WO2023098860A1 - 通信方法和通信装置 - Google Patents

通信方法和通信装置 Download PDF

Info

Publication number
WO2023098860A1
WO2023098860A1 PCT/CN2022/136138 CN2022136138W WO2023098860A1 WO 2023098860 A1 WO2023098860 A1 WO 2023098860A1 CN 2022136138 W CN2022136138 W CN 2022136138W WO 2023098860 A1 WO2023098860 A1 WO 2023098860A1
Authority
WO
WIPO (PCT)
Prior art keywords
information
model
training
indicate
data
Prior art date
Application number
PCT/CN2022/136138
Other languages
English (en)
French (fr)
Inventor
孙琰
柴晓萌
孙雅琪
邱宇珩
吴艺群
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Priority to KR1020247021743A priority Critical patent/KR20240117116A/ko
Priority to EP22900664.8A priority patent/EP4429309A1/en
Publication of WO2023098860A1 publication Critical patent/WO2023098860A1/zh
Priority to US18/679,849 priority patent/US20240320488A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/391—Modelling the propagation channel
    • H04B17/3913—Predictive models, e.g. based on neural network models
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W16/00—Network planning, e.g. coverage or traffic planning tools; Network deployment, e.g. resource partitioning or cells structures
    • H04W16/18—Network planning tools
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/02—Arrangements for optimising operational condition
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W8/00—Network data management
    • H04W8/22—Processing or transfer of terminal data, e.g. status or physical capabilities
    • H04W8/24—Transfer of terminal data

Definitions

  • the present disclosure relates to the field of communication, and more particularly, to a communication method and a communication device.
  • a wireless communication network such as a mobile communication network
  • services supported by the network are becoming more and more diverse, and therefore requirements to be met are becoming more and more diverse.
  • the network needs to be able to support ultra-high speeds, ultra-low latency, and/or very large connections.
  • This feature makes network planning, network configuration and/or resource scheduling more and more complex.
  • functions of the network become more and more powerful, such as supporting higher and higher spectrum, supporting high-order multiple input multiple output (MIMO) technology, supporting beamforming, and/or supporting new technologies such as beam management, etc. technology, making network energy saving a hot research topic.
  • MIMO multiple input multiple output
  • These new requirements, new scenarios, and new features have brought unprecedented challenges to network planning, O&M, and efficient operations.
  • artificial intelligence technology can be introduced into the wireless communication network to realize network intelligence. Based on this, how to effectively implement artificial intelligence in the network is a problem worth studying.
  • the present disclosure provides a communication method and a communication device, which can obtain an intelligent model meeting communication performance requirements while reducing air interface resource overhead.
  • a communication method is provided, and the method may be executed by a terminal or a module (such as a chip) configured (or used) in the terminal.
  • the method includes: receiving first information, the first information being used to indicate a training policy of the first intelligent model; performing model training on the first intelligent model according to the training policy.
  • the network can inform the terminal of the training strategy of the smart model, and the terminal executes the training of the smart model based on the training strategy provided by the network, which reduces the need for the access network node to send the model parameters of the smart model to the terminal (such as the weight value, activation function, bias, etc.) can make the model obtained by terminal training match the model adopted by the network without affecting the transmission efficiency of service data as much as possible, and meet the requirements of the application Performance requirements in actual communication.
  • the model parameters of the smart model such as the weight value, activation function, bias, etc.
  • the training strategy includes one or more of the following:
  • Model training method loss function information, model initialization method, type of model optimization algorithm, or parameters of the optimization algorithm.
  • the above model training method may be, but not limited to, one of supervised learning, unsupervised learning or reinforcement learning.
  • the above loss function information may indicate a loss function used to train the first intelligent model, and the loss function may be but not limited to a cross-entropy loss function or a square loss function. Or the loss function can also be a machine learning intelligent model.
  • the above model initialization manner may be an initialization manner of the weight of each neuron of the first intelligent model.
  • the first information may indicate that the model is initialized in such a way that the initial weight of each neuron is a random value within a preset range.
  • the first information includes a start value and an end value of the preset range.
  • the first information may indicate that the initial weights of neurons are all 0.
  • the first information may indicate that the initial weight of the neuron is generated using a preset function.
  • the first information may indicate identification information of one of multiple predefined initialization manners.
  • model optimization algorithm may be an adaptive momentum estimation (adaptive momentum estimation, ADAM) algorithm, a stochastic gradient descent algorithm (stochastic gradient descent, SGD) or a batch gradient descent algorithm (batch gradient descent, BGD).
  • ADAM adaptive momentum estimation
  • SGD stochastic gradient descent
  • BGD batch gradient descent
  • parameters of the above optimization algorithm may include but not limited to one or more of the following:
  • the method further includes: acquiring second information, where the second information is used to indicate the structure of the first smart model.
  • the access network node notifies the terminal of the structure of the first intelligent model
  • the terminal generates an initial first intelligent model with the structure based on the instruction of the access network node, and trains the first intelligent model based on the indicated training strategy
  • the first intelligent model obtained by the terminal training can be matched with the model adopted by the network device to meet the performance requirements of the access network nodes, so that the first intelligent model can be used to improve the performance of wireless communication.
  • the second information is specifically used to indicate one or more of the following structural information of the first smart model:
  • Network layer structure information information, dimensions of input data or dimensions of output data.
  • the access network node can notify the terminal of one or more of the network layer structure information (such as the number of neural network layers, type, etc.), the dimension of input data, and the dimension of output data through the second information.
  • the amount of information is far less than the amount of information such as the weight value, activation function, and bias of each neuron in the neural network, and the resource occupancy rate can be reduced to a small extent, so that terminal training can obtain the first intelligent model that meets performance requirements.
  • the network layer structure information includes one or more of the following:
  • the number of neural network layers contained in the first intelligent model the type of the neural network layer, the use method of the neural network layer, the cascading relationship between the neural network layers, the dimension of the input data of the neural network layer or the neural network Dimensions of the layer's output data.
  • the method further includes: receiving third information, where the third information is used to indicate information of a training data set, and the training data set is used to train the first intelligence Model.
  • the training data set may include training samples, or training samples and labels.
  • the information of the training data set may indicate the type of the training sample, or the type of the training sample and the type of the label.
  • the information of the training data set may indicate the usage of the first intelligent model.
  • the third information may include a training data set.
  • the access network node may also notify the terminal of the training data set used when training the first intelligent model.
  • the first intelligent model obtained by the terminal training can be closer to the requirement of the access network node, or in other words, better match the intelligent model of the access network node.
  • the method further includes: sending capability information, where the capability information is used to indicate the capability of running the smart model.
  • the terminal can use the capability information to inform the access network node terminal device of the capability of running the smart model, so that the access network node can indicate to the terminal the training strategy and/or model structure that meets the terminal capability requirements based on the terminal capability.
  • the capability information is specifically used to indicate one or more of the following capabilities: whether to support running the smart model, the type of smart model that supports running, data processing capability or storage ability.
  • the method further includes: receiving fourth information, where the fourth information is used to indicate test information, and the test information is used to test the performance of the first smart model.
  • the access network node may also send test information to the terminal, so that the terminal can test the first intelligent model based on the test information.
  • the test information includes one or more of the following:
  • Test data information For evaluation methods or performance evaluation parameters.
  • the method further includes: sending fifth information, where the fifth information is used to indicate a test result of the first smart model.
  • the terminal After testing the first intelligent model based on the test information, the terminal determines whether the trained first intelligent model meets the performance requirement, and notifies the access network node of the test result through the fifth information.
  • the sixth information is sent, the sixth information is used to indicate inference data, the inference data is obtained by reasoning test data of the first intelligent model, and the test information includes The test data.
  • the terminal can reason the test data to obtain the reasoning data, and send the reasoning data to the access network node through the sixth information, so that the access network node can determine whether the first intelligent model trained by the terminal meets the performance requirement based on the reasoning data .
