Prediction model generation device and prediction model generation method

Abstract

This prediction model generation device (100) comprises: a feature vector calculation unit (111) that calculates a simulation feature vector which is a feature vector of input information when performance information about a communication node simulation model is acquired, and an actual communication node feature vector which is a feature vector of input information when performance information about the actual communication node that is an actual machine is acquired; and an actual communication node prediction model generation unit (114) that performs transfer training on a simulation prediction model (121) which predicts performance information about the simulation model on the basis of the simulation feature vector, and generates the simulation prediction model (121) which predicts performance information about the actual communication node on the basis of the actual communication node feature vector.

Description

Prediction model generation device and prediction model generation method

[0001] The present invention relates to a prediction model generation device and a prediction model generation method for generating a prediction model used for predicting the performance of a communication node.

[0002] In conventional network simulations, circuits and programs identical to the network functions of the actual equipment are prepared on the simulator and used for simulation (see Non-Patent Document 1). This requires a long development time and is a factor in increasing costs. In addition, when performing simulations, circuits and programs that reproduce the internal workings of network nodes operate on a packet-by-packet basis, which increases the time required to complete the simulation.

[0003] In response to these challenges, a network node modeling method has been proposed as a simple method for modeling network nodes, in which the performance of a network node is predicted by machine learning using a set of network node configuration information, input traffic information, and the resulting network node performance information as learning data (see Non-Patent Document 2). Note that this method estimates C information (performance value) based on A information + B information, then estimates D information (performance value) based on A information + B information + C information, and the obtained performance value is incorporated into the input for the next estimation, and further estimation is repeated to estimate (predict) the performance value in a chain-like manner.

[0004] Network simulator QualNet / EXata, [online], Structural Planning Institute, [Retrieved July 10, 2023], Internet<URL:https: / / network2.kke.co.jp / wireless-products / qualnet / > Kyota Hattori, et al., "Recursive Router Metrics Prediction Using ML-based Node Modeling for Network Digital Replica," 2022 IEEE Global Communications Conference (GLOBECOM 2022), pp.1006-1012.

[0005] In the method described in Non-Patent Document 2, the estimation accuracy (prediction accuracy) of network node performance can be improved by adding estimated values ​​of highly correlated indicators to training data. However, due to the nature of learning by estimating (predicting) performance values ​​in a chain-like manner, the estimated values ​​contain errors, and adding the estimated performance values ​​to the training data does not necessarily improve the estimation accuracy. For this reason, it is necessary to consider a combination of estimated values ​​that improves estimation accuracy, taking into account the contribution of each performance value, etc., which creates a problem of complex operation. The present invention has been made in light of this background, and an object of the present invention is to enable the accuracy of network node performance prediction to be improved.

[0006] In order to solve the above-mentioned problems, the prediction model generation device of the present invention includes a feature vector calculation unit that calculates a simulation feature vector, which is a feature vector of input information when performance information of a simulation model of a communication node is acquired, and a real communication node feature vector, which is a feature vector of input information when performance information of a real communication node, which is an actual device, is acquired; a simulation prediction model generation unit that uses learning data in which the simulation feature vectors are explanatory variables and the performance information of the simulation model is a target variable to generate a simulation prediction model that predicts the performance information of the simulation model; and a real communication node prediction model generation unit that performs transfer learning of the simulation prediction model using learning data in which the real communication node feature vectors are explanatory variables and the performance information of the real communication node is a target variable to generate a real communication node prediction model that predicts the performance information of the real communication node.

[0007] According to the present invention, it is possible to improve the accuracy of network node performance prediction.

[0008] FIG. 1 is a functional block diagram of a prediction model generation device according to the present embodiment. FIG. 2 is a diagram for explaining a simulation prediction model according to the present embodiment. FIG. 3 is a diagram for explaining a prediction model for actual communication nodes according to the present embodiment. FIG. 4 is a flowchart of a prediction model generation process according to the present embodiment. FIG. 5 is a scatter plot showing the relationship between actual measured values ​​of delay of actual communication nodes and predicted values ​​using a conventional method. FIG. 6 is a scatter plot showing the relationship between actual measured values ​​of delay of actual communication nodes and predicted values ​​using a prediction model for actual communication nodes according to the present embodiment. FIG. 7 is a hardware configuration diagram showing an example of a computer that realizes the functions of the prediction model generation device according to the above-mentioned embodiment.

