A communication frequency prediction method and device, electronic equipment and storage medium

Abstract

The present application relates to the technical field of communication, and especially relates to a communication frequency prediction method and device, electronic equipment and storage medium, the method comprises the following steps: obtaining the aggregation encoding of each initial model in a plurality of model clusters and the fitness of each initial model under each communication scenario; selecting the initial model meeting a preset fitness condition from each model cluster based on the fitness to form a target model set; transforming the base encoding in the aggregation encoding of each initial model to change each initial model, and adding the changed initial model to the target model set; taking all the target models in the added target model set as new initial models, repeating the above steps until a preset iteration stop condition is met; and performing communication frequency prediction based on the final target model and the fitness of each target model under each communication scenario. The present application can predict the communication frequency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to a communication frequency prediction method and device, electronic equipment and storage medium. BACKGROUND

[0002] Communication frequency prediction is a process of modeling and analyzing historical spectrum data to estimate the signal strength, occupancy status or interference level of a specific frequency band in the future. In the field of wireless communication, spectrum resources are extremely scarce, and accurate frequency prediction can bring many benefits. For example, it can effectively avoid channel conflicts, such as interference between 5G communication and radar frequency bands; improve the efficiency of dynamic spectrum access, such as more flexible spectrum allocation in cognitive radio; reduce base station energy consumption, optimize resource utilization by on-demand bandwidth allocation; and ensure the reliability of critical communications, such as reserving necessary resources for emergency frequency bands.

[0003] Given the importance of these, the research and application of communication frequency prediction is particularly urgent. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a communication frequency prediction method and device, electronic equipment and storage medium, which can predict the communication frequency.

[0005] In a first aspect, the embodiments of the present application provide a communication frequency prediction method, which comprises:

[0006] Obtaining the aggregated encoding of each initial model in a plurality of model clusters and the fitness of each initial model in each communication scenario; the model cluster is obtained by clustering all initial models based on the aggregated encoding of each initial model; the aggregated encoding is composed of the base encoding of a plurality of sub-modules in the initial model;

[0007] Based on the fitness, selecting the initial model that meets the preset fitness condition from each model cluster to form a target model set;

[0008] Transforming the base encoding in the aggregated encoding of each initial model to modify each initial model, and adding the modified initial model to the target model set;

[0009] Taking all target models in the added target model set as new initial models, jumping to the step of obtaining the aggregated encoding of each initial model in each model cluster and the fitness of each initial model in each communication scenario to continue execution until the preset iteration stop condition is met;

[0010] Based on the final target model and the fitness of each target model in each communication scenario, the communication frequency is predicted.

[0011] In a possible implementation, the initial models are clustered based on the aggregated encodings of the initial models to obtain a plurality of model clusters by the following steps:

[0012] Similarities between each two initial models are calculated based on the aggregated encodings of the initial models.

[0013] The initial models are clustered based on the similarities to obtain a plurality of model clusters.

[0014] In a possible implementation, the initial models meeting the preset fitness condition are filtered from each model cluster based on the fitness to form the target model set, including:

[0015] For each model cluster, a preset number of initial models with the maximum fitness in the model cluster are determined as first intermediate models.

[0016] The target model set is determined based on the fitness of the first intermediate models and the initial models in each communication scenario.

[0017] In a possible implementation, the target model set is determined based on the fitness of the first intermediate models and the initial models in each communication scenario, including:

[0018] A plurality of initial models with the maximum fitness are filtered from all the initial models according to a preset ratio to obtain second intermediate models.

[0019] All the first intermediate models and all the second intermediate models are merged to obtain the target model set.

[0020] In a possible implementation, the base encoding in the aggregated encoding of any initial model is transformed to modify the initial model by the following steps, including:

[0021] For each sub-module in each initial model, a mutation intensity corresponding to the sub-module in the initial model is calculated according to a gradient mean of the sub-module in the initial model and a learning rate.

[0022] A mutation probability corresponding to the sub-module in the initial model is calculated according to the mutation intensity corresponding to the sub-module in the initial model and a maximum mutation intensity corresponding to all the sub-modules in the initial model.

[0023] The base encoding of each sub-module in the aggregated encoding of the initial model is mutated based on the mutation intensity and the mutation probability corresponding to each sub-module in the initial model, to modify the initial model.

[0024] In a possible implementation, the communication frequency prediction based on the final target model and the fitness of the target model in each communication scenario comprises:

[0025] obtaining a communication frequency of the to-be-predicted communication scenario at a current time;

[0026] selecting a prediction model from all target models based on the fitness of each target model in the to-be-predicted communication scenario;

[0027] inputting the communication frequency at the current time into the prediction model to obtain a communication frequency at a next time.

[0028] In a possible implementation, the selecting a prediction model from all target models based on the fitness of each target model in the to-be-predicted communication scenario comprises:

[0029] calculating an initial selected probability of each target model based on the fitness of each target model in the to-be-predicted communication scenario;

[0030] for each target model, accumulating the initial selected probabilities of all target models before the target model to obtain a target selected probability of the target model; the target models before the target model refer to the models with preset serial numbers less than that of the target model;

[0031] if the target selected probability of the target model is greater than a random selected probability, and the random selected probability is greater than a target selected probability of a last target model corresponding to the target model, the target model is determined as the prediction model.

