Communication method and communication apparatus

Abstract

A communication method and a communication apparatus. When there is a training conflict between a first model and a second model, a model training production entity can report conflict information of the first model to a model training consumption entity. On the basis of the conflict information of the first model, the model training consumption entity can determine that there is a training conflict between the first model and the second model, so as to attempt to solve the training conflict between the first model and the second model, helping to reduce effects between the training of the first model and the training of the second model and thus improving the efficiency of model training.

Description

A communication method and a communication apparatus

[0001] The present application claims priority to the Chinese Patent Application No. 202411399805.8, filed on September 30, 2024, and entitled "A communication method and a communication apparatus", the content of which is incorporated herein by reference in its entirety. TECHNICAL FIELD

[0002] The present application relates to the field of communication, and more particularly, to a communication method and a communication apparatus. BACKGROUND

[0003] In order to improve the intelligent and automated level of the communication network, artificial intelligence (AI) / machine learning (ML) technology is being applied in more and more fields. Model training is an important link in AI / ML, which can be executed in a network environment. Since the training process of the model can affect the performance of the network, when multiple models are trained simultaneously using the same network environment, the training of multiple models can affect each other, affecting the training efficiency. SUMMARY

[0004] Embodiments of the present application provide a communication method and a communication apparatus to improve the model training efficiency.

[0005] In a first aspect, embodiments of the present application provide a communication method, which can be executed by a model training production entity. Unless otherwise specified, the "model training production entity" can refer to the model training production entity itself, a component (such as a circuit, a chip or a chip system (such as a modem chip, also known as a baseband chip, or a system on chip (SoC) chip or a system in package (SIP) chip containing a modem core)) in the model training production entity, or a logic module or software capable of realizing all or part of the functions of the model training production entity.

[0006] The method comprises: determining conflict information of a first model, the conflict information being used to indicate that the first model and a second model exist training conflicts; and sending the conflict information of the first model to a model training consumption entity.

[0007] The conflict information is used to indicate that the first model and the second model exist training conflicts, and can be replaced by: the conflict information is used to determine that the first model and the second model exist training conflicts.

[0008] Based on the method, the model training production entity can report the conflict information of the first model to the model training consumption entity in the case that the first model and the second model exist training conflicts, so that the model training consumption entity determines that the first model and the second model exist training conflicts according to the conflict information of the first model, and attempts to solve the training conflicts between the first model and the second model, which helps to reduce the influence between the training of the first model and the training of the second model, thereby improving the efficiency of model training.

[0009] With reference to the first aspect, in a possible implementation, the conflict information of the first model comprises information of the second model.

[0010] Based on the implementation, the model training production entity can provide the information of the second model that exists training conflicts with the first model to the model training consumption entity, so that the model training consumption entity knows that the first model and the second model exist training conflicts. In this way, the processing of the model training consumption entity is simple.

[0011] With reference to the first aspect or any implementation of the first aspect, in another possible implementation, the conflict information of the first model comprises a conflict identifier; and the method comprises: sending, to the model training consumption entity, the conflict information of the second model, the conflict information of the second model comprising the conflict identifier.

[0012] Based on the implementation, the model training production entity can mark the first model and the second model that exist training conflicts with the same conflict identifier, so that the model training consumption entity can determine that the first model and the second model exist training conflicts through the conflict identifier. In this way, the signaling overhead is small.

[0013] With reference to the first aspect or any implementation of the first aspect, in another possible implementation, the conflict information of the first model further comprises at least one of the following information: time information, used to indicate a time period for training the first model; network information, used to indicate a network location where the first model is trained; a first iteration priority, used to determine an iteration order, the iteration order being an order of alternately training models that exist training conflicts; and a conflict resolution suggestion, the conflict resolution suggestion being a suggestion for resolving the training conflicts between the first model and the second model.

[0014] Based on the implementation, the model training production entity can provide some information related to the training of the first model to the model training consumption entity, so that the model training consumption entity can refer to the information when resolving the training conflicts between the first model and the second model, which helps to better resolve the training conflicts between the first model and the second model.

[0015] With reference to the first aspect or any implementation manner of the first aspect, in a possible implementation manner, the method further includes: receiving first information from the model training consumer entity, the first information being used to request that the first operation be performed on the training of the first model; and performing the first operation on the training of the first model according to the first information. And / or, receiving second information from the model training consumer entity, the second information being used to request that the second operation be performed on the training of the second model; and performing the second operation on the training of the second model according to the second information. The first operation and / or the second operation includes at least one of the following operations: canceling the training; suspending the training; modifying a conflict resolution policy used to resolve a training conflict; or modifying at least one of a time period of the training, a network location where the training is performed, or an iteration priority used to determine an iteration order, the iteration order being an order in which the models having the training conflict are iteratively trained in an alternating manner.

[0016] Based on the above implementation manner, the training conflict between the first model and the second model can be resolved by performing the first operation on the training of the first model and / or performing the second operation on the training of the second model.

[0017] With reference to the first aspect or any implementation manner of the first aspect, in a possible implementation manner, the method further includes: receiving third information from the model training consumer entity, the third information being used to request that the first model be trained; and determining the conflict information of the first model includes: determining the conflict information of the first model according to the third information.

[0018] Based on the above implementation manner, the detection of the training conflict is triggered by the request for training of the first model.

[0019] With reference to the first aspect or any implementation manner of the first aspect, in a possible implementation manner, the method further includes: determining that the first model and the second model have a training conflict.

[0020] With reference to the first aspect or any implementation manner of the first aspect, in a possible implementation manner, the determination that the first model and the second model have a training conflict includes: determining that the first model and the second model have a training conflict according to training information of the first model and training information of the second model, wherein the training information includes at least one of the following information: a network location where the model is trained, a time period of the model training, or a network index affected by the model training.

[0021] In a possible implementation manner of the first aspect or any implementation manner of the first aspect, the determining that the first model and the second model have the training conflict according to the training information of the first model and the training information of the second model comprises: when a network position where the first model is trained and a network position where the second model is trained overlap, a time period when the first model is trained and a time period when the second model is trained overlap, and a network index affected when the first model is trained and a network index affected when the second model is trained interfere with each other, it is determined that the first model and the second model have the training conflict.

[0022] In a possible implementation manner of the first aspect or any implementation manner of the first aspect, before the sending, to the model training consumption entity, of the conflict information of the first model, the method further comprises: obtaining a conflict resolution policy of the first model, the conflict resolution policy being used to resolve the training conflict; and determining that the conflict resolution policy cannot resolve the training conflict between the first model and the second model.

[0023] Based on the above implementation manner, the model training production entity reports the training conflict to the model training consumption entity in a case where the model training production entity cannot resolve the training conflict between the first model and the second model, which can reduce signaling interaction between the model training consumption entity and the model training production entity and reduce processing of the model training consumption entity.

[0024] In a possible implementation manner of the first aspect or any implementation manner of the first aspect, the obtaining of the conflict resolution policy of the first model comprises: receiving the conflict resolution policy from the model training consumption entity.

[0025] Based on the above implementation manner, the model training consumption entity can configure the conflict resolution policy for the model training production entity, and the conflict resolution policy is more in line with training requirements of the first model.

[0026] In a possible implementation manner of the first aspect or any implementation manner of the first aspect, the conflict resolution policy is contained in a machine learning training request MLTrainingRequest management object instance (MOI).

[0027] In a possible implementation manner of the first aspect or any implementation manner of the first aspect, the conflict resolution policy comprises at least one of the following strategies: time-sharing training with a model having a training conflict; partition training with the model having the training conflict; and iterative training alternately with the model having the training conflict.

[0028] With reference to the first aspect or any implementation manner of the first aspect, in a possible implementation manner, the conflict resolution policy comprises alternately performing iterative training on the models that exist training conflicts, and the conflict resolution policy further comprises a second iteration priority, the second iteration priority being used to determine an iteration order, the iteration order being an order of alternately performing iterative training on the models that exist training conflicts.

[0029] With reference to the first aspect or any implementation manner of the first aspect, in a possible implementation manner, the conflict information of the first model is contained in a machine learning training process MOI.

[0030] In a second aspect, embodiments of the present disclosure provide a communication method, which can be performed by a model training consumption entity. Unless otherwise specified, the model training consumption entity can refer to the model training consumption entity itself, a component (for example, a circuit, a chip, or a chip system (such as a modem chip, or an SoC chip or a SIP chip containing a modem core), or a logic module or software capable of realizing all or part of the functions of the model training consumption entity) in the model training consumption entity, or the like. The second aspect is a method on the model training consumption entity side corresponding to the first aspect. The same terms or features in the second aspect or the implementation manners of the second aspect as those in the first aspect or the implementation manners of the first aspect can refer to the first aspect or the implementation manners of the first aspect, and the technical effects of the second aspect or the implementation manners of the second aspect can refer to the technical effects in the first aspect or the implementation manners of the first aspect, which will not be described herein again.

[0031] The method comprises: receiving conflict information of a first model from a model training production entity, the conflict information of the first model being used to indicate that the first model and a second model exist training conflicts; determining, according to the conflict information of the first model, that the first model and the second model exist training conflicts; and resolving the training conflicts between the first model and the second model.

[0032] With reference to the second aspect, in a possible implementation manner, the conflict information of the first model comprises information of the second model.

[0033] With reference to the second aspect or any implementation manner of the second aspect, in a possible implementation manner, the conflict information of the first model comprises a conflict identifier. The method further comprises: receiving conflict information of the second model from the model training production entity, the conflict information of the second model comprising the conflict identifier. The determining, according to the conflict information of the first model, that the first model and the second model exist training conflicts comprises: determining, according to the conflict information of the first model and the conflict information of the second model, that the first model and the second model exist training conflicts.

[0034] With reference to the second aspect or any implementation manner of the second aspect, in a possible implementation manner, the conflict information of the first model comprises at least one of the following information: time information, used to indicate a time period in which the first model is trained; network information, used to indicate a network location in which the first model is trained; first iteration priority, used to determine an iteration order, the iteration order being an order in which models having training conflicts are alternately iteratively trained; and conflict resolution suggestion, the conflict resolution suggestion being a suggestion for resolving the training conflict between the first model and the second model.

[0035] With reference to the second aspect or any implementation manner of the second aspect, in a possible implementation manner, resolving the training conflict between the first model and the second model comprises: sending, to the model training production entity, first information, the first information being used to request that a first operation is performed on the training of the first model; and / or sending, to the model training production entity, second information, the second information being used to request that a second operation is performed on the training of the second model. The first operation and / or the second operation comprises at least one of the following operations: canceling training; suspending training; modifying a conflict resolution strategy, the conflict resolution strategy being used to resolve the training conflict; or modifying at least one of a time period of training, a network location of training, or an iteration priority, the iteration priority being used to determine an iteration order, the iteration order being an order in which models having training conflicts are alternately iteratively trained.

[0036] With reference to the second aspect or any implementation manner of the second aspect, in a possible implementation manner, before the conflict information of the first model is received from the model training production entity, the method further comprises: sending, to the model training production entity, third information, the third information being used to request that the first model is trained.

[0037] With reference to the second aspect or any implementation manner of the second aspect, in a possible implementation manner, the method further comprises: sending, to the model training production entity, a conflict resolution strategy of the first model, the conflict resolution strategy being used to resolve the training conflict.

[0038] With reference to the second aspect or any implementation manner of the second aspect, in a possible implementation manner, the conflict resolution strategy is contained in a machine learning training request MOI (MLTrainingRequest MOI).

[0039] With reference to the second aspect or any implementation manner of the second aspect, in a possible implementation manner, the conflict resolution strategy comprises at least one of the following strategies: time-sharing training with a model having a training conflict; partition training with a model having a training conflict; and alternately iteratively training with a model having a training conflict.

[0040] With reference to the second aspect or any implementation manner thereof, in a possible implementation manner, the conflict resolution policy comprises iteratively training the models with training conflicts alternately, and the conflict resolution policy further comprises a second iteration priority, the second iteration priority being used to determine an iteration order, the iteration order being an order of iteratively training the models with training conflicts alternately.

[0041] With reference to the second aspect or any implementation manner thereof, in a possible implementation manner, the conflict information of the first model is contained in a machine learning training process MOI.

[0042] In a third aspect, a communication method is provided, which can be performed by a model training consumer entity and a model training producer entity. Unless specified otherwise, the model training consumer entity or the model training producer entity can refer to the model training consumer entity or the model training producer entity itself, or a component (for example, a circuit, a chip, or a chip system (such as a modem chip, or an SoC chip or a SIP chip containing a modem core), or a logic module or software capable of implementing all or part of the functions of the model training consumer entity or all or part of the functions of the model training producer entity.

