COMMUNICATION METHOD AND APPARATUS FOR MACHINE LEARNING MODEL TRAINING - Patent application

Abstract

The present application provides a communication method and apparatus for machine learning model training, to improve the effect of machine learning model training and reduce resource waste. The method includes: a first device receives first information from a second device, the first information including at least one of a first indication and a training condition of the machine learning model; a value range of the first indication includes a first value and a second value, the value of the first indication is a first value, and the first value indicates that the first device should perform training on the machine learning model; the first device performs training on the machine learning model based on the first information.

Description

[Technical field]

[0001] The present application relates to the field of mobile communication technologies, and in particular to a communication method and apparatus for machine learning model training. [Background technology]

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to Chinese Patent Application No. 202210313970.1, entitled “COMMUNICATION METHOD FOR MACHINE LEARNING MODEL TRAINING AND APPARATUS,” filed with the China State Intellectual Property Office on March 28, 2022, the entire contents of which are incorporated herein by reference.

[0003] Management data analytics (MDA) provides the ability to process and analyze raw data related to network service events and status, and provide analysis reports to implement required actions of network service operations. Currently, MDA can be combined with artificial intelligence (AI) or machine learning (ML) technologies to improve network service management and coordination for intelligence and automation.

[0004] A management data analytics service (MDAS) is a management service that is enabled for MDA. An MDAS producer can output analysis reports to an MDAS consumer in different scenarios, such as coverage problem analysis scenario, fault event analysis scenario, or mobility management analysis scenario. In different scenarios, the ML models to obtain the analysis reports may be different.

[0005] Currently, an AI / ML model training management service (MnS) producer (AI / ML model training MnS producer) (model training producer for short) may initiate ML model training based on a request of a model training management service consumer (AI / ML model training MnS consumer) (model training consumer for short). Alternatively, a model training producer may initiate ML model training, where the model training producer is the party responsible for training the model. Currently, the effect of improving the performance of the ML model obtained through ML model training is undesirable, and there is a waste of resources. [Prior art documents] [Non-patent literature]

[0006] [Non-Patent Document 1] 3GPP technical specification (TS) 28.554 version (version, V) 17.5.0 [Non-Patent Document 2] 3GPP TS 28.104 V0.3.0 [Non-Patent Document 3] technical report (technical report, TR) 28.809 V17.0.0 [Non-Patent Document 4] 3GPP TS28.532 V16.10.0 [Non-Patent Document 5] TS28.105 V0.1.0 Summary of the Invention

[0007] The present application provides a communication method and apparatus for machine learning model training, to improve the training effect of ML models and reduce resource waste.

[0008] According to a first aspect, the present application provides a communication method for machine learning model training. The method may be implemented by a first device or a component in the first device. The component may be, for example, at least one of a training execution module, a processor, a transceiver, a processing module, or a transceiver module. The first device may be an execution device for ML model training. For example, the communication method is implemented by the first device. The method may include: The first device receives first information from a second device. The first information includes at least one of a first indication and a training condition of the ML model. The value range of the first indication includes a first value and a second value, the first value indicates that the first device should perform training on the ML model, the second value indicates that the first device should not perform training on the ML model, and the value of the first indication is the first value. In addition, the first device further performs training on the ML model based on the first information.

[0009] According to the method, the first device may perform training on the ML model based on first information from the second device. The first information includes at least one of the first indication and the training condition of the ML model. Thus, the training on the ML model performed by the first device is based on the condition determined by the second device or indicated by the second device for starting the training on the ML model by the first device. Thus, it is possible to avoid the case where the first device starts the training on the ML model when the training on the ML model does not need to be performed, thereby reducing the waste of resources.

[0010] In a possible implementation, the first device may be a model training producer, and the second device may be a model training consumer. For example, the model training producer may include a domain management functional unit or a cross-domain management functional unit. For example, the model training consumer may include a domain management functional unit or a cross-domain management functional unit. It may be understood that the model training producer in this application may be an ML model training management service producer (ML model training MnS producer), and the model training consumer may be an ML model training management service consumer (ML model training MnS consumer).

[0011] In a possible implementation, the first device may further transmit the training results of the ML model to the second device.

[0012] The training result may include a trained ML model. Thus, the second device may obtain the trained ML model and perform problem analysis based on the trained ML model to improve analysis performance.

[0013] In a possible implementation, the first information includes a first indication. The first device may start training on the ML model and send a second indication to the second device. The second indication is used to notify the start of training on the ML model. Alternatively, the second indication indicates to the second device that it should determine whether to perform training on the ML model. Alternatively, the second indication indicates to the second device that the first device should determine whether to perform training on the ML model.

[0014] In this implementation, the first device may use the second indication to indicate that training should be started on the ML model (or indicate the start of training on the ML model), or the first device may indicate to the second device that it should determine whether to perform training on the ML model, or the first device may indicate to the second device that it should determine whether the first device should perform training on the ML model. The second device may determine whether the first device performs training on the ML model based on the second indication. If the first device determines to perform training on the ML model, the second device may send a first indication to the first device, the value of which is a first value. Correspondingly, the first device may receive the first indication from the second device, the value of which is a first value, and perform training on the ML model. Optionally, in this implementation, a case in which the first device starts training on the ML model when training on the ML model does not need to be performed can be avoided, thereby improving the training effect of the ML model and reducing resource waste. In addition, in this implementation, the second device can know that the first device starts training on the ML model based on the second indication. In this way, the case where training on the ML model is repeatedly started by the second device and training conflicts caused thereby can be avoided.

[0015] In a possible implementation, the first device may further send a third indication to the second device, where the third indication is used to request training control parameters of the ML model. The first device may further receive training control parameters of the ML model from the second device, where the training control parameters of the ML model are used by the first device to perform training on the ML model.

[0016] In this implementation, the first device may request the training control parameters of the ML model from the second device, and perform training on the ML model based on the training control parameters of the ML model, which may improve the training efficiency of the ML model.

[0017] In a possible implementation, the training control parameters for the ML model include at least one of an identifier for the ML model, type information of the ML model, a number of training iterations for the ML model, performance requirements for the ML model, a training time for the ML model, and data used to train the ML model.

[0018] According to this implementation, the first device performs training on the ML model based on the training control parameters of the ML model, so that the training effect of the ML model can be further improved, and the training efficiency of the ML model can be improved. Optionally, the training control parameters of the ML model may further include parameters such as psychological periodicity to improve the periodicity of training on the ML model.

[0019] In a possible implementation, the data used to train the ML model includes at least one of a data source, updated network states, and updated training data.

[0020] According to this implementation, the data used to train the ML model can be set more accurately, the training effect of the ML model can be further improved, and the training efficiency of the ML model can be improved.

[0021] In a possible implementation, the training conditions of the ML model include at least one of a network load condition, a network coverage condition, and a network performance condition. It should be understood that in this application, the training conditions of the ML model may also be referred to as a training policy of the ML model.

[0022] In this implementation, the second device can accurately and flexibly set the training conditions of the ML model. When the training conditions are met, the first device can start training on the ML model. In this way, the case where the first device starts training on the ML model when training on the ML model does not need to be performed is further avoided, thereby further reducing resource waste.

[0023] In a possible implementation, the network load conditions include at least one of a physical resource block (PRB) utilization threshold and a user connection threshold.

[0024] In this implementation, the network load condition can be set more accurately and flexibly, thus further avoiding the case where the first device starts training on the ML model when training on the ML model does not need to be performed, thereby further reducing resource waste.

[0025] In a possible implementation, the network coverage conditions include at least one of a reference signal received power (RSRP) threshold, a coverage rate threshold for the RSRP threshold in the area, a reference signal received quality (RSRQ) threshold, a coverage rate threshold for the RSRQ threshold in the area, a reference signal signal to interference plus noise ratio (RSSINR) threshold (sometimes referred to as signal to interference plus noise ratio for short), and a coverage rate threshold for the RSSINR threshold in the area.

[0026] In this implementation, the network coverage condition can be set more accurately and flexibly. In this way, the case where the first device starts training on the ML model when training on the ML model does not need to be performed is further avoided, thereby further reducing resource waste.

[0027] In a possible implementation, the network performance conditions include at least one of a handover success rate threshold, a data energy efficiency threshold, and a network slice energy efficiency threshold.

[0028] In this implementation, the network performance conditions can be set more accurately and flexibly, thus further avoiding the case where the first device starts training on the ML model when training on the ML model does not need to be performed, thereby further reducing resource waste.

[0029] In a possible implementation, the first device may further send training status information of the ML model to the second device, where the training status information indicates whether the ML model is trained, or the training status information indicates whether training for the ML model is performed based on at least one of the performance evaluation results of the ML model, the updated network state, and the updated training data.

[0030] In this implementation, the first device may indicate whether the ML model is trained or whether training on the ML model is performed based on the performance evaluation results, the updated network conditions, or the updated training data, so that the second device can know exactly whether the ML model is trained and / or whether training on the ML model has been started based on the performance evaluation results, the updated network conditions, or the updated training data. In this way, the second device can better determine whether the first device performs training on the ML model, and the case where the first device starts training on the ML model when training on the ML model does not need to be performed is further avoided, and repeated training is reduced. Optionally, if the second device determines that the first device performs training on the ML model, the second device may send a first indication to the first device, the value of which is a first value.

[0031] In a possible implementation, the first device may further send second information to the second device, where the second information includes at least one of information indicating the training progress of the ML model, an estimated training time for the training for the ML model, a training execution duration for the ML model, a training start time for the ML model, or a training end time for the ML model.

[0032] Based on this implementation, the first device may use the second information to indicate to the second device the training progress, estimated training time, execution duration, start time, or end time of the ML model, so that the second device can know the training status in a timely manner.

[0033] According to a second aspect, the present application provides a communication method for machine learning model training. The method may be implemented by a second device or a component in the second device. The component may include at least one of a training control module, a processor, a transceiver, a processing module, or a transceiver module. For example, the communication method is implemented by the second device. The method may include: The second device determines first information, the first information including at least one of a first indication and a condition for starting training on the ML model by the first device; A value range of the first indication includes a first value and a second value, the first value indicating that the first device should perform training on the ML model, the second value indicating that the first device should not perform training on the ML model, and the value of the first indication is the first value; The second device further transmits the first information to the first device.

[0034] In a possible implementation, the first device may be a model training producer, and the second device may be a model training consumer. For example, the model training producer may include a domain management functional unit or a cross-domain management functional unit. For example, the model training consumer may include a domain management functional unit or a cross-domain management functional unit. It may be understood that the model training producer in this application may be an ML model training management service producer, and the model training consumer may be an ML model training management service consumer.

[0035] In a possible implementation, the second device may further receive training results of the ML model from the first device.

[0036] In a possible implementation, the first information includes the first indication, and the second device may further receive a second indication from the first device, the second indication being used to notify the start of training on the ML model. Alternatively, the second indication indicates to the second device that the first device should determine whether to perform training on the ML model. Alternatively, the second indication indicates to the second device that the first device should determine whether to perform training on the ML model.

[0037] In a possible implementation, the second device may receive a third indication from the first device, where the third indication is used to request training control parameters of the ML model. The second device may further send the training control parameters of the ML model to the first device, where the training control parameters of the ML model are used by the first device to perform training on the ML model.

[0038] In a possible implementation, the training control parameters for the ML model include at least one of an identifier for the ML model, type information of the ML model, a number of training iterations for the ML model, performance requirements for the ML model, a training time for the ML model, and data used to train the ML model.

[0039] In a possible implementation, the data used to train the ML model includes at least one of a data source, updated network states, and updated training data.

[0040] In a possible implementation, the conditions for starting training on the ML model by the first device include at least one of a network load condition, a network coverage condition, and a network performance condition.

[0041] In a possible implementation, the network load conditions include at least one of a PRB utilization threshold and a user connection volume threshold.

[0042] In a possible implementation, the network coverage conditions include at least one of an RSRP threshold and a coverage rate threshold of the RSRP threshold in the area, an RSRQ threshold and a coverage rate threshold of the RSRQ threshold in the area, and an RSSINR threshold and a coverage rate threshold of the RSSINR threshold in the area.

[0043] In a possible implementation, the network performance conditions include at least one of a handover success rate threshold, a data energy efficiency threshold, and a network slice energy efficiency threshold.

[0044] In a possible implementation, the second device may further receive training status information of the ML model from the first device, where the training status information indicates whether the ML model is trained, or the training status information indicates whether training for the ML model is performed based on at least one of the performance evaluation results of the ML model, the updated network conditions, and the updated training data.

[0045] In a possible implementation, the second device may further receive second information from the first device, where the second information includes at least one of information indicating a training progress of the ML model, an estimated training time for training for the ML model, a training execution duration of the ML model, a training start time of the ML model, or a training end time of the ML model.

[0046] According to a third aspect, the present application provides a communication method for machine learning model training. The method may be implemented by a first device or a component in the first device. The component may be, for example, at least one of a training execution module, a processor, a transceiver, a processing module, or a transceiver module. The first device may be an execution device for ML model training. For example, the communication method is implemented by the first device. The method may include: The first device sends a machine learning ML training indication based on new data to the second device, and the ML training indication based on the new data indicates whether training should be performed based on updated training data of the ML model. Based on the method, the second device may determine whether the first device should perform training based on the updated training data of the ML model based on the machine learning ML training indication based on the new data, and provide the first device with information or parameters required for training in a timely manner to improve training accuracy.

[0047] In a possible implementation, the first device receives training control parameters for the ML model from the second device, and the training control parameters for the ML model are used by the first device to perform training on the ML model.

[0048] In a possible implementation, the first device may further send second information to the second device, where the second information includes at least one of information indicating the training progress of the ML model, an estimated training time for the training for the ML model, a training execution duration for the ML model, a training start time for the ML model, or a training end time for the ML model.

[0049] In a possible implementation, the information indicating the training progress of the ML model includes information indicating that training on the ML model is starting or information indicating that training on the ML model is ending.

[0050] In a possible implementation, the first device may further send training status information of the ML model to the second device, where the training status information indicates the progress of training for the ML model.

[0051] According to a fourth aspect, the present application provides a communication method for machine learning model training. The method may be implemented by a second device or a component in the second device. The component may include at least one of a training control module, a processor, a transceiver, a processing module, or a transceiver module. For example, the communication method is implemented by the second device. The method may include: the second device receives a machine learning ML training indication based on new data from the first device, where the ML training indication based on new data indicates whether training should be performed based on updated training data of the ML model.

[0052] In a possible implementation, the second device may further transmit training control parameters of the ML model to the first device, where the training control parameters of the ML model are used by the first device to perform training on the ML model.

[0053] In a possible implementation, the second device may further receive second information from the first device, where the second information includes at least one of information indicating a training progress of the ML model, an estimated training time for training for the ML model, a training execution duration of the ML model, a training start time of the ML model, or a training end time of the ML model.

