Method for training communication decision-making model, electronic device, computer-readable medium

Abstract

The present disclosure provides a method for training a communication decision model, the method comprising: training an initial model based on a first decision sample to adjust the model parameters of the initial model to obtain a first training model; wherein the first decision sample is a sample of a decision already made by the first communication site; obtaining at least one second model parameter modification value of at least one second training model; wherein each second training model is obtained by training the initial model based on a second decision sample of a second communication site, the second decision sample being a sample of a decision already made by the second communication site; each second model parameter modification value reflects at least a partial modification of the model parameters of the second training model relative to the model parameters of the initial model; adjusting the model parameters of the first training model based on at least a portion of the second model parameter modification values to obtain a communication decision model for the first communication site. The present disclosure also provides an electronic device and a medium.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of communication technology, and in particular to a method, electronic device, and computer-readable medium for training a communication decision model. Background Art

[0002] There are a large number of communication sites in the communication network, and these communication sites often need to make some decisions, such as the direction of business processes and user policies.

[0003] Manually set fixed logic decisions cannot adapt to the different situations of different communication sites, and it takes a long time of manual statistical analysis to make the set parameters relatively accurate.

[0004] The decision model using machine learning methods is not easy to converge and requires a large amount of transmission resources. The difference in the number of samples between different communication sites can easily lead to poor decision-making results for communication sites with a small number of samples. Summary of the Invention

[0005] The embodiments of the present disclosure provide a method for training a communication decision model, an electronic device, and a computer-readable medium.

[0006] In a first aspect, an embodiment of the present disclosure provides a method for training a communication decision model, wherein the communication decision model is used for decision making of a communication site, the method comprising:

[0007] Training an initial model according to a first decision sample to adjust model parameters of the initial model to obtain a first training model; wherein the first decision sample is a sample of a decision already made by the first communication site;

[0008] Obtaining at least one second model parameter modification value of at least one second training model; wherein each second training model is obtained by training the initial model based on a second decision sample of a second communication site, the second decision sample being a sample of a decision already made by the second communication site; and each second model parameter modification value reflects at least a partial modification of a model parameter of the second training model relative to a model parameter of the initial model;

[0009] The model parameters of the first training model are adjusted according to at least part of the second model parameter modification values to obtain a communication decision model for the first communication site.

[0010] In some embodiments, the second model parameter modification value of each second training model is obtained by:

[0011] clustering the second decision samples of the second training model;

[0012] The sum of modifications of all second decision samples in each cluster is determined to be a second model parameter modification value, and the modification of each second decision sample is the modification of the model parameter caused when the initial model is trained according to the second decision sample.

[0013] In some embodiments, adjusting the model parameters of the first training model according to at least part of the second model parameter modification values to obtain a communication decision model for the first communication site includes:

[0014] Calculating, based on the first decision sample, a decision effect of a model adjusted according to each second model parameter modification value on the first training model to determine an accepted second model parameter modification value;

[0015] The model parameters of the first training model are adjusted according to the accepted second model parameter modification value to obtain a communication decision model for the first communication site.

[0016] In some embodiments, adjusting the model parameters of the first training model according to the accepted second model parameter modification value to obtain a communication decision model for the first communication site includes:

[0017] If there are multiple accepted second model parameter modification values, the model parameters of the first training model are adjusted according to each accepted second model parameter modification value and its corresponding weight, and the weight corresponding to each accepted second model parameter modification value is calculated based on the number of second decision samples corresponding to the second model parameter modification value.

[0018] In some embodiments, adjusting the model parameters of the first training model according to the accepted second model parameter modification value to obtain a communication decision model for the first communication site includes:

[0019] Adjust the first training model according to the accepted second model parameter modification value to obtain a first pre-training model;

[0020] Calculating a decision effect of the first pre-trained model according to the first decision sample;

[0021] If the decision effect does not meet the preset conditions, the first training model is determined to be the communication decision model for the first communication site.

[0022] In some embodiments, obtaining the communication decision model for the first communication site further includes:

[0023] The clustering manner of the second decision samples is adjusted according to the second model parameter modification values accepted by the plurality of first communication sites.

[0024] In some embodiments, obtaining at least one second model parameter modification value of at least one second training model includes:

[0025] The second model parameter modification values of all second training models are filtered according to a predetermined filtering rule, and the second model parameter modification values of the second training models that pass the filtering are obtained.

[0026] In some embodiments, after obtaining the communication decision model for the first communication site, the method further includes:

[0027] adjusting the initial model according to the communication decision models of the plurality of first communication sites;

[0028] Return to the step of training the initial model based on the first decision sample.

[0029] In a second aspect, an embodiment of the present disclosure provides an electronic device, comprising:

[0030] one or more processors;

[0031] a memory having one or more programs stored thereon, which, when executed by the one or more processors, causes the one or more processors to perform any of the above-mentioned methods for training a communication decision model;

[0032] One or more I / O interfaces are connected between the processor and the memory, and are used to implement signal interaction between the processor and the memory.