[0009] Overview of Prediction Model Generation Device A prediction model generation device that generates a performance prediction model (also simply referred to as a prediction model) used to predict the performance of a network node in an embodiment for carrying out the present invention will be described. The prediction model generation device generates a prediction model for a real network node (actual communication node) by re-learning a prediction model for a simulation using transfer learning technology. The simulation prediction model and the actual communication node prediction model are machine learning models.

[0010] The explanatory variables of the simulation prediction model are feature vectors calculated using multivariate analysis based on the input data and setting data of the simulation. The objective variable of the simulation prediction model is performance data. The elements of the feature vector are principal components calculated using, for example, partial least squares. The prediction model generation device trains and generates the simulation prediction model using learning data created based on the input data and setting data of the simulation and the performance data resulting from the simulation.

[0011] The explanatory variables of the prediction model for the actual communication node are feature vectors calculated using multivariate analysis based on the input data and setting data used when measuring the performance of the actual communication node, and the specifications of the actual communication node. The objective variable of the prediction model for the actual communication node is performance data. The prediction model generation device creates learning data based on the input data, setting data, specifications, and performance data used when measuring the performance of the actual communication node. The prediction model generation device uses transfer learning technology to retrain the simulation prediction model using the learning data to generate a prediction model for the actual communication node.

[0012] The prediction model generation device identifies the dimensions of the feature vectors that serve as explanatory variables for the simulation prediction model and the feature vectors that serve as explanatory variables for the actual communication node prediction model, such that the dimensions are the same and the distance between the two feature vectors is minimized. By doing so, it is assumed that both feature vectors extract similar features from the input data and setting data, and transfer learning is expected to be highly effective.

[0013] Compared to performance measurements of actual communication nodes, performance measurements using simulations have the advantages of being able to obtain large amounts of data due to lower data collection costs, the absence of noise in the data, and the ability to obtain data in communication environments that cannot be measured with actual communication nodes. For this reason, simulation prediction models can be expected to have high performance prediction accuracy. Prediction models for actual communication nodes generated by transfer learning from such simulation prediction models can also be expected to have high performance prediction accuracy.

[0014] <Configuration of Prediction Model Generation Device> Fig. 1 is a functional block diagram of a prediction model generation device 100 according to this embodiment. The prediction model generation device 100 is a computer and includes a control unit 110, a storage unit 120, and an input / output unit 180. User interface devices such as a display, keyboard, and mouse are connected to the input / output unit 180. The input / output unit 180 may include a communication device, enabling data transmission and reception with a simulator or a device for measuring the performance of actual communication nodes. Furthermore, a media drive may be connected to the input / output unit 180, enabling data exchange using a recording medium.

[0015] <Prediction Model Generation Device: Storage Unit> The storage unit 120 is configured to include storage devices such as a ROM (Read Only Memory), a RAM (Random Access Memory), and an SSD (Solid State Drive). The storage unit 120 stores a simulation result database 130, an actual communication node measurement result database 140, a simulation prediction model 121, an actual communication node prediction model 122, and a program 128. The program 128 includes a description of the prediction model generation process (see FIG. 4), which will be described later.

[0016] <Storage Unit: Simulation Result Database> The simulation result database 130 stores information related to the results of a simulation. More specifically, the simulation result database 130 stores input data, setting data, and performance data related to a simulation using a simulation model of a real communication node. The input data is, for example, input traffic conditions, such as the total number of packets input to the simulation model, the average packet length, and the average input throughput. The setting data is setting information for the simulation model, such as the number of central processing units (CPUs), the size of a routing table, and the number of entries in an access control list (ACL). The performance data is performance information resulting from a simulation using the simulation model, such as throughput, packet loss rate, and delay.