[0043] The method comprises: determining, by the model training producer entity, that a first model and a second model have a training conflict; sending, by the model training producer entity, to the model training consumer entity, conflict information of the first model, the conflict information being used to indicate that the first model and the second model have a training conflict; receiving, by the model training consumer entity, the conflict information of the first model from the model training producer entity; determining, by the model training consumer entity, that the first model and the second model have a training conflict according to the conflict information of the first model; and resolving, by the model training consumer entity, the training conflict between the first model and the second model.

[0044] The steps performed by the model training producer entity in the third aspect or the implementation manners thereof can refer to the first aspect or the implementation manners thereof, the steps performed by the model training consumer entity in the third aspect or the implementation manners thereof can refer to the second aspect or the implementation manners thereof, the same terms or features in the third aspect or the implementation manners thereof as those in the first aspect, the implementation manners of the first aspect, the second aspect, or the implementation manners of the second aspect can refer to the first aspect, the implementation manners of the first aspect, the second aspect, or the implementation manners of the second aspect, and the technical effects of the third aspect or the implementation manners thereof can refer to the technical effects in the first aspect, the implementation manners of the first aspect, the second aspect, or the implementation manners of the second aspect, which will not be described herein again.

[0045] In a fourth aspect, embodiments of the present disclosure provide a communication method, which can be performed by a model training production entity. Unless otherwise specified, the model training production entity can refer to the model training production entity itself, a component (e.g., a circuit, a chip, or a chip system (such as a modem chip, or a SoC chip or a SIP chip containing a modem core)) in the model training production entity, or a logic module or software capable of realizing all or part of the functions of the model training production entity.

[0046] The method comprises: receiving a conflict resolution strategy of a first model from the model training consumption entity, the conflict resolution strategy being used to resolve a training conflict; in the case that there is a training conflict between the first model and a second model, resolving the training conflict between the first model and the second model according to the conflict resolution strategy; and training the first model.

[0047] Based on the above method, the model training production entity can receive a conflict resolution strategy of a first model from the model training consumption entity, and in the case that there is a training conflict between the first model and a second model, resolve the training conflict between the first model and the second model according to the conflict resolution strategy, which helps to reduce the influence between the training of the first model and the training of the second model, thereby improving the efficiency of model training.

[0048] In combination with the fourth aspect, in a possible implementation manner, the resolving the training conflict between the first model and the second model according to the conflict resolution strategy comprises: modifying, according to the conflict resolution strategy, a time period for training the first model, a network position at which the first model is trained, or alternately performing iterative training with the second model.

[0049] In combination with the fourth aspect or any implementation manner thereof, in another possible implementation manner, the conflict resolution strategy is contained in a machine learning training request MLTrainingRequest MOI.

[0050] In combination with the fourth aspect or any implementation manner thereof, in another possible implementation manner, the conflict resolution strategy comprises at least one of the following strategies: time-sharing training with a model that has a training conflict; partition training with a model that has a training conflict; and alternately performing iterative training with a model that has a training conflict.

[0051] In combination with the fourth aspect or any implementation manner thereof, in another possible implementation manner, the conflict resolution strategy comprises alternately performing iterative training with a model that has a training conflict, and the conflict resolution strategy further comprises a second iteration priority, the second iteration priority being used to determine an iteration order, the iteration order being an order of alternately performing iterative training with the model that has a training conflict.

[0052] With reference to the fourth aspect or any implementation thereof, in another possible implementation, the method further includes determining that the first model and the second model have a training conflict.

[0053] With reference to the fourth aspect or any implementation thereof, in another possible implementation, the method further includes receiving third information from the model training consumer entity, the third information being used to request training of the first model; and the determining that the first model and the second model have a training conflict includes determining that the first model and the second model have a training conflict according to the third information.

[0054] Based on the above implementation, the detection of the training conflict is triggered by the training request of the first model.

[0055] With reference to the fourth aspect or any implementation thereof, in another possible implementation, the determining that the first model and the second model have a training conflict includes determining that the first model and the second model have a training conflict according to training information of the first model and training information of the second model, wherein the training information includes at least one of the following: a network location where the model is trained, a time period when the model is trained, or a network index affected when the model is trained.

[0056] With reference to the fourth aspect or any implementation thereof, in another possible implementation, the determining that the first model and the second model have a training conflict according to the training information of the first model and the training information of the second model includes: when the network location where the first model is trained overlaps with the network location where the second model is trained, the time period when the first model is trained overlaps with the time period when the second model is trained, and the network index affected when the first model is trained and the network index affected when the second model is trained interfere with each other, determining that the first model and the second model have a training conflict.

[0057] With reference to the fourth aspect or any implementation thereof, in another possible implementation, after the training of the first model is completed, the method further includes sending, to the model training consumer entity, information of the second model and a reason for the training conflict of the second model.

[0058] With reference to the fourth aspect or any implementation thereof, in another possible implementation, the information of the second model and the reason for the training conflict of the second model are included in a machine learning training report MLTrainingReportMOI.

[0059] Based on the above implementation manner, the model training production entity can report the conflict situation of the training of the first model to the model training consumption entity for subsequent model training reference.

[0060] In a fifth aspect, embodiments of the present application provide a communication method, which can be performed by a model training consumption entity. Unless otherwise specified, the "model training consumption entity" can refer to the model training consumption entity itself, a component (such as a circuit, a chip or a chip system (such as a modem chip, or an SoC chip or a SIP chip containing a modem core), or a logic module or software capable of realizing all or part of the functions of the model training consumption entity) in the model training consumption entity, or the like. The fifth aspect is a method on the model training consumption entity side corresponding to the fourth aspect. The same terms or features in the fifth aspect or its implementation manners as those in the fourth aspect or its implementation manners can refer to the fourth aspect or its implementation manners, and the technical effects of the fifth aspect or its implementation manners can refer to those in the fourth aspect or its implementation manners. The fifth aspect will not be described in detail.

[0061] The method comprises: determining a conflict resolution strategy of the first model, the conflict resolution strategy being used to resolve the training conflict; and sending the conflict resolution strategy to a model training production entity.

[0062] In combination with the fifth aspect, in a possible implementation manner, the conflict resolution strategy is contained in a machine learning training request MLTrainingRequest MOI.

[0063] In combination with the fifth aspect or any implementation manner thereof, in another possible implementation manner, the conflict resolution strategy comprises at least one of the following strategies: time-sharing training with the model having the training conflict; partition training with the model having the training conflict; and iterative training alternately with the model having the training conflict.

[0064] In combination with the fifth aspect or any implementation manner thereof, in another possible implementation manner, the conflict resolution strategy comprises iterative training alternately with the model having the training conflict, and the conflict resolution strategy further comprises a second iteration priority, the second iteration priority being used to determine an iteration order, the iteration order being an order of the iterative training alternately with the model having the training conflict.

[0065] In combination with the fifth aspect or any implementation manner thereof, in another possible implementation manner, the method further comprises: sending third information to the model training production entity, the third information being used to request training of the first model.

[0066] With reference to the fifth aspect or any implementation manner of the fifth aspect, in a possible implementation manner, the method further includes: receiving information of the second model from the model training production entity and a reason for the training conflict of the second model.

[0067] With reference to the fifth aspect or any implementation manner of the fifth aspect, in a possible implementation manner, the information of the second model and the reason for the training conflict of the second model are contained in a machine learning training report, MLTrainingReportMOI.

[0068] The sixth aspect provides a communication method, which can be performed by a model training consumption entity and a model training production entity. Unless otherwise specified, the model training consumption entity or the model training production entity can refer to the model training consumption entity or the model training production entity itself, a component (for example, a circuit, a chip, or a chip system (such as a modem chip, or an SoC chip or an SIP chip containing a modem core), or a logic module or software capable of realizing all or part of the functions of the model training consumption entity or all or part of the functions of the model training production entity.

[0069] The method includes: determining, by the model training consumption entity, a conflict resolution strategy of a first model, the conflict resolution strategy being used to resolve a training conflict; sending, by the model training consumption entity, the conflict resolution strategy to a model training production entity; receiving, by the model training production entity, the conflict resolution strategy from the model training consumption entity; resolving, by the model training production entity, the training conflict between the first model and a second model according to the conflict resolution strategy, in a case where the training conflict exists between the first model and the second model; and training, by the model training production entity, the first model.

[0070] The steps performed by the model training production entity in the sixth aspect or the implementation manner of the sixth aspect can refer to the fourth aspect or the implementation manner of the fourth aspect, the steps performed by the model training consumption entity in the sixth aspect or the implementation manner of the sixth aspect can refer to the fifth aspect or the implementation manner of the fifth aspect, the same terms or features in the sixth aspect or the implementation manner of the sixth aspect as those in the fourth aspect, the implementation manner of the fourth aspect, the fifth aspect, or the implementation manner of the fifth aspect can refer to the fourth aspect, the implementation manner of the fourth aspect, the fifth aspect, or the implementation manner of the fifth aspect, and the technical effects of the sixth aspect or the implementation manner of the sixth aspect can refer to the technical effects of the fourth aspect, the implementation manner of the fourth aspect, the fifth aspect, or the implementation manner of the fifth aspect, which will not be described herein again.

[0071] In a seventh aspect, a communication apparatus is provided. The apparatus is configured to perform the method in any one of the preceding aspects or implementation manners. Specifically, the apparatus can include units and / or modules for performing the method in any one of the preceding aspects or implementation manners, such as a processing unit and / or a transceiving unit. The processing unit is configured to perform the processing steps in the method in any one of the preceding aspects or implementation manners. The transceiving unit is configured to perform the transceiving steps in the method in any one of the preceding aspects or implementation manners.

[0072] In an implementation manner, the apparatus is a model training consumer entity or a model training producer entity. When the apparatus is the model training consumer entity or the model training producer entity, the transceiving unit can be a transceiver, or an input / output interface, or a communication interface; and the processing unit can be at least one processor. Optionally, the transceiver is a transceiving circuit. Optionally, the input / output interface is an input / output circuit.

[0073] In another implementation manner, the apparatus is a chip, a chip system, or a circuit used in a model training consumer entity or a model training producer entity. When the apparatus is the chip, the chip system, or the circuit used in the model training consumer entity or the model training producer entity, the transceiving unit can be an input / output interface, an interface circuit, an output circuit, an input circuit, a pin, or related circuitry, etc. on the chip, the chip system, or the circuit; and the processing unit can be at least one processor, a processing circuit, or a logic circuit, etc.

[0074] In an eighth aspect, a communication apparatus is provided. The apparatus includes a memory configured to store a computer program or instructions; and at least one processor configured to execute the computer program or instructions stored in the memory to perform the method in any one of the preceding aspects or implementation manners.

[0075] In an implementation manner, the apparatus is a model training consumer entity or a model training producer entity.

[0076] In another implementation manner, the apparatus is a chip, a chip system, or a circuit used in a model training consumer entity or a model training producer entity.

[0077] In a ninth aspect, a communication apparatus is provided. The apparatus includes at least one processor and a communication interface. The at least one processor is configured to acquire, through the communication interface, a computer program or instructions stored in a memory to perform the method in any one of the preceding aspects or implementation manners. The communication interface can be implemented by hardware or software.

[0078] In an implementation manner, the apparatus further includes the memory.

[0079] In a tenth aspect, a processor is provided. The processor is configured to perform the method in any one of the preceding aspects.

[0080] For the sending and obtaining / receiving operations involved by the processor, if no special description is made, or if it does not conflict with the actual role or internal logic in the related description, it can be understood as the processor output and receive, input, etc. Operation, but also can be understood as the sending and receiving operations performed by the radio frequency circuit and the antenna, and the present application does not limit this.

[0081] In a eleventh aspect, a computer-readable storage medium is provided, the computer-readable medium storing program code for execution by an apparatus, the program code comprising instructions for performing the method provided by any one of the aspects or implementation manners thereof.

[0082] In a twelfth aspect, a computer program product containing instructions is provided, including a computer program or instructions, when the computer program or instructions are run on a computer, the steps of the method provided by any one of the aspects or implementation manners thereof are implemented.

[0083] In a thirteenth aspect, a chip is provided, the chip comprising a processor and a communication interface, the processor reading instructions stored on a memory through the communication interface, and executing the method provided by any one of the aspects or implementation manners thereof. The communication interface can be implemented by hardware or software.

[0084] Optionally, as an implementation manner, the chip further comprises a memory, the memory storing a computer program or instructions, and the processor is configured to execute the computer program or instructions stored on the memory, and when the computer program or instructions are executed, the processor is configured to execute the method provided by any one of the aspects or implementation manners thereof.

