WO2025170021A1 - Communication control method - Google Patents

Abstract

One aspect of the present invention relates to a communication control method for user equipment in a mobile communication system. The communication control method includes a step for performing fine adjustment on a trained artificial intelligence (AI)/machine learning (ML) model on the basis of a fine adjustment execution time indicating the time it takes for the user equipment to execute the fine adjustment of the trained AI/ML model. The fine adjustment refers to training the trained AI/ML model using training data having a data amount equal to or less than a data amount threshold.

Description

Communication Control Method

This disclosure relates to a communication control method.

In recent years, the Third Generation Partnership Project (3GPP) (registered trademark; the same applies hereinafter), a standardization project for mobile communications systems, has been considering applying artificial intelligence (AI) technology, particularly machine learning (ML) technology, to wireless communications (air interfaces) in mobile communications systems.

3GPP TR 38.843 V18.0.0 (2023-12)

A communication control method according to a first aspect is a communication control method for a user device in a mobile communication system. The communication control method includes a step in which the user device performs fine-tuning on the trained AI/ML model based on a fine-tuning execution time indicating the time required to perform fine-tuning on the trained AI/ML model. The fine-tuning involves training the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

A communication control method according to a second aspect is a communication control method in a network node of a mobile communication system. The communication control method includes a step in which the network node determines fine-tuning execution conditions indicating the conditions for performing fine-tuning on the trained AI/ML model in the user device, based on a fine-tuning execution time indicating the time required to perform fine-tuning on the trained AI/ML model. The communication control method also includes a step in which the network node transmits the fine-tuning execution conditions to the user device. The fine-tuning involves training the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

A communication control method according to a third aspect is a communication control method in a network node of a mobile communication system. The communication control method includes a step in which the network node decides to perform fine-tuning of the trained AI/ML model in the user device based on a fine-tuning execution time indicating the time required to perform fine-tuning of the trained AI/ML model. The communication control method also includes a step in which the network node sends a fine-tuning execution instruction to the user device, instructing the user device to perform fine-tuning. The fine-tuning involves training the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

A communication control method according to a fourth aspect is a communication control method for a user device in a mobile communication system. The communication control method includes a step in which the user device receives, from a network node, fine-tuning execution conditions indicating conditions for performing fine-tuning of a trained AI/ML model. The communication control method also includes a step in which the user device performs inference using the trained AI/ML model. The communication control method further includes a step in which the user device performs fine-tuning. Here, the fine-tuning execution conditions include a performance threshold and a predetermined number of times. The communication control method also includes a step in which the user device performs fine-tuning in response to detecting, the predetermined number of times, that the inference result of the trained AI/ML model is equal to or less than the performance threshold. The fine-tuning is training the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

A communication control method according to a fifth aspect is a communication control method in a network node of a mobile communication system. The communication control method includes a step in which the network node transmits fine-tuning execution conditions indicating conditions for performing fine-tuning of a trained AI/ML model to a user device. The fine-tuning execution conditions include a performance threshold and a predetermined number of times. The fine-tuning is performed in response to the user device detecting that the inference result of the trained AI/ML model is below the performance threshold a predetermined number of times. The fine-tuning is performed by training the trained AI/ML model using training data with a data volume below a data volume threshold.

FIG. 1 is a diagram showing an example of the configuration of a mobile communication system according to the first embodiment. FIG. 2 is a diagram illustrating an example of the configuration of a UE (user equipment) according to the first embodiment. Figure 3 is a diagram showing an example configuration of a gNB (base station) according to the first embodiment. FIG. 4 is a diagram illustrating an example of the configuration of a protocol stack according to the first embodiment. FIG. 5 is a diagram illustrating an example of the configuration of a protocol stack according to the first embodiment. FIG. 6 is a diagram illustrating an example of the configuration of functional blocks of the AI/ML technology according to the first embodiment. FIG. 7 is a diagram illustrating an example of operation in the AI/ML technique according to the first embodiment. FIG. 8 is a diagram illustrating an example of an arrangement of functional blocks of the AI/ML technology according to the first embodiment. FIG. 9 is a diagram illustrating an example of operation according to the first embodiment. FIG. 10 is a diagram illustrating an example of operation according to the first embodiment. FIG. 11 is a diagram illustrating an example of a setting message according to the first embodiment. FIG. 12 is a diagram illustrating an example of operation according to the first embodiment. FIG. 13 is a diagram illustrating an example of operation according to the first embodiment. FIG. 14 is a diagram illustrating an example of operation according to the first embodiment. FIG. 15 is a diagram illustrating an example of operation according to the first embodiment. FIG. 16 is a diagram illustrating an example of operation according to the second embodiment. FIG. 17 is a diagram illustrating an example of operation according to the second embodiment. FIG. 18 is a diagram illustrating an example of operation according to the second embodiment. FIG. 19 is a diagram illustrating an example of operation according to the second embodiment. FIG. 20 is a diagram illustrating an example of operation according to the second embodiment. FIG. 21 is a diagram illustrating an example of operation according to the third embodiment. FIG. 22 is a diagram illustrating an example of operation according to the third embodiment. FIG. 23 is a diagram illustrating an example of operation according to the fourth embodiment. FIG. 24 is a diagram illustrating an example of operation according to the fifth embodiment. FIG. 25 is a diagram illustrating an example of operation according to the fifth embodiment. FIG. 26 is a diagram illustrating an example of operation according to the fifth embodiment.

The purpose of this disclosure is to appropriately perform fine-tuning on trained AI/ML models.

[First embodiment]
The mobile communication system according to the first embodiment will be described with reference to the drawings. In the description of the drawings, the same or similar parts are denoted by the same or similar reference numerals.

(Configuration of mobile communication system)
The configuration of a mobile communication system according to the first embodiment will be described. FIG. 1 is a diagram showing an example of the configuration of a mobile communication system 1 according to the first embodiment. The mobile communication system 1 conforms to the 5th Generation System (5GS) of the 3GPP standard. Although the following description will be given using 5GS as an example, the mobile communication system may also be at least partially applied to an LTE (Long Term Evolution) system. The mobile communication system may also be at least partially applied to a sixth generation (6G) system or later system.

The mobile communication system 1 includes a user equipment (UE) 100, a 5G radio access network (NG-RAN) 10, and a 5G core network (5GC) 20. Hereinafter, the NG-RAN 10 may be simply referred to as the RAN 10. The 5GC 20 may also be simply referred to as the core network (CN) 20. Furthermore, the devices included in the core network CN may also be referred to as core network devices.

UE100 is a mobile wireless communication device. UE100 may be any device that is used by a user. For example, UE100 is a mobile phone terminal (including a smartphone) and/or a tablet terminal, a notebook PC, a communication module (including a communication card or chipset), a sensor or a device provided in a sensor, a vehicle or a device provided in a vehicle (Vehicle UE), or an aircraft or a device provided in an aircraft (Aerial UE).

The NG-RAN 10 includes a base station (called a "gNB" in a 5G system) 200. The gNBs 200 are connected to each other via an Xn interface, which is an interface between base stations. The gNB 200 manages one or more cells. The gNB 200 performs wireless communication with a UE 100 that has established a connection with its own cell. The gNB 200 has a radio resource management (RRM) function, a routing function for user data (hereinafter simply referred to as "data"), and a measurement control function for mobility control and scheduling. The term "cell" is used to indicate the smallest unit of a wireless communication area. The term "cell" is also used to indicate a function or resource for wireless communication with a UE 100. One cell belongs to one carrier frequency (hereinafter simply referred to as "frequency").

In addition, gNBs can also connect to the EPC (Evolved Packet Core), which is the LTE core network. LTE base stations can also connect to 5GC. LTE base stations and gNBs can also be connected via a base station-to-base station interface.

5GC20 includes an AMF (Access and Mobility Management Function) and a UPF (User Plane Function) 300. The AMF performs various mobility controls for UE100. The AMF manages the mobility of UE100 by communicating with UE100 using NAS (Non-Access Stratum) signaling. The UPF controls data forwarding. The AMF and UPF 300 are connected to gNB200 via an NG interface, which is an interface between a base station and a core network. The AMF and UPF 300 may be core network devices included in CN20. The core network device and gNB200 may collectively be referred to as a network device.

FIG. 2 is a diagram showing an example configuration of UE100 (user equipment) according to the first embodiment. UE100 includes a receiver 110, a transmitter 120, and a controller 130. The receiver 110 and transmitter 120 constitute a communication unit that performs wireless communication with gNB200. UE100 is an example of a communication device.

The receiving unit 110 performs various types of reception under the control of the control unit 130. The receiving unit 110 includes an antenna and a receiver. The receiver converts the radio signal received by the antenna into a baseband signal (received signal) and outputs it to the control unit 130.

The transmitting unit 120 performs various transmissions under the control of the control unit 130. The transmitting unit 120 includes an antenna and a transmitter. The transmitter converts the baseband signal (transmission signal) output by the control unit 130 into a radio signal and transmits it from the antenna.

The control unit 130 performs various types of control and processing in the UE 100. Such processing includes processing for each layer, which will be described later. The control unit 130 includes at least one processor and at least one memory. The memory stores programs executed by the processor and information used for processing by the processor. The processor may include a baseband processor and a CPU (Central Processing Unit). The baseband processor performs modulation/demodulation, encoding/decoding, etc. of baseband signals. The CPU executes programs stored in the memory to perform various types of processing. Note that the processing or operations performed by the UE 100 may be performed in the control unit 130.

Figure 3 is a diagram showing an example configuration of a gNB200 (base station) according to the first embodiment. The gNB200 comprises a transmitter 210, a receiver 220, a controller 230, and a backhaul communication unit 250. The transmitter 210 and receiver 220 constitute a communication unit that performs wireless communication with the UE100. The backhaul communication unit 250 constitutes a network communication unit that performs communication with the CN20. The gNB200 is another example of a communication device. Alternatively, the gNB200 may be an example of a network node.

The transmitting unit 210 performs various transmissions under the control of the control unit 230. The transmitting unit 210 includes an antenna and a transmitter. The transmitter converts the baseband signal (transmission signal) output by the control unit 230 into a radio signal and transmits it from the antenna.

The receiving unit 220 performs various types of reception under the control of the control unit 230. The receiving unit 220 includes an antenna and a receiver. The receiver converts the radio signal received by the antenna into a baseband signal (received signal) and outputs it to the control unit 230.

The control unit 230 performs various types of control and processing in the gNB 200. Such processing includes processing for each layer, which will be described later. The control unit 230 includes at least one processor and at least one memory. The memory stores programs executed by the processor and information used in processing by the processor. The processor may include a baseband processor and a CPU. The baseband processor performs modulation/demodulation, encoding/decoding, etc. of baseband signals. The CPU executes programs stored in the memory to perform various types of processing. Note that processing or operations performed in the gNB 200 may be performed by the control unit 230.

The backhaul communication unit 250 is connected to adjacent base stations via an Xn interface, which is an interface between base stations. The backhaul communication unit 250 is connected to the AMF/UPF 300 via an NG interface, which is an interface between a base station and a core network. Note that the gNB 200 may be composed of a central unit (CU) and a distributed unit (DU) (i.e., functionally divided), with the two units connected via an F1 interface, which is a fronthaul interface.

Figure 4 shows an example of the protocol stack configuration for the wireless interface of the user plane that handles data.

The user plane radio interface protocol includes a physical (PHY) layer, a medium access control (MAC) layer, a radio link control (RLC) layer, a packet data convergence protocol (PDCP) layer, and a service data adaptation protocol (SDAP) layer.

The PHY layer performs encoding/decoding, modulation/demodulation, antenna mapping/demapping, and resource mapping/demapping. Data and control information are transmitted between the PHY layer of UE100 and the PHY layer of gNB200 via a physical channel. The PHY layer of UE100 receives downlink control information (DCI) transmitted from gNB200 on the physical downlink control channel (PDCCH). Specifically, UE100 performs blind decoding of the PDCCH using a radio network temporary identifier (RNTI) and acquires successfully decoded DCI as DCI addressed to the UE. The DCI transmitted from gNB200 has CRC (Cyclic Redundancy Code) parity bits scrambled by the RNTI added.

In NR, UE100 can use a bandwidth narrower than the system bandwidth (i.e., the cell bandwidth). gNB200 configures UE100 with a bandwidth portion (BWP) consisting of contiguous PRBs (Physical Resource Blocks). UE100 transmits and receives data and control signals in the active BWP. UE100 may be configured with, for example, up to four BWPs. Each BWP may have a different subcarrier spacing. The BWPs may overlap in frequency. When multiple BWPs are configured for UE100, gNB200 can specify which BWP to apply by controlling the downlink. This allows gNB200 to dynamically adjust the UE bandwidth according to the amount of data traffic of UE100, etc., thereby reducing UE power consumption.

For example, gNB200 can configure up to three control resource sets (CORESETs) for each of up to four BWPs on the serving cell. A CORESET is a radio resource for control information to be received by UE100. Up to 12 or more CORESETs may be configured on the serving cell for UE100. Each CORESET may have an index of 0 to 11 or more. A CORESET may consist of six resource blocks (PRBs) and one, two, or three consecutive Orthogonal Frequency Division Multiplex (OFDM) symbols in the time domain.

The MAC layer performs data priority control, retransmission processing using Hybrid Automatic Repeat reQuest (HARQ), random access procedures, etc. Data and control information are transmitted between the MAC layer of UE100 and the MAC layer of gNB200 via a transport channel. The MAC layer of gNB200 includes a scheduler. The scheduler determines the uplink and downlink transport format (transport block size, modulation and coding scheme (MCS)) and the resource blocks to be allocated to UE100.

The RLC layer uses the functions of the MAC layer and PHY layer to transmit data to the RLC layer on the receiving side. Data and control information are transmitted between the RLC layer of UE100 and the RLC layer of gNB200 via logical channels.

The PDCP layer performs header compression/decompression, encryption/decryption, etc.

The SDAP layer maps IP flows, which are the units by which the core network controls QoS (Quality of Service), to radio bearers, which are the units by which the access stratum (AS) controls QoS. Note that if the RAN is connected to the EPC, SDAP is not necessary.

Figure 5 shows the protocol stack configuration of the radio interface of the control plane, which handles signaling (control signals).

The protocol stack for the radio interface of the control plane includes a Radio Resource Control (RRC) layer and a Non-Access Stratum (NAS) instead of the SDAP layer shown in Figure 4.

RRC signaling for various settings is transmitted between the RRC layer of UE100 and the RRC layer of gNB200. The RRC layer controls logical channels, transport channels, and physical channels in accordance with the establishment, re-establishment, and release of radio bearers. When there is a connection (RRC connection) between the RRC of UE100 and the RRC of gNB200, UE100 is in the RRC connected state. When there is no connection (RRC connection) between the RRC of UE100 and the RRC of gNB200, UE100 is in the RRC idle state. When the connection between the RRC of UE100 and the RRC of gNB200 is suspended, UE100 is in the RRC inactive state.

