System, method, and program

Abstract

A first system (10) comprises: an acquisition unit (11) that acquires data provided by an external server or other data-providing device as inference data with which a second system (20) makes an inference according to an inference model; and a specification unit (12) that specifies, from among data including the inference data acquired by the acquisition unit (11), data to be collected from the second system (20) that has made an inference according to an inference model as training data for a learning model to build an inference model.

Description

SYSTEM, APPARATUS, METHOD, AND NON-TRANSITORY COMPUTER-READABLE MEDIUM

[0001] The present disclosure relates to systems, devices, methods, programs, and non-transitory computer-readable media.

[0002] In recent years, the introduction of 5G (5th Generation) has been promoted as a wireless communication technology that achieves large capacity, low latency, and multiple connectivity. In next-generation wireless communication systems including 5G, the openness of the Radio Access Network (RAN) is being promoted to accommodate increasingly sophisticated and complex systems, and the Open Radio Access Network (O-RAN) Alliance is discussing the openness and intelligentization of the RAN.

[0003] Patent Literature 1 related to RAN describes controlling the RAN using AI / ML (Artificial Intelligence / Machine Learning) and distributing resources for learning. Non-Patent Literature 1 related to O-RAN describes a use case using a Non-RT (Real Time) RIC and a Near-RT RIC as RAN Intelligent Controllers (RICs) that intelligently control the RAN using AI / ML. The Near-RT RIC is placed near E2 nodes including O-DUs (O-RAN Distributed Units) and O-CUs (O-RAN Central Units) and controls the RAN in real time. The Non-RT RIC is placed at a location away from the E2 nodes and controls the RAN in non-real time.

[0004] Japanese Patent Application Laid-Open No. 2021-141419

[0005] O-RAN ALLIANCE, O-RAN Working Group 1,"Use Cases Detailed Specification", Technical Specification, O-RAN.WG1.Use-Cases-Detailed-Specification-v08.00, 2022.04.04

[0006] As described above, Patent Document 1 allows for the distribution of learning resources, while Non-Patent Document 1 allows for inference and learning to be performed using Non-RT RIC and Near-RT RIC. For example, according to Non-Patent Document 1, an inference model that infers RAN control and a learning model that performs learning to build the inference model can be placed in either Near-RT RIC or Non-RT RIC, or can be distributed and placed. This allows for learning using the learning model using operational data while executing control using the inference model. However, in order to learn using the learning model, it is necessary to collect data for learning, and it is also necessary to perform predetermined data processing on the collected data, which can make it difficult to perform learning efficiently.

[0007] In view of the above problems, one of the objects of the present disclosure is to provide a system, an apparatus, a method, and a non-transitory computer-readable medium that enable efficient learning.

[0008] The system disclosed herein comprises an acquisition means for acquiring data provided from a data providing device as inference data for other systems to make inferences using an inference model, and an identification means for identifying data collected from other systems that have made inferences using the inference model as learning data for a learning model that constructs the inference model from data including the acquired inference data.

[0009] The system disclosed herein comprises a collection means for collecting data provided from a data providing device as inference data for making inferences using an inference model, and a transmission means for transmitting data identified from data including the collected inference data as learning data for a learning model that constructs the inference model to another system that performs learning using the learning model.

[0010] The device disclosed herein comprises an acquisition means for acquiring data provided from a data providing device as inference data for other systems to make inferences using an inference model, and an identification means for identifying data collected from other systems that have made inferences using the inference model as learning data for a learning model that constructs the inference model from data including the acquired inference data.

[0011] The device disclosed herein comprises a collection means for collecting data provided from a data providing device as inference data for making inferences using an inference model, and a transmission means for transmitting data identified from data including the collected inference data as learning data for a learning model that constructs the inference model to another system that performs learning using the learning model.

[0012] The method disclosed herein acquires data provided by a data providing device as inference data for other systems to make inferences using an inference model, and identifies data to be collected from other systems that have made inferences using the inference model as learning data for a learning model that constructs the inference model from data that includes the acquired inference data.

[0013] The method disclosed herein involves collecting data provided from a data providing device as inference data for making inferences using an inference model, and transmitting data identified from the collected data including the inference data as learning data for a learning model that constructs the inference model to another system that performs learning using the learning model.

[0014] The non-transitory computer-readable medium of the present disclosure is a non-transitory computer-readable medium that stores a program for causing a computer to execute a process of acquiring data provided from a data providing device as inference data for other systems to perform inference using an inference model, and identifying data to be collected from other systems that have performed inference using the inference model as learning data for a learning model that constructs the inference model from data that includes the acquired inference data.

[0015] The non-transitory computer-readable medium of the present disclosure is a non-transitory computer-readable medium storing a program for causing a computer to execute a process of collecting data provided from a data providing device as inference data for making inferences using an inference model, and transmitting data identified from data including the collected inference data as learning data for a learning model that constructs the inference model to another system that performs learning using the learning model.

[0016] According to the present disclosure, it is possible to provide a system, an apparatus, a method, a program, and a non-transitory computer-readable medium that enable efficient learning.

[0017] 1 is a configuration diagram showing an overview of a first system according to an embodiment. 2 is a configuration diagram showing an overview of a second system according to an embodiment. 3 is a configuration diagram showing an overview of a first device according to an embodiment. 4 is a configuration diagram showing an overview of a second device according to an embodiment. 5 is a flowchart showing an overview of a first method according to an embodiment. 6 is a flowchart showing an overview of a second method according to an embodiment. 7 is a configuration diagram showing an example of a configuration of a RAN system according to embodiment 1. 8 is a diagram for explaining a learning data collection method 1 according to embodiment 1. 9 is a diagram for explaining a learning data collection method 2 according to embodiment 1. 10 is a configuration diagram showing an example of a configuration of a Non-RT RIC according to embodiment 1. 11 is a diagram for explaining an example of a determination of a collection method according to embodiment 1. 12 is a configuration diagram showing an example of a configuration of a Near-RT RIC according to embodiment 1. 13 is a flowchart showing an overview of operations in a RAN system according to embodiment 1. 14 is a sequence diagram showing an example of an operation of an inference phase process according to embodiment 1. 15 is a sequence diagram showing another example of an operation of the inference phase process according to embodiment 1. 16 is a sequence diagram showing an example of an operation of the learning phase process according to embodiment 1. 17 is a sequence diagram showing an example of an operation of the inference phase process according to embodiment 2. 18 is a sequence diagram showing an example of an operation of the learning phase process according to embodiment 2. Fig. 10 is a sequence diagram showing an example of operation of an inference phase process according to embodiment 3. Fig. 11 is a sequence diagram showing an example of operation of a learning phase process according to embodiment 3. Fig. 12 is a diagram for explaining a learning data collection method 3 according to embodiment 4. Fig. 13 is a configuration diagram showing an example of the configuration of a RAN system according to embodiment 5. Fig. 14 is a configuration diagram showing an overview of the hardware of a computer according to an embodiment.

[0018] Hereinafter, embodiments will be described with reference to the drawings. In the drawings, the same elements are denoted by the same reference numerals, and redundant description will be omitted as necessary.

[0019] For example, if an inference model is placed in the Near-RT RIC and a learning model is placed in the Non-RT RIC, one possible method is for the Near-RT RIC to collect data such as wireless quality from the E2 node for inference, and for the Non-RT RIC to collect the same data from the E2 node for learning.In response to this, the inventors of the present disclosure have considered an example in which the data collected by the Near-RT RIC for inference is transferred from the Near-RT RIC to the Non-RT RIC as learning data.

[0020] According to Non-Patent Document 1, a use case is envisioned in which external data is acquired from an external server outside the RAN and used for inference and learning together with data collected within the RAN. In the above-mentioned study example, the inventors of the present disclosure discovered the following problem when attempting to collect external data for inference from an external server. In other words, in this case, it is necessary to combine data acquired from the external server with data collected within the RAN and input it into the learning model of the Non-RT RIC to perform learning processing. This combination process requires shaping processes, such as matching the generation times of each piece of data, which takes time. Therefore, performing the combination process in the Non-RT RIC places a load on the Non-RT RIC. Furthermore, when the amount of data collected from the external server becomes enormous, the load on the network transmitting the data from the Near-RT RIC to the Non-RT RIC becomes heavy. For example, there may be a case in which multimedia data, such as image data or sensor data, is collected from an external server for inference. In this way, in the study example, there are cases where a load is placed on the device that processes the data or on the network that transmits the data, making it difficult to perform learning efficiently. Therefore, in the embodiment, the load on the device and the network is reduced, making it possible to perform learning efficiently.

[0021] (Outline of the embodiment) First, an outline of the embodiment will be described. Fig. 1 shows a schematic configuration of a first system 10 according to the embodiment, and Fig. 2 shows a schematic configuration of a second system 20 according to the embodiment. For example, the first system 10 and the second system 20 constitute a system that controls a wireless network such as a RAN. For example, the first system 10 includes a Non-RT RIC, and the second system 20 includes a Near-RT RIC, but is not limited to these.

[0022] As shown in FIG. 1 , the first system 10 includes an acquisition unit 11 and an identification unit 12. The acquisition unit 11 acquires data provided from a data providing device as inference data for the second system 20 to perform inference using an inference model. For example, the data providing device is a server external to a system including the first system 10 and the second system 20. When the first system 10 and the second system 20 control a RAN, the data providing device is a server external to the RAN. Data provided from the external data providing device is data external to the RAN, such as weather information and traffic information. The inference model infers control of a wireless network, such as a RAN, using the inference data. Control of a wireless network, such as control of the operation of the RAN, is, for example, control of a wireless schedule and beams, which is possible by setting an E2 node.

