​with the corresponding parent identity information to generate phenotypic data of the parent population.

[0018] S14. Integrate phenotypic indicators, selectable parental populations, constraints, genotype data, and phenotypic data in the form of structured tables to construct and output a multi-source breeding dataset.

[0019] Optionally, S2 specifically includes:

[0020] S21. Read the trait indicators and constraints in the multi-source dataset of breeding, preset the weight coefficient for each trait indicator according to the importance of breeding, set the weighted sum of all trait indicators as the optimization objective, transform the constraints into mathematical inequalities or equations, and establish a multi-objective integer programming model with the weighted sum as the optimization objective and the transformed constraints as the boundary, as the mathematical model of genetic breeding.

[0021] S22. A complete breeding strategy is broken down into four decision variables in chronological order: paternal selection, maternal selection, hybridization combination method, and offspring selection criteria.

[0022] S23. Assign an integer gene bit to the parent selection decision variable. The value range of the integer gene bit is from 1 to N, where N is the total number of the available parent population. Each value uniquely corresponds to the index number of a parent in the available parent population. Assign another independent integer gene bit to the maternal selection decision variable. The value range and the correspondence rule between the index number are exactly the same as those for the parent selection decision variable.

[0023] S24. Assign a binary gene bit of length 1 to the decision variable of hybridization combination. When the value of the gene bit is 0, it represents a single cross; when the value of the gene bit is 1, it represents a triple cross.

[0024] S25. Assign a continuous real-numbered gene locus with a preset value range to the standard decision variable for offspring selection. The value of the continuous real-numbered gene locus is used as a cutoff threshold, and only offspring individuals whose fitness score ranking is higher than the proportion corresponding to the cutoff threshold are retained.

[0025] S26. Linearly concatenate the integer gene bits representing the paternal selection, the integer gene bits representing the maternal selection, the binary gene bits representing the hybridization combination, and the continuous real number gene bits representing the offspring selection criteria in a pre-defined order to form a chromosome vector, i.e., the chromosome encoding scheme, and output it.

[0026] Optionally, S3 specifically includes:

[0027] S31. Read the chromosome coding scheme, and randomly generate a preset number of chromosomes according to the gene positions and value ranges defined in the chromosome coding scheme to form an initial population containing multiple different breeding strategies.

[0028] S32. An improved DeepFM model is trained using genotype and phenotypic data from the parent population, and different fitness scores are output using breeding strategies in the initial population.

[0029] Optionally, S32 specifically includes:

[0030] Read the genotype data of the parent population. The genotype data is an M×N 0 / 1 matrix, where M is the number of parents and N is the total number of gene loci. Calculate the linkage disequilibrium r² value between all pairs of gene loci in the 0 / 1 matrix. That is, calculate the product of allele frequencies as the expected frequency of the haplotype. Divide the square of the deviation between the actual observed frequency and the expected frequency by the product of the expected frequencies and then normalize to obtain the r² value.

[0031] Construct an N×N LD matrix, set an LD threshold, set the positions in the LD matrix with r² values ​​greater than the LD threshold to 1, and set the other positions to 0, and generate an N×N adjacency matrix, which is the gene association graph;

[0032] In the improved DeepFM model, the original embedding layer is replaced with a graph embedding layer. The internal structure of the graph embedding layer is a two-layer graph convolutional network. The weight matrix of the graph convolutional network is initialized, and the adjacency matrix is ​​loaded as the set graph structure parameters of the graph convolutional network.

[0033] When processing the paternal decision variable in a breeding strategy, the row vectors corresponding to the values ​​of the paternal decision variables are extracted from the M×N 0 / 1 matrix to obtain a paternal genotype vector of length N, which is then input into the graph embedding layer. The first layer graph convolution operation is performed, and the adjacency matrix and the paternal genotype vector are multiplied and then multiplied by the first layer preset weight matrix to obtain the first layer graph convolution output vector.

[0034] Perform the second-layer graph convolution operation, multiply the adjacency matrix with the output vector of the first-layer graph convolution and multiply by the second-layer preset weight matrix to obtain the output vector of the second-layer graph convolution, and output it as the parent graph embedding vector of the parent decision variable. Repeat the same graph embedding operation on the parent decision variable to generate the parent graph embedding vector of the parent.

[0035] The parental graph embedding vector of the father, the parental graph embedding vector of the mother, the binary scalar representing the hybridization combination mode, and the continuous real scalar representing the offspring selection criteria are concatenated end to end to form a concatenated feature vector.

[0036] The concatenated feature vectors are input into the FM component in the improved DeepFM model. All concatenated feature vectors are combined pairwise, the cross product of each pair of concatenated feature vectors is calculated, and all cross product results are summed to generate a numerical value as a low-order interaction feature.

[0037] The concatenated feature vector is input into the DNN component of the improved DeepFM model. The DNN component consists of three fully connected layers, each followed by a ReLU activation function. Different preset weights are applied sequentially, and the weights and biases are summed and ReLU activation is performed to output high-order nonlinear features.

[0038] The low-order interactive features and high-order nonlinear features are concatenated according to their corresponding features, and then weighted and summed with the preset output layer weights. Finally, the preset output layer bias is added to obtain the fitness score of the current breeding strategy.

