Terminal device for artificial intelligence model handover, base station device, and control method

Abstract

Provided is a terminal device that communicates with a base station device on the basis of a cellular communication standard of the 3rd Generation Partnership Project (3GPP). The terminal device can perform inference for communication using at least one of a plurality of learned models generated by machine learning. On the basis of the fact that a handover to a second cell from a first cell in which the terminal device is located is determined and inference using a first learned model included in the plurality of learned models is being executed in a first communication between a first base station device constituting the first cell and the terminal device, the terminal device notifies a second base station device constituting the second cell of identification information for identifying the first learned model, and receives an instruction, determined by the second base station device on the basis of the notification, regarding execution of inference in a second communication between the second base station device and the terminal device.

Description

Terminal device, base station device, and control method for handover of artificial intelligence model

[0001] The present invention relates to a technology for exchanging information of an artificial intelligence model used between a base station and a terminal in a mobile communication system.

[0002] In the 3rd Generation Partnership Project (3GPP (registered trademark)), utilization of artificial intelligence (AI) and machine learning (ML) for control in mobile communication systems is under study. In inference using AI or ML, an AI / ML model can be used. One example of an AI / ML model is a trained model generated by machine learning using training data. An AI / ML model may also be referred to as an AI model. For example, examples of use cases where an AI model is used in a terminal of a mobile communication system include beam management, positioning, and quality prediction of radio channels. Further, one example of a use case where an AI model is used in both a base station and a terminal of a mobile communication system is compression and restoration of channel state information. Non-Patent Document 1 describes that AI models are used in radio quality estimation, handover determination, and the like, and also describes evaluation methods for AI models.

[0003] 3rd Generation Partnership Project, "TR 38.744 Study on Artificial Intelligence (AI) / Machine Learning (ML) for mobility in NR; (Release 19)"

[0004] In a mobile communication system, due to movement of a terminal or the like, the terminal may perform handover between a first base station and a second base station. At this time, a problem may arise due to a difference between the AI models used before and after the handover is executed. The present invention provides a technique for enabling appropriate selection of an AI model before and after handover execution when a terminal performs handover between a plurality of base stations in a mobile communication system.

[0005] A terminal device according to one aspect of the present invention is a terminal device that communicates with a base station device based on the cellular communication standard of the Third Generation Partnership Project (3GPP), and comprises: an inference means capable of performing inference for the communication using at least one of a plurality of trained models generated by machine learning; a notification means that notifies a second base station device constituting the second cell of specific information identifying the first trained model, based on the fact that it has been decided that the terminal device will perform a handover from a first cell to a second cell in which it is located, and that the inference using the first trained model included in the plurality of trained models has been performed in the first communication between the first base station device constituting the first cell and the terminal device; and a receiving means that receives instructions regarding the execution of the inference in the second communication between the second base station device and the terminal device, which have been determined by the second base station device based on the notification.

[0006] A base station device according to one aspect of the present invention is a base station device that communicates with a terminal device based on the cellular communication standard of the Third Generation Partnership Project (3GPP), wherein in the communication, at least the terminal device is capable of performing inference for the communication using at least one of a plurality of trained models generated by machine learning, and the base station device has receiving means for receiving notification of specific information identifying the first trained model from the terminal device when it is determined that the terminal device will perform a handover from a first cell provided by another base station device in which the terminal device is located to a second cell provided by the base station device, and when a first trained model included in the plurality of trained models is used in a first communication between the other base station device and the terminal device, determining means for determining whether or not to allow the terminal device to continue using the first trained model based on the notification, and transmitting means for transmitting instructions regarding the inference based on the determination to the terminal device.

[0007] According to the present invention, in a mobile communication system, when a terminal performs a handover between multiple base stations, an appropriate AI model is selected both before and after the handover is performed.

[0008] Other features and advantages of this disclosure will become apparent from the following description with reference to the accompanying drawings. In the accompanying drawings, the same or similar components are given the same reference numeral.

[0009] The attached drawings are included in the specification and constitute part thereof, illustrating embodiments of the present disclosure and used together with the description to explain the principles of the present disclosure. Figure 1 is a diagram showing an example configuration of a mobile communication system. Figure 2 is a diagram illustrating the relationship between measured and estimated values ​​of received power at a terminal. Figure 3 is a diagram showing an example hardware configuration of a communication device. Figure 4 is a diagram showing an example functional configuration of a base station. Figure 5 is a diagram showing an example functional configuration of a terminal. Figure 6 is a diagram showing an example sequence of messages exchanged between communication devices. Figure 7 is a diagram showing an example sequence of messages exchanged between communication devices. Figure 8 is a diagram showing an example sequence of messages exchanged between communication devices.

[0010] The embodiments will be described in detail below with reference to the attached drawings. Note that the following embodiments do not limit the invention as defined in the claims, and not all combinations of features described in the embodiments are essential to the invention. Two or more of the features described in the embodiments may be combined in any way. Furthermore, identical or similar configurations will be given the same reference numeral, and redundant descriptions will be omitted.

[0011] (System Configuration) Figure 1 shows an example of the configuration of the mobile communication system according to this embodiment. The mobile communication system of this embodiment is, for example, a cellular communication system compliant with the cellular communication standard of the Third Generation Partnership Project (3GPP®). The cellular communication standard may be Long Term Evolution (LTE), Fifth Generation Mobile Communication System (5G, NR), Beyond 5G, 6G, etc. However, it is not limited to these, and the following discussion can be applied to a mobile communication system compliant with any wireless communication standard. This mobile communication system is composed of, for example, a base station 101, a base station 102, and a terminal 111. Base stations 101 and 102 may be referred to as base station 100 without distinction. Base station 100 exchanges wireless signals with terminal 111 via a wireless medium. The base station 100 includes, for example, a gNB (next Generation Node B), an eNB (evolved Node B), etc. The base station 100 is connected to a core network (not shown). The core network may be, for example, an Evolved Packet System (EPS) or a 5G Core Network (5GC). The terminal 111 is a terminal used by a user and exchanges radio signals with the base station 100 via a radio medium. The terminal 111 may be called User Equipment (UE). The terminal 111 includes, for example, a smartphone, a mobile phone, a personal computer, a tablet terminal, a wearable device, an IoT (Internet of Things) terminal, etc. For example, a base station 100 and a terminal 111 can communicate using the connection established between the terminal 111 and the base station 100 in the cell provided by each base station 100. In Figure 1, base station 101 provides cell 121, and base station 102 provides cell 122. Figure 1 shows an example where there are two base stations 100 and one terminal 111, but in a mobile communication system, there may be three or more base stations 100, and two or more terminals 111. In this case, multiple terminals 111 may be connected to one base station 100, and one terminal 111 may be connected to multiple base stations 100.