[0023] The identification unit 12 identifies data to be collected from the second system 20, which performed inference using the inference model, as learning data for a learning model that constructs an inference model. For example, the identification unit 12 identifies data to be collected as learning data from data including inference data acquired from a data providing device. The learning model, for example, learns control related to a wireless network such as a RAN using the learning data. The learning model is, for example, included in the first system 10, but may be located outside the first system 10. Identifying data to be collected as learning data can also be considered as determining a collection method for collecting the identified data. The identification unit 12 may identify whether data acquired from the data providing device should be collected from the second system 20. For example, when external data acquired from an external server and RAN data acquired from an E2 node are used for inference of the inference model, the identification unit 12 identifies data to be collected from data including the external data and RAN data used for inference.

[0024] The first system 10 may also include a transfer unit that transfers the inference data acquired from the data providing device to the second system 20, and the identification unit 12 may identify whether or not the inference data transferred to the second system 20 is to be collected from the second system. The first system 10 may also include a storage unit that stores the inference data transferred to the second system 20, and the identification unit 12 may identify whether or not the inference data stored in the storage unit is to be collected from the second system. The first system 10 may also include a synthesis unit that, when the inference data from the data providing device is not to be collected from the second system 20, synthesizes the inference data stored in the storage unit with data collected from the second system 20 to generate learning data to be input to the learning model.

[0025] For example, the identification unit 12 may identify data to be collected based on characteristics of the inference data acquired from the data providing device. Alternatively, the identification unit 12 may identify data to be collected based on instructions input by an operator or the load of the RAN system including the first system 10 and the second system 20.

[0026] As shown in FIG. 2 , the second system 20 includes a collection unit 21 and a transmission unit 22. The collection unit 21 collects data provided from a data providing device as inference data for performing inference using an inference model. For example, the collection unit 21 collects external data provided from an external server via the first system 10. The inference model is, for example, included in the second system 20, but may also be located outside the second system 20. For example, the inference model infers RAN control based on external data acquired via the first system 10 and RAN data such as wireless quality collected from an E2 node.

[0027] The transmitter 22 transmits data identified as learning data for the learning model of the first system 10 to the first system 10 that performs learning using the learning model. The transmitter 22 transmits data identified from data including inference data collected by the collector 21. For example, the transmitter 22 transmits data identified by the first system 10 from data including external data and RAN data used for inference to the first system 10.

[0028] The first system 10 and the second system 20 may each be configured by a single device or multiple devices. FIG. 3 shows an example configuration of a first device 30 according to an embodiment, and FIG. 4 shows an example configuration of a second device 40 according to an embodiment. As shown in FIG. 3, the first device 30 may include the acquisition unit 11 and the identification unit 12 shown in FIG. 1. This example is not limited to this example, and the acquisition unit 11 and the identification unit 12 may be implemented in separate devices. As shown in FIG. 4, the second device 40 may include the collection unit 21 and the transmission unit 22 shown in FIG. 2. This example is not limited to this example, and the collection unit 21 and the transmission unit 22 may be implemented in separate devices. As with the first system 10 and the second system 20, for example, the first device 30 may be a non-RT RIC, and the second device 40 may be a near-RT RIC.

[0029] Furthermore, part or all of the first system 10 and the second system 20 may be placed at an edge or in a cloud using virtualization technology or the like. They may be placed in a specific location, or may be distributed across multiple locations. An edge is a location or infrastructure on the base station side that includes an O-DU and an O-CU. A cloud is a location or infrastructure on the core network side that is distant from the base station. For example, the acquisition unit 11 and the identification unit 12 may be placed in the cloud, and the collection unit 21 and the transmission unit 22 may be placed at an edge. Furthermore, the acquisition unit 11, the identification unit 12, the collection unit 21, and the transmission unit 22 may each be distributed.

[0030] Fig. 5 shows a first method according to an embodiment, and Fig. 6 shows a second method according to an embodiment. For example, the first method is performed by the first system 10 of Fig. 1 or the first device 30 of Fig. 3. The second method is performed by the second system 20 of Fig. 2 or the second device 40 of Fig. 4.

[0031] 5, the acquiring unit 11 acquires data provided from the data providing device as inference data for the second system 20 to perform inference using the inference model (S11). Next, the identifying unit 12 identifies, from the data including the acquired inference data, data to be collected from the second system 20 that performed inference using the inference model as learning data to be used by the learning model (S12). For example, the first system 10 transfers the inference data acquired from the data providing device to the second system 20.

[0032] 6 , the collection unit 21 collects inference data for performing inference using the inference model (S21). For example, the collection unit 21 acquires inference data provided from a data providing device via the first system 10. Next, the transmission unit 22 transmits data identified from the collected data including the inference data as learning data for the learning model of the first system 10 to the first system 10, which performs learning using the learning model (S22). The transmission unit 22 transmits the data identified by the first system 10 as learning data to the first system 10. For example, the first system 10 uses the learning data collected from the second system 20 to perform learning using the learning model, and applies the learned learning model to the inference model of the second system 20.

[0033] Thus, in this embodiment, a first system, such as a Non-RT RIC, acquires inference data for an inference model and identifies data collected from a second system, such as a Near-RT RIC, as training data to be used by the training model. Furthermore, the second system acquires inference data from a first system or the like, and transmits the data identified by the first system or the like as training data to be used by the training model. By identifying the data to be collected by the Near-RT RIC from the Non-RT RIC, it is possible to select a system or device that performs data processing, such as synthesizing training data, thereby reducing the load due to data processing. Furthermore, it is possible to adjust the amount of data to be transferred, thereby reducing the load on the network that transfers data. Therefore, training can be performed efficiently.

[0034] (Embodiment 1) Next, a description will be given of embodiment 1. In this embodiment, an example will be described in which a Non-RT RIC acquires external data from an external server and switches the method of collecting learning data depending on the characteristics of the data, etc.

[0035] Fig. 7 shows an example of the configuration of a RAN system 1 according to this embodiment. As shown in Fig. 7, the RAN system 1 according to this embodiment includes a Non-RT RIC 100, a Near-RT RIC 200, an E2 node 300, and an external server 400. The Non-RT RIC 100 is arranged in an SMO (Service Management and Orchestration) 500 that manages and orchestrates the RAN. Note that functions included in the SMO 500 may be described as functions of the Non-RT RIC 100.

[0036] The SMO 500 and the Near-RT RIC 200, and the SMO 500 and the E2 node 300 are communicatively connected via the O1 interface. It can also be said that the Non-RT RIC 100 and the Near-RT RIC 200, and the Non-RT RIC 100 and the E2 node 300 are communicatively connected via the O1 interface via the SMO 500. Note that the Non-RT RIC 100 and the Near-RT RIC 200, and the Non-RT RIC 100 and the E2 node 300 may be described as being connected via the O1 interface. The O1 interface is an interface primarily used to send and receive data and messages required for operation and management. Note that an interface is a connection interface defined by a communication protocol for sending and receiving data and messages, and includes logical transmission paths and networks, as well as physical transmission paths and networks.

[0037] The Non-RT RIC 100 and the Near-RT RIC 200 are communicatively connected via an A1 interface. The Near-RT RIC 200 and the E2 node 300 are connected via an E2 interface. The A1 interface and the E2 interface are interfaces for sending and receiving data and messages required mainly for control purposes.

[0038] In O-RAN, the Policy Management Service (A1-P), Enrichment Information Service (A1-EI), and ML Model Management Service (A1-ML) are defined as services provided over the A1 interface. In the Policy Management Service, the Non-RT RIC provides guidance for RAN optimization, i.e., the A1 policy, to the Near-RT RIC. The A1 policy is a control policy related to RAN control. The Enrichment Information Service optimizes RAN performance by making Enrichment Information, which cannot be collected within the RAN, available to the Near-RT RIC. In the ML Model Management Service, the Non-RT RIC provides Enrichment Information to support inference using an inference model in the Near-RT RIC. In this embodiment, the Enrichment Information Service of the A1 interface is used to transfer Enrichment Information from the Non-RT RIC to the Near-RT RIC during inference of the Near-RT RIC. Note that Enrichment Information is also referred to as EI data.

[0039] The Non-RT RIC 100 and the external server 400 are connected to each other via any interface that is not defined by the O-RAN. The interface between the Non-RT RIC 100 and the external server 400 may be an interface that allows a general application server to provide data. For example, HTTP (Hypertext Transfer Protocol) for a web server or another API (Application Programming Interface) may be used.

[0040] The E2 node 300 is a node that constitutes the RAN and includes an O-DU and an O-CU. The RAN is a wireless network accessed by UEs (User Equipment) and is connected to a core network such as a 5G Core network (5GC) or an Evolved Packet Core (EPC). The RAN may include an O-RAN Remote Unit (O-RU) that constitutes an antenna. The UE is a terminal device that connects to the RAN and performs wireless communication, and may be a mobile phone, smartphone, tablet terminal, IoT (Internet of Things) terminal, or the like. The UE may also be an application device such as a robot, drone, or self-driving car that implements terminal functions.

[0041] The E2 node 300 including the O-DU and O-CU provides base station functionality. The base station is, for example, a gNB (next generation Node B) or an eNB (evolved Node B), but is not limited to these. The O-DU and O-CU are examples of nodes that provide base station functionality, and may be other network nodes.

[0042] The O-DU is a logical node that provides the radio signal control function and Layer 2 control function of the base station. The O-DU accommodates the O-RU, controls the antenna radio signal (beam) in the O-RU, and performs protocol processing such as MAC (Media Access Control) and RLC (Radio Link Control) required between the O-RU and O-CU.

[0043] The O-CU is a logical node that provides the base station's radio resource control function and data processing function higher than Layer 2. The O-CU accommodates the O-DU and performs data transmission and reception via the accommodated O-DU, QoS (Quality of Service) control, cell / UE management, handover control, and protocol processing such as PDCP (Packet Data Convergence Protocol), SDAP (Service Data Adaptation Protocol), and RRC (Radio Resource Control) required between the O-DU and the core network.

[0044] The E2 node 300 may include any number of O-DUs and O-CUs, one or more. In other words, it may include multiple base stations. The E2 node 300 may also be implemented by a virtual machine running on an edge virtualization platform. The edge platform may be MEC (Multi-access Edge Computing). The O-DU and O-CU may be vDUs (virtualized Distributed Units) and vCUs (virtualized Central Units), and may constitute a virtual base station. The O-DU and O-CU may also be physical DUs and CUs. The E2 node 300 may also be a base station device including the functions of the O-DU and O-CU.

