JP7741451B2 - Learning management device, learning management method, and program - Google Patents

Description

This disclosure relates to a learning management device, a learning management method, and a program.

Conventionally, there exists a communication system NS having a communication network NW in which communication between multiple end terminals takes place via multiple forwarding devices. The communication system includes a network database NDB that collects data on parameters (referred to herein as network parameters) that represent the performance of the communication network, and a prediction function LM that predicts future network parameters using machine learning techniques. Figure 8 shows an overview of the communication system NS. In the communication system NS, communication takes place between end terminals ET1 and ET2 via forwarding devices T1 to T4. In the graph of Figure 9, the vertical axis represents traffic flow as a collected network parameter, and the horizontal axis represents time. In the communication system NS, at the time indicated by the arrow in Figure 9, a learning model is created by performing arbitrary machine learning using past traffic flow as learning data, and the prediction function LM can then predict future traffic flow using the learning model. In this way, a method for predicting future network parameters has been devised.

For example, Non-Patent Document 1 discloses a technique for predicting future network parameters using past network parameters. The learning data used to generate a learning model may include past network parameters as well as related setting data. For example, Non-Patent Document 2 discloses performing machine learning using user feature data in addition to network parameters, and predicting future network parameters based on the generated learning model.

Machine learning is periodically repeated to track changes in network parameters. An example of repeated machine learning will be described with reference to Figure 10. Figure 10 shows a graph with the vertical axis representing traffic flow rate and the horizontal axis representing time. First, machine learning is performed using "learning portion A," which represents fluctuations in traffic flow rate before time A, as learning data, to generate learning model A. This learning model A is used to predict traffic flow rate from time A to time B. Learning model A has learned traffic flow rate fluctuations before time A and is capable of predicting similar fluctuations in the future. Because the trend in traffic flow rate fluctuations changes before and after time B, machine learning is performed again using "learning portion B," which represents traffic flow rate fluctuations before time B, to generate learning model B. Here, the traffic flow rate of learning portion B partially overlaps with learning portion A. The generated learning model B has learned traffic flow rate fluctuations before time B and is capable of predicting similar fluctuations in the future.

As an example of iterative machine learning, Non-Patent Document 3 discloses continuous learning, in which tasks are additionally learned as training data as needed. Non-Patent Document 3 also discloses learning new tasks without forgetting past tasks. Furthermore, Non-Patent Document 4 discloses performing new machine learning using past trained models and new tasks. An example of continuous learning will be explained with reference to Figure 11. In Figure 11, the vertical axis represents traffic flow and the horizontal axis represents time. In continuous learning, learning model A is first generated using learning portion A, which is the traffic flow before time A, i.e., task A, as training data. Learning model B is then generated using learning model A, which is the traffic flow from time A to time B. Task B, which is the traffic flow from time A to time B, and learning model A. Thereafter, tasks are additionally used as training data, and machine learning is continuously performed in a similar manner.

R. Vinayakumar, K. P. Soman and P. Poornachandran, “Applying deep learning approaches for network traffic prediction,” 2017 International Conference on Advances in Computing, Communications and Informatics (ICACCI), 2017 A. Azari, P. Papapetrou, S. Denic, and G. Peters, “User traffic prediction for proactive resource management: Learning-powered approaches,” in 2019 IEEE Global Communications Conference (GLOBECOM), pp. 1-6, 2019 Kirkpatrick, James, et al. "Overcoming catastrophic forgetting in neural networks." Proceedings of the national academy of sciences 114.13 (2017): 3521-3526 Masana, Marc, et al. "Class-incremental learning: survey and performance evaluation on image classification." arXiv preprint arXiv:2010.15277 (2020)

When continuous learning is introduced to generate a learning model that predicts future network parameters, the criteria for dividing past network parameters into tasks become important. However, Non-Patent Document 4 suggests that increasing the number of tasks may result in a decrease in prediction accuracy. Thus, there is a need for technology to improve continuous learning methods by using tasks divided into an appropriate number of times.

