Multi-scale multi-objective evolutionary neural architecture search algorithm considering model generalization

Abstract

The invention discloses a multi-scale multi-objective evolutionary neural architecture search algorithm considering model generalization, and belongs to the crossing field of software defined network (SDN) performance prediction and neural architecture search (NAS) technologies. The core of the method is to construct a multi-scale search space covering'unit block combination-network type-parameter configuration 'by randomly segmenting a data set and calculating a space-time two-dimensional similarity weight aiming at the problems of limited feature extraction capability, insufficient generalization and low architecture design efficiency of an existing SDN performance prediction model. An optimal architecture is explored by adopting an evolutionary algorithm of a multi-scale crossover mutation operator, and candidate architecture output is fused in combination with an attention integration strategy based on distance (the flow is as shown in Figure 1). According to the algorithm, automatic optimization of a neural network architecture is realized, global exploration and local optimization are effectively balanced, the prediction precision and generalization stability of the model on different distribution data sets of the same topology are remarkably improved, and the algorithm is suitable for precise prediction tasks of performance indexes such as SDN network delay and the like.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of software-defined networking (SDN) performance prediction and neural architecture search (NAS) technology. Specifically, it involves an automatic neural network architecture generation algorithm based on multi-scale search and multi-objective optimization, which is suitable for accurate prediction of performance indicators such as SDN network latency. Background Technology

[0002] With the development of computer network technology, SDN, with its architectural advantage of separating the control plane and data plane, has become one of the core technologies for network optimization. Network modeling, as a key tool for SDN performance optimization, establishes a mapping relationship between traffic matrices and network performance, providing a basis for network configuration adjustments.

[0003] Traditional network modeling methods, such as queuing theory models and packet-level simulators, suffer from unrealistic assumptions or excessively high computational costs. In recent years, deep learning technology has been widely used in SDN performance prediction due to its data-driven nature, but existing solutions still have significant drawbacks: First, they rely on a single type of neural network (such as CNN or LSTM), making it difficult to capture the complex spatiotemporal relationships in traffic data simultaneously; second, the models have insufficient generalization ability, and prediction performance drops significantly when faced with datasets of the same topology but different distributions; third, the hyperparameters and architecture design of neural networks rely on manual trial and error, which is inefficient and makes it difficult to achieve optimal performance.

[0004] Neural Architecture Search (NAS) technology offers the possibility of automated network architecture design, but existing NAS methods are mainly applied to fields such as image classification and have not yet been customized for the task characteristics of SDN performance prediction. Furthermore, they lack exploration and generalization optimization mechanisms for combinations of multiple neural networks. Therefore, there is an urgent need for a method that can automatically search for neural network architectures that are suitable for SDN performance prediction tasks and possess both high accuracy and strong generalization capabilities. Summary of the Invention

[0005] To address the limitations of existing SDN performance prediction models, such as limited feature extraction capabilities, insufficient generalization, and low architecture design efficiency, this invention provides a multi-scale, multi-objective evolutionary neural architecture search algorithm. This algorithm automates the optimization of neural network architectures, improving prediction accuracy and stability under different data distribution scenarios. By covering all dimensions of neural network parameters in a multi-scale search space, it efficiently explores the optimal architecture using a multi-objective evolutionary algorithm and enhances model generalization through an attention mechanism. The specific technical solution is as follows:

[0006] S1 Dataset Segmentation and Weight Calculation

[0007] The core objective of this module is to achieve fast and unbiased processing of the dataset through random partitioning, while quantifying the similarity between subsets to determine subsequent ensemble weights. The specific process is as follows:

[0008] S1.1 Dataset Splitting

[0009] After shuffling the data, the sizes are adjusted according to a preset ratio. The original dataset Divided into Subdatasets ,in During splitting, ensure that the sample size deviation of each subset does not exceed 5% of the total sample size of the original dataset;

[0010] S1.2 Similarity Weight Calculation

[0011] Based on the spatiotemporal dual-dimensional features of traffic data, two complementary similarity measurement methods are used to quantify the distributional differences among randomly segmented subsets, and these differences are transformed into weights for subsequent model integration. and Specifically, the following two similarity measurement methods are used:

[0012] (1) Spatial Dimension Similarity: Cosine similarity is used for calculation, as shown in the formula below. This measures the "spatial structural consistency" of the flow matrices of two subsets.