[0054] In a possible implementation, the information indicating the training progress of the ML model includes information indicating that training on the ML model is starting or information indicating that training on the ML model is ending.

[0055] In a possible implementation, the second device may further receive training status information of the ML model from the first device, where the training status information indicates the progress of training for the ML model.

[0056] For advantageous effects of the fourth aspect and possible implementations of the fourth aspect, please refer to the description of the advantageous effects in the first aspect or the third aspect.

[0057] According to a fifth aspect, an embodiment of the present application provides a first device comprising a module configured to perform any aspect of the first aspect and steps in any possible implementation of the first aspect.

[0058] For example, the first device may include a transceiver module and a processing module. The transceiver module may be configured to support the first device in performing communication in at least one of a wired manner and a wireless manner. The communication may include transmitting or receiving signals, information, signaling, data, messages, and the like. The processing module may be configured to support the first device in performing processing acts. The processing acts include, but are not limited to, generating content transmitted by the transceiver module, processing content received by the transceiver module, or performing each processing act in the first or third aspect and any possible implementation of the first or third aspect.

[0059] According to a sixth aspect, an embodiment of the present application provides a second device comprising a module configured to perform any aspect of the second aspect and steps in any possible implementation of the second aspect.

[0060] For example, the second device may include a transceiver module and a processing module. The transceiver module may be configured to support the second device in performing communication in at least one of a wired manner and a wireless manner. The communication may include transmitting or receiving signals, information, signaling, data, messages, and the like. The processing module may be configured to support the second device in performing processing acts. The processing acts include, but are not limited to, generating content transmitted by the transceiver module, processing content received by the transceiver module, or performing each processing act in the second or fourth aspect and possible implementations of the second or fourth aspect.

[0061] According to a seventh aspect, an embodiment of the present application provides a communication device, including at least one processing element (e.g., a processor or chip) and at least one storage element (e.g., a memory or storage unit). The at least one storage element is configured to store programs and data, and the at least one processing element is configured to implement the methods provided in the first aspect and the second aspect and any possible design of the first aspect and the second aspect. Alternatively, the communication device may include modules, such as a transceiver module and a processing module, configured to implement steps in any aspect of the first aspect and any possible implementation of the first aspect. The transceiver module may be configured to support the communication device in implementing communication in at least one of a wired manner and a wireless manner. The communication may include transmission or reception of signals, information, signaling, data, messages, etc. The processing module may be configured to support the communication device in implementing the processing acts. The processing act includes, but is not limited to, generating content to be transmitted by the transceiver module, processing content received by the transceiver module, or performing each processing act in any one of the first to fourth aspects and possible implementations of the first to fourth aspects.

[0062] According to an eighth aspect, an embodiment of the present application provides a communication method for machine learning model training. The communication method may be implemented by the first device or a component in the first device in the first aspect and any possible design of the first aspect, and the second device or a component in the second device in the second aspect and any possible design of the second aspect. The communication method may include: The second device transmits first information to the first device, the first information including at least one of a first indication and a condition for starting training on the ML model by the first device. The value range of the first indication includes a first value and a second value, the first value indicating that the first device should perform training on the ML model, the second value indicating that the first device should not perform training on the ML model, and the value of the first indication is the first value. The first device may perform training on the ML model based on the first information.

[0063] In a possible implementation, the first device may be a model training producer, and the second device may be a model training consumer. For example, the model training producer may include a domain management functional unit or a cross-domain management functional unit. For example, the model training consumer may include a domain management functional unit or a cross-domain management functional unit. It may be understood that the model training producer in this application may be an ML model training management service producer, and the model training consumer may be an ML model training management service consumer.

[0064] In a possible implementation, the first device may further transmit the training results of the ML model to the second device. Correspondingly, the second device may receive the training results of the ML model from the first device.

[0065] In a possible implementation, the first information may include a first indication that the first device has started training the ML model and may send a second indication to the second device. In response, the second device may receive a second indication from the first device.

[0066] In a possible implementation, the first device may further send a third indication to the second device, where the third indication is used to request training control parameters of the ML model. Correspondingly, the second device may receive the third indication from the first device. The second device may further send training control parameters of the ML model to the first device, where the training control parameters of the ML model are used by the first device to perform training on the ML model. Correspondingly, the first device may receive the training control parameters of the ML model from the second device.

[0067] In a possible implementation, the training control parameters for the ML model include at least one of an identifier for the ML model, type information of the ML model, a number of training iterations for the ML model, performance requirements for the ML model, a training time for the ML model, and data used to train the ML model.

[0068] In a possible implementation, the data used to train the ML model includes at least one of a data source, updated network states, and updated training data.

[0069] In a possible implementation, the training conditions of the ML model include at least one of a network load condition, a network coverage condition, and a network performance condition. It should be understood that in this application, the training conditions of the ML model may also be referred to as a training policy of the ML model.

[0070] In a possible implementation, the network load conditions include at least one of a PRB utilization threshold and a user connection volume threshold.

[0071] In a possible implementation, the network coverage conditions include at least one of an RSRP threshold and a coverage rate threshold of the RSRP threshold in the area, an RSRQ threshold and a coverage rate threshold of the RSRQ threshold in the area, and an RSSINR threshold and a coverage rate threshold of the RSSINR threshold in the area.

[0072] In a possible implementation, the network performance conditions include at least one of a handover success rate threshold, a data energy efficiency threshold, and a network slice energy efficiency threshold.

[0073] In a possible implementation, the first device may further send training status information of the ML model to the second device, where the training status information indicates whether the ML model is trained, or the training status information indicates whether training for the ML model is performed based on at least one of the performance evaluation result of the ML model, the updated network state, and the updated training data. Correspondingly, the second device may receive the training status information of the ML model from the first device.

[0074] In a possible implementation, the first device may further send second information to the second device, where the second information includes at least one of information indicating the training progress of the ML model, an estimated training time for the training for the ML model, a training execution duration for the ML model, a training start time for the ML model, or a training end time for the ML model. Correspondingly, the second device may receive the second information from the first device.

[0075] Optionally, the method shown in the sixth aspect may further include a method provided in any possible implementation of the first aspect. The method may further include a method provided in any possible implementation of the second aspect.

[0076] Alternatively, the method shown in the eighth aspect may be implemented by a first device or components in the first device in the third aspect and any one of the possible designs of the third aspect, and a second device or components in the second device in the fourth aspect and any one of the possible designs of the fourth aspect. Thus, the method may include acts implemented by a first device in the third aspect and any one of the possible designs of the third aspect, and acts implemented by a second device in the fourth aspect and any one of the possible designs of the fourth aspect.

[0077] According to a ninth aspect, an embodiment of the present application further provides a computer program product, comprising a computer program, which when run on a computer enables the computer to perform the method provided in any one of the first to fourth aspects and possible designs of the first to fourth aspects.

[0078] According to a tenth aspect, an embodiment of the present application further provides a computer readable storage medium, the computer readable storage medium storing a computer program, which when executed by a computer enables the computer to perform the method provided in any one of the first to fourth aspects and possible designs of the first to fourth aspects.

[0079] According to an eleventh aspect, an embodiment of the present application further provides a chip, the chip being configured to read a computer program stored in a memory to perform the method provided in any one of the first to fourth aspects and possible designs of the first to fourth aspects.

[0080] According to a twelfth aspect, an embodiment of the present application further provides a chip system. The chip system includes a processor configured to support a computer device in implementing the method provided in any one of the first to fourth aspects and possible designs of the first to fourth aspects. In a possible design, the chip system further includes a memory, the memory configured to store programs and data required for the computer device. The chip system may include a chip, or may include a chip and another individual component.

[0081] According to a thirteenth aspect, an embodiment of the present application further provides a communication system. The communication system may include a first device provided in any one of the first aspect and possible designs of the first aspect, and a second device provided in any one of the fourth aspect and possible designs of the fourth aspect. The first device may be implemented by using an apparatus provided in the fifth aspect or the seventh aspect. The second device may be implemented by using an apparatus provided in the sixth aspect or the seventh aspect.

[0082] For advantageous effects of the steps provided in the second to thirteenth aspects and in any one of the possible designs of the second to thirteenth aspects, please refer to the advantageous training of an ML model based on a first information effect of the steps provided in the first aspect and in the possible designs of the first aspect. [Brief description of the drawings]

[0083] [Figure 1] 1 is a schematic diagram of the architecture of a service system according to the present application; [Diagram 2] FIG. 2 is a schematic diagram of the architecture of another service system according to the present application. [Diagram 3] FIG. 2 is a schematic diagram of the architecture of another service system according to the present application. [Figure 4]FIG. 1 is a schematic diagram of an architecture for training an ML model according to the present application. [Diagram 5] FIG. 2 is another schematic diagram of an architecture for training an ML model according to the present application. [Figure 6] 1 is a schematic flow chart of a communication method for machine learning model training according to the present application. [Figure 7] 1 is a schematic flow chart of another communication method for machine learning model training according to the present application. [Figure 8] 1 is a schematic flow chart of another communication method for machine learning model training according to the present application. [Figure 9] 1 is a schematic flow chart of another communication method for machine learning model training according to the present application. [Figure 10] 1 is a schematic diagram of an interaction scheme between communication devices according to the present application; [Figure 11] 2 is a schematic diagram of another interaction scheme between communication devices according to the present application; [Figure 12] 2 is a schematic diagram of another interaction scheme between communication devices according to the present application; [Figure 13] 1 is a schematic diagram of the structure of a communication device according to the present application; [Figure 14] 2 is a schematic diagram of the structure of another communication device according to the present application; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0084] Hereinafter, embodiments of the present application will be described in detail with reference to the accompanying drawings herein.

[0085] The embodiments of the present application may be applied to various mobile communication systems, such as new radio (NR) systems, long term evolution (LTE) systems, advanced long term evolution (LTE-A) systems, future communication systems, and other communication systems. In particular, this is not limited herein. For example, the embodiments in the present application may be applied to an NR network management architecture. The NR network management architecture may include a management function (MnF). The MnF is a management entity defined in the 3rd generation partnership project (3GPP). The externally visible behavior and interface of the MnF is defined as a management service (MnS). In the service provision management architecture, the MnF acts as an MnS producer or an MnS consumer. A management service produced by an MnS producer of the MnF may have multiple MnS consumers. An MnF may consume multiple management services from one or more management service producers. As shown in FIG. 1, an MnS provided by an MnF may be used to provide a service to another MnF (e.g., shown as MnF#A). In this case, the MnF may act as an MnS producer, and the MnF#A may act as an MnS consumer. In addition, the MnF may further obtain an MnS service provided by MnF#B from another MnF (shown as MnF#B, which may be the same as or different from MnF#A). In other words, in this case, the MnF shown in FIG. 1 may also act as an MnS consumer of an MnS service provided by MnF#B. In other words, the same MnF may act as an MnS service consumer and an MnS service producer.

[0086] 2 is a schematic diagram of a service-based management architecture to which the method embodiments of the present application are applicable. The service-based management architecture includes a business support system (BSS), a cross-domain management function (CD-MnF) unit, a domain management function (Domain-MnF) unit, and a network element (NE). It should be understood that the cross-domain management function unit may be a node such as a network management system (NMS), an MnS producer, or an MnS consumer, and the domain management function unit may be a node such as a wireless automation engine (MBB automation engine, MAE), an element management system (EMS), an MnS producer, or an MnS consumer.

[0087] In this application, the cross-domain management function unit is configured to manage one or more domain management function units. The domain management function unit may be configured to manage one or more network elements. Each unit is briefly described below.

[0088] If the management service is a management service provided by a cross-domain management functional unit, the cross-domain management functional unit is a management service producer and the business support system is a management service consumer.

[0089] If the management service is a management service provided by a domain management functional unit, the domain management functional unit is a management service producer and the cross-domain management functional unit is a management service consumer.

[0090] If the management service is a management service provided by a network element, the network element is the management service producer and the domain management functional unit is the management service consumer.

[0091] The business support system is adapted to communication services and configured to provide functions and management services such as billing, settlement, accounting, customer service, business, network monitoring, communication service lifecycle management, and service intent transformation. The business support system may be an operator operation system or a vertical industry operation system (vertical OT system).

[0092] The cross-domain management function unit may also be referred to as a network management function (NMF) unit and may be, for example, a network management entity such as an NMS or a network function management service consumer (NFMS_C). The cross-domain management function unit provides one or more of the following management functions or services: network life cycle management, network deployment, network fault management, network performance management, network configuration management, network assurance, network optimization function, translation of network intent from communication service producer (Intent-CSP), and the like.

[0093] The network represented by the management function or management service may include one or more network elements or subnets, or may be a network slice. In other words, the network management function unit may be a network slice management function (NSMF) unit, a cross-domain management data analytics function (MDAF) unit, a cross-domain self-organization network function (SON Function) unit, or a cross-domain intent management function (intent-driven management service, intent-driven MnS) unit.

[0094] In an embodiment of the present application, a network element is an entity that provides a network service. The network element may include a core network element, a radio access network element, a transport network element, etc. For example, in the architecture shown in Figure 2, the domain management functional unit may include a radio access network domain management functional unit, a core network element domain management functional unit, or a transport network domain management functional unit. The radio access network domain management functional unit may be configured to manage the radio access network element, the core network element domain management functional unit may be configured to manage the core network element, and the transport network domain management functional unit may be configured to manage the transport network element.

[0095] For example, core network elements may include, but are not limited to, an access and mobility management function (AMF) entity, a session management function (SMF) entity, a policy control function (PCF) entity, a network data analytics function (NWDAF) entity, a network repository function (NRF), gateways, etc.

[0096] The radio access network elements may include, but are not limited to, various types of base stations (e.g., generation NodeB (gNB) and evolved NodeB (eNB)), a central unit control plane (CUCP), a central unit (CU), a distributed unit (DU), a central unit user plane (CUUP), etc. In this application, the network function NF is also referred to as a network element NE.

[0097] Optionally, in some deployment scenarios, the cross-domain management functional unit may further provide subnet life cycle management, subnet deployment, subnet fault management, subnet performance management, subnet configuration management, subnet guarantee, subnet optimization function, translation of network intent from communication service provider of subnet, translation of network intent from communication service consumer of subnet (Intent-CSC), etc. A subnet in this specification may include multiple small subnets and may be a network slice subnet.

[0098] The domain management functional unit may also be referred to as an NMF or an element management functional unit. For example, the domain management functional unit may be an element management entity such as an MAE, an EMS, or a network function management service provider (NFMS_P).

[0099] The domain management functional unit provides one or more of the following functions or management services: life cycle management of a subnet or network element, deployment of a subnet or network element, fault management of a subnet or network element, performance management of a subnet or network element, assurance of a subnet or network element, optimization function of a subnet or network element, conversion of intent (intent from a network operator, Intent-NOP) from a subnet or network element, etc. A subnet in this specification includes one or more network elements. A subnet may also include a subnet, in other words, one or more subnets form a larger subnet.