[0045] The external server 400 is a server external to the RAN including at least the E2 node 300. The external server 400 can also be considered a server external to a system including the E2 node 300, the Non-RT RIC 100, and the Near-RT RIC 200. The external server 400 is a data providing device that provides EI (Enriched Information) data. The EI data is external data defined by the A1 interface as described above that cannot be collected within the RAN, and is also inference data used by the inference model of the Near-RT RIC 200 for inference. The external server 400 includes an application server that provides various data that can be used for inference of the inference model. For example, the external server 400 may be a web server or an SNS (Social Networking Service) server. The EI data includes, for example, weather information, traffic information, map information, etc. The external server 400 may be, for example, a server on the Internet, as long as it can provide the EI data to the Non-RT RIC 100. The external server 400 may be a physical server or a virtual server on the cloud.

[0046] The Near-RT RIC 200 is a logic function that controls and optimizes the RAN in real time. The Near-RT RIC 200 controls the RAN in a short control period of, for example, 1 second or less. The Near-RT RIC 200 collects and analyzes RAN data from the E2 node 300 via the E2 interface, and controls the E2 node according to the RAN data. In addition, in this embodiment, the Near-RT RIC 200 collects EI data from the external server 400 via the Non-RT RIC 100 via the A1 interface, and controls the E2 node according to the EI data and RAN data. For example, the Near-RT RIC 200 performs control according to the EI data and RAN data in accordance with a control policy, i.e., an A1 policy, acquired from the Non-RT RIC 100 via the A1 interface. The RAN data is radio-related data related to radio communication in the RAN, and may include radio quality data and location information for each UE, and the number of active UEs for each base station (cell). For example, the radio quality data may be acquired from the O-DU, or handover information may be acquired from the O-CU. The Near-RT RIC 200 is located in the same location as the E2 node 300 or in a location close to the E2 node 300. For example, the Near-RT RIC 200 may be implemented in a virtual machine on the same edge as the E2 node 300.

[0047] Some functions of the Near-RT RIC 200 are realized by xApp (Near-RT RIC Application). xApp includes an application that performs analysis and inference of RAN data and EI data. For example, xApp includes an inference unit 210 that performs inference using an inference model, which is a trained model. The inference unit 210 analyzes inference data including RAN data and EI data and controls the RAN using the inference model. The Near-RT RIC 200 also includes an inference data storage unit 220 that stores inference data including EI data and RAN data used by the inference unit 210 for inference.

[0048] The Non-RT RIC 100 is a logic function that controls and optimizes the RAN in non-real time. The Non-RT RIC 100 controls the RAN at a long control period of, for example, 1 second or more. The Non-RT RIC 100 manages control policies, manages the operation of the E2 node 300 and the Near-RT RIC 200, learns (trains) learning models, and updates inference models. For example, the Non-RT RIC 100 generates control policies and notifies the Near-RT RIC 200 of the generated control policies via the A1 interface. The Non-RT RIC 100 also manages and sets the configuration information (Configuration) of the E2 node 300 based on data acquired from the E2 node 300 and the Near-RT RIC 200 via the O1 interface. Furthermore, the Non-RT RIC 100 acquires EI data from the external server 400 and transfers the acquired EI data to the Near-RT RIC 200. The Non-RT RIC 100 and the SMO 500 are located at a location away from the E2 node 300 and the Near-RT RIC 200, for example, on the cloud.

[0049] Some functions of the Non-RT RIC 100 are realized by rApp (Non-RT RIC Application). rApp includes applications that generate control policies and manage the inference model of the Near-RT RIC 200. For example, rApp includes a learner 110 that performs learning using a learning model. The learner 110 generates a learning model that learns RAN control using learning data acquired from the E2 node 300 and the Near-RT RIC 200 via the O1 interface, and applies the generated trained learning model to the xApp of the Near-RT RIC 200. Note that applying a trained learning model to an inference model is also referred to as "deploying." Deploying refers to placing and deploying a model in an application execution environment and making the model executable. The Non-RT RIC 100 also includes an EI data storage unit 120 that stores EI data acquired from the external server 400.

[0050] 8 shows learning data collection method 1 (first collection method) according to this embodiment, and FIG. 9 shows learning data collection method 2 (second collection method) according to this embodiment. In this embodiment, the Non-RT RIC 100 collects learning data from the Near-RT RIC 200 using collection method 1 or collection method 2. It can also be said that learning data is transferred from the Near-RT RIC 200 to the Non-RT RIC 100 using either method.

[0051] 8, collection method 1 is a method for transferring lightened data from the Near-RT RIC 200 to the Non-RT RIC 100. This can reduce the load on the O1 interface.

[0052] In collection method 1, during inference, the Non-RT RIC 100 acquires EI data for inference from the external server 400, stores the acquired EI data in the EI data storage unit 120, and transfers it to the Near-RT RIC 200. The Near-RT RIC 200 performs inference using the inference unit 210 with the EI data acquired from the Non-RT RIC 10 and the RAN data collected from the E2 node 300 as inference data, and stores the inference data used for the inference in the inference data storage unit 220.

[0053] Furthermore, in collection method 1, during learning, when Near-RT RIC 200 transfers the inference data stored in inference data storage unit 220 to Non-RT RIC 100 as learning data, it removes the EI data stored in Non-RT RIC 100 to make the data lighter, and transfers this lighter data to Non-RT RIC 100. Non-RT RIC 100 combines the lighter learning data collected from Near-RT RIC 200 with the EI data stored in EI data storage unit 120, and performs learning using learner 110 using the combined learning data.

[0054] 9, collection method 2 is a method in which all data necessary for learning is transferred from the Near-RT RIC 200 to the Non-RT RIC 100. This makes it possible to reduce the processing load on the Non-RT RIC 100.

[0055] In collection method 2, during inference, the Non-RT RIC 100 acquires EI data for inference from the external server 400 and transfers the acquired EI data to the Near-RT RIC 200 without storing it in the EI data storage unit 120. The Near-RT RIC 200 performs inference using the inference unit 210 with the EI data acquired from the Non-RT RIC 10 and the RAN data collected from the E2 node 300 as inference data, and stores the inference data used for the inference in the inference data storage unit 220.

[0056] Furthermore, in collection method 2, during learning, the Near-RT RIC 200 transfers all inference data, including EI data and RAN data, stored in the inference data storage unit 220 as learning data to the Non-RT RIC 100. The Non-RT RIC 100 performs learning using the learning data including all received data with the learning device 110.

[0057] FIG. 10 shows an example configuration of a Non-RT RIC 100 according to this embodiment. As shown in FIG. 10, the Non-RT RIC 100 includes the above-described learning device 110 and EI data storage unit 120. For example, the learning device 110 includes a learning unit 111 and a model storage unit 112. The Non-RT RIC 100 also includes an O1 communication unit 101, an A1 communication unit 102, an external communication unit 103, a data collection unit 131, a method determination unit 132, a data transfer unit 133, and a system management unit 134. The O1 communication unit 101 may be included in the SMO 500. Note that this configuration is merely an example, and other configurations may be used as long as the operation according to this embodiment, described below, is possible. The Non-RT RIC may also include a configuration for implementing functions required for the Non-RT RIC.

[0058] The O1 communication unit 101 is a communication unit that communicates with the Near-RT RIC 200 via the O1 interface. For example, the O1 communication unit 101 transmits and receives various data including learning data and control messages to and from the Near-RT RIC 200 in accordance with the communication method defined as the O1 interface. The O1 communication unit 101 can also transmit and receive necessary data and control messages to and from the E2 node 300 via the O1 interface.

[0059] The A1 communication unit 102 is a communication unit that communicates with the Near-RT RIC 200 via the A1 interface. For example, the A1 communication unit 102 transmits and receives various data including EI data and control messages including control policies between the Near-RT RIC 200 and the A1 communication unit 102 in accordance with the communication method defined as the A1 interface.

[0060] The external communication unit 103 is a communication unit that communicates with the external server 400 via an arbitrary interface. For example, the external communication unit 103 acquires EI data from the external server 400 according to a predetermined communication method such as HTTP.

[0061] The data collection unit 131 collects data required for learning, management, transfer, etc. from the external server 400, the Near-RT RIC 200, and the E2 node 300. During inference, EI data is acquired from the external server 400 via the external communication unit 103 in accordance with instructions from the Near-RT RIC 200. For example, the data collection unit 131 and the external communication unit 103 correspond to the acquisition unit 11 in FIG. 1 . During learning, learning data transferred from the Near-RT RIC 200 is collected via the O1 communication unit 101 and the O1 interface. Necessary data is also collected from the E2 node 300.

[0062] The method determination unit 132 determines the method of collecting learning data shown in Figures 8 and 9. By determining the collection method, the method determination unit 132 can also be said to identify the data to be collected from the Near-RT RIC 200. For example, the method determination unit 132 corresponds to the identification unit 12 in Figure 1.

[0063] For example, the method determination unit 132 determines the collection method based on the characteristics of EI data acquired for inference from the external server 400 during inference. Determining the collection method also means selecting a collection method. The method determination unit 132 may determine the collection method based on the characteristics of the acquired EI data each time EI data is acquired from the external server 400. Furthermore, the method determination unit 132 may determine the collection method based on the characteristics of the acquired EI data at a specific timing, such as when EI data is acquired for the first time from the external server 400 in response to a request from the Near-RT RIC 200. The collection method may be determined each time EI data is acquired a predetermined number of times, or may be determined each time a predetermined time has elapsed.

[0064] FIG. 11 shows a specific example of determining a collection method based on data characteristics. As shown in FIG. 11, the collection method is determined based on a feature index representing the data characteristics. The feature index includes, for example, data size, number of parameters, and sampling period, but is not limited to these, and other indexes may be used. The collection method may also be determined based on any one of the feature indexes of data size, number of parameters, and sampling period, or may be determined by combining any feature indexes. For example, the collection method may be determined based on the sampling period and data size. The collection method may also be determined based on the sampling period, data size, and number of parameters.