[0012] Base stations 101 and 102 can communicate with each other. For example, base stations 101 and 102 can be connected to each other via a network 131 and communicate using the X2 interface or the Xn interface. Network 131 can be a wired or wireless network. The X2 interface can be used for communication between eNBs. The Xn interface can be used for communication between gNBs. For example, when base station 101 has a terminal 111 connected to its device perform a handover from its device to base station 102, it communicates the information necessary for the handover with base station 102 using the X2 interface or the Xn interface.

[0013] Artificial intelligence (AI) and machine learning (ML) inference can be applied to control systems in mobile communication systems. AI is an abbreviation for Artificial Intelligence. ML is an abbreviation for Machine Learning. When performing inference using AI or ML, AI / ML models can be used. An example of an AI / ML model is a trained model generated by machine learning using training data. AI / ML models can also be called AI models. A model used when inference is performed at either the base station 100 or the terminal 111 in a mobile communication system is called a one-sided model. For example, use cases in which a one-sided model is used at the terminal 111 of a mobile communication system include beam management, positioning, and radio channel quality prediction. On the other hand, a model used when inference is performed at both the base station 100 and the terminal 111 in a mobile communication system is called a two-sided model. One example of a use case where a two-sided model is used is the compression and restoration of channel state information. However, the use cases in mobile communication systems where an AI model is used are not limited to these.

[0014] In a mobile communication system, a handover may be performed by terminal 111 when the terminal 111 moves. For example, if terminal 111 is connected to base station 101, base station 101 has terminal 111 measure and report the radio quality of the cell it provides and surrounding cells. Based on the reported radio quality, base station 101 decides whether or not to hand over terminal 111 to another base station. AI model inference may be used in such radio quality measurement and reporting. For example, terminal 111 can predict future radio quality by applying previously measured radio quality values ​​to an AI model. In this case, terminal 111 may report a predicted future radio quality value to base station 101 instead of the current measured radio quality value. With this configuration, base station 101 can decide whether or not to perform a handover based on the predicted future radio quality value. This makes it possible to speed up the timing of the handover to terminal 111. For example, if terminal 111 is moving at high speed, and the decision to perform a handover is based solely on measured wireless quality, it may not be possible to issue instructions at an appropriate timing according to the speed of terminal 111. In such cases, by deciding whether or not to perform a handover based on an estimated future wireless quality, it becomes possible to issue instructions at an appropriate timing according to the speed of terminal 111.

[0015] Figure 2 shows an example of the relationship between the measured received power of the received signals from each base station 100, as measured by terminal 111, and the estimated future received power of the received signals from each base station 100, as estimated by terminal 111, in a situation where terminal 111 connected to base station 101 is moving toward base station 102. The horizontal axis represents the position of the terminal. The vertical axis represents the measured and estimated received power values ​​by the terminal. Solid lines 211 and 221 show the measured received power values ​​received by terminal 111 from base station 101 and base station 102, respectively, at their respective positions. Dotted lines 212 and 222 show the estimated received power values ​​after a predetermined time has elapsed, estimated using the measured received power values ​​received by terminal 111 from base station 101 and base station 102, respectively, at their respective positions. For example, the estimated received power value after a predetermined time has elapsed is the estimated received power value at the position to which terminal 111 has moved after the predetermined time has elapsed. In Figure 2, the arrow pointing to the right indicates that terminal 111 is moving from base station 101 to base station 102. For example, terminal 111 can measure the received power of signals received from base station 101 and base station 102 in response to receiving measurement settings (Measurement Configuration, MeasConfig) from base station 101. Furthermore, terminal 111 can estimate future received power from base station 101 and base station 102 by applying the measurement results to an AI model. As an example, terminal 111 can periodically measure the received power from base station 101 and base station 102 and retain one or more measurement values. By inputting one or more of the retained received power measurement values ​​from base station 101 and base station 102, along with terminal 111's movement speed, into an AI model, terminal 111 can obtain estimated values ​​of the received power from base station 101 and base station 102 after a predetermined time has elapsed as output. In this case, it can be interpreted that the AI ​​model outputs an estimated value of the received power at a location to which the terminal 111 has moved after a predetermined time has elapsed, rather than the location where the received power was measured. If the terminal 111 is far from the base station 101, the estimated value of the received power from the base station 101 after a predetermined time has elapsed may be lower than the current measured value of the received power.Furthermore, when terminal 111 is approaching base station 102, the estimated value of the received power from base station 102 after a predetermined time has elapsed may be higher than the current measured value of the received power. For example, in Figure 2, comparing the solid line 211 and the dotted line 212 at the same position of terminal 111, the estimated value is lower than the measured value. Also, comparing the solid line 221 and the dotted line 222 at the same position of terminal 111, the estimated value is higher than the measured value.

[0016] When determining whether or not to have terminal 111 perform a handover, base station 101 may use the received power after a predetermined period has elapsed, as estimated by terminal 111. This makes it possible for base station 101 to give terminal 111 an early handover instruction that takes into account the movement of terminal 111. For example, in the measurement setting report setting (Report Configuration, ReportConfig), base station 101 instructs terminal 111 to report an estimated value of the received power. Based on the instruction from base station 101, terminal 111 reports an estimated value in addition to or instead of the measured value. Based on the estimated value reported by terminal 111, base station 101 determines whether or not the conditions for handover are met. For example, suppose the condition for determining that a handover should be performed (handover condition) is that the received power from the target cell at terminal 111 exceeds the received power from the source cell by a predetermined value. The power received from the source cell is the power received from the source base station 101, and the power received from the target cell is the power received from the target base station 102. Here, we compare the determination of the handover condition based on measured values ​​and the determination of the handover condition based on estimated values ​​using Figure 2. If we determine whether the handover condition is met at the location of terminal 111 in Figure 2 based on the measured value of the received power, the handover condition will not be met. This is because, comparing the value of solid line 211 and the value of solid line 221 at the location of terminal 111, the value of solid line 211 is higher. On the other hand, if we determine whether the handover condition is met at the location of terminal 111 in Figure 2 based on the estimated value of the received power, it may be determined that the handover condition is met. This is because, comparing the value of dotted line 212 and the value of dotted line 222 at the location of terminal 111, the value of dotted line 222 is higher. If it is estimated that the handover condition is met, base station 101 may initiate the procedure to have terminal 111 perform a handover.