[0032] In a second aspect, an embodiment of the present application further provides a communication frequency prediction device, and the device comprises:

[0033] an obtaining module, configured to obtain an aggregated encoding of each initial model in a plurality of model clusters and a fitness of each initial model in each communication scenario; the model clusters are obtained by clustering all initial models based on the aggregated encoding of each initial model; the aggregated encoding is composed of base encodings of a plurality of sub-modules in the initial model;

[0034] a screening module, configured to screen out initial models meeting a preset fitness condition from each model cluster based on the fitness to form a target model set;

[0035] a transformation module, configured to transform the base encodings in the aggregated encoding of each initial model to change each initial model, and add the changed initial model to the target model set;

[0036] a jumping module, configured to take all the target models in the target model set after the addition as new initial models, and jump to the step of acquiring the aggregated encoding of each initial model in each model class cluster and the fitness of each initial model in each communication scenario, to continue the execution until the preset iteration stop condition is met;

[0037] a prediction module, configured to perform communication frequency prediction based on the final target model and the fitness of each target model in each communication scenario.

[0038] In a third aspect, an electronic device is provided, which includes a processor, a storage medium and a bus. The storage medium stores machine readable instructions executable by the processor. When the electronic device is running, the processor communicates with the storage medium through the bus. The processor executes the machine readable instructions to perform the steps of the communication frequency prediction method according to any one of the first aspect.

[0039] In a fourth aspect, a computer readable storage medium is provided, which stores a computer program. When the computer program is run by a processor, the steps of the communication frequency prediction method according to any one of the first aspect are performed.

[0040] The embodiments of the present application provide a communication frequency prediction method, device, electronic device and storage medium. The method includes: acquiring the aggregated encoding of each initial model in a plurality of model class clusters and the fitness of each initial model in each communication scenario; selecting initial models that meet a preset fitness condition from each model class cluster based on the fitness to form a target model set; transforming the base encoding in the aggregated encoding of each initial model to change each initial model, and adding the changed initial model to the target model set; taking all the target models in the target model set after the addition as new initial models, and repeating the above steps until a preset iteration stop condition is met; and performing communication frequency prediction based on the final target model and the fitness of each target model in each communication scenario. The communication frequency can be predicted by the embodiments of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be considered as limiting the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0042] Figure 1 a flowchart of a communication frequency prediction method provided by the embodiments of the present application is shown;

[0043] Figure 2 A splitting schematic diagram of an initial model provided by the embodiment of the application is shown;

[0044] Figure 3 A splitting schematic diagram of a model body provided by the embodiment of the application is shown;

[0045] Figure 4 A variation schematic diagram of an initial model provided by the embodiment of the application is shown;

[0046] Figure 5 A crossing schematic diagram of an initial model provided by the embodiment of the application is shown;

[0047] Figure 6 A model iteration flowchart provided by the embodiment of the application is shown;

[0048] Figure 7 A communication frequency prediction flowchart provided by the embodiment of the application is shown;

[0049] Figure 8 A structure schematic diagram of a communication frequency prediction device provided by the embodiment of the application is shown;

[0050] Figure 9 A structure schematic diagram of an electronic device provided by the embodiment of the application is shown. DETAILED DESCRIPTION

[0051] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of illustration and description, and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn according to the actual proportions. The flowcharts show the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowcharts can not be implemented in sequence, and the steps without logical context relationship can be reversed in sequence or implemented simultaneously. In addition, one or more other operations can be added to the flowcharts or one or more operations can be removed from the flowcharts under the guidance of the content of the present application.

[0052] In addition, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0053] In order to enable those skilled in the art to use the content of the present application, the following implementation is given in combination with a specific application scenario "communication technology field". For those skilled in the art, the general principles defined herein can be applied to other embodiments and application scenarios without departing from the spirit and scope of the present application. Although the present application is mainly described in connection with the "communication technology field", it should be understood that this is only an exemplary embodiment.

[0054] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0055] The communication frequency prediction method provided by the embodiments of the present application is described in detail below.

[0056] Referring to Figure 1 The flowchart of the communication frequency prediction method provided by the embodiments of the present application is shown in FIG. 1. The exemplary steps of the embodiments of the present application are described below.

[0057] S101, obtaining the aggregated encoding of each initial model in a plurality of model clusters and the fitness of each initial model in each communication scenario.

[0058] In the embodiments of the present application, the model clusters are obtained by clustering all initial models based on the aggregated encoding of each initial model; the initial models in the same model cluster have similar functional characteristics. The aggregated encoding is composed of the basic encoding of all sub-modules in the initial model. The fitness of the initial model in the communication scenario is used to evaluate the performance of the model, which covers the accuracy, resource consumption and competition between other initial models when the initial model predicts the communication frequency in the communication scenario. The basic encoding of the sub-module includes the preset identification string corresponding to the sub-module and the preset serial number corresponding to the initial model. The preset identification strings corresponding to the sub-modules with the same network structure are different, and the preset identification strings corresponding to the sub-modules with the same network structure are the same.