[0065] In an example using data size, the method determination unit 132 selects collection method 1 or collection method 2 depending on whether the data size is large or small. Specifically, if the data size of the acquired EI data is large, collection method 1 is selected, in which the EI data is stored in the Non-RT RIC 100. If the data size of the EI data accounts for a large proportion of the overall learning data, it is determined that it is appropriate to use collection method 1 to reduce the amount of data transferred from the Near-RT RIC 200 to the Non-RT RIC 100. For example, collection method 1 is selected when the data size of the EI data is larger than a predetermined threshold. The total data size required for learning may be acquired from a learning device, etc., and collection method 1 may be selected when the proportion of the EI data to the total data size is larger than a predetermined threshold. Examples of EI data with large data sizes include image data and map information covering a wide area or with high accuracy. Note that, since the data size of these data is expected to be large, collection method 1 may be selected when the EI data type is map information, multimedia data, or the like.

[0066] On the other hand, if the data size of the acquired EI data is smaller than a predetermined threshold, the method determination unit 132 selects collection method 2, which does not require the Non-RT RIC 100 to store the EI data. Collection method 2 may be selected if the ratio of the EI data to the total data size required for learning is smaller than a predetermined threshold. If the data size of the EI data is small, there is little effect in reducing the amount of data transferred from the Near-RT RIC 200 to the Non-RT RIC 100. Therefore, it is determined that it is more appropriate to reduce the processing cost of combining the data collected from the Near-RT RIC 200 and the EI data stored in the Non-RT RIC 100 as learning data to be used in the Non-RT RIC 100, and the data used for inference in the Near-RT RIC 200 is used directly for learning. Examples of EI data with small data sizes include time-series log data of a specific application and log data from a sensor device. In addition, since the data size of these data is expected to be small, collection method 2 may be selected when the data type of EI data is log data, text data, or the like.

[0067] Furthermore, in an example in which the number of parameters is used, the method determination unit 132 selects collection method 1 or collection method 2 depending on whether the number of parameters is large or small. Specifically, if the number of parameters of the acquired EI data is greater than a predetermined threshold, collection method 2 is selected. The number of parameters is the number of parameters that make up the EI data, such as the number of variables or the number of data. For example, if the number of parameters is enormous, such as in the case of camera sensor data, it is determined that it is more appropriate to reduce the processing cost of combining the data collected from the near-RT RIC 200 and the EI data stored in the non-RT RIC 100 as learning data to be used in the non-RT RIC 100.

[0068] On the other hand, if the number of parameters of the acquired EI data is less than a predetermined threshold, the method determination unit 132 selects collection method 1. For example, if the number of parameters of the EI data is small, the processing cost of combining the data collected from the Near-RT RIC 200 and the EI data stored in the Non-RT RIC 100 as learning data to be used in the Non-RT RIC 100 is small, and therefore it is determined that it is appropriate to reduce the amount of data transferred from the Near-RT RIC 200 to the Non-RT RIC 100.

[0069] In this example, when a sampling period is used, the method determination unit 132 determines the collection method based on the sampling period and data size. For example, collection method 1 or collection method 2 is selected depending on whether the data volume determined by the sampling period and data size is large or small. The collection method may also be determined based solely on the sampling period. The sampling period is the data collection period, collection interval, or number of collections in a predetermined period. Specifically, the total data volume or the amount of data collected in a predetermined period is calculated from the data size and sampling period. If the calculated total data volume or the amount of data in a predetermined period is greater than a predetermined threshold, collection method 1 is selected. For example, even if the data size is small, collecting data at short intervals increases the final data size. In this case, the proportion of the data in the entire training data is large, so it is determined that collection method 1 is selected and it is appropriate to reduce the amount of data transferred from the Near-RT RIC 200 to the Non-RT RIC 100.

[0070] On the other hand, if the amount of data calculated from the data size and sampling period is smaller than a predetermined threshold, the method determination unit 132 selects collection method 2. Even if the data size is large, if data is collected at long intervals, there is little effect from reducing the amount of data transferred from the Near-RT RIC 200 to the Non-RT RIC 100, so collection method 2 is selected.

[0071] In this way, based on the characteristics of the data, one of the collection methods is selected so that learning processing can be performed efficiently in the entire system including the Non-RT RIC 100 and the Near-RT RIC 200. Note that, as shown in Fig. 11, the characteristic index and the collection method may be associated in advance, and the collection method may be determined on a rule-based basis using an associated table or the like, or the relationship between the characteristic index and the optimal collection method may be machine-learned, and the collection method may be determined on a machine-learning basis using the learned learning model.

[0072] The method determination unit 132 may determine the collection method based on other conditions, not limited to the characteristics of the EI data. For example, the collection method may be switched based on instructions from an operator. Alternatively, the collection method may be set for each time period and switched according to the time. Furthermore, the collection method may be selected according to the load on the RAN system 1. For example, the system management unit 134 may determine the load on each device and each interface based on the data collected from the E2 node 300 and the Near-RT RIC 200, and select the collection method according to the determined load. When the load on the O1 interface is high, collection method 1 may be selected to reduce the load on the O1 interface. When the load on the Non-RT RIC 100 is high, collection method 2 may be selected to reduce the load on the Non-RT RIC 100.

[0073] The data transfer unit 133 transfers the EI data for inference acquired from the external server 400 to the Near-RT RIC 200 via the A1 interface via the A1 communication unit 102. When transferring, the acquired EI data is stored in the EI data storage unit 120 in accordance with the collection method determined by the method determination unit 132. The method determination unit 132 notifies the Near-RT RIC 200 of the collection method corresponding to the EI data by transmitting the determined collection method together with the EI data to be transferred.

[0074] The system management unit 134 manages the settings and operations of the RAN system including the E2 node 300 and the Near-RT RIC 200. The functions of the system management unit 134 may be realized by executing an rApp for system management processing. For example, the system management unit 134 is a policy generation unit that generates a control policy. The system management unit 134 may generate the control policy based on instructions input from an operator or an external device, or may generate the control policy based on data acquired from the E2 node 300 and the Near-RT RIC 200. The system management unit 134 notifies the Near-RT RIC 200 of the generated control policy via the A1 communication unit 102 and the A1 interface.

[0075] The model storage unit 112 stores a learning model for constructing an inference model of the Near-RT RIC 200. The learning model learns RAN control according to RAN data and EI data. The learning model is, for example, a model that performs learning to analyze and predict time-series data. The learning model may be a CNN (Convolutional Neural Network), an RNN (Recurrent Neural Network), an LSTM (Long-Short Term Model), or other neural networks. The learning model is not limited to a neural network, and may be other machine learning models.

[0076] The learning unit 111 performs machine learning using training data collected from the Near-RT RIC 200 according to the collection method. The function of the learning unit 111 may be realized by executing an rApp for learning processing. The learning unit 111 performs necessary data processing to input the acquired training data into a training model. For example, when collecting training data using collection method 1, the learning unit 111 combines the training data acquired from the Near-RT RIC 200 with the EI data stored in the EI data storage unit 120. That is, the learning unit 111 includes a data synthesis unit that synthesizes the training data. The data synthesis process includes shaping processes such as matching the occurrence times of each piece of data. The learning unit 111 performs machine learning such as deep learning to generate a trained training model. The learning unit 111 inputs the training data into the training model in the model storage unit 112 and trains the training model. For example, the learning data includes EI data from the external server 400 and RAN data from the O-DU and O-CU, and by using these data, analysis and control according to the RAN data are learned. Furthermore, the learning data may include an inference result inferred by the Near-RT RIC 200, and the inference result may be used to train the learning model. The learning unit 111 stores the learned learning model in the model storage unit 112, and further transmits the learned learning model to the Near-RT RIC 200, where it is applied to the inference model.

[0077] FIG. 12 shows an example configuration of a Near-RT RIC 200 according to this embodiment. As shown in FIG. 12, the Near-RT RIC 200 includes the above-described inference unit 210 and inference data storage unit 220. For example, the inference unit 210 includes an inference unit 211 and a model storage unit 212. The Near-RT RIC 200 also includes an E2 communication unit 201, an O1 communication unit 202, an A1 communication unit 203, a data collection unit 231, a data extraction unit 232, and a data transfer unit 233. Note that this configuration is merely an example, and other configurations may be used as long as the operation according to this embodiment, which will be described later, is possible. Furthermore, a configuration for realizing functions required for the Near-RT RIC may also be included.

[0078] The E2 communication unit 201 is a communication unit that communicates with the E2 node 300 via the E2 interface. For example, the E2 communication unit 201 transmits and receives various data including RAN data and control messages between the E2 node 300, which is an O-DU or an O-CU, in accordance with a communication method defined as the E2 interface.

[0079] The O1 communication unit 202 is a communication unit that communicates with the Non-RT RIC 100 via the O1 interface. For example, the O1 communication unit 202 transmits and receives various data including learning data and control messages to and from the Non-RT RIC 100 in accordance with the communication method defined as the O1 interface.

[0080] The A1 communication unit 203 is a communication unit that communicates with the Non-RT RIC 100 via the A1 interface. For example, the A1 communication unit 203 transmits and receives various data including EI data and control messages including control policies to and from the Non-RT RIC 100 in accordance with the communication method defined as the A1 interface.

[0081] The data collection unit 231 collects data necessary for inference, control, and the like from the Non-RT RIC 100 and the E2 node 300. During inference, the data collection unit 231 collects EI data from the external server 400 via the A1 interface via the A1 communication unit 203 and the Non-RT RIC 100. For example, the data collection unit 231 and the A1 communication unit 203 correspond to the collection unit 21 in FIG. 2. The data collection unit 231 also collects RAN data from the E2 node 300 via the E2 communication unit 201 and the E2 interface. The data collection unit 231 periodically collects EI data and RAN data as inference data used for inference by the inference model of the inference unit 210. The data collection unit 231 may instruct the Non-RT RIC 100 and the E2 node 300 on the data to be collected and the collection period. The data collection unit 231 outputs the collected EI data and RAN data to the inference unit 211 as inference data, and stores the data in the inference data storage unit 220. For example, a collection method is specified for the EI data acquired from the Non-RT RIC 100, and the EI data and the collection method are associated with each other and stored in the inference data storage unit 220.