[0017] In this case, if terminal 111 is instructed to perform a handover based on estimated values, a situation may arise where, immediately after base station 101 instructs terminal 111 to perform a handover from base station 101 to base station 102, base station 102 instructs terminal 111 to perform a handover from base station 102 to base station 101. For example, this situation may occur if a different AI model is used in communication with base station 102 after terminal 111 has performed a handover, or if base station 102 does not use inference based on the AI ​​model in its communication with terminal 111. As an example, if inference based on the AI ​​model is not used in communication between base station 102 and terminal 111, base station 102 may notify terminal 111 of a reporting setting instructing it to report the measured values ​​of the received signals from each base station. Then, base station 101 may determine whether the handover conditions are met based on the reported measured values. In this case, comparing the value of solid line 211 and the value of solid line 221 at the location of terminal 111 in Figure 2, the value of solid line 211 is higher, so the handover condition from base station 102 to base station 101 may be met. If the handover condition is met, base station 102 may initiate the procedure to have terminal 111 perform a handover. In this way, a ping-pong problem can occur in which terminal 111 repeatedly performs a handover between base station 101 and base station 102. Similarly, if the AI ​​model that base station 102 has terminal 111 use to estimate received power is a different model from the AI ​​model used for communication between base station 101 and terminal 111, such as a model that does not include the terminal 111's movement speed as input, the ping-pong problem can also occur.

[0018] Such a ping-pong problem can be resolved by using the same AI model used in communication between base station 101 and terminal 111 for communication between base station 102 and terminal 111. However, there may be cases where the same AI model used in communication between base station 101 and terminal 111 should not be used for communication between base station 102 and terminal 111. For example, in a mobile communication system, a common AI model may be provided to each base station 100, and each base station 100 may fine-tune this common AI model according to the environment in which its device is located. In this case, base station 100 can communicate with terminal 111 using the fine-tuned AI model for its environment. By using the fine-tuned AI model, base station 100 and terminal 111 can perform highly accurate inferences according to the environment of the location where each base station 100 is located. For example, base station 100 can generate a fine-tuned AI model by using data obtained from actual communication with terminals at its location as training data to perform further machine learning on the common AI model. If finely tuned AI models are used at both base station 101 and base station 102, the wireless quality may deteriorate when the AI ​​model finely tuned at base station 101 is used in communication between base station 102 and terminal 111. This is because the AI ​​model finely tuned in the environment in which base station 101 is deployed is used in an environment different from the environment in which base station 101 is deployed, which increases the error in the AI ​​model's predictions.

[0019] In light of these circumstances, in this embodiment, terminal 111 is configured to notify the second base station device 102, which constitutes the second cell, of specific information identifying the first trained model, based on the fact that it has been decided to perform a handover from the first cell to the second cell while in the service area, and that inference using the first trained model included in the multiple trained models has been performed in the first communication between terminal 111 and the first base station 101, which constitutes the first cell. Furthermore, in this embodiment, when base station 102 receives the notification of the specific information, it makes a decision based on the notification whether or not to allow terminal 111 to continue using the first trained model, and sends the above-mentioned inference instructions based on that decision to terminal 111. With this configuration, terminal 111 can notify base station 102, which is the destination after the handover, of the AI ​​model that was used in the communication with base station 101 before the handover. The base station 102 can identify the AI ​​model that the terminal 111, which establishes a connection with the base station 102 through the handover, was using in its communication with the base station 102 before the handover. The base station 102 then determines whether or not that AI model should be used in the communication between the base station 102 and the terminal 111, and based on the result of that determination, it can select an appropriate AI model to use for communication with the terminal 111.

[0020] For example, if the instructions from the base station 102 include information indicating that inference should be performed using the first trained model, terminal 111 will continue to perform inference using the first trained model in the second communication with the base station 102. Also, if the instructions from the base station 102 include information indicating that inference should be performed using a second trained model different from the first trained model, terminal 111 will perform inference using the second trained model in the second communication with the base station 102. If terminal 111 has the second trained model, it will perform inference using that second trained model. If terminal 111 does not have the second trained model, it will obtain the second trained model from the base station 102 and perform inference using the obtained second trained model. Furthermore, if the instructions from the base station 102 include information indicating that inference should not be performed, terminal 111 will not perform inference.

[0021] Terminal 111 may notify specific information using a random access preamble in a four-step random access procedure. For example, if each of several trained models is associated with each of the patterns in the random access preamble, terminal 111 may notify specific information by transmitting the random access preamble corresponding to the first trained model. Also, if a specific radio resource is allocated for transmitting the random access preamble used to notify specific information, terminal 111 may notify specific information by transmitting the random access preamble using the specific radio resource. Terminal 111 may also notify specific information using message 3 in a four-step random access procedure. Alternatively, terminal 111 may notify specific information using message A in a two-step random access procedure. Furthermore, if base station 101 is performing inference for the first communication using a third trained model corresponding to the first trained model, terminal 111 may obtain the third trained model from base station 101 and provide it to base station 102. As a result, even if the base station 102 does not have a third trained model, the terminal 111 can continue to use inference using the first trained model. An example of the configuration and operation of the terminal 111 and base station 100 operating in this manner is described below.

[0022] (Circuit Configuration) An example configuration of the base station 100 and terminal 111 will be described. Figure 3 is a diagram showing the hardware configuration of the base station 100 and terminal 111. In one example, the base station 100 and terminal 111 are configured to include a processor 301, ROM 302, RAM 303, storage device 304, and communication circuit 305. The processor 301 is a computer configured to include one or more processing circuits, such as a general-purpose CPU (Central Processing Unit) or ASIC (Application-Specific Integrated Circuit). The processor 301 performs the overall processing of the device and the above-mentioned processing by reading and executing programs stored in the ROM 302 and storage device 304. The ROM 302 is a read-only memory in which information such as programs and various parameters related to the processing performed by the base station 100 and terminal 111 is recorded. The RAM 303 functions as a workspace when the processor 301 executes programs and is a random access memory in which temporary information is recorded. The storage device 304 is configured, for example, by a removable external storage device. The communication circuit 305 is configured to include, for example, circuits for wired or wireless communication between the base station 100 and the terminal 111. For example, the base station 100 and the terminal 111 can communicate with the other party's communication device using the communication circuit 305 for LTE or 5G and an antenna (not shown). The base station 100 can also communicate with other base stations using the communication circuit 305 for wired communication.

[0023] (Functional Configuration) Figure 4 shows an example of the functional configuration of the base station 100. The base station 100 is configured to include, for example, a model information receiving unit 401, a model usage determination unit 402, a determination transmission unit 403, a model management unit 404, a model request unit 405, and a model supply unit 406. Figure 4 shows the functional configuration of the base station 100 in this embodiment, and the general configuration of the base station 100 is omitted. These functional units can be realized, for example, by the processor 301 executing a program stored in the ROM 302 or storage device 304 and controlling the communication circuit 305 as needed. However, it is not limited to this, and for example, dedicated hardware for realizing each function may be provided. The following describes an example of the functional configuration of the base station 102.