[0059] Here, before obtaining the aggregated encoding of each initial model in multiple model clusters and the fitness of each initial model in various communication scenarios, it is necessary to pre-build an initial model library. Specifically, this includes: selecting at least one initial model from the time series prediction models of each preset model category and adding it to the initial model library; the preset model categories refer to the model categories that can be used for communication frequency prediction, including regression mobility models, time series decomposition models, and neural network models. For example, regression mobility models include ARIMA, SARIMAX, VAR, VARMAX, ARMD, Holt-Winters, etc.; time series decomposition models include Prophet, SEATS, STL, TSR, X11 decomposition, etc.; and neural network models include TCN, LSTM, Timer-XL, Autoformer, Transformer and its variants, etc.

[0060] Furthermore, due to the significant differences in model types and structures among different initial models, it is not conducive to subsequent iterative evolution operations. Therefore, this application designs a model gene encoding strategy to uniformly modularize (block) the models in the model library. Specifically, the gene encoding of each initial model in the initial model library is performed through the following steps to obtain the aggregate encoding of each initial model:

[0061] Step 1: Divide the initial model into a preprocessing module, the main model body, and a postprocessing module.

[0062] In the embodiments of this application, reference is made to Figure 2 The diagram illustrates the breakdown of the initial model provided in this embodiment. The preprocessing module and postprocessing module do not participate in the training of the initial model and are each treated as an independent block within the initial model. Specifically, the preprocessing module is the initial part of the initial model without trainable model parameters, used for operations such as differencing, channel expansion, and loop stacking of the frequency sequence. The postprocessing module is the later part of the initial model without trainable model parameters, used to restore the output of the initial model to a frequency sequence. The main body of the model is the portion between the preprocessing and postprocessing modules, containing the trainable model parameters.

[0063] Step 2: First, decompose the main body of the initial model into the smallest recurring functional units to obtain multiple smallest functional modules; then, take the non-recurring key units in the decomposed main body of the model as a module to obtain non-recurring key modules.

[0064] Typical examples of recurring minimum functional units include: residual blocks in ResNet, Inception blocks in GoogLeNet, and Transformer coding blocks.

[0065] In the embodiments of the present application, referring to Figure 3 Fig. 1 shows a split schematic diagram of a model subject provided by an embodiment of the present application. Non-repeated key modules are generally input processing modules and output heads, thus, after splitting, the sequence includes non-repeated key module 1, minimum function module 1, minimum function module 2, …, minimum function module n, non-repeated key module 2.

[0066] For example, the model subject includes an input processing block, nine consecutive residual blocks and an output head in sequence, then the sub-modules of the model subject after splitting include the input processing block, residual block 1, residual block 2, residual block 3, residual block 4, residual block 5, residual block 6, residual block 7, residual block 8, residual block 9 and the output head in sequence.

[0067] Step three, the pre-processing module, post-processing module, minimum function module and non-repeated key module obtained after splitting are all taken as sub-modules of the initial model.

[0068] Step four, for each sub-module, the preset identification string corresponding to the sub-module is taken as a prefix of the preset serial number corresponding to the initial model, to obtain the basic code corresponding to the sub-module.

[0069] For example, assuming that the pre-processing module has a preset identification string y as a prefix, and the post-processing module has a preset identification string h as a prefix. Then the basic code of the pre-processing module of the initial model with preset serial number 01 is y01, and the basic code of the post-processing module with preset serial number 02 is h02.

[0070] Step five, arrange the basic codes of the sub-modules according to the positions of the sub-modules in the initial model, to obtain the aggregate code of the initial model.

[0071] In the embodiments of the present application, through the aforementioned gene code processing, all initial models in the initial model library can be expressed in block units, for example, an initial model including a pre-processing module + model subject (input processing module + residual block + Inception block + output head) + post-processing module, the aggregate code of which can be represented as y01i01r01in01o01h01, wherein y01 is the pre-processing module of the initial model with preset serial number 01, i01 is the input processing module of the initial model with preset serial number 01, r01 is the residual block of the initial model with preset serial number 01, in01 is the Inception block of the initial model with preset serial number 01, o01 is the output head of the initial model with preset serial number 01, and h01 is the post-processing module of the initial model with preset serial number 01.

[0072] Further, all initial models are clustered based on the aggregate codes of the initial models through the following steps to obtain a plurality of model clusters:

[0073] Step one, based on the aggregate encoding of each initial model, calculate the similarity between each two initial models.