[0039] Optionally, S4 specifically includes:

[0040] S41. Input the chromosome vector of each breeding strategy in the current population into the improved DeepFM model to obtain the fitness score of each breeding strategy, and sort the breeding strategies in the current population from high to low.

[0041] S42. Initialize the parameters of the CMA-ES algorithm. Set the arithmetic mean of the chromosome vectors of the top 10 breeding strategies with the highest fitness scores in the current population as the mean vector, set the step size parameter, initialize the identity matrix as the covariance matrix, and initialize a zero vector with the same dimension as the chromosome vector as the evolutionary path.

[0042] S43. From a multivariate normal distribution with the mean vector as the distribution center and the covariance matrix and step size parameter as the common determination, randomly sample to generate a preset number of new candidate chromosome vectors to obtain a preset number of candidate offspring breeding strategies and input them into the improved DeepFM model to obtain the fitness score of each candidate offspring breeding strategy and sort them from high to low, and select the top preset number of candidate offspring breeding strategies with the highest fitness scores.

[0043] S44. Calculate the new mean vector by multiplying the chromosome vectors of the selected first preset candidate offspring breeding strategies by preset weight coefficients that decrease in rank, and summing the weighted chromosome vectors element by element to obtain the new mean vector.

[0044] S45. Update the evolutionary path. Calculate the difference vector between the new mean vector of the current generation and the old mean vector of the previous generation. Divide the difference vector by the current step size parameter to obtain the normalized movement vector. Multiply the evolutionary path vector of the previous generation by the preset decay coefficient and add it to the result of multiplying the normalized movement vector by the preset amplification coefficient to generate the new generation's evolutionary path.

[0045] S46. Adaptively update the covariance matrix by performing an outer product operation on the new generation evolution path vector and its own transpose vector to generate a matrix with rank 1. For the first preset number of candidate offspring breeding strategies, calculate the difference between each chromosome vector and the old mean vector, divide each difference vector by the current step size parameter to normalize it, and perform an outer product operation on it with its own transpose vector to generate a preset number of matrices with rank 1.

[0046] S47. The matrix with a rank of 1 is weighted and summed according to the preset weights to generate a weighted matrix. The covariance matrix of the previous generation is multiplied by the preset retention coefficient, and then the result of multiplying the matrix with a rank of 1 by the first learning rate and the result of multiplying the weighted matrix by the second learning rate are added to obtain the updated covariance matrix.

[0047] S48. Use the updated mean vector and covariance matrix as parameters for the new generation CMA-ES algorithm, and use the preset number of candidate offspring breeding strategies generated by the new parameters as the offspring population output.

[0048] Optionally, S5 specifically includes:

[0049] S51. Read the preset iteration number threshold, and compare the actual number of executions of the current evolution iteration with the iteration number threshold to determine whether the current actual number of executions has reached the iteration number threshold.

[0050] S52. Read the fitness scores of all breeding strategies in the offspring population, find the highest fitness score, and record the change value of the highest fitness score in the most recent consecutive preset generations, and determine whether the change value is less than the preset convergence threshold.

[0051] S53. If the iteration number threshold is reached or less than the convergence threshold, the evolution iteration operation is determined to meet the preset termination condition, the iteration is terminated, and the breeding strategy with the highest fitness score is selected from the offspring population as the optimal solution breeding strategy output.

[0052] S54. If the iteration number threshold is not reached but is not less than the convergence threshold, the evolutionary iteration operation is determined to not meet the preset termination condition. The offspring population is taken as the new generation current population, and the next evolutionary iteration operation begins.

[0053] Optionally, S6 specifically includes: decoding the optimal breeding strategy using a chromosome encoding scheme, mapping the values ​​of integer gene bits back to specific parent indices, parsing binary gene bits into hybridization methods, converting consecutive real gene bits into offspring screening thresholds, and combining them to generate and output specific breeding strategy schemes.

[0054] The beneficial effects of this invention are:

[0055] First, this invention constructs a multi-source dataset that integrates parental genotypes, phenotypes, and breeding constraints, and designs a hybrid coded chromosome containing decision variables. This provides a high-dimensional, structured data foundation for the quantitative characterization and optimization of breeding strategies, effectively solving the problems of traditional breeding decision-making relying on experience and information fragmentation.

[0056] Secondly, the core of this invention lies in the use of an improved DeepFM model for fitness evaluation. This model replaces the traditional embedding layer with a graph convolutional network, utilizing a gene association graph constructed from gene linkage disequilibrium as prior knowledge, enabling in-depth mining of complex topological relationships and higher-order interaction effects between gene loci. This improvement significantly enhances the biological rationality and accuracy of evaluating the potential of breeding strategies, overcoming the evaluation bias caused by neglecting internal gene associations in existing methods, and providing reliable navigation signals for subsequent optimization.