[0085] When the method provided by the present application is executed by a chip, the present application does not limit the number of chips that specifically implement the method of the present application, for example, it can be executed by one chip, or two or more chips. And when the number of chips that implement the method of the present application is two or more, the chip manufacturer is not limited, which can be the same manufacturer or different manufacturers.

[0086] In a fourteenth aspect, a communication system is provided, comprising at least one of the model training consumer entity or the model training producer entity described above.

[0087] In a fifteenth aspect, a computer program is provided, when it is run on a computer, the method provided by any one of the aspects or implementation manners thereof is executed. BRIEF DESCRIPTION OF DRAWINGS

[0088] Fig. 1 is a schematic diagram of the life cycle of AI / ML.

[0089] Fig. 2 is a schematic structural diagram of a system architecture suitable for the embodiments of the present application.

[0090] Figure 3 is another schematic structural diagram of a system architecture to which embodiments of the present application are applicable.

[0091] Figure 4 is a schematic flow chart of a communication method 400 provided by the present application.

[0092] Figure 5 is a schematic flow chart of a communication method 500 provided by the present application.

[0093] Figure 6 is a schematic flow chart of a communication method 600 provided by the present application.

[0094] Figure 7 is a schematic flow chart of a communication method 700 provided by the present application.

[0095] Figure 8 is a schematic structural diagram of an apparatus provided by an embodiment of the present application.

[0096] Figure 9 is another schematic structural diagram of an apparatus provided by an embodiment of the present application.

[0097] Figure 10 is a schematic diagram of a chip system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0098] Before introducing embodiments of the present application, the following explanations are made.

[0099] "Indicate" or "indicating" can include for direct indication and for indirect indication, or "indicate" or "indicating" can indicate explicitly and / or implicitly. The first, second, and the like various numerical designations are only for the convenience of description and do not limit the scope of the embodiments of the present application, for example, to distinguish different messages, different information, and the like. "Predefined" can be achieved by pre-storing corresponding codes, tables or other means for indicating related information in the device, and the specific implementation manner is not limited in the present application. The "protocol" referred to can refer to a standard protocol in the communication field, for example, can include a long term evolution (LTE) protocol, a new radio (NR) protocol, and a related protocol applied in a future communication system, and the present application is not limited thereto. The words "example", "for example", "exemplarily", "as (another) example" and the like are used to indicate as an example, illustration or description. Any embodiment or design scheme described as "example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. The terms "include", "contain", "have" and their variants mean "include but are not limited to", unless otherwise specifically emphasized. "Multiple" refers to two or more. "And / or" describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist together, and B exists alone, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects. "At least one of the following" or the like means any combination of these items, including any combination of single item or multiple items. For example, at least one of a, b and c can represent: a, or b, or c, or a and b, or a and c, or b and c, or a, b and c. Where a, b and c can be single or multiple. The description related to the network element A sending a message, information or data to the network element B, and the network element B receiving the message, information or data from the network element A, is intended to indicate which network element the message, information or data is intended to send to, and does not limit whether they are directly sent or indirectly sent via other network elements. "When", "in the case of", "if" and the like all refer to the objective situation in which the device will make corresponding processing, and are not limited to time, and do not require the device to have a judgment action when implemented, nor does it mean that there are other limitations.

[0100] In order to facilitate the understanding of the embodiments of the present application, first, some terms related to the embodiments of the present application are explained.

[0101] 1. Artificial intelligence (AI) / Machine learning (ML)

[0102] To improve the intelligent and automated level of the network, AI / ML technology is being applied in more and more fields.

[0103] Figure 1 is a schematic diagram of the life cycle of AI / ML. As shown in Figure 1, the operation workflow of AI / ML mainly includes: ML model training, ML model testing, ML model inference emulation, ML model deployment, and AI / ML model inference, etc.

[0104] ML model training: including initial training and retraining for a ML model or a group of ML models. ML model training can also include validation of the ML model to evaluate the performance of the ML model when executed on training data and validation data, and the ML model needs to be retrained if the validation result does not meet the expectation.

[0105] ML model testing: testing the validated ML model to evaluate the performance of the trained ML model on the test data. If the test result meets the expectation, the ML model can proceed to the next step. If the test result does not meet the expectation, the ML model needs to be retrained.

[0106] ML model inference emulation: running the ML model for inference in a simulation environment, the purpose of which is to evaluate the inference performance of the ML model in the simulation environment before applying the ML model to the target network or system. This process is an optional process.

[0107] ML model deployment: including the ML model loading process to make the trained ML model available for the target AI / ML inference function. In some cases, ML model deployment can not be performed, for example, when the training function and the inference function coexist, ML model deployment can not be performed.

[0108] AI / ML model inference: using the trained ML model to perform inference through the AI / ML inference function.

[0109] 2. Reinforcement learning (RL)

[0110] RL is one of the AI / ML technologies, in which an agent (e.g., an RL agent) determines a corresponding reward according to a state change of an environment caused by an action, and then makes the action, and the training goal of RL is to maximize the reward.

[0111] Taking a coverage problem analytics use case in a management data analytics (MDA) use case as an example, when RL is applied to the coverage problem analytics use case, the RL agent can be an ML model for coverage problem analytics; the environment can be a real network or a simulation environment; the action can be a value of a tunable parameter in the network, such as a recommended action (e.g., changing the transmission power of the NR sector carrier frequency); the state can be a performance measurement (PM) / key performance indicator (KPI) of the network, such as a reference signal received power (RSRP) distribution, etc.; and the reward can be a score of an RL performance indicator, used to evaluate the PM / KPI.

[0112] 3. ML model training

[0113] The ML model training can be initiated by a model training consumer entity to a model training producer entity, and the description of the model training consumer entity and the model training producer entity can be referred to the description in the system architecture below. Specifically, the model training consumer entity can request the model training producer entity to create a model training request instance (e.g., MLTrainingRequest MOI) and a model training process instance (e.g., MLTrainingProcess MOI) through a model training request (e.g., MLTrainingRequest); the model training producer entity creates the model training request instance and the model training process instance according to the model training request of the model training consumer entity to perform model training, and reports the information of the related instances to the model training consumer entity; and the model training consumer entity can manage the model training process according to the received information, such as performing no operation, cancellation, or suspension, etc.

[0114] Exemplarily, the attributes in the MLTrainingRequest IOC are shown in Table 1, where M represents mandatory, CM represents conditional mandatory, and T represents true, and the description of the related attributes can be referred to the protocol TS 28.105.

[0115] Table 1 Attributes contained in the MLTrainingRequest IOC

[0116] In addition, the MLTrainingRequest IOC can also include environment-related attributes, such as an RL environment attribute, for describing a network environment on which model training is based. Illustratively, this attribute can describe the location of the network environment, devices in the network, and the like. Illustratively, this attribute can describe whether the network environment is a simulated network or a real network.

[0117] Illustratively, the attributes in the MLTrainingProcess IOC are shown in Table 2, where M indicates mandatory, CM indicates conditional mandatory, and T indicates true, and the description of the relevant attributes can refer to the protocol TS 28.105.

[0118] Table 2 Attributes contained in the MLTrainingProcess IOC

[0119] The above describes the related terms involved in the embodiments of the present application, which will not be explained hereinafter.

[0120] The system architecture to which the present application is applicable is described below.

[0121] Embodiments of the present application can be applied to various communication systems, including but not limited to: a 5th generation (5G) system or NR system, an LTE system, a long term evolution-advanced (LTE-A) system, an LTE frequency division duplex (FDD) system, an LTE time division duplex (TDD) system, etc. It can also be applied to future communication systems. In addition, it can also be applied to device to device (D2D) communication, vehicle-to-everything (V2X) communication, machine to machine (M2M) communication, machine type communication (MTC), an internet of things (IoT) communication system, a narrow band-internet of things (NB-IoT) system, or other communication systems. In addition, it can also be extended to similar wireless communication systems, such as wireless-fidelity (WiFi), worldwide interoperability for microwave access (WIMAX), and 3rd generation partnership project (3GPP) related communication systems, etc., without limitation.

[0122] Exemplarily, FIG. 2 is a schematic structural diagram of a system architecture to which embodiments of the present application are applicable.

[0123] The system architecture shown in FIG. 2 includes a model training consumption entity 210 and a model training production entity 220.

[0124] The model training consumer entity 210 is configured to invoke the model training service. The model training consumer entity can also be replaced by a model training invoker entity, a model training consumer, a model training consumer network element, a machine learning training (MLT) user, a machine learning training management service consumer (MLT MnS consumer or ML training MnS consumer), a MLT consumer, a network management system (NMS), a service management and orchestration function (SMO), a cross-domain management node, a cross-domain management system, a cross-domain management function, or a management service consumer (MnS consumer), etc. The model training producer entity 220 is configured to provide the model training service. The model training producer entity can also be replaced by a model training provider entity, a model training provider, a model training producer network element, a MLT provider, a machine learning training management service producer (MLT MnS producer or ML training MnS producer), a MLT producer, an element management system (EMS), a domain management node, a domain management system, a domain management function, or a management service producer (MnS producer), etc. In future communication systems, the model training consumer entity and / or the model training producer entity can also have other names, which are not particularly limited in the present application.

[0125] The model training consumer entity 210 can invoke the model training function or service provided by the model training producer entity 220 through a network interface (e.g., a service based interface).

[0126] The above entities can be network elements in a hardware device, or software functions running on a dedicated hardware, or virtualized functions instantiated on a platform (e.g., a cloud platform). It can be understood that the above entities can be implemented by one device, or can be implemented by multiple devices together. In addition, the above entities can also be functional modules within a system, such as functional modules in a network management system (NMS) or an element management system (EMS), or functional modules within a device, such as one or more functional modules within a network equipment (NE), which can be an access network device or a core network element. As an example, the above entities can be deployed in different EMSs. As another example, the above entities can be deployed in different NEs. The NMS is responsible for the operation, management and maintenance functions of the network, and can also be referred to as a cross-domain management system. The EMS is used to manage one or more network elements of a certain category, and can also be referred to as a domain management system or a single-domain management system. The access network device can be a device that connects the fixed part and the wireless part in a mobile communication system, and is connected to a mobile terminal through an air interface, such as a base station, an evolved NodeB (eNodeB), an access point (AP), a transmission reception point (TRP), a next generation NodeB (gNB), a base station in a future mobile communication system, or an access node in a WiFi system.

[0127] Exemplarily, in a 3rd generation partnership project (3GPP) network domain, the model training consumer entity can be a network management system (NMS). The model training producer entity can be an element management system (EMS), or a network element managed by the EMS, such as a node or device in a radio access network (RAN) (e.g., a base station), or a function or network element of a core network (CN) (e.g., a network data analytics function (NWDAF) having AI training, inference, and other intelligent computing functions).

[0128] Exemplarily, in an open RAN (O-RAN or ORAN) domain, a service management and orchestration (SMO) function can be a model training consumer entity, and network elements (which can be heterogeneous, such as gNBs, NWDAFs, etc.) directly managed by the SMO can be model training producer entities. The role of the SMO in the network architecture is similar to that of the NMS, and the SMO is responsible for the operation, management, and maintenance of various network services and orchestration functions. The network elements directly managed by the SMO can be heterogeneous, such as EMSs, gNBs, NWDAFs, etc.

[0129] For ease of description, the model training consumer entity is referred to as a consumer entity, and the model training producer entity is referred to as a producer entity.

[0130] FIG. 3 is another schematic structural diagram of a system architecture to which embodiments of the present application are applicable.

[0131] FIG. 3 shows several deployment modes of ML training functions and AI / ML inference functions in a 3GPP network domain, in which the RAN domain management function can correspond to the EMS described above. As shown in (a) of FIG. 3, both the ML training function and the AI / ML inference function are located in the 3GPP management system (such as the RAN domain management function). As shown in (b) of FIG. 3, the ML training function is located in the 3GPP management system (such as the RAN domain management function), and the AI / ML inference function is located in the access network device. As shown in (c) of FIG. 3, both the ML training function and the AI / ML inference function are located in the access network device.

[0132] It should be understood that the network architecture and business scenarios described in the embodiments of the present application are for more clearly illustrating the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. It can be known by those skilled in the art that, as the network architecture evolves and new business scenarios appear, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0133] It should also be understood that some embodiments herein take the 5G system as an example to introduce specific solution details. It can be understood that when the solution is used in other communication systems, for example, the LTE system, or future communication systems, the messages, information, or attributes in the solution can be replaced by messages, information, or attributes in other communication systems that can achieve corresponding functions, and the present application does not limit this.