[0074] In the embodiments of the present application, the aggregate encoding of each two initial models is substituted into the following formula to obtain the similarity between the two initial models:

[0075]

[0076] wherein, is the aggregate encoding of the first initial model in each two initial models, is the aggregate encoding of the second initial model in each two initial models, is the similarity between the two initial models; is and the number of the same preset identification string in is and the number of the string in (twice the total number of the sub-modules in the two initial models, including all preset identification strings and all preset serial numbers of the initial models), is X the number of the string in (twice the number of the sub-modules in the first initial model), is Y the number of the string in (twice the number of the sub-modules in the second initial model).

[0077] Step two, based on the similarity, cluster all the initial models to obtain a plurality of model clusters.

[0078] In the embodiments of the present application, two initial models with a similarity greater than a preset similarity threshold (such as 80%) can be in the same model cluster. DBSCAN can also be used to cluster all the models according to the similarity between each two initial models, and thereby form a plurality of model clusters.

[0079] Here, imitating the example of evolution in nature, the initial models in the same model cluster are in the same niche level, so it can be considered that there is a strong competitive relationship between the initial models in the same model cluster, and the competitive relationship between the models in different model clusters is weaker. The distribution of the model clusters will affect the subsequent model iteration process.

[0080] Further, each initial model in the initial model library is allowed to perform a communication frequency prediction task in a plurality of different communication environments, and the fitness of each initial model in each communication environment is evaluated according to the prediction result. The fitness of any initial model in any communication scenario is calculated by the following formula: ​

[0081] ;

[0082] wherein, is the fitness of the initial model under the communication scenario, is a first preset weight, is the accuracy of the initial model in predicting the communication frequency under the communication scenario, is a preset accuracy reference value (such as 0.95), is a second preset weight, is the resource consumption of the initial model in predicting the communication frequency under the communication scenario (the present embodiment represents it by “the time length required for performing a communication frequency task once”), is a preset maximum acceptable resource overhead (such as 200 ms), is a third preset weight, is the number of initial models in the model cluster to which the initial model belongs (i.e. the number of initial models that compete with the initial model), is a preset maximum number of competing models.

[0083] S102, based on the fitness, select the initial models meeting the preset fitness condition from each model cluster to form a target model set.

[0084] In the present embodiment, in order to avoid losing specific types of model structures and models with excellent performance in the iteration process, in each iteration process, based on the fitness, the initial models meeting the preset fitness condition are forcibly selected from each model cluster, which can maximize the retention of the most comprehensive and high-quality target models and improve the accuracy of subsequent communication frequency prediction.

[0085] Specifically, based on the fitness, select the initial models meeting the preset fitness condition from each model cluster to form a target model set, including:

[0086] Step one, for each model cluster, determine the preset number of initial models with the maximum fitness in the model cluster as first intermediate models.

[0087] In the present embodiment, in order to avoid losing specific types of model structures in the iteration process, the preset number of initial models with the maximum fitness in each model cluster are first selected to determine the first intermediate models.

[0088] For example, there are model cluster A and model cluster B, and the preset number is 3. Then, the 3 initial models with the maximum fitness in model cluster A are selected to determine the first intermediate models, and the 3 initial models with the maximum fitness in model cluster B are selected to determine the first intermediate models.

[0089] Continuing the example, assume that the model cluster A contains initial model 1, initial model 2, initial model 3, initial model 4 and initial model 5. The fitness ranking is initial model 1 > initial model 3 > initial model 5 > initial model 2 > initial model 4. Therefore, initial model 1, initial model 3 and initial model 5 in the model cluster A are determined as the first intermediate models.

[0090] Step two, determine the target model set based on the fitness of each initial model in each communication scenario and the first intermediate models.

[0091] In the embodiments of the present application, in order to avoid the loss of outstanding models in the iteration process, it is also necessary to select outstanding models from all initial models. Specifically, a plurality of initial models with the largest fitness are selected from all initial models according to a preset proportion, to obtain second intermediate models; and all first intermediate models and all second intermediate models are merged to obtain the target model set.

[0092] For example, the preset proportion can be 10%, and the initial models with the top 10% fitness are selected from all initial models as the second intermediate models.

[0093] Optionally, in order to avoid the same model structure between the first intermediate models and the second intermediate models, it is necessary to remove the duplicate target models in the target model set to obtain the final target model set.

[0094] S103, transform the base encoding in the aggregate encoding of each initial model to change each initial model, and add the changed initial model to the target model set.

[0095] In the embodiments of the present application, in order to reduce the calculation workload and improve the model iteration efficiency, the base encoding in the aggregate encoding of the initial model with the largest fitness can be transformed according to a preset transformation proportion to change each initial model.

[0096] For example, the base encoding in the aggregate encoding of the initial model with the top 70% fitness is transformed, and the remaining 30% initial models are discarded.

[0097] Specifically, the base encoding in the aggregate encoding of any initial model is transformed to change the initial model by the following steps, including:

[0098] Step one, for each sub-module in each initial model, the mutation strength corresponding to the sub-module in the initial model is calculated according to the gradient mean value of the sub-module in the initial model and the learning rate.