[0100] It should be understood that an intent in this application may be an expectation of an intent producer (e.g., a network element) or a system (e.g., a network or a subnet) in which the intent producer is located, and may include requirements, targets, constraints, etc. The translation of intent is the process of determining a policy of the intent. For example, the policy may indicate a condition that does not satisfy the intent. For example, when the intent is energy saving, policy A may be that the energy consumption is abnormal (i.e., non-energy saving) when the energy consumption is greater than a first threshold, and policy B may be that the energy consumption is abnormal (i.e., non-energy saving) when the energy consumption is greater than a second threshold. It may be understood that even in the case of the same intent, the intent can be satisfied and the determined solution by using different policies may be different.

[0101] Optionally, the subnet in this specification may alternatively be a network slice subnet. The domain management system may be a network slice subnet management function (NSSMF) unit, a domain management data analytics function (MDAF) unit, a domain self-organization network function (SON Function), a domain intent management function unit, etc.

[0102] The domain management functional units can be classified in the following manner:

[0103] According to the network type, the domain management function unit can be classified into the following types: radio access network (RAN) domain management function (RAN domain MnF) unit, core network domain management function (CN domain MnF) unit, and transport network domain management function (TN domain MnF) unit. It should be noted that the domain management function unit may also be a domain network management system and may manage one or more of an access network, a core network, or a transport network.

[0104] According to administrative regions, the domain management functional units may be classified as regional domain management functional units, for example, city A domain management functional units and city B domain management functional units.

[0105] A network element is an entity that provides network services and includes a core network element, an access network element, etc. The core network element includes an access and mobility management function (AMF), a session management function (SMF), a policy control function (PCF), a network data analytics function (NWDAF), a network repository unit (NF repository function, NRF), a gateway, etc. The access network element includes a base station (e.g., gNB or eNB), a central unit control plane (CUCP), a central unit (CU), a distribution unit (DU), a central unit user plane (CUUP), etc.

[0106] The network element may provide one or more of the following management functions or services: network element life cycle management, network element deployment, network element fault management, network element performance management, network element assurance, network element optimization functions, intent translation from the network element, and the like.

[0107] For example, as shown in Figure 3, in an interaction between a 3GPP cross-domain MDA MnS producer and a RAN domain MDA, the cross-domain MDA MnS producer may consume a RAN domain MDA MnS. Alternatively, the cross-domain MDA MnS producer may consume an MDA MnS provided by the RAN domain and produce an MDA MnS that can be consumed by a 3GPP cross-domain MDA MnS consumer.

[0108] FIG. 4 shows an example of a schematic diagram of an architecture of an ML model-based MDA system. As shown in FIG. 4, the system may include a management function and an MDAS consumer. The management function may include an MDAS producer, or the management function may be implemented by the MDAS producer. In an embodiment of the present application, the MDAS producer may also be referred to as an MDA MnS producer or an MnS producer, and the MDAS consumer may also be referred to as an MDA MnS consumer or an MnS consumer. In the present application, for an MDAS, the MDAS consumer is an MDAS requester device, and the MDAS producer is an MDAS provider device.

[0109] In the architecture shown in Figure 4, the management function may collect and provide MDA analytics input to the MDAS producer. The MDAS consumer may request MDA reports from the MDAS producer and control the MDAS producer's MDA reporting. The MDAS consumer controls the MDAS producer's MDA reporting. For example, the MDAS consumer may send control parameters to the MDAS producer that are used to obtain the MDA reports.

[0110] The management function (or MDAS producer) may perform inference on the analytics input of the MDA based on the ML model and generate analytics / inference output. The ML model may be trained. The management function (or MDAS producer) may further generate MDA reports based on the analytics / inference output and provide the MDA reports to the MDAS consumer.

[0111] It should be appreciated that the lines depicted in FIGS. 3 and 4, including diagrams including circles and arcs, may represent service-based interfaces.

[0112] Currently, training on an ML model can be triggered in two ways. It should be understood that triggering in this application may be replaced with starting, commencing, or beginning to perform.

[0113] Currently, the manner of triggering training for an ML model is that a model training consumer triggers training. It should be understood that the model training consumer and the model training producer may be deployed in a cross-domain management functional unit and / or a domain management functional unit. For example, the model training producer may include a domain management functional unit or a cross-domain management functional unit, and the model training consumer may include a domain management functional unit or a cross-domain management functional unit. In addition, the model training consumer includes a machine learning model training (ML model training, MLMT) management service consumer (MLMT MnS consumer). The MDAS producer used to perform training on the ML model may be referred to as a model training producer, for example, a machine learning model training management service producer (MLMT MnS producer). For example, in an MDAS scenario, the model training consumer may include an MDAS consumer, and the model training producer may include an MDAS producer. It should be understood that model training in this application may include training for an AI / ML model or training for an application (APP), or in other words, an AI / ML model may include an application. AI / ML model training may also be referred to as AI / ML training, AI / ML application training, AI / ML APP training, etc.

[0114] It may be understood that the model training producer in this application may be an ML model training management service producer, and the model training consumer may be an ML model training management service consumer.

[0115] In this triggering scheme, training on an ML model may be triggered by an ML model training request from one or more model training consumers, as shown in Figure 5. Optionally, the model training consumer may provide a data source as training data used to perform training on the ML model. The model training producer may return a response to the model training consumer to inform the model training consumer whether the training request has been accepted.

[0116] In another triggering scheme, the model training producer triggers training on the ML model based on feedback received from the model training consumer or performance evaluation results of the ML model, or the model training producer triggers training on the ML model when there is an updated network state or updated training data is available. In addition, as shown in Figure 5, when the MDAMnS producer completes training on the ML model, the MDAMnS producer may send the training result of the ML model to the MDAMnS consumer. In this application, the training result of the ML model may also be referred to as the training result for short.

[0117] In this application, the feedback, or what is referred to as a feedback report, may include the ML model type and the ML model fitness, or indicate whether a trained ML model is used. For example, the ML model fitness included in the feedback may indicate a fitness determined in a process that uses the ML model. If the ML model fitness is not high (e.g., lower than a fitness threshold), the model training producer may start training on the ML model.

[0118] It should be appreciated that the precision of an ML model may be defined as the amount of true positives (TP) divided by the sum of the amount of true positives and the amount of false positives (FP).

[0119] Regarding the interpretation of the amount of true positives and the amount of false positives, the true positives are the amount of times that positive data in the test data are identified as positive by the ML model, and the false positives are the amount of times that the ML model erroneously identifies negative data in the test data as positive. For example, the precision is 90%. Alternatively, the precision may be expressed by using a decimal. For example, the precision is 0.9. Positives are types of identification results among various types of possible identification results of the ML model, and another type of identification result is assumed to be negative. Positive data are data for which the identification result should be positive, and positive data are data for which the identification result should be negative. For example, positive integers and non-positive integers are two types of identification results of the ML model. Positive integers are positive, non-positive integers are negative, and it is assumed that positive data includes, for example, 1, and negative data includes, for example, 0. When the ML model identifies 1 as a positive integer, the amount of true positives is increased by 1. When the ML model identifies 0 as a positive integer, the amount of false positives is increased by 1.

[0120] In another description of the amount of true positives and the amount of false positives, the ML model may be configured to identify whether data meets a requirement. If a piece of data to be identified should be identified as meeting the requirement, and in the process of identifying data by the ML model, the data is actually identified as meeting the requirement, the amount of true positives of the ML model is increased by 1. If a piece of data to be identified should be identified as meeting the requirement, but in the process of identifying data by the ML model, the data is actually identified as not meeting the requirement, the amount of false positives of the ML model is increased by 1.

[0121] The performance evaluation result of the ML model may be used to measure the performance of the ML model. If the performance evaluation result of the ML model indicates that the ML model performs poorly (e.g., the time consumed for each calculation of the ML model exceeds a requirement or the calculation error in the process using the ML model exceeds an error threshold), the model training producer may trigger training for the ML model. Optionally, the feedback may also include the performance evaluation result of the ML model.

[0122] The updated network state may be a changed network state. It may be understood that the change in the network state may change the training data used to perform training on the ML model. Thus, when the network state is updated, the model training producer may trigger training on the ML model. In this application, the updated network state may also be referred to as a new network state. The updated network state may include, for example, some cells entering an energy saving state or the transmit power of some cells being adjusted.

[0123] The updated training data may be data in which parameters such as data distribution and data value change, or data that changes due to changes in network conditions, etc. In this application, the updated training data may also be referred to as new training data. The updated training data may be data obtained from an external server such as a 5G core network, or an application program such as another external application (APP). This is not particularly limited in this application.

[0124] Currently, the effect of improving the performance of the ML model obtained through training on the ML model is not good, and there is resource waste. For example, when a model training producer starts training on an ML model, the model training producer may start training on the ML model when training on the ML model does not need to be performed. For example, if the performance of the ML model can still meet the usage requirements of a model training consumer, training on the ML model does not need to be performed. In this case, when a model training producer starts training on an ML model, the effect of improving the performance of the ML model obtained through training is not ideal, and there is resource waste.

[0125] In order to avoid a case where a first device performs training on an ML model when training on an ML model does not need to be performed and to reduce resource waste, an embodiment of the present application provides a communication method for machine learning model training.

[0126] The communication method may be implemented by a first device and a second device. The first device may be a model training producer or a training execution module in the model training producer, and the second device may be a model training consumer or a training control module in the model training consumer. For example, the model training producer may include an MLMT MnS producer, and the model training consumer may include an MLMT MnS consumer. In particular, the MLMT MnS producer may be an MDAS producer, and the MLMT MnS consumer may be an MDAS consumer.

[0127] Optionally, the first device in the present application may include a domain management functional unit such as a RAN domain MnF, and the second device may include a cross-domain management functional unit. The MLMT MnS consumer and / or the MDAS consumer may be located in the domain management functional unit or the cross-domain management functional unit. The MLMT MnS producer and / or the MDAS producer may be located in the domain management functional unit or the cross-domain management functional unit.

[0128] As shown in FIG. 6, the communication method for machine learning model training provided in an embodiment of the present application may include steps shown in S101 to S103.

[0129] S101: A second device determines first information, where the first information includes at least one of a first indication and a training condition of the ML model.

[0130] The value range of the first indication includes a first value and a second value, the first value indicating that the first device should perform training on the ML model and the second value indicating to the first device that training should not be performed on the ML model, and the value of the first indication is the first value.

[0131] In this application, the training conditions of the ML model may also be referred to (or may be understood as) the training policy of the ML model. The training policy of the ML model may include a condition for starting training for the ML model. When the training policy of the ML model is satisfied, the first device may start training for the ML model. Satisfying the training policy of the ML model may be understood as the condition for starting training for the ML model and included in the training policy of the ML model being satisfied.

[0132] In the present application, subsequently, an embodiment in which the first information includes a first indication and an embodiment in which the first information includes a training policy for the ML model will be described separately with reference to respective flowcharts, and details will not be temporarily described in this specification.

[0133] S102: The second device transmits first information to the first device.

[0134] In response, the first device receives the first information from the second device.

[0135] S103: The first device trains the ML model based on the first information.

[0136] In S103, when the first information includes a first indication, the first device performing training on the ML model based on the first information may be understood as the first device starting or continuing training on the ML model after the first device receives the first indication whose value is a first value. When the first device receives the first indication whose value is a second value, the first device may stop, cancel, or no longer perform training on the ML model. When the first information includes a training policy for the ML model, the first device performing training on the ML model based on the first information may be understood as the first device starting or performing training on the ML model when the training policy is satisfied.

[0137] It should be noted that the algorithm used by the first device to perform training on the ML model includes, but is not limited to, a simulated annealing (SA) algorithm, a tabu search (TS) algorithm, or another algorithm that can be used for model training, which is not particularly limited in this application.

[0138] In this embodiment of the present application, the first device may perform training on the ML model based on at least one of the first indication received from the second device and the training policy of the ML model. Thus, the training on the ML model performed by the first device is determined by the second device or is of the ML model and is performed according to the training policy indicated by the second device. Thus, when training on the ML model does not need to be performed, the case where the first device performs training on the ML model can be avoided, and resource waste can be reduced. It should be understood that reducing resource waste in the present application may also be described as saving training resources, etc. This is not particularly limited.

[0139] In the implementation of the first information in S102, the first information may include a first indication. When the value of the first indication is a first value, it indicates that the second device determines that the first device can perform training on the ML model. When the value of the first indication information is a second value, it indicates that the second device determines that the first device does not perform training on the ML model. It may be known that the first device may determine whether training on the ML model needs to be performed based on the first indication. For example, when the second device considers that training on the ML model does not need to be performed based on the current energy consumption or based on the performance of the ML model, the second device sends a first indication whose value is a second value. Correspondingly, the first device may not perform training on the ML model. Alternatively, when the second device determines that the current fit rate of the ML model is insufficient, the second device sends a first indication whose value is a first value to enable the first device to perform training on the ML model. Correspondingly, after receiving the first indication whose value is the first value, the first device may determine that the second device has determined that the first device will perform training on the ML model, or after receiving the first indication whose value is the first value, the first device may determine to perform training on the ML model.

[0140] For example, the first indication may be indication information such as "yes" or "no". For example, the bit value of the indication information "yes" is "1", in other words, the first value is "1", and the bit value of the indication information "no" is "0", in other words, the second value is "1". When the bit value is "1", it indicates "yes". When the bit value is "0", it indicates "no". Alternatively, the first value may include another indication indicating "yes" that is not "1", and the second value may include another indication indicating "no" that is not "0". This is not particularly limited in the present application.

[0141] In a possible example, the first indication in the present application does not belong to or include a request to train an ML model sent by the second device to the first device.

[0142] In a possible example, the first indication in S102, whose value is a first value, is a positive response made by the second device to the second indication from the first device. The second indication can be used to notify the ML model to start training, or in other words, the second indication can be used to notify the start of training for the ML model. It should be understood that the second indication in this application can further indicate to the second device that the first device should determine whether training for the ML model is performed by the first device. In another example, the second indication indicates to the second device that it should determine whether training for the ML model is performed by the first device. After receiving the second indication, the second device can determine whether training for the ML model is performed by the first device, or the second device can determine that the value of the first indication is the first value or the second value.

[0143] Therefore, in the present application, the first device may use the second indication to indicate to the second device that training should be started on the ML model, or to indicate to the second device that the first device should determine whether to perform training on the ML model. This can avoid the case where the first device starts training on the ML model when training on the ML model does not need to be performed, and reduce resource waste. In addition, based on the second indication, the second device can know about the training of the ML model started by the first device, and avoid training conflicts caused when the second device repeatedly requests the first device to start training on the ML model.