[0134] Currently, different ML models can be trained at the same network location, i.e., different ML models can share the network location, such as the case of online training of RL type models sharing the network location. The training process of the ML models can affect the network indicators, and therefore, when multiple ML models are trained at the same network location, the training of the multiple ML models can affect each other, affecting the training efficiency. For example, a mobility robustness optimization (MRO) model and a load balancing optimization (LBO) model are trained at the same network location, where the MRO model is used to manage the switching of terminals between multiple radio base stations / cells to ensure the service stability of the network to the terminals, and the LBO model is used to manage the switching of terminals between multiple radio base stations / cells to avoid overloading of some base stations. Since the training objectives of the two models are to find the optimal switching strategy, the training processes of the two models affect each other, causing both training processes to fail to converge, and the training to fail to complete.

[0135] To solve the above problems, the present application provides a communication method and device to improve the model training efficiency.

[0136] The method embodiments of the present application are described below.

[0137] FIG. 4 is a schematic flowchart of a communication method 400 provided by the present application.

[0138] The method shown in FIG. 4 can be performed by a consumer entity and a producer entity. Without special description, the “consumer entity” or “producer entity” can refer to the consumer entity or producer entity itself, or a component (such as a circuit, a chip or a chip system (such as a modem chip, or a SoC chip or a SIP chip containing a modem core), or a logic module or software capable of realizing all or part of the functions of the consumer entity or all or part of the functions of the producer entity.

[0139] The method 400 includes at least part of the following content.

[0140] In step 401, the producer entity determines the conflict information of the first model.

[0141] The conflict information of the first model is used to indicate that the first model and the second model have training conflicts. Through the conflict information of the first model, the first model or the training of the first model can be associated with the second model or the training of the second model. The conflict information is used to indicate that the first model and the second model have training conflicts, and can be replaced by: the conflict information is used to determine that the first model and the second model have training conflicts.

[0142] The production entity determines the conflict information of the first model, which can be replaced by the production entity creating the conflict information of the first model, or the production entity generating the conflict information of the first model, and the like.

[0143] The first model can be a model corresponding to a first inference function, and the second model can be a model corresponding to a second inference function. The inference function can include at least one of the following inference functions: an inference function corresponding to the values of the MDA type, an inference function corresponding to the analytics ID(s) of NWDAF, a RAN inference function, or a vendor’s specific extensions. The values of the MDA type can refer to the protocol 3GPP TS 28.104 [2], and the analytics ID(s) of NWDAF can refer to the protocol 3GPP TS 23.288 [3]. For example, the first model is a model associated with an LBO inference function in the RAN inference function, and the second model is a model associated with an MRO inference function in the RAN inference function.

[0144] Embodiments of the present application do not limit the types of the first model and the second model. The type of the first model and the type of the second model can be the same or different. In embodiments of the present application, the model is used to implement one or more functions, so the model can also be replaced by the function. The second model can include one or more models.

[0145] In one example, the first model and the second model have training conflicts, which can mean that the network location where the first model is trained and the network location where the second model is trained overlap, the time period for training the first model and the time period for training the second model overlap, and the network indicators affected when training the first model and the network indicators affected when training the second model affect each other.

[0146] It should be noted that the training of the first model and the training of the second model are in conflict, which can also be replaced by: the training of the first model and the training of the second model are associated with each other, the training of the first model and the training of the second model influence each other, the training request of the first model and the training request of the second model are in conflict, the training request of the first model and the training request of the second model are associated with each other, or the training request of the first model and the training request of the second model influence each other, etc. The "overlap" can mean all or partial overlap. All overlap can also be understood as "same". "Overlap" can also be replaced by having an intersection, covering each other, etc. The "time period" can also be replaced by "training time".

[0147] In another example, the training of the first model and the training of the second model are in conflict, which can also mean that the training process of the first model and the training process of the second model are in conflict. Illustratively, the MLTrainingProcess MOI corresponding to the first model and the MLTrainingProcess MOI corresponding to the second model are in conflict. For example, the MLTrainingProcess MOI corresponding to the model associated with the LBO inference function and the MLTrainingProcess MOI corresponding to the model associated with the LBO inference function are in training conflict.

[0148] It should be noted that the training process of the first model and the training process of the second model are in conflict, which can also be replaced by: the training process of the first model and the training process of the second model are associated with each other, or the training process of the first model and the training process of the second model influence each other, etc.

[0149] The network location where the model is trained can be understood as the network environment, network area, or network range corresponding to the device or entity used to perform the model training, for example, assuming that the device is a base station, the network location corresponding to the device can be the coverage range of the base station. Further, the device or entity used to perform the model training can collect input data for model training and / or output data for model training from the environment corresponding to the network location. The network location where the model is trained can also be described as: the network location used by the trained model, the network location based on the trained model, the network location where the model is trained, the network location used by the model training, or the network location based on the model training. The network location can also be replaced by network environment, network area, or network range, etc. Illustratively, the network location can be identified using an RL environment index.

[0150] Further, the network locations can overlap, which can be understood as: the network environment, the network area, or the network range can overlap. For example, the coverage range of the base station A used for training the model A overlaps with the coverage range of the base station B used for training the model B. It should be understood that the overlapping part between the network location where the first model is trained and the network location where the second model is trained is the network location shared by the first model and the second model.

[0151] The network indicators affected when training the model can include network performance indicators and / or network parameters. For example, the stability of the service of the terminal is affected when training the MRO model. The network indicators can be associated with each other, which can also be replaced by: the network indicators affect each other.

[0152] The network indicators can be associated with each other, which can also be replaced by: the network indicators affect each other.

[0153] The embodiments of the present application do not limit the implementation of the conflict information of the first model.

[0154] In a possible implementation, the conflict information of the first model comprises information of the second model. Exemplarily, the information of the second model can be a name of the second model, an ID of the second model, etc. For example, the information of the second model can be aIMLInferenceName. When the second model comprises a plurality of models, the information of the second model can comprise information of the plurality of models, for example, the information of the second model can be in the form of a list comprising information of the plurality of models, such as the list can comprise a plurality of aIMLInferenceName.

[0155] Corresponding to the implementation, step 401 specifically comprises: the production entity adding or indicating, in the MLTrainingProcess MOI of the first model, information of the second model that has a training conflict with the first model. Similarly, the production entity can add or indicate, in the MLTrainingProcess MOI of the second model, information of the first model that has a training conflict with the second model.

[0156] In another possible implementation, the conflict information of the first model comprises a conflict identifier. Models corresponding to or having the same conflict identifier have a training conflict. In this implementation, the method 400 further comprises: the production entity determining the conflict information of the second model, wherein the conflict information of the second model comprises the same conflict identifier as in the conflict information of the first model. In other words, the production entity adds the same identifier for models or training requests that have a training conflict. Exemplarily, the conflict identifier can be a virtual identifier, which has no actual meaning and is a private implementation of the production entity of each vendor. The conflict identifier can comprise one or more identifiers. When the conflict identifier comprises a plurality of identifiers, it indicates that the first model belongs to a plurality of conflict groups, for example, the first model has a training conflict with the second model, and the production entity assigns the same identifier 1 for the first model and the second model, the first model also has a training conflict with the third model, and the production entity assigns the same identifier 2 for the first model and the third model, in which case the conflict information of the first model comprises identifier 1 and identifier 2. Exemplarily, the conflict identifier can be in the form of a list.

[0157] Corresponding to the implementation, step 401 specifically comprises: the production entity configuring the same conflict identifier for the MLTrainingProcess MOI of the first model and the second model. Wherein, “configuring” can also be described as “generating”, “determining”, “adding”, “creating”, etc.

[0158] In a possible implementation, before the production entity determines the conflict information of the first model, the method 400 further includes: determining, by the production entity, that the first model has a training conflict with the second model. The determination that the first model has a training conflict with the second model can also be replaced with: discovering that the first model has a training conflict with the second model, detecting that the first model has a training conflict with the second model, and the like. Before the production entity determines the conflict information of the first model, the production entity determines that the first model has a training conflict with the second model, which can also be described as: in the case that the first model has a training conflict with the second model, the production entity determines the conflict information of the first model, and the like.

[0159] The embodiments of the present application do not limit the implementation of the production entity determining that the first model has a training conflict with the second model.

[0160] In a possible implementation, the production entity can maintain training information of each model, and the training information can include at least one of the following: a network location where the model is trained, a time period of training the model, or a network index affected when the model is trained. For example, for each model, the production entity can maintain a list containing information such as the network location where the model is trained, the time period of training the model, or the network index affected when the model is trained. The training information of the model can be included in the MLTrainingProcess MOI corresponding to the model, or can be stored separately from the MLTrainingProcess MOI corresponding to the model, without limitation. The network location where the model is trained, the time period of training the model, or the network index affected when the model is trained can also be replaced with: the network location where the model is trained, the time period of training the model, or the network index affected when the model is trained.

[0161] Based on the above implementation, the production entity determines that the first model has a training conflict with the second model, including: the production entity can determine that the first model has a training conflict with the second model according to the training information of the first model and the training information of the second model. Specifically, when the network location where the first model is trained overlaps with the network location where the second model is trained, the time period of training the first model overlaps with the time period of training the second model, and the network index affected when the first model is trained and the network index affected when the second model is trained affect each other, it is determined that the first model has a training conflict with the second model.

[0162] In another possible implementation, the production entity determines that the first model and the second model have a training conflict, including: when allocating network positions and training time periods for the training of the first model, the production entity finds that part or all of the network positions and / or part or all of the training time that are most suitable for the training of the first model have been allocated to the second model, and determines that the network indicators affected when training the first model and the network indicators affected when training the second model affect each other, in which case the production entity determines that the first model and the second model have a training conflict. Optionally, the production entity further determines that the priority of training the second model is higher than the priority of training the first model, and / or further determines that the time of requesting training of the second model is earlier than the time of requesting training of the first model. The priority of training a model can be determined by the production entity when creating the MLTrainingProcess MOI, such as the priority indicated by the priority attribute in the MLTrainingProcess MOI.

[0163] For example, the production entity can first allocate the training time periods and the network positions used for training according to preset rules or default rules for the training of the second model with higher priority, and then attempt to allocate the training time periods and the network positions used for training for the training of the first model with lower priority, and if it is found that part or all of the network positions and / or part or all of the training time that are most suitable for the training of the first model have been allocated to the second model, the production entity determines that the first model and the second model have a training conflict.

[0164] For example, the production entity can first allocate the training time periods and the network positions used for training according to preset rules or default rules for the training of the second model with higher priority, and then attempt to allocate the training time periods and the network positions used for training for the training of the first model with lower priority, and if it is found that part or all of the network positions and / or part or all of the training time that are most suitable for the training of the first model have been allocated to the second model, the production entity determines that the first model and the second model have a training conflict.

[0165] Embodiments of the present application do not limit the timing at which the production entity determines that the first model and the second model have a training conflict.

[0166] In a possible implementation, the production entity can detect whether the first model and the second model have a training conflict when receiving the training request of the first model. That is, the detection of the training conflict is triggered by the training request of the first model. In this case, the method 400 further includes: the consumption entity sends third information to the production entity, and accordingly, the production entity receives the third information from the consumption entity, where the third information is used to request training of the first model. The production entity determines that the first model and the second model have a training conflict, specifically including: the production entity determines that the first model and the second model have a training conflict according to the third information.

[0167] In another possible implementation, the production entity can detect whether the first model has a training conflict with the second model when creating the model training MOI of the first model. That is, the detection of the training conflict is triggered by the creation of the model training MOI of the first model. In this case, the method 400 further includes that the consumption entity sends third information to the production entity, and accordingly, the production entity receives the third information from the consumption entity, where the third information is used to request training of the first model. The production entity determines that the first model has a training conflict with the second model, specifically including that the production entity creates the model training MOI of the first model according to the third information and determines that the first model has a training conflict with the second model in the process of creating the model training MOI of the first model.

[0168] In another possible implementation, the production entity can detect whether the first model has a training conflict with the second model after creating the model training MOI of the first model. That is, the detection of the training conflict is triggered by the completion of the creation of the model training MOI of the first model. In this case, the method 400 further includes that the consumption entity sends third information to the production entity, and accordingly, the production entity receives the third information from the consumption entity, where the third information is used to request training of the first model; and the production entity creates the model training MOI of the first model according to the third information. The production entity determines that the first model has a training conflict with the second model, specifically including that the production entity determines that the first model has a training conflict with the second model after the completion of the creation of the model training MOI of the first model.