[0099] In the embodiments of the present application, the mutation strength refers to the degree of mutation of the sub-module. Reasonable setting of the mutation strength is crucial for the performance of the final target model. If the mutation strength is too high, it can lead to jumping out of the local optimal solution too early, so that the target model with optimal performance cannot be obtained; if the mutation strength is too low, it can lead to falling into a local optimal solution, and a better target model cannot be explored. Therefore, the embodiments of the present application dynamically calculate the mutation strength corresponding to the sub-module in the initial model according to the gradient mean of the sub-module in the initial model and the learning rate.

[0100] Specifically, the gradient mean of the sub-module in the initial model and the learning rate are substituted into the following formula to obtain the mutation strength corresponding to the sub-module in the initial model:

[0101] ;

[0102] wherein, is the mutation strength corresponding to the sub-module in the initial model, is the learning rate of the sub-module in the initial model, is the gradient mean of the sub-module in the initial model, is a normal distribution with a mean of 0 and a standard deviation of is a normal distribution for introducing the randomness of the mutation strength.

[0103] Here, the gradient mean is an important concept of deep learning, which is usually used to describe the statistical characteristics of the gradient in the training process of the model, and is used to describe the average level of the gradient in the training process of the model. Specifically, the gradient mean of the sub-module refers to the average value of all parameter gradients in the sub-module in one iteration. The learning rate of the sub-module refers to the learning rate set for a certain sub-module (or sub-network) in the model during the training process. The learning rate is a hyperparameter that determines the step size of the model parameter update in each iteration.

[0104] Step two, and according to the mutation strength corresponding to the sub-module in the initial model and the maximum mutation strength corresponding to all sub-modules in the initial model, calculate the mutation probability corresponding to the sub-module in the initial model.

[0105] In this embodiment, the maximum mutation intensity corresponding to all sub-modules in the initial model refers to the maximum value among all mutation intensities corresponding to all sub-modules. The mutation probability corresponding to a sub-module is used to characterize the likelihood of the sub-module mutating. If the mutation probability is too high, it may cause premature exit from a local optimum, thus failing to obtain the optimal target model; if the mutation probability is too low, it may cause getting trapped in a local optimum, failing to explore a better target model. Therefore, this embodiment dynamically calculates the mutation probability corresponding to the sub-module in the initial model based on the mutation intensity corresponding to the sub-module in the initial model and the maximum mutation intensity corresponding to all sub-modules in the initial model.

[0106] Specifically, by substituting the mutation intensity corresponding to the submodule in the initial model and the maximum mutation intensity corresponding to all submodules in the initial model into the following formula, the mutation probability corresponding to the submodule in the initial model is obtained:

[0107] ;

[0108] in, This represents the mutation probability corresponding to this submodule in the initial model. This is a preset upper limit for the mutation probability. This represents the lower bound of the mutation probability. This represents the mutation intensity corresponding to this submodule in the initial model. This represents the maximum mutation intensity corresponding to all submodules in the initial model.

[0109] Step 3: Based on the mutation intensity and mutation probability of each sub-module in the initial model, mutate the basic encoding of each sub-module in the aggregate encoding of the initial model.

[0110] In the embodiments of this application, mutations include the addition, deletion, and replacement of submodules. These three mutation operations are independent of each other and may occur simultaneously. (See also...) Figure 4 The diagram shown illustrates the variations of the initial model provided in this embodiment. In the initial model structure, unvariated blocks are labeled A, and variated blocks are labeled B. Therefore, Figure 4 The initial model consists of nine blocks, A1 to A9. In mutation 1, a replacement occurred, specifically A1 was replaced with B1 and A9 was replaced with B9. In mutation 2, an addition occurred, with B8_9 added between A8 and A9. In mutation 3, both replacement and deletion occurred, with A2 and A7 being replaced with B2 and B7, and A8 being deleted.

[0111] Here, by the mutation strength and mutation probability of the gradient guiding sub-module, the initial model can be guided to preferentially mutate the part with greater impact on the result, accelerate the model structure exploration, and enrich the model system.

[0112] In addition, the base encoding in the aggregate encoding of any initial model is transformed to change the initial model, and the base encoding in the aggregate encoding of each two initial models is crossed to change the initial models.

[0113] In the embodiments of the present application, referring to Figure 5 The crossing of the initial models provided by the embodiments of the present application is shown in FIG. 1. The first initial model for crossing is composed of a total of nine blocks from A1 to A9, and the second initial model is composed of a total of nine blocks from B1 to B9. In crossing 1, the front and rear parts of the two initial models are exchanged, specifically, the A1-A5 part of the first initial model is exchanged with the B1-B5 part of the second initial model. In crossing 2, the middle parts of the two initial models are exchanged, and the A3-A7 part of the first initial model is exchanged with the B3-B7 part of the second initial model.

[0114] Here, 15% to 25% of the total initial models can be randomly selected for crossing operation each time, and the crossing position is also randomly determined. Here, no guidance is provided, and diversity is enhanced by completely random means.

[0115] S104, all target models in the target model set after being added are taken as new initial models, and jumping to the step of obtaining the aggregate encoding of each initial model in each model class cluster and the fitness of each initial model under each communication scenario to continue execution until a preset iteration stop condition is met.