[0013]

[0014] (2) Temporal similarity: The Pearson correlation coefficient was used for calculation, as shown in the formula below. This coefficient measures the consistency of the temporal fluctuation trends of the flow sequences in two subsets of data.

[0015]

[0016] The corresponding similarity is calculated using the above formula. and Then, the similarity weights are normalized to obtain the spatial weights. and time weight :

[0017]

[0018]

[0019] S2 Search Space Initialization

[0020] By constructing a multi-scale search space (MS³) covering all dimensions of neural networks, we can achieve efficient exploration of "unit block combination - network type - parameter configuration".

[0021] S2.1 Define the individual coding format

[0022] Each individual in the population is set as the neural network is decomposed into... Each functionally independent unit block, in which Each unit block serves as a basic functional unit of the architecture, and its corresponding network type and parameters can be independently configured, denoted as . ,in Representing the The network type contained in each unit block Network layers And detailed network configuration; simultaneously, one topology control unit (T) is set up, using cell blocks to define the connection relationships between cell blocks, and conceptualizing the network composed of cell blocks as a graph. Each cell block serves as a node in the graph. ;

[0023] S2.2 Define the multi-scale coding rules for the search space

[0024] Based on individual coding methods, coding is performed at three scales: macroscopic, mesoscopic, and microscopic. The macroscopic scale encodes the overall connectivity of the network; the mesoscopic scale encodes the network type and number of layers of network units; and the microscopic scale encodes the specific configuration of the units. Specifically, the coding rules are as follows:

[0025] (1) Macro-scale coding: using a length of The binary string represents the connection relationship between unit blocks at the macro scale, directly reflecting the information transmission path between unit blocks through binary bits. The nodes representing unit blocks are... and The connection between them can be used This means that if the node and There is a connection. =1, otherwise 0. Each unit block can selectively receive the output of the preceding unit block. Therefore, the connection status of a neural network consisting of n unit blocks can be represented by a binary string. express;

[0026] (2) Mesoscale encoding: Defining the representable unit blocks at the mesoscale. Neural Network Types and network layers Each unit block can be and Choose a value from the range as your configuration. and ;

[0027] (3) Microscale coding: At the microscale, according to the network type defined at the mesoscale, the parameter coding DP specific to that network type is configured;

[0028] S3 Population Evolution Iteration

[0029] S3.1 Initialize the population

[0030] Initialize population Set the population size to NP, and generate population individuals based on a multi-scale search space.

[0031] S3.2 Fitness Assessment

[0032] For each individual (candidate architecture) in each population, respectively in The individual is trained on several subsets of datasets, and the mean squared error between the predicted and true values ​​is calculated on the same subset of validation sets. The mean squared error is used as the fitness vector of the individual to evaluate its performance on different randomly split subsets of datasets.

[0033] S3.2 Population Crossover

[0034] For population The process involves performing a crossover operation based on a multi-scale search space design to generate corresponding offspring individuals, resulting in an evolved offspring population. The multi-scale crossover operator is a scale-aware crossover operator. By quantifying the differences between parent individuals at macro, meso, and micro scales, this operator selects the scale with the greatest difference for the crossover operation, thus ensuring more efficient exploration of the search space during evolution. Specifically, the crossover operator of the multi-scale genetic operator is as follows:

[0035] S3.2.1 Calculating Scale Differences. First, for the selected parent individuals, the following indicators are used to calculate differences across three scales:

[0036] (1) Macro-scale difference: Macro-scale difference is measured by the Hamming distance between the binary topological strings of two parent nodes. The Hamming distance is calculated by the number of bits that differ between corresponding positions in the two binary strings, directly reflecting the difference in the unit block connection pattern.