  • the method further includes: receiving seventh information, where the seventh information is used to indicate the updated training strategy of the first intelligent model, and/or, using Indicates the updated structure of the first smart model.
  • the access network node determines that the first intelligent model trained by the terminal does not meet the performance requirements, the access network node notifies the terminal of the updated training strategy and/or the updated model structure through the seventh information, and the terminal can base on The seventh information trains the first intelligent model again.
  • the seventh information is specifically used to indicate at least one change amount of the training strategy, and/or at least one change amount of the structure of the first intelligent model.
  • the seventh information may specifically indicate at least one variation of the training strategy and/or at least one variation of the model structure, which can reduce information overhead.
  • a communication method is provided, and the method may be executed by an access network node or a module (such as a chip) configured (or used) in the access network node.
  • the method includes: sending first information, where the first information is used to indicate a training policy of the first intelligent model.
  • the training strategy is the same as the training strategy introduced in the first aspect, and for the sake of brevity, details are not repeated here.
  • the method further includes: sending second information, where the second information is used to indicate the structure of the first smart model.
  • the structural information specifically indicated by the second information is the same as that introduced in the first aspect, and for the sake of brevity, details are not repeated here.
  • the method further includes: sending third information, where the third information is used to indicate information used for a training data set, and the training data set is used to train the first An intelligent model.
  • the method further includes: receiving capability information, where the capability information is used to indicate the capability of running the smart model.
  • the training strategy and/or the structure of the first intelligent model can be determined according to the capability information.
  • the capabilities specifically indicated by the capability information are the same as those described in the first aspect, and will not be repeated here for brevity.
  • determining the training strategy and/or the structure of the first intelligent model according to the capability information includes: determining the first intelligent model according to the capability information and/or, train the first intelligent model according to the training strategy.
  • the method further includes:
  • test information is the same as that introduced in the first aspect, and will not be repeated here for brevity.
  • the method further includes:
  • the fifth information is used to indicate the test result of the first intelligent model.
  • the sixth information is received, the sixth information is used to indicate reasoning data, the reasoning data is obtained by reasoning the test data of the first intelligent model, and the test information includes the test data.
  • the seventh information is specifically used to indicate at least one parameter change amount of the training strategy, and/or at least one parameter change amount of the structure of the first intelligent model .
  • a communication method is provided.
  • the method can be executed by a terminal or a module (such as a chip) configured on (or used for) the terminal.
  • the method is executed by the terminal as an example for description below.
  • the method includes: receiving second information, the second information is used to indicate the structure of the first intelligent model, the second information includes network layer structure information of the first intelligent model, dimensions of input data or dimensions of output data One or more types of structural information; according to the second information, determine the structure of the first intelligent model; perform model training on the first intelligent model.
  • the information about the network layer structure is the same as that introduced in the first aspect, and for the sake of brevity, details are not repeated here.
  • the method further includes: acquiring first information, the first information being used to indicate the training strategy of the first intelligent model; and, the first intelligent model Performing model training includes: performing model training on the first intelligent model according to the training strategy.
  • the training strategy is the same as that introduced in the first aspect, and for the sake of brevity, details are not repeated here.
  • the method further includes: receiving third information, where the third information is used to indicate information of a training data set, and the training data set is used to train the first intelligence Model.
  • the method further includes: sending capability information, where the capability information is used to indicate the capability of running the smart model.
  • the capability information is the same as that introduced in the first aspect, and will not be repeated here for brevity.
  • the method further includes: receiving fourth information, where the fourth information is used to indicate test information, and the test information is used to test the performance of the first smart model.
  • test information is the same as that introduced in the first aspect, and will not be repeated here for brevity.
  • the method further includes: sending fifth information, where the fifth information is used to indicate the test result of the first intelligent model; and/or, sending sixth information , the sixth information is used to indicate reasoning data, the reasoning data is obtained by reasoning the test data of the first intelligent model, and the test information includes the test data.
  • the method further includes: receiving seventh information, where the seventh information is used to indicate the updated training strategy of the first intelligent model, and/or, using Indicates the updated structure of the first smart model.
  • the seventh information is specifically used to indicate at least one change amount of the training strategy, and/or at least one change amount of the structure of the first intelligent model.
  • a communication method is provided, and the method may be executed by an access network node or a module (such as a chip) configured (or used) in the access network node.
  • the method includes: sending second information, the second information is used to indicate the structure of the first intelligent model, the second information includes network layer structure information of the first intelligent model, dimensions of input data or dimensions of output data One or more structural information.
  • the information about the network layer structure is the same as that introduced in the first aspect, and for the sake of brevity, details are not repeated here.
  • the method further includes: sending first information, where the first information is used to indicate a training policy of the first intelligent model.
  • the training strategy is the same as that introduced in the first aspect, and for the sake of brevity, details are not repeated here.
  • the method further includes: sending third information, the third information is used to indicate the information of the training data set, and the training data set is used to train the first intelligence Model.
  • the method further includes: receiving capability information, where the capability information is used to indicate the capability of running the smart model.
  • the capability information is the same as that introduced in the first aspect, and will not be repeated here for brevity.
  • the method further includes: sending fourth information, where the fourth information is used to indicate test information, and the test information is used to test the performance of the first smart model.
  • test information is the same as that introduced in the first aspect, and will not be repeated here for brevity.
  • the method further includes: receiving fifth information, where the fifth information is used to indicate the test result of the first intelligent model; and/or, receiving sixth information , the sixth information is used to indicate reasoning data, the reasoning data is obtained by reasoning the test data of the first intelligent model, and the test information includes the test data.
  • the method further includes: sending seventh information, where the seventh information is used to indicate the updated training strategy of the first intelligent model, and/or, using Indicates the updated structure of the first smart model.
  • the seventh information is specifically used to indicate at least one change amount of the training strategy, and/or at least one change amount of the structure of the first intelligent model.
  • a communication device may include a one-to-one corresponding module for performing the method/operation/step/action described in the first aspect.
  • the module may be a hardware circuit, or a However, software may also be realized by combining hardware circuits with software.
  • the device includes: a transceiver unit, configured to receive first information, the first information being used to indicate a training strategy of the first intelligent model; a processing unit, configured to, according to the training strategy, Do model training.
  • the sixth aspect provides a communication device.
  • the device may include a one-to-one corresponding module for executing the method/operation/step/action described in the second aspect.
  • the module may be a hardware circuit or a However, software may also be realized by combining hardware circuits with software.
  • the device includes: a processing unit, configured to determine the training strategy of the first intelligent model; a transceiver unit, configured to send first information, the first information being used to indicate the training strategy of the first intelligent model.
  • a seventh aspect provides a communication device.
  • the device may include a one-to-one corresponding module for executing the method/operation/step/action described in the third aspect.
  • the module may be a hardware circuit or However, software may also be realized by combining hardware circuits with software.
  • the device includes: a transceiver unit, configured to receive second information, the second information is used to indicate the structure of the first intelligent model, the second information includes network layer structure information of the first intelligent model, input One or more structural information in the dimensions of the data or the dimensions of the output data; the processing unit is used to determine the structure of the first intelligent model according to the second information; Model for model training.
  • a communication device may include a one-to-one corresponding module for executing the method/operation/step/action described in the fourth aspect.
  • the module may be a hardware circuit or However, software may also be realized by combining hardware circuits with software.
  • the device includes: a processing unit, configured to determine the structure of the first intelligent model; a transceiver unit, configured to send second information, the second information is used to indicate the structure of the first intelligent model, and the second information It includes one or more structural information of the network layer structure information of the first intelligent model, the dimension of input data or the dimension of output data.
  • a communication device including a processor.
  • the processor may implement the first aspect or the third aspect and the method in any possible implementation manner of the first aspect or the third aspect.
  • the communication device further includes a memory, and the processor is coupled to the memory, and can be used to execute instructions in the memory, so as to realize any possibility in the first aspect or the third aspect and the first aspect or the third aspect method in the implementation.