[0017] <Storage Unit: Actual Communication Node Measurement Result Database> The actual communication node measurement result database 140 stores information related to the results of measurements (actual measurements) using actual communication nodes. More specifically, the actual communication node measurement result database 140 stores input data, setting data, specifications, and performance data related to actual measurements of actual communication nodes. The input data is, for example, input traffic conditions, such as the total number of packets input to the actual communication node, the average packet length, and the average input throughput. The setting data is setting information for the actual communication node, such as the size of the routing table and the number of ACL entries. The performance data is performance information obtained by actual measurements, such as throughput, packet loss rate, and delay.

[0018] <Storage Unit: Simulation Prediction Model> The simulation prediction model 121 is a machine learning model used to predict the performance of a simulation model. The objective variable of the simulation prediction model 121 is performance information (performance data) of the simulation model. The explanatory variables of the simulation prediction model 121 are feature vectors calculated based on input data and setting data of the simulation model. Details of the feature vectors will be described later.

[0019] 2 is a diagram illustrating a simulation prediction model 121 according to this embodiment. A feature vector calculation unit 111, which will be described later, calculates a feature vector 312 (a simulation feature vector, which will be described later) based on input data and setting data of the simulation model. A performance prediction unit 115, which will be described later, calculates performance information 313 (a target variable) using the simulation prediction model 121 with the feature vector 312 as an input (explanatory variable). The input data and setting data of the simulation model, which are the basis for calculating the feature vector 312, are also referred to as input information 311. It is assumed that a simulation prediction model 121 exists for each piece of performance information 313, such as throughput, packet loss rate, and delay.

[0020] <<Storage Unit: Prediction Model for Actual Communication Node>> The prediction model for actual communication node 122 is a machine learning model used to predict the performance of an actual communication node. The objective variable of the prediction model for actual communication node 122 is performance information (performance data) of the actual communication node. The explanatory variables of the prediction model for actual communication node 122 are feature vectors calculated based on the input data, specifications, and setting data of the actual communication node.

[0021] 3 is a diagram for explaining the prediction model 122 for an actual communication node according to this embodiment. The feature vector calculation unit 111, which will be described later, calculates a feature vector 322 (feature vector for an actual communication node, which will be described later) based on the input data, specifications, and setting data of the actual communication node. The performance prediction unit 115, which will be described later, calculates performance information 323 (objective variable) using the prediction model 122 for an actual communication node with the feature vector 322 as input (explanatory variable). The input data, specifications, and setting data of the actual communication node, which are the basis for calculating the feature vector 322, are also referred to as input information 321. It is assumed that a prediction model 122 for an actual communication node exists for each piece of performance information 323, such as throughput, packet loss rate, and delay.

[0022] <Prediction Model Generating Device: Control Unit> Returning to Fig. 1 , the control unit 110 will be described. The control unit 110 is configured to include a CPU, and is equipped with a feature vector calculation unit 111, a dimension calculation unit 112, a simulation prediction model generation unit 113, an actual communication node prediction model generation unit 114, and a performance prediction unit 115. The control unit 110 may be configured to include a GPU (Graphics Processing Unit), an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), etc.

[0023] <<Control Unit: Feature Vector Calculation Unit>> The feature vector calculation unit 111 calculates feature vectors that serve as explanatory variables for the simulation prediction model 121 based on the input data and setting data of the simulation model. For example, the feature vector calculation unit 111 calculates feature vectors for the input data and setting data using partial least squares (PLS) with the input data and setting data as explanatory variables (components) and the performance data as the objective variable. Each element of the feature vector is a latent variable (principal component). The feature vector calculation unit 111 repeatedly calculates latent variables until a predetermined number of latent variables (dimensions of the feature vector) is reached or until the accuracy of the PLS regression model reaches a predetermined accuracy, thereby calculating the feature vector. The order of the elements of the feature vector is the order of the calculated latent variables.

[0024] Similarly, the feature vector calculation unit 111 calculates feature vectors that serve as explanatory variables of the prediction model for the actual communication nodes, based on the input data, specifications, and setting data of the actual communication nodes. Hereinafter, feature vectors that serve as explanatory variables of the prediction model for the simulation may be referred to as simulation feature vectors, and feature vectors that serve as explanatory variables of the prediction model for the actual communication nodes may be referred to as actual communication node feature vectors.