[0116] In the embodiments of the present application, the preset iteration stop condition can be that the maximum number of iterations is reached, a target model greater than a fitness threshold is obtained, or the increase rate of the maximum fitness is less than a preset increase rate threshold.

[0117] Here, in S101 to S104, the initial model library is regarded as a population, each initial model in the library is regarded as an individual, and the communication frequency prediction is regarded as an ecological environment. The natural selection and genetics mechanism of Darwin's biological evolution theory are simulated, the model population is repeatedly iterated, and the optimal target model set is constructed.

[0118] In addition, in the integration and optimization process of the deep learning model, the problem of data format mismatch between sub-modules is often encountered, for example, one sub-module outputs a size of 256x256, while the next sub-module needs an input of 128x128. This difference hinders the direct combination between sub-modules. To overcome this challenge, the method introduces AI programming software such as RooCode to assist in executing the iterative process. Specifically, before S105 performs communication frequency prediction based on the final target model set and the fitness of each target model in each communication scenario, the method further includes: for any target model in the target model set, generating a transition block that adapts the data format between adjacent sub-modules in the target model, such as an up-sampling or down-sampling module, through large model coding. This method can effectively eliminate the problem of data format mismatch between sub-modules. In this way, not only can the seamless cooperation of each part of the model be ensured, but also the model access iterative process can be accelerated, realizing a fast model system integration evolution closed loop, thereby improving the adaptability and flexibility of the model and speeding up the development and iteration process.

[0119] In summary, referring to Figure 6 the model iteration flowchart provided by the embodiments of the present application.

[0120] S105, based on each target model and the fitness of each target model in each communication scenario, performs communication frequency prediction.

[0121] In the embodiments of the present application, the communication frequency of the to-be-predicted communication scenario at the current time is obtained; a prediction model is selected from all target models based on the fitness of each target model in the to-be-predicted communication scenario; and the communication frequency at the current time is input into the prediction model to obtain the communication frequency at the next time.

[0122] Further, selecting a prediction model from all target models based on the fitness of each target model in the to-be-predicted communication scenario includes:

[0123] Step one, based on the fitness of each target model in the to-be-predicted communication scenario, calculate the initial selection probability of each target model.

[0124] In the embodiments of the present application, the fitness of each target model in the to-be-predicted communication scenario is substituted into the following formula to obtain the initial selection probability of any target model:

[0125] ;

[0126] wherein, is the initial selection probability of the target model with a preset sequence number i, is the fitness of the target model with a preset sequence number i in the to-be-predicted communication scenario, the number of target models, the fitness of the target model with preset sequence number k in the communication scenario to be predicted.

[0127] Step two, for each target model, accumulate the initial selected probability of all target models before the target model to obtain the target selected probability of the target model.

[0128] In the embodiments of the present application, the target model before the target model refers to the model with a preset sequence number smaller than the preset sequence number of the target model. The initial selected probability of all target models before the target model is accumulated to obtain the target selected probability of the target model by the following formula:

[0129] ;

[0130] wherein, the target selected probability of the target model with preset sequence number i, the initial selected probability of the target model with preset sequence number k.

[0131] Step three, if the target selected probability of the target model is greater than the random selected probability, and the random selected probability is greater than the target selected probability of the last target model corresponding to the target model, the target model is determined as the prediction model.

[0132] In the embodiments of the present application, a random number r in the interval [0, 1) is generated as the random selected probability. If, the target model with preset sequence number i is selected to determine the prediction model for prediction.

[0133] Here, in the model selection, the roulette selection method is adopted, the probability of each model being selected is generated based on the fitness ratio, the high fitness model occupies a larger interval of the roulette, and the probability of being selected is higher; the low fitness model still has a small probability of being selected, maintaining diversity, and will not be limited to selecting the model with the highest fitness, which is more suitable for the idea of model integration evolution.

[0134] Referring to Figure 7 , a flowchart of prediction of communication frequency provided by the embodiments of the present application is shown. The whole process starts from the current T time, first uses the time-frequency analysis method to process the signal data, extracts the signal frequency; then inputs the data into the flowchart, selects the prediction model from all target models in the model library through the "selector" module; then the data flows into the core integrated model library, and the selected prediction model is called to perform prediction, and the predicted value of the communication frequency at T+1 time is output, that is, a single prediction is completed.

[0135] At T+1, the output prediction result is compared with the actual data, and an evaluator is input to generate evaluation results (such as accuracy, resource consumption, etc.), and the generated evaluation results are used as feedback signals to drive a new round of iterative evolution, form a "prediction-evaluation-optimization" closed loop, and continuously enrich the model system in the model library to improve the ability to adapt to non-stationary environments.

[0136] In summary, the application integrates multiple models to construct a model library, and uses a heuristic evolutionary learning strategy to enhance non-stationary environment adaptability by enriching the model system. Multiple models are integrated to compensate for the shortcomings of a single model, and the diversity of the model system is continuously enriched through iteration. The key core is the integrated construction and iterative evolution mechanism of the model system in the model library part.