The purpose of this disclosure, made in light of these circumstances, is to provide technology that improves continuous learning methods by using tasks divided into appropriate numbers of times.

To solve the above problem, the learning management device disclosed herein comprises a log acquisition unit that displays a log of network parameters generated at predetermined time intervals and acquires time-series log information used for continuous learning; a judgment unit that determines whether a trend change in the network parameters has occurred based on the log information; a task generation unit that, if a trend change has occurred, separates the subsequent log information from past tasks as a new task; a memory unit that stores the tasks; and a deletion unit that deletes the past task if the new task is similar to the past task stored in the memory unit.

In addition, to solve the above-mentioned problems, the learning management method disclosed herein is a learning management method executed by a learning management device equipped with a memory unit, and includes a log acquisition step of acquiring time-series log information that shows a log of network parameters generated at predetermined time intervals and is used for continuous learning; a judgment step of determining whether a trend change has occurred in the network parameters based on the log information; a task generation step of separating the subsequent log information from past tasks as a new task if the trend change has occurred; a storage step of storing the task in the memory unit; and a deletion step of deleting the past task if the new task is similar to the past task stored in the memory unit.

To solve the above problem, the program disclosed herein causes a computer to function as a learning management device disclosed herein.

The learning management device, learning management method, and program disclosed herein can provide technology that improves continuous learning methods by using tasks divided into appropriate numbers of times.

1 is a diagram showing a schematic configuration of a system according to an embodiment of the present invention. 1 is a diagram illustrating an example of the configuration of a learning management device according to an embodiment of the present invention. FIG. 10 is a diagram illustrating an example of log information. FIG. 10 is a diagram for explaining trend changes of network parameters. FIG. 10 is a diagram for explaining tasks to be divided. FIG. 10 is a diagram for explaining a determination of whether a new task is similar to a past task. 10 is a flowchart illustrating an example of an operation of the learning management device according to the present embodiment. 10 is a flowchart illustrating an example of an operation of the learning management device according to the present embodiment. FIG. 1 is a schematic diagram illustrating a conventional communication system. FIG. 1 is a diagram for explaining a conventional technique for predicting network parameters. FIG. 1 is a diagram illustrating an example of conventional machine learning that is performed repeatedly. FIG. 1 is a diagram illustrating an example of conventional continuous learning. FIG. 10 is a diagram illustrating an example in which the learning management device according to the present embodiment is not used.

Embodiments of the present disclosure will be described below with reference to the drawings as appropriate. In each drawing, identical or corresponding parts are designated by the same reference numerals. In describing the present embodiments, descriptions of identical or corresponding parts will be omitted or simplified as appropriate. The embodiments described below are examples of the configuration of the present disclosure, and the present invention is not limited to the following embodiments.

<General configuration of system 1>
FIG. 1 is a diagram showing the configuration of a system 1 according to this embodiment. As shown in FIG. 1, the system 1 includes a network parameter collection device N and a learning management device 10. Each device is communicably connected to a network 30, which may include, for example, the Internet and a mobile communication network, via wired or wireless communication. The communication method for transmitting and receiving information between each device is not particularly limited. Each device may be integrated. That is, for example, the learning management device 10 may have the functions of the network parameter collection device N. Each device communicates with each other via the network 30.

The network parameter collection device N and the learning management device 10 are computers such as servers belonging to a cloud computing system or other computing system.

Network 30 includes the Internet, at least one Wide Area Network (WAN), at least one Metropolitan Area Network (MAN), or any combination thereof. Network 30 may also include at least one wireless network, at least one optical network, or any combination thereof. The wireless network may be, for example, an ad hoc network, a cellular network, a wireless local area network (LAN), a satellite communication network, or a terrestrial microwave network.