[0033] In a third aspect, an embodiment of the present disclosure provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned methods for training a communication decision model.

[0034] In the method, electronic device, and computer-readable medium for training a communication decision model according to the embodiments of the present disclosure, each communication site uses the decision samples of that communication site to train the model, thereby avoiding the difficulty in convergence when using the decision samples of all communication sites to train a unified model, as well as the problem of poor decision results for communication sites with fewer decision samples due to large differences in the number of decision samples of different communication sites. At the same time, the first training model of the first communication site is modified using the model modification values of the training models of other communication sites, thereby avoiding the problem of the first training model entering a local optimal solution due to the small number of training samples when the first decision samples are small. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In the accompanying drawings of the embodiments of the present disclosure:

[0036] Figure 1 A flowchart of a method for training a communication decision model provided in an embodiment of the present disclosure;

[0037] Figure 2A flowchart of some steps in another method for training a communication decision model provided by an embodiment of the present disclosure;

[0038] Figure 3 A flowchart of another method for training a communication decision model provided by an embodiment of the present disclosure;

[0039] Figure 4 A flowchart of some steps in another method for training a communication decision model provided by an embodiment of the present disclosure;

[0040] Figure 5 A flowchart of a method for training a communication decision model provided in an embodiment of the present disclosure;

[0041] Figure 6 A block diagram of an electronic device according to an embodiment of the present disclosure;

[0042] Figure 7 A block diagram of the composition of a computer-readable medium provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0043] In order to enable those skilled in the art to better understand the technical solutions of the embodiments of the present disclosure, the method for training a communication decision model, electronic device, and computer-readable medium provided by the embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0044] The embodiments of the present disclosure will be described more fully below with reference to the accompanying drawings, but the illustrated embodiments may be embodied in different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully enable those skilled in the art to understand the scope of this disclosure.

[0045] The accompanying drawings of the embodiments of the present disclosure are used to provide a further understanding of the embodiments of the present disclosure and constitute a part of the specification. Together with the embodiments of the present disclosure, they are used to explain the present disclosure and do not constitute a limitation of the present disclosure. By describing the detailed exemplary embodiments with reference to the accompanying drawings, the above and other features and advantages will become more apparent to those skilled in the art.

[0046] The embodiments of the present disclosure may be described with reference to plan views and / or cross-sectional views by way of ideal schematic views of the present disclosure. Therefore, the exemplary illustrations may be modified according to manufacturing techniques and / or tolerances.

[0047] In the absence of conflict, the various embodiments of the present disclosure and the various features therein may be combined with each other.

[0048] The terms used in this disclosure are only used to describe specific embodiments and are not intended to limit the disclosure. As used in this disclosure, the term "and / or" includes any and all combinations of one or more related enumerated items. As used in this disclosure, the singular forms "a" and "the" are also intended to include plural forms, unless the context clearly indicates otherwise. As used in this disclosure, the terms "comprising" and "made of" specify the presence of the features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or groups thereof.

[0049] Unless otherwise defined, all terms (including technical and scientific terms) used in this disclosure have the same meanings as those commonly understood by those skilled in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and this disclosure, and will not be interpreted as having an idealized or overly formal meaning unless expressly defined in this disclosure.

[0050] The embodiments of the present disclosure are not limited to the embodiments shown in the drawings, but include modifications of the configurations formed based on the manufacturing process. Therefore, the regions illustrated in the drawings are schematic in nature, and the shapes of the regions shown in the drawings illustrate the specific shapes of the regions of the elements, but are not intended to be limiting.

[0051] There are a large number of communication sites (such as base stations and specific communication equipment) in the communication network. These communication sites are used to carry out corresponding communication services. In the process of carrying out communication services, some "decisions" (i.e., deciding "how" to carry out communication services) may need to be made, such as making decisions on the direction of business processes and user strategies.

[0052] In some related technologies, communication sites use fixed logic for decision-making. Different communication sites use the same logic, but can set different logical parameters. Because the conditions of different communication sites vary greatly, the same logic may not be suitable for the specific conditions of each communication site. Moreover, because the logical parameters are manually set, it takes a long time of statistical analysis to find relatively accurate logical parameters.

[0053] In some related technologies, communication sites can use machine learning methods to make decisions, collect sample data on which decisions have been made from multiple communication sites, and use machine learning methods to learn these data from multiple communication sites to form corresponding machine learning models. Each communication site makes decisions based on the obtained machine learning model.

[0054] Since there may be large differences between different communication sites, their corresponding sample data will also have large differences. The essence of machine learning is to learn from sample data. When the number of sample data between different communication sites varies greatly, it may cause the machine learning model to hardly learn the characteristics of the small number of sample data, resulting in poor decision-making results for the communication sites with small amounts of sample data.