  • the communication device further includes a communication interface, and the processor is coupled to the communication interface.
  • the communication interface may be a transceiver, a pin, a circuit, a bus, a module or other types of communication interfaces, without limitation.
  • the communication device is a terminal.
  • the communication interface may be a transceiver, or an input/output interface.
  • the communication device is a chip configured in a terminal.
  • the communication interface may be an input/output interface.
  • the transceiver may be a transceiver circuit.
  • the input/output interface may be an input/output circuit.
  • a communication device including a processor.
  • the processor may implement the method in the second aspect or the fourth aspect and any possible implementation manner of the second aspect or the fourth aspect.
  • the communication device further includes a memory, the processor is coupled to the memory, and can be used to execute instructions in the memory, so as to realize the second aspect or the fourth aspect and any possibility in the second aspect or the fourth aspect method in the implementation.
  • the communication device further includes a communication interface, and the processor is coupled to the communication interface.
  • the communication device is an access network node.
  • the communication interface may be a transceiver, or an input/output interface.
  • the communication device is a chip configured in an access network node.
  • the communication interface may be an input/output interface.
  • the transceiver may be a transceiver circuit.
  • the input/output interface may be an input/output circuit.
  • a processor including: an input circuit, an output circuit, and a processing circuit.
  • the processing circuit is configured to receive a signal through the input circuit and transmit a signal through the output circuit, so that the processor executes the method in any possible implementation manner of the first aspect to the fourth aspect and the first aspect to the fourth aspect .
  • the above-mentioned processor can be one or more chips
  • the input circuit can be an input pin
  • the output circuit can be an output pin
  • the processing circuit can be a transistor, a gate circuit, a flip-flop and various logic circuits, etc. .
  • the input signal received by the input circuit may be received and input by, for example but not limited to, the receiver
  • the output signal of the output circuit may be, for example but not limited to, output to the transmitter and transmitted by the transmitter
  • the circuit may be the same circuit, which is used as an input circuit and an output circuit respectively at different times.
  • the present disclosure does not limit the specific implementation manners of the processor and various circuits.
  • a computer program product includes: a computer program (also referred to as code, or an instruction), which, when the computer program is executed, causes the computer to perform the above-mentioned first to fourth aspects. aspect and the method in any possible implementation manner of the first aspect to the fourth aspect.
  • a computer-readable storage medium stores a computer program (also referred to as code, or instruction) when it is run on a computer, so that the computer executes the above-mentioned first aspect to the fourth aspect and the method in any possible implementation manner of the first aspect to the fourth aspect.
  • a computer program also referred to as code, or instruction
  • a communication system including the aforementioned at least one device for implementing the method for a terminal and at least one device for implementing the method for an access network node.
  • FIG. 1 is a schematic diagram of a communication system provided by the present disclosure
  • FIG. 2A is an example diagram of an application framework of AI in a communication system provided by the present disclosure
  • FIGS. 2B to 2E are schematic diagrams of the network architecture provided by the present disclosure.
  • Fig. 2F is a schematic diagram of the layer relationship of the neural network provided by the present disclosure.
  • Fig. 3 is a schematic diagram of the training process of the intelligent model provided by the present disclosure
  • Fig. 4 is a schematic flowchart of the communication method provided by the present disclosure.
  • FIG. 5A is a schematic diagram of the input-output relationship of neurons when the dropout method is not used in the present disclosure
  • FIG. 5B is a schematic diagram of the input-output relationship of neurons when the dropout method is adopted in the present disclosure
  • FIG. 5C is another schematic flowchart of the communication method provided by the present disclosure.
  • Fig. 6 is another schematic flowchart of the communication method provided by the present disclosure.
  • Fig. 7 is a schematic block diagram of an example of a communication device provided by the present disclosure.
  • Fig. 8 is a schematic structural diagram of an example of a terminal device provided by the present disclosure.
  • Fig. 9 is a schematic structural diagram of an example of a network device provided by the present disclosure.
  • At least one (item) can also be described as one (item) or multiple (items), and multiple (items) can be two (items), three (items), four (items) or more Multiple (items), without limitation.
  • “/" can indicate that the associated objects are an "or” relationship, for example, A/B can indicate A or B; "and/or” can be used to describe that there are three relationships between associated objects, for example, A and / Or B, can mean: A alone exists, A and B exist at the same time, and B exists alone, where A and B can be singular or plural.
  • words such as “first”, “second”, “A”, or “B” may be used to distinguish technical features with the same or similar functions.
  • the words “first”, “second”, “A”, or “B” do not limit the quantity and execution order.
  • words such as “first”, “second”, “A”, or “B” are not necessarily different.
  • Words such as “exemplary” or “such as” are used to indicate examples, illustrations or illustrations, and any design described as “exemplary” or “such as” should not be construed as being more preferred or better than other design solutions.
  • Advantage. The use of words such as “exemplary” or “for example” is intended to present related concepts in a specific manner for easy understanding.
  • FIG. 1 is a schematic structural diagram of a communication system 1000 to which the present disclosure can be applied.
  • the communication system includes a radio access network (radio access network, RAN) 100 and a core network (core network, CN) 200.
  • the communication system 1000 may also include the Internet 300 .
  • the radio access network 100 may include at least one access network node (such as 110a and 110b in FIG. 1 ), and may also include at least one terminal (such as 120a-120j in FIG. 1 ).
  • the terminal is connected to the access network node through wireless communication.
  • the access network nodes are connected to the core network through wireless communication or wired communication.
  • the core network device and the access network node may be independent and different physical devices; or may be the same physical device that integrates the functions of the core network device and the access network node; or may be other possible situations, such as a
  • the function of the access network node and some functions of the core network equipment can be integrated on the physical device, and another physical device realizes the rest of the functions of the core network equipment.
  • the present disclosure does not limit the physical existence forms of core network devices and access network nodes. Terminals may be connected to each other in a wired or wireless manner.
  • the access network nodes may be connected to each other in a wired or wireless manner.
  • FIG. 1 is only a schematic diagram, and the communication system may also include other network devices, such as wireless relay devices and wireless backhaul devices.
  • the access network node may be an access network device, such as a base station (base station), a node B (Node B), an evolved node B (evolved NodeB, eNodeB or eNB), a transmission reception point (transmission reception point, TRP), a The next generation Node B (next generation NodeB, gNB) in the fifth generation (5th generation, 5G) mobile communication system, the access network node in the open radio access network (open radio access network, O-RAN or open RAN), The next-generation base station in the sixth generation (6th generation, 6G) mobile communication system, the base station in the future mobile communication system, or the access node in the wireless fidelity (wireless fidelity, WiFi) system, etc.
  • base station base station
  • Node B node B
  • eNodeB or eNB evolved node B
  • TRP transmission reception point
  • the access network node may be a module or unit that completes some functions of the base station, for example, it may be a centralized unit (central unit, CU), a distributed unit (distributed unit, DU), a centralized unit control plane (CU control plane, CU-CP) module, or centralized unit user plane (CU user plane, CU-UP) module.
  • the access network node may be a macro base station (such as 110a in Figure 1), a micro base station or an indoor station (such as 110b in Figure 1), or a relay node or a donor node.
  • the 5G system can also be called a new radio (new radio, NR) system.
  • the device for implementing the function of the access network node may be an access network node; it may also be a device capable of supporting the access network node to realize the function, such as a chip system, a hardware circuit, a software module, or a hardware A circuit plus a software module, the device can be installed in the access network node or can be matched with the access network node.
  • the system-on-a-chip may be composed of chips, and may also include chips and other discrete devices.
  • the technical solution provided by the present disclosure is described below by taking the device for realizing the function of the access network node as the access network node, and optionally taking the access network node as the base station as an example.
  • a terminal may also be called terminal equipment, user equipment (user equipment, UE), mobile station, mobile terminal, and so on.