The NAS, which is located above the RRC layer, performs session management, mobility management, etc. NAS signaling is transmitted between the NAS of UE100 and the NAS of AMF300. In addition to the radio interface protocol, UE100 also has an application layer, etc. The layer below the NAS is called the AS (Access Stratum).

(AI/ML technology)
Next, the AI/ML technology according to the embodiment will be described. Fig. 6 is a diagram showing an example of the configuration of functional blocks of the AI/ML technology in the mobile communication system 1 according to the first embodiment.

The functional block configuration example shown in Figure 6 includes a data collection unit (Data Collection) A1, a model training unit (Model Training) A2, a model inference unit (Inference) A3, a management unit (Management) A5, and a model storage unit (Model Storage) A6.

The example functional block configuration shown in Figure 6 represents the functional framework of general AI/ML technology. Therefore, depending on the hypothetical use case, some of the example functional block configuration (such as the model recording unit A6) may not be included in the example functional block configuration. Furthermore, the example functional block configuration shown in Figure 6 may be distributed between the UE 100 and the network-side device. Alternatively, some functions of the example functional block configuration (such as the model learning unit A2 or model inference unit A3) may be located in both the UE 100 and the network-side device.

The data collection unit A1 provides input data to the model learning unit A2, the model inference unit A3, and the management unit A5. The input data includes training data for the model learning unit A2, inference data for the model inference unit A3, and monitoring data for the management unit A5.

Learning data is data that is required as input when an AI/ML model is learning. Inference data is data that is required as input when an AI/ML model is making inferences. Monitoring data is data that is required as input when managing an AI/ML model.

Note that data collection may refer to the process of collecting data in a network node, management entity, or UE 100, for example, to train an AI/ML model, manage an AI/ML model, and perform inference on an AI/ML model.

The model training unit A2 performs AI/ML model training, AI/ML model validation, and AI/ML model testing.

AI/ML model training is the process of learning an AI/ML model from input/output relationships to obtain a trained AI/ML model that can be used for inference. For example,
y = ax + b
In this sense, the process of optimizing a (slope) and b (intercept) by providing input (x) and output (y) (i.e., providing learning data) may be AI/ML model learning.

Generally, machine learning is divided into supervised learning, unsupervised learning, and reinforcement learning. Supervised learning is a method that uses correct answer data for training data. Unsupervised learning is a method that does not use correct answer data for training data. For example, unsupervised learning memorizes feature points from large amounts of training data and determines the correct answer (estimates the range). Reinforcement learning is a method that assigns a score to the output result and learns how to maximize the score. Supervised learning is explained below, but either unsupervised learning or reinforcement learning can also be applied to machine learning.

The model learning unit A2 outputs the trained AI/ML model (Trained Model) obtained by AI/ML model learning to the model recording unit A6. The model learning unit A2 also outputs the updated AI/ML model (Updated Model) obtained by relearning the trained AI/ML model to the model recording unit A6.

In the following, AI/ML model learning may be referred to as "model learning" or "learning."

The model inference unit A3 performs AI/ML model inference. Specifically, the model inference unit A3 applies the inference data provided by the data collection unit A1 to the trained AI/ML model (or the updated AI/ML model) to obtain inference output data (Inference Output). For example,
y = ax + b
In this case, x corresponds to the inference data and y corresponds to the inference output data. Note that "y = ax + b" is an AI/ML model. A model with optimized slope and intercept, for example, "y = 5x + 3", is a trained AI/ML model. There are various model approaches, including linear regression analysis, neural networks, and decision tree analysis. The above "y = ax + b" can also be considered a type of linear regression analysis.

The model inference unit A3 outputs inference output data to the management unit A5. The model inference unit A3 also receives management instructions from the management unit A5. For example, management instructions include selecting an AI/ML model, activating (deactivating) an AI/ML model, switching between AI/ML models, and fallback (performing inference without using an AI/ML model). The model inference unit A3 performs model inference in accordance with the management instructions.

Note that AI/ML model inference is, for example, the process of obtaining a set of outputs from a set of inputs using a trained AI/ML model (or an updated AI/ML model). Alternatively, model inference may be the process of obtaining inference output data from inference data using a trained AI/ML model (or an updated AI/ML model). Below, AI/ML model inference may be referred to as "model inference" or "inference."

In the following, an AI/ML model that is currently being trained (or updated) may be referred to as a training AI/ML model (or an updating AI/ML model). In the following, when there is no need to distinguish between a training (or updating) AI/ML model and a trained (or updated) AI/ML model, they may simply be referred to as an "AI/ML model."

The management unit A5 supervises operations on the AI/ML model (selection, activation, deactivation, switching, fallback, etc.). The management unit A5 also supervises monitoring of the AI/ML model. The management unit A5 can also perform operations to ensure appropriate inference operations based on monitoring data and inference output data. To this end, the management unit A5 outputs a model transfer and/or model delivery request (Model Transfer/Delivery Request) to the model recording unit A6, and causes the trained (or updated) AI/ML model recorded in the model recording unit A6 to be output to the model inference unit A3. The management unit A5 also outputs management instructions to the model inference unit A3 and supervises operations on the AI/ML model. Furthermore, the management unit A5 can output performance feedback and a re-learning request to the model learning unit A2, causing the model learning unit A2 to re-learn the AI/ML model (i.e., update the learned AI/ML model).

Figure 7 shows an example of the operation of the AI/ML technology according to the first embodiment.

In FIG. 7, the transmitting entity TE is an entity capable of performing model inference and transmitting inference output data to the receiving entity RE. On the other hand, the receiving entity RE is an entity capable of receiving inference output data from the transmitting entity TE. Model training may be performed in the transmitting entity TE. The model training may also be performed in the receiving entity RE. When model training is performed in the receiving entity RE, the trained AI/ML model may be transmitted from the receiving entity RE to the transmitting entity TE.

Note that an entity may be, for example, a device. The entity may also be a functional block included in the device. The entity may also be a hardware block included in the device.

For example, the transmitting entity TE may be UE100, and the receiving entity RE may be gNB200 or a core network device. Alternatively, the transmitting entity TE may be gNB200 or a core network device, and the receiving entity RE may be UE100.

As shown in FIG. 7, in step S1, the transmitting entity TE transmits control data related to AI/ML technology to the receiving entity RE and receives the control data from the receiving entity RE. The control data may be an RRC message, which is signaling of the RRC layer (i.e., layer 3). The control data may be a MAC Control Element (CE), which is signaling of the MAC layer (i.e., layer 2). The control data may be downlink control information (DCI), which is signaling of the PHY layer (i.e., layer 1). The downlink signaling may be UE-specific signaling. The downlink signaling may be broadcast signaling. The control data may be a control message in a control layer (e.g., the AI/ML layer) specialized for artificial intelligence or machine learning. Alternatively, the control data may be a NAS message in the NAS layer. The control data may include a performance feedback request and/or a re-learning request sent from the management unit A5 to the model learning unit A2. Alternatively, the control data may include a model transfer request and/or a model delivery request sent from the management unit A5 to the model recording unit A6. Alternatively, the control data may include a management instruction sent from the management unit A5 to the model inference unit A3.

(Deployment examples and use cases)
Next, a description will be given of how the functional blocks shown in Fig. 6 are arranged in the mobile communication system 1. Below, an example of the arrangement of the functional blocks will be described along with a specific use case.

For example, there are three use cases where AI/ML technology can be applied:

(X1.1) "CSI (Channel State Information) Feedback Enhancement"

(X1.2) "Beam Management"

(X1.3) “Positioning accuracy enhancement”

(X1.1) Example of functional block arrangement in "CSI feedback improvement""CSI feedback improvement" represents a use case in which AI/ML technology is applied to CSI fed back from UE100 to gNB200, for example. CSI is information about the channel state in the downlink between UE100 and gNB200. The CSI includes at least one of a channel quality indicator (CQI), a precoding matrix indicator (PMI), and a rank indicator (RI). The gNB200 performs, for example, downlink scheduling based on the CSI feedback from UE100.

Figure 8 is a diagram showing an example of the arrangement of each functional block in "CSI feedback improvement." In the example of "CSI feedback improvement" shown in Figure 8, the data collection unit A1, model learning unit A2, and model inference unit A3 are included in the control unit 130 of the UE 100. On the other hand, the data processing unit A4 is included in the control unit 230 of the gNB 200. In other words, model learning and model inference are performed in the UE 100. Figure 8 shows an example in which the transmitting entity TE is the UE 100 and the receiving entity RE is the gNB 200.

In "improved CSI feedback," the gNB 200 transmits a reference signal for the UE 100 to estimate the downlink channel state. The following describes the reference signal using a CSI reference signal (CSI-RS) as an example, but the reference signal may also be a demodulation reference signal (DMRS).

First, in model learning, UE100 (receiving unit 110) receives a first reference signal from gNB200 using a first resource. Then, UE100 (model learning unit A2) uses learning data including the first reference signal and CSI to derive a learned model for inferring CSI from the reference signal. Such a first reference signal is sometimes referred to as a full CSI-RS.

For example, the CSI generation unit 131 performs channel estimation using the received signal (CSI-RS) received by the receiving unit 110 to generate CSI. The transmitting unit 120 transmits the generated CSI to the gNB 200. The model learning unit A2 performs model learning using a set of the received signal (CSI-RS) and CSI as learning data, and derives a learned model for inferring CSI from the received signal (CSI-RS).

Second, in model inference, the receiver 110 receives a second reference signal from the gNB 200 using second resources that are fewer than the first resources. Then, the model inference unit A3 uses the trained model to infer CSI as inference result data using the second reference signal as inference data. Hereinafter, such a second reference signal may be referred to as a partial CSI-RS or a punctured CSI-RS.

For example, the model inference unit A3 inputs the partial CSI-RS received by the receiver 110 as inference data into the trained model and infers CSI from the CSI-RS. The transmitter 120 transmits the inferred CSI to the gNB 200.

This enables UE100 to feed back (or transmit) accurate (complete) CSI to gNB200 from the limited CSI-RS (partial CSI-RS) received from gNB200. For example, gNB200 can reduce (puncture) CSI-RS when intended to reduce overhead. It also enables UE100 to respond to situations where the radio conditions deteriorate and some CSI-RS cannot be received normally.

Figure 9 shows an example of operation for "CSI feedback improvement" according to the first embodiment.

As shown in FIG. 9, in step S10, gNB200 may notify or set the CSI-RS transmission pattern (puncture pattern) in inference mode to UE100 as control data. For example, gNB200 transmits to UE100 the antenna ports and/or time-frequency resources that will or will not transmit CSI-RS in inference mode.

In step S11, gNB200 may send a switching notification to UE100 to start learning mode.

In step S12, UE100 starts learning mode.

In step S13, gNB200 transmits full CSI-RS. UE100's receiver 110 receives the full CSI-RS, and CSI generator 131 generates (or estimates) CSI based on the full CSI-RS. In learning mode, data collector A1 collects full CSI-RS and CSI. Model learning unit A2 uses the full CSI-RS and CSI as learning data to create a learned AI/ML model.

In step S14, UE100 transmits the generated CSI to gNB200.

Then, in step S15, when the model learning is completed, UE100 transmits a completion notification to gNB200 indicating that the model learning is completed. UE100 may also transmit a completion notification when the creation of the trained model is completed.

In step S16, in response to receiving the completion notification, gNB200 transmits a switching notification to UE100 to switch from learning mode to inference mode.

In step S17, in response to receiving the switching notification, UE100 switches from learning mode to inference mode.

In step S18, gNB200 transmits partial CSI-RS. Receiver 110 of UE100 receives the partial CSI-RS. In inference mode, data collector A1 collects the partial CSI-RS. Model inference unit A3 inputs the partial CSI-RS as inference data into the trained model, and obtains CSI as the inference result.

In step S19, UE100 feeds back (or transmits) the CSI, which is the inference result, to gNB200 as inference result data. UE100 can generate a trained model with a predetermined accuracy or higher by repeating model learning during learning mode. It is expected that the inference results using the trained model generated in this way will also have a predetermined accuracy or higher.

In addition, in step S20, if UE100 determines by itself that model learning is necessary, it may transmit a notification indicating that model learning is necessary to gNB200 as control data.

In the example shown in Figure 9, the training data is "(full) CSI-RS" and "CSI," and the inference data is "(partial) CSI-RS." Hereinafter, the training data and/or the inference data may be referred to as a "dataset."

In "CSI Feedback Improvement," in addition to "CSI-RS" and "CSI," at least one of the following data or information may also be used as a data set:

(Y1) RSRP (Reference Signals Received Power), RSRQ (Reference Signal Received Quality), SINR (Signal-to-Interference-plus-Noise Ratio), or AD converter output waveform (These measurements may be CSI-RS. These measurements may also be other received signals received from gNB200.)

(Y2) Bit error rate (BER) or block error rate (BLER) (BER (or BLER) may be measured based on CSI-RS, assuming the total number of transmitted bits (or total number of transmitted blocks) is known.)

(Y3) Moving speed of UE 100 (may be measured by a speed sensor in UE 100)
It may be set what is used as the dataset used in machine learning. For example, the following processing may be performed. That is, UE100 transmits capability information indicating what type of input data it can handle in machine learning to gNB200 as control data. The capability information may represent, for example, any of the data or information shown in (Y1) to (Y3). The capability information may be information in which learning data and inference data are separately specified. Then, gNB200 transmits data type information to be used as the dataset to UE100 as control data. The data type information may represent, for example, any of the data or information shown in (Y1) to (Y3). Furthermore, the data type information may separately specify data type information used as learning data and data type information used as inference data.

The above describes an example of the functional block layout in (X1.1) "CSI Feedback." The above layout example is just one example, and in 3GPP, functional block layout examples are still under consideration. Similarly, (X1.2) "Beam Management" and (X1.3) "Position Accuracy Improvement" are also still under consideration.

(X1.4) Example of Model Transfer Next, we will explain the transfer of an AI/ML model (Model Transfer). Note that the terms "transfer of an AI/ML model" and "delivery of an AI/ML model" have the same meaning.

(X1.4.1) First operation pattern related to model forwarding Figure 10 is a diagram showing an example of an operation of the first operation pattern related to model forwarding according to the first embodiment. In the example shown in Figure 10, the receiving entity RE will be described as mainly being the UE 100, but the receiving entity RE may be the gNB 200 or the AMF 300. Also, in the example shown in Figure 10, the transmitting entity TE will be described as being the gNB 200, but the transmitting entity TE may be the UE 100 or the AMF 300.