[0057] Furthermore, this invention employs the CMA-ES algorithm as the optimization engine, achieving efficient global search in a complex hybrid coding space through adaptive updates of the covariance matrix and evolutionary path. This mechanism can dynamically adjust the search step size and direction according to the population distribution, effectively avoiding the shortcomings of traditional genetic algorithms that are prone to getting trapped in local optima and have slow convergence speed, and significantly improving the efficiency and stability of finding the globally optimal breeding strategy.

[0058] In summary, this invention constructs a data-driven, self-evolving intelligent breeding optimization closed loop by deeply integrating the improved DeepFM model and the CMA-ES algorithm. It achieves full-link intelligence from strategy evaluation to optimization, significantly improving the scientific nature, efficiency, and success rate of breeding decisions, and providing strong technical support for solving complex breeding problems. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of a genetic breeding strategy optimization method based on evolutionary algorithms proposed in this invention;

[0061] Figure 2 This is a flowchart of the fitness evaluation process for the improved DeepFM model based on graph convolutional networks proposed in this invention.

[0062] Figure 3 This is a flowchart of the population evolution and iterative optimization process of the CMA-ES algorithm based on covariance matrix adaptation proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figures 1-3 A genetic breeding strategy optimization method based on evolutionary algorithms includes the following steps:

[0065] S1. Collect breeding multi-source data and output breeding multi-source dataset. The breeding multi-source dataset includes phenotypic indicators for defining breeding objectives, optional parental populations as decision variables, and constraints of the breeding process. At the same time, obtain the genotype and phenotypic data of the parental population.

[0066] S2. Based on the multi-source dataset of breeding, construct a mathematical model of genetic breeding, and design a chromosome coding scheme to represent the complete breeding strategy, map decision variables to genes on chromosomes, and output the chromosome coding scheme.

[0067] S3. Using a chromosome encoding scheme, an initial population containing multiple different breeding strategies is randomly initialized. An improved DeepFM model is trained using genotype and phenotypic data based on the parent population. The FM component is used to learn the low-order interaction effects between decision variables in the breeding strategy, and the DNN component is used to mine high-order nonlinear features and output the fitness score of the breeding strategy.

[0068] S4. Perform evolutionary iteration operation, input the current population into the improved DeepFM model, obtain the fitness score of each breeding strategy, and use the CMA-ES algorithm to dynamically adjust the CMA-ES algorithm by adaptively updating the covariance matrix of the multivariate normal distribution, so as to efficiently recombine and mutate the breeding strategy population represented in the form of continuous variables to generate offspring population.

[0069] S5. Determine whether the evolutionary iteration operation meets the preset termination condition. If it does, terminate the iteration and select the breeding strategy with the highest fitness score from the current offspring population as the optimal solution breeding strategy output. If it does not meet the condition, take the offspring population as the new generation current population and repeat the evolutionary iteration operation.

[0070] S6. Decode the optimal breeding strategy using the chromosome coding scheme, reverse-parse the chromosome code into a specific breeding strategy scheme containing parental pairing, hybridization method and offspring selection criteria, and output it.

[0071] This implementation significantly improves the efficiency and accuracy of breeding strategy optimization. By constructing a multi-source dataset integrating genotype, phenotype, and constraints, and designing a hybrid coding chromosome, quantitative characterization of breeding strategies is achieved. The core lies in employing an improved DeepFM model, utilizing graph convolutional networks to deeply mine complex gene interactions, thereby providing a more accurate and biologically interpretable assessment of the breeding strategy potential. Combined with the adaptive optimization capability of the CMA-ES algorithm, efficient global search in complex strategy spaces is achieved. This method constructs a data-driven, self-evolving intelligent breeding decision-making closed loop, effectively overcoming the shortcomings of traditional methods such as reliance on experience, biased evaluation, and blind optimization, significantly improving the success rate and efficiency of breeding superior varieties.

[0072] In this embodiment, S1 specifically includes:

[0073] S11. Obtain the breeding project plan, and extract from the breeding project plan the trait indicators used to define the breeding objectives, the optional parent populations as decision variables, and the constraints of the breeding process.

[0074] S12. For each parent in the selectable parent population, collect leaf tissue samples, perform genotyping and sequencing on the leaf tissue samples, and obtain genotypic data of the parent population.

[0075] S13. Record multiple key growth stages of the selectable parent population, use measuring instruments to record the values ​​of plant height, ear length, thousand-grain weight and yield-related traits of each parent, and associate the recorded values ​​with the corresponding parent identity information to generate phenotypic data of the parent population.

[0076] S14. Integrate phenotypic indicators, selectable parental populations, constraints, genotype data, and phenotypic data in the form of structured tables to construct and output a multi-source breeding dataset.

[0077] In this embodiment, S2 specifically includes:

[0078] S21. Read the trait indicators and constraints in the multi-source dataset of breeding, preset the weight coefficient for each trait indicator according to the importance of breeding, set the weighted sum of all trait indicators as the optimization objective, transform the constraints into mathematical inequalities or equations, and establish a multi-objective integer programming model with the weighted sum as the optimization objective and the transformed constraints as the boundary, as the mathematical model of genetic breeding.

[0079] S22. A complete breeding strategy is broken down into four decision variables in chronological order: paternal selection, maternal selection, hybridization combination method, and offspring selection criteria.