[0144] Optionally, the first device may send a second indication after starting training on the ML model. In the present application, the first device starting training on the ML model may include the following cases:

[0145] Case 1: The first device may start training on an ML model based on feedback received from the second device or performance evaluation results of the ML model, or the first device may start training on an ML model based on updated network conditions or updated training data.

[0146] For example, the second device may obtain feedback or performance evaluation results of the ML model by using the ML model provided by the first device. The first device may further send the feedback or performance evaluation results of the ML model to the first device to enable the first device to start training on the ML model based on the feedback or performance evaluation results of the ML model. For example, if the feedback from the second device indicates that the fitting rate of the ML model is not high, the first device may start training on the ML model. In another example, if the performance evaluation results of the ML model from the second device indicate that the performance of the ML model is poor, the first device may start training on the ML model.

[0147] In addition, when the first device obtains updated network conditions and / or updated training data, the first device may further start training the ML model based on the updated network conditions and / or the updated training data.

[0148] Case 2: The first device may receive the training policy of the ML model from the second device, and start training for the ML model after determining that the training policy of the ML model is satisfied. The training policy of the ML model may be from the second device. For the training policy of the ML model, please refer to the description of the training policy of the ML model in S101 and S301 of this application. This is not limited in this application.

[0149] In another implementation of the first information in S101, when the first information includes a training policy of the ML model, the training policy of the ML model may include at least one of a network load condition, a network coverage condition, and a network performance condition. In other words, the condition for starting training for the ML model may include at least one of a network load condition, a network coverage condition, and a network performance condition. Based on this implementation, in the present application, the training condition of the ML model can be set accurately and flexibly to further reduce resource waste.

[0150] The network load condition is related to a current network load. When the current network load satisfies the network load condition, the first device may start training the ML model. In other words, in this case, the first device may perform training on the ML model.

[0151] For example, the network load condition may include at least one of a PRB utilization threshold and a user connection amount threshold. The PRB utilization threshold is, for example, a percentage of PRB utilization in an area (which may be referred to as an analysis area). If the percentage of PRB utilization in the area exceeds or reaches the threshold, the first device may start training the ML model. It may be understood that the analysis area in this application may include a specific area (e.g., a geographic location area) and an area in which a specific cell or a specific base station is located, or may include multiple specific areas and an area in which a specific cell or multiple specific base stations are located. This is not particularly limited in this application. It should be understood that the PRB utilization threshold may be a threshold applicable to a specific area, a specific cell, or a specific base station, or may be a threshold applicable to multiple areas, multiple cells, and multiple base stations. In other words, the training granularity of the ML model may be area granularity, cell granularity, or base station granularity. This is not limited in this application. In this example, the training policy of the ML model may include at least one of a physical resource block utilization threshold and a user connection amount threshold.

[0152] In this application, the training granularity of an ML model is the effective scope of a trained ML model, and the trained ML model may be configured to provide analytics services within the effective scope. For example, area granularity, cell granularity, and base station granularity respectively mean that a trained ML model may be configured to provide analytics services in one or more areas, one or more cells, or the coverage area of ​​one or more base stations.

[0153] The network load condition may further include a user connection amount threshold. When the user connection amount in the area exceeds or reaches the user connection amount threshold, the first device may start training for the ML model. It should be understood that the user connection amount threshold may be a threshold applicable to a specific area, a specific cell, or a specific base station, or may be a threshold applicable to multiple areas, multiple cells, and multiple base stations. In other words, the training granularity may be area granularity, cell granularity, or base station granularity. This is not limited in the present application.

[0154] The network coverage condition relates to the coverage rate that the signal measurement is reached and is in the area. The coverage rate is the ratio of the range in the area where the signal quality meets the signal quality threshold to the total range of the area. The signal measurement may include RSRP, RSRQ, and RSSINR. For example, the network coverage condition in this application may include at least one of the following: RSRP threshold and coverage rate threshold of RSRP threshold in the area, RSRQ threshold and coverage rate threshold of RSRQ threshold in the area, or RSSINR threshold and coverage rate threshold of RSSINR threshold in the area.

[0155] For example, if the signal measurement threshold is reached and the coverage rate of the one in the area reaches (or exceeds) the coverage rate threshold corresponding to the signal measurement, training for the ML model does not need to be performed, and the first device does not start training for the ML model. If the coverage rate of the one in the area reaching the signal measurement threshold does not reach (or exceed) the coverage rate threshold of the signal measurement, training for the ML model needs to be performed, and the first device may start training for the ML model. For example, the signal measurement is RSRP, the RSRP threshold is -110db, and the coverage rate threshold corresponding to the RSRP threshold is 80%. If the one in the area and the range where the measured RSRP reaches the RSRP threshold (for example, the range may be the amount of measurement, the amount of users, or the amount of grid boxes, which is not limited herein in the present invention) occupies 82% of the total range of the area, the first device may determine that training for the ML model does not need to be performed, and the first device does not start training for the ML model. The grid box volume refers to the volume of the grid box where the signal measure in the grid box reaches the grid box signal measure threshold value after the area is divided into grid boxes with equal areas.

[0156] If the range where the measured RSRP reaches the RSRP threshold in another area having the same RSRP threshold and the same RSRP coverage rate threshold is 75% of the total range of the area, the first device may determine that training of the ML model needs to be performed. It should be understood that the manner in which the first device acquires signal measurements such as RSRP is not limited in this application.

[0157] It should be understood that the signal measurement threshold and the signal measurement coverage rate may be indicated separately or may be indicated by using the same signaling, message, or information.

[0158] The network performance condition is related to the network performance. If the current network performance is poor, the first device may perform training on the ML model to optimize the network performance by training the ML model. If the current network performance is good, the first device does not need to start training on the ML model.

[0159] For example, the network performance conditions may include at least one of a handover success rate threshold, a data energy efficiency threshold, and a network slice energy efficiency threshold.

[0160] In this application, the handover success rate is also referred to as the radio access network handover success rate, which indicates the handover success rate of the handover between base stations and for the user equipment (UE) in the radio access network. When the handover success rate does not reach or exceed the handover success rate threshold, the first device may perform training on the ML model to optimize the handover success rate by training the ML model. When the handover success rate reaches or exceeds the handover success rate threshold, the first device does not need to perform training on the ML model.

[0161] The data energy efficiency may include an indicator of data energy efficiency of the mobile network in the radio access network. When the data energy efficiency does not reach or exceed the data energy efficiency threshold, the first device may perform training on the ML model to optimize the data energy efficiency by training the ML model. When the data energy efficiency reaches or exceeds the data energy efficiency threshold, the first device does not need to perform training on the ML model.

[0162] Network slice energy efficiency is related to network slice performance and network slice energy consumption. When the network slice energy efficiency does not reach or exceed the network slice energy efficiency threshold, the first device may perform training on the ML model to optimize the network slice energy efficiency by training the ML model. When the network slice energy efficiency reaches or exceeds the data energy efficiency threshold, the first device does not need to perform training on the ML model.

[0163] It should be understood that for the definition and determination manner of handover success rate, data energy efficiency, and network slice energy efficiency in this application, please refer to the description in Chapter 6 of Non-Patent Document 1. In addition, one or more of the handover success rate, data energy efficiency, and network slice energy efficiency may be replaced with another indicator used to measure network performance and included in Non-Patent Document 1. In other words, the network performance condition may further include another indicator used to measure network performance and included in Non-Patent Document 1 that is not the handover success rate, data energy efficiency, and network slice energy efficiency.

[0164] Therefore, in each embodiment of the present application, the network load conditions, network coverage conditions, and network performance conditions can be set more accurately and flexibly to further avoid the case where the first device starts training on an ML model when training on an ML model does not need to be performed, and the waste of resources can be further reduced.

[0165] Optionally, the training policy corresponds to the ML model. In other words, the training policy is determined for the ML model, or in other words, different ML models may have different training policies. In particular, the second device may determine the training policy of the ML model. For example, when the ML model is used in different scenarios, the training policy of the ML model is different. In an example, in a coverage problem analysis scenario, the training policy of the ML model includes at least a network coverage condition. In another scenario, the network coverage condition is an optional condition for starting training on the ML model by the first device.

[0166] 7 shows a procedure of a communication method for training a machine learning model according to an embodiment of the present application. In the procedure, a first device receives a first indication from a second device and performs training on an ML model based on the first indication. It can be understood that the procedure shown in FIG. 7 is an implementation in which the first information in FIG. 6 includes the first indication. In this case, S102 in FIG. 6 can be implemented by S202 in FIG. 7.

[0167] After starting training for the ML model based on the procedure shown in FIG. 7, the first device may send a second indication to the second device, and the second device sends a first indication to the first device based on the second indication. Thus, it is possible to avoid a case where the first device starts training for the ML model when training for the ML model does not need to be performed, thereby reducing resource waste. In addition, in this implementation, the second device can learn that the first device starts training for the ML model based on the second indication. In this way, it is possible to avoid training conflicts caused by repeated initiation of training for the ML model by the second device.

[0168] 7, in S201, the first device may send a second indication to the second device after starting training on the ML model. Optionally, in the example, the first device may send a first notification (or a first message) to the second device, and the first notification may include the second indication.

[0169] It should be understood that the second indication may be specific information in the first notification, for example, a field or bit with a specific value. In other words, the field or bit with a specific value may indicate to the second device that the first device should determine whether to train on the ML model. For example, the second indication indicates the Xth bit in the first notification. When the value of the bit is 1, the bit indicates to the second device that the first device should determine whether to train on the ML model, where 0≦X, and X is a positive integer. For example, when the second device receives a notification from the first device and determines that the value of the Xth bit in the notification is 1, the notification or bit indicates to the second device that the first device should determine whether to train on the ML model. In other words, in this case, the second device may determine whether the first device should train on the ML model, and the notification is the first notification. For example, in the notification received by the second device from the first device, if the value of the Xth bit is 0, the notification does not indicate to the second device that the first device should decide whether to train the ML model, or the second device does not need to decide whether the first device trains the ML model based on the notification. The notification may have another meaning, which is not specifically required in this application.

[0170] In addition, the second indication may alternatively be the first notification. In other words, the notification may indicate to the second device whether the first device performs training on the ML model. For example, a notification with a specific name, a notification with a specific attribute, or a notification sent through a specific interface may be used as the first notification. The first notification may not include an attribute, an information field, or a bit indicating whether training is performed on the ML model. The following will be described in detail with reference to attributes that may be included in the first notification. Details will not be described herein.

[0171] 7, in S202, the second device may send a first indication to the first device. Optionally, the second device may send a second notification (or a second message) to the first device, and the second notification may include the first indication. In other words, S202 may be used as an implementation of S101.

[0172] Specifically, after receiving the second indication shown in S201, the second device may determine whether the first device performs training on the ML model, and perform S202 based on the determined result. The value range of the first indication in S202 includes a first value and a second value. For details, please refer to the description of the first indication in S101.

[0173] If the second device determines that the first device will train on the ML model, the value of the first indication in S202 is a first value, in other words, the second device sends to the first device a first indication whose value is the first value. If the second device determines that the first device will not train on the ML model, the value of the first indication in S202 is a second value, in other words, the second device sends to the first device a first indication whose value is the second value.

[0174] Correspondingly, the first device may receive a first indication from the second device, and determine whether to perform training on the ML model based on the value of the first indication. If the value of the first indication is a first value, S203 is performed, in other words, the first device performs training on the ML model. It should be understood that S203 in FIG. 7 is an example of S103 shown in FIG. 6. In addition, if the value of the first indication is a second value, the first device does not perform training on the ML model.

[0175] In another implementation, if the second device determines that the first device does not train the ML model, the second device does not perform S202 or does not send the first indication. Correspondingly, if the first device does not receive the second notification, the first device does not need to perform S203.

[0176] 7, optionally, the first device may further send training status information of the ML model to the second device, so that the second device may determine whether the first device performs training on the ML model based on the training status information of the ML model. For example, in S201, the first device may send a second indication and training status information of the ML model to the second device.

[0177] In a possible implementation, the training status information of the ML model may indicate whether the ML model is trained. For example, the training status information of the ML model may indicate that the ML model is currently being trained (i.e., indicates that training for the ML model is running), that the ML model is being trained, etc. It should be understood that training in this specification includes the first training, retraining (ML model retraining), or the like of the ML model. This is not particularly limited.

[0178] For example, in this implementation, when the training status information of the ML model indicates that the ML model is being trained, the second device may avoid repeatedly training the ML model to avoid training conflicts and improve the training accuracy of the ML model. Thus, in this case, the second device may determine that the first device does not train the ML model.

[0179] In another possible implementation, the training status information of the ML model may indicate whether training is performed on the ML model based on the performance evaluation result or feedback of the ML model. For example, when a first device receives a performance evaluation result or feedback of the ML model from a second device and triggers training on the ML model based on the performance evaluation result or feedback, the training status information of the ML model may indicate that the current training is performed based on the performance evaluation result or feedback.

[0180] In another possible implementation, the training status information of the ML model may indicate whether training for the ML model is performed based on updated network conditions and updated training data. For example, when the first device determines that training should be performed for the ML model based on updated network conditions, the training status information of the ML model may indicate that training for the ML model has been performed based on updated network conditions. As another example, when the first device determines that training should be performed for the ML model based on updated training data, the training status information of the ML model may indicate that training for the ML model has been performed based on updated training data. As another example, when the first device determines that training should be performed for the ML model based on updated network conditions and updated training data, the training status information of the ML model may indicate that training for the ML model has been performed based on updated network conditions and updated training data.

[0181] Further, optionally, the training status information of the ML model may further include information indicating whether training is performed based on updated training data transmitted by the second device, or may further include information indicating whether training for the ML model is performed based on updated training data from another source (e.g., training data obtained by the first device from another channel).

[0182] It should be understood that if the training state information of the ML model indicates that the training on the ML model has been performed based on the performance evaluation result (or feedback) of the ML model, or indicates that the training on the ML model has been performed based on at least one of an updated network state and updated training data, the second device may determine that the first device performs training on the ML model regardless of whether the ML model has been trained. This is because in this case, even if the ML model has been trained before, the accuracy of the previously trained ML model may not meet the current requirements, and the ML model needs to be retrained to improve the accuracy.

[0183] In the present application, the second device may determine whether an ML model has been acquired or whether the accuracy of the acquired ML model meets a requirement based on the training status information, and the second device may perform training on the ML model again to further avoid a case where the first device performs training on the ML model when training does not need to be performed on the ML model.

[0184] Optionally, the training status information of the ML model may be specific to one or more specific analysis scenarios (or referred to as use cases). For example, the training status information of the ML model indicates whether the ML model is trained in the scenario, or indicates whether training for the ML model is performed in a particular scenario based on at least one of the performance evaluation results of the ML model, the updated network state, and the updated training data. Thus, this can help the second device to more appropriately determine whether the first device performs training on the ML model while referring to the scenario. For example, different scenarios have different requirements on the performance of the ML model. Thus, the second device may determine whether the first device performs training on the ML model based on the requirements of the scenario on the performance of the ML model.