[0082] The data extraction unit 232 extracts data to be transferred as learning data from the inference data stored in the inference data storage unit 220. The data extraction unit 232 extracts data according to the collection method set in the stored EI data. As a result, data specified by the collection method determined by the Non-RT RIC 100 is extracted. That is, the data extraction unit 232 extracts data excluding EI data specified in collection method 1 as learning data. In other words, if the EI data is from collection method 1, the EI data is not extracted, and if the EI data is from collection method 2, the EI data is extracted. As a result, the E1 data specified in collection method 2 and the RAN data are extracted as learning data. Note that the learning data to be transferred may include the inference result data of the inference unit 210.

[0083] The data transfer unit 233 transfers the extracted learning data to the Non-RT RIC 100 via the O1 communication unit 202 and the O1 interface. For example, the data transfer unit 233 and the O1 communication unit 202 correspond to the transmission unit 22 in FIG. 2. The data transfer unit 233 transmits learning data in accordance with an instruction from the Non-RT RIC 100. Learning data specified by the Non-RT RIC 100 may be transmitted at a specified timing. Furthermore, learning data may be transmitted depending on the communication status of the O1 interface.

[0084] The model storage unit 212 stores an inference model used in the inference process by the inference unit 211. The inference model is a trained model that infers control of the E2 node 300 according to the RAN data and the EI data. The inference model is the same model as the trained model of the Non-RT RIC 100, and is a model that can analyze and predict time-series data, for example.

[0085] The inference unit 211 analyzes the collected RAN data and EI data and infers (specifies) control of the E2 node based on the analysis results. The function of the inference unit 211 may be realized by executing an xApp for inference processing. The inference unit 211 analyzes data using an inference model stored in the model storage unit 212 and specifies control content (control information). The inference unit 211 inputs the collected RAN data and EI data into the inference model and specifies control content for the E2 node 300 based on the RAN data and EI data. The inference unit 211 may also infer (predict) multiple control content and specify the control content to be used for control according to a control policy. The inference unit 211 outputs the result of specifying the control content, i.e., the inference result, as control information. For example, the inference unit 211 predicts future wireless quality around the UE based on wireless quality and weather information, specifies the wireless intensity and modulation method to be set for the E2 node 300 based on the predicted wireless quality, and outputs control information for setting the corresponding E2 node 300. The inference unit 211 transmits control information indicating the identified control content to the O-DU or O-CU of the E2 node 300 via the E2 interface via the E2 communication unit 201. Furthermore, the inference unit 211 may store the inference result (control information) in the inference data storage unit 220.

[0086] 13 shows an overview of the operation of the RAN system 1 according to this embodiment. In this example, the inference phase process is followed by the learning phase process, but the inference phase process and the learning phase process may be executed in parallel.

[0087] As shown in FIG. 13, the RAN system 1 executes inference phase processing (S101). The Near-RT RIC 200 collects inference data from the Non-RT RIC 100 and the E2 node 300, and uses the collected inference data to infer control of the RAN. The Near-RT RIC 200 controls the E2 node 300 based on the inference results. The Near-RT RIC 200 repeatedly collects inference data and executes inference. The Near-RT RIC 200 also stores the inference data used for inference. The Near-RT RIC 200 may start storing inference data when instructed by the Non-RT RIC 100.

[0088] Next, the RAN system 1 determines whether to start collecting training data (S102). If the training data collection is started, the RAN system 1 executes a training phase process (S103). For example, when training of the training model becomes necessary, the RAN system 1 may start accumulating inference data to be used as training data, and when the accumulation of inference data necessary for training is completed, the RAN system 1 may start collecting training data. For example, the non-RT RIC 100 may determine that training of the training model is necessary when an instruction is input from an operator or an external device, when the on-site environment including the UE changes, periodically, or when the accuracy of the inference model decreases. The change in the environment may be detected from a change in wireless quality, or a signal indicating an environmental change such as a layout change may be input. The accuracy of the inference model may be determined from the inference result of the inference model and RAN data, etc.

[0089] When a predetermined amount of data or an amount of data instructed by the Non-RT RIC 100 has been accumulated in the Near-RT RIC 200 for inference, the Non-RT RIC 100 determines to start collecting learning data. The Non-RT RIC 100 may also determine to start collecting learning data when a predetermined accumulation period or an accumulation period instructed by the Non-RT RIC 100 has ended. For example, when the Near-RT RIC 200 notifies the Non-RT RIC 100 that accumulation of inference data has been completed, the Non-RT RIC 100 starts collecting learning data and trains a learning model using the collected learning data. For example, one hour's worth of learning data may be collected and used for training. The trained learning model generated by learning is applied to the Near-RT RIC 200 as an inference model. When learning of the learning model is necessary, the Non-RT RIC 100 repeatedly collects learning data and learns the learning model in steps S102 and S103.

[0090] FIG. 14 is a sequence diagram showing an example of the operation of the inference phase process (S101) of FIG. 13. FIG. 14 shows an example in which the Non-RT RIC 100 determines the collection method each time it repeatedly acquires external data from the external server 400. Note that FIG. 14 is just an example, and the order of some of the processes may be changed, or some of the processes may be executed in parallel. For example, S202 may be executed after S201, or S201 and S202 may be executed in parallel. S203 to S207 may be executed after S208, or S203 to S207 and S208 may be executed in parallel.

[0091] As shown in FIG. 14 , the Near-RT RIC 200 transmits an RIC subscription message to the E2 node 300 via the E2 interface to request RAN data for inference (S201). For example, the data collection unit 231 requests the transfer of RAN data to collect RAN data to be used for inference. The RIC subscription message is a message specified by the E2 interface and requests the periodic transfer of RAN data. The data collection unit 231, for example, specifies information identifying the data to be transferred and the timing of data transfer in the RIC subscription message. The information identifying the specified data may indicate the ID or name of the data, or may include the size of the data. In the case of data for each UE or data with a base station (cell), information identifying the UE or base station may be specified. The specified timing may include the transfer cycle or interval, the transfer time, the number of transfers, and the transfer period. The RIC subscription message may request multiple pieces of RAN data. The RIC subscription message may include information for identifying the data transfer source. The information for identifying the data transfer source may be information for identifying an O-DU or an O-CU. Thereafter, the E2 node 300 repeatedly transfers the RAN data at the timing specified in the RIC subscription message (S208).

[0092] Next, the Near-RT RIC 200 sends a Create EI Job message to the Non-RT RIC 100 via the A1 interface to request EI data for inference (S202). For example, the data collection unit 231 requests the transfer of EI data to collect EI data to be used for inference. The Create EI Job message is a message specified by the A1 interface and requests the periodic transfer of EI data. For example, the data collection unit 231 specifies, in the Create EI Job message, information identifying the data provider, information identifying the data to be transferred, and the timing of the data transfer. The information identifying the data provider may be a URL or IP address identifying the external server 400. The information identifying the data may indicate the ID or name of the data, or may include the size of the data. The timing may include the transfer cycle or interval, the transfer time, the number of transfers, and the transfer period. The Create EI Job message may request multiple EI data. Thereafter, the Non-RT RIC 100 repeatedly acquires and transfers the EI data at the timing specified in the Create EI Job message (S203 to S207). Note that the EI data may be requested and transferred between the Near-RT RIC 200 and the Non-RT RIC 100 via the O1 interface.

[0093] Next, when the Non-RT RIC 100 receives the Create EI Job message from the Near-RT RIC 200, it requests EI data for inference from the external server 400 via the interface with the external server 400 (S203). For example, the data collection unit 131 requests the data specified in the Create EI Job message from the external server 400 upon receiving the Create EI Job message or at the timing specified in the Create EI Job message. The data collection unit 131 transmits a data request message that can be used via the interface with the external server 400. The data request message may be, for example, an HTTP Get request message. The data collection unit 131 may specify information identifying the data provider or information identifying the data to be transferred in the data request message. For example, the information identifying the data provider or the information identifying the data may be the information specified in the Create EI Job message. That is, the information identifying the data provider may be a URL or IP address that identifies the external server 400. The information identifying the data may indicate the ID or name of the data, or may include the size of the data. A data request message may request multiple pieces of data. Note that the data request message may specify the timing for transferring the data, and the external server 400 may periodically transmit the data at the specified timing.

[0094] Next, when the external server 400 receives a request for EI data for inference from the Non-RT RIC 100, it transfers the requested EI data to the Non-RT RIC 100 via the interface with the Non-RT RIC 100 (S204). When the external server 400 receives the data request message, it transmits the EI data specified in the received data request message to the Non-RT RIC 100. The external server 400 transmits a data transfer message that can be used in the interface with the Non-RT RIC 100. The data transfer message may be, for example, an HTTP Get response message. A plurality of data may be transferred in the data transfer message in accordance with the data request message.

[0095] Next, upon receiving EI data from the external server 400, the Non-RT RIC 100 determines a collection method (S205). For example, the method determination unit 132 determines the collection method based on data characteristics of the acquired EI data, such as the data size, number of parameters, and sampling period. For example, the method determination unit 132 may extract the data size from the acquired EI data, and select collection method 1 if the extracted data size is greater than a predetermined threshold, and select collection method 2 if the data size is less than the predetermined threshold. Alternatively, the method determination unit 132 may extract the number of parameters from the acquired EI data, and select collection method 2 if the number of extracted parameters is greater than a predetermined threshold, and select collection method 1 if the number of parameters is less than the predetermined threshold. Alternatively, the method determination unit 132 may extract the data size from the acquired EI data, and, using the collection period instructed by the Near-RT RIC as the sampling period, select collection method 1 if the amount of data calculated from the data size and sampling period is greater than a predetermined threshold, and select collection method 2 if the calculated amount of data is less than the predetermined threshold. When multiple EI data are acquired, the collection method may be determined for each EI data based on the characteristics of the data, or the collection method may be determined for each of the multiple EI data based on the characteristics of the entire data including the multiple EI data.