[0024] The model information receiving unit 401 receives information from the terminal 111 about the AI ​​model used in the communication between the base station 101 and the terminal 111. For example, the model information receiving unit 401 receives a notification from the terminal 111 that includes specific information identifying the first AI model used in the first communication between the base station 101 and the terminal 111. The model usage decision unit 402 decides whether or not to continue using the first AI model in the second communication between the base station 102 and the terminal 111. For example, the model usage decision unit 402 identifies the first AI model based on the specific information included in the notification from the terminal 111 and decides whether or not to continue using that first AI model in the second communication between the base station 102 and the terminal 111. The decision communication unit 403 transmits the decision made by the model usage decision unit 402. For example, the decision communication unit 403 transmits an instruction including the decision to the terminal 111.

[0025] The model management unit 404 manages the AI ​​models owned by the device. For example, the model management unit 404 holds identification information for the AI ​​models owned by the device and identification information for the AI ​​models used by the terminal 111 connected to the device. When the model management unit 404 receives identification information from the terminal 111 that identifies a first AI model, it compares this information with the identification information for the AI ​​models held by the device and determines whether the device has a third AI model corresponding to the first AI model. The determination result may be used by the model request unit 405. The model request unit 405 requests the terminal 111 to provide a third AI model if the model usage determination unit 402 has determined that the first AI model will be used continuously in the second communication between the device and the terminal 111, and the model management unit 404 has determined that the device does not have a third AI model corresponding to the first AI model. When a third AI model is supplied from terminal 111, the model management unit 404 manages the identification information of that AI model.

[0026] Figure 5 shows an example of the functional configuration of terminal 111. Terminal 111 is configured to include, for example, a model information notification unit 501, an instruction receiving unit 502, an inference execution unit 503, a model management unit 504, a model acquisition unit 505, and a model provision unit 506. Figure 5 shows the functional configuration of terminal 111 in this embodiment, and omits, for example, the general configuration of terminal 111. These functional units can be realized, for example, by the processor 301 executing a program stored in ROM 302 or storage device 304 and controlling the communication circuit 305 as needed. However, it is not limited to this, and for example, dedicated hardware for realizing each function may be provided.

[0027] The model information notification unit 501 notifies the base station 102 of the AI ​​model used in the communication between the base station 101 and the terminal 111. For example, the model information notification unit 501 sends a notification to the base station 102 that includes specific information identifying the first AI model used in the first communication between the base station 101 and the terminal 111. As an example, the model information notification unit 501 may use message 1 or message 3 in a four-step random access procedure or message A in a two-step random access procedure to send a notification that includes specific information identifying the first AI model. The instruction receiving unit 502 receives instructions from the base station 102 regarding the execution of inference in the second communication between the base station 102 and the terminal 111. For example, instructions from the base station 102 may include whether to perform inference using the first trained model, whether to perform inference using the second trained model, or whether to not perform inference.

[0028] The inference execution unit 503 performs inference based on instructions received from the base station 102. For example, if the instructions received from the base station 102 include information indicating that inference should be performed using the second trained model, the inference execution unit 503 performs inference using the second trained model in the second communication. If the instructions received from the base station 102 include information indicating that inference should not be performed, the inference execution unit 503 does not perform inference in the second communication.

[0029] The model management unit 504 manages the AI ​​models owned by the device. For example, the model management unit 404 holds identification information for the AI ​​models owned by the device and identification information for the AI ​​models used by base stations 101 and 102 to which the device is connected. The model acquisition unit 505 acquires the AI ​​models used by the device and the AI ​​models used by base station 100 to which the device is connected. For example, when the model acquisition unit 505 is performing inference using the first AI model in the first communication with base station 101, it acquires the third AI model that corresponds to the first AI model and is used by base station 101. The model provision unit 506 provides the AI ​​models owned by the device to base station 100. For example, the model provision unit 506 provides the third AI model acquired during the first communication with base station 101 to base station 102.

[0030] (Processing Flow) The operation of each device when a terminal 111 connected to base station 101 performs a handover to base station 102 will be explained. Figure 6 shows an example of the processing performed by each device when a terminal 111 connected to base station 101 performs a handover to base station 102. Base station 101 to which terminal 111 performing the handover is connected is called the source base station. Base station 102 to which terminal 111 connects after the handover is called the target base station. First, let's assume that the source base station 101 instructs terminal 111 to measure the radio quality of the cells provided by its device and the cells (peripheral cells) around terminal 111. For example, the source base station 101 may instruct terminal 111 to perform the measurement by notifying terminal 111 of an RRC message including the measurement settings (MeasConfig). RRC is an abbreviation for Radio Resource Control. Furthermore, the source base station 101 may instruct the terminal 111 to use measured radio quality values ​​to estimate whether predetermined conditions are met. The predetermined conditions may include one or more of the following: the radio quality of the serving cell provided by the source base station 101 is better than a threshold (event A1), the radio quality of the serving cell is worse than a threshold (event A2), the radio quality of the surrounding cell is better than the radio quality of the serving cell by more than a threshold (event A3), the radio quality of the surrounding cell is better than a threshold (event A4), or the radio quality of the serving cell is worse than a first threshold and the radio quality of the surrounding cell is better than a second threshold (event A5). Radio quality is, for example, received power. Values ​​indicating received power include RSRP, RSRQ, RSSI, SINR, etc. RSRP is an abbreviation for Reference Signal Received Power. RSRQ is an abbreviation for Reference Signal Received Quality. RSSI is an abbreviation for Received Signal Strength Indicator. SINR is an abbreviation for Signal to Interference plus Noise Ratio.The method by which terminal 111 determines whether predetermined conditions are met and reports this to source base station 101 may be called event trigger reporting. The determination of whether predetermined conditions are met may be performed by source base station 101. The following explanation uses an example in which terminal 111 performs event trigger reporting. For example, terminal 111 may be notified of predetermined conditions in the measurement settings. Terminal 111 performs measurement and reports the estimated results based on instructions from source base station 101. Based on the meeting of predetermined conditions, the start of the procedure for performing handover is triggered (S601). For example, terminal 111 measures the radio quality based on the measurement settings notified by source base station 101, estimates whether predetermined conditions are met after a predetermined time has elapsed using the measured values, and reports to source base station 101 if it is estimated that predetermined conditions are met. The report by terminal 111 may include information identifying the surrounding cells in which the predetermined conditions are met. In this example, it may include information identifying the target base station 102. Based on the report from terminal 111, source base station 101 decides whether or not to perform a handover to target base station 102. For example, source base station 101 may decide to perform a handover based on receiving a report from terminal 111. This triggers the start of the procedure for terminal 111 to perform the handover.