As shown in FIG. 10, in step S25, gNB200 transmits a capability inquiry message to UE100 to request transmission of a message including an information element (IE) indicating the execution capability for the learning process. UE100 receives the capability inquiry message. However, gNB200 may also transmit the capability inquiry message when it executes the learning process (when it determines that it will execute the learning process).

In step S26, UE100 transmits a message to gNB200 that includes information elements indicating its execution capability for the learning process (or, from another perspective, the execution environment for the learning process). gNB200 receives the message. The message may be an RRC message (e.g., a "UE Capability" message or a newly defined message (e.g., a "UE AI Capability" message, etc.)). Alternatively, the transmitting entity TE may be AMF300, and the message may be a NAS message. Alternatively, if a new layer is defined for performing or controlling the learning process (AI/ML process), the message may be a message of the new layer.

The information element indicating the execution capability for the learning process may be an information element indicating the processor's capability for executing the learning process and/or an information element indicating the memory's capability for executing the learning process. Specifically, the information element indicating the processor's capability may be an information element indicating the product number (or model number) of the AI processor. Specifically, the information element indicating the memory's capability may be an information element indicating the memory capacity.

Alternatively, the information element indicating the execution capability for the learning process may be an information element indicating the execution capability for the inference process (model inference). Specifically, the information element indicating the execution capability for the inference process may be an information element indicating whether a deep neural network model is supported, or an information element indicating the time (or response time) required to execute the inference process.

Alternatively, the information element indicating the execution capacity for the learning process may be an information element indicating the execution capacity of the learning process (model learning). Specifically, the information element indicating the execution capacity of the learning process may be an information element indicating the number of learning processes being executed simultaneously, or an information element indicating the processing capacity of the learning process.

In step S27, gNB200 determines the model to be configured (or deployed) in UE100 based on the information elements included in the message received in step S26.

In step S28, gNB200 transmits a message including the model determined in step S27 to UE100. UE100 receives the message and performs a learning process (i.e., a model learning process and/or a model inference process) using the model included in the message. A specific example of step S28 will be described in the second operation pattern below.

(X1.4.2) Second Operation Pattern Related to Model Transfer FIG. 11 is a diagram showing an example of a configuration message including a model and additional information according to the first embodiment. The configuration message may be an RRC message transmitted from the gNB 200 to the UE 100 (for example, an "RRC Reconfiguration" message, or a newly defined message (for example, an "AI Deployment" message or an "AI Reconfiguration" message, etc.). Alternatively, the configuration message may be a NAS message transmitted from the AMF 300 to the UE 100. Alternatively, when a new layer for performing or controlling machine learning processing (AI/ML processing) is defined, the message may be a message of the new layer.

In the example of Figure 11, the setting message includes three models (Model #1 to #3). Each model is included as a container in the setting message. However, the setting message may include only one model. The setting message further includes, as additional information, three individual additional information pieces (Info #1 to #3) that correspond to each of the three models (Model #1 to #3), and common additional information (Meta-Info) that is commonly associated with all three models (Model #1 to #3). Each of the individual additional information pieces (Info #1 to #3) includes information unique to the corresponding model. The common additional information (Meta-Info) includes information common to all models in the setting message.

The individual additional information may be a model index that indicates an index (index number) assigned to each model. The individual additional information may also be model execution conditions that indicate the performance (e.g., processing delay) required to apply (execute) the model.

The individual additional information or common additional information may be a model application that specifies the function to which the model is to be applied (e.g., "CSI feedback," "beam management," "positioning," etc.). The individual additional information or common additional information may also be a model selection criterion that applies (executes) the corresponding model when a specified criterion (e.g., movement speed) is met.

(Fine-tuning and re-learning according to the first embodiment)
Depending on the use case or scenario, transferring the trained AI/ML model to the UE 100 may be beneficial, even at the expense of overhead and delay. However, in general, it may not be appropriate in consideration of the storage capacity of the UE 100.

Here, when focusing on fine-tuning and/or relearning of a trained AI/ML model, it is conceivable that fine-tuning and/or relearning may replace the transfer of the trained AI/ML model. In other words, it is conceivable that if fine-tuning and/or relearning are performed appropriately, the transfer of the trained AI/ML model may not be necessary. This makes it possible, for example, in mobile communication system 1, to perform appropriate model inference operations using the updated AI/ML model after updating, without having to consider the storage capacity of UE 100 or network overhead and delays.

Therefore, in the first embodiment, we focus on fine-tuning and relearning. In particular, in the first embodiment, we focus on fine-tuning. First, we will explain how relearning is handled in 3GPP.

(Relearning according to the first embodiment)
3GPP defines life cycle management (LCM) for AI/ML models. From the perspective of LCM, processes are performed on an AI/ML model in the following order: model learning (i.e., "learning"), model inference (i.e., "inference"), model monitoring, and model update. Although each process is not strictly defined in 3GPP, it is assumed that re-learning is performed at the model update stage after model monitoring. In other words, it is assumed that model monitoring is performed on a trained AI/ML model, and that re-learning of the trained AI/ML model is performed based on the results of model monitoring.

On the other hand, in the functional block configuration shown in Figure 6, the management unit (Management) A5 issues a re-learning request to the model training unit (Model Training) A2, which causes re-learning to be performed in the model training unit A2. Therefore, from the perspective of the functional block configuration shown in Figure 6, after management in the management unit A5, re-learning is performed in the model training unit A2.

This will be explained specifically using a use case. Figures 12 and 13 are diagrams showing an example of operation according to the first embodiment. In particular, Figures 12 and 13 show an example of operation in a UE-side model (a model in which inference is performed on the UE 100 side) in the use case of "Beam management." Figures 12 and 13 also show an example of operation from inference to relearning.

As shown in Figures 12 and 13, after inference (step S30 in Figure 12), management (step S40) is performed, and then relearning (step S50 in Figure 13) is performed. In particular, in the example shown in Figures 12 and 13, monitoring is performed in management (step S40). In the use case of "beam management" in the UE-side model, 3GPP specifies cases where monitoring is performed by the UE 100 and cases where it is performed by the gNB 200. Specifically, there are the following three cases:

(Z1) Monitoring is performed on the UE 100 side, and control of the UE's trained AI/ML model (selection, (de)activation, switching, fallback, etc.; hereinafter, sometimes referred to as "model control") is also performed on the UE 100 side.

(Z2) Monitoring is performed on the UE 100 side, and model control of the UE's trained AI/ML model is performed on the gNB 200 side.

(Z3) When gNB200 monitors and controls the trained AI/ML model on the UE side In FIG. 12, steps S41 and S42 represent the above-described case of (Z1). That is, in step S41, the control unit 130 of UE100 monitors the performance of the inference (step S30) performed by UE100. Then, in step S42, the control unit 130 performs model control based on the monitoring result. The control unit 130 may perform model control based on the monitoring result that the inference performance is equal to or less than a performance threshold. As model control, the control unit 130 may perform re-learning of the trained AI/ML model. Here, the performance threshold is a threshold used to determine whether the inference performance (or accuracy) of the trained AI/ML model is good or bad. The control unit 130 may compare the inference results (or monitoring results, or the difference between the inference results and the monitoring results) of the trained AI/ML model with a performance threshold to determine the performance of the trained AI/ML model.

Steps S43 to S45 represent the case of (Z2) described above. That is, in step S43, the control unit 130 of UE100 monitors the performance of the inference (step S30) performed by UE100. In step S44, the transmission unit 120 of UE100 transmits the monitoring results to gNB200. The transmission unit 120 may transmit control data including the monitoring results to gNB200. The reception unit 220 of gNB200 receives the monitoring results. In step S45, the control unit 230 of gNB200 decides to perform model control based on the monitoring results and instructs UE100 to perform model control. The transmission unit 210 of gNB200 transmits the instruction to UE100. The instruction may also be transmitted included in the control data. The reception unit 110 of UE100 receives the instruction. The control unit 130 of the UE 100 may decide to re-learn the trained AI/ML model in accordance with the instruction to execute model control.

Steps S46 to S49 represent the case of (Z3) described above. That is, in step S46, the control unit 130 of UE100 calculates a performance metric (or performance metrics) related to the performance of the inference (step S30) performed by UE100. The performance metric represents, for example, an evaluation index of the trained AI/ML model when inference is performed. Specifically, the performance metric is a value for analyzing the accuracy of the current inference, i.e., the confidence level, and may be squared generalized cosine similarity (SGCS), throughput, block error rate (BLER), etc. The performance metric may also include CPU usage, memory usage, inference time, and/or data usage during inference. The transmission unit 120 of UE100 transmits the performance metric to gNB200. The transmission unit 120 may also transmit control data including the performance metric to gNB200. The receiver 220 of the gNB 200 receives the performance metric. In step S48, the control unit 230 of the gNB 200 uses the performance metric to monitor the performance of the inference (step S30). If the control unit 230 obtains a detection result based on the monitoring result that the inference performance of the trained AI/ML model on the UE 100 side is below the performance threshold, it instructs the UE 100 to perform model control. The model control instruction may be a re-learning instruction, as in step S45. The receiver 110 of the UE 100 receives the model control instruction. The control unit 130 of the UE 100 may decide to perform re-learning in accordance with the model control instruction.

Then, in step S50 shown in FIG. 13, re-learning is performed. Re-learning may be performed in UE 100. That is, in step S51, control unit 130 of UE 100 generates learning data. Then, in step S52, control unit 130 uses the learning data to perform re-learning on the trained AI/ML model.

Alternatively, the re-learning may be performed by the OTT (Over The Top) server 500. The trained AI/ML model may be stored in the OTT server 500. Note that the OTT server may be an external server located outside the mobile communication system 1. In this case, in step S53, the control unit 130 of the UE 100 generates training data, and in step S54, the transmission unit 120 of the UE 100 transmits the training data to the OTT server 500. In step S55, the OTT server 500 performs the re-learning.

(Communication control method according to the first embodiment)
The operation examples shown in Figures 12 and 13 are examples of operation related to re-learning. Currently, 3GPP does not strictly distinguish between re-learning and fine-tuning. In the first embodiment, fine-tuning is distinguished from re-learning, and fine-tuning of a trained AI/ML model is appropriately performed in the mobile communication system 1. This makes it possible, for example, in the mobile communication system 1, to create an updated AI/ML model more quickly than in the case of re-learning, and to perform inference using the updated AI/ML model more quickly than in the case of re-learning.

In other words, the first embodiment aims to enable appropriate fine-tuning of a trained AI/ML model.

In the first embodiment, to distinguish between fine-tuning and relearning, fine-tuning and relearning are defined as follows:

In other words, fine-tuning, for example, means training a trained AI/ML model using training data with a data volume equal to or less than a data volume threshold. Alternatively, fine-tuning may mean training a trained AI/ML model using training data with a data volume equal to or less than a data volume threshold, and further training layers including the final layer that are equal to or less than a layer threshold (hereinafter sometimes referred to as a "predetermined number of layers") out of multiple layers up to the final layer that make up the trained AI/ML model. By fine-tuning a predetermined number of layers (parameters of the predetermined number of layers) out of multiple layers included in the trained AI/ML model, it becomes possible to adapt it to a specific new task or a specific new dataset.

On the other hand, re-learning, for example, refers to training a trained AI/ML model using training data with a data volume greater than the data volume threshold. Alternatively, re-learning may refer to training a trained AI/ML model using training data with a data volume greater than the data volume threshold, and may further refer to training layers, including the final layer, that are greater than the layer threshold and that are part of multiple layers up to the final layer that make up the trained AI/ML model.

When comparing fine-tuning and relearning, the following can be said:

First, fine-tuning requires less data to train than re-training. Therefore, fine-tuning can obtain an updated AI/ML model more quickly than re-training.

Second, fine-tuning allows training to be performed on fewer layers than re-training. Therefore, fine-tuning makes it possible to adapt an updated AI/ML model to a specific new task or a specific new dataset. On the other hand, re-training changes parameters for many layers of a trained AI/ML model compared to fine-tuning. Therefore, re-training makes it possible to create a new, updated AI/ML model that corresponds to a new dataset compared to fine-tuning.

Thirdly, re-learning takes longer to process than fine-tuning, and may require a lot of time. On the other hand, fine-tuning can be performed faster and more easily than re-learning.

Fine-tuning may also be called model fine-tuning. Re-learning may also be called model re-learning.

In the first embodiment, in order to distinguish fine-tuning from relearning and to ensure that fine-tuning can be performed appropriately, fine-tuning is performed as follows. That is, a user device (e.g., UE100) performs fine-tuning on a trained AI/ML model based on a fine-tuning execution time that indicates the time required to fine-tune the trained AI/ML model. Here, fine-tuning means training the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

In this way, in the first embodiment, fine-tuning is performed in UE 100 based on the fine-tuning execution time. For example, it is possible to perform fine-tuning when the fine-tuning execution time is equal to or less than the fine-tuning execution time threshold, and not perform fine-tuning when the fine-tuning execution time requires more time than the fine-tuning execution time threshold. This allows fine-tuning to be performed in UE 100 under certain judgment, for example, making it possible to appropriately perform fine-tuning on a trained AI/ML model.

Furthermore, for example, although it is not necessary to train the entire trained AI/ML model to obtain an updated AI/ML model, training the entire trained AI/ML model may not always be efficient. By performing fine-tuning, it is possible to train only a portion of the trained AI/ML model and efficiently obtain an updated AI/ML model.

(Operation example according to the first embodiment)
Next, an example of operation according to the first embodiment will be described.

FIGS. 14 and 15 are diagrams showing an example of operation according to the first embodiment. Note that while FIG. 14 and FIG. 15 show a gNB 200 and a network device 400, "gNB 200" indicates a case where processing is performed by gNB 200 alone among the network devices 400. "gNB 200" may also be a network node. On the other hand, "network device 400" indicates a case where processing is performed by any of the network devices 400, including gNB 200. Furthermore, FIG. 14 and FIG. 15 are a use case of beam management, and show an example of operation in the case of a UE-side model.

In addition, in Figures 14 and 15, information such as conditions and instructions, data, or messages are transmitted and received between UE 100 and gNB 200. This information may be transmitted and received using the control data described above. In the following explanation, the fact that transmission and reception are performed using control data may be omitted.

Furthermore, in Figures 14 and 15, information, data, messages, etc., such as AI/ML models, are transmitted and received between UE 100 and network device 400. This information, etc., may be transmitted and received using a predetermined message according to a predetermined protocol set between UE 100 and network device 400. For example, the predetermined message may be control data if network device 400 is a gNB 200, or a NAS message if network device 400 is an AMF. In the following explanation, the fact that transmission and reception are performed using a predetermined message may be omitted.

As shown in FIG. 14, in step S60, the network device 400 transmits (transfers) the trained AI/ML model to the UE 100. The receiver 110 of the UE 100 receives the trained AI/ML model.