[0096] Next, if the determined collection method is collection method 1, the Non-RT RIC 100 stores the acquired EI data in the EI data storage unit 120 (S206). For example, the data transfer unit 133 determines the collection method when transferring EI data, and if collection method 1 is selected, stores the acquired EI data in the EI data storage unit 120, and if collection method 2 is selected, does not store the acquired EI data in the EI data storage unit 120.

[0097] Next, the Non-RT RIC 100 transmits a Deliver EI Job result message to the Near-RT RIC 200 via the A1 interface to transfer the EI data for inference (S207). For example, the data transfer unit 133 transfers the EI data in accordance with the instructions of the Create EI Job message. The Deliver EI Job result message is a message specified by the A1 interface and is a message for transferring EI data. The data transfer unit 133 repeatedly transmits the data specified in the Create EI Job message at the timing specified in the Create EI Job message. Multiple pieces of EI data may be transferred in the Deliver EI Job result message in accordance with the Create EI Job message. Furthermore, the data transfer unit 133 specifies the collection method determined by the method determination unit 132 in the Deliver EI Job result message together with the EI data for inference. In other words, it specifies whether the data is to be extracted as learning data or whether the data is to be made lighter when transferred as learning data. For example, in the Deliver EI Job result message, it specifies a flag indicating collection method 1 or collection method 2, or a flag indicating whether the data is to be extracted as learning data. Note that the collection method may be notified in a message separate from the Deliver EI Job result message.

[0098] Furthermore, when the E2 node 300 receives a RIC subscription message from the Near-RT RIC 200, it transmits a RIC Indication message to the Near-RT RIC 200 via the E2 interface in accordance with the specifications of the RIC subscription message, and transfers RAN data for inference (S208). The RIC Indication message is a message specified by the E2 interface and is a message for transferring RAN data. If the data transfer source is specified in the RIC subscription message, the specified E2 node 300 transmits the RIC subscription message. The E2 node 300 repeatedly transmits the data specified in the RIC subscription message at the timing specified in the RIC subscription message. According to the RIC subscription message, multiple RAN data may be transferred in the RIC Indication message.

[0099] Next, the Near-RT RIC 200 acquires EI data from the Non-RT RIC 100 and RAN data from the E2 node 300, and stores the acquired EI data and RAN data as inference data in the inference data storage unit 220 (S209). For example, the data collection unit 231 associates the EI data acquired from the Non-RT RIC 100 with a flag specifying the collection method and stores them in the inference data storage unit 220, and stores the RAN data acquired from the E2 node 300 in the inference data storage unit 220. Note that when storing EI data, only data for which collection method 2 is specified may be stored in the inference data storage unit 220, and when transferring learning data, all data stored in the inference data storage unit 220 may be used as learning data.

[0100] Next, the Near-RT RIC 200 performs inference processing using the acquired EI data and RAN data as inference data (S210). For example, the inference unit 211 inputs the acquired EI data and RAN data into a learning model and infers RAN control according to the EI data and RAN data.

[0101] Next, based on the inference result, the Near-RT RIC 200 transmits a RAN Control message to the E2 node 300 via the E2 interface to set radio control parameters (S211). For example, the inference unit 211 generates radio control parameters for controlling the E2 node 300 based on the inference result of the inference model, and transmits the generated radio control parameters. The RAN Control message is a message specified by the E2 interface and is a message for controlling the E2 node. The inference unit 211 may specify information for identifying the E2 node 300, information for identifying the radio control parameters, values ​​of the radio control parameters, etc. in the RAN Control message. The information for identifying the E2 node 300 may be information for identifying an O-DU or an O-CU. The information for identifying the radio control parameters may indicate the ID or name of the parameter. Multiple radio control parameters may be set in the RAN Control message.

[0102] For example, the Near-RT RIC 200 stores inference data and performs inference processing each time it receives EI data and RAN data. When receiving EI data, the Near-RT RIC 200 may perform inference processing using the received EI data and previously received RAN data. When receiving RAN data, the Near-RT RIC 200 may perform inference processing using the received RAN data and previously received EI data. The unit for performing inference processing is not limited to one EI data and one RAN data. Inference processing may be performed using any number of EI data or more and any number of RAN data or more. When reception of a predetermined number of EI data and reception of a predetermined number of RAN data are completed, inference processing may be performed using the predetermined number of EI data and the predetermined number of RAN data. For example, inference processing may be performed using multiple EI data acquired from multiple external servers 400 and multiple RAN data acquired from multiple E2 nodes 300 including O-DUs and O-CUs.

[0103] Furthermore, upon receiving the RAN Control message, the E2 node 300 sets radio control parameters in accordance with the specifications of the RAN Control message. The E2 node 300 may transmit the setting results of the radio control parameters to the Near-RT RIC 200. If the setting of the radio control parameters fails, the Near-RT RIC 200 may set the same radio control parameters again for the E2 node 300, or may redo the inference process. The Near-RT RIC 200 may store the received setting results together with the inference results in the inference data storage unit 220.

[0104] After S201 and S202, the data collection loop of S203 to S211 is repeated. In the example of Fig. 14, during the data collection loop, the collection method is determined each time EI data is acquired from the external server 400. In other words, the collection method is determined according to changes in the data acquired multiple times. This makes it possible to switch the collection method for each piece of data acquired from the external server 400. For example, of the repeatedly collected data, some data may be stored in the Non-RT RIC 100 using Collection Method 1, and the other data may be collected from the Near-RT RIC 200 using Collection Method 2.

[0105] Fig. 15 is a sequence diagram showing another example of the operation of the inference phase process (S101) of Fig. 13. Fig. 15 shows an example in which the Non-RT RIC 100 determines the collection method before repeatedly acquiring external data from the external server 400. In this way, if it is known in advance that there will be no significant fluctuations in the characteristics of the data transferred from the external server, it is not necessary to determine the collection method each time EI data for inference is transferred to the Near-RT RIC.

[0106] In the example of FIG. 15, the timing of determining the collection method (S205) differs from that of FIG. 14, but the other processing is the same as that of FIG. 14. That is, in the example of FIG. 15, when the Non-RT RIC 100 receives a Create EI Job message from the Near-RT RIC 200 in S202, it determines the collection method in S205. For example, the collection method is determined when a request for EI data is received from the Near-RT RIC 200. For example, if the characteristics of the EI data to be acquired are set in advance, the method determination unit 132 determines the collection method based on the set information. For example, the data size and number of parameters may be associated with each EI data and set, and the collection method may be determined using the data size and number of parameters corresponding to the data specified in the Create EI Job message. Note that the collection method may also be determined based on other information specified in the Create EI Job message. For example, the collection method may be determined based on information identifying the specified data provider, i.e., the external server 400. Identification information of the external server 400 may be associated with the collection method and set, and the collection method may be determined based on the collection method corresponding to the identification information of the data provider specified in the Create EI Job message.

[0107] It is also possible to select whether to determine the collection method at the timing shown in FIG. 15 or at the timing shown in FIG. 14. For example, the Near-RT RIC 200 may specify the timing for determining the collection method in a Create EI Job message. The Non-RT RIC 100 may select the timing for determining the collection method according to the data specified in the Create EI Job message. Data may be classified in advance into data with no fluctuations in data characteristics and data with fluctuations in data characteristics, and when collecting data with no fluctuations in data characteristics, the collection method may be determined at the timing shown in FIG. 15, and when collecting data with fluctuations in data characteristics, the collection method may be determined at the timing shown in FIG. 14.

[0108] In the example of FIG. 15 , after the collection method is determined, a data collection loop is repeatedly executed (S203 to S204, S206 to S211) as in FIG. 14 . In the data collection loop, the Non-RT RIC 100 acquires EI data from the external server 400 and stores the EI data according to the predetermined collection method. Because the collection method does not change within the data collection loop, it is not necessary to notify the collection method each time the Non-RT RIC 100 transfers EI data to the Near-RT RIC 200 in S207. For example, the collection method may be notified in the first Deliver EI Job result message sent after the Create EI Job message, and the notification of the collection method may be omitted in subsequent Deliver EI Job result messages. For example, the collection method may be notified if the collection method has changed since the previous notification.

[0109] The collection method may be determined at other times. The collection method may be determined before S201, i.e., before the inference phase processing. If the EI data to be collected is determined in advance, the collection method may be determined based on the characteristics of the EI data to be collected. Furthermore, the collection method may be determined at any time in response to an instruction from an operator or the load on the RAN system 1.

[0110] Fig. 16 is a sequence diagram showing an example of the operation of the learning phase process (S103) in Fig. 13. After the inference phase process in Fig. 14 or Fig. 15 has collected and stored inference data, the process in Fig. 16 is executed.

[0111] As shown in FIG. 16 , when the Non-RT RIC 100 starts collecting learning data (S301), it requests the transfer of the learning data via the O1 interface (S302). For example, the learning unit 111 requests the transfer of learning data when it determines that collection of learning data is necessary. When the Near-RT RIC 200 stores a predetermined amount of inference data, the Near-RT RIC 200 notifies the Non-RT RIC 100 that storage is complete. Upon receiving the storage completion notification, the Non-RT RIC may determine to start collecting learning data and send a learning data transfer request. The learning unit 111 may send the learning data transfer request using any message specified in the O1 interface. The transfer timing, etc. may be specified in the learning data transfer request. Learning data may also be requested and transferred between the Non-RT RIC 100 and the Near-RT RIC 200 via the A1 interface.