[0055] Furthermore, due to the significant differences between communication sites, if one hopes to train a model that performs effectively across all communication sites, the resulting machine learning model may be complex and difficult to converge during training. Furthermore, a more complex machine learning model is inherently larger than a simpler one, and therefore requires more transmission resources to transmit.

[0056] First, refer to Figure 1 , an embodiment of the present disclosure provides a method for training a communication decision model, where the communication decision model is used for decision-making of multiple communication sites in a communication network.

[0057] Each communication site corresponds to a service node, which is used to process the services of the communication site, such as making decisions. Multiple communication sites correspond to a central node, which is used to process services between different communication sites, such as data exchange between different communication sites.

[0058] Different business nodes can be different devices or located at different locations, such as the business node is its corresponding communication site; different business nodes can also be the same device or located at the same location, such as the business node and the central node are the same server, that is, one server completes the entire process of the method of the embodiment of the present disclosure.

[0059] In other words, the various steps of the method of the embodiment of the present disclosure may be performed in different devices or in one device. When different steps are performed in different devices, different data transmission processes may be required. However, as long as the substantive process of the various steps of the method of the embodiment of the present disclosure is included, it falls within the protection scope of the embodiment of the present disclosure.

[0060] The method of the embodiment of the present disclosure specifically includes:

[0061] S101. Train an initial model according to a first decision sample to adjust model parameters of the initial model to obtain a first training model; wherein the first decision sample is a sample of a decision that has been made by the first communication site.

[0062] The service node corresponding to the first communication site collects the first decision samples of the first communication site. After collecting a certain number of first decision samples, the initial model is trained according to the first decision samples, or the model parameters of the initial model are adjusted according to the first decision samples to obtain a first training model.

[0063] The first decision sample is a sample of a decision made by the first communication site, which may specifically include feature information (or index information), the decision made, and the effect of the decision. The initial model is sent from the central node to each service node. It may specifically be the model structure and corresponding model parameters of a machine learning model (such as an artificial intelligence model) that has been pre-trained using big data. The initial model is the same for different communication sites.

[0064] S102. Obtain at least one second model parameter modification value of at least one second training model; wherein each second training model is obtained by training the initial model based on a second decision sample of a second communication site, and the second decision sample is a sample of a decision already made by the second communication site; each second model parameter modification value reflects at least a partial modification of the model parameters of the second training model relative to the model parameters of the initial model.

[0065] The central node obtains at least one modified value of a second model parameter of a second training model at at least one second communication site and sends the modified value to a service node corresponding to the first communication site. Of course, if the central node and the service node are the same communication site or device, the subsequent processing is performed directly without the sending step.

[0066] It should be understood that in the implementation of the present disclosure, the second communication site is relative to the first communication site, that is, when the method of the embodiment of the present disclosure is executed with a certain communication site as the first communication site, the other communication sites are all second communication sites.

[0067] That is to say, the first communication site and the second communication site are relative to the process of training the communication decision model, rather than specifically referring to one or some communication sites. That is, the first communication site in the process of training the communication decision model once may be the second communication site in the process of training the communication decision model next time. Similarly, the second communication site in the process of training the communication decision model once may be the first communication site in the process of training the communication decision model next time.

[0068] Similar to the first communication site, the business node corresponding to the second communication site collects the second decision samples of the second communication site, and trains the initial model based on the second decision samples (that is, the samples on which decisions have been made at the second communication site), or adjusts the model parameters of the initial model to obtain a second training model (for the second communication site itself, it is also the "first communication site", and this model is its own "first training model)".

[0069] The adjustment of the model parameters of the initial model based on the second decision sample, i.e., the modification of the model parameters of the second training model relative to the model parameters of the initial model, is referred to as the second model parameter modification value. Each second training model has at least one second model parameter modification value. Each second model parameter modification value represents a partial modification of the model parameters of the second training model relative to the model parameters of the initial model, and all second model parameter modification values represent a complete modification of the model parameters of the second training model relative to the model parameters of the initial model.

[0070] S103: Adjust the model parameters of the first training model according to at least part of the second model parameter modification value to obtain a communication decision model for the first communication site.

[0071] The service node corresponding to the first communication site adjusts the model parameters of the first training model obtained through training according to at least part of the acquired second model parameter modification value, and finally obtains a communication decision model for communication decision of the first communication site.

[0072] In the method for training a communication decision model in the embodiment of the present disclosure, each communication site uses the decision samples of that communication site to train the model, thereby avoiding the problem of difficulty in convergence when using the decision samples of all communication sites to train a unified model, as well as the problem of poor decision results for communication sites with fewer decision samples due to large differences in the number of decision samples of different communication sites. At the same time, the embodiment of the present disclosure also modifies the first training model of the first communication site using the model modification values of the training models of other communication sites, thereby avoiding the problem of the first training model entering a local optimal solution due to the small number of training samples when the first decision samples are small.