[0185] Specifically, the scenarios may include coverage problem analysis, slice coverage analysis, paging optimization analysis, fault analysis (alarm incident analysis), fault prediction analysis, end-to-end (E2E) latency analysis, energy saving analysis, mobility performance analysis, network slice load analysis, network slice throughput analysis, and key performance indicator (KPI) anomaly analysis, radio access network node software upgrade analysis, etc.

[0186] Coverage problem analysis refers to analyzing coverage such as weak coverage, excess coverage, and overshoot coverage, and providing coverage problem analysis results. Slice coverage problem analysis is used to analyze slice coverage, slice availability, and slice prediction information. Paging optimization analysis is used to analyze subscriber paging data and optimize the current network paging process. Fault analysis is used to analyze network warnings, faults, or events and provide solution suggestions. Fault prediction analysis is used to analyze and predict potential network faults and provide solution suggestions. End-to-end latency analysis is used to analyze RAN latency problems, core network (CN) latency problems, transport network (TN) latency problems, UE latency problems, and service provider latency problems to support service level specification (SLS) guarantees. Energy saving analysis is used to analyze RAN or CN energy consumption and energy efficiency problems to reduce network energy consumption and improve energy efficiency. Mobility analysis is used to analyze network mobility performance, such as premature and delayed handovers, to improve handover success rates. Network slice load analysis is used to analyze KPI degradation caused by performance policies and load issues (e.g., radio resource utilization) to provide suggestions for network slice load issues. Network slice throughput analysis is used to analyze and predict network slice throughput to improve or ensure network throughput. KPI anomaly analysis is used to analyze cross-domain and single-domain KPIs to identify root causes and provide correction suggestions. A root cause is the cause of a KPI anomaly. For example, root causes include parameter configuration anomalies, software anomalies, hardware device anomalies, or increased service requirements.The radio access network node software upgrade analysis can be used to analyze and optimize the time to upgrade the RAN node software. The expected impact is small in time, or the expected operating cost and data loss are small in time, so that service interruption is avoided and operating cost is saved. In other words, by using the radio access network node software upgrade analysis, the software upgrade can be automatically started in good time when the expected operating cost and data loss are small. For the above scenario, please refer to the description in Chapter 7 of Non-Patent Document 2 and Chapter 6 of Non-Patent Document 3. The details will not be described again in this specification.

[0187] It should be understood that in the embodiment of the present application, optionally, the second indication of the ML model and the training state information may be carried in the same message (e.g., the first notification) or may be sent separately, which is not particularly limited.

[0188] The following describes a possible implementation in which the training state information of the ML model is included in the first notification with reference to Table 1. Details will not be described herein.

[0189] FIG. 8 shows a procedure of another communication method for training a machine learning model according to an embodiment of the present application. In this procedure, a first device receives a training policy of an ML model from a second device. The training policy of the ML model is a condition for starting training on the ML model by the first device. Therefore, the procedure shown in FIG. 8 can avoid a case where the first device starts training on the ML model when training does not need to be performed on the ML model, and reduce resource waste. It can be understood that the procedure shown in FIG. 8 is an implementation used when the first information includes a training policy of the ML model, which is that of the procedure shown in FIG. 6. In this case, S102 in FIG. 6 can be implemented by S301 in FIG. 8. For example, the procedure shown in FIG. 8 can include S301 and S302.

[0190] 8, in S301, the second device may send a training policy of the ML model to the first device. In other words, S301 may be used as an implementation of S101 shown in FIG.

[0191] Optionally, in S301, the second device may send the training policy of the ML model to the first device. Optionally, when the first device determines that training needs to be performed on the ML model, the second device may send the training policy of the ML model to the first device. The first device may determine that training needs to be performed on the ML model, for example, when the performance of the ML model does not satisfy the second device's requirements for the performance of the ML model. Alternatively, the second device may determine that training needs to be performed on the ML model after a specified time after a previous training on the ML model.

[0192] At S302, the first device may perform training on the ML model according to a training policy for the ML model.

[0193] It should be understood that S302 is an example of S103.

[0194] As described in S103, the first device performing training on the ML model in accordance with the training policy of the ML model may be understood as the first device performing training on the ML model after the training policy of the ML model is satisfied.

[0195] In a possible implementation of the procedure shown in FIG. 8, the first device may further request the second device to determine whether the first device performs training on the ML model, and to prevent the first device from starting training on the ML model when training does not need to be performed on the ML model. For example, S302 may further include the first device sending a second indication to the second device after determining that the training policy of the ML model is satisfied. Optionally, the second indication may be implemented by using S201 of the procedure shown in FIG. 7. Therefore, please refer to the description of S201. The first device may further receive a first indication from the second device. When the value of the first indication is a first value, the first device may perform training on the ML model. Otherwise, when the value of the first indication is a second value, or when the first device does not receive a first indication whose value is a first value, the first device stops or does not perform training on the ML model. Optionally, the first indication may be implemented by using S202 of the procedure shown in Figure 7. Therefore, please refer to the description of S202.

[0196] In the embodiment of the present specification, the first device may further receive training control parameters of the ML model from the second device, and the training control parameters are used to perform training on the ML model. For example, in S103, S203, and S302, the first device may perform training on the ML model based on the training control parameters of the ML model.

[0197] For example, the training control parameters of the ML model include at least one of parameters such as an identifier of the ML model, type information of the ML model, the number of training iterations for the ML model, performance requirements of the ML model, training time of the ML model, and data used to train the ML model. For example, the first device performing training on the ML model based on the training control parameters of the ML model means that the first device performs training on the ML model based on at least one of parameters such as the number of training iterations for the ML model, performance requirements of the ML model, training time of the ML model, and data used to train the ML model. Optionally, at least one of the identifier of the ML model and the type information of the ML model may be used to determine the ML model to be trained.

[0198] According to this example, the first device performs training on the ML model based on the training control parameters of the ML model, so that the training effect of the ML model can be further improved, and the training efficiency of the ML model can be improved.

[0199] The identifier of the ML model may indicate the ML model to which the training control parameters are applicable, and the first device may then use the identifier to perform training on the ML model based on the training control parameters.

[0200] The type information of the ML model may indicate a type of ML model or a training type to which the training control parameters are applicable. The type of ML model or the training type may correspond to a scenario to which the ML model is applicable. For example, the type of ML model may include types such as coverage problem analysis and slice coverage problem analysis. For details, please refer to the scenario description in this specification. Therefore, the first device may perform training on the ML model of that type.

[0201] Optionally, when the training control parameters do not include an identifier of the ML model or type information of the ML model, the first device may apply the training control parameters for training all ML models. The number of training iterations for the ML model may indicate a maximum amount of training to be performed on the ML model.

[0202] The performance requirement of the ML model may include, for example, an accuracy requirement or a confidence requirement of the ML model. The accuracy requirement of the ML model may include, for example, an accuracy requirement and an error rate requirement. The accuracy rate represents the accuracy of the trained ML model, and the error rate represents the error rate of the trained ML model. The confidence is the certainty of the trained ML model. For example, the performance requirement of the ML model may indicate that the accuracy requirement of the ML model is 99% and the confidence requirement is 90%.

[0203] The training time of the ML model may indicate a period of time required for training the ML model. For example, the training time of the ML model may indicate that the first device needs to train the ML model for a day or N hours (N is greater than 0).

[0204] The data used to train the ML model may include at least one of a data source, an updated network state, and an updated training data. Thus, the training control parameters of the ML model may more accurately set the data used to train the ML model to further improve the training effect of the ML model, or may indicate that the first device should perform training on the ML model based on a portion of the data to improve the training efficiency of the ML model.

[0205] A data source is data that is obtained based on a data source (candidate data source) address and is used by a first device to perform training on an ML model. The data source address may be from the first device. The data source may be pre-located in the device to which the data source address belongs. In this application, there is no specific requirement on the data contained in the data source.

[0206] The updated network state may include information about changes in the network state, such as an indication that the network state has changed. For example, when some cells in the analysis area enter an energy saving state, the updated network state may indicate that these cells enter an energy saving state.

[0207] The updated training data may specifically be data, data values, etc. obtained after a change in data distribution.

[0208] It should be understood that the data used to train the ML model may include data from the second device, data collected by the first device, or data obtained by the first device from another device. For example, the other device may include a 5G core network, an application program such as an APP, or an external server. This is not particularly limited.

[0209] Optionally, at least one of the parameters, such as the number of training iterations for the ML model, the performance requirements of the ML model, the training time of the ML model, and the data used to train the ML model, may be a general training control parameter, i.e., applicable to the training process of multiple ML models. At least one of the parameters, such as the number of training iterations for the ML model, the performance requirements of the ML model, the training time of the ML model, and the data used to train the ML model, may alternatively be a training control parameter for one ML model or multiple ML models. In other words, the training control parameter of an ML model is applicable only to the training process of one ML model or to the training process of multiple ML models. The type information of the multiple ML models may be the same. When the training control parameter of an ML model includes an identifier of the ML model and / or the type information of the ML model, the first device may perform training on the ML model having the identifier of the ML model and / or the type information of the ML model based on at least one of the parameters, such as the number of training iterations for the ML model, the performance requirements of the ML model, the training time of the ML model, and the data used to train the ML model.

[0210] In addition, the training control parameters of the ML model may further include a training periodicity parameter to improve the training periodicity of the ML model.

[0211] It can be appreciated that the training control parameters used to train the ML model in different analysis scenarios may be different because the ML models for different scenarios may be different, in other words, the training control parameters of the ML model may be set or determined for a particular scenario.

[0212] Optionally, the training control parameters of the ML model may be transmitted by the second device to the first device. For example, in the procedure shown in FIG. 7, the first device may transmit a first indication whose value is a first value and the training control parameters of the ML model in S202. For example, the first indication and the training control parameters of the ML model may be carried in the same notification (e.g., the first notification) or message, or the first indication, and the training control parameters of the ML model may be carried in a different notification or message. As another example, in the procedure shown in FIG. 8, the first device may transmit a training policy and training control parameters of the ML model to the second device in S301. For example, the training policy and training control parameters of the ML model may be carried in the same notification or message, or the training policy and training control parameters of the ML model may be carried in a different notification or message. Alternatively, regardless of whether S202 and S301 are performed, the second device may transmit the training control parameters of the ML model to the first device via another notification or message.

[0213] In a possible example, in each embodiment of the present application, the first device may request the training control parameters of the ML model from the second device to accurately obtain the training control parameters required for training the ML model, and improve the training efficiency. Correspondingly, the second device may transmit the training control parameters of the ML model to the first device based on the request of the first device.

[0214] For example, the first device may send a third indication to the second device, where the third indication may be used to request training control parameters of the ML model, so that the second device sends the training control parameters of the ML model to the first device after receiving the third indication. For example, the third indication may include at least one of an identifier of the ML model and type information of the ML model. It should be understood that the third indication may be sent before or after the first device receives the first information in S102. In a possible implementation, the third indication may be sent by the first device for the current training. For example, the first device sends the third indication after receiving the first information. In another implementation, the third indication may be sent in a previous training process of the ML model as an alternative. This is not particularly limited in the present application. For example, in the procedure shown in FIG. 7, the first device may send the second indication and the third indication in S201. The second indication and the third indication may be carried in the same notification (e.g., the first notification) or message, or the second indication and the third indication may be carried in different notifications or messages. As another example, in the procedure shown in FIG. 8, after receiving the training policy of the ML model shown in S301, the first device may send a third indication to the second device, for example, a notification or another message carrying the third indication. Alternatively, regardless of whether S201 and S301 are performed, the first device may send the third indication to the second device via a notification or message.

[0215] In another possible example, the training control parameters of the ML model may be pre-configured in the first device or may be obtained by the first device from another device, which is not particularly limited in this application.

[0216] In an embodiment of the present application, after performing training on the ML model, the first device may further transmit the training result of the ML model to the second device. Correspondingly, the second device receives the training result of the ML model from the first device, and performs problem analysis based on the trained ML model to improve analysis performance. For example, optionally, the procedure shown in FIG. 7 may further include S204. In S204, the first device transmits the training result of the ML model to the second device. As another example, the procedure shown in FIG. 8 may further include S303. In S303, the first device transmits the training result of the ML model to the second device. It should be understood that the dashed arrows in FIG. 7 and FIG. 8 represent optional steps.

[0217] Optionally, the first device may transmit the training result of the ML model to the second device based on a training result reporting method (reporting method) configuration of the ML model. For example, the training result reporting method configuration of the ML model may include configurations such as a training result reporting periodicity of the ML model or a training result reporting manner of the ML model. In this manner, the first device may report the training result of the ML model to the first device based on at least one of the periodicity or the reporting manner indicated by the training result reporting method configuration. Optionally, the training result reporting method configuration may include at least one of an identifier of the ML model and type information of the ML model to indicate a specific ML model, where the training result reporting method configuration is applicable to reporting the training result of the specific ML model.

[0218] The training result reporting scheme of the ML model may include a file data reporting scheme (or, for short, file reporting) scheme, a streaming data reporting scheme (or, for short, streaming reporting) scheme, or a notification reporting scheme.

[0219] In an example, when a first device transmits a training result of an ML model to a second device in a streaming reporting manner, the first device may transmit a notification to the second device including fields such as streaming connection establishment (establish a streaming connection), streaming connection termination (terminate a streaming connection), and streaming data report (report streaming data). The streaming data report field may include the training result of the ML model.

[0220] In another example, when a first device transmits training results to a second device in a file reporting manner, the first device may send a notification to the second device including fields such as file subscription (Subscribe), file unsubscription (Unsubscribe), and available file enumeration (Enumerate available files). The available file enumeration field may include the training results of the ML model.

[0221] Alternatively, in another example, when the first device transmits the training result to the second device in a file reporting manner, the first device may transmit a file ready notification (notify file ready) message to the second device, where the message may include the training result of the ML model. In addition, in the file reporting process, notifying file preparation error (notify file preparation error) represents reporting an error.

[0222] In another example, when a first device sends training results to a second device in a notification reporting manner, the first device may send the training results of the ML model to the second device via the original notification or the newly defined notification.

[0223] It should be understood that for the file data reporting method or streaming data reporting method, reference should be made to the description in Non-Patent Document 4.

[0224] In a possible implementation, the first device obtains the training result reporting method configuration from the second device. For example, the first device may send an indication (sometimes referred to as a fourth indication) used to request the training result reporting method configuration of the ML model to the second device. After receiving the fourth indication, the second device sends the training result reporting method configuration of the ML model to the first device. Optionally, the fourth indication and the second indication may be the same or different. In other words, the first device may request the training control parameters of the ML model and the training result reporting method configuration of the ML model from the second device by using the same indication (or information or request). For example, in the procedure shown in FIG. 7, the first device may send the second indication and the fourth indication in S201. The second indication and the fourth indication may be carried in the same notification (e.g., the first notification) or message, or the second indication and the fourth indication may be carried in different notifications or messages. As another example, in the procedure shown in FIG. 8, after receiving the training policy of the ML model shown in S301, the first device may send a fourth indication to the second device, for example, a notification or another message carrying the fourth indication. Alternatively, regardless of whether S201 and S301 are performed, the first device may send the second indication to the second device via another notification or message.