The network parameter collection device N is composed of a computer equipped with memory, a controller, and a communication interface. The memory may be composed of a hard disk drive (HDD), a solid state drive (SSD), an electrically erasable programmable read-only memory (EEPROM), a read-only memory (ROM), a random access memory (RAM), etc. The controller may be composed of dedicated hardware such as an application specific integrated circuit (ASIC) or a field-programmable gate array (FPGA), a processor, or a combination of both. The communication interface may use standards such as Ethernet (registered trademark), FDDI (Fiber Distributed Data Interface), or Wi-Fi (registered trademark).

The network parameter collection device N collects network parameters, which are parameters that represent the performance of the network within the network 30. Network parameters are parameters that are collected in any unit, such as per interface or per flow, at any time interval. Network parameters include, for example, average transmission traffic flow rate, resource utilization rate in the transfer device, and delay time. In this embodiment, the network parameters refer to the average transmission traffic flow rate. The network parameter collection device N transmits log information indicating a log of the collected network parameters to the learning management device 10.

The learning management device 10 generates tasks as learning data based on log information as described below, and performs continuous learning using the tasks. The learning management device 10 generates a learning model that outputs predicted values of future network parameters, and can predict future network parameters using the learning model.

<Configuration of learning management device 10>
The learning management device 10 will be described with reference to Fig. 2. The learning management device 10 includes a control unit 11, a storage unit 12, a communication unit 13, an input unit 14, and an output unit 15.

The memory unit 12 includes one or more memories, and may include, for example, semiconductor memory, magnetic memory, optical memory, etc. Each memory included in the memory unit 12 may function, for example, as a main memory device, an auxiliary memory device, or a cache memory. The memory unit 12 stores any information used in the operation of the learning management device 10. The memory unit 12 does not necessarily have to be provided inside the learning management device 10, but may also be configured to be provided external to the learning management device 10. The memory unit 12 stores tasks generated as described below.

The communication unit 13 includes one or more communication interfaces that connect to the network 30. The communication interfaces correspond to, for example, mobile communication standards, wired LAN standards, or wireless LAN standards, but are not limited to these and may correspond to any communication standard. The communication unit 13 receives information used in the operation of the learning management device 10 and transmits information obtained by the operation of the learning management device 10.

The input unit 14 includes at least one input interface. The input interface may be, for example, a physical key, a capacitive key, a pointing device, a touch screen integrated with a display, or a microphone. The input unit 14 accepts operations to input information used in the operation of the learning management device 10. The input unit 14 may be connected to the learning management device 10 as an external input device, instead of being provided in the learning management device 10. Any connection method may be used, for example, USB (Universal Serial Bus), HDMI (High-Definition Multimedia Interface) (registered trademark), or Bluetooth (registered trademark).

The output unit 15 includes at least one output interface. The output interface is, for example, a display or a speaker. The display is, for example, an LCD (Liquid Crystal Display) or an organic EL (Electro Luminescence) display. The output unit 15 outputs information obtained by the operation of the learning management device 10. The output unit 15 may be connected to the learning management device 10 as an external output device instead of being provided in the learning management device 10. Any connection method can be used, for example, USB, HDMI (registered trademark), or Bluetooth (registered trademark).

The control unit 11 is realized by a control and arithmetic circuit (controller). The control and arithmetic circuit may be configured with dedicated hardware such as an ASIC or FPGA, or may be configured with a processor, or may be configured to include both. The control unit 11 controls each part of the learning management device 10 and executes processes related to the operation of the learning management device 10. The control unit 11 can send and receive information to and from external devices via the communication unit 13 and the network 30.

The control unit 11 includes a log acquisition unit 111, a judgment unit 112, a task generation unit 113, a deletion unit 114, and a learning prediction unit 115.

In the following description, the learning management device 10 is assumed to have each unit, but each unit may also be configured to be provided on a device separate from the learning management device 10. For example, a first device separate from the learning management device 10 may be configured to have the continuous learning function of the learning prediction unit 115, and a second device separate from the learning management device 10 may be configured to have the prediction function of the learning prediction unit 115.