[0073] Reference Figure 3 In some embodiments, obtaining at least one second model parameter modification value of at least one second training model (S102) includes:

[0074] S1021: Filter the second model parameter modification values of all second training models according to a predetermined filtering rule, and obtain the second model parameter modification values of the second training models that pass the filtering.

[0075] After receiving the second model parameter modification values sent by each business node, the central node filters the second model parameter modification values according to certain rules, such as filtering the second model parameter modification values based on the acceptance of the second model parameter modification values of the business node by other business nodes during the previous training communication decision model (that is, whether to modify its first training model according to the second model parameter modification values of the business node).

[0076] A specific method for determining whether to filter the modified value of the second model parameter may be clustering, calculating the average sample distance, calculating the overlap ratio of the mean square error of the sample distance, and the like.

[0077] By filtering, the second model parameter modification value can be reduced. When the central node and the service node are not in the same location and data needs to be transmitted, the waste of transmission resources can be reduced.

[0078] Reference Figure 2 In some embodiments, the modified value of the second model parameter of each second training model is obtained by:

[0079] S201: Cluster the second decision samples of the second training model.

[0080] The service node corresponding to the second communication site clusters the second decision samples used to train the initial model in the second communication site, and clusters the second decision samples into several categories with similar characteristics.

[0081] The clustering method and clustering parameters used are not limited and can be flexibly selected according to actual conditions, such as uniformly clustering the second decision samples according to density based on the feature information of the second decision samples, setting the number of clusters according to the decision effect of the second decision samples, clustering the second decision samples based on the initial clustering points, etc.

[0082] S202: Determine the sum of modifications of all second decision samples in each cluster as a second model parameter modification value, where the modification of each second decision sample is the modification of the model parameter caused when the initial model is trained according to the second decision sample.

[0083] When training the initial model based on the second decision samples, the change in model parameters before and after each second decision sample is input into the initial model is the modification of the second decision sample. A second model parameter modification value is determined based on the modification of each second decision sample in a cluster. For example, the sum of the modifications of each second decision sample in each cluster is the second model parameter modification value, or the average of the modifications of each second decision sample in each cluster is the second model parameter modification value.

[0084] The service node corresponding to the second communication site determines a plurality of second model parameter modification values according to the second decision samples in each cluster and sends them to the service node corresponding to the first communication site through the central node.

[0085] Each cluster corresponds to a second model parameter modification value. On the one hand, this can reduce the number of second model parameter modification values, that is, reduce the amount of calculation in subsequent steps. On the other hand, the second decision samples of the same cluster have similar characteristics, so their modifications to the model parameters of the initial model are also similar. The sum or average value of the modifications of each second decision sample can represent the modifications of these decision samples to the model parameters of the initial model. If the second decision samples do not have similar characteristics, their modifications to the model parameters of the initial model may be very different. Using the sum or average value of the modifications cannot replace the actual modifications.

[0086] Reference Figure 3 In some embodiments, adjusting the model parameters of the first training model according to at least part of the second model parameter modification value to obtain the communication decision model for the first communication site (S103) includes:

[0087] S1031. Calculate, based on the first decision sample, the decision effect of the model after adjusting the first training model according to each second model parameter modification value, to determine the accepted second model parameter modification value.

[0088] The business node corresponding to the first communication site verifies the multiple second model parameter modification values obtained one by one to obtain their decision effects. If the decision effect meets certain conditions, the second model parameter modification value is determined to be an accepted second model parameter modification value. If not, the second model parameter modification value can be discarded.

[0089] Specifically, the first training model is modified according to a modified value of a second model parameter to obtain an adjusted first training model. The feature information of the first decision sample is input into the adjusted first training model to obtain a decision result. The obtained decision result is then compared with the decision of the first decision sample to obtain the decision effect of the adjusted first training model. The specific comparison method is not limited and can be used according to actual circumstances, such as averaging, confidence calculation, and statistical key indicator calculation.

[0090] If the decision effect meets certain conditions (such as being higher than a certain threshold), the second model parameter modification value is determined to be the accepted second model parameter modification value. According to the same method, all second model parameter modification values are verified to determine the accepted second model parameter modification value.

[0091] S1032. Adjust the model parameters of the first training model according to the accepted second model parameter modification value to obtain a communication decision model for the first communication site.

[0092] The service node corresponding to the first communication site adjusts the model parameters of the first training model according to the accepted second model parameter modification value, such as making corresponding modifications to the first training model according to the accepted second model parameter modification value to obtain a communication decision model for the first communication site.

[0093] Since the second model parameter modification value is a modification of the initial model by the second decision sample, the second decision sample may be different from the first decision sample, or the sample of the first communication site. Therefore, the second model parameter modification value is not always suitable for the first training model of the first communication site. By screening the second model parameter modification value according to the first decision sample, the second model parameter modification value that is not suitable for the first training model can be eliminated, so that the obtained communication decision model is more suitable for the first communication site.