[0037]

[0038] (2) Mesoscale differences: Mesoscale differences are assessed by comparing parameter differences between corresponding cell blocks from two parents. For each cell block at the same location, if or If any parameter differs, the difference counter increments by 1. If both parameters are the same, the counter remains unchanged. After counting the differences across all corresponding cell blocks, the total count is normalized to obtain the mesoscale difference value:

[0039]

[0040] (3) Microscale differences: Microscale differences are identified by comparing two parents. and The corresponding unit blocks have the same The calculation is based on the differences in the module parameter sets. For modules with the same... The module we extract and The parameter set is defined. For each parameter in the set, if a difference exists, the difference counter is incremented by 1. If all parameters are the same, the counter remains unchanged. The total count is then normalized to obtain the microscale difference values:

[0041]

[0042] in It represents the number of network unit blocks of the same type. Number of parameter types;

[0043] S3.2.2 Multi-scale crossover operator

[0044] After calculating the differences across the three scales, the scale with the largest difference is selected as the crossover target. Then, the crossover operation is performed using the crossover operator corresponding to the selected scale. The specific operators for each scale are described below:

[0045] (1) Macroscale crossover operator: The crossover operation at the macro level uses binary topological strings. For objects. This operation involves two core steps: First, from the binary topology string of the parent individual... A random position is selected as the intersection point; then, the parent individual swaps the segments after the intersection point in its topological string.

[0046] (2) Mesoscale Crossover Operator: Mesoscale crossover enables unit block-specific swapping. First, a unit block position is randomly selected from the parent individuals, and this position remains consistent between the parent individuals to ensure the alignment and swapping of corresponding unit blocks. Second, the parent individual swaps the entire unit block at the selected position;

[0047] (3) Microscale crossover operator: The microscale crossover operator realizes the internal parameter exchange of specific unit blocks, preserving network type consistency while optimizing parameters. First, it identifies the common parameters between two parents. The unit blocks are then selected, and a parameter category is randomly chosen from these matching unit blocks. Next, the selected parameter values ​​are swapped between the matching unit blocks of the two parents.

[0048] S3.3 Population Variation

[0049] After crossover occurs in the population, the mutation process begins. For the population... The mutation operation, based on a multi-scale search space design, is performed to generate corresponding offspring individuals, resulting in an evolved offspring population. The mutation operator is a multi-scale random mutation operator. During the mutation process, this operator randomly selects a scale in the search space to perform the corresponding mutation. Specifically, the specific mutation mechanism for each scale is as follows:

[0050] (1) Macroscale variation: The macroscale method takes the binary topological string of an individual as the research object, first randomly selects mutation sites, and then performs bit flipping operation;

[0051] (2) Mesoscale variation: The mesoscale method takes the unit block characteristics of an individual as its object. First, it randomly selects a unit block, and then reselects the unit block containing... or And then according to the new and To redefine the specific parameters at the microscopic level;

[0052] (3) Microscale variation: The mesoscale method takes the specific parameters of the unit block of an individual as the object. First, a unit block is randomly selected, and then the variation is determined based on the unit block's... and Redetermine the specific parameters DP at the micro level;

[0053] S3.3 Evaluate the offspring population

[0054] For each individual in the evolved offspring population, respectively in The individual is trained on several subsets of datasets and the mean squared error of prediction is calculated on the same subset of validation sets. The mean squared error value is used as the fitness vector of the individual to evaluate its performance on different randomly split subsets of datasets and select the best individuals to enter the next evolution.

[0055] S3.4 Termination Judgment

[0056] Determine if a single evolutionary generation has been reached. If the target is reached, evolution stops, and the final population is non-dominated sorted to extract the Pareto optimal solution for the first frontier. From this solution set, the individual with the smallest mean squared error in each subset is selected, for a total of [number missing]. Each candidate architecture is selected to ensure optimal performance on its corresponding randomly partitioned subset of the dataset.

[0057] S4 Distance-Based Attention Integration Strategy

[0058] The core objective of this module is to address the problem that "a single architecture struggles to adapt to datasets with different distributions within the same topology," by fusing data through an attention mechanism. The output of each candidate architecture is used to generate a final model that combines high accuracy and strong generalization.

[0059] S4.1 Mixed Weight Calculation

[0060] The spatial and temporal weights generated by S1 are linearly weighted and fused to obtain the hybrid weights:

[0061]

[0062] in These are the weighting coefficients for spatial weights. The weighting coefficient for time weighting;

[0063] S4.2 Attention-Weighted Integration

[0064] For each input sample (traffic matrix) in the global test set, input respectively 1 candidate architecture, obtained Predicted delay value The final predicted value is calculated using an attention mechanism. :

[0065]

[0066] The present invention has the following advantages and effects compared with the prior art:

[0067] (1) This invention proposes a multi-scale, multi-objective evolutionary neural architecture search algorithm, MS-MO-ENAS, which uses neural architecture search (NAS) technology to automatically search for the optimal network structure among various neural network combinations, thus solving the challenges in SDN performance prediction tasks. Based on the NSGA-II multi-objective optimization algorithm, it takes "minimizing the prediction error of multiple subset datasets" as the objective function, automatically exploring the optimal structure among neural network combinations without the need for manual intervention in architecture design and parameter adjustment.