The log acquisition unit 111 acquires time-series log information that shows a log of network parameters generated at a predetermined time interval and is used for continuous learning. An example of log information is shown in Figure 3. In Figure 3, the horizontal axis represents time and the vertical axis represents traffic flow. The predetermined time interval may be set arbitrarily, for example, several milliseconds or several minutes. In this embodiment, the network parameter log is a log of traffic flow on a predetermined network, but is not limited to this and may also be resource utilization rate, delay time, etc. in a transfer device at predetermined time intervals. The format of the log information is not limited to a graph and may be any format, such as a table. The log acquisition unit 111 outputs the acquired log information to the determination unit 112.

The judgment unit 112 judges whether a trend change has occurred in the network parameters based on the log information. In this embodiment, a trend change refers to a fluctuation in the rate of increase or decrease of traffic flow over a specified period of time. Any method may be used to determine whether a trend change has occurred. For example, the judgment unit 112 may determine that a trend change has occurred based on a method using a moving average, seasonal difference, or the like, which is commonly used in time series analysis.

In methods using moving averages, the average value of time series data over a specified period is plotted as the average value at that point in time, generating a moving average line that smooths the time series data. Seasonal differencing refers to the difference from the same period of the previous year, for example, and using seasonal differencing with a period p is a method that can visualize fluctuations in time series data with seasonal fluctuations removed. Moving averages and seasonal differencing are well-known technologies, so a detailed explanation will be omitted.

In this embodiment, the judgment unit 112 applies a moving average line to the acquired log information using a moving average method. Figure 4 shows the log information with the moving average line applied. Based on the log information in Figure 4, the judgment unit 112 determines whether there is a change in the trend of the traffic flow log. Based on the moving average line shown in Figure 4, the judgment unit 112 determines that there is a change in the trend of the traffic flow between the log before time A and the log after time A. Similarly, the judgment unit 112 determines that there is a change in the trend between the log before time B and the log after time B, and further determines that there is a change in the trend between the log before time C and the log after time C.

The timing at which the judgment unit 112 judges a change in trend may be set freely. For example, the judgment unit 112 may be set to judge a change in trend every few hours, or every few days.

The conditions under which the judgment unit 112 determines a trend change may be set, for example, according to the amount of network communication traffic available during a specified period contracted by each of multiple users of the network 30. For example, the judgment unit 112 may determine that a trend change has occurred when the contracted network communication traffic exceeds the upper limit. In this way, the judgment unit 112 can determine that a trend change is the excess or deficiency of the contracted network communication traffic relative to user demand.

When a trend change occurs, the task generation unit 113 separates the subsequent log information as a new task from past tasks. For example, assume that the judgment unit 112 determines, based on the log information in Figure 4, that a change in the trend of the traffic flow log occurred at time A. In this case, as shown in Figure 5, the task generation unit 113 separates the log information from time A onwards as a new task from the past log from the time log acquisition began to time A. This past log is called task A. Similarly, assume that the judgment unit 112 determines that there is a change in the trend of the traffic flow log at times B and C. In this case, the task generation unit 113 separates the log from time A to time B, and further separates the log from time B onwards to time C. The log from time A to time B is called task B, and the log from time B to time C is called task C. In this way, the task generation unit 113 generates task A, task B, and task C. Similarly, when the task generation unit 113 determines that there is a trend change in the logs from time C onward, it divides the subsequent log information from past tasks and generates tasks.

The task generation unit 113 stores past tasks in the memory unit 12. The task generation unit 113 also outputs the new tasks it generates to the deletion unit 114. Past tasks may exist for the period for which the log acquisition unit 111 acquired logs. For example, at time B, the task generation unit 113 stores task A in the memory unit 12 as a past task, and outputs task B to the deletion unit 114 as a generated task. Then, at time C, the task generation unit 113 adds task B to task A as a past task, stores task B in the memory unit 12, and outputs task C to the deletion unit 114 as a generated task.