[0169] In another possible implementation, the production entity can periodically detect whether the model training maintained by the production entity has a training conflict. The training information of the first model and / or the training information of the second model can change over time. For example, in time period A, the training of the first model and the training of the second model are both performed at network location A, and in time period B, the training of the first model is no longer performed at location A but at network location B, so that in time period B, the first model and the second model will no longer have a training conflict. In this case, the production entity periodically detects whether the first model has a training conflict with the second model, which can more accurately determine whether the first model has a training conflict with the second model.

[0170] Step 402, the production entity sends the conflict information of the first model to the consumption entity, and accordingly, the consumption entity receives the conflict information of the first model from the production entity.

[0171] Step 403, the consumption entity determines that the first model has a training conflict with the second model according to the conflict information of the first model.

[0172] In a case that the conflict information of the first model includes information of the second model, the consuming entity knows that there is a training conflict between the first model and the second model according to that the conflict information of the first model includes information of the second model. It should be noted that in this case, the operation of the consuming entity reading the information of the second model from the conflict information of the first model can be regarded as the operation of determining that there is a training conflict between the first model and the second model, or in this case, the step 403 can not be performed, that is, the consuming entity can not have the operation of determining that there is a training conflict between the first model and the second model.

[0173] In a case that the conflict information of the first model includes the conflict identifier, the consuming entity further receives the conflict information of the second model. The step 403 specifically includes: the consuming entity determines that there is a training conflict between the first model and the second model according to the conflict information of the first model and the conflict information of the second model. For example, when the conflict information of the first model and the conflict information of the second model carry the same conflict identifier, the consuming entity determines that there is a training conflict between the first model and the second model.

[0174] The step 404 is that the consuming entity solves the training conflict between the first model and the second model.

[0175] Based on the method 400, the producing entity can report the conflict information of the first model to the consuming entity in a case that there is a training conflict between the first model and the second model, and the consuming entity can determine that there is a training conflict between the first model and the second model according to the conflict information of the first model, so as to attempt to solve the training conflict between the first model and the second model, which helps to reduce the influence between the training of the first model and the training of the second model, thereby improving the efficiency of model training.

[0176] In a possible implementation, the conflict information of the first model can further include information related to the training of the first model. For example, the conflict information of the first model can include at least one of the following information: time information, network information, first iteration priority, or conflict resolution suggestion. In other words, the producing entity can provide some information related to the training of the first model to the consuming entity for reference when the consuming entity solves the training conflict between the first model and the second model.

[0177] The time information is used to indicate a time period for training the first model. For example, the time information can include a start time of training the first model and an estimated training duration. The estimated training duration can be estimated according to a number of completed epochs of training the first model, a total number of epochs required for training the first model, and a time length consumed for training the first model. One epoch can refer to a process of training once using all samples in a training set.

[0178] The network information is used to indicate a network location where the first model is trained. For example, the network information includes information of a network range (e.g., coverage range) corresponding to a device or entity used to perform the training of the first model. If the device is a base station, the network range can be an identifier of a cell or a geographic coordinate, which is not limited.

[0179] The first iteration priority is used to determine an iteration order. The iteration order can be an order in which the models with training conflicts are alternately trained. The alternately training the models with training conflicts can refer to sequentially performing training on the models with training conflicts in iteration rounds X. For example, if model #1 and model #2 have training conflicts, X rounds of training on model #1 can be performed first, X rounds of training on model #2 can be performed second, X rounds of training on model #1 can be performed third, and so on. In some implementations, X can be 1 by default.

[0180] The conflict resolution suggestion is used to resolve the training conflicts between the first model and the second model. For example, the suggestion can be to cancel or suspend the training of the model with a lower priority. Specifically, if the priority of the training of the first model is lower than the priority of the training of the second model, the consumer entity can cancel or suspend the training of the first model according to the conflict resolution suggestion.

[0181] Similarly, if the production entity sends the conflict information of the second model to the consumer entity, the conflict information of the second model can include information related to the training of the second model, which can be specifically referred to the conflict information of the first model.

[0182] Embodiments of the present application do not limit the sending manner of the conflict information of the first model.

[0183] In a possible implementation, the conflict information of the first model can be contained in the MLTrainingProcess MOI. For example, Table 3 shows attributes contained in the MLTrainingProcess IOC, and the MLTrainingProcess MOI can be instantiated after the MLTrainingProcess IOC is instantiated. In the table, the conflictflag is used to carry the conflict identifier. For example, the conflictflag can be a list, and each table entry can be of int type, such as a value range of 0-65525. The conflictmlModelRef is used to carry the information of the second model. For example, the type of the conflictmlModelRef can be the same as that of the mLModelRef. The MLTrainingProcess IOC can include the conflictflag or the conflictmlModelRef. Optionally, at least one of the time information, the network information, the first iteration priority, or the conflict resolution suggestion can be carried in the conflictmlModelRef. Optionally, at least one of the time information, the network information, the first iteration priority, or the conflict resolution suggestion can also be carried in other newly added attributes. It should be understood that the names of the attributes in Table 3 are only examples.

[0184] Table 3 Attributes contained in the MLTrainingProcess IOC

[0185] It should be noted that the conflictmlModelRef in Table 3 can also be writable. When the conflictmlModelRef is writable, the consuming entity can request the producing entity to modify the attributes in the conflictmlModelRef. For example, the attribute constraints of the conflictflag or the conflictmlModelRef can be as shown in Table 4.

[0186] Table 4 Attribute constraints

[0187] In another possible implementation, the conflict information of the first model can be independent of the MLTrainingProcess MOI of the first model. The conflict information of the first model and the MLTrainingProcess MOI of the first model can be carried in the same message or different messages.

[0188] It should be noted that the above conflict information can also be referred to as information #1, associated information, impact information, etc., and the conflict identifier can also be referred to as an associated identifier, an impact identifier, a first identifier, etc.

[0189] In a possible implementation, before sending the conflict information, the production entity can also attempt to solve the training conflict between the first model and the second model according to a conflict resolution policy. In this case, before sending the conflict information of the first model to the consumption entity, the method 400 further includes: the production entity obtaining a conflict resolution policy of the first model, the conflict resolution policy being used to solve the training conflict; and the production entity determining that the conflict resolution policy cannot solve the training conflict between the first model and the second model. Alternatively, the method 400 further includes: the production entity obtaining a conflict resolution policy of the first model, the conflict resolution policy being used to solve the training conflict; the production entity attempting to solve the training conflict between the first model and the second model; and step 402 specifically includes: in a case where it is determined that the conflict resolution policy cannot solve the training conflict between the first model and the second model, the production entity sending the conflict information of the first model to the consumption entity, and correspondingly, the consumption entity receiving the conflict information of the first model from the production entity.

[0190] The conflict resolution policy can also be described as: solving a requirement, the requirement representing a requirement of solving the training conflict.

[0191] Embodiments of the present application do not limit the implementation of the production entity obtaining the conflict resolution policy of the first model. In a possible implementation, the conflict resolution policy described above can be preconfigured or internally implemented by the production entity, and in this case, the production entity obtaining the conflict resolution policy of the first model can refer to reading the conflict resolution policy of the first model in the production entity. In another possible implementation, the conflict resolution policy of the first model can be provided by the consumption entity, and in this case, the method 400 further includes: the consumption entity sending the conflict resolution policy of the first model to the production entity; and the production entity obtaining the conflict resolution policy of the first model, including: the production entity receiving the conflict resolution policy of the first model from the consumption entity.

[0192] Embodiments of the present application do not limit the specific policy of the conflict resolution policy of the first model. In a possible implementation, the conflict resolution policy of the first model includes at least one of the following policies:

[0193] Policy 1: time-sharing training with a model that has a training conflict, that is, the first model and the second model are trained in different time periods;

[0194] Policy 2: partition training with a model that has a training conflict, that is, the first model and the second model are trained in different network locations;

[0195] Policy 3: Iterative training of the models with training conflicts alternately, that is, without changing the time period and network position of training the first model and training the second model, the first model and the second model are iteratively trained alternately.

[0196] Optionally, policy 1 is suitable for a scenario where the model performance requirement is high. Policy 2 is suitable for a scenario where the model requires rapid online.

[0197] Optionally, when the conflict resolution policy of the first model includes the above-mentioned policy 3, the conflict resolution policy of the first model further includes a second priority, the second priority being used to determine an iteration order, the iteration order being an order of iterative training of the models with training conflicts alternately. In other words, the iteration priority of iterative training alternately can be specified by the consumption entity, so that the production entity determines the order of iterative training alternately according to the priority.

[0198] Embodiments of the present application do not limit the sending manner of the conflict resolution policy of the first model.

[0199] In a possible implementation, the conflict resolution policy of the first model can be contained in a machine learning training request MLTrainingRequest MOI. Illustratively, Table 5 shows attributes contained in the MLTrainingRequest IOC, and the MLTrainingRequest IOC can be instantiated to obtain the MLTrainingRequest MOI. Among them, the conflictPolicy is used for the conflict resolution policy of the first model. Optionally, the MLTrainingRequest IOC further includes an rLEnvironment, the rLEnvironment being used to describe a network environment based on which the model is trained. It should be understood that the names of various attributes in Table 5 are only examples.

[0200] Table 5 Attributes contained in the MLTrainingRequest MOI

[0201] Illustratively, the attribute constraints of the conflictPolicy can be as shown in Table 6.

[0202] Table 6 Attribute constraints

[0203] In another possible implementation, the conflict resolution policy of the first model can be independent of the MLTrainingRequest MOI of the first model. The conflict resolution policy of the first model and the MLTrainingRequest MOI of the first model can be carried in the same message or different messages.

[0204] Embodiments of the present application do not limit the implementation manner in which the consuming entity resolves the training conflict between the first model and the second model.

[0205] In a possible implementation manner, the consuming entity can request the producing entity to perform a first operation on the training of the first model and / or perform a second operation on the training of the second model. The first operation includes at least one of the following operations: canceling the training of the first model, suspending the training of the first model, modifying a conflict resolution strategy of the first model, modifying a time period for training the first model, modifying a network location where the first model is trained, or modifying an iteration priority of the first model, wherein the conflict resolution strategy of the first model is used to resolve the training conflict of the first model, and the iteration priority of the first model is used to determine the iteration order of the first model in the alternating iterative training. The second operation includes at least one of the following operations: canceling the training of the second model, suspending the training of the second model, modifying a conflict resolution strategy of the second model, modifying a time period for training the second model, modifying a network location where the second model is trained, or modifying an iteration priority of the second model, wherein the conflict resolution strategy of the second model is used to resolve the training conflict of the second model, and the iteration priority of the second model is used to determine the iteration order of the second model in the alternating iterative training.

[0206] It should be noted that the consuming entity can implement the operation on the model training by invoking the general service. For example, the consuming entity can construct the MLTrainingRequest by invoking the createMOI operation (createMOI operation) to request the producing entity to create the MLTrainingRequest MOI on the producing entity. For another example, the consuming entity can construct the MLTrainingRequest by invoking the modifyMOIAttributes operation (modifyMOIAttributes operation) to request the producing entity to modify the MLTrainingRequest MOI on the producing entity. For another example, the consuming entity can construct the MLTrainingProcess by invoking the modifyMOIAttributes operation (modifyMOIAttributes operation) to request the producing entity to modify the MLTrainingProcess MOI on the producing entity.

[0207] Exemplarily, Table 7 shows the correspondence between the general service and the model training management service.

[0208] Table 7 Correspondence between general service and model training management service

[0209] Based on the above description, the consumer entity requesting the producer entity to perform the first operation on the training of the first model and / or perform the second operation on the training of the second model can include: the consumer entity sending first information to the producer entity, and correspondingly, the producer entity receiving the first information from the consumer entity, where the first information is used to request the first operation on the training of the first model; and / or, the consumer entity sending second information to the producer entity, and correspondingly, the producer entity receiving the second information from the consumer entity, where the second information is used to request the second operation on the training of the second model. After receiving the first information, the producer entity can perform the first operation on the training of the first model according to the first information, and / or after receiving the second information, the producer entity can perform the second operation on the training of the second model according to the second information. Wherein, the consumer entity can send the first information and / or the second information to the producer entity by calling the general service implementation, thereby realizing the operation on the model training.

[0210] Exemplarily, when the conflict information of the first model does not include time information, network information and iteration priority, the first operation can be canceling the training of the first model, or suspending the training of the first model, or modifying the conflict resolution strategy of the first model. When the conflict information of the first model includes at least one of the time information, the network information or the iteration priority, the first operation can be canceling the training of the first model, or suspending the training of the first model, or modifying the conflict resolution strategy of the first model, or modifying at least one of the time period of training the first model, the network location where the first model is trained or the iteration priority of the first model.