[0137] Based on the same inventive concept, the embodiments of the application also provide a communication frequency prediction device corresponding to the communication frequency prediction method. Since the principle of solving problems in the device of the embodiments of the application is similar to the above-mentioned communication frequency prediction method of the embodiments of the application, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described here.

[0138] Referring to Figure 8 FIG. 1 is a schematic diagram of a communication frequency prediction device provided by an embodiment of the application. The device includes:

[0139] The acquisition module 801 is configured to acquire the aggregate encoding of each initial model in a plurality of model clusters and the fitness of each initial model in each communication scenario. The model clusters are obtained by clustering all initial models based on the aggregate encoding of each initial model. The aggregate encoding is composed of the base encoding of a plurality of sub-modules in the initial model.

[0140] The screening module 802 is configured to screen the initial models that meet a preset fitness condition from each model cluster based on the fitness, to constitute a target model set.

[0141] The transformation module 803 is configured to transform the base encoding in the aggregate encoding of each initial model to change each initial model, and add the changed initial model to the target model set.

[0142] The jump module 804 is configured to use all target models in the added target model set as new initial models, jump to the acquisition of the aggregate encoding of each initial model in each model cluster and the fitness of each initial model in each communication scenario, and continue to execute until a preset iteration stop condition is met.

[0143] The prediction module 805 is configured to perform communication frequency prediction based on the final target model and the fitness of each target model in each communication scenario.

[0144] The device provided by the application continuously iterates the diversity of the model system, and the key core is the integrated construction and iterative evolution mechanism of the model system in the model library part, which can predict the communication frequency.

[0145] As shown in Figure 9 The electronic device 900 provided by the embodiment of the application includes a processor 901, a memory 902 and a bus, the memory 902 stores machine readable instructions executable by the processor 901, when the electronic device is running, the processor 901 and the memory 902 communicate through the bus, and the processor 901 executes the machine readable instructions to perform the steps of the above-mentioned communication frequency prediction method.

[0146] Specifically, the above-mentioned memory 902 and processor 901 can be general memory and processor, which are not specifically limited here, and when the processor 901 runs the computer program stored in the memory 902, the above-mentioned communication frequency prediction method can be executed.

[0147] Corresponding to the above-mentioned communication frequency prediction method, the embodiment of the application further provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by the processor to perform the steps of the above-mentioned communication frequency prediction method.

[0148] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-mentioned system and device can refer to the corresponding process in the method embodiment, which will not be repeated in the application. In several embodiments provided by the application, it should be understood that the disclosed system, device and method can be implemented by other ways. The above-mentioned device embodiment is only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division way, and for example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed each other can be through some communication interface, indirect coupling or communication connection between the devices or modules, which can be electrical, mechanical or other forms.

[0149] The modules described as separate components can or can not be physically separated, and the components shown as modules can or can not be physical units, that is, they can be located in one place, or they can be distributed on a plurality of network units. According to the actual needs, part or all of the units can be selected to achieve the purpose of the embodiment of the application.

[0150] In addition, each function unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit.

[0151] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a nonvolatile computer readable storage medium executable by a processor. Based on this understanding, the technical solutions of the present application or the parts of the technical solutions that essentially contribute to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the information processing method described in each embodiment of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a ROM, a RAM, a magnetic disk or an optical disk, and various media that can store program codes.

[0152] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method of predicting a communication frequency, characterized by, The method comprises: obtaining the aggregate encoding of each initial model in each model cluster and the fitness of each initial model in each communication scenario; the model cluster is obtained by clustering all initial models based on the aggregate encoding of each initial model; the aggregate encoding is composed of the base encoding of multiple sub-modules in the initial model; selecting the initial model that meets the preset fitness condition from each model cluster based on the fitness to form a target model set; transforming the base encoding in the aggregate encoding of each initial model to modify each initial model, and adding the modified initial model to the target model set; taking all target models in the target model set after addition as new initial models, and jumping to the step of obtaining the aggregate encoding of each initial model in each model cluster and the fitness of each initial model in each communication scenario to continue execution until a preset iteration stop condition is met; performing communication frequency prediction based on the final target model and the fitness of each target model in each communication scenario; transforming the base encoding in the aggregate encoding of any initial model to modify the initial model by the following steps, including: for each sub-module in each initial model, substituting the gradient mean and learning rate of the sub-module in the initial model into the following formula to obtain the mutation intensity corresponding to the sub-module in the initial model; ； wherein, is the mutation strength of the sub-module in the initial model, is the learning rate of the sub-module in the initial model, is the gradient mean of the sub-module in the initial model, is a normal distribution with mean 0 and standard deviation is a normal distribution with mean 0 and standard deviation is used to introduce mutation strength randomness; and substituting the mutation intensity corresponding to the sub-module in the initial model and the maximum mutation intensity corresponding to all sub-modules in the initial model into the following formula to obtain the mutation probability corresponding to the sub-module in the initial model; ； wherein, is a mutation probability corresponding to the sub-module in the initial model, is a preset upper limit of the mutation probability, is a mutation lower limit of the mutation probability, is a mutation intensity corresponding to the sub-module in the initial model, is a maximum mutation intensity corresponding to all sub-modules in the initial model; based on the mutation intensity and mutation probability corresponding to each sub-module in the initial model, mutating the base encoding of each sub-module in the aggregate encoding of the initial model to modify the initial model.