[0094] Reference Figure 4 In some embodiments, adjusting the model parameters of the first training model according to the accepted second model parameter modification value to obtain the communication decision model for the first communication site (S1032) includes:

[0095] S10321. If there are multiple accepted second model parameter modification values, the model parameters of the first training model are adjusted according to each accepted second model parameter modification value and its corresponding weight. The weight corresponding to each accepted second model parameter modification value is calculated based on the number of second decision samples corresponding to the second model parameter modification value.

[0096] If the service node corresponding to the first communication site determines multiple accepted second model parameter modification values, the number of second decision samples in the cluster corresponding to each second model parameter modification value is calculated, and the ratio of the number of second decision samples corresponding to these multiple second model parameter modification values is calculated. The first training model is modified according to the second model parameter modification value using the ratio as a weight to obtain a communication decision model for the first communication site.

[0097] If there are three second model parameter modification values, and the ratio of the number of corresponding second decision samples is 3:4:3, then the first training model is modified according to 30% of the first second model parameter modification value, 40% of the second second model parameter modification value, and 30% of the third second model parameter modification value to obtain a communication decision model for the first communication site.

[0098] When there are abnormal samples, due to their small number, the first training model parameters are modified according to the sample number ratio, and the modification to the model will become correspondingly small, avoiding a large impact on the model and ultimately affecting the decision-making effect of the model.

[0099] Reference Figure 4 In some embodiments, adjusting the model parameters of the first training model according to the accepted second model parameter modification value to obtain the communication decision model for the first communication site (S1032) includes:

[0100] S10322. Adjust the first training model according to the accepted second model parameter modification value to obtain a first pre-training model.

[0101] The service node corresponding to the first communication site adjusts the model parameters of the first training model according to the accepted second model parameter modification value, such as modifying the first training model accordingly according to the accepted second model parameter modification value to obtain a first pre-trained model.

[0102] S10323. Calculate the decision effect of the first pre-training model based on the first decision sample.

[0103] The service node corresponding to the first communication site inputs the feature information of the first decision sample into the first pre-training model to obtain a decision result, compares the obtained decision result with the decision of the first decision sample, and obtains the decision effect of the first pre-training model.

[0104] The specific comparison method is not limited, and different comparison methods can be used according to actual conditions, such as the average method, confidence calculation, and key indicator calculation based on statistics.

[0105] S10324: If the decision result does not meet the preset conditions, determine the first training model as the communication decision model for the first communication site.

[0106] If the decision effect does not meet the preset conditions (such as being lower than the preset threshold), the accepted second model parameter modification value adjustment is discarded, that is, the model parameters of the first training model are not adjusted according to the accepted second model parameter modification value, and the first training model is used as the communication decision model for the first communication site.

[0107] Alternatively, the modification ratio of the accepted second model parameter modification value to the first training model is reduced, such as modifying the first training model according to 80% of the accepted second model parameter modification value to obtain a modified model and verifying the decision effect of the modified model. If the preset conditions are met, the first training model is modified according to the ratio and the received second model parameter modification value; if the preset conditions are not met, the modification and verification are continued with the ratio reduced until the decision effect of the modified model meets the preset conditions or reaches the minimum ratio and the accepted second model parameter modification value is discarded.

[0108] Verifying the first training model, i.e., the first pre-training model, after adjusting the second model parameter modification value accepted according to the first decision sample can ensure that the communication decision model finally obtained for the first communication site has a good decision effect at least on the first decision sample.

[0109] Reference Figure 3 In some embodiments, after obtaining the communication decision model for the first communication site (S1032), the process further includes:

[0110] S1033. Adjust the clustering method of the second decision samples according to the second model parameter modification values accepted by the multiple first communication sites.

[0111] After determining the accepted second model parameter modification value, the business node corresponding to the first communication site sends the specific acceptance status of the second model parameter modification value (such as whether it is accepted, the acceptance ratio, etc.) to the central node. The central node adjusts the clustering method of the second decision sample according to the specific acceptance status of the second model parameter modification values sent by multiple business nodes and sends it to the business node.

[0112] For example, the specific acceptance of the modified value of the second model parameter and its corresponding clustering method are used as input to perform machine learning (such as deep neural networks, decision trees, etc.) to adjust the clustering method of the second decision samples of each business node. There is no limitation on the machine learning method, and different methods can be flexibly selected according to the specific situation.

[0113] The modified sum or average of each second decision sample in each cluster is a modified second model parameter value. The modified second model parameter value is related to the clustering method and clustering parameters of the second decision sample. Therefore, the acceptance of the modified second model parameter value also reflects the quality of the clustering method and clustering parameters of the second decision sample. Based on the modified second model parameter value, the clustering method and clustering parameters of the second decision sample are adjusted to determine the appropriate clustering method and clustering parameters for the second decision sample.

[0114] Reference Figure 3In some embodiments, after obtaining the communication decision model for the first communication site (S103), the method further includes:

[0115] S1041. Adjust the initial model according to the communication decision models of multiple first communication sites.