  • Terminals can be widely used in various scenarios for communication.
  • the scenario includes, but is not limited to, at least one of the following scenarios: enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), large-scale machine type communication ( massive machine-type communications (mMTC), device-to-device (D2D), vehicle to everything (V2X), machine-type communication (MTC), Internet of Things (Internet of Things) , IOT), virtual reality, augmented reality, industrial control, automatic driving, telemedicine, smart grid, smart furniture, smart office, smart wear, smart transportation, or smart city, etc.
  • eMBB enhanced mobile broadband
  • URLLC ultra-reliable low-latency communication
  • mMTC massive machine-type communications
  • D2D device-to-device
  • V2X vehicle to everything
  • MTC Internet of Things
  • IOT Internet of Things
  • virtual reality augmented reality
  • the terminal can be a mobile phone, a tablet computer, a computer with wireless transceiver function, a wearable device, a vehicle, a drone, a helicopter, an airplane, a ship, a robot, a mechanical arm, or a smart home device, etc.
  • the present disclosure does not limit the specific technology and specific equipment form adopted by the terminal.
  • the device for realizing the function of the terminal may be a terminal; it may also be a device capable of supporting the terminal to realize the function, such as a chip system, a hardware circuit, a software module, or a hardware circuit plus a software module. Installs in the terminal or can be used with the terminal.
  • the technical solution provided by the present disclosure is described below by taking the terminal as an example of the device for realizing the function of the terminal, and optionally taking the UE as an example.
  • Base stations and/or terminals may be fixed or mobile. Base stations and/or terminals can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; or can be deployed on water; or can be deployed on aircraft, balloons, and satellites in the air.
  • the present disclosure does not limit the environment/scene where the base station and the terminal are located.
  • the base station and the terminal can be deployed in the same or different environments/scenarios, for example, the base station and the terminal are deployed on land at the same time; or, the base station is deployed on land, and the terminal is deployed on water, etc., no more examples are given here.
  • base station and terminal may be relative.
  • the helicopter or UAV 120i in FIG. 1 can be configured as a mobile base station.
  • the terminal 120i is a base station; but for the base station 110a, 120i It may be a terminal, that is, communication between 110a and 120i may be performed through a wireless air interface protocol.
  • 110a and 120i communicate through an interface protocol between base stations.
  • relative to 110a, 120i is also a base station. Therefore, both base stations and terminals can be collectively referred to as communication devices (or communication devices), 110a and 110b in FIG. 1 can be referred to as communication devices with base station functions, and 120a-120j in FIG. 1 can be referred to as communication devices with terminal functions. device.
  • the protocol layer structure between the access network node and the terminal may include an artificial intelligence (AI) layer, which is used to transmit data related to the AI function.
  • AI artificial intelligence
  • an independent network element such as an AI entity, AI network element, AI node, or AI device, etc.
  • the AI network element can be directly connected to the base station, or can be indirectly connected to the base station through a third-party network element.
  • the third-party network element may be a core network element such as an access and mobility management function (access and mobility management function, AMF) network element or a user plane function (user plane function, UPF) network element.
  • AI entities or may be called AI modules or other names
  • the network element configured with the AI entity may be a base station, a core network device, or a network management (operation, administration and maintenance, OAM), etc.
  • OAM is used to operate, manage and/or maintain core network equipment (network management of core network equipment), and/or is used to operate, manage and/or maintain access network nodes (network management of access network nodes).
  • an AI entity may be integrated in a terminal or a terminal chip.
  • the AI entity may also be called by other names, such as AI module or AI unit, etc., which are mainly used to implement AI functions (or called AI-related operations), and this disclosure does not limit its specific name.
  • an AI model is a specific method for realizing an AI function.
  • the AI model represents the mapping relationship between the input and output of the model.
  • the AI model can be a neural network, linear regression model, decision tree model, clustering SVD model or other machine learning models. Among them, the AI model may be referred to as an intelligent model, a model or other names for short, without limitation.
  • AI-related operations may include at least one of the following: data collection, model training, model information release, model testing (or called model verification), model inference (or called model reasoning, reasoning, or prediction, etc.), or reasoning Result release etc.
  • FIG. 2A is an example diagram of the first application framework of AI in a communication system.
  • the data source is used to store training data and inference data.
  • the model training host (model training host) obtains the AI model by analyzing or training the training data provided by the data source, and deploys the AI model in the model inference host (model inference host).
  • the AI model represents the mapping relationship between the input and output of the model. Learning the AI model through the model training node is equivalent to using the training data to learn the mapping relationship between the input and output of the model.
  • the model inference node uses the AI model to perform inference based on the inference data provided by the data source, and obtains the inference result.
  • the model inference node inputs the inference data into the AI model, and obtains an output through the AI model, and the output is the inference result.
  • the inference result may indicate: configuration parameters used (executed) by the execution object, and/or operations performed by the execution object.
  • the reasoning result can be uniformly planned by the execution (actor) entity, and sent to one or more execution objects (for example, network elements) for execution.
  • the model inference node can feed back its inference results to the model training node. This process can be called model feedback.
  • the parameters fed back are used for the model training node to update the AI model, and the updated AI model is deployed on the model inference node. in the node.
  • the execution object can feed back the collected network parameters to the data source. This process can be called performance feedback, and the fed-back parameters can be used as training data or inference data.
  • the application framework shown in FIG. 2A can be deployed on the network element shown in FIG. 1 .
  • the application framework in FIG. 2A may be deployed on at least one of the terminal device, access network device, core network device, or independently deployed AI network element (not shown) in FIG. 1 .
  • the AI network element (which can be regarded as a model training node) can analyze or train the training data (training data) provided by the terminal device and/or the access network device to obtain a model.
  • At least one of the terminal device, the access network device, or the core network device (which can be regarded as a model reasoning node) can use the model and reasoning data to perform reasoning and obtain the output of the model.
  • the reasoning data may be provided by the terminal device and/or the access network device.
  • the input of the model includes inference data
  • the output of the model is the inference result corresponding to the model.
  • At least one of the terminal device, the access network device, or the core network device (which may be regarded as an execution object) may perform a corresponding operation according to the inference data and/or the inference result.
  • the model inference node and the execution object may be the same or different, without limitation.
  • the network architecture to which the communication solution provided by the present disclosure can be applied is introduced below with reference to FIGS. 2B to 2E .
  • the access network device includes a near real-time access network intelligent controller (RAN intelligent controller, RIC) module for model learning and reasoning.
  • RAN intelligent controller RIC
  • near real-time RIC can be used to train an AI model and use that AI model for inference.
  • the near real-time RIC can obtain network-side and/or terminal-side information from at least one of CU, DU, or RU, and the information can be used as training data or inference data.
  • the near real-time RIC may submit the reasoning result to at least one of CU, DU or RU.
  • the inference results can be exchanged between the CU and the DU.
  • the reasoning results can be exchanged between the DU and the RU, for example, the near real-time RIC submits the reasoning result to the DU, and the DU forwards it to the RU.
  • a non-real-time RIC is included outside the access network (optionally, the non-real-time RIC can be located in the OAM or in the core network device) for model learning and reasoning.
  • non-real-time RIC is used to train an AI model and use that model for inference.
  • the non-real-time RIC can obtain network-side and/or terminal-side information from at least one of CU, DU, or RU, which can be used as training data or inference data, and the inference results can be submitted to CU, DU, or RU At least one of the .
  • the inference results can be exchanged between the CU and the DU.
  • the reasoning results can be exchanged between the DU and the RU, for example, the non-real-time RIC submits the reasoning result to the DU, and the DU forwards it to the RU.
  • the access network equipment includes a near-real-time RIC, and the access network equipment includes a non-real-time RIC (optionally, the non-real-time RIC can be located in the OAM or the core network equipment middle).