In step S61, the network device 400 transmits parameters for calculating the fine-tuning execution time (hereinafter, sometimes referred to as "fine-tuning execution time calculation parameters") to the UE 100.

First, the parameters for calculating the fine-tuning execution time may include information representing the type of trained AI/ML model (step S60). The information representing the type may be, for example, ensemble learning, neural network, etc. Ensemble learning is, for example, a machine learning method that combines multiple trained AI/ML models to obtain output. Random forest, which is one method of ensemble learning, may be indicated as information representing the type of model. Random forest is a machine learning method that uses ensemble learning with decision trees.

Second, the parameters for calculating the fine-tuning execution time may include the complexity of the trained AI/ML model (step S60). Specifically, the complexity may be expressed as the number of neurons and hidden layers of the trained AI/ML model, the number of data items to be handled, the number of loss function calculations, and/or the number of gradient calculations.

Third, a sample may be transmitted from the network device 400 to the UE 100 along with parameters for calculating the fine-tuning execution time. The sample represents, for example, the fine-tuning execution time when the CPU (or GPU) performs fine-tuning using a predetermined amount of learning data with a predetermined memory capacity.

Fourth, the parameters for calculating the fine-tuning execution time may include identification information for the trained AI/ML model for which the fine-tuning execution time is to be calculated. The identification information may be the model ID, function name, or model name of the trained AI/ML model.

Note that the trained AI/ML model (step S60) and the parameters for calculating the fine-tuning execution time (step S61) may be sent in a single message, or may be sent in separate messages, as shown in Figure 14.

In step S62, the OTT server 500 may transmit the trained AI/ML model to the UE 100. Also, in step S63, the OTT server 500 may transmit parameters for calculating the fine-tuning execution time to the UE 100. In step S63, the OTT server may transmit the above-mentioned sample to the UE 100. Taking into account the processing load on the network side, the OTT server may train the AI/ML model and generate a trained AI/ML model. Taking such cases into account, steps S62 and S63 may be performed. The OTT server 500 may perform steps S62 and S63 using messages according to a protocol established between the OTT server 500 and the UE 100. The receiver 110 of the UE 100 receives the trained AI/ML model and the parameters for calculating the fine-tuning execution time.

In step S64, the control unit 130 of the UE 100 calculates the fine-tuning execution time for the trained AI/ML model using the parameters for calculating the fine-tuning execution time. The control unit 130 may perform fine-tuning on the trained AI/ML model and obtain the fine-tuning execution time by calculating (or measuring) it through actual measurement. The fine-tuning execution time may represent the time from the start to the end of the fine-tuning.

In step S65, the transmitter 120 of the UE 100 transmits the fine-tuning execution time to the gNB 200. The receiver 220 of the gNB 200 receives the fine-tuning execution time.

In step S66, the control unit 130 of the UE 100 performs inference on the trained AI/ML model received in step S60 (or step S62).

In step S67, management is performed in the mobile communication system 1.

In the first embodiment, management will be explained for each of the cases (Z1) to (Z3).

(1-1) When monitoring and model control are performed on the UE side (case (Z1))

In the case of (Z1), steps S68 to S71 are executed.

Specifically, in step S68, the control unit 230 of the gNB 200 determines fine-tuning execution conditions indicating the conditions for performing fine-tuning of the trained AI/ML model in the UE 100, based on the fine-tuning execution time received in step S65.

First, the fine-tuning execution condition may include a fine-tuning execution time threshold. In UE 100, fine-tuning may be performed when the fine-tuning execution time is equal to or less than the fine-tuning execution time threshold, and fine-tuning may not be performed when the fine-tuning execution time takes longer than the fine-tuning execution time threshold. Alternatively, the reverse may be true. The fine-tuning execution condition may include information indicating that fine-tuning is performed if the fine-tuning execution time is equal to or less than the fine-tuning execution time threshold (or information indicating that fine-tuning is not performed if the fine-tuning execution time exceeds the fine-tuning execution time threshold).

Second, the fine-tuning execution conditions may include information regarding model control when fine-tuning is not performed. The information regarding model control may be, for example, information indicating re-learning, falling back to legacy (i.e., obtaining output without using a trained AI/ML model), or doing nothing.

Third, the fine-tuning execution conditions may include the conditions for using the updated AI/ML model after fine-tuning has been performed. The conditions for use may be expressed, for example, by time and/or location. The location may be expressed, for example, by any of a cell ID, an RNA (RAN-based Notification Area), and a PLMN (Public Land Mobile Network).

Fourth, the fine-tuning execution conditions may include a request to transmit the dataset (learning data) used in the fine-tuning. If a transmission request is included, the transmitter 120 of the UE 100 will transmit the dataset used in the fine-tuning to the gNB 200.

Fifth, the fine-tuning execution conditions may include information regarding whether or not to register the updated AI/ML model after fine-tuning has been performed. If the fine-tuning execution conditions include this information, UE100 will transmit information indicating whether or not to register the updated AI/ML model to gNB200.

Sixth, the fine-tuning execution conditions may include information instructing that fine-tuning be performed. Alternatively, the fine-tuning execution conditions may include information instructing that fine-tuning not be performed. The control unit 130 of UE100 may decide to perform or not perform fine-tuning in accordance with the instruction information. In this case, the control unit 230 of gNB200 may decide to perform (or not perform) fine-tuning in UE100 based on the fine-tuning execution time received from UE100 in step S65.

Seventh, the fine-tuning execution conditions may include identification information (such as a model ID or function name) of the trained AI/ML model on which fine-tuning is to be performed.

In step S69, the transmitter 210 of the gNB 200 transmits the fine-tuning execution conditions to the UE 100. The receiver 220 of the UE 100 receives the fine-tuning execution conditions.

In step S70, the control unit 130 of the UE 100 monitors the performance of the inference of the trained AI/ML model (step S66). Then, based on the monitoring results, the control unit 130 detects that the performance (or accuracy) of the inference of the trained AI/ML model is poor, i.e., that the monitoring results are below the performance threshold. The monitoring results indicate, for example, data measured using a legacy method that does not use AI/ML functions (in the "beam management" use case, RSRP (Reference Signal Received Power) or beam angle, etc.). The monitoring results may also be CPU (or GPU) usage, memory usage, free memory capacity, and/or monitoring time. The performance threshold may be a value corresponding to these indicators of the monitoring results. The control unit 130 may execute the subsequent step S71 when the monitoring results are below the performance threshold.

In step S71, the control unit 130 of the UE 100 determines whether to perform fine-tuning based on the fine-tuning execution conditions received in step S69. Specifically, the control unit 130 determines whether the fine-tuning execution time satisfies the fine-tuning execution conditions based on the fine-tuning execution time calculated in step S64 and the fine-tuning execution conditions received in step S69, and if it determines that the fine-tuning execution conditions are met, it decides to perform fine-tuning. For example, the control unit 130 may decide to perform fine-tuning if the fine-tuning execution time is equal to or less than the fine-tuning execution time threshold included in the fine-tuning execution conditions. On the other hand, if the control unit 130 determines that the fine-tuning execution time does not satisfy the fine-tuning execution conditions, it decides not to perform fine-tuning. For example, the control unit 130 may decide not to perform fine-tuning if the fine-tuning execution time requires more time than the fine-tuning execution time threshold. If the fine-tuning execution conditions include information instructing the execution of fine-tuning (or information instructing not to perform fine-tuning), the control unit 130 may decide whether to perform fine-tuning in accordance with that instruction information.

(1-2) When monitoring is performed on the UE 100 side and model control is performed on the gNB 200 side (case (Z2))

In the case of (Z2), steps S75 to S78 shown in Figure 15 are executed.

In other words, in step S75, the control unit 130 of the UE 100 monitors the performance of the inference (step S66) for the trained AI/ML model.

In step S76, the transmitter 120 of UE100 transmits the monitoring results to gNB200. The receiver 220 of gNB200 receives the monitoring results. The control unit 230 of gNB200 detects, based on the monitoring results, that the performance of the trained AI/ML model is poor, i.e., that the monitoring results are below the performance threshold. The control unit 230 may perform step S77 when the monitoring results are below the performance threshold.

In step S77, the control unit 230 of the gNB 200 determines whether or not to perform fine-tuning in the UE 100 based on the fine-tuning execution time received in step S65. The determination itself may be the same as step S71. That is, the control unit 230 may decide to perform fine-tuning if the fine-tuning execution time is equal to or less than the fine-tuning execution time threshold, and may decide not to perform fine-tuning if the fine-tuning execution time takes longer than the fine-tuning execution time threshold. The fine-tuning execution time threshold may be stored in advance in the memory of the gNB 200. Here, the following description will be given assuming that the control unit 230 decides to perform fine-tuning based on the fine-tuning execution time.

In step S78, the transmitter 210 of the gNB 200 transmits a fine-tuning execution instruction to the UE 100, instructing the UE 100 to perform fine-tuning. The receiver 110 of the UE 100 receives the fine-tuning execution instruction. The control unit 130 of the UE 100 may decide to perform fine-tuning in accordance with the fine-tuning execution instruction. The fine-tuning execution instruction may include identification information (such as a model ID or function name) of the trained AI/ML model that is the target of fine-tuning execution. The transmitter 210 of the gNB 200 may instruct the UE 100 not to perform fine-tuning. The control unit 130 of the UE 100 may decide not to perform fine-tuning in response to receiving the instruction.

(1-3) When monitoring and model control are performed on the gNB200 side (case (Z3))

In the case of (Z3), steps S79 to S83 are executed.

In other words, in step S79, the control unit 130 of the UE 100 calculates the inference metric by performing inference on the trained AI/ML model (step S66). The inference metric may be the same as that in step S46 of FIG. 12.

In step S80, the transmitter 120 of the UE 100 transmits the inference metric to the gNB 200. The receiver 220 of the gNB 200 receives the performance metric.

In step S81, the control unit 230 of the gNB 200 monitors the performance of the inference performed by the UE 100 (step S66). As in step S76, the control unit 230 detects, based on the monitoring results, that the inference performance of the trained AI/ML model is poor (i.e., the monitoring results are below the performance threshold). At this time, the control unit 230 may monitor the inference performance using performance metrics received in advance from the UE 100. The control unit 230 may perform the subsequent step S82 by detecting that the monitoring results are below the performance threshold.

In step S82, the control unit 230 of the gNB 200 determines whether or not to perform fine-tuning in the UE 100 based on the fine-tuning execution time received in step S65. The determination itself may be the same as in step S77. Here, the following description will be given assuming that the control unit 230 decides to perform fine-tuning.

In step S83, the transmitter 210 of the gNB 200 transmits a fine-tuning execution instruction to the UE 100, instructing the UE 100 to perform fine-tuning. The transmission of the fine-tuning execution instruction itself may be the same as step S78. The receiver 110 of the UE 100 receives the fine-tuning execution instruction. The control unit 130 of the UE 100 may decide to perform fine-tuning in accordance with the fine-tuning execution instruction.

(1-4) Fine Adjustment Then, in step S84, fine adjustment is performed.

First, fine-tuning may be performed on the UE 100 side. That is, in step S85, the control unit 130 of the UE 100 generates learning data to be used for fine-tuning. In step S86, the control unit 130 uses the learning data to perform fine-tuning on the trained AI/ML model.

Secondly, the fine-tuning may be performed on the OTT server 500 side. The OTT server 500 may store the trained AI/ML model to be fine-tuned. In this case, in step S87, the control unit 130 of the UE 100 generates training data to be used for the fine-tuning. In step S88, the transmission unit 120 of the UE 100 transmits the training data to the OTT server 500. The OTT server 500 receives the training data and uses the training data to perform fine-tuning on the trained AI/ML model.

(Another Operation Example 1 According to the First Embodiment)
In the first embodiment, a UE-side model in which inference is performed using a trained AI/ML model on the UE 100 side has been described as an example, but the type of inference is not limited to the UE-side model.

For example, the first embodiment can also be implemented in a two-sided model in which inference is performed on both the UE100 side and the gNB200 side. In this case, implementation is possible by applying the operation examples shown in Figures 14 and 15 to the inference performed on the UE100 side. Furthermore, for the inference performed on the gNB200 side, the control unit 230 calculates and monitors the fine-tuning execution time for the trained AI/ML model on the gNB200 side, and when performance is below the performance threshold, determines whether the fine-tuning execution condition is met (or whether the fine-tuning execution time is below the fine-tuning execution time threshold). Then, if the control unit 230 performs fine-tuning according to the determination result, implementation is also possible for the inference performed on the gNB200 side. Furthermore, the first embodiment can also be applied to a network-side model in which inference is performed on the network side (gNB200, core network, LMF (Location Management Function), or OTT server). For example, in Figures 14 and 15, by replacing "UE 100" with "network side," it is also possible to implement a network-side model. In this case, steps S60 and S61 do not need to be performed, and the network device 400 itself may hold the model, parameters, etc.

(Another Operation Example 2 According to the First Embodiment)
In the first embodiment, an example (step S65) in which UE100 transmits the fine-tuning execution time to gNB200 is described, but UE100 does not have to transmit the fine-tuning execution time to gNB200.

For example, when monitoring and model control are performed on the UE100 side (steps S68 to S71 in FIG. 14), the control unit 130 of UE100 may store the fine-tuning execution time in a memory or the like without transmitting it to the gNB200, and after monitoring (step S70), may determine whether to perform fine-tuning based on the fine-tuning execution conditions received from the gNB200 (step S69).

(Another Operation Example 3 According to the First Embodiment)
In the first embodiment, when monitoring is performed on the UE 100 side and model control is performed on the gNB 200 side (steps S75 to S78 in FIG. 15), an example has been described in which the UE 100 transmits the fine-tuning execution time (step S65) and the monitoring result (step S76) separately. For example, instead of transmitting the fine-tuning execution time in step S65, the UE 100 may transmit the fine-tuning execution time together with the monitoring result in step S76. Since the gNB 200 can acquire the fine-tuning execution time, the subsequent steps can be performed in the same manner as in the first embodiment.

(Another Operation Example 4 According to the First Embodiment)
In the first embodiment, beam management has been described as an example of a use case, but the use case is not limited to beam management. For example, the first embodiment can also be applied to CSI feedback improvement (X1.1) or position accuracy improvement (X1.3).

Second Embodiment
Next, a second embodiment will be described, focusing on the differences from the first embodiment.

In the first embodiment, we have described performing fine-tuning while taking into account the time required to perform the fine-tuning. In the second embodiment, we will describe evaluating the performance of a trained AI/ML model and performing fine-tuning based on the evaluation results. In particular, in the second embodiment, when a performance evaluation is performed, the time required for the performance evaluation is calculated, and a decision is made as to whether or not to perform the performance evaluation based on the time required for the performance evaluation.