[0068] (2) This invention proposes a multi-scale search space and a corresponding multi-scale genetic strategy for the search process of neural network architectures. The multi-scale search space covers possible combinations of different types of neural networks, while the corresponding multi-scale genetic strategy enables the MS-MO-ENAS algorithm to comprehensively explore neural network combination architectures from multiple perspectives, including macroscopic, mesoscopic, and microscopic.

[0069] (3) This invention proposes a distance-based attention generalization strategy to enhance the model's generalization ability on datasets with different distributions, ensuring that the model performs well on datasets with the same topology but different distributions. A two-dimensional similarity metric is used to capture the spatial structural differences of the traffic matrix of the subsets through cosine similarity and to capture the temporal fluctuation trend differences of the traffic sequence through Pearson correlation coefficient. An adaptive weight generation is used, based on the normalized similarity results, and then a mixed weight is generated by linear weighting, giving higher weights to the architectures corresponding to subsets with large distribution differences. Attached Figure Description

[0070] Figure 1 is a flowchart of the overall process of the multi-scale, multi-objective evolutionary neural architecture search algorithm.

[0071] Figure 2 is a schematic diagram of the multi-scale search space.

[0072] Figure 3 is a schematic diagram of a multi-scale mesoscopic crossover strategy.

[0073] Figure 4 is a schematic diagram of the multi-scale dimension crossover strategy.

[0074] Figure 5 is a schematic diagram of multi-scale mesoscopic variation strategies.

[0075] Figure 6 is a schematic diagram of a multi-scale micro-variation strategy. Detailed Implementation

[0076] Figure 1 This illustrates the framework and flow of the algorithm from dataset partitioning to final output. The specific implementation of this invention will be described in detail below with reference to the accompanying drawings and practical application scenarios. Other drawings are intended to aid in understanding the core logic and execution flow of the technical solution.

[0077] S1 Dataset Segmentation and Weight Calculation

[0078] The core objective of this module is to achieve fast and unbiased processing of the dataset through random partitioning, while quantifying the similarity between subsets to determine subsequent ensemble weights. The specific process is as follows:

[0079] S1.1 Dataset Splitting

[0080] After shuffling the data, the sizes are adjusted according to a preset ratio. The original dataset Divided into 4 subsets During splitting, ensure that the sample size deviation of each subset does not exceed 5% of the total sample size of the original dataset;

[0081] S1.2 Similarity Weight Calculation

[0082] Based on the spatiotemporal dual-dimensional features of traffic data, two complementary similarity measurement methods are used to quantify the distributional differences among randomly segmented subsets, and these differences are transformed into weights for subsequent model integration. and Specifically, the following two similarity measurement methods are used:

[0083] (1) Spatial Dimension Similarity: Cosine similarity is used for calculation, as shown in the formula below. This measures the "spatial structural consistency" of the flow matrices of two subsets.

[0084]

[0085] (2) Temporal similarity: The Pearson correlation coefficient was used for calculation, as shown in the formula below. This coefficient measures the consistency of the temporal fluctuation trends of the flow sequences in two subsets of data.

[0086]

[0087] The corresponding similarity is calculated using the above formula. and Then, the similarity weights are normalized to obtain the spatial weights. and time weight :

[0088]

[0089]

[0090] S2 Search Space Initialization

[0091] By constructing a multi-scale search space (MS³) covering all dimensions of neural networks, efficient exploration of "unit block combination - network type - parameter configuration" can be achieved;

[0092] S2.1 Define the individual coding format

[0093] Each individual in the population is set as the neural network is decomposed into... Each functionally independent unit block, in which Each unit block serves as a basic functional unit of the architecture, and its corresponding network type and parameters can be independently configured, denoted as . ,in Representing the The network type contained in each unit block Network layers And detailed network configuration; simultaneously, one topology control unit (T) is set up, using cell blocks to define the connection relationships between cell blocks, and conceptualizing the network composed of cell blocks as a graph. Each cell block serves as a node in the graph. ;