[0025] <<Control Unit: Dimension Calculation Unit>> The dimension calculation unit 112 calculates the dimensions of the simulation feature vector and the actual communication node feature vector that minimize the distance between the simulation feature vector and the actual communication node feature vector. Hereinafter, a vector of only the first element of the simulation feature vector calculated by the feature vector calculation unit 111 will be referred to as a one-dimensional simulation feature vector. Also, a vector of only the first and second elements of the simulation feature vector calculated by the feature vector calculation unit 111 will be referred to as a two-dimensional simulation feature vector. Hereinafter, similarly, these will be referred to as a three-dimensional simulation feature vector, a four-dimensional simulation feature vector, etc. The actual communication node feature vector will also be similarly referred to as a one-dimensional actual communication node feature vector, a two-dimensional actual communication node feature vector, etc.

[0026] The dimension calculation unit 112 calculates the distance between the one-dimensional simulation feature vector and the one-dimensional real communication node feature vector, the distance between the two-dimensional simulation feature vector and the two-dimensional real communication node feature vector, the distance between the three-dimensional simulation feature vector and the three-dimensional real communication node feature vector, etc., to find the dimension that minimizes the distance. Hereinafter, the dimension that minimizes this distance will be referred to as the minimum distance dimension. The upper limit of the dimension is the number of latent variables calculated by the feature vector calculation unit 111, and is the smaller of the dimension of the simulation feature vector and the dimension of the real communication node feature vector.

[0027] The distance between the one-dimensional simulation feature vector and the one-dimensional real communication node feature vector is, for example, the distance between the average vector of the one-dimensional simulation feature vector and the average vector of the one-dimensional real communication node feature vector. It may also be the average of the distances between each one-dimensional simulation feature vector and each one-dimensional real communication node feature vector. The same applies to other dimensions.

[0028] <<Control Unit: Simulation Prediction Model Generation Unit>> The simulation prediction model generation unit 113 generates a simulation prediction model 121 using training data in which a simulation feature vector (see feature vector 312 in FIG. 2 ), which is the minimum distance dimension, is an explanatory variable and performance information 313 is an objective variable. The simulation prediction model 121 is, for example, a neural network.

[0029] <<Control Unit: Real Communication Node Prediction Model Generation Unit>> The real communication node prediction model generation unit 114 generates the real communication node prediction model 122 by transfer learning the simulation prediction model 121. The real communication node prediction model generation unit 114 generates the real communication node prediction model 122 by, for example, re-learning the weighting between the input layer and the next layer of the simulation prediction model 121, which is a neural network, and the weighting between the output layer and the layer before that. The training data used for re-learning is training data with the real communication node feature vector (see 322 in FIG. 3 ), which is the minimum distance dimension, as the explanatory variable and performance information 323 as the objective variable. Note that the real communication node prediction model generation unit 114 may re-learn by increasing the number of layers. For example, the real communication node prediction model generation unit 114 may re-learn the weighting between the output layer and the layer before that of the simulation prediction model 121.

[0030] <<Control Unit: Performance Prediction Unit>> The performance prediction unit 115 calculates performance information of the actual communication node using an actual communication node prediction model 122 based on input data, specifications, and setting data for the actual communication node. More specifically, the performance prediction unit 115 instructs the feature vector calculation unit 111 to acquire an actual communication node feature vector (see feature vector 322) based on input data, specifications, and setting data for the actual communication node (see input information 321 in FIG. 3 ). Next, the performance prediction unit 115 calculates performance information 323 of the actual communication node using the actual communication node prediction model 122 based on the actual communication node feature vector. The performance prediction unit 115 can also calculate performance information of a simulation model using a simulation prediction model 121 based on a simulation feature vector.

[0031] The calculation of performance information by the performance prediction unit 115 may be described as the prediction model for actual communication nodes 122 predicting performance information of the actual communication nodes, with the prediction model for actual communication nodes 122 being regarded as the subject. Also, the calculation of performance information by the performance prediction unit 115 may be described as the prediction model for simulation nodes 121 predicting performance information of the simulation model.

[0032] <Prediction Model Generation Process> Fig. 4 is a flowchart of the prediction model generation process according to this embodiment. The process of generating the actual communication node prediction model 122 will be described with reference to Fig. 4 .