Specifically, first, the user device (e.g., UE100) performs a performance evaluation based on a performance evaluation execution time that indicates the time required to evaluate the performance of the trained AI/ML model. Second, the user device performs fine-tuning of the trained AI/ML model based on the results of the performance evaluation. Here, fine-tuning means training the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

In this way, UE100 performs performance evaluation based on the performance evaluation execution time of the trained AI/ML model. Therefore, for example, UE100 can perform performance evaluation when the performance evaluation execution time is equal to or less than the performance evaluation time threshold, and can also not perform performance evaluation when the performance evaluation execution time is longer than the performance evaluation time threshold. This makes it possible, for example, for UE100 to reduce the performance evaluation execution time. Furthermore, by reducing the performance evaluation execution time, UE100 can also reduce the time until the fine-tuning execution is completed, making it possible to perform fine-tuning appropriately from a time perspective.

Furthermore, by evaluating the performance of the trained AI/ML model, UE100 can determine when a trained AI/ML model obtains optimal inference result data at a specific time or location, but does not necessarily obtain optimal inference result data at other times or locations (this is generally referred to as "overlearning"). Generally, when a trained AI/ML model is trained using training data obtained at a specific time or location, it is expected that optimal inference results will be obtained at that specific time or location. Performance evaluation can also capture the characteristics of such a trained AI/ML model. Furthermore, in UE100, for example, by performing fine-tuning when the evaluation result of the performance evaluation of the trained AI/ML model is equal to or greater than a threshold (e.g., a performance evaluation result threshold), the likelihood of obtaining optimal inference result data from the updated AI/ML model increases to a certain extent. Therefore, fine-tuning can be performed appropriately in mobile communication system 1.

(Evaluation index used in the second embodiment)
Here, the evaluation index of the trained AI/ML model used in the second embodiment will be described.

In general, regression evaluation and classification evaluation are sometimes used as evaluation indicators for AI/ML models.

Regression evaluation is, for example, evaluation using the difference between a predicted value (e.g., inference result data) and an actual value (e.g., an output result obtained without using an AI/ML model (i.e., legacy)). Examples of regression evaluation include mean squared error (MSE), normalized mean squared error (NMSE), and mean absolute error (MAE). MSE, for example, represents the root mean square of the difference between the predicted value and the actual value. NMSE, for example, represents the normalized root mean square of the difference between the predicted value and the actual value. NMSE, for example, represents the root mean square of the difference between the predicted value and the actual value. MAE, for example, represents the root mean square of the difference between the true value and the predicted value. All evaluation indices range from 0 to 1, with values closer to 0 indicating better performance and values closer to 1 indicating worse performance.

On the other hand, classification evaluation is a method of evaluation using, for example, the proportion of correct answers to predictions. Examples of classification evaluation include precision, accuracy, and recall. Accuracy, for example, represents the proportion of correct answers out of all predictions. Precision, for example, represents the proportion of those predicted as positive that are actually correct. Recall, for example, represents the proportion of those predicted as correct out of those that are actually positive.

In the second embodiment, evaluating a trained AI/ML model is referred to as "performance evaluation." In the second embodiment, the evaluation method used in "performance evaluation" is described as using NMSE, but evaluation methods other than NMSE may also be used.

As described above, in the second embodiment, the mobile communication system 1 determines whether to perform a performance evaluation based on the performance evaluation execution time required for the performance evaluation. If the mobile communication system 1 determines to perform a performance evaluation, it performs a performance evaluation on the trained AI/ML model and performs fine-tuning based on the evaluation results.

(Operation example according to the second embodiment)
Next, an example of operation according to the second embodiment will be described.

FIGS. 16 to 20 are diagrams showing examples of operation according to the second embodiment. In FIGS. 16 to 20, a use case of beam management is described, illustrating an example of operation in the case of a UE-side model. Furthermore, as shown in FIGS. 17 to 19, management is described using examples (Z1) to (Z3), similar to the first embodiment.

As shown in FIG. 16, in step S90, the network device 400 transfers (or transmits) the trained AI/ML model to the UE 100. The receiver 110 of the UE 100 receives the trained AI/ML model.

In step S91, the network device 400 transmits performance evaluation time calculation parameters to the UE 100. The performance evaluation calculation parameters represent parameters used to calculate the performance evaluation execution time.

First, the performance evaluation calculation parameters may include, for example, the number of datasets (or amount of data) used in the performance evaluation. The number of datasets used in the performance evaluation may be the same as or smaller than the number of training datasets used in the trained AI/ML model. In UE100, if the dataset contains data for which a performance evaluation index can be calculated using a purely mathematical method such as NMSE, for example, a predicted value (a value obtained by inference) and a correct value (ground truth data), the performance evaluation index may be calculated using a mathematical method, and the calculation time may be taken as the performance evaluation execution time. Furthermore, in UE100, if the dataset is input data and a correct value, the same number of datasets as the number of datasets may be input into the trained AI/ML model to perform inference, and performance evaluation may be performed on the inference results and the correct value to calculate the performance evaluation execution time. The performance evaluation execution time may represent the time from the start of inference with the trained AI/ML model (or the start of evaluation of the trained AI/ML model) to the acquisition of the performance evaluation results for the trained AI/ML model.

Second, the performance evaluation calculation parameters may include an evaluation method (such as NMSE) used in performance evaluation. UE 100 uses this evaluation method to perform performance evaluation and calculate the performance evaluation execution time. As in the first embodiment, network device 400 may transmit samples calculated using the performance evaluation time calculation parameters to UE 100.

Note that model transfer and transmission of performance evaluation calculation parameters and samples may be performed from the OTT server rather than from the network device 400 (steps S92 and S93).

In step S94, the network device 400 transmits a performance evaluation time calculation dataset to the UE 100. The performance evaluation time calculation dataset is a dataset used to calculate the performance evaluation execution time. As described above, the UE 100 may calculate a performance evaluation index using a mathematical method (e.g., NMSE) and use the calculation time as the performance evaluation execution time. The UE 100 may also perform inference to calculate the performance evaluation execution time. The receiving unit 110 of the UE 100 receives the performance evaluation time calculation dataset.

Note that if the amount of data in the dataset is equal to or greater than a predetermined threshold, UE100 may not be able to use the dataset to calculate the performance evaluation execution time. Alternatively, UE100 may find that the size of the trained AI/ML model received in step S90 is larger than a predetermined value and cannot be stored in memory. Therefore, in step S95, UE100 may transmit the received dataset to gNB200. This is to allow gNB200 to calculate the performance evaluation execution time. Furthermore, in step S96, a network device 400 other than gNB200 may transmit the dataset to gNB200 to allow gNB200 to calculate the performance evaluation execution time.

Alternatively, the data set may be transmitted from the OTT server 500 to the UE 100 (step S97). In this case, the UE 100 may transmit the received data set to the gNB 200 (step S98), causing the gNB 200 to calculate the performance evaluation execution time.

In step S99, the network device 400 transmits the current index to the UE 100. The current index represents the evaluation result when performance evaluation is currently performed. For example, if performance evaluation is performed using NMSE, the current index would be NMSE = 0.2, etc. If learning is performed in the network device 400, the network device 400 possesses a data set of predicted values and correct values, and therefore is able to evaluate the performance of the model. The network device 400 can transmit the acquired evaluation result as the current index to the UE 100. The UE 100 can use the current index in subsequent performance evaluation (step S117 in Figure 17).

Note that the performance evaluation execution time may be calculated in gNB200, and taking such cases into consideration, the UE may transmit the current indicators received in step S99 to gNB200 (step S100). Network devices 400 other than gNB200 may also transmit the current indicators to gNB200 (step S101).

The current indicator may also be transmitted from the OTT server 500 to the UE 100 (step S102), and the UE 100 may transmit the current indicator received from the OTT server 500 to the gNB 200 in order to cause the gNB 200 to perform calculation of the performance evaluation execution time (step S103).

In step S104, the control unit 130 of the UE 100 calculates the execution time of the performance evaluation using the parameters for calculating the performance evaluation time. As described above, the control unit 130 may calculate the performance evaluation index using a mathematical method (e.g., NMSE) and use the calculation time as the performance evaluation execution time. The control unit 130 may also perform inference to calculate the performance evaluation execution time. The control unit 130 may calculate (or measure) the time required to execute the performance evaluation and use the calculated time as the performance evaluation execution time.

In step S105, the transmitter 120 of the UE 100 transmits the performance evaluation execution time to the gNB 200. The receiver 220 of the gNB 200 receives the performance evaluation execution time.

In step S106, the control unit 130 of the UE 100 performs inference on the trained AI/ML model received in step S90.

Then, in step S110 (Figure 17), management is performed in the mobile communication system 1.

(2-1) When monitoring and model control are performed on the UE 100 side (case (Z1))

In the case of (Z1), steps S111 to S118 are executed.

That is, in step S111, the control unit 230 of the gNB 200 determines performance evaluation execution conditions that indicate the conditions for performing performance evaluation in the UE 100. The control unit 230 may determine the performance evaluation execution conditions based on the performance evaluation execution time received in step S105.

First, the performance evaluation execution condition may include a performance evaluation time threshold. In UE 100, performance evaluation may be performed when the performance evaluation execution time is equal to or less than the performance evaluation time threshold, and performance evaluation may not be performed when the performance evaluation execution time is longer than the performance evaluation time threshold. The performance evaluation execution condition may include information indicating that performance evaluation is to be performed when the performance evaluation time is equal to or less than the performance evaluation time threshold (or information indicating that performance evaluation is not to be performed when the performance evaluation time is longer than the performance evaluation time threshold).

Second, the performance evaluation execution conditions may include information instructing that performance evaluation be performed. Alternatively, the performance evaluation execution conditions may include information instructing that performance evaluation not be performed. The control unit 230 of the gNB 200 may decide to perform (or not perform) performance evaluation in the UE 100 based on the performance evaluation execution time received in step S105.

Third, the performance evaluation conditions may include information regarding the evaluation method (e.g., NMSE) used when performing the performance evaluation. UE 100 performs the performance evaluation using this evaluation method.

Fourth, the performance evaluation conditions may include identification information (such as a model ID or function name) of the trained AI/ML model for which performance evaluation is to be performed.

In step S112, the transmitter 210 of the gNB 200 transmits the performance evaluation execution conditions to the UE 100. The receiver 110 of the UE 100 receives the performance evaluation execution conditions.

In step S113, the control unit 230 of the gNB 200 determines fine-tuning execution conditions that indicate the conditions for performing fine-tuning in the UE 100. The fine-tuning execution conditions may be the same as the fine-tuning execution conditions described in the first embodiment (step S68 in FIG. 14).

However, the fine-tuning execution conditions according to the second embodiment include a performance evaluation result threshold that indicates a threshold for the evaluation result of the performance evaluation. For example, the UE 100 uses the performance evaluation result threshold to determine whether or not to perform fine-tuning. The fine-tuning execution conditions according to the second embodiment may include information indicating that fine-tuning is to be performed when the evaluation result of the performance evaluation is equal to or greater than the performance evaluation result threshold (or information indicating that fine-tuning is not to be performed when the evaluation result is less than the performance evaluation result threshold). Alternatively, they may include information instructing that relearning be performed if fine-tuning is not to be performed.

Note that the fine-tuning execution conditions in the second embodiment do not necessarily have to include the fine-tuning execution time threshold described in the first embodiment.

In step S114, the transmitter 210 of the gNB 200 transmits the fine-tuning execution conditions to the UE 100. The receiver 110 of the UE 100 receives the fine-tuning execution conditions.

In step S115, the control unit 130 of the UE 100 monitors the inference of the trained AI/ML model (step S106). As a result of the monitoring, the control unit 130 detects that the performance of the trained AI/ML model is below the performance threshold (i.e., the performance is poor). The control unit 130 may perform the subsequent step S116 by detecting that the performance of the trained AI/ML model is below the performance threshold.

In step S116, the control unit 130 of the UE 100 determines whether or not to perform performance evaluation based on the performance evaluation execution conditions received in step S112. Specifically, the control unit 130 determines whether or not to perform performance evaluation based on the performance evaluation execution time calculated in step S104 and the performance evaluation execution conditions received in step S112. For example, the control unit 130 performs performance evaluation when the performance evaluation execution time is equal to or less than the performance evaluation time threshold included in the performance evaluation execution conditions, and does not perform performance evaluation when the performance evaluation execution time exceeds the performance evaluation time threshold. In the following description, the control unit 130 is assumed to perform performance evaluation.

In step S117, the control unit 130 of the UE 100 performs a performance evaluation of the trained AI/ML model. For example, the control unit 130 performs the performance evaluation using an evaluation method (e.g., NMSE) included in the performance evaluation conditions.

In step S118, the control unit 130 of the UE 100 determines whether to perform fine-tuning. Specifically, the control unit 130 determines whether to perform fine-tuning based on the evaluation result of the performance evaluation performed in step S117 and the fine-tuning execution conditions. For example, the control unit 130 performs fine-tuning when the evaluation result of the performance evaluation is equal to or greater than the performance evaluation result threshold (i.e., the evaluation result is good), and does not perform fine-tuning when the evaluation result is less than the performance evaluation result threshold (i.e., the evaluation result is poor). This is because, for example, if fine-tuning is performed on the trained AI/ML model when the evaluation result is less than the performance evaluation result threshold, the impact on the trained AI/ML model (step S115) that is below the performance threshold will be greater than a certain level, and it is predicted that fine-tuning will not improve the trained AI/ML model, or that over-training will occur, resulting in a decrease in the overall performance of the AI/ML model. In such cases, re-training may be preferable to fine-tuning. On the other hand, if the evaluation result is equal to or greater than the performance evaluation result threshold, even if fine-tuning is performed on the trained AI/ML model, the impact can be kept below a certain level, and it is predicted that performance will be improved. The following explanation assumes that the control unit 130 decides to perform fine-tuning in step S118.

(2-2) When monitoring is performed on the UE 100 side and model control is performed on the gNB 200 side (case (Z2))

In the case of (Z2), steps S120 to S127 shown in Figure 18 are executed.

That is, in step S120, the control unit 130 of the UE 100 monitors the inference of the trained AI/ML model.

In step S121, the transmitter 120 of the UE 100 transmits the monitoring results to the gNB 200. The receiver 220 of the gNB 200 receives the monitoring results.

In step S122, the control unit 230 of the gNB 200 detects, based on the monitoring results, that the performance of the trained AI/ML model is below a performance threshold (i.e., performance is poor). The control unit 230 then decides whether to cause the UE 100 to perform performance evaluation. Specifically, the control unit 230 decides whether to perform performance evaluation based on the performance evaluation execution time received in step S105 (Figure 16). The decision on whether to perform performance evaluation may be the same as that in step S116 (Figure 17). That is, the control unit 230 may decide to perform performance evaluation when the performance evaluation execution time received in step S105 is below the performance evaluation time threshold, and may decide not to perform performance evaluation when the performance evaluation execution time received in step S105 exceeds the performance evaluation time threshold. The following description is based on the assumption that the control unit 230 decides to perform performance evaluation based on the performance evaluation execution time.