[0094] S2.2 Define the multi-scale coding rules for the search space

[0095] Based on individual coding methods, coding is performed at three scales: macroscopic, mesoscopic, and microscopic. The macroscopic scale encodes the overall connectivity of the network; the mesoscopic scale encodes the network type and number of layers of network units; and the microscopic scale encodes the specific configuration of the units. Specifically, the coding rules are as follows:

[0096] (1) Macro-scale coding: using a length of The binary string represents the connection relationship between unit blocks at the macro scale, directly reflecting the information transmission path between unit blocks through binary bits. The nodes representing unit blocks are... and The connection between them can be used This means that if the node and There is a connection. =1, otherwise 0. Each unit block can selectively receive the output of the preceding unit block. Therefore, the connection status of a neural network consisting of 4 unit blocks can be represented by a binary string. express;

[0097] (2) Mesoscale encoding: Defines four types of neural networks that can be represented by unit blocks at the mesoscale. and network layers =2. Each unit block can be... and Choose a value from the range as your configuration. and ;

[0098] (3) Microscale coding: At the microscale, based on the network type defined at the mesoscale, parameter coding is configured specifically for that network type. For different... Different parameters were set: for CNNs, the DP parameter is the kernel size CK. For LSTM, the DP parameter is the size of the hidden nodes. ,in Take 5. Set the value to 20; for GRU, the DP parameter is also the size of the hidden nodes. ,in Take 5. Set the value to 20; for Transformer, the DP parameter is the hidden state dimension. and the number of attention heads ;

[0099] S3 Population Evolution Iteration

[0100] S3.1 Initialize the population

[0101] Initialize population Set the population size to NP=50 and generate population individuals based on a multi-scale search space;

[0102] S3.2 Fitness Assessment

[0103] For each individual (candidate architecture) in each population, respectively in The individual is trained on several subsets of datasets, and the mean squared error between the predicted and true values ​​is calculated on the same subset of validation sets. The mean squared error is used as the fitness vector of the individual to evaluate its performance on different randomly split subsets of datasets.

[0104] S3.2 Population Crossover

[0105] For population The process involves performing a crossover operation based on a multi-scale search space design to generate corresponding offspring individuals, resulting in an evolved offspring population. The multi-scale crossover operator is a scale-aware crossover operator. By quantifying the differences between parent individuals at macro, meso, and micro scales, this operator selects the scale with the greatest difference for the crossover operation, thus ensuring more efficient exploration of the search space during evolution. Specifically, the crossover operator of the multi-scale genetic operator is as follows:

[0106] S3.2.1 Calculating Scale Differences. First, for the selected parent individuals, the following indicators are used to calculate differences across three scales:

[0107] (1) Macro-scale difference: Macro-scale difference is measured by the Hamming distance between the binary topological strings of two parent nodes. The Hamming distance is calculated by the number of bits that differ between corresponding positions in the two binary strings, directly reflecting the difference in the unit block connection pattern.

[0108]

[0109] (2) Mesoscale differences: Mesoscale differences are assessed by comparing parameter differences between corresponding cell blocks from two parents. For each cell block at the same location, if or If any parameter differs, the difference counter increments by 1. If both parameters are the same, the counter remains unchanged. After counting the differences across all corresponding cell blocks, the total count is normalized to obtain the mesoscale difference value:

[0110]

[0111] (3) Microscale differences: Microscale differences are identified by comparing two parents. and The corresponding unit blocks have the same The calculation is based on the differences in the module parameter sets. For modules with the same... The module we extract and The parameter set is defined. For each parameter in the set, if a difference exists, the difference counter is incremented by 1. If all parameters are the same, the counter remains unchanged. The total count is then normalized to obtain the microscale difference values:

[0112]

[0113] in It represents the number of network unit blocks of the same type. Number of parameter types;

[0114] S3.2.2 Multi-scale crossover operator

[0115] After calculating the differences across the three scales, the scale with the largest difference is selected as the crossover target. Then, the crossover operation is performed using the crossover operator corresponding to the selected scale. The specific operators for each scale are described below:

[0116] (1) Macroscale crossover operator: The crossover operation at the macro level uses binary topological strings. For objects. This operation involves two core steps: First, from the binary topology string of the parent individual... A random position is selected as the intersection point; then, the parent individual swaps the segments after the intersection point in its topological string.