[0211] Exemplarily, when the conflict information of the second model does not include time information, network information and iteration priority, the second operation can be canceling the training of the first model, or suspending the training of the first model, or modifying the conflict resolution strategy of the first model. When the conflict information of the second model includes at least one of the time information, the network information or the iteration priority, the second operation can be canceling the training of the first model, or suspending the training of the first model, or modifying the conflict resolution strategy of the first model, or modifying at least one of the time period of training the first model, the network location where the first model is trained or the iteration priority of the first model.

[0212] It should be noted that performing the first operation on the training of the first model and / or performing the second operation on the training of the second model can also be described as performing the first operation on the training process of the first model and / or performing the second operation on the training process of the second model, or performing the first operation on the training request of the first model and / or performing the second operation on the training request of the second model.

[0213] The embodiments of the present application do not limit the implementation manner of the consumer entity determining the first operation and / or the second operation.

[0214] In a possible implementation, the consumption entity determines the first operation and / or the second operation according to the priority of the training of the first model and the priority of the training of the second model. Illustratively, the consumption entity can give priority to the training of a high-priority model and request the production entity to suspend or cancel the training of a low-priority model. If the priority of the training of the second model is higher than the priority of the training of the first model, the consumption entity can request the production entity to suspend or cancel the training of the first model. If the priority of the training of the second model is the same as the priority of the training of the first model, the consumption entity can give priority to the training of the model in training (that is, the consumption entity gives priority to the model training request in an earlier time).

[0215] In another possible implementation, the consumption entity determines the first operation and / or the second operation according to the training requirement of the first model and the training requirement of the second model, where the training requirement can include a time requirement, a performance requirement, and the like of the model training. Illustratively, the consumption entity determines the first operation and / or the second operation that can satisfy the training requirement of the maximum number of models according to the training requirement of the first model and the training requirement of the second model. For example, if it is determined according to the training requirement of the first model that the training of the first model can be performed after the training of the second model to satisfy the training requirement of the first model, the consumption entity can request the production entity to modify the time period for training the first model, or the consumption entity can request the production entity to suspend the training of the first model and perform the training of the first model after the training of the second model is completed. If it is determined according to the training requirement of the first model that the training of the first model cannot be performed after the training of the second model to satisfy the training requirement of the first model, the consumption entity determines whether there is another network location that can be used for the training of the first model. If there is another network location that can be used for the training of the first model, the consumption entity can request the production entity to modify the network location for training the first model. If there is no other network location that can be used for the training of the first model, the consumption entity gives priority to the training of a high-priority model and requests the production entity to suspend the training of a low-priority model. If the priority of the training of the first model is higher than the priority of the training of the second model, the consumption entity can request the production entity to suspend the training of the second model. If the priority of the training of the models in conflict is the same, the consumption entity gives priority to the training of the model in training (that is, the consumption entity gives priority to the model training request in an earlier time).

[0216] In another possible implementation, when the conflict information of the first model further includes at least one of time information, network information, a first iteration priority, or a conflict resolution suggestion, the consumption entity determines the first operation and / or the second operation according to the time information, the network information, the first iteration priority, or the conflict resolution suggestion.

[0217] Exemplarily, corresponding to the first operation, after obtaining the time period allocated by the production entity for the training of the first model, if the consumption entity finds that the training demand of the first model cannot be met according to the time period reported by the production entity, the consumption entity can change the conflict resolution strategy of the first model to partition training with the model that exists training conflict or iterative training with the model that exists training conflict alternately, and if the change is to iterative training with the model that exists training conflict alternately, the consumption entity further configures the iteration priority of the iterative training according to the priority provided by the production entity (such as the priority indicated by the priority attribute in the MLTrainingProcess MOI). The determination manner of the second operation is similar and will not be described in detail.

[0218] Exemplarily, if the consumption entity finds that the network location allocated for the first model is too different from the network location for inference of the first model according to the network information provided by the production entity, the consumption entity can change the conflict resolution strategy of the first model to time-sharing training with the model that exists training conflict or iterative training with the model that exists training conflict alternately, and if the change is to iterative training with the model that exists training conflict alternately, the consumption entity further configures the iteration priority of the iterative training according to the priority provided by the production entity (such as the priority indicated by the priority attribute in the MLTrainingProcess MOI). The determination manner of the second operation is similar and will not be described in detail.

[0219] Exemplarily, if the consumption entity finds that there are too many models or processes that perform iterative training alternately, it indicates that the load of the network location is too heavy, which may affect the network performance, therefore the consumption entity can change the training process (such as the training of the first model) that has relatively low requirement on the online time to time-sharing training, and change the training process (such as the training of the second model) corresponding to the model that is not sensitive to the network location to partition training.

[0220] Exemplarily, the conflict resolution suggestion of the first model is to change the conflict resolution strategy of the first model to partition training with the model that exists training conflict, and if the consumption entity determines that the training demand of the first model can still be met after the change, the consumption entity changes the conflict resolution strategy of the first model to partition training with the model that exists training conflict. The determination manner of the second operation is similar and will not be described in detail.

[0221] The scheme of FIG. 4 is that the production entity reports to the consumption entity that the first model and the second model exist training conflict, and the consumption entity solves the training conflict between the first model and the second model based on the information reported by the production entity. The present application also provides a scheme in which the consumption entity provides a conflict resolution strategy to the production entity, and the production entity solves the training conflict according to the conflict resolution strategy. The scheme will be described below in combination with FIG. 5.

[0222] FIG. 5 is a schematic flowchart of a communication method 500 provided by the present application.

[0223] The method shown in FIG. 5 can be performed by a consumer entity and a producer entity. Without special description, the "consumer entity" or "producer entity" can refer to the consumer entity or producer entity itself, or a component (such as a circuit, a chip or a chip system (such as a modem chip, or a SoC chip or a SIP chip containing a modem core)) in the consumer entity or producer entity, or a logic module or software capable of realizing all or part of the functions of the consumer entity or all or part of the functions of the producer entity.

[0224] The method 500 includes at least part of the following.

[0225] Step 501: The consumer entity determines a conflict resolution strategy of the first model.

[0226] The conflict resolution strategy of the first model is used to resolve the training conflict related to the first model.

[0227] Embodiments of the present application do not limit the specific strategy of the conflict resolution strategy of the first model. In one possible implementation, the conflict resolution strategy of the first model includes at least one of the following strategies:

[0228] Strategy 1: Time-sharing training with the model that has a training conflict, that is, training the first model and the second model in different time periods;

[0229] Strategy 2: Partition training with the model that has a training conflict, that is, training the first model and the second model in different network locations;

[0230] Strategy 3: Iterative training alternately with the model that has a training conflict, that is, without changing the time period and network location of training the first model and training the second model, the first model and the second model are iteratively trained alternately.

[0231] Optionally, strategy 1 is suitable for a scenario where the model has a higher performance requirement. Strategy 2 is suitable for a scenario where the model requires rapid online.

[0232] Optionally, when the conflict resolution strategy of the first model includes the above-mentioned strategy 3, the conflict resolution strategy of the first model further includes a second priority, and the second priority is used to determine an iteration order, and the iteration order is the order of iteratively training alternately with the model that has a training conflict. In other words, the iteration priority of iteratively training alternately can be specified by the consumer entity, so that the producer entity determines the order of iteratively training alternately according to the priority.

[0233] The embodiments of the present application do not limit the implementation manner of the consumption entity determining the conflict resolution strategy of the first model. In one possible implementation manner, the consumption entity can determine the conflict resolution strategy of the first model according to the training requirement of the first model, wherein the training requirement of the first model can include a time requirement and / or a performance requirement. For example, when it is required to guarantee the model performance of the first model, the conflict resolution strategy of the first model can be strategy 1. For example, when it is required to guarantee the online time of the first model, the conflict resolution strategy of the first model can be strategy 2.

[0234] In step 502, the consumption entity sends the conflict resolution strategy of the first model to the production entity, and correspondingly, the production entity receives the conflict resolution strategy of the first model from the consumption entity.

[0235] The embodiments of the present application do not limit the sending manner of the conflict resolution strategy of the first model. In one possible implementation manner, the conflict resolution strategy of the first model can be contained in the machine learning training request MLTrainingRequest MOI. In another possible implementation manner, the conflict resolution strategy of the first model can be independent of the MLTrainingRequest MOI of the first model. For details, reference can be made to the related description in method 400.

[0236] In step 503, in the case that there is a training conflict between the first model and the second model, the production entity resolves the training conflict between the first model and the second model according to the conflict resolution strategy of the first model.

[0237] The embodiments of the present application do not limit the implementation manner of the production entity resolving the training conflict between the first model and the second model according to the conflict resolution strategy of the first model.

[0238] As an example, if the conflict resolution strategy of the first model is strategy 1, the production entity arranges the training of the first model after the training of the second model.

[0239] As another example, if the conflict resolution strategy of the first model is strategy 2, the production entity finds a network location closest to the default assigned network location for the first model. For example, if the default assigned network location for the training of the first model is base station 1, the production entity finds a base station 2 most similar to the data distribution of base station 1. The similarity of the data distribution can be determined by selecting data related to the training of the first model in a plurality of base stations including base station 2 in a period of time, calculating the Kullback-Leible (KL) divergence of the corresponding data, obtaining the similarity of the data distribution of base station 1 and the plurality of base stations, and thus the plurality of base stations can be sorted according to the similarity of the data distribution.

[0240] As another example, if the conflict resolution strategy of the first model is strategy 3, the production entity trains the first model and the second model alternately according to the second iteration priority in the conflict resolution strategy of the first model and the second iteration priority in the conflict resolution strategy of the second model (the second iteration priority in the conflict resolution strategy of the second model is obtained in the same way as the second iteration priority in the conflict resolution strategy of the first model is obtained).

[0241] At step 504, after resolving the training conflict between the first model and the second model, the production entity trains the first model.

[0242] Based on the method 500, the consumption entity can provide the conflict resolution strategy to the production entity, and the production entity resolves the training conflict between the first model and the second model according to the conflict resolution strategy in the case where it is determined that there is a training conflict between the first model and the second model, which helps to reduce the influence between the training of the first model and the training of the second model, thereby improving the efficiency of model training.

[0243] In a possible implementation, the method 500 further includes that the production entity determines that there is a training conflict between the first model and the second model, and the detailed description can be referred to step 401.

[0244] In a possible implementation, after the training of the first model is completed, the production entity can send the information of the second model and the reason for the training conflict of the second model to the consumption entity. In this case, the method 500 further includes step 505: the production entity sends the information of the second model and the reason for the training conflict of the second model to the consumption entity, and correspondingly, the consumption entity receives the information of the second model and the reason for the training conflict of the second model from the production entity.

[0245] Illustratively, the information of the second model and the reason for the training conflict of the second model can be carried in the MLTrainingReport MOI. Table 8 shows the attributes included in the MLTrainingReport IOC, and the MLTrainingReport MOI can be obtained after the MLTrainingReport IOC is instantiated. Among them, the conflictmlModelRef is used to carry the information of the second model and the reason for the training conflict of the second model. Illustratively, the type of the conflictmlModelRef can be the same as the type of the mLModelRef. It should be understood that the names of the attributes in Table 8 are only examples.

[0246] Table 8 Attributes included in the MLTrainingReport IOC

[0247] Exemplarily, the attribute constraints of the conflictmlModelRef can be as shown in Table 9.

[0248] Table 9 Attribute constraints

[0249] It should be noted that steps 501 and 502 can also not be performed, such as when the conflict resolution policy of the first model is a conflict resolution policy pre-configured in the production entity or implemented internally by the production entity. It should also be noted that the names of various terms in the embodiments of the present application are exemplary, and can be other names in specific implementations or other communication systems. For example, the conflict resolution suggestion can also be referred to as a first suggestion, an associated resolution suggestion, etc.

[0250] The method 400 will be described below in conjunction with specific examples.

[0251] The following examples take the consumer entity as the NMS and the production entity as the EMS to describe the scheme. It should be understood that the method is also applicable to other system architectures mentioned in the present application, such as the system architecture shown in (b) of FIG. 3, the system architecture shown in (c) of FIG. 3, and the system architecture of ML training and AI / ML inference co-deployment within an ORAN management domain.

[0252] It should be noted that the descriptions of the terms and operations in the following examples can be mutually referred to and referenced with the descriptions of the terms or operations in FIG. 4 and FIG. 5 above.

[0253] Example 1

[0254] FIG. 6 is a schematic flowchart of a communication method 600 provided by the present application.

[0255] The method 600 includes at least part of the following contents.

[0256] Step 0, the NMS receives a training request.