In step S123, the transmitter 210 of the gNB 200 transmits a performance evaluation execution instruction (or performance evaluation execution instruction) to the UE 100, instructing the UE 100 to perform a performance evaluation. The performance evaluation execution instruction may include identification information (e.g., a model ID or function name) of the trained AI/ML model for which performance evaluation is to be performed. The performance evaluation execution instruction may also be instruction information instructing the UE 100 not to perform performance evaluation. The receiver 110 of the UE 100 receives the performance evaluation execution instruction.

In step S124, the control unit 130 of the UE 100 performs performance evaluation in response to receiving the performance evaluation execution instruction. The performance evaluation itself may be performed in the same manner as in step S117 (Figure 17).

In step S125, the transmitter 120 of the UE 100 transmits the performance evaluation results to the gNB 200. The receiver 220 of the gNB 200 receives the evaluation results.

In step S126, the control unit 230 of the gNB 200 determines whether or not to perform fine-tuning in the UE 100 based on the evaluation result of the performance evaluation received in step S125. The decision on whether or not to perform fine-tuning may be the same as that in step S118 (Figure 17). That is, the control unit 230 may decide to perform fine-tuning when the evaluation result received in step S125 is equal to or greater than the performance evaluation result threshold, and may decide not to perform fine-tuning when the evaluation result received in step S125 is less than the performance evaluation result threshold. Alternatively, the control unit 230 may decide to perform re-learning if fine-tuning is not to be performed. In the following, it is assumed that the control unit 230 decides to perform fine-tuning based on the evaluation result.

In step S127, the transmitter 210 of the gNB 200 transmits a fine-tuning execution instruction to the UE 100 to instruct the UE 100 to perform fine-tuning. The fine-tuning execution instruction may include identification information (e.g., a model ID or a function name) of the trained AI/ML model for which fine-tuning is to be performed. The fine-tuning execution instruction may also be instruction information instructing the UE 100 not to perform fine-tuning. The receiver 110 of the UE 100 receives the fine-tuning execution instruction. The controller 130 of the UE 100 may decide to perform fine-tuning in response to receiving the fine-tuning execution instruction.

(2-3) When monitoring and model control are performed on the gNB200 side (case (Z3))

In the case of (Z3), steps S130 to S138 shown in Figure 19 are executed.

That is, in step S130, the control unit 130 of the UE 100 calculates a performance metric for the inference of the trained AI/ML model (step S106).

In step S131, the transmitter 120 of the UE 100 transmits the performance metric to the gNB 200. The receiver 220 of the gNB 200 receives the performance metric.

In step S132, the control unit 230 of the gNB 200 uses the performance metrics to monitor the inference being performed by the UE 100 (step S106). The control unit 230 detects that the monitoring results are below the performance threshold, i.e., that the performance of the trained AI/ML model is poor.

In step S133, the control unit 230 of the gNB 200 determines whether or not to perform performance evaluation. This determination may be the same as that in step S122. That is, the control unit 230 may determine whether or not to perform performance evaluation using the performance evaluation execution time and the performance evaluation time threshold.

The subsequent steps S134 to S138 are the same as those in the example of case (Z2) (steps S123 to S127). In response to receiving the fine-tuning execution instruction (step S138), the control unit 130 of the UE 100 may decide to execute the fine-tuning.

(2-4) Fine Adjustment Then, in step S140 of Fig. 20, fine adjustment is performed. The fine adjustment itself may be performed on the UE 100 side (steps S141 and S142) as in the first embodiment, or may be performed on the OTT server 500 side (steps S143 to S145). Each operation (steps S141 to S145) is also the same as each operation (steps S85 to S89) in the first embodiment.

(Another Operation Example 1 According to the Second Embodiment)
In the second embodiment, an example has been described in which the model transfer (step S90), the transmission of the parameters for calculating the performance evaluation time (step S91), the transmission of the data set for calculating the performance evaluation time (step S94), and the transmission of the current index (step S99) are performed using separate messages. For example, the model transfer and at least one of the parameters, the data set, and the index may be performed using a single message.

(Another Operation Example 2 According to the Second Embodiment)
In the second embodiment, an example of a UE-side model has been described, but this is not limited thereto and can also be implemented in the case of a two-side model. That is, the inference performed on the UE 100 side can be implemented by applying the operation examples shown in FIGS. 16 to 20. Furthermore, for the inference performed on the gNB 200 side, the control unit 230 calculates the performance evaluation execution time for the trained AI/ML model on the gNB 200 side, performs monitoring, and when the monitoring result is equal to or less than the performance threshold, determines whether the performance evaluation execution condition is met (or whether the performance evaluation execution time is equal to or less than the performance evaluation time threshold). Then, the control unit 230 performs performance evaluation according to the determination result, and, for example, as in step S126, compares the evaluation result of the performance evaluation with the performance evaluation result threshold and determines whether to perform fine-tuning. Furthermore, the second embodiment can also be applied to a network-side model. For example, in FIGS. 16 to 20, the network-side model can also be implemented by replacing "UE 100" with "network side." In this case, model transfer (step S90), parameter transmission (step S91), data set transmission (step S94), and current indicator transmission (step S99) do not need to be performed, and the network device 400 itself may hold the trained AI/ML model, parameters, data set, and current indicator.

(Another Operation Example 3 According to Second Embodiment)
In the second embodiment, an example has been described in which the UE 100 transmits the performance evaluation execution time to the gNB 200, but this is not limiting. For example, when monitoring and model control are performed on the UE 100 side (FIG. 17), the control unit 130 of the UE 100 may store the performance evaluation execution time in a memory or the like without transmitting it, and after monitoring (step S120), may determine whether or not to perform performance evaluation based on the performance evaluation execution conditions (step S114) received from the gNB 200.

(Another Operation Example 4 According to the Second Embodiment)
The transmitter 120 of the UE 100 may transmit the performance evaluation execution time together with the monitoring result in step S121, instead of transmitting the performance evaluation execution time in step S105. The UE 100 can transmit the monitoring result and the performance evaluation execution time by one message.

(Another Operation Example 5 According to the Second Embodiment)
In the second embodiment, beam management is used as an example of a use case, but the use case is not limited to beam management. For example, the first embodiment can also be applied to CSI feedback improvement (X1.1) or position accuracy improvement (X1.3).

[Third embodiment]
Next, a third embodiment will be described, focusing on the differences from the first embodiment.

In the first embodiment, we described performing fine-tuning based on the fine-tuning execution time (steps S68 to S71). In the third embodiment, we describe an example in which we determine whether to perform fine-tuning based on the number of times it is detected that the inference results of the trained AI/ML model are below the performance threshold (i.e., the inference accuracy is poor).

Specifically, first, the user device (e.g., UE100) receives fine-tuning execution conditions from a network node (e.g., gNB200) that indicate the conditions for performing fine-tuning of the trained AI/ML model. Second, the user device performs inference using the trained AI/ML model. Third, the user device performs fine-tuning. Here, the fine-tuning execution conditions include a performance threshold and a predetermined number of times. Fourth, fine-tuning is performed in response to detecting a predetermined number of times that the inference result of the trained AI/ML model is below the performance threshold. Here, fine-tuning means training the trained AI/ML model using training data with a data volume below a data volume threshold.

In this way, UE100 performs fine-tuning when the inference result of the trained AI/ML model is below the performance threshold, i.e., when it detects a predetermined number of times that the performance of the trained AI/ML model is poor. This eliminates the need for UE100 to determine whether to perform fine-tuning (steps S70 and S71 in FIG. 14) each time it detects that the performance of the trained AI/ML model is poor, thereby reducing the processing load of UE100. Furthermore, for example, UE100 can determine whether to perform fine-tuning in accordance with instructions from gNB200, making it possible to perform fine-tuning appropriately.

(Operation example according to the third embodiment)
Next, an example of operation according to the third embodiment will be described.

FIG. 21 shows an example of operation according to the third embodiment.

As shown in FIG. 21, in step S150, the network device 400 transfers (or transmits) the trained AI/ML model to the UE 100. The receiver 110 of the UE 100 receives the trained AI/ML model.

In step S151, the transmitter 210 of the gNB 200 transmits the fine-tuning execution conditions to the UE 100.

First, the fine-tuning execution conditions include the above-mentioned performance threshold. As in the first embodiment, the performance threshold is, for example, a threshold for determining the performance of the trained AI/ML model. For example, UE100 determines that the performance of the trained AI/ML model is poor (in the third embodiment, the inference accuracy is poor) when the inference result of the trained AI/ML model is below the performance threshold, and determines that the inference result of the trained AI/ML model is good (in the third embodiment, the inference accuracy is good) when the inference result is greater than the performance threshold. Here, the inference result may include, for example, the CPU usage rate or memory usage rate when the inference was performed, the amount of inference data used when the inference was performed, or the amount of inference result data. Furthermore, the performance threshold may be a threshold corresponding to each indicator of the inference result.

Second, the fine-tuning execution conditions include a predetermined number of times representing a threshold for the number of times. UE100 executes fine-tuning when it detects that the inference result of the trained AI/ML model is below the performance threshold a predetermined number of times. The predetermined number of times may be a consecutive number of times or a cumulative number of times. In the case of a cumulative number of times, the fine-tuning execution conditions may also include a period of time.

Third, the fine-tuning execution condition may include a report instruction (or report request) that instructs (or requests) a report on whether or not fine-tuning has been performed. When fine-tuning has been performed, UE 100 transmits information indicating that fine-tuning has been performed to network device 400 in accordance with the report instruction. The report instruction may include a destination indicating to which node the report should be sent. Alternatively, the report instruction may include information instructing transmission of the inference data set used for inference, along with whether or not fine-tuning has been performed.

Fourth, the fine-tuning execution conditions may include the same information as the fine-tuning execution conditions described in the first embodiment (or second embodiment).

In step S152-1, the control unit 130 of the UE 100 performs inference using the trained AI/ML model received in step S150.

In step S153-1, the control unit 130 of the UE 100 evaluates the inference result. Step S153-1 may be the monitoring (step S70) in the first embodiment.

In step S154-1, the control unit 130 of the UE 100 detects that the inference result is below the performance threshold.

The control unit 130 of the UE 100 performs steps S152-1 through S154-1 n times (n is a natural number) (steps S152-n through S154-n).

In other words, in step S155, the control unit 130 of the UE 100 detects that the inference result is below the performance threshold a predetermined number of times (e.g., n times).

In step S156, the control unit 130 of UE 100 performs fine-tuning on the trained AI/ML model received in step S150 because the fine-tuning execution conditions are met (the inference result has been detected to be below the performance threshold a predetermined number of times).

In step S157, the transmitter 120 of the UE 100 transmits information indicating that fine-tuning has been performed to the gNB 200. The transmitter 120 may transmit the information in accordance with a reporting instruction included in the fine-tuning execution conditions.

In step S158, the transmitter 120 of the UE 100 transmits to the network device 400 information indicating that fine-tuning has been performed and a data set of the inference data used in the inference. The transmitter 120 may transmit this information in accordance with a reporting instruction included in the fine-tuning execution conditions.

(Another Operation Example 1 According to the Third Embodiment)
In the third embodiment, an example of an operation related to fine-tuning has been described. For example, the third embodiment can apply the performance evaluation described in the second embodiment. That is, in the UE 100, whether or not to perform performance evaluation is determined based on the number of times that the inference result of the trained AI/ML model is detected to be equal to or less than the performance threshold.

Specifically, first, the user equipment (e.g., UE100) receives performance evaluation execution conditions from a network node (e.g., gNB200) that indicate the conditions for performing performance evaluation of the trained AI/ML model. Second, the user equipment performs inference using the trained AI/ML model. Third, the user equipment performs performance evaluation in response to detecting a predetermined number of times that the inference result of the trained AI/ML model is below the performance threshold. Here, the performance evaluation execution conditions include the performance threshold and the predetermined number of times.

In this way, UE100 performs performance evaluation when it detects that the inference result of the trained AI/ML model is below the performance threshold, i.e., when it detects a predetermined number of times that the performance of the trained AI/ML model is poor. This eliminates the need for UE100 to determine whether or not to perform performance evaluation each time it detects that the performance of the trained AI/ML model is poor (steps S115 and S116 in FIG. 17), thereby reducing the processing load of UE100. Furthermore, for example, UE100 can determine whether or not to perform performance evaluation in accordance with instructions from gNB200, making it possible to appropriately perform performance evaluation.

FIG. 22 shows another example of operation according to the third embodiment.

After performing model transfer (step S150 in FIG. 21), the transmitter 210 of gNB 200 transmits the performance evaluation execution conditions to UE 100 in step S160. The receiver 110 of UE 100 receives the performance evaluation execution conditions.

First, the performance evaluation execution conditions include a performance threshold. As in the first embodiment, the performance threshold is, for example, a threshold for determining the inference performance of the trained AI/ML model. The performance determination method may be the same as in the third embodiment.

Second, the performance evaluation execution conditions include a predetermined number of times that represents a threshold for the number of times. UE100 executes performance evaluation, for example, when it detects that the inference result of the trained AI/ML model is below the performance threshold a predetermined number of times. The predetermined number of times may be a consecutive number of times or a cumulative number of times. In the case of a cumulative number of times, the performance evaluation conditions may also include a period of time.

Third, the performance evaluation execution conditions may include a reporting instruction to report whether or not performance evaluation has been performed. The reporting instruction may include information to report the results of the performance evaluation as the report content.

Fourth, the performance evaluation execution conditions may include the evaluation method (e.g., NMSE) to be used when performing the performance evaluation.

Fifth, the performance evaluation execution conditions may include identification information (e.g., model ID or function name) of the trained AI/ML model for which performance evaluation is to be performed.

Steps S161-1 to S164 are the same as steps S152-1 to S155 in the third embodiment.

In step S164, the control unit 130 of the UE 100 detects that the inference result is below the performance threshold a predetermined number of times (e.g., n times).

In step S165, the control unit 130 of the UE 100 performs performance evaluation to satisfy the performance evaluation execution conditions.

In step S166, the transmitter 120 of the UE 100 transmits to the gNB 200 information indicating that the inference result is below the performance threshold (i.e., the accuracy of the inference is poor) and the evaluation result. The transmitter 120 may transmit this information in accordance with a reporting instruction included in the performance evaluation execution conditions. The receiver 220 of the gNB 200 receives the information indicating that the inference result is below the performance threshold and the evaluation result.