[0031] Terminal 111 can estimate wireless quality by using an AI model for inference during communication with source base station 101. For example, when terminal 111 receives measurement settings from source base station 101, it periodically measures received power according to the received measurement settings. Terminal 111 can store a history of measured received power values ​​for signals received from source base station 101 and surrounding cells as data. Terminal 111 can input each of the stored measured received power values ​​from source base station 101, as well as the movement speed of its own device, to each input port of the AI ​​model. The AI ​​model can output one or more estimated values ​​based on the measured values ​​input by terminal 111. For example, the AI ​​model can output estimated received power values ​​for signals received from source base station 101 at intervals of unit time. For example, if the unit time is T seconds, the AI ​​model can output estimated received power values ​​when 1 × T seconds, 2 × T seconds, 3 × T seconds, ... have elapsed since the measurement was performed. Similarly, terminal 111 can input the measured values ​​of received power it holds and the movement speed of its own device to each input port of the AI ​​model for each of the surrounding cells. The AI ​​model can output one or more estimated values ​​based on the measured values ​​input by terminal 111. In this way, terminal 111 can estimate future received power from source base station 101 and surrounding cells. Terminal 111 may also input both the measured values ​​of received power of the signal received from source base station 101 and the measured values ​​of received power of the signal received from surrounding cells into the AI ​​model to estimate the received power from source base station 101 and surrounding cells. Terminal 111 uses the estimated received power values ​​to determine whether predetermined conditions are met. For example, if terminal 111 makes a determination using event A3, the predetermined conditions may be met if terminal 111 is located at the position shown in Figure 2. Terminal 111 can use an AI model in communication with source base station 101 or select an AI model to use based on instructions from source base station 101. For example, the source base station 101 may notify the terminal 111 of information indicating whether or not the use of an AI model is permitted, or information identifying the AI ​​model to be used. The terminal 111 may then perform inference using the AI ​​model based on the notification.

[0032] When the source base station 101 is triggered to initiate the procedure for performing a handover, it sends a handover request to the target base station 102 (S602). For example, the source base station 101 may send a handover request using the X2 interface or Xn interface on the network 131. As an example, the handover request may be sent as a Handover Request message. When the target base station 102 receives the handover request, it sends a response to this handover request (S603). As an example, the response to the handover request may be sent as a Handover Request Acknowledge message. The Handover Request Acknowledge message may include communication parameters for terminal 111 to communicate with target base station 102. When source base station 101 receives a response to a handover request from target base station 102, it instructs terminal 111 to perform a handover (S604). For example, source base station 101 may give this instruction via an RRC reconfiguration message containing a Handover Command. The Handover Command may include cell identification information provided by target base station 102, which will be the handover destination.

[0033] Terminal 111 performs a procedure to establish a connection with the target base station 102 based on the information contained in Handover Command. For example, terminal 111 may establish a connection with the target base station 102 using a random access procedure. First, terminal 111 sends message 1 containing a Random Access Preamble (S605). For example, terminal 111 performs cell synchronization using a synchronization signal transmitted by base station 102 and identifies the radio resource to be used for random access by receiving a system information block transmitted by base station 102. As an example, the synchronization signal is a Synchronization Signal Block (SSB) and the system information block is a System Information Block (SIB). The radio resource to be used for random access may be called a RACH. RACH is an abbreviation for Random Access Channel. Terminal 111 sends message 1, which includes a random access preamble, using the identified RACH.

[0034] When the target base station 102 receives message 1, it sends message 2, which includes a Random Access Response. The Random Access Response contains information that identifies the radio resources that terminal 111 can use to send message 3. When terminal 111 receives message 2 from the target base station 102, it sends message 3 using the radio resources identified in message 2. When the target base station 102 receives message 3 from terminal 111, it sends message 4, which indicates that the random access was successful. For example, message 4 may include a temporary network identifier to be assigned to terminal 111. This completes the random access procedure.

[0035] The target base station 102 sends an RRC Reconfiguration message to the terminal 111 to configure radio resources and communication parameters (S609). When the terminal 111 has completed the configuration based on the RRC Reconfiguration message, it sends an RRC Reconfiguration Complete message (S610). As a result, the terminal 111 establishes a connection with the target base station 102. The state in which the terminal 111 has established a connection with the target base station 102 is called the RRC Connected state.

[0036] While executing a random access procedure with the target base station 102, terminal 111 may notify the target base station 102 of information identifying the AI ​​model used in communication with the source base station 101. For example, if terminal 111 performed inference using an AI model in communication with the source base station 101, it will retain information identifying that AI model. The information identifying the AI ​​model will be described later. For example, if the information identifying the AI ​​model can be expressed with a small amount of information, terminal 111 may notify the target base station 102 of the information identifying the AI ​​model using a radio resource or preamble pattern for transmitting a random access preamble. As an example, if each AI model is assigned an identifier, and a radio resource or preamble pattern for random access is associated with that identifier, terminal 111 may notify the target base station 102 of the information identifying that AI model by transmitting a random access preamble using a radio resource or preamble pattern corresponding to the identifier of the AI ​​model used in communication with the source base station 101. Note that a radio resource used to notify the information identifying the AI ​​model, which is not used in a normal random access procedure, may be provided. In this case, terminal 111 can use this radio resource to transmit a random access preamble, thereby notifying target base station 102 that it was performing inference using an AI model in its communication with source base station 101. Furthermore, terminal 111 can use this radio resource to transmit a random access preamble of a preamble pattern associated with the identifier of the AI ​​model used by its device, thereby notifying target base station 102 of information identifying the AI ​​model used by its device.

[0037] On the other hand, if terminal 111 requires a certain amount of information to identify the AI ​​model, it may notify the target base station 102 of the information to identify the AI ​​model using message 3 or the like. For example, if terminal 111 is provided with the aforementioned radio resources used to notify information to identify the AI ​​model and not used in normal random access procedures, it may notify the target base station 102 that it was performing inference using the AI ​​model in its communication with source base station 101 by transmitting a random access preamble using these radio resources. When target base station 102 determines that terminal 111 was performing inference using the AI ​​model in its communication with source base station 101, it allocates radio resources to terminal 111 for notifying information to identify the AI ​​model and transmits information indicating this allocation of radio resources in message 2. Terminal 111 may then transmit information to identify the AI ​​model using the radio resources allocated to its device as indicated in message 2.

[0038] Terminal 111 may also notify the target base station 102 of information identifying the AI ​​model after the random access procedure is completed. For example, terminal 111 may notify the AI ​​model using an RRC Reconfiguration Complete message or a UE Assistance Information message.