  • non-real-time RIC can be used for model learning and reasoning; and/or, like the first possible implementation above, near real-time RIC can be used for model learning and reasoning; and/or, Near-real-time RIC can obtain AI model information from non-real-time RIC, and obtain network-side and/or terminal-side information from at least one of CU, DU or RU, and use this information and the AI model information to obtain inference results, optional Yes, the near-real-time RIC can submit the inference results to at least one of CU, DU, or RU.
  • the inference results can be exchanged between CU and DU.
  • the inference results can be exchanged between DU and RU.
  • the near real-time RIC submits the inference result to the DU, and the DU forwards it to the RU.
  • near real-time RIC is used to train model A and use model A for inference.
  • non-real-time RIC is used to train Model B and utilize Model B for inference.
  • the non-real-time RIC is used to train the model C, and the model C is submitted to the near-real-time RIC, and the near-real-time RIC uses the model C for inference.
  • FIG. 2C is an example diagram of a network architecture to which the method provided by the present disclosure can be applied. Compared with FIG. 2B , in FIG. 2B CU is separated into CU-CP and CU-UP.
  • FIG. 2D is an example diagram of a network architecture to which the method provided by the present disclosure can be applied.
  • the access network device includes one or more AI entities, and the functions of the AI entities are similar to the near real-time RIC described above.
  • the OAM includes one or more AI entities, and the functions of the AI entities are similar to the non-real-time RIC described above.
  • the core network device includes one or more AI entities, and the functions of the AI entities are similar to the above-mentioned non-real-time RIC.
  • both the OAM and the core network equipment include AI entities, the models trained by their respective AI entities are different, and/or the models used for reasoning are different.
  • the different models include at least one of the following differences: the structural parameters of the model (such as the number of layers of the model, the width of the model, the connection relationship between layers, the weight of neurons, the activation function of neurons, or the at least one of the bias), the input parameters of the model (such as the type of the input parameter and/or the dimension of the input parameter), or the output parameters of the model (such as the type of the output parameter and/or the dimension of the output parameter).
  • the structural parameters of the model such as the number of layers of the model, the width of the model, the connection relationship between layers, the weight of neurons, the activation function of neurons, or the at least one of the bias
  • the input parameters of the model such as the type of the input parameter and/or the dimension of the input parameter
  • the output parameters of the model such as the type of the output parameter and/or the dimension of the output parameter.
  • FIG. 2E is an example diagram of a network architecture to which the method provided by the present disclosure can be applied.
  • the access network devices in FIG. 2E are separated into CU and DU.
  • the CU may include an AI entity, and the function of the AI entity is similar to the above-mentioned near real-time RIC.
  • the DU may include an AI entity, and the function of the AI entity is similar to the above-mentioned near real-time RIC.
  • both the CU and the DU include AI entities, the models trained by their respective AI entities are different, and/or the models used for reasoning are different.
  • the CU in FIG. 2E may be further split into CU-CP and CU-UP.
  • one or more AI models may be deployed in the CU-CP.
  • one or more AI models can be deployed in CU-UP.
  • the OAM of the access network device and the OAM of the core network device may be deployed separately and independently.
  • the core network equipment may include one or more modules, for example including
  • a model can infer one parameter, or infer multiple parameters.
  • the learning process of different models can be deployed in different devices or nodes, and can also be deployed in the same device or node.
  • the inference process of different models can be deployed in different devices or nodes, or can be deployed in the same device or node.
  • Artificial intelligence can make machines have human intelligence, for example, it can make machines use computer software and hardware to simulate certain intelligent behaviors of humans.
  • machine learning methods or other methods may be adopted without limitation.
  • Neural network is a specific implementation of machine learning. According to the general approximation theorem, the neural network can theoretically approximate any continuous function, so that the neural network has the ability to learn any mapping. Therefore, neural networks can accurately and abstractly model complex high-dimensional problems. That is, intelligent models can be realized through neural networks.
  • the idea of a neural network is derived from the neuronal structure of brain tissue.
  • Each neuron performs a weighted sum operation on its input values, and passes the weighted sum result through an activation function to generate an output.
  • the weighted And the bias is b.
  • the form of the activation function can be varied.
  • the output of the neuron is:
  • the element x i of the input x of the neuron, the element w i of the weight value w, or the bias b can be various possible values such as decimals, integers (including 0, positive integers or negative integers, etc.), or complex numbers.
  • the activation functions of different neurons in a neural network can be the same or different.
  • a neural network generally includes a multi-layer structure, and each layer may include one or more neurons.
  • Increasing the depth and/or width of a neural network can improve the expressive power of the neural network, providing more powerful information extraction and abstract modeling capabilities for complex systems.
  • the depth of the neural network refers to the number of layers included in the neural network, and the number of neurons included in each layer can be called the width of the layer.
  • FIG. 2F it is a schematic diagram of the layer relationship of the neural network.
  • a neural network includes an input layer and an output layer.
  • the input layer of the neural network processes the input received by neurons, and then passes the result to the output layer, and the output layer obtains the output result of the neural network.
  • a neural network in another implementation, includes an input layer, a hidden layer, and an output layer.
  • the input layer of the neural network processes the input received by neurons, and then passes the result to the hidden layer in the middle, and the hidden layer then passes the calculation result to the output layer or the adjacent hidden layer, and finally the output layer obtains the result of the neural network. Output the result.
  • a neural network may include one or more sequentially connected hidden layers without limitation.
  • a loss function can be defined.
  • the loss function describes the gap or difference between the output value of the AI model and the ideal target value, and the present disclosure does not limit the specific form of the loss function.
  • the training process of the AI model is the process of adjusting some or all parameters of the AI model so that the value of the loss function is less than the threshold value or meets the target requirements. If the AI model is a neural network, one or more of the following parameters can be adjusted during training: the number of layers of the neural network, the width of the neural network, the connection relationship between layers, the weight value of neurons, the activation function of neurons, or A bias in the activation function such that the difference between the output of the neural network and the ideal target value is as small as possible.
  • Fig. 3 is an example diagram of the training process of the intelligent model.
  • the intelligent model is taken as an example of a neural network model, and the neural network f ⁇ ( ⁇ ) can include a 4-layer convolutional layer network and a 4-layer fully connected layer network as an example for introduction.
  • the AI entity can obtain training data from the basic data set (such as a collection of channel data).
  • the training data can include training samples and labels.
  • the sample x is used as input and processed by the neural network f ⁇ ( ⁇ ) to output the inference result f ⁇ (x) , the loss function calculates the error between the inference result and the label, and the AI entity can use the backpropagation optimization algorithm (which can be called the model optimization algorithm) to optimize the network parameter ⁇ based on the error obtained by the loss function.
  • the backpropagation optimization algorithm which can be called the model optimization algorithm
  • a large amount of training data is used to train the neural network, so that the difference between the output of the neural network and the label is less than a preset value, and then the training of the neural network is completed.
  • the training process shown in Figure 3 adopts a supervised learning training method, that is, based on samples and labels, the loss function is used to realize the training of the model.
  • the training process of the intelligent model can also use unsupervised learning, using the algorithm to learn the internal mode of the sample, and realize the training of the intelligent model based on the sample.
  • the training process of the intelligent model can also use reinforcement learning to obtain the excitation signal of environmental feedback through interaction with the environment, so as to learn the strategy to solve the problem and realize the optimization of the model.
  • the disclosure does not limit the training method of the model and the type of the model.
  • the trained AI model can perform reasoning tasks. After the actual data is input into the AI model for processing, the corresponding inference results are obtained. Optionally, an AI model can infer one parameter, or infer multiple parameters.
  • the application of artificial intelligence in wireless communication systems can significantly improve the performance of communication systems.
  • the network side and the terminal need to use matching artificial intelligence models to improve wireless communication performance.
  • the terminal can compress and encode the uplink information by using a compression encoder model and send it to the network side, and the network side can decode the uplink information sent by the terminal by using a matching decoder model.
  • the application of artificial intelligence in the field of wireless communication often involves complex nonlinear function fitting.