[0117] (2) Mesoscale Crossover Operator: Mesoscale crossover enables unit block-specific swapping. First, a unit block position is randomly selected from the parent individuals, and this position remains consistent between the parent individuals to ensure the alignment and swapping of corresponding unit blocks. Second, the parent individual swaps the entire unit block at the selected position, such as... Figure 3 As shown;

[0118] (3) Microscale crossover operator: The microscale crossover operator realizes the internal parameter exchange of specific unit blocks, preserving network type consistency while optimizing parameters. First, it identifies the common parameters between two parents. The unit blocks are then selected, and a parameter category is randomly chosen from these matching unit blocks. Next, the selected parameter values ​​are swapped between the matching unit blocks of the two parents, such as... Figure 4 As shown;

[0119] S3.3 Population Variation

[0120] After crossover occurs in the population, the mutation process begins. For the population... The mutation operation, based on a multi-scale search space design, is performed to generate corresponding offspring individuals, resulting in an evolved offspring population. The mutation operator is a multi-scale random mutation operator. During the mutation process, this operator randomly selects a scale in the search space to perform the corresponding mutation. Specifically, the specific mutation mechanism for each scale is as follows:

[0121] (1) Macroscale variation: The macroscale method takes the binary topological string of an individual as the research object, first randomly selects mutation sites, and then performs bit flipping operation;

[0122] (2) Mesoscale variation: The mesoscale method takes the unit block characteristics of an individual as its object. First, it randomly selects a unit block, and then reselects the unit block containing... or And then according to the new and Redetermine the specific parameters at the micro level, such as Figure 5 As shown;

[0123] (3) Microscale variation: The mesoscale method takes the specific parameters of the unit block of an individual as the object. First, a unit block is randomly selected, and then the variation is determined based on the unit block's... and Redefine the specific parameters DP at the micro level, such as Figure 6 As shown;

[0124] S3.3 Evaluate the offspring population

[0125] For each individual in the evolved offspring population, respectively in The individual is trained on several subsets of datasets and the mean squared error of prediction is calculated on the same subset of validation sets. The mean squared error value is used as the fitness vector of the individual to evaluate its performance on different randomly split subsets of datasets and select the best individuals to enter the next evolution.

[0126] S3.4 Termination Judgment

[0127] Determine if a single evolutionary generation has been reached. If the target is reached, evolution stops, and the final population is non-dominated sorted to extract the Pareto optimal solution for the first frontier. From this solution set, the individual with the smallest mean squared error in each subset is selected, for a total of [number missing]. Each candidate architecture is selected to ensure optimal performance on its corresponding randomly partitioned subset of the dataset.

[0128] S4 Distance-Based Attention Integration Strategy

[0129] The core objective of this module is to address the problem that "a single architecture struggles to adapt to datasets with different distributions within the same topology," by fusing data through an attention mechanism. The output of each candidate architecture is used to generate a final model that combines high accuracy and strong generalization.

[0130] S4.1 Mixed Weight Calculation

[0131] The spatial and temporal weights generated by S1 are linearly weighted and fused to obtain the hybrid weights:

[0132]

[0133] in This is the weighting coefficient for spatial weights, taken as 0.6. The weighting factor for time weighting is set to 0.4;

[0134] S4.2 Attention-Weighted Integration

[0135] For each input sample (traffic matrix) in the global test set, input respectively 1 candidate architecture, obtained Predicted delay value The final predicted value is calculated using an attention mechanism. :

[0136]