[0225] In another possible implementation, when the first device requests the training control parameters of the ML model from the second device, e.g., when the first device sends a second indication to the second device, the second device may send the training control parameters of the ML model and the training result reporting method configuration of the ML model to the first device by default. In addition, the training result reporting method configuration may alternatively be sent independently of the training control parameters.

[0226] In addition, the ML model training result reporting method configuration may be pre-configured in the first device or may be obtained by the first device from a device other than the second device, which is not particularly limited in the present application.

[0227] In an embodiment of the present application, by using a specific interface operation or instance object class (IOC), for example, an ML training indication, or by reusing an existing MnS operation, the first device may send at least one of the second indication, the third indication, or the fourth indication to the second device, or may send a notification or message including at least one of the second indication, the third indication, or the fourth indication to the second device.

[0228] For example, a notification of a specific format may be defined as a first notification carrying a second indication in S201. For example, a specific format is embodied because the notification has a specific name or because the notification includes a specific attribute. For example, a specific notification may include attributes shown in Table 1. In this case, the notification is the second indication in S201 shown in FIG. 7, or the notification including the attributes shown in Table 1 may be used as the second indication. As shown in Table 1, the notification includes a training type, an ML training state attribute, and an ML training control request attribute. When receiving a notification carrying the aforementioned attributes, the second device may learn that the notification indicates to the second device that the first device should decide whether to perform training on the ML model, in other words, the notification is a second indication. Thus, after receiving the notification, the second device may determine whether the first device performs training on the ML model and send a first indication to the first device. For the process of determining whether the first device performs training on the ML model by the second device and the corresponding process of sending the first indication, please refer to the description of the embodiment shown in FIG. 7.

[0229] Further optionally, the ML training control request attribute may be used as a third indication.

[0230] [Table 1]

[0231] In addition, the first notification may further include information such as an identifier of the ML model, which is not particularly limited in this application.

[0232] Optionally, the ML training states in Table 1 may include the attributes shown in Table 2.

[0233] [Table 2]

[0234] Optionally, the ML training control request attributes in Table 1 may include the attributes shown in Table 3.

[0235] [Table 3]

[0236] Based on an implementation similar to that of Tables 1 to 3, the second device may send at least one of the first indication and the training control parameters to the second device by using a specific interface operation or instance object class, or by reusing an existing MnS operation. For example, when sending the training control parameters to the first device, the second device may send a notification carrying the attributes shown in Table 3. In this case, the attributes shown in Table 3 may carry the corresponding training control parameters. For example, the iteration count attribute may carry the number of iterations of training for the ML model, and the training performance may carry the performance requirements of the ML model.

[0237] In an embodiment of the present application, optionally, the first device may show the training progress of the ML model to the second device, so that the second device learns the training status in time.

[0238] For example, the first device may transmit second information to the second device, where the second information includes at least one of information indicating a training progress of the ML model, an estimated training time of the ML model, a training execution duration of the ML model, a training start time of the ML model, and a training end time of the ML model. Optionally, the second information may include an identifier of the ML model.

[0239] The information indicating the training progress of the ML model may indicate that training on the ML model starts or ends. For example, the message (e.g., a notification) or information indicates that the first device starts training on the ML model. When starting training on the ML model, the first device sends the message or information to the second device. As another example, a field or bit with a particular value indicates that training on the ML model starts. When the first device starts training on the ML model, the value of the field or bit is set to that value, indicating that the first device starts performing training on the ML model.

[0240] Alternatively, the information indicating the training progress of the ML model may include a percentage value indicating the completed (or incomplete) progress of the current training for the ML model, for example, the value conveyed may indicate that the completed progress is 50%.

[0241] The estimated training time for training on an ML model may indicate an estimated duration of training on an ML model, i.e., an estimated time between a training start time and a training end time. For example, the estimated training time for training on an ML model may indicate that the estimated duration of the current training is 6 hours.

[0242] The training duration of the ML model may indicate how long the current training for the ML model has been running. For example, the training duration of the ML model may indicate that the first device has been training the ML for 3 hours.

[0243] The training start time of an ML model may indicate a training start time of a current training for an ML model. For example, the training start time of an ML model may indicate that the current training starts at 7:00 a.m. (ante meridiem, AM).

[0244] The training end time of the ML model may indicate an estimated time to complete training for the ML model. For example, the estimated training time of the training for the ML model may include a completion time, which may indicate, for example, that the training for the ML model is estimated to be completed at 3:00 PM (post meridiem). As another example, the estimated training time of the training for the ML model may include an estimated remaining time. For example, the estimated remaining time may indicate that the estimated remaining duration of the training is 3 hours, or that the training is estimated to be completed in another 3 hours.

[0245] Optionally, the second information may be carried in the newly defined notification or the existing notification, or the second information may be included in the newly defined notification or the existing notification, which is not limited in this application.

[0246] For example, the ML model training start notification may indicate an estimated duration of the ML model training. For example, the ML model training start notification may include the attributes shown in Table 4.

[0247] [Table 4]

[0248] As another example, the ML model training end notification may indicate the progress percentage of the ML model and the estimated end time of training. For example, the ML model training end notification may include the attributes shown in Table 5.

[0249] [Table 5]

[0250] As another example, the ML model training progress notification may indicate the progress percentage of the ML model training, the training execution duration, and the estimated completion time. For example, the ML model training completion notification may include the attributes shown in Table 6.

[0251] [Table 6]

[0252] It should be understood that the act of transmitting the second information by the first device may not depend on the procedure shown in Fig. 6. For example, the first device performs training on the ML model between the first device and the second device in the ML model training trigger manner described in Chapter 6.2.2 of Non-Patent Document 5. The first device may further transmit the second information to the second device to indicate the training progress of the ML model. This may avoid repeated training on the ML model.

[0253] In an example, when the second information is implemented with reference to the procedure shown in FIG. 7, the procedure shown in FIG. 7 may further include S205. In S205, the first device may transmit the second information to the second device. In addition, when the second information is implemented with reference to the procedure shown in FIG. 8, the procedure shown in FIG. 8 may further include S304. In S304, the first device may transmit the second information to the second device. For example, the second information in S205 or S304 includes an ML model training start notification shown in Table 4, an ML model training end notification shown in Table 5, or an ML model training progress notification shown in Table 6.

[0254] In another example, as shown in FIG. 9, in S401, the first device may send second information to the second device to indicate the training progress of the ML model. The first device may perform S401 at least once in the training process of the ML model or when starting training for the ML model after completing the training for the ML model. For example, the first device transmits to the second device second information indicating that training for the ML model starts when training for the ML model starts, transmits second information indicating the training progress of the ML model in the training process of the ML model, and transmits second information indicating that training for the ML model ends after the training for the ML model ends. As another example, the first device transmits to the second device second information indicating that training for the ML model starts when training for the ML model starts, and transmits second information indicating that training for the ML model ends after the training for the ML model ends. As another example, the first device transmits second information indicating the training progress of the ML model in the training process of the ML model, and transmits second information indicating that training for the ML model ends after the training for the ML model ends. As another example, the first device transmits to the second device, when training for the ML model starts, second information indicating that training for the ML model starts, and in the training process of the ML model, transmits second information indicating the training progress of the ML model. As another example, the first device transmits to the second device, when training for the ML model starts, second information indicating that training for the ML model starts, in the training process of the ML model, transmits second information indicating the training progress of the ML model, or transmits second information indicating that training for the ML model ends after training for the ML model ends.

[0255] The training process of the ML model is between the start of training for the ML model and the end of training for the ML model. Optionally, in the training process of the ML model, the first device may send the second information to the second device at least once. It should be understood that in FIG. 9, the dashed arrow indicates that the first device may perform S401 continuously or multiple times. The content of the second information may be different each time S401 is performed.

[0256] For example, when the first device transmits the second information when starting training for the ML model, the second information may include at least one of information indicating that training for the ML model starts, an estimated training time for the training for the ML model, or a training start time for the ML model. When the first device transmits the second information during training processing for the ML model, the second information may include at least one of a percentage value indicating a completed (or incomplete) processing of training for the ML model, an estimated training time for the training for the ML model, or a training execution duration for the ML model, and a training start time for the ML model. When the first device transmits the second information after training for the ML model is completed, the second information may include at least one of information indicating that training for the ML model is completed, a training execution duration for the ML model, or a training end time for the ML model.

[0257] Based on the above method embodiment, the embodiment of the present application further provides a communication device configured to implement the communication method used for machine learning model training. When the communication device is configured to implement the method implemented by the first device, the communication device may include a training execution module. When the communication device is configured to implement the method implemented by the second device, some devices may include a training control module.

[0258] For example, an example is used in which the first device is an MLMT MnS producer and the second device is an MLMT MnS consumer. When the procedure shown in FIG. 7 is implemented, the training execution module in the MLMT MnS producer may perform the steps performed by the first device in the procedure shown in FIG. 7, and the training control module in the MLMT MnS consumer may perform the steps performed by the second device in the procedure shown in FIG. 7. For details, please refer to the description in the procedure in FIG. 7. As shown in FIG. 10, the training execution module may send a second indication in S201 to the training control module, and the training control module may send a first indication in S202 to the training execution module. In addition, the training execution module may further trigger training for the ML model and perform training for the ML model. Alternatively, the training execution module may send at least one of the training result of the ML model in S204 and the second information in S205 to the training control module.

[0259] When the procedure shown in FIG. 8 is implemented, the training execution module in the MLMT MnS producer may perform the steps performed by the first device in the procedure shown in FIG. 8, and the training control module in the MLMT MnS consumer may perform the steps performed by the second device in the procedure shown in FIG. 8. For details, please refer to the description in the procedure in FIG. 8. For example, as shown in FIG. 11, the training control module may send the training policy of the ML model in S301 to the training execution module, and the training execution module may perform S302, in other words, perform training on the ML model according to the training policy of the ML model. In addition, the training execution module may send at least one of the training result of the ML model in S303 and the second information in S304 to the training control module.

[0260] When the procedure shown in Fig. 9 is implemented, a training execution module in the MLMT MnS producer may perform the steps performed by the first device in the procedure shown in Fig. 9, and a training control module in the MLMT MnS consumer may perform the steps performed by the second device in the procedure shown in Fig. 9. For details, please refer to the description in the procedure in Fig. 9. For example, as shown in Fig. 12, the training control module may send the second information in S401 to the training execution module to indicate the training progress of the ML model.

[0261] Optionally, when an ML model-related indication (e.g., any one of the first indication to the fourth indication), ML model-related information (e.g., any one of the first information and the second information), or an ML model-related notification (e.g., any one of the first notification and the second notification) is transmitted between a first device and a second device in the present application, at least one of an ML model identifier and ML model type information may be carried or transmitted together to indicate which ML model or type of ML model the corresponding indication, information, or notification is associated with. For example, in communication processes such as S102 in the procedure shown in Fig. 6, S201, S202, S204, and S205 in the procedure shown in Fig. 7, S301, S303, or S304 in the procedure shown in Fig. 8, and S401 in Fig. 9, at least one of an ML model identifier and ML model type information may be transmitted between the first device and the second device to indicate an ML model or a type of ML model related to the related interaction process. For example, in S101, the second device may transmit the first information and at least one of the ML model identifier and ML model type information to the first device, indicating one or more ML models to which the first information is specific, so that the first device can perform training on an ML model having at least one of the ML model identifier and ML model type information based on the first information.

[0262] Based on the aforementioned method embodiment, the embodiment of the present application further provides a communication device. As shown in FIG. 13, the structure of the device includes a transceiver module (or communication module) 1301 and a processing module 1302. The transceiver module 1301 may include at least one of a receiving module and a transmitting module, and is configured to implement at least one of information reception and information transmission. The processing module 1302 may be configured to generate information transmitted by the transceiver module 1301 or process information received by the transceiver module 1301. The processing module 1302 may be configured to perform acts other than receiving and transmitting in the aforementioned method embodiment. The device 1300 may be used in at least one of a first device and a second device. The first device and the second device may be used in the communication system shown in FIG. 1 to FIG. 5, and may implement the communication method used for machine learning model training in the aforementioned drawings.

[0263] When the communication device is configured to implement the first device provided in this embodiment of the present application, the communication device may include the first device. In other words, the first device may include a transceiver module 1301 and a processing module 1302. The transceiver module 1301 may be configured to receive first information from a second device, where the first information includes at least one of a first indication and a training condition of the ML model. The value range of the first indication includes a first value and a second value, where the first value indicates that the communication device should perform training on the ML model, and the second value indicates that the communication device should not perform training on the ML model, and the value of the first indication is the first value. The processing module may be configured to perform training on the ML model based on the first information. The communication device includes a model training producer, and the second device includes a model training consumer.

[0264] In a possible implementation, the transceiver module 1301 may further transmit training results of the ML model to the second device.

[0265] In a possible implementation, the first information may include the first indication, and the processing module 1302 may further start training on the ML model. The transceiver module 1301 may be further configured to send a second indication to the second device, where the second indication is used to notify the start of training on the ML model, indicate to the second device that it should determine whether to perform training on the ML model, or indicate to the second device that the first device should determine whether to perform training on the ML model.

[0266] In a possible implementation, the transceiver module 1301 may further transmit a third indication to the second device, where the third indication is used to request training control parameters of the ML model. The transceiver module 1301 receives the training control parameters of the ML model from the second device, where the training control parameters of the ML model are used by the communication device to perform training on the ML model.

[0267] In a possible implementation, the training control parameters for the ML model include at least one of an identifier for the ML model, type information of the ML model, a number of training iterations for the ML model, performance requirements for the ML model, a training time for the ML model, or data used to train the ML model.

[0268] In a possible implementation, the data used to train the ML model includes at least one of a data source, updated network states, and updated training data.

[0269] In a possible implementation, the training conditions for the ML model include at least one of a network load condition, a network coverage condition, and a network performance condition.

[0270] In a possible implementation, the network load conditions include at least one of a PRB utilization threshold and a user connection volume threshold.

[0271] In a possible implementation, the network coverage conditions include at least one of an RSRP threshold and a coverage rate threshold of the RSRP threshold in the area, an RSRQ threshold and a coverage rate threshold of the RSRQ threshold in the area, or an RSSINR threshold and a coverage rate threshold of the RSSINR threshold in the area.

[0272] In a possible implementation, the network performance conditions include at least one of a handover success rate threshold, a data energy efficiency threshold, or a network slice energy efficiency threshold.