[0080] S23. Assign an integer gene bit to the parent selection decision variable. The value range of the integer gene bit is from 1 to N, where N is the total number of the available parent population. Each value uniquely corresponds to the index number of a parent in the available parent population. Assign another independent integer gene bit to the maternal selection decision variable. The value range and the correspondence rule between the index number are exactly the same as those for the parent selection decision variable.

[0081] S24. Assign a binary gene bit of length 1 to the decision variable of hybridization combination. When the value of the gene bit is 0, it represents a single cross; when the value of the gene bit is 1, it represents a triple cross.

[0082] S25. Assign a continuous real-numbered gene locus with a preset value range to the standard decision variable for offspring selection. The value of the continuous real-numbered gene locus is used as a cutoff threshold, and only offspring individuals whose fitness score ranking is higher than the proportion corresponding to the cutoff threshold are retained.

[0083] S26. Linearly concatenate the integer gene bits representing the paternal selection, the integer gene bits representing the maternal selection, the binary gene bits representing the hybridization combination, and the continuous real number gene bits representing the offspring selection criteria in a pre-defined order to form a chromosome vector, i.e., the chromosome encoding scheme, and output it.

[0084] In this embodiment, S3 specifically includes:

[0085] S31. Read the chromosome coding scheme, and randomly generate a preset number of chromosomes according to the gene positions and value ranges defined in the chromosome coding scheme to form an initial population containing multiple different breeding strategies.

[0086] S32. An improved DeepFM model is trained using genotype and phenotypic data from the parent population, and different fitness scores are output using breeding strategies in the initial population.

[0087] In this embodiment, S32 specifically includes:

[0088] Read the genotype data of the parent population. The genotype data is an M×N 0 / 1 matrix, where M is the number of parents and N is the total number of gene loci. Calculate the linkage disequilibrium r² value between all pairs of gene loci in the 0 / 1 matrix. That is, calculate the product of allele frequencies as the expected frequency of the haplotype. Divide the square of the deviation between the actual observed frequency and the expected frequency by the product of the expected frequencies and then normalize to obtain the r² value.

[0089] Construct an N×N LD matrix, set an LD threshold, set the positions in the LD matrix with r² values ​​greater than the LD threshold to 1, and set the other positions to 0, and generate an N×N adjacency matrix, which is the gene association graph;

[0090] In the improved DeepFM model, the original embedding layer is replaced with a graph embedding layer. The internal structure of the graph embedding layer is a two-layer graph convolutional network. The weight matrix of the graph convolutional network is initialized, and the adjacency matrix is ​​loaded as the set graph structure parameters of the graph convolutional network.

[0091] When processing the paternal decision variable in a breeding strategy, the row vectors corresponding to the values ​​of the paternal decision variables are extracted from the M×N 0 / 1 matrix to obtain a paternal genotype vector of length N, which is then input into the graph embedding layer. The first layer graph convolution operation is performed, and the adjacency matrix and the paternal genotype vector are multiplied and then multiplied by the first layer preset weight matrix to obtain the first layer graph convolution output vector.

[0092] Perform the second-layer graph convolution operation, multiply the adjacency matrix with the output vector of the first-layer graph convolution and multiply by the second-layer preset weight matrix to obtain the output vector of the second-layer graph convolution, and output it as the parent graph embedding vector of the parent decision variable. Repeat the same graph embedding operation on the parent decision variable to generate the parent graph embedding vector of the parent.

[0093] The parental graph embedding vector of the father, the parental graph embedding vector of the mother, the binary scalar representing the hybridization combination mode, and the continuous real scalar representing the offspring selection criteria are concatenated end to end to form a concatenated feature vector.

[0094] The concatenated feature vectors are input into the FM component in the improved DeepFM model. All concatenated feature vectors are combined pairwise, the cross product of each pair of concatenated feature vectors is calculated, and all cross product results are summed to generate a numerical value as a low-order interaction feature.

[0095] The concatenated feature vector is input into the DNN component of the improved DeepFM model. The DNN component consists of three fully connected layers, each followed by a ReLU activation function. Different preset weights are applied sequentially, and the weights and biases are summed and ReLU activation is performed to output high-order nonlinear features.

[0096] The low-order interactive features and high-order nonlinear features are concatenated according to their corresponding features, and then weighted and summed with the preset output layer weights. Finally, the preset output layer bias is added to obtain the fitness score of the current breeding strategy.

[0097] This implementation method achieves accurate assessment of breeding strategy fitness by introducing an improved DeepFM model. First, linkage disequilibrium is calculated based on parental genotype data, and a gene association graph is constructed as prior knowledge, replacing the original embedding layer with a two-layer graph convolutional network. By constructing the gene association graph and utilizing the graph convolutional network, complex gene linkage and interaction effects are successfully quantified into learnable features, completely overcoming the fundamental deficiency of traditional models that treat genes as independent variables. When processing breeding strategies, this method extracts parental genotype vectors and inputs them into the graph embedding layer. Graph convolution operations are used to capture the complex topological relationships between gene loci, generating biologically meaningful parental graph embedding vectors. Subsequently, the graph embedding vectors are concatenated with other decision variables and input into the FM component to learn low-order interaction effects, while the DNN component mines high-order nonlinear features. This approach can deeply mine gene linkage and interaction effects ignored by traditional models, making fitness scores closer to true genetic potential, providing reliable and interpretable navigation signals for subsequent evolutionary optimization, and significantly improving the scientific rigor and accuracy of breeding strategy assessment.