In step S167, the control unit 230 of gNB200 determines to perform fine-tuning based on the fact that the inference result is below the performance threshold and the evaluation result, and the transmission unit 210 transmits a fine-tuning execution instruction to UE100 instructing it to perform fine-tuning. Alternatively, the control unit 230 may determine to perform re-learning instead of fine-tuning based on the evaluation result, and the transmission unit 210 may transmit a re-learning execution instruction to UE100 indicating that re-learning will be performed. This is because, if the evaluation result is worse than a certain level, it may be better for UE100 to perform re-learning rather than fine-tuning.

[Fourth embodiment]
Next, a fourth embodiment will be described.

In the fourth embodiment, an example will be described in which the UE100 reports the recorded log to the gNB200. Specifically, first, the user equipment (e.g., the UE100) receives reporting conditions from the network node (e.g., the gNB200) indicating the conditions for reporting information related to fine-tuning (and/or information related to performance evaluation). Second, the user equipment transmits information related to fine-tuning (and/or information related to performance evaluation) to the network node in accordance with the reporting conditions.

In this way, in the fourth embodiment, UE100 can appropriately report information regarding fine-tuning (and/or information regarding performance evaluation) to gNB200 in accordance with the reporting conditions received from gNB200. Furthermore, gNB200 can also change the execution conditions for fine-tuning and/or the execution conditions for performance evaluation based on the reported information. This makes it possible, for example, to appropriately execute use cases using an AI/ML model in mobile communication system 1.

(Operation example according to the fourth embodiment)
FIG. 23 is a diagram illustrating an example of operation according to the fourth embodiment.

As shown in FIG. 23, in step S170, the network device 400 transfers (or transmits) the trained AI/ML model to the UE 100. The receiver 110 of the UE 100 receives the trained AI/ML model.

In step S171, gNB200 transmits the reporting conditions to UE100. The receiver 110 of UE100 receives the reporting conditions.

First, the reporting conditions may include information indicating the type of information to be reported. Examples of information indicating the type include information indicating whether information related to fine-tuning should be reported, information related to performance evaluation, or information related to both fine-tuning and performance evaluation should be reported. The transmitter 210 of the gNB 200 may transmit information indicating the type depending on whether the UE 100 is performing fine-tuning, performance evaluation, or both.

Secondly, the reporting conditions may include report target information indicating information to be reported. The report target information may include at least any of the following. Note that the report target information may represent information to be transmitted as log information by the UE 100. The report target information may also be information related to performance evaluation. The report target information may also be information related to fine-tuning.

(U1) Whether or not performance evaluation and/or fine-tuning is performed,

(U2) Time required for performance evaluation and/or fine-tuning,

(U3) Datasets used for performance evaluation and/or fine-tuning,

(U4) Factors (or causes) for performance evaluation and/or fine-tuning,

(U5) Timestamp,

(U6) Location information,

(U7) Identification information of the trained AI/ML model used for performance evaluation and/or fine-tuning (e.g., model name, model ID, or function name),

(U8) Evaluation method used for performance evaluation Third, the reporting conditions may include information indicating that a log is to be transmitted in response to a log transmission request from gNB200.

Fourth, the reporting conditions may include log transmission conditions that indicate the conditions for sending a log. The log transmission condition may be that at least one of the pieces of information to be reported has been acquired. The log transmission condition may also be that the data volume of at least one of the pieces of information to be reported has reached or exceeded a data volume threshold. In the latter case, the data volume threshold may be included in the log transmission condition.

Fifth, the reporting conditions may include information indicating whether the report should be made when UE100 is in the RRC connected state, or when UE100 is in the RRC idle state or the RRC inactive state. In the former case, the reporting conditions may be included in the reporting configuration (MeansConfig) of an individual RRC message (such as an RRCReconfiguration message or an RRCResume message) and transmitted as an immediate MDT (Minimalization of Drive Tests). In the latter case, the reporting conditions may be included in a logged measurement configuration (LoggedMeasurementConfiguration) message and transmitted as a logged MDT.

In step S172, the control unit 130 of the UE 100 performs inference on the trained AI/ML model received in step S170.

In step S173, the control unit 130 of the UE 100 performs performance evaluation based on the inference results.

In step S174, the control unit 130 of the UE 100 performs fine-tuning on the trained AI/ML model.

In step S175, the control unit 130 of the UE 100 records in memory the log used during performance evaluation and fine-tuning. The control unit 130 may record the log in memory in accordance with the report target information included in the reporting conditions.

In step S176, the control unit 130 of the UE 100 detects that the log transmission conditions are met.

In step S177, the transmitter 210 of the gNB 200 transmits a log transmission request to the UE 100. The receiver 110 of the UE 100 receives the log transmission request.

In step S178, in response to receiving the log transmission request, the transmitter 120 of the UE 100 transmits the log recorded in step S175 to the gNB 200. The transmitter 120 may also transmit the log together with a measurement report (MeasurementReport).

Fifth Embodiment
Next, a fifth embodiment will be described.

The fifth embodiment is an embodiment that combines the first embodiment (the fine-tuning embodiment) and the second embodiment (the performance evaluation embodiment). That is, the fifth embodiment is an example in which the UE 100 determines whether to perform fine-tuning based on the fine-tuning execution time, the performance evaluation execution time, and the evaluation results of the performance evaluation.

Specifically, a user device (e.g., UE100) performs fine-tuning of a first trained AI/ML model (e.g., a trained AI/ML model on the UE100 side) based on a first time required to perform fine-tuning of the first trained AI/ML model, a second time required to perform a performance evaluation of the first trained AI/ML model, and the evaluation results of the performance evaluation. Here, fine-tuning means training the first trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

In this way, in the fifth embodiment, whether or not to perform fine-tuning is determined based on the first and second times and the evaluation results of the performance evaluation, making it possible to perform fine-tuning in less time than when fine-tuning is performed without considering these parameters. Also, in the fifth embodiment, for example, fine-tuning may be performed when the first time is equal to or less than the fine-tuning execution time threshold, the second time is also equal to or less than the performance evaluation time threshold, and the evaluation result is equal to or greater than the performance threshold, and fine-tuning may not be performed under other conditions. Therefore, in the mobile communication system 1, fine-tuning is performed under certain judgments, making it possible to appropriately perform fine-tuning on the trained AI/ML model.

(Operation example according to the fifth embodiment)
Next, an example of operation according to the fifth embodiment will be described.

FIGS. 24 to 26 are diagrams showing an example of operation according to the fifth embodiment. Note that the example of operation shown in FIG. 24 to 26 is an example of CSI compression, which is a sub-use case of CSI feedback, and represents an example of a two-sided model.

For example, CSI compression is performed by UE100 transmitting compressed CSI to gNB200, which then reconstructs the pre-compressed CSI. The AI/ML method is used for CSI compression and CSI reconstruction, and a two-sided model is constructed by an inference unit on the UE100 side that performs CSI compression and an inference unit on the UE100 side that performs CSI reconstruction. The CSI to be compressed is, for example, CSI feedback information generated by legacy processing by CSI generation unit 131.

As shown in FIG. 24, for a two-sided model, the network device 400 transmits a trained AI/ML model for the UE 100 (e.g., a first trained AI/ML model) to the UE 100 (step S180), and transmits a trained AI/ML model for the gNB 200 (e.g., a second trained AI/ML model) to the gNB 200 (step S181).

As in the first embodiment, the network device 400 transmits a parameter for calculating the fine-tuning execution time (e.g., the first parameter) to the UE 100 (step S182). The network device 400 also transmits a parameter for calculating the fine-tuning execution time (e.g., the third parameter) to the gNB 200 in order to calculate the fine-tuning execution time at the gNB 200 (step S185).

The control unit 130 of the UE 100 calculates the fine-tuning execution time (e.g., the first time) using the fine-tuning execution time calculation parameters (step S183), and the transmission unit 120 of the UE 100 transmits the fine-tuning execution time on the UE 100 side to the gNB 200 (step S184). Meanwhile, the control unit 230 of the gNB 200 also calculates the fine-tuning execution time (e.g., the third time) using the fine-tuning execution time calculation parameters (step S186).

As in the second embodiment, the network device 400 transmits parameters for calculating the performance evaluation time (e.g., second parameters), a dataset for calculating the performance evaluation time, and the current indicators to the UE 100 (steps S187 to S189). The control unit 130 of the UE 100 uses these received parameters to calculate the performance evaluation execution time (e.g., second time) for the trained AI/ML model on the UE 100 side (step S190). The transmission unit 120 of the UE 100 transmits the performance evaluation execution time on the UE 100 side to the gNB 200 (step S191).

Meanwhile, since gNB200 also performs inference on the trained AI/ML model for gNB200 and calculates the performance evaluation execution time, network device 400 transmits parameters for calculating the performance evaluation time (e.g., the fourth parameter), a dataset for calculating the performance evaluation time, and the current indicators to gNB200 (steps S192 to S194). The control unit 230 of gNB200 uses these received parameters to calculate the performance evaluation execution time (e.g., the fourth time) for the trained AI/ML model on the gNB200 side (step S195).

In step S200 (Figure 25), inference is performed in the mobile communication system 1. Because it is a two-sided model, inference is performed in the UE 100 and the gNB 200 (steps S202 and S204). In the example of Figure 25, the control unit 130 of the UE 100 generates inference data to be used in the trained AI/ML model on the UE side (step S201) and performs inference using the inference data (step S202). Then, the transmission unit 120 of the UE 100 transmits the inference result data to the gNB 200 as inference data on the gNB 200 side (step S203). The control unit 230 of the gNB 200 inputs the inference result data as inference data for the trained AI/ML model on the gNB 200 side and performs inference (step S204).

In step S205, monitoring is performed in the mobile communication system 1. Because of the two-sided model, the control unit 130 of the UE 100 monitors the inference on the UE 100 side (step S202) (step S206), and the control unit 230 of the gNB 200 monitors the inference on the gNB 200 side (step S204) (step S208). The transmission unit 120 of the UE 100 transmits its own monitoring results to the gNB 200.

In step S210 (Figure 26), management is performed in the mobile communication system 1. In the fifth embodiment, there is fine-tuning of the trained AI/ML model on the UE100 side, and fine-tuning of the trained AI/ML model on the gNB200 side. However, with regard to the fine-tuning performed on the UE100 side, there are cases where the gNB200 makes the fine-tuning decision, and cases where the UE100 itself makes the fine-tuning decision.

(V1) When gNB200 determines fine-tuning on the UE100 side When gNB200 determines fine-tuning on the UE100 side, gNB200 performs steps S211 to S213.

That is, in step S211, the control unit 230 of the gNB 200 determines whether or not to cause the UE 100 to perform performance evaluation. The determination itself may be the same as step S122 (Figure 18) of the second embodiment. That is, based on the monitoring results received in step S207, the control unit 230 detects that the performance of the trained AI/ML model on the UE 100 side is below the performance threshold. Then, the control unit 230 determines whether or not to cause the UE 100 to perform performance evaluation based on the performance evaluation execution time and performance evaluation time threshold received in step S191.

In step S212, the control unit 230 of the gNB 200 determines whether or not to perform fine-tuning. The control unit 230 performs fine-tuning when the performance evaluation execution time is equal to or less than the performance evaluation time threshold (i.e., to perform performance evaluation), the fine-tuning execution time is equal to or less than the fine-tuning execution time threshold, and the evaluation result from the performance evaluation is equal to or greater than the performance evaluation result threshold (i.e., the evaluation result is good). In this way, the control unit 230 combines the performance evaluation execution time, the fine-tuning execution time, and the evaluation result from the performance evaluation to determine whether or not to perform fine-tuning.

Note that the evaluation results from the performance evaluation are evaluation results for the trained AI/ML model on the UE100 side. Therefore, the control unit 130 of the UE100 evaluates the trained AI/ML model, and the transmission unit 120 of the UE100 transmits the evaluation results to the gNB200. This allows the gNB200 to obtain the evaluation results of the trained AI/ML model on the UE100 side. The control unit 230 of the gNB200 can then use the evaluation results to determine whether or not to perform fine-tuning.

In the following, it is assumed that the control unit 230 has determined to perform fine adjustment.

In step S213, the transmitter 210 of the gNB 200 transmits a fine-tuning execution instruction to the UE 100 to instruct the UE 100 to perform fine-tuning. The receiver 110 of the UE 100 receives the fine-tuning execution instruction, and the controller 130 of the UE 100 decides to perform fine-tuning.

(V2) When the UE 100 Determines Fine Adjustment on the UE 100 Side by Itself When the UE 100 determines fine adjustment on the UE 100 side by itself, the UE 100 performs step S214 and step S215.

In other words, in step S214, the control unit 130 of the UE 100 determines whether or not to perform performance evaluation. The determination of whether or not to perform performance evaluation may be the same as step S211. That is, as a result of monitoring the trained AI/ML model on the UE side (step S206), the control unit 130 detects that the performance of the model is below the performance threshold, and compares the performance evaluation execution time calculated in step S190 with the performance evaluation time threshold to determine whether or not to perform performance evaluation. Here, the following explanation will be given assuming that the control unit 130 has determined to perform performance evaluation.

In step S215, the control unit 130 of the UE 100 determines whether or not to perform fine-tuning. The determination of whether or not to perform fine-tuning may itself be the same as in step S212. That is, the control unit 130 may perform fine-tuning when the performance evaluation execution time is equal to or less than the performance evaluation time threshold (i.e., performance evaluation is to be performed), the fine-tuning execution time is equal to or less than the fine-tuning execution time threshold, and the evaluation result from the performance evaluation is equal to or greater than the performance evaluation result threshold (i.e., the evaluation result is good). In the following description, it is assumed that the control unit 130 determines to perform fine-tuning and decides to perform fine-tuning.

The above describes management (step S210). Note that the control unit 230 of gNB200 may also determine whether or not to perform fine-tuning on its own trained AI/ML model. In this case, when the control unit 230 monitors the trained AI/ML model on the gNB200 side (step S208) and detects that the performance of the trained AI/ML model is below the performance threshold (i.e., poor performance), it determines whether or not to perform fine-tuning. That is, the control unit 230 may determine whether or not to perform fine-tuning on the trained AI/ML model based on the fine-tuning execution time (e.g., the third time) calculated in step S186, the performance evaluation execution time (e.g., the fourth time) calculated in step S195, and the evaluation results of the performance evaluation of the trained AI/ML model on the gNB200 side. Specifically, the control unit 230 performs fine-tuning on the trained AI/ML model on the gNB 200 side when the performance evaluation execution time is equal to or less than the performance evaluation time threshold, the fine-tuning execution time is equal to or less than the fine-tuning execution time threshold, and the evaluation result from the performance evaluation is equal to or greater than the performance evaluation result threshold.

In step S216, fine-tuning is performed in the mobile communication system 1. That is, on the UE100 side, the control unit 130 generates learning data (step S217) and uses the generated learning data to perform fine-tuning on the trained AI/ML model on the UE100 side (step S218). Meanwhile, on the gNB200 side, the control unit 230 generates learning data (step S219) and uses the generated learning data to perform fine-tuning on the trained AI/ML model on the gNB200 side (step S220).