  • the scale of the intelligent model is often large, for example, the number of model layers is large, and the number of model parameters is large.
  • a convolutional neural network (CNN) for an encoder model implementing CSI feedback compression might include 15 neural network layers.
  • the intelligent model used by the terminal is usually specified by the network.
  • the network side may select a model from multiple predefined models, and notify the terminal of the identifier of the model, so that the terminal device can determine the adopted model based on the identifier.
  • the network side can also deliver the intelligent model to the terminal, but the parameters of the intelligent model are huge and need to occupy a large amount of air interface resources, which will affect the transmission efficiency of service data.
  • this disclosure proposes that the network can inform the terminal of the training strategy of the intelligent model, and the terminal executes the training of the intelligent model based on the training strategy provided by the network, so that the model obtained by the terminal training can match the model adopted by the network and achieve the expected performance need.
  • the communication method provided by the present disclosure will be described below with reference to the accompanying drawings. It should be noted that this article takes the interaction between a terminal and an access network node as an example, but this application is not limited thereto.
  • the communication method provided in this disclosure can be used between any two nodes, and one node communicates with another node Obtain a training policy and/or model structure of the smart model, and perform model training based on the training policy and/or model structure.
  • one or more modules of the access network node (such as RU, DU, CU, CU-CP, CU-UP or near real-time RIC) can realize the corresponding method of the access network node or operate.
  • Fig. 4 is a schematic flowchart of a communication method 400 provided in the present disclosure.
  • the access network node sends first information to the terminal, where the first information is used to indicate a training policy of the first intelligent model.
  • the terminal receives the first information from the access network node.
  • the terminal determines a training strategy for the first smart model according to the first information.
  • the training strategy includes one or more of the following:
  • Model training method loss function information, model initialization method, type of model optimization algorithm, or parameters of the optimization algorithm.
  • the first information may indicate a training manner for the terminal to train the first smart model, and the terminal may use the model training manner indicated by the first information to train the first smart model.
  • the terminal may use the model training manner indicated by the first information to train the first smart model.
  • the trained first intelligent model match the intelligent model on the network side. Achieve the effect of improving wireless communication performance.
  • the first information may instruct the terminal to train the first intelligent model in one of supervised learning, unsupervised learning, or reinforcement learning. But the present disclosure is not limited thereto.
  • the first information may include loss function information. After receiving the first information, the terminal may obtain a loss function for training the first intelligent model based on the loss function information.
  • the access network node may notify the terminal of the loss function adopted by the loss function information.
  • the loss function information may directly indicate the loss function, or the loss function information may indicate identification information of a loss function among a plurality of predefined loss functions.
  • the terminal may train the first intelligent model by means of supervised learning, for example, it may be a pre-defined training method using the supervised learning, or the access network node indicates that the model training method is supervised learning through the first information. Way.
  • the training goal of the supervised learning model training method is to minimize the value of the loss function.
  • the model training can be completed when the value of the loss function obtained through training is less than or equal to the first threshold.
  • the first threshold may be predefined or indicated by the loss function information.
  • the loss function can be the following cross-entropy loss function:
  • log(A) means to calculate the logarithm of A
  • ⁇ n B n means to sum B n based on the value range of n.
  • the loss function can be a squared loss function as follows:
  • x n is the sample
  • y n is the label of the nth sample
  • f ⁇ (x n ) is the output result of the model for the input data x n .
  • the terminal may train the first smart model in an unsupervised learning manner, and for unsupervised learning, the loss function may be a function used to evaluate model performance.
  • the training goal of the unsupervised learning model training method is to maximize the value of the loss function.
  • the model training can be completed when the value of the loss function obtained through training is greater than or equal to the second threshold.
  • the second threshold may be predefined or indicated by the loss function information.
  • the first intelligent model is used to infer the transmit power based on the channel response, then the input of the first intelligent model is the channel response h n , and the reasoning result of the first intelligent model is the transmit power P n , then the loss function can be calculated data throughput
  • the function, such as the loss function can be as follows:
  • log(A) means to calculate the logarithm of A
  • N 0 is the noise power
  • the loss function may be the above mathematical expression, and optionally, the loss function may also be a machine learning intelligent model.
  • the complexity and parameter quantity of the loss function model are much smaller than the complexity and parameter quantity of the first intelligent model, and the loss function information may include structural information and/or parameter information of the loss function, so that the terminal can obtain the first information after receiving the first information
  • a loss function model is used to train the first intelligent model based on the loss function, which can improve the effect of terminal model training.
  • the loss function may be a nonlinear performance measurement model such as system throughput, signal bit error rate, and signal-to-interference-noise ratio.
  • the first information may indicate a model initialization manner, and the terminal may initialize the first smart model based on the model initialization manner indicated by the first information.
  • the first intelligent model may be a neural network model
  • the first information may indicate an initialization manner of neuron weights in the first intelligent model.
  • the access network node may instruct the terminal that the weight of each neuron is randomly selected, or that the weight values of the neurons are all 0, or that the weight values of the neurons are generated by a preset function, and the like.
  • the access network node may use the first information to instruct the terminal to randomly select a value within the interval [z min , z max ] as the weight of the neuron, for example, the first information may indicate z min and z max .
  • the terminal selects a value for each neuron within the interval [z min , z max ] as the initial weight value of the neuron.
  • the first information indicates the identification information of one of the predefined initialization methods
  • the first information includes a 2-bit indication field, which is used to indicate the initialization method of the neuron weight, and the 2 bits indicate "00" Indicates that the initial value of the weight of the neuron is randomly selected.
  • the terminal can select a weight value for each neuron within a preset value range, or the first information indicates this method.
  • the terminal is the weight value of the neuron within the value interval indicated by the first information;
  • the 2-bit indication "01” means that the initial value of the neuron is set to 0;
  • the 2-bit indication "10” Indicates that the preset function is used to generate a sequence of initial values, and each value in the sequence is used as the initial value of a neuron.
  • the training device can use the model optimization algorithm based on the output value of the loss function to obtain the optimized parameters of the intelligent model, and after updating the intelligent model based on the optimized parameters, use the intelligent model after the updated parameters to perform the following steps: A model training.
  • the access network node may notify the terminal of the optimization algorithm of the first intelligent model through the first information.
  • the first information may indicate one type of optimization algorithm among predefined types of model optimization algorithms, and the terminal may use the model optimization algorithm indicated by the first information to perform model optimization when training the first intelligent model.
  • the model optimization algorithm may include but not limited to one or more of the following algorithms:
  • ADAM Adaptive momentum estimation
  • SGD stochastic gradient descent
  • BGD batch gradient descent
  • the communication protocol may predefine multiple types of model optimization algorithms, and the parameters of each type of model optimization algorithm may also be predefined, and the terminal may, based on the identification information of the type of model optimization algorithm indicated by the first information, The type of the optimization algorithm and the predefined parameters of the optimization algorithm are determined.
  • the terminal uses the model optimization algorithm to update the model weights.
  • the parameters of the optimization algorithm may also be indicated by the first information, and reference may be made to the description below.
  • the first information may indicate parameters of the model optimization algorithm, where the type of the model optimization algorithm may be indicated by the access network node through the first information, or the type of the model optimization algorithm may be preconfigured.
  • the parameters of the model optimization algorithm may include but not limited to one or more of the following parameters:
  • the learning rate is a parameter that characterizes the step size of gradient descent in model training.
  • the learning rate may be a fixed value, and the access network node may notify the terminal of the value of the learning rate through the first information.
  • the learning rate can be set to 0.001.
  • the learning rate can also be set to a value that gradually decreases with the iterative steps, and the first information can indicate the learning rate and the value that the learning rate decreases for each iteration.
  • the first information can indicate that the initial value is 0.1, and each iteration decreases Small 0.01.
  • the number of iterations refers to the number of times the training data set is traversed.