[0098] In this embodiment, S4 specifically includes:

[0099] S41. Input the chromosome vector of each breeding strategy in the current population into the improved DeepFM model to obtain the fitness score of each breeding strategy, and sort the breeding strategies in the current population from high to low.

[0100] S42. Initialize the parameters of the CMA-ES algorithm. Set the arithmetic mean of the chromosome vectors of the top 10 breeding strategies with the highest fitness scores in the current population as the mean vector, set the step size parameter, initialize the identity matrix as the covariance matrix, and initialize a zero vector with the same dimension as the chromosome vector as the evolutionary path.

[0101] S43. From a multivariate normal distribution with the mean vector as the distribution center and the covariance matrix and step size parameter as the common determination, randomly sample to generate a preset number of new candidate chromosome vectors to obtain a preset number of candidate offspring breeding strategies and input them into the improved DeepFM model to obtain the fitness score of each candidate offspring breeding strategy and sort them from high to low, and select the top preset number of candidate offspring breeding strategies with the highest fitness scores.

[0102] S44. Calculate the new mean vector by multiplying the chromosome vectors of the selected first preset candidate offspring breeding strategies by preset weight coefficients that decrease in rank, and summing the weighted chromosome vectors element by element to obtain the new mean vector.

[0103] S45. Update the evolutionary path. Calculate the difference vector between the new mean vector of the current generation and the old mean vector of the previous generation. Divide the difference vector by the current step size parameter to obtain the normalized movement vector. Multiply the evolutionary path vector of the previous generation by the preset decay coefficient and add it to the result of multiplying the normalized movement vector by the preset amplification coefficient to generate the new generation's evolutionary path.

[0104] S46. Adaptively update the covariance matrix by performing an outer product operation on the new generation evolution path vector and its own transpose vector to generate a matrix with rank 1. For the first preset number of candidate offspring breeding strategies, calculate the difference between each chromosome vector and the old mean vector, divide each difference vector by the current step size parameter to normalize it, and perform an outer product operation on it with its own transpose vector to generate a preset number of matrices with rank 1.

[0105] S47. The matrix with a rank of 1 is weighted and summed according to the preset weights to generate a weighted matrix. The covariance matrix of the previous generation is multiplied by the preset retention coefficient, and then the result of multiplying the matrix with a rank of 1 by the first learning rate and the result of multiplying the weighted matrix by the second learning rate are added to obtain the updated covariance matrix.

[0106] S48. Use the updated mean vector and covariance matrix as parameters for the new generation CMA-ES algorithm, and use the preset number of candidate offspring breeding strategies generated by the new parameters as the offspring population output.

[0107] This implementation method achieves efficient adaptive evolution of breeding strategy populations by introducing the CMA-ES algorithm. First, the population is sorted according to fitness scores, and algorithm parameters are initialized, constructing a multivariate normal distribution using the mean and covariance of superior strategies. A new generation of candidate strategies is generated by sampling from this distribution, and the search direction and step size are dynamically adjusted using an adaptive update mechanism of the evolutionary path and covariance matrix. This approach can intelligently learn the shape and direction of the optimal search region based on the population's evolutionary history, avoiding the blindness of fixed operators in traditional genetic algorithms. This enables the algorithm to quickly approach the global optimum in complex strategy spaces containing both continuous and discrete variables, effectively avoiding getting trapped in local optima, significantly improving the evolutionary convergence speed and the quality of the final solution, and ensuring the efficiency and robustness of breeding strategy optimization.

[0108] In this embodiment, S5 specifically includes:

[0109] S51. Read the preset iteration number threshold, and compare the actual number of executions of the current evolution iteration with the iteration number threshold to determine whether the current actual number of executions has reached the iteration number threshold.

[0110] S52. Read the fitness scores of all breeding strategies in the offspring population, find the highest fitness score, and record the change value of the highest fitness score in the most recent consecutive preset generations, and determine whether the change value is less than the preset convergence threshold.

[0111] S53. If the iteration number threshold is reached or less than the convergence threshold, the evolution iteration operation is determined to meet the preset termination condition, the iteration is terminated, and the breeding strategy with the highest fitness score is selected from the offspring population as the optimal solution breeding strategy output.

[0112] S54. If the iteration number threshold is not reached but is not less than the convergence threshold, the evolutionary iteration operation is determined to not meet the preset termination condition. The offspring population is taken as the new generation current population, and the next evolutionary iteration operation begins.

[0113] In this embodiment, S6 specifically includes: decoding the optimal breeding strategy using a chromosome encoding scheme, mapping the values ​​of integer gene bits back to specific parent indices, parsing binary gene bits into hybridization methods, converting consecutive real gene bits into offspring screening thresholds, and combining them to generate and output specific breeding strategy schemes.