Claims

1. A multi-scale, multi-objective evolutionary neural architecture search algorithm considering model generalization, comprising the following steps: S1 Dataset Segmentation and Weight Calculation The core objective of this module is to achieve fast and unbiased processing of the dataset through random partitioning, while quantifying the similarity between subsets to determine subsequent ensemble weights. The specific process is as follows: S1.1 Dataset Splitting After shuffling the data, the sizes are adjusted according to a preset ratio. The original dataset Divided into Subdatasets ,in During splitting, ensure that the sample size deviation of each subset does not exceed 5% of the total sample size of the original dataset; S1.2 Similarity Weight Calculation Based on the spatiotemporal dual-dimensional features of traffic data, two complementary similarity measurement methods are used to quantify the distributional differences among randomly segmented subsets, and these differences are transformed into weights for subsequent model integration. and ; S2 Search Space Initialization By constructing a multi-scale search space (MS³) covering all dimensions of neural networks, we can achieve efficient exploration of "unit block combination - network type - parameter configuration"; S2.1 Define the individual coding format Each individual in the population is set as the neural network is decomposed into... Each functionally independent unit block, in which Each unit block serves as a basic functional unit of the architecture, and its corresponding network type and parameters can be independently configured, denoted as . ,in Representing the The network type contained in each unit block Network layers And detailed network configuration; simultaneously, one topology control unit (T) is set up, using cell blocks to define the connection relationships between cell blocks, and conceptualizing the network composed of cell blocks as a graph. Each cell block serves as a node in the graph. ; S2.2 Define the multi-scale coding rules for the search space Based on individual coding forms, coding is performed at three scales: macro, meso, and micro. The macro scale encodes the overall connectivity of the network, the meso scale encodes the network type and number of layers of network unit blocks, and the micro scale encodes the specific configuration of unit blocks. S3 Population Evolution Iteration S3.1 Initialize the population Initialize population Set the population size to NP, and generate population individuals based on a multi-scale search space; S3.2 Fitness Assessment For each individual (candidate architecture) in each population, respectively in The individual is trained on several subsets of datasets, and the mean squared error between the predicted and true values ​​is calculated on the same subset of validation sets. The mean squared error is used as the fitness vector of the individual to evaluate its performance on different randomly split subsets of datasets. S3.2 Population Crossover For population The algorithm performs crossover operations based on a multi-scale search space design to generate corresponding offspring individuals, resulting in an evolved offspring population. The multi-scale crossover operator is a scale-aware crossover operator. By quantifying the differences between parent individuals at macro, meso, and micro scales, this operator selects the scale with the greatest difference for crossover operations, thereby ensuring more efficient exploration of the search space during the evolutionary process. S3.3 Population Variation After crossover occurs in the population, the mutation process begins. For the population... The mutation operation, based on a multi-scale search space design, is performed to generate corresponding offspring individuals, resulting in an evolved offspring population. The mutation operator is a multi-scale random mutation operator. During the mutation process, this operator randomly selects a scale in the search space to perform the corresponding mutation. S3.3 Evaluate the offspring population For each individual in the evolved offspring population, respectively in The individual is trained on several subsets of datasets and the mean squared error of prediction is calculated on the same subset of validation sets. The mean squared error value is used as the fitness vector of the individual to evaluate its performance on different randomly split subsets of datasets and select the best individuals to enter the next evolution. S3.4 Termination Judgment Determine if a single evolutionary generation has been reached. If the target is reached, evolution stops, and the final population is non-dominated sorted to extract the Pareto optimal solution for the first frontier. From this solution set, the individual with the smallest mean squared error in each subset is selected, for a total of [number missing]. Each candidate architecture is selected to ensure optimal performance on its corresponding randomly partitioned subset of the dataset. S4 Distance-Based Attention Integration Strategy The core objective of this module is to address the problem that "a single architecture struggles to adapt to datasets with different distributions within the same topology," by fusing data through an attention mechanism. The output of each candidate architecture is used to generate a final model that combines high accuracy and strong generalization. S4.1 Mixed Weight Calculation The spatial and temporal weights generated by S1 are linearly weighted and fused to obtain the hybrid weights. ,in These are the weighting coefficients for spatial weights. The weighting coefficient for time weighting; S4.2 Attention-Weighted Integration For each input sample (traffic matrix) in the global test set, input respectively 1 candidate architecture, obtained Predicted delay value The final predicted value is calculated using an attention mechanism. .

2. The multi-scale, multi-objective evolutionary neural architecture search algorithm considering model generalization as described in claim 1, characterized in that, S1 uses multiple similarity measurement methods to calculate similarity. Specifically, the following two similarity measurement methods are used: (1) Spatial Dimension Similarity: Calculated using cosine similarity, the formula is as follows: This is used to measure the "spatial structural consistency" of the flow matrices between two subsets of data. (2) Time-dimensional similarity: Calculated using the Pearson correlation coefficient, the formula is as follows: Used to measure the "consistency of time fluctuation trends" between flow sequences of two subsets of data; The corresponding similarity is calculated using the above formula. and Then, the similarity weights are normalized to obtain the spatial weights. and time weight .