[0273] In a possible implementation, the transceiver module 1301 may further transmit training status information of the ML model to the second device, where the training status information indicates whether the ML model is trained or whether training for the ML model is performed based on at least one of the performance evaluation results of the ML model, the updated network conditions, and the updated training data.

[0274] In a possible implementation, the transceiver module 1301 may further transmit second information to the second device, where the second information includes at least one of information indicating the training progress of the ML model, an estimated training time for the training for the ML model, a training execution duration for the ML model, a training start time for the ML model, or a training end time for the ML model.

[0275] When the communication device is configured to implement the second device provided in this embodiment of the present application, the communication device may include the second device. In other words, the second device may include a transceiver module 1301 and a processing module 1302. The processing module 1302 may be configured to determine first information, where the first information includes at least one of a first indication and a training condition of the ML model. The value range of the first indication includes a first value and a second value, where the first value indicates that the first device should perform training on the ML model, the second value indicates that the first device should not perform training on the ML model, and the value of the first indication is the first value. The transceiver module 1301 is configured to transmit the first information to the first device. The first device includes a model training producer, and the second device includes a model training consumer.

[0276] In a possible implementation, the transceiver module 1301 may be further configured to receive training results of the ML model from the first device.

[0277] In a possible implementation, the first information includes a first indication, and the transceiver module 1301 may be further configured to receive a second indication from the first device, the second indication being used to notify the start of training on the ML model, indicating to the second device that it should determine whether to perform training on the ML model, or indicating to the second device that the first device should determine whether to perform training on the ML model.

[0278] In a possible implementation, the transceiver module 1301 may be further configured to receive a third indication from the first device, where the third indication is used to request training control parameters of the ML model. The transceiver module 1301 may be further configured to transmit the training control parameters of the ML model to the first device, where the training control parameters of the ML model are used by the first device to perform training on the ML model.

[0279] In a possible implementation, the training control parameters for the ML model include at least one of an identifier for the ML model, type information of the ML model, a number of training iterations for the ML model, performance requirements for the ML model, a training time for the ML model, and data used to train the ML model.

[0280] In a possible implementation, the data used to train the ML model includes at least one of a data source, updated network states, and updated training data.

[0281] In a possible implementation, the training conditions for the ML model include at least one of a network load condition, a network coverage condition, and a network performance condition.

[0282] In a possible implementation, the network load conditions include at least one of a PRB utilization threshold and a user connection volume threshold.

[0283] In a possible implementation, the network coverage conditions include at least one of an RSRP threshold and a coverage rate threshold of the RSRP threshold in the area, an RSRQ threshold and a coverage rate threshold of the RSRQ threshold in the area, and an RSSINR threshold and a coverage rate threshold of the RSSINR threshold in the area.

[0284] In a possible implementation, the network performance conditions include at least one of a handover success rate threshold, a data energy efficiency threshold, and a network slice energy efficiency threshold.

[0285] In a possible implementation, the transceiver module 1301 may further receive training status information of the ML model from the first device, where the training status information indicates whether the ML model is trained, and where the training status information indicates whether the ML model is trained based on at least one of the performance evaluation results of the ML model, the updated network conditions, and the updated training data.

[0286] In a possible implementation, the transceiver module 1301 may be further configured to receive second information from the first device, where the second information includes at least one of information indicating the training progress of the ML model, an estimated training time for the training for the ML model, a training execution duration of the ML model, a training start time of the ML model, and a training end time of the ML model.

[0287] In a possible implementation, when implementing the function of the first device, the transceiver module 1301 may be further configured to send a machine learning ML training indication based on new data to the second device, where the ML training indication based on new data indicates whether training should be performed based on updated training data of the ML model.

[0288] Optionally, the transceiver module 1301 may be further configured to receive training control parameters of the ML model from the second device, where the training control parameters of the ML model are used by the communication device to perform training on the ML model.

[0289] Optionally, the transceiver module 1301 may further transmit second information to the second device, where the second information includes at least one of information indicating the training progress of the ML model, an estimated training time for the training for the ML model, a training execution duration of the ML model, a training start time of the ML model, or a training end time of the ML model.

[0290] Optionally, the information indicating the training progress of the ML model includes information indicating that training on the ML model starts or information indicating that training on the ML model ends.

[0291] Optionally, the transceiver module 1301 may further transmit training status information of the ML model to the second device, where the training status information indicates the running of training for the ML model.

[0292] In a possible implementation, when implementing the functionality of the second device, the transceiver module 1301 may be further configured to receive a machine learning ML training indication based on new data from the first device, where the ML training indication based on new data indicates whether training should be performed based on updated training data of the ML model.

[0293] Optionally, the transceiver module 1301 may be further configured to transmit training control parameters of the ML model to the first device, where the training control parameters of the ML model are used by the communication device to perform training on the ML model.

[0294] Optionally, the transceiver module 1301 may be further configured to receive second information from the first device, where the second information includes at least one of information indicating a training progress of the ML model, an estimated training time for training for the ML model, a training execution duration of the ML model, a training start time of the ML model, or a training end time of the ML model.

[0295] Optionally, the information indicating the training progress of the ML model includes information indicating that training on the ML model starts or information indicating that training on the ML model ends.

[0296] Optionally, the transceiver module 1301 may be further configured to receive training status information of the ML model from the first device, where the training status information indicates the running of training for the ML model.

[0297] Based on the above embodiment, the embodiment of the present application further provides a communication device. The device is configured to implement the communication method for machine learning model training in the above figure. See Figure 14. The communication device 1400 may include at least one of a transceiver 1401, a processor 1402, and a memory 1403. The transceiver 1401, the processor 1402, and the memory 1403 are connected to each other.

[0298] Optionally, the transceiver 1401, the processor 1402, and the memory 1403 are connected to each other via a bus 1404. The bus 1404 may be a peripheral component interconnect (PCI) bus, an extended industry standard architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of indication, the bus is represented by only one bold line in FIG. 14. However, it does not indicate that there is only one bus or only one type of bus.

[0299] The transceiver 1401 is configured to receive and send data to implement communication with another device. For example, the transceiver 1401 may be configured to perform the functions of the transceiver module 1301.

[0300] Optionally, the transceiver may include a communication interface that may be used for communication of the communication device 1400. For example, the communication interface may be configured to implement the functionality of the transceiver module 1301 shown above in a wired manner.

[0301] The processor 1402 is configured to implement the communication method for machine learning model training in the above figures. For details, please refer to the description in the above embodiment. For example, the processor 1402 may be configured to perform the function of the processing module 1302.

[0302] The memory 1403 is configured to store program instructions and the like. Specifically, the program instructions may include program code, and the program code includes computer operation instructions. The memory 1403 may include a random access memory (RAM) and may further include a non-volatile memory, for example, at least one magnetic disk memory. The processor 1402 executes the program instructions stored in the memory 1403 to implement the aforementioned functions and implement the communication method for machine learning model training provided in the aforementioned embodiment.

[0303] It should be understood that at least one of the communication module 1301 and the transceiver 1401 shown above may be configured to perform an act of transmitting information, notification, or message by the first device to the second device or receiving information, notification, or message by the second device from the first device, for example, performing any one of S201, S204, S205, S303, and S304. At least one of the communication module 1301 and the transceiver 1401 may be further configured to perform an act of transmitting information, notification, or message by the second device to the first device, or may be configured to perform an act of receiving information, notification, or message by the first device from the second device, for example, performing any one of S102, S202, and S301.

[0304] At least one of the processing module 1302 and the processor 1402 may be configured to perform processing actions of the first device or the second device. For example, the processing actions of the first device include S103, S203, and S302. The processing actions of the first device and the second device may further include triggering training on an ML model, generating information, notifications, or messages sent by at least one of the communication module 1301 and the transceiver 1401, or processing information, notifications, or messages received by at least one of the communication module 1301 and the transceiver 1401.

[0305] It should be further understood that in the steps performed by at least one of the transceiver module 1301, the transceiver 1401, the processing module 1302, and the processor 1402, for related technical terms, nouns, act implementations, etc., refer to the descriptions of the corresponding technical terms, nouns, and act implementations in Figures 6 to 9 in the method embodiment of the present application. For example, in the act of transmitting a second indication performed by the communication module 1301 and the transceiver 1401, for the meaning and function of the second indication, refer to the description of S201 in Figure 7 of the present application. As another example, for the execution manner of training the ML model based on the first information by at least one of the processing module 1302 and the processor 1402, refer to the description of S103 shown in Figure 6, S203 shown in Figure 7, and S302 shown in Figure 8 of the present application. No further examples are provided.

[0306] Based on the above embodiment, the embodiment of the present application further provides a communication method for machine learning model training. The method is implemented by a first device and a second device. The second device can be configured to send first information to the first device, and the first device can be configured to receive the first information from the second device and perform training on the ML model based on the first information. For details of the method, please refer to the implementation of this specification.

[0307] Based on the above embodiment, an embodiment of the present application further provides a system. The system may include the above-mentioned first device and second device.

[0308] Based on the above embodiment, an embodiment of the present application further provides a computer program, which, when running on a computer, enables the computer to perform the communication method for machine learning model training provided in the above embodiment.

[0309] Based on the above embodiment, an embodiment of the present application further provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a computer, the computer is enabled to perform the communication method for machine learning model training provided in the above embodiment.

[0310] Based on the above embodiment, an embodiment of the present application further provides a chip, which is configured to read a computer program stored in a memory to implement the communication method for machine learning model training provided in the above embodiment. The chip may include a processor.

[0311] Based on the above embodiment, an embodiment of the present application provides a chip system. The chip system includes a processor configured to support a computer device in implementing functions related to at least one of the first device and the second device in the above embodiment. In a possible design, the chip system further includes a memory, and the memory is configured to store programs and data required for the computer device. The chip system may include a chip, or may include a chip and another discrete component.

[0312] In conclusion, the embodiment of the present application provides a communication method for machine learning model training. In this method, a first device (e.g., a model training producer) may train an ML model based on first information from a second device (e.g., a model training consumer). The first information includes at least one of a first indication and a condition for starting training on the ML model. Thus, the training on the ML model performed by the first device is based on the condition for starting training on the ML model determined or indicated by the second device. Thus, the case where the first device starts training on the ML model when training does not need to be performed on the ML model is avoided, and as a result, the waste of resources is reduced.

[0313] It should be understood by those skilled in the art that the embodiments of the present application may be provided as a method, a system, or a computer program product. Thus, the present application may use the form of a hardware-only embodiment, a software-only embodiment, or an embodiment with a combination of software and hardware. In addition, the present application may use the form of a computer program product embodied on one or more computer usable storage media (including, but not limited to, disk memory, CD-ROM, optical memory, etc.) that includes computer usable program code.

[0314] The present application will be described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that computer program instructions can be used to implement each process and / or each block in the flowcharts and / or block diagrams, and combinations of processes and / or blocks in the flowcharts and / or block diagrams. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or any other programmable data processing device to create a machine, such that the instructions executed by the processor of the computer or any other programmable data processing device create an apparatus for implementing a particular function in one or more processes in the flowcharts and / or one or more blocks in the block diagrams.

[0315] These computer program instructions may be stored in a computer readable memory that can indicate to a computer or any other programmable data processing device to act in a particular manner, such that the instructions stored in the computer readable memory generate an artifact that includes an instruction apparatus that implements a particular function in one or more processes in the flowcharts and / or one or more blocks in the block diagrams.

[0316] The computer program instructions may alternatively be loaded onto a computer or another programmable data processing device such that a sequence of operations and steps are performed on the computer or another programmable device to generate a computer-implemented process. Thus, the instructions that run on the computer or another programmable device provide steps for implementing a particular function in one or more procedures in the flowcharts and / or one or more blocks in the block diagrams.

[0317] It is obvious that those skilled in the art can make various modifications and variations to the embodiments of the present application without departing from the scope of the embodiments of the present application. The present application intends to cover these modifications and variations, provided that they fall within the scope of protection defined by the following claims and their equivalent technologies.

[0318] In this application, "plurality" means two or more than two. "And / or" describes an association relationship between related objects and indicates that three relationships may exist. For example, A and / or B may represent the following cases: only A is present, both A and B are present, only B is present, where A and B may be singular or plural. In the text description of this application, the character " / " generally indicates an "or" relationship between related objects. In the formula of this application, the character " / " indicates a "split" relationship between related objects. "Comprising at least one of A, B, or C" may represent the following cases: including A, including B, including C, including A and B, including A and C, including B and C, and including A, B, and C.

[0319] It can be understood that various numbers in the embodiments of the present application are merely used for distinction to simplify the description, and are not used to limit the scope of the embodiments of the present application. The sequence numbers of the above processes do not imply the execution order, and the execution order of the processes should be determined based on the functions and internal logic of the processes.

Claims

1. 1. A communication method for machine learning model training, comprising: receiving, by a first device, first information from a second device, the first information including at least one of a first indication and a training condition of a machine learning ML model, a value range of the first indication including a first value and a second value, the first value indicating that the first device should perform training on the ML model, the second value indicating that the first device should not perform training on the ML model, and a value of the first indication being the first value; training, by the first device, the ML model based on the first information; A communication method comprising:

2. The first information includes the first indication, and the method further comprises: starting the training of the ML model by the first device; sending, by the first device, a second indication to the second device, the second indication being used to notify the start of the training for the ML model; The method of claim 1 further comprising:

3. The method comprises: sending, by the first device, a third indication to the second device, the third indication being used to request training control parameters of the ML model; receiving, by the first device, the training control parameters of the ML model from the second device, the training control parameters of the ML model being used by the first device to perform training on the ML model; The method of claim 1 or 2, further comprising:

4. The training control parameters of the ML model are an identifier for the ML model; Type information of the ML model; the number of training iterations for the ML model; The performance requirements of the ML model; the training time of the ML model, or The data used to train the ML model The method of claim 3 , comprising at least one of:

5. The method of claim 4 , wherein the data used to train the ML model includes at least one of a data source, updated network conditions, and updated training data.

6. The method of claim 1 , wherein the training conditions of the ML model include at least one of a network load condition, a network coverage condition, and a network performance condition.