[0039] This section describes the information used to identify an AI model. An AI model can be identified by its configuration, parameters, and training data. The configuration of an AI model may include rules, formulas, and the structure of a neural network used to output inference results based on input data. The parameters of an AI model may be, for example, the weight coefficients of a trained model generated by performing machine learning using the training data. For example, by being notified of the formulas that constitute the AI ​​model and the values ​​of each coefficient included in the formulas, the target base station 102 can identify the AI ​​model used in communication between the source base station 101 and the terminal 111. Furthermore, by being notified of the structure of the neural network used in the AI ​​model and the weights of each neuron, the target base station 102 can identify the AI ​​model used in communication between the source base station 101 and the terminal 111. In addition, the training data used to generate the AI ​​model may be provided to the target base station 102. In this case, the target base station 102 can generate an AI model by performing machine learning using the provided training data. Alternatively, the AI ​​model itself may be provided to the target base station 102. On the other hand, if each AI model is assigned identification information that allows it to be identified, that identification information can be notified to the target base station 102. For example, if a specific server in a mobile communication system manages AI models used in communication devices including base stations 100 and terminals 111, and provides AI models to each communication device, then each AI model may be assigned identification information that is commonly used in the mobile communication system. For example, different identification information may be assigned to each AI model with a different purpose, such as an AI model used for predicting radio quality, an AI model used for beam management, an AI model used for predicting terminal positioning, and an AI model used for compressing and restoring channel state information. The identification information may be associated with predetermined attribute information. The identification information may include predetermined attribute information. For example, the attribute information may include information to specify the purpose of the AI ​​model or information to specify the environment in which the AI ​​model should be used.Furthermore, if multiple AI models exist for the same purpose, different identification information may be assigned to each AI model. For example, for AI models used to estimate wireless quality, different identification information may be assigned to an AI model optimized for estimating wireless quality in urban areas with densely packed buildings and multipath environments, and to an AI model optimized for estimating wireless quality in rural areas with low building density and clear lines of sight. Also, when a particular AI model is updated, update information may be assigned. The update information may be a version number. For example, an AI model may be updated using additional training data. Different version numbers may be assigned to each AI model to distinguish between updated and unupdated AI models. For example, a version number corresponding to the training data used to generate the AI ​​model may be assigned. In addition, the date and time the AI ​​model was updated may be added as update information for the AI ​​model. The information notified to the target base station 102 as information identifying the AI ​​model is not limited to the above, and any information that the target base station 102 can use to identify the AI ​​model may be used.

[0040] When the target base station 102 receives information from the terminal 111 that identifies an AI model, it determines whether to continue using the identified AI model in communication between its device and the terminal 111. For example, the target base station 102 may decide to continue using the identified AI model based on the fact that the identified AI model is used for estimating radio quality to determine whether to perform a handover. This can resolve the ping-pong problem in which the terminal 111 repeatedly performs handovers between base station 101 and base station 102. On the other hand, the target base station 102 may decide not to continue using the identified AI model based on the fact that the identified AI model is used for estimating radio quality to select the optimal MCS, and the environment in which the identified AI model should be used is different from the environment in which its device is deployed. MCS is an abbreviation for Modulation and Coding Scheme. For example, if the information identifying the AI ​​model is identification information for the AI ​​model, the target base station 102 may determine the intended use of the AI ​​model and the environment in which the AI ​​model should be used based on this identification information and attribute information included in or associated with the identification information. The target base station 102 may also acquire information indicating the intended use of the AI ​​model and the environment in which it should be used, in conjunction with the information identifying the AI ​​model. Furthermore, if the identified AI model cannot be used in the communication of its own device, the target base station 102 may determine that it will not continue to use that AI model. For example, the target base station 102 may determine that it will not continue to use that AI model based on factors such as the system being different from the source base station 101 (e.g., differences between NR and LTE), the device type being different (e.g., Massive MIMO base station and omni-antenna base station), the manufacturer being different, the frequency being used being different (e.g., differences in frequency bands such as 800 MHz, 3.7 GHz, and 4.0 GHz, differences between TDD and FDD frequencies), or the version of the software being used being different.

[0041] The target base station 102 may notify the terminal 111 of the result of its determination of whether or not to continue using the identified AI model. For example, if the received message 1 contains information identifying the AI ​​model, the target base station 102 may notify the terminal 111 of the determination result using message 2 or message 4. Also, if the received message 3 contains information identifying the AI ​​model, the target base station 102 may notify the terminal 111 of the determination result using message 4. The target base station 102 may also notify the terminal 111 of the determination result using an RRC Reconfiguration message. When the terminal 111 notifies the target base station 102 of information identifying the AI ​​model, the target base station 102 can determine which AI model to use in communication with the terminal 111, including that information. Therefore, instead of notifying the terminal 111 of the determination result from the target base station 102, the target base station 102 may notify the terminal 111 of instructions on whether or not to perform inference using the AI ​​model, or information indicating which AI model should be used for inference.

[0042] In the above example, terminal 111 established a connection with target base station 102 using a four-step random access procedure. However, terminal 111 can also establish a connection with target base station 102 using a two-step random access procedure. In this case, terminal 111 can notify target base station 102 of information identifying the AI ​​model used in communication with source base station 101 using message A, which corresponds to messages 1 and 3 in the four-step random access procedure. For example, terminal 111 can notify target base station 102 of information identifying the AI ​​model using a radio resource or preamble pattern that transmits a random access preamble. Alternatively, terminal 111 can notify target base station 102 by sending message A, which contains information identifying the AI ​​model, such as AI model identification information. In this case, target base station 102 can notify terminal 111 of the above-mentioned determination result using message B.

[0043] (Modification 1) In the above, the operation of terminal 111 when it performs a handover from source base station 101 to target base station 102 was explained as an example. In this modification, the operation of terminal 111 when it performs cell reselection will be explained. For example, terminal 111 may transition from a connected state with base station 101 (RRC Connected state) to an unconnected state such as an idle state (RRC Idle state) or an inactive state (RRC Inactive state). Then, terminal 111 may perform cell reselection when it reconnects to the network after a certain period of time has elapsed. Also, if a link failure (Radio Link Failure) occurs in the link established between source base station 101 and terminal 111, terminal 111 may perform a network reconnection process (RRC Reestabilizement). In these cases, terminal 111 may connect to base station 102 instead of base station 101. In this situation, the same problem as described above may occur due to a difference between the AI ​​model used by terminal 111 for communication with base station 101 and the AI ​​model used for communication with base station 102. The above technology can be applied even in such a situation.