3. The multi-scale, multi-objective evolutionary neural architecture search algorithm considering model generalization as described in claim 1, characterized in that, S2 employs a multi-scale coding rule, specifically as follows: (1) Macro-scale coding: using a length of The binary string represents the connection relationship between unit blocks at the macro scale, directly reflecting the information transmission path between unit blocks through binary bits. The nodes representing unit blocks are... and The connection between them can be used This means that if the node and There is a connection. =1, otherwise 0. Each unit block can selectively receive the output of the preceding unit block. Therefore, the connection status of a neural network consisting of n unit blocks can be represented by a binary string. express; (2) Mesoscale encoding: Defining the representable unit blocks at the mesoscale. Neural Network Types and network layers Each unit block can be and Choose a value from the range as your configuration. and ; (3) Microscale coding: At the microscale, based on the network type defined at the mesoscale, the parameter coding DP specific to that network type is configured.

4. The multi-scale, multi-objective evolutionary neural architecture search algorithm considering model generalization as described in claim 1, characterized in that, S3 designs a scale-aware crossover operator based on a multi-scale search space. Specifically, the steps of the scale-aware crossover operator of the multi-scale genetic operator are as follows: First, scale differences were calculated. For selected parental individuals, the following indicators were used to calculate differences across three scales: (1) Macroscale difference: Macroscale difference is measured by the Hamming distance between the binary topological strings of two parent nodes, and the formula is as follows: Hamming distance, by counting the number of bits that differ between corresponding positions of two binary strings, directly reflects the difference in the unit block connection pattern; (2) Mesoscale differences: Mesoscale differences are assessed by comparing the parameter differences between corresponding cell blocks of two parents. For each cell block at the same location, if or If any parameter differs, the difference counter increments by 1. If both parameters are the same, the counter remains unchanged. After counting the differences across all corresponding cell blocks, the total count is normalized to obtain the mesoscale difference value, using the formula: ; (3) Microscale differences: Microscale differences are identified by comparing two parents. and The corresponding unit blocks have the same The calculation is based on the differences in the module parameter sets. For modules with the same... The module we extract and The parameter set is defined. For each parameter in the set, if a difference exists, the difference counter is incremented by 1. If all parameters are the same, the counter remains unchanged. The total count is then normalized to obtain the microscale difference value, as shown in the formula: ,in It represents the number of network unit blocks of the same type. Number of parameter types; After calculating the differences across the three scales, the scale with the largest difference is selected as the crossover target. Then, the crossover operation is performed using the crossover operator corresponding to the selected scale. The specific operators for each scale are described below: (1) Macroscale crossover operator: The crossover operation at the macro level is performed using binary topological strings. For objects. This operation involves two core steps: First, from the binary topology string of the parent individual... A random position is selected as the intersection point; then, the parent individual swaps the segments after the intersection point in its topological string. (2) Mesoscale crossover operator: Mesoscale crossover enables unit block-specific swapping. First, a unit block position is randomly selected from the parent individuals, and this position remains consistent between the parent individuals to ensure the alignment and swapping of the corresponding unit blocks. Second, the parent individual swaps the entire unit block at the selected position; (3) Microscale crossover operator: The microscale crossover operator realizes the internal parameter exchange of specific unit blocks, preserving network type consistency while optimizing parameters. First, it identifies the two parents with the same The unit blocks are then matched, and a parameter category is randomly selected from these matching unit blocks. Next, the selected parameter values ​​are swapped between the matching unit blocks of the two parents.

5. The multi-scale, multi-objective evolutionary neural architecture search algorithm considering model generalization as described in claim 1, characterized in that, S3 designs a multi-scale random mutation operator based on a multi-scale search space. Specifically, the mutation mechanism for each scale is as follows: (1) Macroscale variation: The macroscale method takes the binary topological string of an individual as the research object, first randomly selects mutation sites, and then performs bit flipping operation; (2) Mesoscale variation: The mesoscale method takes the unit block characteristics of an individual as the object, first randomly selects a unit block, and then reselects the unit block containing... or And then according to the new and To redefine the specific parameters at the microscopic level; (3) Microscale variation: The mesoscale method takes the specific parameters of the unit block of an individual as the object. First, a unit block is randomly selected, and then the parameters of the unit block are... and The specific parameters DP at the microscopic level are redefined.

Images