  • the amount of batch processed data refers to the amount of training data selected in the training data set for each model training and used for gradient descent update.
  • the training strategy of the first intelligent model indicated by the first information mentioned above may include but not limited to one or more strategies in (1) to (5) above. So that the terminal device can train the first intelligent model based on the training strategy, and obtain an intelligent model matching the network.
  • the terminal can use but not limited to the following methods to determine the training Strategy.
  • the terminal may be pre-configured with one or more of the above training strategies. If the access network node does not indicate one or more of the strategies, the terminal performs model training based on the pre-configured training strategy. train.
  • the pre-configured training strategy may be defined by the protocol, or the pre-configured training strategy may be configured by the manufacturer based on terminal production and implementation.
  • the terminal may determine one or more training strategies not indicated by the access network node based on the one or more training strategies indicated by the access network node. That is to say, multiple training strategies may have an association relationship, and the access network node implicitly indicates other training strategies related to the training strategy by indicating a training strategy.
  • the association relationship may be stipulated in the agreement, or the access network node notifies the terminal in advance, and is not limited.
  • the access network node may use the first information to indicate the model training mode without indicating the loss function information, and the terminal determines the loss function based on the model training mode indicated by the access network node. If the first information indicates that the model training method is a supervised learning method, the terminal may determine that the loss function is a cross-entropy loss function, or the terminal may determine that the loss function is a square loss function. But the present application is not limited thereto.
  • the access network node may also send second information to the terminal, where the second information is used to indicate the structure of the first intelligent model.
  • the access network node In addition to notifying the terminal of the training policy of the first intelligent model, the access network node also notifies the terminal of the structure of the first intelligent model through the second information. After receiving the second information, the terminal generates an initial first intelligent model with this structure (or called the first intelligent model before training) based on the second information, and adopts the training strategy indicated by the first information to the first intelligent model. The intelligent model is trained, so that the trained first intelligent model can match the model adopted by the access network node.
  • the access network node may indicate the identification information of a structure among the structures of multiple predefined intelligent models, and notify the terminal to adopt the intelligent model of the structure.
  • the access network node may indicate the structural parameters of the first intelligent model.
  • the second information may specifically be used to indicate one or more of the following structural information of the first smart model:
  • Network layer structure information information, dimensions of input data or dimensions of output data.
  • the second information may indicate the dimension of the input data of the first intelligent model, that is, the dimension of a training sample.
  • the first intelligent model is a neural network model, including an input layer and an output layer
  • the dimension of the input data is the dimension of the input data of the input layer of the first intelligent model.
  • the second information may indicate the dimension of the output data of the first intelligent model, that is, the dimension of the inference data output by the first intelligent model. If the first intelligent model is a neural network model, the dimension of the output data is the dimension of the output data of the output layer of the first intelligent model.
  • the second information may include the network layer structure information of the first intelligent model, and the network layer structure information may include but not limited to one or more of the following:
  • the number of neural network layers contained in the first intelligent model the type of neural network layers, the use of neural network layers, the cascading relationship between neural network layers, the dimension of input data of neural network layers or the output of neural network layers Dimensions of the data.
  • the access network node may indicate the number of neural network layers included in the first intelligent model, and may indicate the type of each neural network layer, such as CNN, FNN, dimension transformation neural network layer, and the like.
  • CNN may be implemented by a filter.
  • the access network node may also notify the terminal of the type of filter (Filter) implementing the CNN through the network layer structure information.
  • the type of Filter can be the dimension of the filter.
  • the network layer structure information may indicate that the dimension of the filter is 3 ⁇ 3, 5 ⁇ 5 or 3 ⁇ 3 ⁇ 3, etc.
  • the dimension transformation neural network layer is used for the transformation of the data dimension.
  • the dimension transformation neural network layer can be located between two neural network layers, and the dimension transformation neural network layer is used to transform the input data (ie, the output data of the previous neural network)
  • the dimension of the input data of a neural network layer is obtained after the dimension transformation, and the data after the dimension transformation is output to the next neural network layer.
  • the dimension of the output data of the neural network layer A is 3 ⁇ 3 ⁇ 10
  • the dimension of the input data of the neural network layer B is required to be 30 ⁇ 3.
  • a dimension transformation neural network can be set between the two neural networks to convert the 3 ⁇ 3 ⁇ 10 data output by the neural network layer A into 30 ⁇ 3 data, and output the data to the neural network layer B.
  • the data can be processed by the neural network layer B after dimension transformation.
  • the access network node may also notify the terminal of the use mode of the neural network layer through the second information, for example, may notify the terminal of whether one or more neural network layers in the neural network use a temporary dropout mode.
  • the dropout method means that during model training, one or more neurons in the neural network layer can be randomly selected not to participate in this training, and the one or more neurons can be restored to participate in model training in the next training.
  • the input-output relationship of neurons in the neural network layer may be as shown in FIG. 5A .
  • the dropout method is adopted, the input-output relationship of the neurons in the neural network layer may be as shown in FIG. 5B .
  • the terminal can randomly select multiple neurons in the CNN neural network layer not to participate in model training, the output of the neurons is set to 0, and the neurons not participating in model training will not be finer in this training. , continue to use this weight in the next participating model training.
  • the access network node may also notify the terminal of the dimension of the input data and/or the dimension of the output data of the neural network layer in the first intelligent model. And/or, the access network node may also notify the terminal of the cascading relationship between the neural network layers, for example, the cascading relationship between the neural network layers may be an arrangement order of the neural network layers in the first smart model.
  • Table 1 shows an example of a neural network structure, and the access network node may notify the terminal of the neural network structure shown in Table 1, but the disclosure is not limited thereto.
  • the second information indicates that the number of neural network layers of the first intelligent model is 6, and notify the type of each neural network layer, such as the type of neural network layer includes input layer, CNN, shape transformation, FNN and output layer
  • the second information may also indicate the cascading relationship between the neural network layers.
  • the second information may indicate that the arrangement sequence of the six neural network layers may be input layer, CNN, CNN, shape transformation, FNN, and output layer.
  • the second information may also indicate the input dimension and output dimension of each neural network layer as shown in Table 2, wherein NT is the number of antennas, and N sc is the number of subcarriers.
  • the second information may also notify the terminal to implement a 3 ⁇ 3 filter in the second and third CNN layers, and the 2nd, 3rd, 4th, and 5th neural network layers may use dropout.
  • the present disclosure is not limited thereto, and the structure of the first intelligent model may be partly predefined and indicated by the access network node through the second information.
  • the access network node may notify the terminal of some or all of the structural parameters shown in Table 1. Wherein, N/A shown in Table 1 indicates that this row is not applicable (or not configured) for this parameter.
  • the terminal may generate an initial first smart model based on the structure of the first smart model shown in Table 1 indicated by the second information.
  • the terminal can first determine that the first intelligent model includes 6 neural network layers based on the second information, and indicate the type, input dimension, and output dimension of each of the 6 neural network layers based on the second information, and generate a satisfying type and Dimensionally required 6 neural network layers, namely input layer, CNN layer, CNN layer, shape transformation layer, FNN layer and output layer.
  • the terminal implements two CNN layers therein through a 3 ⁇ 3 filter based on the indication of the second information.
  • the terminal determines the arrangement order of the 6 neural network layers based on the cascading relationship indicated by the second information, and obtains the initial first intelligent model.
  • the 2, 3, 4, and 5 neural network layers are used in the way of dropout, and the neurons in the network layer are randomly selected in each training to not participate in this training.
  • the access network node sends third information to the terminal, where the third information is used to indicate information of a training data set, and the training data set is used to train the first intelligent model.
  • the third information may indicate the identification information of the preset training data set, so that the terminal uses the training data set corresponding to the identification information to train the first smart model after receiving the third information.
  • the protocol can predefine multiple training data sets, and each training data set corresponds to a piece of identification information, and the access network node can indicate the identification information of one training data set among the multiple training data sets through the third information, and the terminal according to The identification information determines that the training data set corresponding to the identification information is used to train the first intelligent model.