[0044] Figure 7 shows an example of the processing performed by each device when terminal 111, which is connected to base station 101, becomes idle and then connects to base station 102 through cell reselection.In Figure 7, the same processing as in Figure 6 is given the same reference number and its explanation is omitted. Assume that terminal 111 is connected to base station 101.For example, base station 101 may transition terminal 111 to an idle state (S701) based on the fact that there has been no communication traffic with terminal 111 for a certain period of time, or that the received power from terminal 111 has fallen below a predetermined threshold.For example, base station 101 may transition terminal 111 to an idle state by sending an RRC Release message. Terminal 111 transitions to an idle state based on instructions from base station 101.At this time, base station 101 may instruct terminal 111 to retain the AI ​​model that was used for communication with its own device.As a result, when terminal 111 returns to a connected state with its own device, inference using the AI ​​model can be performed quickly. For example, base station 101 may notify terminal 111 of the deadline for retaining the AI ​​model. For instance, if terminal 111 reconnects to the network at a location far from base station 111, it is possible that the location is unsuitable for inference using the AI ​​model held by terminal 111. Setting a deadline for terminal 111 to retain the AI ​​model prevents terminal 111 from attempting to use the same AI model when it moves to a different environment. In this case, when terminal 111 reconnects to the network, it determines whether it holds the AI ​​model it should use, and if so, whether the deadline for retaining that AI model has expired. If the deadline for retaining that AI model has not expired, terminal 111 can continue to use that AI model. For example, in a random access procedure, terminal 111 notifies base station 102 of information that identifies the AI ​​model. On the other hand, if the AI ​​model that terminal 111 holds has expired, it will not continue to use that AI model. In this case, terminal 111 does not notify base station 102 of information that identifies the AI ​​model, for example, in a random access procedure.Note that if the retention period of the held AI model has expired, the terminal 111 may discard said AI model. This eliminates the need for the terminal 111 to determine whether information specifying the AI model should be notified when the terminal 111 performs cell reselection.

[0045] Terminal 111 attempts to connect to the network, for example, when data to be transmitted from its own device is generated. If terminal 111 is idle, it performs a cell reselection procedure. For example, terminal 111 may identify a base station to connect to based on the received power of signals received from surrounding cells. For example, terminal 111 may determine that it should connect to base station 102 based on the fact that the received power of the signal received from base station 102 is higher than the received power of signals received from other base stations. Terminal 111 performs cell synchronization using the synchronization signal transmitted by base station 102 and identifies the radio resources to be used for random access by receiving the system information block transmitted by base station 102. Terminal 111 may connect to a cell if the cell reselection conditions indicated in the system information block are met. If the cell reselection conditions are not met, terminal 111 does not connect to that cell. Terminal 111 establishes a connection with base station 102 using the random access procedure (S605-S610). At this time, terminal 111 determines whether or not it possesses the AI ​​model that was used in communication with base station 101. If terminal 111 possesses the AI ​​model that was used in communication with base station 101, it may notify base station 102 of information identifying that AI model during the random access procedure with base station 102. Terminal 111 may also determine that it does not possess the AI ​​model if the period for which it should retain that AI model has expired. Terminal 111 may notify base station 102 of information identifying the AI ​​model using the method described in Figure 6. Furthermore, if terminal 111 does not possess the AI ​​model that was used in communication with base station 101, it will not notify base station 102 of information identifying the AI ​​model during the random access procedure with base station 102. For example, terminal 111 may discard the AI ​​model if the period for which it should retain the AI ​​model that was used in communication with base station 101 has expired.

[0046] Furthermore, in the reconnection process when a link break occurs between the source base station 101 and the terminal 111, the terminal 111 may notify the target base station 102 of information identifying the AI ​​model used in communication with base station 101 when performing random access as described above. For example, when the terminal 111 detects that a link break has occurred, it may identify the base station to connect to based on the received power of signals received from surrounding cells. Once the terminal 111 has identified the base station to connect to, it establishes a connection with base station 102 using a random access procedure (S605-S610). In the random access procedure with base station 102, the terminal 111 may notify base station 102 of information identifying its AI model. For example, the terminal 111 may notify base station 102 of information identifying its AI model using the method described in Figure 6.

[0047] (Modification 2) In the above explanation, an example of a one-sided model was used in which only terminal 111 performs inference using the AI ​​model. In this case, by notifying the target base station 102 of specific information that identifies the AI ​​model used in the communication between source base station 101 and terminal 111, terminal 111 can continue to use the same AI model in the communication between target base station 102 and terminal 111. On the other hand, in the case of a two-sided model in which inference using the AI ​​model is performed in both base station 100 and terminal 111, even if the AI ​​model is identified, if the target base station 102 does not have that AI model, terminal 111 cannot use the same AI model in the communication between target base station 102 and terminal 111. For example, if inference using an AI model is used for CSI compression and decompression, and the AI ​​model used for the encoder for CSI compression used by terminal 111 does not correspond to the AI ​​model used for the decoder for CSI decompression used by target base station 102, then the information compressed in CSI at terminal 111 will not be properly restored at target base station 102. In such a situation, terminal 111 may supply the AI ​​model used by source base station 101 to target base station 102 in communication between itself and source base station 101. By using the AI ​​model supplied by terminal 111, target base station 102 can perform inference using the same AI model in communication between itself and terminal 111. For example, if the AI ​​model for CSI decompression that corresponds to the AI ​​model for CSI compression used by terminal 111 is used by target base station 102, then the information compressed in CSI at terminal 111 will be properly restored at target base station 102.

[0048] FIG. 8 shows an example of processing executed in each apparatus when terminal 111 connected to base station 101 executes handover to base station 102. In FIG. 8, operations similar to those in FIG. 6 are assigned the same reference numerals, and descriptions thereof are omitted. Terminal 111 establishes a connection with target base station 102 using a random access procedure (S605 to S610). During this period, terminal 111 notifies target base station 102 of information identifying an AI model that has been used in communication between terminal 111 itself and source base station 101. Target base station 102 determines whether or not to continuously use the identified AI model in communication between the base station itself and terminal 111. When target base station 102 determines to continue using said AI model, and detects that said AI model is a Two-sided model and the base station itself does not have an AI model corresponding to said AI model, it requests terminal 111 to supply the AI model (S801). Note that target base station 102 does not request terminal 111 to supply the AI model if target base station 102 determines not to continuously use the identified AI model in communication between the base station itself and terminal 111, if the identified AI model is a One-sided model, or if the base station itself has an AI model corresponding to the identified AI model.