[0114] Example 1:

[0115] To verify the feasibility of this invention in the field of complex crop breeding optimization, it was deployed in the intelligent maize breeding decision-making platform of a top agricultural research institute in China. This platform manages the institute's core breeding projects in multiple ecological zones across the country, possessing over 5,000 meticulously identified maize inbred line germplasm resources. Genotypic data covers over 600,000 SNP markers, and phenotypic data encompasses over 20 key agronomic traits, including plant height, ear height, ear length, ear diameter, 1000-kernel weight, number of kernels per row, growth period, and disease resistance. Traditional breeding strategies rely heavily on the experience and judgment of breeding experts, combined with simple combining ability analysis and phenotypic selection. This results in a long breeding cycle, low decision-making efficiency, and difficulty in systematically exploring and utilizing complex gene interaction effects, leading to slow progress in the selection of superior varieties and an inability to cope with increasing market demands and environmental changes.

[0116] In practical deployment, the method of this invention first integrates breeding project data from the Academy of Sciences over the years, genotype and phenotypic data from the germplasm resource bank, and environmental constraints from different ecological zones to construct a structured multi-source breeding dataset. Next, based on the breeding objectives, this invention designs a hybrid coding chromosome that includes paternal selection, maternal selection, hybridization method, and offspring selection criteria, quantifying a complete breeding strategy into a vector that the algorithm can process. Subsequently, the improved DeepFM model proposed in this invention is used to evaluate the fitness of the breeding strategy. This model constructs a gene association graph by calculating linkage disequilibrium between gene loci and uses a two-layer graph convolutional network to deeply learn the complex genetic relationships between parents, effectively capturing higher-order interaction effects ignored by traditional models. Finally, the CMA-ES algorithm is used as the optimization engine, dynamically adjusting the search direction and step size based on the fitness score, efficiently iterating and optimizing within the vast strategy space, and ultimately outputting the optimal combination of breeding strategies. Table 1 below shows the comparison data of the method of this invention and the traditional expert experience method on key performance indicators in a two-year breeding project for a high-yield, disease-resistant maize variety:

[0117] Table 1. Performance Comparison Data of the Invention and Traditional Methods in High-Yield and Disease-Resistant Maize Breeding Projects

[0118]

[0119] Based on the comparative data shown in Table 1, it can be seen that the genetic breeding strategy optimization method based on evolutionary algorithm proposed in this invention has significant performance advantages over traditional expert experience methods in the breeding of new maize varieties, especially in key indicators such as screening efficiency, variety performance, resource utilization and cost-effectiveness.

[0120] In terms of hybrid combination screening efficiency, this invention significantly increases the success rate of screening superior combinations from 1.31% to 6.74%, achieving a leap of more than four times. This is thanks to the improved DeepFM model, which can accurately predict the genetic potential of hybrid combinations, thereby freeing breeding resources from thousands of ineffective combinations and focusing on a few high-value candidates, completely changing the traditional inefficient broad-based approach.

[0121] In terms of the final performance of the varieties, the advantages of this invention are particularly prominent. Through systematic optimization, the average yield of the tested varieties increased by 14.2%, and the proportion of ultra-high-yield combinations surged from 8.3% to 24.1%. At the same time, the comprehensive disease resistance index also increased by 43.5%, indicating that this invention can synergistically optimize multiple complex traits such as yield and resistance, overcoming the problem of neglecting one aspect for another in traditional breeding, and cultivating varieties with a greater comprehensive competitive advantage.

[0122] In terms of resource utilization efficiency and cost control, the economic benefits of this invention are significant. The output per unit area of ​​superior combinations increased by 322.4%, meaning that the utilization efficiency of scarce resources such as land has increased several times. At the same time, the total project cycle was shortened by 37.5%, and the cultivation cost of a single superior combination was reduced by 75.9%, which greatly accelerated the breeding process, reduced labor and time costs, and brought huge economic returns to breeding institutions.

[0123] Overall, this invention, through the deep integration of data modeling and intelligent optimization algorithms, constructs an efficient, accurate, and economical breeding decision-making closed loop. It not only significantly improves the breeding efficiency and overall performance of superior varieties, but also fundamentally transforms the traditional breeding decision-making paradigm, providing strong technical support for addressing future food security challenges.

[0124] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A genetic breeding strategy optimization method based on evolutionary algorithms, characterized in that, Includes the following steps: S1. Collect breeding multi-source data, output breeding multi-source dataset, and simultaneously obtain genotype and phenotypic data of parental populations; S2. Based on the multi-source dataset of breeding, construct a mathematical model of genetic breeding, design a chromosome coding scheme, map decision variables to genes on chromosomes, and output the chromosome coding scheme. S3. Using the chromosome coding scheme, the initial population is randomly initialized, and the improved DeepFM model is trained using genotype and phenotypic data based on the parent population to output the fitness score of the breeding strategy. S4. Perform evolutionary iteration operation, input the current population into the improved DeepFM model, obtain the fitness score of each breeding strategy, and use the CMA-ES algorithm to efficiently recombine and mutate the breeding strategy population represented in continuous variable form to generate offspring population. S5. Determine whether the evolutionary iteration operation meets the preset termination condition. If it does, terminate the iteration and select the breeding strategy with the highest fitness score from the current offspring population as the optimal solution breeding strategy output. If it does not meet the condition, take the offspring population as the new generation current population and repeat the evolutionary iteration operation. S6. Use the chromosome encoding scheme to decode the optimal breeding strategy, reverse analyze the chromosome encoding into a specific breeding strategy scheme and output it.