The deletion unit 114 first determines whether the output task is similar to a past task. Specifically, the deletion unit 114 reads the past task from the memory unit 12 and compares it with the new task output from the task generation unit 113. For example, assume that at time C, task C is output to the deletion unit 114 from the task generation unit 113 as a new task. At this time, the memory unit 12 stores task A and task B. The deletion unit 114 reads task A and task B from the memory unit 12 and compares them with task C output from the task generation unit 113.

Any method may be used to determine whether tasks are similar. For example, the deletion unit 114 can use dynamic time warping (DTW) to determine whether tasks are similar.

Dynamic time warping calculates the similarity between two waveform patterns. In dynamic time warping, multiple arbitrary points are taken on each of two time-varying waveform patterns, the distance (cost) between these points is calculated in a brute-force manner, and the total value of the elements that make up the cost path is calculated as the DTW distance. The smaller the DTW distance, the greater the similarity between the two patterns; the larger the DTW distance, the less similarity there is. Dynamic time warping is a known technology, so a detailed explanation will be omitted.

The deletion unit 114 may calculate the similarity between tasks using similarity levels that are preset as graded values according to the value of the DTW distance. For example, when dynamic time warping is used for two tasks, the deletion unit 114 may calculate the similarity as "high" if the DTW distance value is less than a first predetermined value, calculate the similarity as "medium" if the DTW distance value is equal to or greater than the first predetermined value and less than a second predetermined value, and calculate the similarity as "low" if the DTW distance value is equal to or greater than the second predetermined value. The deletion unit 114 may determine that the output task is similar to a past task if the similarity is "high" or "medium."

In this embodiment, it is assumed that the deletion unit 114 compares task A, which is a past task shown in Figure 6, with the output task C and calculates that the similarity is "high." In this case, the deletion unit 114 determines that task A and task C are similar. It is assumed that the deletion unit 114 further compares task B, which is a past task, with task C and calculates that the similarity is "low." In this case, the deletion unit 114 determines that task B and task C are not similar.

Not limited to the above, the deletion unit 114 may also be able to determine whether two tasks are similar using an arbitrary correlation function.

If the new task is similar to a past task stored in the memory unit 12, the deletion unit 114 deletes the past task. In this embodiment, the deletion unit 114 determines that past task A and output task C are similar. Therefore, the deletion unit 114 deletes task A from the memory unit 12. The deletion unit 114 also stores task C in the memory unit 12.

As described above, the deletion unit 114 deletes a past task if the new task is similar to a past task stored in the memory unit 12. The deletion unit 114 calculates the similarity between the new task and the past task using dynamic time warping.

The learning prediction unit 115 performs continuous learning using tasks stored in the memory unit 12 to predict future network parameters. Continuous learning involves re-learning using a trained model and data subsequent to the training data used to generate the trained model as new training data to generate a new model. Continuous learning can regenerate a trained model in response to changes in the training data, just as if reviewing a trained model. Any machine learning method, such as a neural network or deep learning, may be used for continuous learning.

In this embodiment, the learning prediction unit 115 reads out task B and task C, which have not been deleted from the memory unit 12, and uses them as learning data to generate a new learning model that outputs future network parameters. The new learning model recognizes tasks with changing trends and clearly dissimilar logs as separate tasks and uses them as learning data, enabling more accurate predictions when similar trends exist for these tasks. Furthermore, since the learning model is generated using only the minimum number of tasks necessary, without using similar tasks as learning data, it is possible to suppress a decline in prediction accuracy compared to when multiple similar tasks are used in duplicate as learning data.

The learning prediction unit 115 applies the newly generated learning model to predict future network parameters. Predictions using this new learning model can more accurately predict the trends in network parameters observed in the task A and task B sections.

<Program>
A computer capable of executing program instructions can be used to function as the learning management device 10. Here, the computer may be a general-purpose computer, a dedicated computer, a workstation, a personal computer (PC), an electronic notepad, etc. The program instructions may be program code, code segments, etc. for performing the necessary tasks.