[0049] When terminal 111 receives a request for an AI model from target base station 102, it supplies the requested AI model (S802). For example, terminal 111 may obtain the AI ​​model used by source base station 101 from source base station 101 during communication with source base station 101. For example, if a two-sided AI model is used, terminal 111 may obtain not only the AI ​​model used in its own device but also the AI ​​model used by base station 101. By holding the AI ​​model obtained in this way, terminal 111 can supply that AI model to target 102. For example, terminal 111 may supply the target base station 102 with information indicating the configuration and parameters of the AI ​​model. For example, terminal 111 may notify the mathematical formulas that constitute the AI ​​model and the values ​​of each coefficient included in the formulas. Terminal 111 may also notify the configuration of the neural network used in the AI ​​model and the weights of each neuron. Furthermore, terminal 111 may provide the target base station 102 with the training data used to generate the AI ​​model. In this case, the target base station 102 can generate an AI model using the provided training data. When the target base station 102 is ready to perform inference using the AI ​​model supplied by terminal 111, it instructs terminal 111 to perform inference using that AI model (S803). In this way, by acquiring the AI ​​model from terminal 111, the target base station 102 can perform inference using that AI model in communication with terminal 111.

[0050] As described above, according to this embodiment, terminal 111 decides to perform a handover from the first cell in the area to the second cell, and based on the fact that inference using the first trained model included in the plurality of trained models is performed in the first communication between terminal 111 and the first base station 101 constituting the first cell, terminal 111 notifies the second base station device 102 constituting the second cell of specific information that identifies the first trained model. Furthermore, when base station 102 in this embodiment receives the notification of the specific information, it decides whether or not to allow terminal 111 to continue using the first trained model based on the notification, and transmits the above-mentioned inference instructions based on that decision to terminal 111. With this configuration, terminal 111 can notify base station 102, the destination after the handover, of the AI ​​model that was used in the communication with base station 101 before the handover. Base station 102 can identify the AI ​​model used by terminal 111, which establishes a connection with its device through handover, in communication with base station 101 before the handover. Base station 102 then determines whether or not to use that AI model in communication between its device and terminal 111, and based on the determination result, can select an appropriate AI model to use for communication with terminal 111. As a result, when terminal 111 performs a handover between multiple base stations, the appropriate AI will be selected both before and after the handover, thereby stabilizing the wireless quality provided to terminal 111. Thus, it will be possible to contribute to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs), "Build resilient infrastructure, promote sustainable industrialization and foster innovation."

[0051] The invention is not limited to the embodiments described above, and various modifications and changes are possible within the scope of the gist of the invention.

[0052] This application claims priority based on Japanese Patent Application No. 2025-037367, filed on March 10, 2025, and all of its contents are incorporated herein by reference.

Claims

1. A terminal device that communicates with a base station device based on the cellular communication standard of the Third Generation Partnership Project (3GPP), comprising: an inference means capable of performing inference for the communication using at least one of a plurality of trained models generated by machine learning; a notification means that notifies a second base station device constituting the second cell of specific information identifying the first trained model, based on the fact that it has been decided that the terminal device will perform a handover from a first cell to a second cell in which it is located, and that the inference using the first trained model included in the plurality of trained models has been performed in the first communication between the first base station device constituting the first cell and the terminal device; and a receiving means that receives instructions regarding the execution of the inference in the second communication between the second base station device and the terminal device, which have been determined by the second base station device based on the notification.

2. The terminal device according to claim 1, wherein the inference means continues to perform the inference using the first trained model in the second communication if the instruction includes information indicating that the inference should be performed using the first trained model.

3. The terminal device according to claim 1, wherein the inference means performs the inference using the second trained model in the second communication if the instruction includes information indicating that the inference should be performed using a second trained model different from the first trained model.

4. The terminal device according to claim 3, wherein the inference means performs the inference using the second trained model if the terminal device has the second trained model, and if the terminal device does not have the second trained model, it obtains the second trained model from the second base station device and performs the inference using the obtained second trained model.

5. The terminal device according to claim 1, wherein the inference means does not perform the inference if the instruction includes information indicating that the inference should not be performed.

6. The terminal device according to any one of claims 1 to 5, wherein the notification means provides notification of the specific information using a random access preamble in a four-step random access procedure.

7. The terminal device according to claim 6, wherein each of the plurality of trained models is associated with each of the patterns of the random access preamble, and the notification means notifies the specific information by transmitting the random access preamble of the pattern corresponding to the first trained model.

8. The terminal device according to claim 6, wherein a specific radio resource is allocated for transmitting the random access preamble used for notifying the specific information, and the notification means notifies the specific information by transmitting the random access preamble using the specific radio resource.

9. The terminal device according to any one of claims 1 to 5, wherein the notification means notifies the specific information using message 3 in a four-step random access procedure.

10. The terminal device according to any one of claims 1 to 5, characterized in that the notification means notifies the specific information using message A in a two-step random access procedure.

11. The terminal device according to any one of claims 1 to 10, further comprising: an acquisition means for acquiring a third trained model from the first base station device when the first base station device is performing the inference for the first communication using a third trained model corresponding to the first trained model; and a providing means for providing the acquired third trained model to the second base station device.

12. A base station device that communicates with a terminal device based on the cellular communication standard of the Third Generation Partnership Project (3GPP), wherein, in the communication, at least the terminal device is capable of performing inference for the communication using at least one of a plurality of trained models generated by machine learning, and the base station device has receiving means for receiving notification of specific information identifying the first trained model from the terminal device when it is determined that the terminal device will perform a handover from a first cell provided by another base station device in which the terminal device is located to a second cell provided by the base station device, and when a first trained model included in the plurality of trained models is used in a first communication between the other base station device and the terminal device, determining means for determining whether or not to allow the terminal device to continue using the first trained model based on the notification, and transmitting means for transmitting instructions regarding the inference based on the determination to the terminal device.

13. A control method performed by a terminal device that communicates with a base station device based on the cellular communication standard of the Third Generation Partnership Project (3GPP), comprising: performing inference for the communication using at least one of a plurality of trained models generated by machine learning; notifying a second base station device constituting the second cell of specific information identifying the first trained model, based on the fact that the terminal device has decided to perform a handover from a first cell to a second cell in which it is located, and that the inference using a first trained model included in the plurality of trained models has been performed in a first communication between the first base station device constituting the first cell and the terminal device; and receiving instructions regarding the execution of the inference in a second communication between the second base station device and the terminal device, which have been determined by the second base station device based on the notification.

14. A control method performed by a base station device that communicates with a terminal device based on the cellular communication standard of the Third Generation Partnership Project (3GPP), wherein, in the communication, at least the terminal device is capable of performing inference for the communication using at least one of a plurality of trained models generated by machine learning, and the control method is characterized by: receiving notification of specific information identifying the first trained model from the terminal device when it is determined that the terminal device will perform a handover from a first cell provided by another base station device in which the terminal device is located to a second cell provided by the base station device, and when a first trained model included in the plurality of trained models is used in a first communication between the other base station device and the terminal device; making a decision based on the notification whether or not to allow the terminal device to continue using the first trained model; and transmitting instructions regarding the inference based on the decision to the terminal device.

Images