[0112] Next, upon receiving a request to transfer training data from the Non-RT RIC 100, the Near-RT RIC 200 extracts training data from the inference data storage unit 220 according to the collection method (S303). The data extraction unit 232 determines the collection method set for each piece of inference data stored in the inference data storage unit 220 and extracts the training data according to the determined collection method. The data extraction unit 232 does not extract inference data for collection method 1, but extracts inference data for collection method 2 to generate training data to be transferred. For example, EI data designated for collection method 1 is excluded, and the remaining EI data and RAN data for collection method 2 are extracted as training data. This generates lightweight training data excluding the EI data held in the Non-RT RIC 100.

[0113] Next, the Near-RT RIC 200 transfers the extracted training data to the Non-RT RIC 100 via the O1 interface (S304). For example, when training data is extracted from the inference data in the inference data storage unit 220, the data transfer unit 233 transfers the training data that has been reduced in size according to the collection method. The data transfer unit 233 may transfer the training data using any message specified by the O1 interface. If a transfer timing is specified in the training data transfer request, the training data may be transmitted at the specified timing. The training data may be transmitted when the transmission bandwidth of the O1 interface is available.

[0114] Next, when the Non-RT RIC 100 receives the training data from the Near-RT RIC 200, it combines the EI data stored in the EI data storage unit 120 with the received training data according to the collection method (S305). For example, the training unit 111 may determine whether EI data is stored in the EI data storage unit 120 to determine the collection method. If EI data is stored in the EI data storage unit 120, the training unit 111 determines that the training data is of collection method 1 or that data corresponding to collection method 1 is included, and combines the EI data stored in the EI data storage unit 120 with the training data received via the O1 interface to form data necessary for input into the training model. If no EI data is stored in the EI data storage unit 120, the training data is determined to be of collection method 2 or that there is no data corresponding to collection method 1, and the training data is not combined.

[0115] Furthermore, the learning unit 111 may determine the collection method set for each EI data. The collection method for EI data may be retained when the method determination unit 132 determines the collection method. If there is EI data set to collection method 1, the learning unit 111 acquires the corresponding EI data from the EI data storage unit 120 and combines the acquired EI data with the learning data received via the O1 interface. If there is no EI data set to collection method 1, the learning unit 111 does not combine the learning data.

[0116] Next, the Non-RT RIC 100 performs a learning process using the learning data synthesized according to the collection method (S306). For example, when EI data is stored in the EI data storage unit 120, i.e., in the case of collection method 1, the learning unit 111 trains the learning model using synthesized data obtained by synthesizing the EI data stored in the EI data storage unit 120 and the learning data received from the Near-RT RIC 200. When EI data is not stored in the EI data storage unit 120, i.e., in the case of collection method 2, the learning unit 111 trains the learning model using the learning data received from the Near-RT RIC 200. When learning is completed, the learning unit 111 stores the trained learning model in the model storage unit 112 and transmits the trained learning model to the Near-RT RIC 200. The Near-RT RIC 200 applies the received trained learning model to the inference model and performs inference processing using the updated inference model.

[0117] As described above, in this embodiment, the Non-RT RIC collects EI data for inference from an external server, and selects a collection method for collecting training data depending on the characteristics of the collected EI data. By selecting collection method 1, the Non-RT RIC stores the data collected for inference, and transfers lightweight training data from the Near-RT RIC to the Non-RT RIC, the network load on the O1 interface that transfers the training data can be reduced. Furthermore, by selecting collection method 2 and transferring all data necessary for training from the Near-RT RIC to the Non-RT RIC, the load on the training data formatting process in the Non-RT RIC can be reduced. Therefore, depending on the characteristics of the data collected for inference and used for training, the load on the O1 interface or the processing load on the Non-RT RIC can be reduced, thereby streamlining the training data collection process and enabling efficient generation of training data.

[0118] (Embodiment 2) Next, embodiment 2 will be described. In this embodiment, an example will be described in which learning data is collected by fixing the collection method to collection method 1. Note that this embodiment can be implemented in combination with embodiment 1, and may be implemented by using the configuration of embodiment 1 as appropriate. For example, the configuration of this embodiment is the same as embodiment 1, and therefore a description thereof will be omitted. In this embodiment, the method determination unit 132 in the Non-RT RIC 100 of FIG. 10 may be omitted.

[0119] FIG. 17 shows an example of the operation of the inference phase processing (S101) in this embodiment. In this embodiment, since it is sufficient to be able to implement at least collection method 1, unlike embodiment 1, the Non-RT RIC 100 does not determine the collection method (S205). Furthermore, in S206, the EI data acquired from the external server 400 is stored without determining the collection method. Furthermore, in S207, when transferring EI data from the Non-RT RIC 100 to the Near-RT RIC 200, there is no need to notify the collection method. The rest is the same as embodiment 1. Note that when storing inference data in S209, the EI data collected from the Non-RT RIC 100 may not be stored in the inference data storage unit 220, and instead, when transferring learning data, the RAN data stored in the inference data storage unit 220 may be transferred as learning data.

[0120] FIG. 18 shows an example of the operation of the learning phase process (S103) in this embodiment. In this embodiment, it is sufficient to be able to implement at least collection method 1. Therefore, compared to embodiment 1, the Near-RT RIC 200 extracts learning data in S303 without determining the collection method. That is, from the data stored in the inference data storage unit 220, excluding the EI data collected via the Non-RT RIC 100, extracts only the RAN data collected from the E2 node 300 to generate learning data. Furthermore, in S305, the Non-RT RIC 100 combines the learning data without determining the collection method. That is, the EI data stored in the EI data storage unit 120 is combined with the learning data collected via the O1 interface. The rest is the same as embodiment 1.

[0121] In this way, by using collection method 1, the Non-RT RIC stores the data collected for inference and transfers the reduced-weight learning data from the Near-RT RIC to the Non-RT RIC, the network load on the O1 interface that transfers the learning data can be reduced.

[0122] (Embodiment 3) Next, embodiment 3 will be described. In this embodiment, an example will be described in which learning data is collected by fixing the collection method to collection method 2. This embodiment can be implemented in combination with either embodiment 1 or 2, and may be implemented by using the configuration of either embodiment 1 or 2 as appropriate. For example, the configuration of this embodiment is the same as embodiment 1, so its description will be omitted. In this embodiment, the method determination unit 132 and the EI data storage unit 120 in the Non-RT RIC 100 of FIG. 10 may be omitted.

[0123] 19 shows an example of the operation of the inference phase processing (S101) in this embodiment. In this embodiment, it is sufficient to be able to implement at least collection method 2, so unlike embodiment 1, the Non-RT RIC 100 does not determine the collection method (S205) or store the EI data (S206). That is, when the Non-RT RIC 100 acquires EI data from the external server 400 in S204, it transfers the EI data for inference to the Near-RT RIC 200 in S207. Furthermore, when transferring EI data from the Non-RT RIC 100 to the Near-RT RIC 200 in S207, there is no need to notify the collection method. The rest is the same as embodiment 1.

[0124] FIG. 20 shows an example of the operation of the learning phase process (S103) in this embodiment. In this embodiment, it is sufficient to be able to implement at least collection method 2, so compared to embodiment 1, the Near-RT RIC 200 extracts learning data in S303 without determining the collection method. That is, all data stored in the inference data storage unit 220, including EI data and RAN data, is extracted to generate learning data. Furthermore, the Non-RT RIC 100 does not synthesize the learning data (S305). That is, the Non-RT RIC 100 performs learning processing using the learning data received from the Near-RT RIC 200 in S306. The rest is the same as in embodiment 1.

[0125] In this way, by using collection method 2 to transfer all data necessary for learning from the Near-RT RIC to the Non-RT RIC, the load of the learning data formatting process on the Non-RT RIC can be reduced.

[0126] (Fourth Embodiment) Next, a fourth embodiment will be described. In this embodiment, an example in which learning data is collected by collection method 3 will be described. This embodiment can be implemented in combination with any of the first to third embodiments, and may be implemented by using any of the configurations of the first to third embodiments as appropriate. For example, the configuration of this embodiment is the same as that of the first embodiment, and therefore description thereof will be omitted.

[0127] 21 shows a third collection method of learning data according to this embodiment. As shown in FIG. 21, collection method 3 is a method in which learning data is transferred from the E2 node 300 to the Non-RT RIC 100. This reduces the load on the Near-RT RIC 200.

[0128] In collection method 3, during inference, the Non-RT RIC 100 acquires EI data for inference from the external server 400, stores the acquired EI data in the EI data storage unit 120, and transfers the EI data to the Near-RT RIC 200. The Near-RT RIC 200 performs inference using the inference unit 210 with the EI data acquired from the Non-RT RIC 10 and the RAN data collected from the E2 node 300 as inference data.

[0129] Furthermore, in collection method 3, during learning, the Non-RT RIC 100 collects RAN data from the E2 node 300 via the O1 interface as learning data. The Non-RT RIC 100 combines the RAN data collected from the E2 node 300 with the EI data stored in the EI data storage unit 120, and performs learning using the combined learning data. Note that, as with collection method 2, during learning, the EI data used for inference may be transferred from the Near-RT RIC 200 to the Non-RT RIC 100.

[0130] In this way, the training data may be transferred from the E2 node to the Non-RT RIC using collection method 3. Any of collection methods 1 to 3 may be selected to collect the training data. As in the first embodiment, the method determination unit 132 may select any of collection methods 1 to 3 depending on the characteristics of the data, instructions from the operator, and the load on the RAN system. Since the collection path for the training data changes in collection method 3, it can also be said that the method determination unit 132 selects the collection path for the training data. For example, collection method 3 may be selected depending on the characteristics of the RAN data collected from the E2 node. Furthermore, collection method 3 may be selected when the load on the Near-RT RIC is heavy or when the E2 node has spare resources. This allows training data to be collected using an appropriate method depending on various situations.

[0131] (Fifth Embodiment) Next, a fifth embodiment will be described. In this embodiment, an example in which an external server and a Near-RT RIC are directly connected will be described. Note that this embodiment can be implemented in combination with any of the first to fourth embodiments, and may also be implemented by appropriately using any of the configurations of the first to fourth embodiments.