7. The network load condition is: a physical resource block (PRB) utilization threshold, or User Connection Threshold The method of claim 6 , comprising at least one of:

8. The network coverage condition is: a reference signal received power (RSRP) threshold and a coverage rate threshold for said RSRP threshold in an area; a reference signal reception quality RSRQ threshold and a coverage rate threshold for said RSRQ threshold in an area; or A reference signal to interference plus noise ratio (RSSINR) threshold and a coverage rate threshold for said RSSINR threshold in an area The method according to claim 6 or 7, comprising at least one of the following:

9. The network performance condition is: Handover success rate threshold, Data Energy Efficiency Threshold, or Network Slice Energy Efficiency Threshold The method according to any one of claims 6 to 8, comprising at least one of the following:

10. The method comprises: The method further includes transmitting, by the first device, training status information of the ML model to the second device, the training status information comprising: whether the ML model has been trained, or Whether the ML model is trained based on at least one of a performance evaluation result of the ML model, the updated network state, and the updated training data. The method according to any one of claims 1 to 9, wherein

11. The method comprises: The method further includes transmitting, by the first device, second information to the second device, the second information comprising: information indicative of the training progress of the ML model; an estimated training time for the training of the ML model; the duration of the training run of the ML model; the start time of training of the ML model, or The time to finish training the ML model The method according to claim 1 , comprising at least one of the following:

12. 1. A communication method for machine learning model training, comprising: determining, by a second device, first information, the first information including at least one of a first indication and a training condition of a machine learning ML model, a value range of the first indication including a first value and a second value, the first value indicating that the first device should perform training on the ML model, the second value indicating that the first device should not perform training on the ML model, and a value of the first indication being the first value; transmitting, by the second device, the first information to the first device; A communication method comprising:

13. The first information includes the first indication, and the method further comprises: and receiving, by the second device, a second indication from the first device, the second indication being used to notify the start of the training for the ML model. The method of claim 12.

14. The method comprises: receiving, by the second device, a third indication from the first device, the third indication being used to request training control parameters for the ML model; transmitting, by the second device, the training control parameters of the ML model to the first device, the training control parameters of the ML model being used by the first device to perform training on the ML model; The method of claim 12 or 13, further comprising:

15. The training control parameters of the ML model are an identifier for the ML model; Type information of the ML model; the number of training iterations for the ML model; The performance requirements of the ML model; the training time of the ML model, or The data used to train the ML model The method of claim 14 , comprising at least one of:

16. 16. The method of claim 15, wherein the data used to train the ML model includes at least one of a data source, updated network conditions, and updated training data.

17. The method of claim 12 , wherein the training conditions of the ML model include at least one of a network load condition, a network coverage condition, and a network performance condition.

18. The network load condition is: a physical resource block (PRB) utilization threshold, or User Connection Threshold 20. The method of claim 17, comprising at least one of:

19. The network coverage condition is: a reference signal received power (RSRP) threshold and a coverage rate threshold for said RSRP threshold in an area; a reference signal reception quality RSRQ threshold and a coverage rate threshold for said RSRQ threshold in an area; or A reference signal to interference plus noise ratio (RSSINR) threshold and a coverage rate threshold for said RSSINR threshold in an area 19. The method of claim 17 or 18, comprising at least one of the following:

20. The network performance condition is: Handover success rate, Data Energy Efficiency, or Network Slice Energy Efficiency 20. The method of claim 17, comprising at least one of the following:

21. The method comprises: The method further includes receiving, by the second device, training state information of the ML model from the first device, the training state information comprising: whether the ML model has been trained, or Whether the ML model is trained based on at least one of a performance evaluation result of the ML model, the updated network state, and the updated training data.

21. The method according to claim 12, wherein:

22. The method comprises: The method further includes receiving, by the second device, second information from the first device, the second information comprising: information indicative of the training progress of the ML model; an estimated training time for the training of the ML model; the duration of the training run of the ML model; the start time of training of the ML model, or The time to finish training the ML model 22. The method of claim 12, comprising at least one of the following:

23. 1. A first device for machine learning model training, comprising: a transceiver module configured to receive first information from a second device, the first information including at least one of a first indication and a training condition of a machine learning ML model, a value range of the first indication including a first value and a second value, the first value indicating that the first device should perform training on the ML model, the second value indicating that the first device should not perform training on the ML model, and a value of the first indication being the first value; a processing module configured to train the ML model based on the first information; A first device comprising:

24. The first information includes the first indication, and the processing module: further configured to start the training for the ML model; The transceiver module includes: and further configured to send a second indication to the second device, the second indication being used to notify the start of the training for the ML model.

24. The first device of claim 23.

25. The transceiver module includes: sending a third indication to the second device, the third indication being used to request training control parameters for the ML model; and receiving the training control parameters of the ML model from the second device, the training control parameters of the ML model being used by the first device to perform training on the ML model; 25. The first device of claim 23 or 24, further configured to:

26. The training control parameters of the ML model are an identifier for the ML model; Type information of the ML model; the number of training iterations for the ML model; The performance requirements of the ML model; the training time of the ML model, or The data used to train the ML model 26. The first device of claim 25, comprising at least one of:

27. 27. The first device of claim 26, wherein the data used to train the ML model includes at least one of a data source, updated network conditions, and updated training data.

28. 28. The first device of claim 23, wherein the training conditions of the ML model include at least one of a network load condition, a network coverage condition, and a network performance condition.

29. The network load condition is: a physical resource block (PRB) utilization threshold, or User Connection Threshold 30. The first device of claim 28, comprising at least one of:

30. The network coverage condition is: a reference signal received power (RSRP) threshold and a coverage rate threshold for said RSRP threshold in an area; a reference signal reception quality RSRQ threshold and a coverage rate threshold for said RSRQ threshold in an area; or A reference signal to interference plus noise ratio (RSSINR) threshold and a coverage rate threshold for said RSSINR threshold in an area 30. The first device according to claim 28 or 29, comprising at least one of:

31. The network performance condition is: Handover success rate threshold, Data Energy Efficiency Threshold, or Network Slice Energy Efficiency Threshold 31. A first device according to any one of claims 28 to 30, comprising at least one of:

32. The transceiver module includes: and transmitting training state information of the ML model to the second device, the training state information comprising: whether the ML model has been trained, or Whether the ML model is trained based on at least one of a performance evaluation result of the ML model, the updated network state, and the updated training data.

32. A first device according to any one of claims 23 to 31, wherein:

33. The transceiver module includes: and further configured to transmit second information to the second device, the second information comprising: information indicative of the training progress of the ML model; an estimated training time for the training of the ML model; the duration of the training run of the ML model; the start time of training of the ML model, or The time to finish training the ML model 33. A first device according to any one of claims 23 to 32, comprising at least one of:

34. A second device for machine learning model training, comprising: a processing module configured to determine first information, the first information including at least one of a first indication and a training condition of a machine learning ML model, a value range of the first indication including a first value and a second value, the first value indicating that the first device should train on the ML model, the second value indicating that the first device should not train on the ML model, and a value of the first indication being the first value; a transceiver module configured to transmit the first information to the first device; A second device comprising:

35. The first information includes the first indication, and the transceiver module: and further configured to receive a second indication from the first device, the second indication being used to notify the start of the training for the ML model.

35. A second device according to claim 34.

36. The transceiver module includes: receiving a third indication from the first device, the third indication being used to request training control parameters for the ML model; transmitting the training control parameters of the ML model to the first device, the training control parameters of the ML model being used by the first device to perform training on the ML model; 36. The second device of claim 34 or 35, further configured to:

37. The training control parameters of the ML model are an identifier for the ML model; Type information of the ML model; the number of training iterations for the ML model; The performance requirements of the ML model; the training time of the ML model, or The data used to train the ML model 37. The second device of claim 36, comprising at least one of:

38. 38. The second device of claim 37, wherein the data used to train the ML model includes at least one of a data source, updated network conditions, and updated training data.

39. The second device of claim 34 , wherein the training conditions of the ML model include at least one of a network load condition, a network coverage condition, and a network performance condition.

40. The network load condition is: a physical resource block (PRB) utilization threshold, or User Connection Threshold 40. The second device of claim 39, comprising at least one of:

41. The network coverage condition is: a reference signal received power (RSRP) threshold and a coverage rate threshold for said RSRP threshold in an area; a reference signal reception quality RSRQ threshold and a coverage rate threshold for said RSRQ threshold in an area; or A reference signal to interference plus noise ratio (RSSINR) threshold and a coverage rate threshold for said RSSINR threshold in an area 41. The second device according to claim 39 or 40, comprising at least one of:

42. The network performance condition is: Handover success rate, Data Energy Efficiency, or Network Slice Energy Efficiency 42. A second device according to any one of claims 39 to 41, comprising at least one of:

43. The transceiver module includes: and further configured to receive training state information of the ML model from the first device, the training state information comprising: whether the ML model has been trained, or Whether the ML model is trained based on at least one of a performance evaluation result of the ML model, the updated network state, and the updated training data.

43. A second device according to any one of claims 34 to 42, wherein:

44. The transceiver module includes: and further configured to receive second information from the first device, the second information comprising: information indicative of the training progress of the ML model; an estimated training time for the training of the ML model; the duration of the training run of the ML model; the start time of training of the ML model, or The time to finish training the ML model 44. A second device according to any one of claims 34 to 43, comprising at least one of:

45. 1. A communication method for machine learning model training, comprising: sending, by the first device, to the second device, a machine learning ML training indication based on new data, the new data-based ML training indication indicating whether training is performed based on updated training data of the ML model; Communication methods.

46. The method comprises: and receiving, by the first device, training control parameters of the ML model from the second device, the training control parameters of the ML model being used by the first device to perform training on the ML model.

46. ​​The method of claim 45.

47. The method comprises: The method further includes transmitting, by the first device, second information to the second device, the second information comprising: information indicative of the training progress of the ML model; an estimated training time for the training of the ML model; the duration of the training run of the ML model; the start time of training of the ML model, or The time to finish training the ML model 47. The method of claim 45 or 46, comprising at least one of:

48. 48. The method of claim 47, wherein the information indicating the training progress of the ML model includes information indicating that the training for the ML model is commencing or information indicating that the training for the ML model is terminated.

49. The method comprises: and transmitting, by the first device, training status information of the ML model to the second device, the training status information indicating the progress of the training for the ML model.

49. A method according to any one of claims 45 to 48.

50. 1. A communication method for machine learning model training, comprising: receiving, by the second device, from the first device, a machine learning ML training indication based on new data, the new data-based ML training indication indicating whether training is performed based on updated training data of the ML model; Communication methods.

51. The method comprises: and transmitting, by the second device, the training control parameters of the ML model to the first device, the training control parameters of the ML model being used by the first device to perform training on the ML model.

51. The method of claim 50.

52. The method comprises: The method further includes receiving, by the second device, second information from the first device, the second information comprising: information indicative of the training progress of the ML model; an estimated training time for the training of the ML model; the duration of the training run of the ML model; the start time of training of the ML model, or The time to finish training the ML model 52. The method of claim 50 or 51, comprising at least one of:

53. 53. The method of claim 52, wherein the information indicating the training progress of the ML model includes information indicating that the training for the ML model is commencing or information indicating that the training for the ML model is terminated.

54. The method comprises: and receiving, by the second device, training status information of the ML model from the first device, the training status information indicating the progress of the training for the ML model.

54. A method according to any one of claims 50 to 53.

55. A communication device for machine learning model training, comprising: a communication module configured to send a new data-based machine learning ML training indication to a second device, the new data-based ML training indication indicating whether training is performed based on updated training data of the ML model; Communications equipment.

56. The communication module includes: and further configured to receive training control parameters of the ML model from the second device, the training control parameters of the ML model being used by the communication device to perform training on the ML model.

56. The apparatus of claim 55.

57. The communication module includes: and further configured to transmit second information to the second device, the second information comprising: information indicative of the training progress of the ML model; an estimated training time for the training of the ML model; the duration of the training run of the ML model; the start time of training of the ML model, or The time to finish training the ML model 57. The apparatus of claim 55 or 56, comprising at least one of:

58. 58. The apparatus of claim 57, wherein the information indicating the training progress of the ML model comprises information indicating that the training for the ML model is commencing or information indicating that the training for the ML model is terminated.

59. The communication module includes: and transmitting training status information of the ML model to the second device, the training status information indicating the progress of the training for the ML model.

59. Apparatus according to any one of claims 55 to 58.

60. A communication device for machine learning model training, comprising: a communication module configured to receive a machine learning ML training indication based on new data from a first device, the new data-based ML training indication indicating whether training is performed based on updated training data of the ML model; Communications equipment.

61. The communication module includes: and transmitting training control parameters of the ML model to the first device, the training control parameters of the ML model being used by the first device to perform training on the ML model.

61. The apparatus of claim 60.

62. The communication module includes: and further configured to receive second information from the first device, the second information comprising: information indicative of the training progress of the ML model; an estimated training time for the training of the ML model; the duration of the training run of the ML model; the start time of training of the ML model, or The time to finish training the ML model 62. The apparatus of claim 60 or 61, comprising at least one of the following:

63. 63. The apparatus of claim 62, wherein the information indicating the training progress of the ML model comprises information indicating that the training for the ML model is commencing or information indicating that the training for the ML model is terminated.

64. The communication module includes: and further configured to receive training status information of the ML model from the first device, the training status information indicating the running of the training for the ML model.

64. Apparatus according to any one of claims 60 to 63.

65. 54. A computer readable storage medium, the computer readable storage medium being configured to store a computer program, the computer program being enabled, when run on a computer, to perform a method according to any one of claims 1 to 11, a method according to any one of claims 12 to 22, a method according to any one of claims 45 to 49, or a method according to any one of claims 50 to 54.

66. 54. A computer program product, which when said computer program product is run on a computer, enables said computer to carry out the method according to any one of claims 1 to 11, the method according to any one of claims 12 to 22, the method according to any one of claims 45 to 49 or the method according to any one of claims 50 to 54.

67. A communication system comprising a first device according to any one of claims 23 to 33 and a second device according to any one of claims 34 to 44, or A communication system comprising a communication device according to any one of claims 55 to 59 and a communication device according to any one of claims 60 to 64.

68. 1. A communication method comprising: transmitting, by a second device, first information to a first device, the first information including at least one of a first indication and a training condition of a machine learning ML model, a value range of the first indication including a first value and a second value, the first value indicating that the first device should perform training on the ML model, the second value indicating that the first device should not perform training on the ML model, and a value of the first indication being the first value; training, by the first device, the ML model based on the first information; A communication method comprising:

69. 1. A communication method comprising: sending, by the first device, to the second device, a machine learning ML training indication based on new data, the ML training indication based on new data indicating whether training is performed based on updated training data of the ML model; receiving, by the second device, from the first device, a machine learning ML training indication based on the new data; A communication method comprising:

70. A communications device comprising a processor configured to execute a computer program or instructions stored in a memory to implement a method according to any one of claims 1 to 11, a method according to any one of claims 12 to 22, a method according to any one of claims 45 to 49 or a method according to any one of claims 50 to 54.

71. 71. The communication device of claim 70, further comprising the memory and / or a transceiver, the transceiver being used by the communication device to perform reception and / or transmission.

72. A communications device configured to perform the operations of the method according to any one of claims 1 to 11 or any one of claims 12 to 22 or claims 45 to 54.

73. A chip comprising a processor configured to perform the operations of the method of any one of claims 1 to 11 or any one of claims 12 to 22 or claims 45 to 54.

Images