2. The genetic breeding strategy optimization method based on evolutionary algorithm according to claim 1, characterized in that, S1 specifically includes: S11. Obtain the breeding project plan, and extract from the breeding project plan the trait indicators used to define the breeding objectives, the optional parent populations as decision variables, and the constraints of the breeding process. S12. For each parent in the selectable parent population, collect leaf tissue samples, perform genotyping and sequencing on the leaf tissue samples, and obtain genotypic data of the parent population. S13. Record multiple key growth stages of the selectable parent population, use measuring instruments to record the values ​​of plant height, ear length, thousand-grain weight and yield-related traits of each parent, and associate the recorded values ​​with the corresponding parent identity information to generate phenotypic data of the parent population. S14. Integrate phenotypic indicators, selectable parental populations, constraints, genotype data, and phenotypic data in the form of structured tables to construct and output a multi-source breeding dataset.

3. The genetic breeding strategy optimization method based on evolutionary algorithm according to claim 1, characterized in that, S2 specifically includes: S21. Read the trait indicators and constraints in the multi-source dataset of breeding, preset the weight coefficient for each trait indicator according to the importance of breeding, set the weighted sum of all trait indicators as the optimization objective, transform the constraints into mathematical inequalities or equations, and establish a multi-objective integer programming model with the weighted sum as the optimization objective and the transformed constraints as the boundary, as the mathematical model of genetic breeding. S22. A complete breeding strategy is broken down into four decision variables in chronological order: paternal selection, maternal selection, hybridization combination method, and offspring selection criteria. S23. Assign an integer gene bit to the parent selection decision variable. The value range of the integer gene bit is from 1 to N, where N is the total number of the available parent population. Each value uniquely corresponds to the index number of a parent in the available parent population. Assign another independent integer gene bit to the maternal selection decision variable. The value range and the correspondence rule between the index number are exactly the same as those for the parent selection decision variable. S24. Assign a binary gene bit of length 1 to the decision variable of hybridization combination. When the value of the gene bit is 0, it represents a single cross; when the value of the gene bit is 1, it represents a triple cross. S25. Assign a continuous real-numbered gene locus with a preset value range to the standard decision variable for offspring selection. The value of the continuous real-numbered gene locus is used as a cutoff threshold, and only offspring individuals whose fitness score ranking is higher than the proportion corresponding to the cutoff threshold are retained. S26. Linearly concatenate the integer gene bits representing the paternal selection, the integer gene bits representing the maternal selection, the binary gene bits representing the hybridization combination, and the continuous real number gene bits representing the offspring selection criteria in a pre-defined order to form a chromosome vector, i.e., the chromosome encoding scheme, and output it.

4. The genetic breeding strategy optimization method based on evolutionary algorithm according to claim 1, characterized in that, S3 specifically includes: S31. Read the chromosome coding scheme, and randomly generate a preset number of chromosomes according to the gene positions and value ranges defined in the chromosome coding scheme to form an initial population containing multiple different breeding strategies. S32. An improved DeepFM model is trained using genotype and phenotypic data from the parent population, and different fitness scores are output using breeding strategies in the initial population.

5. The genetic breeding strategy optimization method based on evolutionary algorithm according to claim 4, characterized in that, Specifically, S32 includes: Read the genotype data of the parent population. The genotype data is an M×N 0 / 1 matrix, where M is the number of parents and N is the total number of gene loci. Calculate the linkage disequilibrium r² value between all pairs of gene loci in the 0 / 1 matrix. That is, calculate the product of allele frequencies as the expected frequency of the haplotype. Divide the square of the deviation between the actual observed frequency and the expected frequency by the product of the expected frequencies and then normalize to obtain the r² value. Construct an N×N LD matrix, set an LD threshold, set the positions in the LD matrix with r² values ​​greater than the LD threshold to 1, and set the other positions to 0, and generate an N×N adjacency matrix, which is the gene association graph; In the improved DeepFM model, the original embedding layer is replaced with a graph embedding layer. The internal structure of the graph embedding layer is a two-layer graph convolutional network. The weight matrix of the graph convolutional network is initialized, and the adjacency matrix is ​​loaded as the set graph structure parameters of the graph convolutional network. When processing the paternal decision variable in a breeding strategy, the row vectors corresponding to the values ​​of the paternal decision variables are extracted from the M×N 0 / 1 matrix to obtain a paternal genotype vector of length N, which is then input into the graph embedding layer. The first layer graph convolution operation is performed, and the adjacency matrix and the paternal genotype vector are multiplied and then multiplied by the first layer preset weight matrix to obtain the first layer graph convolution output vector. Perform the second-layer graph convolution operation, multiply the adjacency matrix with the output vector of the first-layer graph convolution and multiply by the second-layer preset weight matrix to obtain the output vector of the second-layer graph convolution, and output it as the parent graph embedding vector of the parent decision variable. Repeat the same graph embedding operation on the parent decision variable to generate the parent graph embedding vector of the parent. The parental graph embedding vector of the father, the parental graph embedding vector of the mother, the binary scalar representing the hybridization combination mode, and the continuous real scalar representing the offspring selection criteria are concatenated end to end to form a concatenated feature vector. The concatenated feature vectors are input into the FM component in the improved DeepFM model. All concatenated feature vectors are combined pairwise, the cross product of each pair of concatenated feature vectors is calculated, and all cross product results are summed to generate a numerical value as a low-order interaction feature. The concatenated feature vector is input into the DNN component of the improved DeepFM model. The DNN component consists of three fully connected layers, each followed by a ReLU activation function. Different preset weights are applied sequentially, and the weights and biases are summed and ReLU activation is performed to output high-order nonlinear features. The low-order interactive features and high-order nonlinear features are concatenated according to their corresponding features, and then weighted and summed with the preset output layer weights. Finally, the preset output layer bias is added to obtain the fitness score of the current breeding strategy.