[0132] Fig. 22 shows a configuration example of a RAN system 1 according to the present embodiment. As shown in Fig. 22, in the RAN system 1 according to the present embodiment, compared to the first embodiment, the external server 400 and the Near-RT RIC 200 are directly connected. The Near-RT RIC 200 may include an external communication unit, similar to the Non-RT RIC 100. The other configurations are the same as those of the first embodiment, for example.

[0133] The Near-RT RIC 200 and the external server 400 are communicatively connected via any interface, similar to the connection between the Non-RT RIC 100 and the external server 400. During inference, the Near-RT RIC 200 directly collects EI data for inference from the external server 400 via the interface with the external server 400. The Near-RT RIC 200 performs inference using the EI data collected from the external server 400 and the RAN data collected from the E2 node 300. Note that in this example, it is not necessary to transfer the EI data for inference from the Non-RT RIC 100 to the Near-RT RIC 200. For example, the Non-RT RIC 100 may acquire the EI data for inference from the external server 400 at any timing.

[0134] In this way, the Near-RT RIC may directly acquire inference data from an external server. This configuration allows the Near-RT RIC to control the RAN in more real time. For example, this improves the ability to respond to sudden changes in the wireless environment.

[0135] The present disclosure is not limited to the above-described embodiment, and can be modified as appropriate within the scope of the present disclosure.

[0136] Each component in the above-described embodiments may be configured with hardware, software, or both, and may be configured with a single piece of hardware or software, or may be configured with multiple pieces of hardware or software. Each device, including the Non-RT RIC and Near-RT RIC, and each function (processing) may be implemented by a computer 50 having a network interface 51, a processor 52 such as a CPU (Central Processing Unit), and a memory 53 serving as a storage device, as shown in FIG. 23 . The network interface 51 may include a network interface card (NIC) for communicating with devices including network nodes. For example, a program for performing the method in the embodiment may be stored in the memory 53, and each function may be implemented by executing the program stored in the memory 53 by the processor 52.

[0137] These programs include instructions (or software code) that, when loaded into a computer, cause the computer to perform one or more functions described in the embodiments. The programs may be stored on a non-transitory computer-readable medium or a tangible storage medium. By way of example and not limitation, computer-readable media or tangible storage media include random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technology, CD-ROM, digital versatile disc (DVD), Blu-ray disc or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device. The programs may also be transmitted on a transitory computer-readable medium or communication medium. By way of example and not limitation, transitory computer-readable media or communication media include electrical, optical, acoustic, or other forms of propagated signals.

[0138] Although the present disclosure has been described above with reference to the embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure.

[0139] Some or all of the above embodiments can be described as, but are not limited to, the following supplementary notes. (Supplementary Note 1) A system comprising: an acquisition means for acquiring data provided from a data providing device as inference data for another system to perform inference using an inference model; and an identification means for identifying, from data including the acquired inference data, data to be collected from another system that has performed inference using the inference model as learning data for a learning model that constructs the inference model. (Supplementary Note 2) The system according to Supplementary Note 1, comprising: a transfer means for transferring the inference data acquired from the data providing device to the other system, wherein the identification means identifies whether or not the inference data transferred to the other system should be collected from the other system. (Supplementary Note 3) The system according to Supplementary Note 2, comprising: a storage means for storing the inference data transferred to the other system, wherein the identification means identifies whether or not the stored inference data should be collected from the other system. (Supplementary Note 4) The system according to Supplementary Note 3, further comprising a synthesis means for synthesizing the inference data stored in the storage means with data collected from the other system when the inference data is not collected from the other system, and generating learning data to be input to the learning model. (Supplementary Note 5) The system according to any one of Supplements 1 to 4, wherein the identification means identifies a path for collecting the learning data. (Supplementary Note 6) The system according to any one of Supplements 1 to 5, wherein the identification means identifies data to collect from the other system based on characteristics of the inference data acquired from the data providing device. (Supplementary Note 7) The system according to Supplementary Note 6, wherein the characteristics of the inference data include data size, number of parameters, or data acquisition period. (Supplementary Note 8) The system according to Supplementary Note 6 or 7, wherein the acquisition means acquires inference data from the data providing device multiple times, and identifies data to collect from the other system according to changes in the inference data acquired multiple times. (Supplementary Note 9) The system according to any one of Supplementary Notes 1 to 8, wherein the specifying unit specifies the data to be collected from the other system based on an input instruction.(Supplementary Note 10) The system according to any one of Supplements 1 to 9, wherein the specifying means specifies data to be collected from the other system based on a load of a system including the system and the other system. (Supplementary Note 11) The system according to any one of Supplements 1 to 10, wherein the data providing device is a server external to the system including the system and the other system. (Supplementary Note 12) The system according to any one of Supplements 1 to 11, wherein the inference model infers control related to a wireless network according to the inference data, and the learning model learns control related to the wireless network according to the learning data. (Supplementary Note 13) The system according to any one of Supplements 1 to 12, wherein the system and the other system include a RAN Intelligent Controller (RIC) that controls a RAN (Radio Access Network). (Supplementary Note 14) The system according to Supplementary Note 13, wherein the system includes a Non-RT (Real Time) RIC, and the other system includes a Near-RT RIC. (Supplementary Note 15) A system comprising: a collection means for collecting data provided from a data providing device as inference data for making inferences using an inference model; and a transmission means for transmitting data identified from data including the collected inference data to another system that performs learning using the learning model as learning data for a learning model that constructs the inference model. (Supplementary Note 16) The system according to Supplementary Note 15, wherein the collection means collects the inference data via the other system, and the identified data is data identified by the other system. (Supplementary Note 17) A system comprising: an acquisition means for acquiring data provided from a data providing device as inference data for another system that has performed inference using the inference model; and an identification means for identifying data to be collected from another system that has performed inference using the inference model as learning data for a learning model that constructs the inference model from data including the acquired inference data.(Supplementary Note 18) A device comprising: a collection means for collecting data provided from a data providing device as inference data for making inferences using an inference model; and a transmission means for transmitting data identified from data including the collected inference data as learning data for a learning model that constructs the inference model to another system that learns using the learning model. (Supplementary Note 19) A method for acquiring data provided from a data providing device as inference data for another system to make inferences using an inference model, and identifying data to be collected from another system that has made inferences using the inference model as learning data for the learning model that constructs the inference model from the acquired data including the inference data. (Supplementary Note 20) A method for collecting data provided from a data providing device as inference data for making inferences using an inference model, and transmitting data identified from data including the collected inference data as learning data for the learning model that constructs the inference model to another system that learns using the learning model. (Supplementary Note 21) A non-transitory computer-readable medium storing a program for causing a computer to execute the following processes: acquiring data provided from a data providing device as inference data for another system to make inference using an inference model, and identifying data to be collected from another system that has made inference using the inference model as learning data for a learning model that constructs the inference model from data including the acquired inference data. (Supplementary Note 22) A non-transitory computer-readable medium storing a program for causing a computer to execute the following processes: collecting data provided from a data providing device as inference data for making inference using an inference model, and transmitting data identified from data including the collected inference data to another system that learns using the learning model as learning data for a learning model that constructs the inference model.

[0140] 1 RAN system 10 First system 11 Acquisition unit 12 Identification unit 20 Second system 21 Collection unit 22 Transmission unit 30 First device 40 Second device 50 Computer 51 Network interface 52 Processor 53 Memory 100 Non-RT RIC 101 O1 communication unit 102 A1 communication unit 103 External communication unit 110 Learning device 111 Learning unit 112 Model storage unit 120 EI data storage unit 131 Data collection unit 132 Method determination unit 133 Data transfer unit 134 System management unit 200 Near-RT RIC 201 E2 communication unit 202 O1 communication unit 203 A1 communication unit 210 Inference unit 211 Inference unit 212 Model storage unit 220 Inference data storage unit 231 Data collection unit 232 Data extraction unit 233 Data transfer unit 300 E2 node 400 External server 500 SMO

Claims

1. An acquisition means for acquiring data provided from a data providing device as inference data for another system to perform inference using an inference model; A means for identifying data collected from another system that has performed inference using the inference model as learning data for a learning model that constructs the inference model, from among data including the acquired inference data; A system comprising:

2. a transfer means for transferring the inference data acquired from the data providing device to the other system, The determination means determines whether or not the inference data transferred to the other system is to be collected from the other system. The system of claim 1 .

3. A storage means for storing the inference data transferred to the other system, The determination means determines whether or not the stored inference data is to be collected from the other system. The system of claim 2.

4. a synthesis means for synthesizing the inference data stored in the storage means and data collected from the other system when the inference data is not collected from the other system, and generating learning data to be input to the learning model; The system of claim 3.

5. The identification means identifies a route for collecting the learning data. A system according to any one of claims 1 to 4.

6. A collection means for collecting data provided from a data providing device as inference data for performing inference using an inference model; A transmission means for transmitting data identified from the collected data including the inference data as learning data for a learning model that constructs the inference model to another system that performs learning using the learning model; A system comprising:

7. The data provided by the data providing device is acquired as inference data for another system to perform inference using an inference model; Identifying data to be collected from other systems that have performed inference using the inference model as learning data for a learning model that constructs the inference model from among the data including the acquired inference data; method.

8. Collecting data provided from a data providing device as inference data for performing inference using an inference model; Transmitting data identified from the collected data including the inference data as learning data for a learning model that constructs the inference model to another system that performs learning using the learning model; method.

9. The data provided by the data providing device is acquired as inference data for another system to perform inference using an inference model; Identifying data to be collected from other systems that have performed inference using the inference model as learning data for a learning model that constructs the inference model from among the data including the acquired inference data; A program that causes a computer to execute a process.

10. Collecting data provided from a data providing device as inference data for performing inference using an inference model; Transmitting data identified from the collected data including the inference data as learning data for a learning model that constructs the inference model to another system that performs learning using the learning model; A program that causes a computer to execute a process.