A computer comprises a processor, a memory unit, an input unit, an output unit, and a communication interface. The processor may be a CPU (Central Processing Unit), MPU (Micro Processing Unit), GPU (Graphics Processing Unit), DSP (Digital Signal Processor), SoC (System on a Chip), etc., and may be composed of multiple processors of the same or different types. The processor controls the above components and performs various arithmetic processing by reading and executing programs from the memory unit. Note that at least some of these processing contents may be implemented in hardware. The input unit is an input interface that accepts user input operations and acquires information based on the user operations, such as a pointing device, keyboard, or mouse. The output unit is an output interface that outputs information, such as a display or speaker. The communication interface is an interface for communicating with external devices.

The program may be recorded on a computer-readable recording medium. Using such a recording medium, the program can be installed on a computer. Here, the recording medium on which the program is recorded may be a non-transitory recording medium. Non-transitory recording media are not particularly limited, but may be, for example, a CD-ROM, DVD-ROM, or USB memory. The program may also be downloaded from an external device via a network.

<Operation of System 1>
Next, the operation of the learning management device 10 according to this embodiment will be described with reference to Figures 7A and 7B. The operation of the learning management device 10 corresponds to the learning management method according to this embodiment.

In step S1, the log acquisition unit 111 of the learning management device 10 acquires time-series log information that indicates a log of network parameters generated at a predetermined time interval and is used for continuous learning. In this embodiment, the log acquisition unit 111 acquires the log information by receiving it from the network parameter collection device N via the communication unit 13. The log acquisition unit 111 outputs the acquired log information to the judgment unit 112.

In step S2, the judgment unit 112 judges whether a change in the trend of the network parameters has occurred based on the log information. Any method may be used to determine whether a trend change has occurred. If the judgment unit 112 determines that there has been no change in the trend of the network parameter log, the processing of the learning management device 10 returns to step S1. If the judgment unit 112 determines that there has been a change in the trend of the network parameter log, the processing of the learning management device 10 proceeds to step S3.

In step S3, the task generation unit 113 separates the subsequent log information from past tasks as a new task. In this example, it is assumed that the task generation unit 113 determines, based on the log information in FIG. 4, that a change in the trend of the traffic flow log occurred at time A. In this case, as shown in FIG. 5, the task generation unit 113 separates the log information from time A onwards as a new task from past logs from the time log acquisition started to time A. This past log is referred to as task A. Similarly, it is assumed that the judgment unit 112 determines that a change in the trend of the traffic flow log occurred at times B and C. In this case, the task generation unit 113 separates the log from time A to time B, and further separates the log from time B onwards to time C. It is assumed that the log from time A to time B is called task B, and the log from time B to time C is called task C. In this way, when a trend change occurs, the task generation unit 113 separates the subsequent log information from past tasks as a new task.

In step S4, the task generation unit 113 stores the past tasks in the memory unit 12 and outputs the newly generated task to the deletion unit 114. In this example, at time C, the task generation unit 113 adds task B to task A as a past task, stores task B in the memory unit 12, and outputs task C to the deletion unit 114 as a generated task.

In step S5, the deletion unit 114 determines whether the output task and the past task are similar. Any method may be used to determine whether they are similar. For example, the deletion unit 114 can use dynamic time warping to determine whether the tasks are similar. If it is determined that the tasks are similar, the processing of the learning management device 10 proceeds to step S6. If it is determined that the tasks are not similar, the processing of the learning management device 10 proceeds to step S7. In this example, as shown in FIG. 6, the deletion unit 114 determines whether the output task C and the past task A are similar, and whether the output task C and the past task B are similar.

In step S6, if the new task is similar to a past task stored in the memory unit 12, the deletion unit 114 deletes the past task. In this example, it is determined that task C and task A are similar. Therefore, the deletion unit 114 deletes task A from the memory unit 12. The deletion unit 114 also stores task C in the memory unit 12.

As shown in steps S5 and S6, if the new task is similar to a past task stored in the memory unit 12, the deletion unit 114 deletes the past task. The deletion unit 114 calculates the similarity between the new task and the past task using dynamic time warping.