[0033] In step S11, the feature vector calculation unit 111 calculates a simulation feature vector using input data, setting data, and performance data related to a simulation using a simulation model, which are stored in the simulation result database 130. In step S12, the feature vector calculation unit 111 calculates an actual communication node feature vector using input data, specifications, setting data, and performance data related to actual measurement of the actual communication node, which are stored in the actual communication node measurement result database 140.

[0034] In step S13, the dimension calculation unit 112 starts the process of repeating steps S14 to S16 while increasing the dimension by one from 1 to the smaller of the dimension of the simulation feature vector and the dimension of the actual communication node feature vector. Hereinafter, the dimension that is the target of this repeated process will be referred to as the target dimension.

[0035] In step S14, the dimension calculation unit 112 calculates the average vector of the simulation feature vectors of the processing target dimension. In step S15, the dimension calculation unit 112 calculates the average vector of the actual communication node feature vectors of the processing target dimension. In step S16, the dimension calculation unit 112 calculates the distance between the average vectors calculated in steps S14 and S15.

[0036] In step S17, the dimension calculation unit 112 determines the dimension (target dimension) for which the distance calculated in step S16 is smallest as the minimum distance dimension. In step S18, the simulation prediction model generation unit 113 generates the simulation prediction model 121 using training data in which the simulation feature vector, which is the minimum distance dimension, is an explanatory variable and the performance information is an objective variable.

[0037] In step S19, the actual communication node prediction model generation unit 114 performs transfer learning on the simulation prediction model 121 to generate an actual communication node prediction model 122. The learning data used for re-learning is learning data in which the actual communication node feature vector, which is the minimum distance dimension, is used as an explanatory variable and performance information is used as a target variable.

[0038] <Evaluation of Prediction Accuracy> The prediction accuracy of the prediction model for actual communication nodes 122 will be described below. FIG. 5 is a scatter diagram showing the relationship between the actual measured values ​​of delays of actual communication nodes and the predicted values ​​obtained by the conventional method. The coefficient of determination (R ２ ) is 0.57. FIG. 6 is a scatter plot showing the relationship between the actual measured value of delay of an actual communication node and the predicted value using the prediction model for actual communication nodes 122 according to this embodiment. The coefficient of determination is 0.97. When the prediction model for actual communication nodes 122 is used, the coefficient of determination for throughput and packet loss rate is 0.99, which is a good result. This shows that the use of the prediction model for actual communication nodes 122 makes it possible to predict performance information with high accuracy.

[0039] <Features of the Prediction Model Generation Device> The prediction model generation device 100 generates a prediction model for actual communication nodes, namely, a prediction model for actual communication nodes, namely, a prediction model for actual communication nodes, namely, a simulation prediction model 121, by using transfer learning technology to re-learn the simulation prediction model 121, which is a prediction model for a simulation model. The dimensions of the feature vectors serving as explanatory variables of the simulation prediction model 121 and the feature vectors serving as explanatory variables of the actual communication node prediction model 122 are dimensions in which the distance between the two feature vectors is minimum (minimum distance dimension). Therefore, it is assumed that the two feature vectors extract similar features of the input data and setting data, and the effect of transfer learning can be expected to be high.

[0040] Compared to performance measurements of actual communication nodes, performance measurements using simulations have the advantages of being able to obtain large amounts of data due to lower data collection costs, the absence of noise in the data, and the ability to obtain data in communication environments that cannot be measured with actual communication nodes, and these advantages are expected to result in high performance prediction accuracy for simulation prediction models.Prediction models for actual communication nodes generated by transfer learning from such simulation prediction models are also expected to have high prediction accuracy.

[0041] Although several embodiments of the present invention have been described above, these embodiments are merely illustrative and do not limit the technical scope of the present invention. The present invention can take on various other embodiments, and various modifications such as omissions and substitutions can be made without departing from the spirit of the present invention. These embodiments and their modifications are included within the scope and spirit of the invention described in this specification, etc., and are also included in the invention described in the claims and their equivalents.