[0116] The central node obtains the model parameters of the communication decision model of each service node and can adjust the initial model according to these model parameters and the second model parameter modification value of each service node.

[0117] For example, when the decision-making effect of most business nodes is not good, the central node judges the proportion of the influence of each part of the model structure of the initial model on the final decision effect (such as the influence of the connection weights of different parts of the neural network on the decision effect) based on the difference between the model parameters of the initial model and the model parameters of multiple communication decision models, and manually decides whether to remove the model structure that has less influence on the final decision effect.

[0118] The second decision samples and the corresponding second model parameter modification values are counted to assist manual decision-making in increasing the model structure of the initial model, that is, making part of the model structure of the initial model more complex.

[0119] Specifically, all second model parameter modification values and the corresponding number of second decision samples are obtained, and the distribution of the second decision samples is counted.

[0120] The model structure for densely distributed second decision samples may need to be more complex to enable the model to more precisely distinguish densely distributed second decision samples and achieve better decision results. For example, the number of rules or thresholds in the decision tree corresponding to densely distributed second decision samples may be increased, or the number of intermediate layer node connections corresponding to densely distributed second decision samples may be increased in the neural network. The model structure for sparsely distributed second decision samples does not need to be more complex.

[0121] It should be noted that when increasing the complexity of the model structure, the computing power burden of the business node should be taken into consideration. The computing power used in the decision-making process of the initial model after adding the model structure should be maintained below the original load level of the business node or the preset maximum computing power limit.

[0122] S1042: Return to the step of training the initial model according to the first decision sample.

[0123] After the central node completes the model structure adjustment and forms a new initial model, it sends the new initial model to the business node.

[0124] All service nodes use the same initial model. Of course, an initial model corresponding to each service node can also be formed separately for different situations of different service nodes.

[0125] By removing model structures that have little impact on the final decision-making effect, the model parameters are reduced and the requirements for decision-making computing power are lowered; and by increasing the complexity of the model structure corresponding to the densely distributed second decision samples, the sensitivity of the structure at specific positions in the model structure of the initial model can be enhanced, so as to more finely distinguish the densely distributed decision samples and achieve better decision-making results. Specific embodiment 1:

[0127] Reference Figure 5 The method for training a communication decision model according to an embodiment of the present disclosure can train a communication decision model for deciding whether a terminal should switch to a neighboring cell, which specifically includes:

[0128] A01. Collect sample data.

[0129] Sample data is collected in each cell, and the sample data includes the decisions made and the effects of the decisions. The decisions made include switching to a neighboring cell and not switching to a neighboring cell. The effects of the decisions can be obtained through an evaluation function. The effects of the decisions corresponding to the samples of switching to a neighboring cell can be obtained by inputting whether the cell is subsequently switched back to the current cell, the time interval between switching back to the current cell and switching out of the current cell, and the feedback after switching out of the current cell into the evaluation function. The effects of the decisions corresponding to the samples of not switching to a neighboring cell can be obtained by inputting whether an abnormal interruption occurs in the communication process, the duration of communication before the abnormal interruption, the throughput during the communication period, and the corresponding power / scheduling cost into the evaluation function.

[0130] A02. Train the initial model.

[0131] A deep neural network model is established in each cell. The structure of each deep neural network model is the same, but the parameters can be different.

[0132] The deep neural network model in each cell performs local learning based on the collected sample data to obtain the training model corresponding to the cell, and records the model parameter modification values of the model parameters of the initial model for each sample data.

[0133] A03. Sample clustering.

[0134] Taking a base station as a unit, sample data of cells included in the base station and corresponding model parameter modification values are collected, and the collected sample data are clustered using a k-center point algorithm.

[0135] After the corresponding clusters are obtained, the model parameter modification values corresponding to all sample data in each cluster are calculated as the arithmetic average and used as the model parameter modification value corresponding to the cluster.

[0136] A04. Obtain a communication decision model.

[0137] Each cell in the base station incorporates the modified model parameter values into its own training model based on the modified model parameter values of the base station and the number of sample data corresponding to the cluster. The specific process can be: all the modified model parameter values corresponding to the base station are incorporated into its own training model to obtain an adjusted model, and the decision effect of the adjusted model is verified based on the sample data of its own cell. If the decision effect is not good, all the modified model parameter values are not accepted and the model is restored to the original training model; if the decision effect is good, the modified model parameter values are incorporated into its own training model based on the ratio of the number of sample data corresponding to each model parameter modification value.

[0138] If a large proportion of all cells of the base station adopt the base station's model parameter modification values, all model parameter modification values of the base station are sent to the central node, and of course can also be sent to the EMS (element management system) or other background nodes.

[0139] The central node forwards all model parameter modification values of the base station to other base stations, so that cells in other base stations can process the base station model parameter modification values using a method similar to that of the cell of the base station.