(Another Operation Example 1 According to Fifth Embodiment)
Although the fifth embodiment has been described in the case of a two-side model, this is not limiting. For example, the fifth embodiment can also be applied to a UE-side model. In this case, in FIGS. 24 to 26, some of the processing performed on the gNB 200 side (such as steps S186, S195, S219, and S220) can be prevented from being performed. The fifth embodiment can also be applied to a network-side model. For example, in FIGS. 24 to 26, the network-side model can also be implemented by replacing "UE 100" with "network side." In this case, model transfer (step S180), parameter transmission (steps S182 and S187), data set transmission (step S188), and current indicator transmission (step S189) do not need to be performed, and the network device 400 itself may retain the trained AI/ML model, parameters, data set, and current indicator.

Furthermore, in the fifth embodiment, a use case for CSI compression was described, but it can also be applied to use cases other than CSI compression.

(Another Operation Example 2 According to Fifth Embodiment)
The fifth embodiment can be applied to the third embodiment. That is, in the mobile communication system 1, it may be determined whether to perform fine-tuning based on the number of times that the inference result (or monitoring result) of the trained AI/ML model is detected to be equal to or lower than the performance threshold (i.e., the inference accuracy is poor). Specifically, steps S151 to S158 shown in FIG. 21 may be executed instead of steps S211 to S215 in FIG. 26.

Furthermore, the fifth embodiment can be applied to the fourth embodiment. That is, gNB200 may transmit reporting conditions to UE100, and UE100 may transmit the recorded log to gNB200 in accordance with the reporting conditions. Specifically, step S171 (transmission of reporting conditions) in FIG. 23 may be executed between step S180 (model transfer) in FIG. 24 and step S200 (inference) in FIG. 25, and step S175 (log recording) and subsequent steps in FIG. 23 may be executed after step S216 (fine-tuning) in FIG. 26.

Note that other operational examples according to the first embodiment and other operational examples according to the second embodiment can also be applied to the fifth embodiment.

[Other embodiments]
In the first embodiment described above, supervised learning has been mainly described, but the present invention is not limited to this. For example, unsupervised learning or reinforcement learning may be applied to the first embodiment.

The above-mentioned operational flows do not necessarily have to be implemented separately and independently, but can also be implemented by combining two or more operational flows. For example, some steps in one operational flow may be added to another operational flow, or some steps in one operational flow may be replaced with some steps in another operational flow. In each flow, it is not necessary to execute all steps; only some steps may be executed.

In the above-described embodiments and examples, an example was described in which the base station is an NR base station (gNB), but the base station may also be an LTE base station (eNB) or a 6G base station. Furthermore, the base station may also be a relay node such as an IAB (Integrated Access and Backhaul) node. The base station may also be a DU of an IAB node. Furthermore, UE100 may also be an MT (Mobile Termination) of an IAB node.

In other words, UE100 may be a terminal function unit (a type of communication module) that allows a base station to control a repeater that relays signals. Such a terminal function unit is referred to as an MT. In addition to IAB-MT, examples of MT include NCR (Network Controlled Repeater)-MT and RIS (Reconfigurable Intelligent Surface)-MT.

Furthermore, the term "network node" primarily refers to a base station, but may also refer to a core network device or part of a base station (CU, DU, or RU). A network node may also be composed of a combination of at least part of a core network device and at least part of a base station.

A program may be provided that causes a computer to execute each process performed by UE100, gNB200, or network device 400. The program may be recorded on a computer-readable medium. Using the computer-readable medium, the program can be installed on a computer. Here, the computer-readable medium on which the program is recorded may be a non-transitory recording medium. The non-transitory recording medium is not particularly limited, but may be, for example, a CD-ROM and/or DVD-ROM. In addition, circuits that execute each process performed by UE100, gNB200, or network device 400 may be integrated, and at least a portion of UE100, gNB200, or network device 400 may be configured as a semiconductor integrated circuit (chipset, SoC: System on a chip).

As used in this disclosure, the terms "based on" and "depending on/in response to" do not mean "based only on" or "depending only on," unless expressly stated otherwise. The term "based on" means both "based only on" and "based at least in part on." Similarly, the term "depending on" means both "depending only on" and "depending at least in part on." The terms "include," "comprise," and variations thereof do not mean including only the listed items, but may mean including only the listed items or including additional items in addition to the listed items. Additionally, as used in this disclosure, the term "or" is not intended to mean an exclusive or. Furthermore, any reference to elements using designations such as "first," "second," etc., as used in this disclosure does not generally limit the quantity or order of those elements. These designations may be used herein as a convenient method of distinguishing between two or more elements. Thus, a reference to a first and a second element does not imply that only two elements may be employed therein, or that the first element must precede the second element in some way. In this disclosure, where articles are added by translation, such as a, an, and the in English, these articles shall include the plural unless the context clearly indicates otherwise.

The functions performed by UE 100 or base station 200 (network node) may be implemented in circuitry or processing circuitry, including general-purpose processors, application-specific processors, integrated circuits, ASICs (Application Specific Integrated Circuits), CPUs (Central Processing Units), conventional circuits, and/or combinations thereof, programmed to perform the described functions. A processor includes transistors and/or other circuits and is considered to be circuitry or processing circuitry. A processor may also be a programmed processor that executes programs stored in memory.

In this specification, a circuit, unit, or means refers to hardware that is programmed to realize or executes the described functions. The hardware may be any hardware disclosed in this specification or any hardware known to be programmed to realize or execute the described functions.

If the hardware is a processor, which is considered to be a type of circuitry, the circuitry, means, or unit is the combination of the hardware and the software used to configure the hardware and/or processor.

Although one embodiment has been described in detail above with reference to the drawings, the specific configuration is not limited to that described above, and various design changes can be made without departing from the spirit of the invention. Furthermore, the various embodiments, operational examples, and processes can be combined as appropriate within the scope of consistent principles.

This application claims priority from Japanese Patent Application No. 2024-018853 (filed February 9, 2024), the entire contents of which are incorporated herein by reference.

(Additional Note)
The above can be summarized as follows:

(Appendix 1)
A communication control method in a user device of a mobile communication system, comprising:
The user device performs the fine-tuning on the trained AI/ML model based on a fine-tuning execution time indicating a time required to perform the fine-tuning of the trained AI/ML model;
The fine-tuning is to train the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

(Appendix 2)
The fine-tuning is to train the trained AI/ML model using the training data, and further to perform the training on layers that are equal to or less than a layer threshold and include a final layer among multiple layers up to the final layer that constitute the trained AI/ML model.

(Appendix 3)
the user device calculating the fine-tuning execution time;
The communication control method according to claim 1 or 2, further comprising the step of the user equipment transmitting the fine-tuning execution time to a network node.

(Appendix 4)
The method further comprises receiving, by the user device, the trained AI/ML model and parameters for calculating the fine-tuning execution time from the network node;
The communication control method according to any one of Supplementary Note 1 to Supplementary Note 3, wherein the calculating step includes a step in which the user device calculates the fine-tuning execution time using the parameters.

(Appendix 5)
The method further comprises receiving, by the user equipment, a fine-tuning execution condition indicating a condition for performing the fine-tuning from the network node;
5. The communication control method according to claim 1, wherein the performing step comprises the user device performing the fine-tuning based on the fine-tuning execution condition.

(Appendix 6)
The method further comprises receiving, by the user equipment, a fine-tuning execution instruction from the network node, the fine-tuning execution instruction instructing the user equipment to perform the fine-tuning;
6. The communication control method according to claim 1, wherein the performing step comprises the user device performing the fine-tuning in accordance with the fine-tuning execution instruction.

(Appendix 7)
receiving, by the user equipment, a reporting condition from a network node indicating a condition for reporting information related to the fine-tuning;
The communication control method according to any one of Supplementary Note 1 to Supplementary Note 6, further comprising the step of the user equipment transmitting information relating to the fine-tuning to the network node in accordance with the reporting condition.

(Appendix 8)
A communication control method in a network node of a mobile communication system, comprising:
a step in which the network node determines fine-tuning execution conditions indicating conditions for performing the fine-tuning on the trained AI/ML model in the user device based on a fine-tuning execution time indicating the time required to perform the fine-tuning of the trained AI/ML model;
the network node transmitting the fine-tuning execution condition to the user equipment;
The fine-tuning is to train the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

(Appendix 9)
9. The communication control method according to any one of Supplementary Note 1 to Supplementary Note 8, further comprising the step of the network node receiving the fine-tuning execution time from the user equipment.

(Appendix 10)
The communication control method according to any one of Supplementary Note 1 to Supplementary Note 9, further comprising the step of the network node transmitting to the user equipment a reporting condition indicating a condition for reporting information related to the fine-tuning.

(Appendix 11)
A communication control method for a network node in a mobile communication system, comprising:
The network node determines to perform fine-tuning of the trained AI/ML model in the user device based on a fine-tuning execution time indicating a time required to perform fine-tuning of the trained AI/ML model;
the network node transmitting a fine-tuning execution instruction to the user equipment instructing the user equipment to perform the fine-tuning;
The fine-tuning is to train the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

(Appendix 12)
12. The communication control method according to any one of Supplementary Note 1 to Supplementary Note 11, further comprising the step of the network node receiving the fine-tuning execution time from the user equipment.

(Appendix 13)
The communication control method according to any one of Supplementary Note 1 to Supplementary Note 12, further comprising the step of the network node transmitting to the user equipment a reporting condition indicating a condition for reporting information related to the fine-tuning.

(Appendix 14)
A communication control method in a user device of a mobile communication system, comprising:
The user equipment receives, from a network node, a fine-tuning execution condition indicating a condition for executing fine-tuning of the trained AI/ML model;
The user device performs inference using the trained AI/ML model;
and performing the fine-tuning by the user device;
the fine-tuning execution conditions include a performance threshold and a predetermined number of times;
the performing step includes a step of performing the fine-tuning in response to the user device detecting that the inference result of the trained AI/ML model is equal to or less than the performance threshold a predetermined number of times;
The fine-tuning is to train the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

(Appendix 15)
A communication control method in a network node of a mobile communication system, comprising:
The network node transmits to the user device a fine-tuning execution condition indicating a condition for performing fine-tuning of the trained AI/ML model;
the fine-tuning execution conditions include a performance threshold and a predetermined number of times;
In the user device, the fine-tuning is performed in response to detecting that the inference result of the trained AI/ML model is equal to or less than the performance threshold a predetermined number of times;
The fine-tuning is to train the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.

1: Mobile communication system 20: 5GC (CN)
100:UE
110: Receiving unit 120: Transmitting unit 130: Control unit 200: gNB
210: Transmitter 220: Receiver 230: Controller 400: Network device 500: OTT server

Claims (15)

A communication control method in a user device of a mobile communication system, comprising:
The user device performs the fine-tuning on the trained AI (Artificial Intelligence)/ML (Machine Learning) model based on a fine-tuning execution time indicating a time required to perform the fine-tuning of the trained AI/ML model;
The fine-tuning is to train the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold. 2. The communication control method according to claim 1, wherein the fine-tuning involves training the trained AI/ML model using the training data, and further comprising performing the training on layers including the final layer and below a layer threshold, among multiple layers up to the final layer that constitute the trained AI/ML model. the user device calculating the fine-tuning execution time;
The communication control method according to claim 1 , further comprising: the user equipment transmitting the fine-tuning execution time to a network node. The user device further comprises receiving the trained AI/ML model and parameters for calculating the fine-tuning runtime from the network node;
The communication control method according to claim 3 , wherein the calculating step includes the user device calculating the fine-tuning execution time using the parameters. The method further includes receiving, by the user equipment, a fine-tuning execution condition from the network node, the fine-tuning execution condition indicating a condition for performing the fine-tuning;
The communication control method according to claim 3 , wherein the executing step comprises the user device executing the fine-tuning based on the fine-tuning execution condition. The method further includes receiving, by the user equipment, a fine-tuning execution instruction from the network node, the fine-tuning execution instruction instructing the user equipment to perform the fine-tuning;
The communication control method according to claim 3 , wherein the executing step comprises the user device executing the fine-tuning in accordance with the fine-tuning execution instruction. receiving, by the user equipment, a reporting condition from a network node indicating a condition for reporting information related to the fine-tuning;
The communication control method according to claim 1 , further comprising: the user equipment transmitting information relating to the fine-tuning to the network node in accordance with the reporting condition. A communication control method in a network node of a mobile communication system, comprising:
The network node determines a fine-tuning execution condition indicating a condition for executing the fine-tuning of the trained AI/ML model in the user device based on a fine-tuning execution time indicating a time required to execute the fine-tuning of the trained AI/ML model;
the network node transmitting the fine-tuning execution condition to the user equipment;
The fine-tuning is to train the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold. The communication control method according to claim 8 , further comprising: the network node receiving the fine-tuning execution time from the user equipment. The communication control method according to claim 8 , further comprising the network node transmitting, to the user equipment, a reporting condition indicating a condition for reporting information related to the fine-tuning. A communication control method in a network node of a mobile communication system, comprising:
The network node determines to perform fine-tuning of the trained AI/ML model in the user device based on a fine-tuning execution time indicating a time required to perform fine-tuning of the trained AI/ML model; and
the network node transmitting a fine-tuning execution instruction to the user equipment instructing the user equipment to perform the fine-tuning;
The fine-tuning is to train the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold. The communication control method according to claim 11 , further comprising the network node receiving the fine-tuning execution time from the user equipment. The communication control method according to claim 11 , further comprising: the network node transmitting, to the user equipment, a reporting condition indicating a condition for reporting information related to the fine-tuning. A communication control method in a user device of a mobile communication system, comprising:
The user equipment receives, from a network node, a fine-tuning execution condition indicating a condition for performing fine-tuning of a trained AI (Artificial Intelligence)/ML (Machine Learning) model;
The user device performs inference using the trained AI/ML model;
the user device performing the fine-tuning;
the fine-tuning execution conditions include a performance threshold and a predetermined number of times;
The performing includes performing the fine-tuning in response to the user device detecting that the inference result of the trained AI/ML model is equal to or less than the performance threshold a predetermined number of times;
The fine-tuning is to train the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold. A communication control method in a network node of a mobile communication system, comprising:
The network node transmits, to a user equipment, a fine-tuning execution condition indicating a condition for performing fine-tuning of a trained AI (Artificial Intelligence)/ML (Machine Learning) model;
the fine-tuning execution conditions include a performance threshold and a predetermined number of times;
In the user device, the fine-tuning is performed in response to detecting that the inference result of the trained AI/ML model is equal to or less than the performance threshold a predetermined number of times;
The fine-tuning is to train the trained AI/ML model using training data with a data volume equal to or less than a data volume threshold.