[0257] The training request can be a training request for an RL function in the network, for requesting training of a model A. The training request can include a time requirement and / or a performance requirement of the model. Exemplarily, the training request can be issued by a user to the NMS through human-computer interaction.

[0258] Step 601, the NMS initiates an MLTrainingRequest to the EMS to request training of the model A.

[0259] When initiating the MLTrainingRequest, the NMS provides the EMS with an initialized conflict resolution policy for resolving the training conflict related to model A. The conflict resolution policy can take a value of time-sharing training with the model that has a training conflict, partition training with the model that has a training conflict, or iterative training alternately with the model that has a training conflict. For the policy of iterative training alternately with the model that has a training conflict, the NMS provides the EMS with an iteration priority.

[0260] The manner in which the NMS provides the EMS with the initialized conflict resolution policy can refer to the method 400, and will not be described in detail.

[0261] Step 602, the EMS creates the MLTrainingRequest MOI and the MLTraningProcess MOI.

[0262] The EMS can receive multiple requests related to the MLTraningRequest including the request in step 601, and for each request (or model or function), the EMS can perform the following operations:

[0263] 1) create the MLTraniningRequest MOI and the MLTraningProcess MOI, wherein the MLTrainingProcess MOI contains the priority of the request in local training;

[0264] 2) maintain a list related to the request, which describes the training environment allocated by the production entity for the request, and can include at least one of the following information: network location for model training, time period for model training, or network index affected by model training.

[0265] Step 603, the production entity detects whether multiple requests affect each other.

[0266] Exemplarily, the production entity determines whether multiple requests conflict according to the lists corresponding to the multiple requests respectively. In the following cases, the production entity determines that two requests affect each other: the network locations for model training corresponding to the two requests overlap, the time periods for model training corresponding to the two requests overlap, and the network indexes affected by model training corresponding to the two requests affect each other. The network indexes affect each other can mean that the network indexes overlap or conflict, for example, the training goals of the energy saving model and the bandwidth guarantee model are contrary.

[0267] The implementation of step 603 can refer to step 401, and will not be described in detail.

[0268] Step 604, if the production entity detects that some or all of the requests in the plurality of requests affect each other, the production entity performs conflict avoidance operation according to the conflict resolution policy provided by the NMS, and generates conflict information. The requests affecting each other include the requests in step 601.

[0269] The implementation of the production entity performing conflict avoidance operation according to the conflict resolution policy provided by the NMS can refer to step 503, and the description of the conflict information can refer to the description in method 400, which will not be described in detail.

[0270] Step 605, the EMS sends a notification of the plurality of MLTrainingProcess creations to the NMS.

[0271] The notification of the plurality of MLTrainingProcess creations corresponds to the plurality of requests respectively. The request affecting other requests corresponds to the notification containing the conflict information. The implementation of the notification operation in step 605 can refer to the subscription / notification in TS28.532.

[0272] Step 606, the NMS judges whether the training of the model corresponding to each request meets the training requirement.

[0273] Step 607, if it is judged that the training of the model corresponding to some or all of the plurality of requests does not meet the training requirement, the NMS triggers to modify the attribute (such as cancel request attribute, suspend request attribute, or conflict resolution policy) in the MLTrainingRequest MOI of the plurality of requests or modify the attribute (such as training time period, network location, or iteration / update logic) in the MLTrainingProcess MOI.

[0274] Step 608, the EMS performs model training.

[0275] Step 609, the EMS returns the training result to the NMS.

[0276] Optionally, the training result includes a conflict list, and the conflict list includes the information of other training conflicting with the training, such as the information of the model of the other training. Optionally, the training result includes the reason of the training conflict.

[0277] Exemplarily, the EMS returns the training result to the NMS through the notification of the MLTraningReport IOC creation.

[0278] Based on the method 600, the NMS can configure a conflict resolution policy to the EMS, and the EMS attempts to resolve the training conflict according to the conflict resolution policy and negotiates with the NMS, which helps to avoid multi-model ML training conflicts caused by shared environments.

[0279] Example 2

[0280] Unlike example 1, in example 2, the NMS does not provide a conflict resolution policy to the EMS, and after the EMS detects a training conflict, the conflict is reported to the NMS, and the NMS resolves the training conflict.

[0281] FIG. 7 is a schematic flowchart of a communication method 700 provided by the present application.

[0282] The method 700 includes at least part of the following contents.

[0283] Step 0, the NMS receives a training request.

[0284] The training request can be a training request for an RL function in the network, which is used to request training of a model A. The training request can include time requirements and / or performance requirements of the model. Illustratively, the training request can be issued by a user to the NMS through human-computer interaction.

[0285] Step 701, the NMS initiates an MLTrainingRequest to the EMS to request training of the model A.

[0286] When initiating the MLTrainingRequest, the NMS provides an initialized conflict resolution policy to the EMS, which is used to resolve training conflicts related to the model A. The value of the conflict resolution policy can be time-sharing training with a model that has a training conflict, partition training with a model that has a training conflict, or iterative training alternately with a model that has a training conflict. For the strategy of iterative training alternately with a model that has a training conflict, the NMS provides an iteration priority to the EMS.

[0287] The way in which the NMS provides the initialized conflict resolution policy to the EMS can refer to the method 400, which will not be described in detail.

[0288] Step 702, the EMS creates an MLTrainingRequest MOI and an MLTraningProcess MOI.

[0289] The EMS can receive multiple requests related to the MLTraningRequest including the request in step 601, and for each request (or model or function), the EMS can perform the following operations:

[0290] 1) create MLTraniningRequest MOI and MLTraningProcess MOI, wherein the MLTrainingProcess MOI contains the priority of the request in local training;

[0291] 2) maintain a list related to the request, which describes the training environment allocated by the production entity for the request, and can include at least one of the following information: network location where model training is located, time period of model training, or network indicators affected by model training.

[0292] Step 703, the production entity detects whether multiple requests affect each other.

[0293] Exemplarily, the production entity determines whether multiple requests conflict according to the lists corresponding to the multiple requests respectively. In the following cases, the production entity determines that two requests affect each other: the network locations where the two requests correspond to model training overlap, the time periods of model training corresponding to the two requests overlap, and the network indicators affected by the model training corresponding to the two requests affect each other. Network indicators affecting each other can be that there is overlap between network indicators, or that there is conflict between network indicators, for example, the training goals of energy saving model and bandwidth guarantee model are contrary.

[0294] The implementation of step 703 can refer to step 401, which will not be described in detail.

[0295] Step 704, if the production entity detects that part or all of the multiple requests affect each other, the production entity generates conflict information. The mutually affecting requests include the requests in step 701. The description of the conflict information can refer to the description in method 400, which will not be described in detail.

[0296] Step 705, the EMS sends a notification of multiple MLTrainingProcess creations to the NMS.

[0297] Among them, the notification of multiple MLTrainingProcess creations corresponds to multiple requests respectively. The notification corresponding to the request affecting other requests contains conflict information. The implementation of the notification operation in step 605 can refer to the subscription / notification of TS28.532.

[0298] Step 706, the NMS modifies part or all of the requests according to the status of all existing requests, so that the training needs of the maximum number of requests are met.

[0299] For example, if request 1 can be executed after request 2 and the execution after request 2 can meet the training requirement of request 1, the NMS adjusts the attributes in the MLTrainingProcess MOI to suspend request 1, if request 1 cannot be executed after request 2 and / or the execution after request 2 cannot meet the training requirement of request 1, the NMS tries to adjust the environment (such as rLEnvironment) in the MLTrainingRequest MOI, if there is no other available environment, the NMS prioritizes request 2 with high priority and suspends request 1 with low priority. If the priorities of request 1 and request 2 are the same, the NMS prioritizes the request in training (that is, the NMS prioritizes the request with earlier time).

[0300] In step 707, the NMS triggers to modify the attributes (such as cancel request, suspend request, or conflict resolution strategy) in the MLTrainingRequest MOI of part or all of the requests or modify the attributes (such as training time period, network location, or iteration / update logic) in the MLTrainingProcess MOI.

[0301] In step 708, the EMS performs model training.

[0302] In step 709, the EMS returns the training result to the NMS.

[0303] Optionally, the training result includes a conflict list, and the conflict list includes information of other training in conflict with the current training, such as information of the model of the other training. Optionally, the training result includes the reason for the training conflict.

[0304] For example, the EMS returns the training result to the NMS through the notification created by the MLTraningReport IOC.

[0305] Based on the method 700, the EMS can present the training conflict, and the NMS can adjust according to the conflict, which helps to avoid the multi-model ML training conflict caused by the shared network environment.

[0306] The above describes the method embodiments provided by the present application in detail in combination with FIG. 1 to FIG. 7, and the device embodiments of the present application will be described in combination with FIG. 8 to FIG. 10.

[0307] It can be understood that, in order to implement the functions in the above embodiments, the apparatuses in FIGS. 8 to 10 include hardware structures and / or software modules corresponding to the respective functions, and the apparatuses can be used to implement the functions of the production entity or the consumption entity in the above method embodiments, and thus the beneficial effects of the above method embodiments can also be achieved. Those skilled in the art should clearly understand that, in combination with the units and method steps of the examples described in the embodiments disclosed in the present application, the present application can be implemented in the form of hardware or a combination of hardware and computer software.

[0308] FIG. 8 is a structural schematic diagram of an apparatus provided by an embodiment of the present application.

[0309] The embodiments of the present application can divide the functional units of the production entity or the consumption entity according to the above method examples, for example, each functional unit can be divided according to each function, or two or more functions can be integrated in one unit, and each function can be implemented in the form of hardware or in the form of a software function module. It should be noted that the division shown in FIG. 8 is illustrative, and is only a logical functional division, and another division mode can be used in actual implementation.

[0310] As shown in FIG. 8, the apparatus 10 includes a transceiver unit 11 and a processing unit 12.

[0311] When the apparatus 10 is used to implement the functions of the production entity in the above method embodiments, the transceiver unit 11 is configured to perform the transceiving steps of the production entity, such as steps 402, 502, 504, 601, 605, 607, 609, 701, 706, 707, and 709, and the processing unit 12 is configured to perform the processing steps of the production entity, such as steps 401, 503, 602, 603, 604, 608, 702, 703, 704, and 708. When the apparatus 10 is used to implement the functions of the consumption entity in the above method embodiments, the transceiver unit 11 is configured to perform the transceiving steps of the consumption entity, such as steps 0, 402, 502, 504, 601, 605, 607, 609, 701, 706, 707, and 709, and the processing unit 12 is configured to perform the processing steps of the consumption entity, such as steps 403, 404, 501, and 606.

[0312] Optionally, the apparatus 10 further includes a storage unit 13 configured to store instructions and / or data.

[0313] For more detailed description of the transceiver unit 11 and the processing unit 12, reference can be made to the related description in the above method embodiments, which will not be described herein.

[0314] FIG. 9 is another structural schematic diagram of an apparatus provided by an embodiment of the present application.

[0315] The apparatus 20 comprises a processor 21. The processor 21 is coupled to a memory 23 for storing instructions. When the apparatus 20 is configured to implement the method described above, the processor 21 is configured to execute the instructions in the memory 23 to implement the functions of the processing unit 12 described above.

[0316] Optionally, the apparatus 20 further comprises a memory 23 for implementing the functions of the storage unit 13 described above.

[0317] Optionally, the apparatus 20 further comprises an interface circuit 22. The interface circuit can be referred to as a communication interface. The processor 21 and the interface circuit 22 are coupled to each other. It can be understood that the interface circuit 22 can be a transceiver or an input / output interface. When the apparatus 20 is configured to implement the method described above, the processor 21 is configured to execute the instructions to implement the functions of the processing unit 12 described above, and the interface circuit 22 is configured to implement the functions of the transceiver unit 11 described above.

[0318] Optionally, the apparatus 20 can be a production entity or a consumption entity, and accordingly, the interface circuit can be a transceiver.

[0319] Optionally, the apparatus 20 can be a chip applied to a production entity or a consumption entity, and accordingly, the interface circuit can be an input / output interface.

[0320] Illustratively, when the apparatus 20 is a chip applied to a production entity or a consumption entity, the chip implements the functions of the production entity or the consumption entity in the method embodiments described above. The chip receives information from other modules (such as a radio frequency module or an antenna) in the production entity or the consumption entity, and the information is sent by other devices to the production entity or the consumption entity; or the chip sends information to other modules (such as a radio frequency module or an antenna) in the production entity or the consumption entity, and the information is sent by the production entity or the consumption entity to other devices.

[0321] FIG. 10 is a schematic diagram of a chip system according to an embodiment of the present application. The chip system 30 (or also referred to as a processing system) comprises a logic circuit 31 and an input / output interface 32.