[0140] Similarly, each cell in the base station will also receive model parameter modification values sent by other base stations, and use the same method to process the received model parameter modification values and make corresponding adjustments to the training model to obtain the communication decision model corresponding to the cell.

[0141] A05. Adjust the initial model.

[0142] The central node collects the models of each cell of each base station, including the average value and mean square error information of each model parameter, and refers to the reported cluster sample information to prune and supplement the deep neural network, and sends the pruned and supplemented deep neural network model to each cell of each base station.

[0143] At the same time, the central node adjusts the clustering method of the sample data according to the specific acceptance of the model parameter modification values of all base stations (such as whether it is accepted, the acceptance ratio, etc.), and also sends the modified clustering method to each cell of each base station. Specific embodiment 2:

[0145] Different from the specific embodiment 1, the communication decision model of this embodiment is used for query prediction, that is, when the current query used by the user or the previous queries are known, the user's next query or the next few queries are predicted and pre-queried so that the results can be returned as soon as possible when the user uses it. Figure 5 , the specific steps of the embodiment of the present disclosure include:

[0146] B01. Collect sample data.

[0147] Collect the query commands of each user's application on the data source. Generally speaking, the query command format used by the application has a paradigm. All query paradigms of the application can be summarized through pattern matching, and the index value of the corresponding query can be extracted by analyzing the range clause. The query paradigm and index value are the sample data.

[0148] Each query performed by a user can be considered as a sample for learning, and the paradigm and index value of the user's next query can be used as annotations for the predicted results to judge the quality of the decision results.

[0149] B02. Train the initial model.

[0150] The initial model for each user's business node consists of two Bayesian belief networks: one for predicting the next query pattern and the other for predicting the query index. These two Bayesian belief networks are trained based on the collected sample data to generate training models, and the network parameter modifications for each Bayesian belief network are recorded for each sample.

[0151] Among them, the Bayesian belief network used to predict the paradigm of the next query makes decisions based on the paradigms and viewing order of the user's previous views, and can also record the values of certain fields of the previous few results of the current query.

[0152] After determining the paradigm of the next query, the Bayesian belief network used to predict the index value of the query predicts the index value based on this paradigm. In addition to the paradigms of the previous queries, the input of the Bayesian belief network also includes the index values used in the previous queries and the values of certain fields of the first few results of the previous queries.

[0153] B03. Sample clustering.

[0154] The sample data corresponding to each user is clustered using the k-center point algorithm. The index used for sample clustering can be the query type and query index information of the previous queries or the user's next query type, etc.

[0155] After the corresponding clusters are obtained, the model parameter modification values corresponding to all sample data in each cluster are calculated as the arithmetic average and used as the model parameter modification value corresponding to the cluster.

[0156] B04. Obtain the communication decision model.

[0157] The service node corresponding to each user sends all model parameter modification values to the central node, and of course can also be sent to the EMS (element management system) or other background nodes.

[0158] The central node forwards all model parameter modification values corresponding to the user to the business nodes corresponding to other users. Each business node corresponding to other users merges the model parameter modification values into its own training model based on the model parameter modification values corresponding to the user and the number of sample data corresponding to its cluster. The specific process can be: all model parameter modification values corresponding to the user are merged into its own training model to obtain an adjusted model, and the decision effect of the adjusted model is verified based on the sample data of its own user. If the decision effect is not good, all model parameter modification values are not accepted and the model is restored to the original training model; if the decision effect is good, the model parameter modification values are merged into its own training model based on the ratio of the number of sample data corresponding to each model parameter modification value.

[0159] Similarly, the service node corresponding to the user will also receive the model parameter modification values sent by the service nodes corresponding to other users, and use the same method to process the received model parameter modification values and make corresponding adjustments to the training model to obtain the communication decision model corresponding to the cell.

[0160] B05. Adjust the initial model.

[0161] The central node collects the communication decision model corresponding to each user, including the average value and mean square error information of each model parameter. At the same time, it refers to the reported cluster sample information to adjust the input items and the intermediate judgment nodes, and sends the adjusted model to the user node corresponding to each user.

[0162] At the same time, the central node adjusts the clustering method of the sample data according to the specific acceptance of all model parameter modification values corresponding to all users (such as whether to accept, the acceptance ratio, etc.), and also sends the modified clustering method to the user nodes corresponding to each user.

[0163] Secondly, refer to Figure 6 , an embodiment of the present disclosure provides an electronic device, comprising:

[0164] one or more processors,

[0165] A memory having one or more programs stored thereon, which, when executed by one or more processors, causes the one or more processors to implement any one of the above methods for training a communication decision model;

[0166] One or more I / O interfaces are connected between the processor and the memory to implement information exchange between the processor and the memory.

[0167] Among them, the processor is a device with data processing capabilities, including but not limited to the central processing unit (CPU); the memory is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically such as SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read-write interface) is connected between the processor and the memory to realize information exchange between the memory and the processor, including but not limited to the data bus (Bus), etc.