6. The genetic breeding strategy optimization method based on evolutionary algorithm according to claim 1, characterized in that, S4 specifically includes: S41. Input the chromosome vector of each breeding strategy in the current population into the improved DeepFM model to obtain the fitness score of each breeding strategy, and sort the breeding strategies in the current population from high to low. S42. Initialize the parameters of the CMA-ES algorithm. Set the arithmetic mean of the chromosome vectors of the top 10 breeding strategies with the highest fitness scores in the current population as the mean vector, set the step size parameter, initialize the identity matrix as the covariance matrix, and initialize a zero vector with the same dimension as the chromosome vector as the evolutionary path. S43. From a multivariate normal distribution with the mean vector as the distribution center and the covariance matrix and step size parameter as the common determination, randomly sample to generate a preset number of new candidate chromosome vectors to obtain a preset number of candidate offspring breeding strategies and input them into the improved DeepFM model to obtain the fitness score of each candidate offspring breeding strategy and sort them from high to low, and select the top preset number of candidate offspring breeding strategies with the highest fitness scores. S44. Calculate the new mean vector by multiplying the chromosome vectors of the selected first preset candidate offspring breeding strategies by preset weight coefficients that decrease in rank, and summing the weighted chromosome vectors element by element to obtain the new mean vector. S45. Update the evolutionary path. Calculate the difference vector between the new mean vector of the current generation and the old mean vector of the previous generation. Divide the difference vector by the current step size parameter to obtain the normalized movement vector. Multiply the evolutionary path vector of the previous generation by the preset decay coefficient and add it to the result of multiplying the normalized movement vector by the preset amplification coefficient to generate the new generation's evolutionary path. S46. Adaptively update the covariance matrix by performing an outer product operation on the new generation evolution path vector and its own transpose vector to generate a matrix with rank 1. For the first preset number of candidate offspring breeding strategies, calculate the difference between each chromosome vector and the old mean vector, divide each difference vector by the current step size parameter to normalize it, and perform an outer product operation on it with its own transpose vector to generate a preset number of matrices with rank 1. S47. The matrix with a rank of 1 is weighted and summed according to the preset weights to generate a weighted matrix. The covariance matrix of the previous generation is multiplied by the preset retention coefficient, and then the result of multiplying the matrix with a rank of 1 by the first learning rate and the result of multiplying the weighted matrix by the second learning rate are added to obtain the updated covariance matrix. S48. Use the updated mean vector and covariance matrix as parameters for the new generation CMA-ES algorithm, and use the preset number of candidate offspring breeding strategies generated by the new parameters as the offspring population output.

7. The genetic breeding strategy optimization method based on evolutionary algorithm according to claim 1, characterized in that, S5 specifically includes: S51. Read the preset iteration number threshold, and compare the actual number of executions of the current evolution iteration with the iteration number threshold to determine whether the current actual number of executions has reached the iteration number threshold. S52. Read the fitness scores of all breeding strategies in the offspring population, find the highest fitness score, and record the change value of the highest fitness score in the most recent consecutive preset generations, and determine whether the change value is less than the preset convergence threshold. S53. If the iteration number threshold is reached or less than the convergence threshold, the evolution iteration operation is determined to meet the preset termination condition, the iteration is terminated, and the breeding strategy with the highest fitness score is selected from the offspring population as the optimal solution breeding strategy output. S54. If the iteration number threshold is not reached but is not less than the convergence threshold, the evolutionary iteration operation is determined to not meet the preset termination condition. The offspring population is taken as the new generation current population, and the next evolutionary iteration operation begins.

8. The genetic breeding strategy optimization method based on evolutionary algorithm according to claim 1, characterized in that, S6 specifically includes: decoding the optimal breeding strategy using a chromosome encoding scheme, mapping the values ​​of integer gene bits back to specific parent indices, parsing binary gene bits into hybridization methods, converting consecutive real gene bits into offspring screening thresholds, and combining them to generate and output specific breeding strategy schemes.

Images