2. The communication frequency prediction method according to claim 1, characterized by, The following steps are used to cluster all initial models based on the aggregate encoding of each initial model to obtain multiple model clusters: based on the aggregate encoding of each initial model, calculating the similarity between each two initial models; based on the similarity, clustering all initial models to obtain multiple model clusters.

3. The communication frequency prediction method according to claim 1, characterized by, The following steps are used to select the initial model that meets the preset fitness condition from each model cluster based on the fitness to form a target model set: for each model cluster, determining the preset number of initial models with the maximum fitness in the model cluster as first intermediate models; determining the target model set based on the first intermediate models and the fitness of each initial model in each communication scenario.

4. The communication frequency prediction method according to claim 3, characterized by, The following steps are used to determine the target model set based on the first intermediate models and the fitness of each initial model in each communication scenario: selecting the multiple initial models with the maximum fitness from all initial models in a preset proportion to obtain second intermediate models; merging all first intermediate models and all second intermediate models to obtain the target model set.

5. The communication frequency prediction method according to claim 1, characterized by, The following steps are used to perform communication frequency prediction based on the final target model and the fitness of each target model in each communication scenario: obtaining the communication frequency of the to-be-predicted communication scenario at the current time; selecting a prediction model from all the target models based on fitness of each target model under the communication scenario to be predicted; inputting the communication frequency at the current time into the prediction model to obtain the communication frequency at the next time.

6. The communication frequency prediction method according to claim 5, characterized by, The selecting a prediction model from all the target models based on fitness of each target model under the communication scenario to be predicted comprises: calculating an initial selection probability of each target model based on fitness of each target model under the communication scenario to be predicted; for each target model, accumulating the initial selection probabilities of all the target models before the target model to obtain a target selection probability of the target model; the target models before the target model refer to the models with preset serial numbers less than that of the target model; if the target selection probability of the target model is greater than a random selection probability, and the random selection probability is greater than a target selection probability of a last target model corresponding to the target model, the target model is determined as the prediction model.

7. A communication frequency prediction apparatus characterized by comprising: The apparatus comprises: an acquisition module, configured to acquire an aggregated encoding of each initial model in a plurality of model clusters and fitness of each initial model under each communication scenario; the model clusters are obtained by clustering all the initial models based on the aggregated encoding of each initial model; the aggregated encoding is composed of base encodings of a plurality of sub-modules in the initial model; a screening module, configured to screen out initial models satisfying a preset fitness condition from each model cluster based on the fitness to form a target model set; a transformation module, configured to transform the base encodings in the aggregated encoding of each initial model to modify each initial model, and add the modified initial model to the target model set; a jump module, configured to take all the target models in the added target model set as new initial models, jump to the acquisition of the aggregated encoding of each initial model in each model cluster and the fitness of each initial model under each communication scenario to continue execution until a preset iteration stop condition is satisfied; a prediction module, configured to perform communication frequency prediction based on the final target model and the fitness of each target model under each communication scenario. The transformation module is specifically configured to transform the base encodings in the aggregated encoding of any initial model to modify the initial model by the following steps, comprising: for each sub-module in each initial model, substituting the gradient mean value and the learning rate of the sub-module in the initial model into the following formula to obtain a mutation intensity corresponding to the sub-module in the initial model; ； wherein, is the mutation strength of the sub-module in the initial model, is the learning rate of the sub-module in the initial model, is the gradient mean of the sub-module in the initial model, is a normal distribution with mean 0 and standard deviation is a normal distribution with mean 0 and standard deviation is used to introduce mutation strength randomness; and substituting the mutation intensity corresponding to the sub-module in the initial model and the maximum mutation intensity corresponding to all the sub-modules in the initial model into the following formula to obtain a mutation probability corresponding to the sub-module in the initial model; ； wherein, is a mutation probability corresponding to the sub-module in the initial model, is a preset upper limit of the mutation probability, is a mutation lower limit of the mutation probability, is a mutation intensity corresponding to the sub-module in the initial model, is a maximum mutation intensity corresponding to all sub-modules in the initial model; based on the mutation intensity and the mutation probability corresponding to each sub-module in the initial model, the base encodings of each sub-module in the aggregated encoding of the initial model are mutated to modify the initial model.

8. An electronic device, comprising: comprises: A processor, a storage medium storing machine readable instructions executable by the processor, and a bus for communication between the processor and the storage medium when the electronic device is running, the processor executing the machine readable instructions to perform the steps of the communication frequency prediction method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, A computer program stored on the computer readable storage medium, the computer program being executed by the processor to perform the steps of the communication frequency prediction method of any one of claims 1 to 6.

Images