[0042] <Hardware Configuration> The predictive model generating device 100 according to the above-described embodiment is realized by a computer 900 having a configuration as shown in FIG. 7 , for example. FIG. 7 is a hardware configuration diagram showing an example of the computer 900 that realizes the functions of the predictive model generating device 100 according to the above-described embodiment. The computer 900 includes a CPU 901, a ROM 902, a RAM 903, an SSD 904, an input / output interface 905 (referred to as an input / output I / F (Interface) in FIG. 7 ), a communication interface 906 (referred to as a communication I / F in FIG. 7 ), and a media interface 907 (referred to as a media I / F in FIG. 7 ). The computer 900 may include a hard disk drive (HDD) instead of the SSD 904, or may include an HDD in addition to the SSD 904.

[0043] The CPU 901 operates based on programs stored in the ROM 902 or the SSD 904, and performs control by the control unit 110 in Fig. 1. The ROM 902 stores a boot program executed by the CPU 901 when the computer 900 starts up, programs related to the hardware of the computer 900, and the like. The CPU 901 controls an input device 910 such as a mouse or keyboard, and an output device 911 such as a display or printer, via an input / output interface 905. The CPU 901 acquires data from the input device 910 and outputs generated data to the output device 911 via the input / output interface 905.

[0044] The SSD 904 stores programs executed by the CPU 901 and data used by the programs. The communication interface 906 receives data from other devices (not shown) (e.g., a simulator or a device for measuring the performance of a real communication node) via a communication network and outputs the data to the CPU 901. It also transmits data generated by the CPU 901 to other devices via the communication network. The media interface 907 reads programs or data stored in a recording medium 912 and outputs the programs to the CPU 901 via the RAM 903. The CPU 901 loads the programs from the recording medium 912 onto the RAM 903 via the media interface 907 and executes the loaded programs. The recording medium 912 may be an optical recording medium such as a DVD (Digital Versatile Disk), a magneto-optical recording medium such as an MO (Magneto Optical Disk), a magnetic recording medium, a conductive memory tape medium, or a semiconductor memory.

[0045] For example, when the computer 900 functions as the predictive model generation device 100 according to the embodiment described above, the CPU 901 of the computer 900 executes the program 128 (see FIG. 1 ) loaded onto the RAM 903, thereby realizing the functions of the predictive model generation device 100. The CPU 901 reads and executes the program from the recording medium 912. Alternatively, the CPU 901 may read the program from another device via a communications network, or may install and execute the program 128 from the recording medium 912 onto the SSD 904.

[0046] <Effects> The effects of the device will be described below.

[0047] The prediction model generation device 100 according to the embodiment described above includes a feature vector calculation unit 111 that calculates a simulation feature vector, which is a feature vector 312 of input information 311 (input data, setting data) when performance information 313 (see FIG. 2) of a simulation model of a communication node is acquired, and a real communication node feature vector, which is a feature vector 322 of input information 321 when performance information 323 (see FIG. 3) of a real communication node, which is an actual device, is acquired.

[0048] The prediction model generating device 100 also includes a simulation prediction model generating unit 113 that generates a simulation prediction model 121 that predicts the performance information of the simulation model using learning data that uses a simulation feature vector (see feature vector 312) as an explanatory variable and performance information 313 of the simulation model as a target variable.

[0049] Furthermore, the prediction model generating device 100 includes a real communication node prediction model generating unit 114 that performs transfer learning of the simulation prediction model 121 using learning data that has the feature vector for the real communication node (see feature vector 322) as an explanatory variable and performance information 323 of the real communication node as a target variable, to generate a real communication node prediction model 122 that predicts the performance information of the real communication node.

[0050] Such a prediction model generation device 100 can generate a prediction model 122 for an actual communication node based on the simulation prediction model 121. Because the training data for the simulation prediction model 121 has advantages such as being able to be prepared in large quantities and having little noise, the prediction accuracy of the simulation prediction model 121 can be expected to be high. The prediction accuracy of the actual communication node prediction model 122 generated by transfer learning of such a simulation prediction model 121 can also be expected to be high.