[0168] Thirdly, refer to Figure 7 , an embodiment of the present disclosure provides a computer-readable medium having a computer program stored thereon, which implements any of the above-mentioned methods for training a communication decision model when the program is executed by a processor.

[0169] Among them, the processor is a device with data processing capabilities, including but not limited to the central processing unit (CPU); the memory is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically such as SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read-write interface) is connected between the processor and the memory, which can realize information exchange between the memory and the processor, including but not limited to the data bus (Bus), etc.

[0170] Those skilled in the art will appreciate that all or some of the steps, systems, and functional modules / units in the apparatus disclosed above may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0171] In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be performed by several physical components in cooperation.

[0172] Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit (CPU), a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or temporary media). As is well known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, random access memory (RAM, more specifically SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory (FLASH) or other disk storage; compact disc (CD-ROM), digital versatile disc (DVD) or other optical disc storage; magnetic cassettes, tapes, disk storage or other magnetic storage; any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0173] The present disclosure has disclosed example embodiments, and although specific terms are employed, they are used and should be interpreted only in a general illustrative sense and not for purposes of limitation. In some instances, it will be apparent to those skilled in the art that, unless otherwise expressly indicated, features, characteristics, and / or elements described in conjunction with a particular embodiment may be used alone or in combination with features, characteristics, and / or elements described in conjunction with other embodiments. Therefore, it will be understood by those skilled in the art that various changes in form and detail may be made without departing from the scope of the present disclosure as set forth in the appended claims.

Claims

1. A method for training a communication decision model, wherein the communication decision model is used for communication site decision making, the method comprising: Training an initial model based on a first decision sample to adjust model parameters of the initial model to obtain a first training model; wherein the first decision sample is a sample of a decision already made by the first communication site; the first decision sample includes at least one of the following: feature information, index information, a decision made, and an effect of the decision; Obtaining at least one second model parameter modification value of at least one second training model; wherein each second training model is obtained by training the initial model based on a second decision sample of a second communication site, the second decision sample being a sample of a decision already made by the second communication site; and each second model parameter modification value reflects at least a partial modification of a model parameter of the second training model relative to a model parameter of the initial model; adjusting model parameters of the first training model according to at least part of the second model parameter modification values to obtain a communication decision model for the first communication site; The step of adjusting the model parameters of the first training model according to at least part of the second model parameter modification values to obtain a communication decision model for the first communication site includes: Calculating, based on the first decision sample, a decision effect of a model adjusted according to each second model parameter modification value on the first training model to determine an accepted second model parameter modification value; Adjusting the model parameters of the first training model according to the accepted second model parameter modification value to obtain a communication decision model for the first communication site; The modified value of the second model parameter of each second training model is obtained by: clustering the second decision samples of the second training model; The sum of modifications of all second decision samples in each cluster is determined to be a second model parameter modification value, and the modification of each second decision sample is the modification of the model parameter caused when the initial model is trained according to the second decision sample.

2. The method according to claim 1, wherein The step of adjusting the model parameters of the first training model according to the accepted second model parameter modification value to obtain a communication decision model for the first communication site includes: If there are multiple accepted second model parameter modification values, the model parameters of the first training model are adjusted according to each accepted second model parameter modification value and its corresponding weight, and the weight corresponding to each accepted second model parameter modification value is calculated based on the number of second decision samples corresponding to the second model parameter modification value.

3. The method according to claim 1, wherein The step of adjusting the model parameters of the first training model according to the accepted second model parameter modification value to obtain a communication decision model for the first communication site includes: Adjust the first training model according to the accepted second model parameter modification value to obtain a first pre-training model; Calculating a decision effect of the first pre-trained model according to the first decision sample; If the decision effect does not meet the preset conditions, the first training model is determined to be the communication decision model for the first communication site.

4. The method according to claim 1, wherein After obtaining the communication decision model for the first communication site, the method further includes: The clustering manner of the second decision samples is adjusted according to the second model parameter modification values accepted by the plurality of first communication sites.

5. The method according to claim 1, wherein The obtaining of at least one second model parameter modification value of at least one second training model comprises: The second model parameter modification values of all second training models are filtered according to a predetermined filtering rule, and the second model parameter modification values of the second training models that pass the filtering are obtained.

6. The method according to claim 1, wherein After obtaining the communication decision model for the first communication site, the method further includes: adjusting the initial model according to the communication decision models of the plurality of first communication sites; Return to the step of training the initial model based on the first decision sample.

7. An electronic device comprising: one or more processors; a memory having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method for training a communication decision model according to any one of claims 1 to 6; One or more I / O interfaces are connected between the processor and the memory, and are used to implement information exchange between the processor and the memory.

8. A computer-readable medium having a computer program stored thereon, wherein when the program is executed by a processor, the method for training a communication decision model according to any one of claims 1 to 6 is implemented.

Images