In step S7, the learning prediction unit 115 generates a new learning model that outputs future network parameters. In this embodiment, the learning prediction unit 115 reads out task B and task C that have not been deleted from the memory unit 12, and uses these as learning data to generate a new learning model that outputs future network parameters.

In step S8, the learning prediction unit 115 applies the newly generated learning model in step S8 to predict future network parameters. Then, the operation of the system 1 ends.

As described above, the learning management device 10 of this embodiment includes a log acquisition unit 111 that shows a log of network parameters generated at predetermined time intervals and acquires time-series log information used for continuous learning, a judgment unit 112 that determines whether a change in the trend of the network parameters has occurred based on the log information, a task generation unit 113 that, if a change in trend has occurred, separates the subsequent log information from past tasks as a new task, a memory unit 12 that stores the tasks, and a deletion unit 114 that deletes the past task if the new task is similar to a past task stored in the memory unit 12.

According to this embodiment, only if a newly generated task is determined to be dissimilar to a past task, the newly generated task is stored in the memory unit 12 as a past task, and if it is determined to be similar, the past task is deleted. Therefore, only tasks that are dissimilar to each other remain in the memory unit 12, and continuous learning is performed using these tasks as learning data.

Figure 12 shows an example of dividing tasks at a single time interval, without using trend changes in network parameters as a division criterion, as in this embodiment. In Figure 12, tasks A' to C' are generated by dividing them between times A' and C'. Comparing this with Figure 5 according to this embodiment, it can be seen that task B' contains a mixture of network parameter logs with different trends. If an attempt is made to generate a new learned model B' by performing continuous learning using trained model A' and task B' as training data for a new task, task B' will contain data for multiple trends, and model B' may not be able to properly learn each trend.

On the other hand, when tasks are divided based on trend changes as in this embodiment, task B will only contain logs of network parameters with the same trend. Therefore, learning model B, which is generated using learning model A, a trained model, and task B, can accurately learn the trends contained in task B. This makes it possible to achieve highly accurate predictions when data with a trend similar to that contained in task B appears in the future.

In this way, in this embodiment, tasks are divided into the optimal number of times based on the effective criterion of trend changes, which prevents a decrease in the accuracy of predictions of the learning model due to a large number of tasks and reduces the load on machine learning. Therefore, it is possible to provide a technology that improves continuous learning methods by using tasks that are divided into an appropriate number of times.

As described above, the learning management device 10 of this embodiment further includes a learning prediction unit 115 that performs continuous learning using tasks stored in the memory unit 12 and predicts future network parameters.

According to this embodiment, the learning prediction unit 115 can efficiently perform continuous learning using only tasks that are not similar to past tasks and that have not been deleted by the deletion unit 114. The learning prediction unit 115 performs additional continuous learning each time a trend change occurs, generates a new learning model, and uses the newly generated learning model to more accurately predict future network parameters. Therefore, a technology can be provided that can improve continuous learning methods by using tasks divided into an appropriate number of times.

As described above, in the learning management device 10 of this embodiment, the deletion unit 114 calculates the similarity between a new task and a past task using dynamic time warping.

According to this embodiment, the deletion unit 114 can accurately determine the similarity between two tasks even when the logs are of different lengths and there is a time offset. Therefore, it is possible to provide a technology that improves continuous learning methods by using tasks divided into an appropriate number of times.

As described above, in the learning management device 10 of this embodiment, the network parameter is traffic flow rate.

According to this embodiment, it is possible to generate a more accurate learning model for traffic flow, which is one of the network parameters that is closely related to the content of a user's communication service contract.