[0051] The prediction model generating device 100 includes a dimension calculation unit 112 that calculates a minimum distance dimension, which is the dimension of a simulation feature vector (see feature vector 312) and an actual communication node feature vector (see feature vector 322) that minimizes the distance between the simulation feature vector and the actual communication node feature vector. The dimensions of the simulation feature vector and the actual communication node feature vector, which are explanatory variables of the training data, are the minimum distance dimension.

[0052] According to this predictive model generation device 100, the dimension of the feature vectors 312 and 322 is the dimension (minimum distance dimension) in which the distance between the feature vectors 312 and 322 is minimum. Therefore, it is assumed that the feature vectors 312 and 322 extract similar features from the input data and setting data, and transfer learning can be expected to be highly effective.

[0053] The simulation prediction model 121 is a neural network. The actual communication node prediction model generation unit 114 performs transfer learning on the simulation prediction model 121 by re-learning the weightings for the input layer and the output layer of the simulation prediction model 121 using the learning data, thereby generating the actual communication node prediction model 122.

[0054] In transfer learning, the output layer (weighting between the output layer and the layer before it) is generally retrained (weighting parameters are readjusted). In contrast, the prediction model generation device 100 retrains the input layer (weighting between the input layer and the layer after it) in addition to the output layer. This enables transfer learning between the simulation model and the actual communication node, even when different input data or different setting data are used.

[0055] The feature vector calculation unit 111 calculates a simulation feature vector and the feature vector for the actual communication node using the partial least squares method.

[0056] According to this prediction model generating device 100, it is possible to calculate feature vectors 312, 322 of input information (see input information 311, 321) taking into account performance information 313, 323 (see FIGS. 2 and 3).

[0057] 100 Prediction model generation device 110 Control unit 111 Feature vector calculation unit 112 Dimension calculation unit 113 Simulation prediction model generation unit 114 Real communication node prediction model generation unit 115 Performance prediction unit 120 Storage unit 121 Simulation prediction model 122 Real communication node prediction model 128 Program 130 Simulation result database 140 Real communication node measurement result database 180 Input / output unit 311, 321 Input information 312 Feature vector (simulation feature vector) 313, 323 Performance information 322 Feature vector (real communication node feature vector)

Claims

1. A prediction model generation device comprising: a feature vector calculation unit that calculates a simulation feature vector which is a feature vector of input information when performance information of a simulation model of a communication node is obtained, and a real communication node feature vector which is a feature vector of input information when performance information of a real communication node which is an actual device is obtained; a simulation prediction model generation unit that generates a simulation prediction model that predicts the performance information of the simulation model using learning data which has the simulation feature vector as an explanatory variable and the performance information of the simulation model as an objective variable; and a real communication node prediction model generation unit that transfer learns the simulation prediction model using learning data which has the real communication node feature vector as an explanatory variable and the performance information of the real communication node as an objective variable, to generate a real communication node prediction model that predicts the performance information of the real communication node.

2. A predictive model generating device as described in claim 1, further comprising a dimension calculation unit that calculates a minimum distance dimension, which is the dimension of the simulation feature vector and the feature vector for the actual communication node that minimizes the distance between the simulation feature vector and the feature vector for the actual communication node, and the dimension of the simulation feature vector and the feature vector for the actual communication node, which are explanatory variables of the learning data, is the minimum distance dimension.

3. The prediction model generating device according to claim 1, wherein the simulation prediction model is a neural network, and the actual communication node prediction model generating unit generates the actual communication node prediction model by transfer learning the simulation prediction model by re-learning weightings related to an input layer and an output layer of the simulation prediction model using the learning data.

4. A prediction model generating method in which a prediction model generating device executes the steps of: calculating a simulation feature vector which is a feature vector of input information when performance information of a simulation model of a communication node is obtained, and a real communication node feature vector which is a feature vector of input information when performance information of a real communication node which is an actual device is obtained; generating a simulation prediction model which predicts the performance information of the simulation model using learning data which has the simulation feature vector as an explanatory variable and the performance information of the simulation model as an objective variable; and transferring learning the simulation prediction model using learning data which has the real communication node feature vector as an explanatory variable and the performance information of the real communication node as an objective variable, to generate a real communication node prediction model which predicts the performance information of the real communication node.

Images