[0322] The logic circuit 31 can be a processing circuit in the chip system 30. The logic circuit 31 can be coupled to a storage unit to call instructions in the storage unit, so that the chip system 30 can implement the methods and functions of the embodiments of the present application. The input / output interface 32 can be an input / output circuit in the chip system 30, which outputs information processed by the chip system 30 or inputs data or signaling information to be processed by the chip system 30 for processing.

[0323] As an option, the chip system 30 can further comprise a storage unit.

[0324] As a solution, the chip system 30 is configured to implement the operations performed by the production entity or the consumption entity in the above method embodiments.

[0325] For example, the logic circuit 31 is configured to implement the processing-related operations performed by the production entity or the consumption entity in the above method embodiments; and the input / output interface 32 is configured to implement the sending and / or receiving-related operations performed by the production entity or the consumption entity in the above method embodiments.

[0326] The present application also provides a communication apparatus, comprising a processing circuit and a memory, the memory being configured to store computer programs or instructions and / or data, and the processing circuit being configured to execute the computer programs or instructions stored in the memory, or read the data stored in the memory, to implement the methods in the above method embodiments. Optionally, the processing circuit is one or more. Optionally, the communication apparatus comprises the memory. Optionally, the memory is one or more. Optionally, the memory is integrated with the processing circuit, or is separately arranged.

[0327] The present application also provides a chip, comprising a processing circuit and a memory, the memory being configured to store computer programs or instructions, and the processing circuit being configured to execute the computer programs or instructions stored in the memory to implement the methods performed by the production entity or the consumption entity in the above method embodiments. The memory can be located in the chip, or can be independent of the chip, located outside the chip, which is not limited herein.

[0328] The present application also provides a computer readable storage medium, which stores computer instructions for implementing the methods performed by the production entity or the consumption entity in the above method embodiments.

[0329] The present application also provides a computer program product, comprising instructions, which are executed by a computer to implement the methods performed by the production entity or the consumption entity in the above method embodiments.

[0330] The present application also provides a computer program, which is executed by a computer to implement the methods performed by the production entity or the consumption entity in the above method embodiments.

[0331] The present application also provides a communication system, which comprises at least one of the production entity or the consumption entity in the above embodiments.

[0332] The explanations and beneficial effects of the related contents in any of the above apparatuses can refer to the corresponding method embodiments provided above, which will not be repeated here.

[0333] It can be understood that the processing circuit in the embodiments of the present application can be a processor or a circuit in the processor for performing processing operations, which can include one or a combination of a central processing unit (CPU), a digital signal processor (DSP), a microprocessor unit (MPU), a microcontroller unit (MCU), a graphics processing unit (GPU), a field programmable gate array (FPGA), an artificial intelligence processor (AI processor) or a neural processing unit (NPU).

[0334] The above-mentioned memory can include one or more of the following storage media: random access memory (RAM), static RAM (SRAM), dynamic RAM (DRAM), phase-change memory (PCM), resistive RAM (ReRAM), magnetoresistive RAM (MRAM), ferroelectric RAM (FRAM), cache, register, read-only memory (ROM), flash memory, erasable programmable ROM (EPROM), hard disk, etc. In one example, computer program instructions for executing the above-mentioned embodiments can be stored on a non-volatile memory, such as at least part of the above-mentioned memory 23 or storage unit (such as one or more of ROM, flash memory, EPROM, or hard disk).

[0335] The method steps in the embodiments of the present application can be realized by hardware or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, which can be stored in a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, a register, a hard disk, a mobile hard disk, a compact disc read-only memory (CD-ROM), or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor, so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an application specific integrated circuit (ASIC). In addition, the ASIC can be located in a production entity or a consumption entity. Of course, the processor and the storage medium can also exist as discrete components in the production entity or the consumption entity.

[0336] In the above embodiments, all or part of the embodiments can be realized by software, hardware, firmware, or any combination thereof. When realized by software, all or part of the embodiments can be realized in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When loaded and executed by a computer, all or part of the processes or functions described in the embodiments are executed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, a network device, a user equipment, or other programmable apparatus. The computer program or instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another, for example, the computer program or instructions can be transferred from one website, computer, server, or data center to another by wired or wireless means. The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be a magnetic medium, such as a floppy disk, a hard disk, a magnetic tape; an optical medium, such as a digital video disc; or a semiconductor medium, such as a solid state disk.

[0337] In various embodiments of the present application, the terms and / or descriptions of different embodiments are consistent and can be referred to each other if there is no special description and logical conflict. The technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationship.

[0338] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. It is to be understood that the above description is intended to be illustrative and not restrictive. The examples set forth herein are not intended to be exhaustive or to be unduly limit the scope of the application. Many modifications and variations to the examples described herein will be apparent to those of ordinary skill in the art upon employing the teachings of the application presented herein. It is, therefore, to be understood that changes can be made in the form, details, and / or arrangement of parts without departing from the scope of the application.

Claims

1. A communication method characterized by comprising: The method comprises: determining conflict information of the first model, the conflict information being used to indicate that the first model and the second model have training conflicts; sending the conflict information of the first model to a model training consumption entity.

2. The method of claim 1, wherein the conflict information of the first model comprises information of the second model.

3. The method of claim 1, wherein the conflict information of the first model comprises a conflict identifier; the method comprises: sending the conflict information of the second model to the model training consumption entity, the conflict information of the second model comprising the conflict identifier.

4. The method according to claim 2 or 3, characterized in that, The conflict information of the first model further comprises at least one of the following information: time information, used to indicate a time period for training the first model; network information, used to indicate a network location where the first model is trained; a first iteration priority, used to determine an iteration order, the iteration order being an order of alternately performing iteration training on models having training conflicts; a conflict resolution suggestion, the conflict resolution suggestion being a suggestion for resolving training conflicts between the first model and the second model.

5. The method according to any one of claims 1 to 4, characterized in that, The method further comprises: receiving first information from the model training consumption entity, the first information being used to request a first operation to be performed on training of the first model; performing the first operation on training of the first model according to the first information; and / or, receiving second information from the model training consumption entity, the second information being used to request a second operation to be performed on training of the second model; performing the second operation on training of the second model according to the second information; wherein the first operation and / or the second operation comprises at least one of the following operations: canceling training; suspending training; modifying a conflict resolution strategy, the conflict resolution strategy being used to resolve training conflicts; or modifying at least one of a time period for training, a network location where training is performed, or an iteration priority, the iteration priority being used to determine an iteration order, the iteration order being an order of alternately performing iteration training on models having training conflicts.

6. The method of any one of claims 1 to 5, wherein the method further comprises: receiving third information from the model training consumption entity, the third information being used to request training of the first model; the determining of the conflict information of the first model comprises: determining the conflict information of the first model according to the third information.

7. The method according to any one of claims 1 to 6, characterized in that, The method further comprises: determining that the first model and the second model have training conflicts.

8. The method of claim 7, wherein, The determining that the first model and the second model have training conflicts comprises: determining that the first model and the second model have training conflicts according to training information of the first model and training information of the second model, wherein the training information comprises at least one of the following information: a network location where a model is trained, a time period for training a model, or a network indicator affected during training of a model.

9. The method of claim 8, wherein, The determining that the first model and the second model have training conflicts according to the training information of the first model and the training information of the second model comprises: When there is an overlap between a network location where the first model is trained and a network location where the second model is trained, an overlap between a time period when the first model is trained and a time period when the second model is trained, and an interaction between a network index affected when the first model is trained and a network index affected when the second model is trained, it is determined that there is a training conflict between the first model and the second model.

10. The method according to any one of claims 1 to 9, characterized in that, Before the model training consumption entity is sent the conflict information of the first model, the method further includes: obtaining a conflict resolution strategy of the first model, the conflict resolution strategy being used to resolve the training conflict; determining that the conflict resolution strategy cannot resolve the training conflict between the first model and the second model.

11. The method of claim 10, wherein, The obtaining of the conflict resolution strategy of the first model includes: receiving the conflict resolution strategy from the model training consumption entity.

12. The method according to claim 10 or 11, characterized in that, The conflict resolution strategy includes at least one of the following strategies: time-sharing training with a model that has a training conflict; partition training with a model that has a training conflict; iterative training alternately with a model that has a training conflict.

13. The method of claim 12, wherein, The conflict resolution strategy includes iterative training alternately with a model that has a training conflict, and the conflict resolution strategy further includes a second iteration priority, the second iteration priority being used to determine an iteration order, the iteration order being an order of iterative training alternately with a model that has a training conflict.

14. A communication method, comprising: The method includes: receiving conflict information of a first model from a model training production entity, the conflict information of the first model being used to indicate that the first model has a training conflict with a second model; determining, according to the conflict information of the first model, that the first model has a training conflict with the second model; resolving the training conflict between the first model and the second model.

15. The method of claim 14, wherein: the conflict information of the first model includes information of the second model.

16. The method of claim 14, wherein: the conflict information of the first model includes a conflict identifier; the method further includes: receiving conflict information of the second model from the model training production entity, the conflict information of the second model including the conflict identifier; the determining, according to the conflict information of the first model, that the first model has a training conflict with the second model includes: determining, according to the conflict information of the first model and the conflict information of the second model, that the first model has a training conflict with the second model.

17. The method according to claim 15 or 16, characterized in that, The conflict information of the first model includes at least one of the following information: time information, used to indicate a time period when the first model is trained; network information, used to indicate a network location where the first model is trained; a first iteration priority, used to determine an iteration order, the iteration order being an order of iterative training alternately with a model that has a training conflict; a conflict resolution suggestion, the conflict resolution suggestion being a suggestion for resolving the training conflict between the first model and the second model.

18. The method according to any one of claims 14 to 17, characterized in that, The resolving of the training conflict between the first model and the second model includes: sending first information to the model training production entity, the first information being used to request that a first operation be performed on the training of the first model; and / or, sending second information to the model training production entity, the second information being used to request that a second operation be performed on the training of the second model; wherein the first operation and / or the second operation comprises at least one of the following operations: canceling the training; suspending the training; modifying a conflict resolution policy, the conflict resolution policy being used to resolve a training conflict; or modifying at least one of a time period of the training, a network location where the training is performed, or an iteration priority, the iteration priority being used to determine an iteration order, the iteration order being an order in which the models having the training conflict are iteratively trained in an alternating manner.

19. The method according to any one of claims 14 to 17, characterized in that, Before the receiving the conflict information of the first model from the model training production entity, the method further comprises: sending third information to the model training production entity, the third information being used to request that the first model be trained.

20. The method of claim 19, wherein, The method further comprises: sending a conflict resolution policy of the first model to the model training production entity, the conflict resolution policy being used to resolve a training conflict.

21. The method of claim 20, wherein, The conflict resolution policy comprises at least one of the following policies: training the models having the training conflict in a time-sharing manner; training the models having the training conflict in a partitioned manner; iteratively training the models having the training conflict in an alternating manner.

22. The method of claim 21, wherein, The conflict resolution policy comprises iteratively training the models having the training conflict in an alternating manner, and the conflict resolution policy further comprises a second iteration priority, the second iteration priority being used to determine an iteration order, the iteration order being an order in which the models having the training conflict are iteratively trained in an alternating manner.

23. A method of communication, comprising: The method comprises: a model training production entity determining conflict information of a first model, the conflict information being used to indicate that the first model has a training conflict with a second model; the model training production entity sending the conflict information of the first model to a model training consumption entity; the model training consumption entity receiving the conflict information of the first model from the model training production entity; the model training consumption entity determining, according to the conflict information of the first model, that the first model has a training conflict with the second model; the model training consumption entity resolving the training conflict between the first model and the second model.

24. A communications device, characterized by The apparatus comprises modules or units for implementing the method of any one of claims 1 to 23.

25. A communications device, characterized by The apparatus comprises a processor and an interface circuit for receiving signals from other apparatuses outside the communication apparatus and transmitting signals to the processor or sending signals from the processor to other apparatuses outside the communication apparatus, the processor being used to implement the method of any one of claims 1 to 23 by means of logic circuit or executing code instructions.

26. The communication apparatus according to claim 25, wherein The communication apparatus is a chip or a chip system.

27. A computer-readable storage medium, characterized in that, The storage medium stores a computer program or instructions, which, when executed by the communication apparatus, implement the method of any one of claims 1 to 23.

28. A communication system, characterized by comprising at least one of: a communication device for performing the method according to any one of claims 1 to 13, or a communication device for performing the method according to any one of claims 14 to 22.

29. A computer program product, characterised in that, computer program or instructions, which, when executed in a computer, implement the steps of the method according to any one of claims 1 to 23.

Images