The following additional notes are provided regarding the above-described embodiments.
[Additional note 1]
A memory and a control unit connected to the memory,
The control unit
Showing a log of network parameters generated at a predetermined time interval, and obtaining time-series log information for use in continuous learning;
determining whether a trend change of the network parameter has occurred based on the log information;
When a change in the trend occurs, subsequent log information is separated from past tasks as a new task;
If the new task is similar to the previous task stored in the memory, deleting the previous task;
The memory stores the task.
Learning management device.
[Additional note 2]
2. The learning management device according to claim 1, wherein the control unit performs continuous learning using the tasks stored in the memory and predicts future network parameters.
[Additional note 3]
3. The learning management device according to claim 1, wherein the control unit calculates the similarity between the new task and the past task using a dynamic time warping method.
[Additional note 4]
4. The learning management device according to any one of claims 1 to 3, wherein the network parameter is a traffic flow rate.
[Additional note 5]
A learning management method executed by a learning management device having a memory,
a log acquisition step of acquiring time-series log information that indicates a log of network parameters generated at a predetermined time interval and is used for continuous learning;
a determining step of determining whether a trend change of the network parameter has occurred based on the log information;
a task generation step of dividing subsequent log information into a new task and past tasks when a change in the trend occurs;
a storing step of storing the task in the memory;
a deleting step of deleting the previous task stored in the memory if the new task is similar to the previous task;
Learning management methods, including:
[Additional note 6]
6. The learning management method according to claim 5, further comprising a learning prediction step of performing continuous learning using the tasks stored in the memory and predicting future network parameters.
[Additional note 7]
7. The learning management method according to claim 5, wherein the deleting step includes a step of calculating a similarity between the new task and the past task using a dynamic time warping method.
[Additional Note 8]
A non-transitory storage medium storing a program executable by a computer,
A non-transitory storage medium storing a program that causes the computer to function as the learning management device described in any one of appendixes 1 to 4.

The above-described embodiments have been described as representative examples, but it will be apparent to those skilled in the art that numerous modifications and substitutions are possible within the spirit and scope of the present disclosure. Therefore, the present invention should not be construed as being limited to the above-described embodiments, and various modifications and alterations are possible without departing from the scope of the claims. For example, it is possible to combine multiple building blocks shown in the configuration diagrams of the embodiments into one, or to divide one building block.

1 System 10 Learning management device 11 Control unit 12 Storage unit 13 Communication unit 14 Input unit 15 Output unit 30 Network 111 Log acquisition unit 112 Determination unit 113 Task generation unit 114 Deletion unit 115 Learning prediction unit

Claims (8)

a log acquisition unit that acquires time-series log information indicating a log of network parameters generated at a predetermined time interval and that is used for continuous learning;
a determination unit that determines whether a trend change of the network parameter has occurred based on the log information;
a task generation unit that, when a change in the trend occurs, separates subsequent log information from past tasks as a new task;
a storage unit that stores the task;
a deletion unit that deletes the past task stored in the storage unit if the new task is similar to the past task;
A learning management device comprising: The learning management device described in claim 1 further comprises a learning prediction unit that performs continuous learning using the tasks stored in the memory unit and predicts future network parameters. A learning management device as described in claim 1 or 2, wherein the deletion unit calculates the similarity between the new task and the past task using dynamic time warping. A learning management device as described in claim 1 or 2, wherein the network parameter is traffic flow rate. A learning management method executed by a learning management device having a storage unit,
a log acquisition step of acquiring time-series log information that indicates a log of network parameters generated at a predetermined time interval and is used for continuous learning;
a determining step of determining whether a trend change of the network parameter has occurred based on the log information;
a task generation step of dividing subsequent log information into a new task and past tasks when a change in the trend occurs;
a storage step of storing the task in the storage unit;
a deletion step of deleting the past task stored in the storage unit if the new task is similar to the past task;
Learning management methods, including: The learning management method described in claim 5 further includes a learning prediction step of performing continuous learning using the tasks stored in the memory unit and predicting future network parameters. A learning management method as described in claim 5 or 6, wherein the deletion step includes a step of calculating the similarity between the new task and the past task using dynamic time warping. A program for causing a computer to function as a learning management device described in claim 1 or 2.