  • the information of the training data set may indicate the type of the training sample, or the type of the training sample and the type of the label.
  • the training method of the first intelligent model is an unsupervised learning method
  • the third information may indicate the type of training samples, for example, the first intelligent model may infer the transmission power based on the input channel data. Then the third information may indicate that the type of the training sample is channel data, and the terminal may train the first intelligent model according to the channel data.
  • the channel data may be obtained by the terminal through measurement based on a reference signal sent by the access network node, or may be pre-configured on the terminal. But the present application is not limited thereto.
  • the training method of the first intelligent model is an unsupervised learning method
  • the third information may indicate the type of the training sample and the type of the label
  • the terminal may determine the training data set based on the type of the training sample and the type of the label.
  • the data contained in the data set may be data collected during communication between the terminal and the access network node, or may be data of a corresponding type preconfigured on the terminal.
  • the first intelligent model implements the channel coding function of the data, that is, infers and outputs coded data based on the input data before coding
  • the third information can indicate that the type of training samples is data before channel coding, and the type of label is after channel coding The data.
  • the information of the training data set may also indicate the use of the first intelligent model, for example, the information of the training data set indicates that the first intelligent model is used for compression coding, decoding or transmission power.
  • the terminal determines the corresponding training data set according to the purpose of the first intelligent model.
  • the third information may indicate that the purpose of the first intelligent model is compression coding, and the training method of the first intelligent model is supervised learning, then the terminal may determine the method for training the first intelligent model based on the purpose of the first intelligent model.
  • the type of the training sample is information data before compression encoding, and the label is compressed data after compression encoding, so as to determine the training data set of the first intelligent model including the training sample and the label.
  • the data included in the training data set may be data collected during communication between the terminal and the access network node, or may be data of a corresponding type preconfigured on the terminal. But the present application is not limited thereto.
  • the third information may include a training data set
  • the training data set may include training samples, or training samples and labels, that is to say, the terminal may obtain the training data set from the network.
  • the training method of the first intelligent model is a supervised learning method
  • the training data set in the third information includes training samples and labels.
  • the training method of the first intelligent model is an unsupervised learning method, and the training data set in the third information includes training samples but does not include labels.
  • the terminal can train the first intelligent model while downloading the training data in the training data set with a small resource bandwidth, which can reduce resource occupation and avoid affecting the transmission rate of service data.
  • the access network node may determine the training strategy and model of the first intelligent model based on the capability information of the terminal, and/or, based on the environmental data (such as channel data, etc.) between the access network node and the terminal. One or more items in the structure or training data set.
  • the access network node may acquire the training policy and/or model structure of the first smart model from a third-party node in the network, and forward it to the terminal. That is to say, the above information (including but not limited to the first information, second information, and third information) sent by the access network node to the terminal may be generated by the access network node, or may be obtained by the access network node from a third party. Obtained by the node and forwarded to the terminal.
  • the third-party node may be a node in the network that has an AI function or is configured with an AI entity.
  • the third-party node may be a non-real-time RIC, and the non-real-time RIC may be included in the OAM.
  • the third-party node may actually be an OAM.
  • the access network node receives capability information from the terminal, where the capability information is used to indicate the capability of the terminal to run the smart model.
  • the capability information is specifically used to indicate one or more of the following capabilities of the terminal:
  • the capability information may indicate whether the terminal supports running an intelligent model.
  • the access network node After the access network node receives the capability information and determines that the terminal supports the intelligent model based on the capability information, the access network node sends the first information to the terminal, In order to notify the terminal of the strategy for training the first intelligent model.
  • the access network node needs to notify the terminal of the structure of the first intelligent model through the second information, and the capability information may indicate the operator library information of the machine learning model supported by the terminal.
  • the operator library information may indicate a set of basic intelligent model computing units that the terminal can support, or the operator library information may indicate basic intelligent model computing units not supported by the terminal.
  • the access network node obtains the model operator database information supported by the terminal by using the capability information, so as to determine the structure of the first intelligent model.
  • the intelligent model operation unit may include but not limited to one or more of the following:
  • 1D CNN unit (conv1d), 2D CNN unit (conv2d), 3D CNN unit (conv3d), pooling layer, pooling operation, or sigmoid activation function.
  • the capability information indicates the data processing capability of the terminal.
  • the capability information may indicate the type of processor, the operation speed of the processor, the amount of data that the processor can process, and the like.
  • the type of the processor may include a type of a graphics processing unit (graphics processing unit, GPU) and/or a type of a central processing unit (central processing unit, CPU).
  • the access network node may determine one or more of the training strategy, model structure, or training data set of the first intelligent model based on the data processing capability of the terminal.
  • the access network node may train to obtain terminal capabilities and/or channel conditions based on the capability information of the terminal, and/or, based on environmental data (such as channel data, etc.) between the access network node and the terminal. smart model.
  • the access network node notifies the terminal of the training strategy for training the smart model, so that the terminal adopts the training strategy to train the first smart model, so that the first smart model trained by the terminal is as close as possible to the smart model trained by the access network node.
  • the first intelligent model applied by the terminal and the access network node in the communication process can match the intelligent model used by the access network node, so as to achieve the effect of improving communication performance.
  • the first intelligent model is a compression coding model
  • the access network node can train a matching compression decoding model used by the access network node and a compression coding model used by the terminal based on the capability information and channel data of the terminal.
  • the access network node notifies the terminal of the training strategy used by the access network node to train the compression coding model.
  • the access network node notifies the terminal of the initialization method used by the access network node to randomly select values within the interval [z min , z max ]
  • the model training method is supervised learning
  • the loss function is cross-entropy loss function
  • the optimization algorithm is stochastic gradient descent algorithm.
  • the terminal After obtaining the training strategy of the access network node, the terminal trains the compression coding model based on the training strategy, so that the compression coding model trained by the terminal is as identical as possible to the compression coding model obtained by the training of the access network node, thereby realizing the
  • the compression coding model used matches the compression decoding model used by the access network nodes in communication.
  • the access network node may also notify the terminal of the structure of the intelligent model as the structure of the first intelligent model, and/or, the access network node may also instruct the terminal to train the intelligent model. Information about the dataset.
  • a third-party node in the network may obtain terminal capability information from an access network node, and/or obtain environmental data between the access network node and the terminal, and obtain Match the intelligent model of the access network node application and the intelligent model of the terminal application, and send the intelligent model of the access network node application to the access network node by the third-party node, and notify the access network node terminal to train the intelligent model
  • the training policy of is forwarded to the terminal by the access network node through the first information.
  • the third-party node may also notify the terminal of the structure of the intelligent model and/or the training data set information through the access network node.
  • the terminal performs model training on the first smart model according to the training strategy.
  • the terminal can determine the training strategy of the first intelligent model, such as model training method, loss function, model initialization method, type of model optimization algorithm, parameters of the optimization algorithm, etc., part or all of the training strategy can be based on the terminal
  • the training strategy not indicated by the access network node determined by the first information of the node may be obtained by the terminal based on pre-configured information or based on an associated training strategy indicated by the access network node.
  • the terminal may acquire the structure of the first intelligent model, and determine the first intelligent model based on the structure.
  • the terminal may receive second information from the access network node, the second information is used to indicate the structure of the first intelligent model, and the terminal determines the initial first intelligent model (or called first intelligent model before training).
  • the terminal obtains the structure of the first smart model from the preconfigured information, and determines the initial first smart model.
  • the terminal may receive third information from the access network node, where the third information is used to indicate the information of the training data set,
  • the terminal acquires the training data set according to the third information, and performs model training on the first intelligent model by using the training strategy indicated by the first information based on the training data set.
  • the present disclosure is not limited thereto, and the terminal may also perform model training on the first smart model based on the data set stored in the terminal.
  • the terminal After the terminal acquires the training policy of the first intelligent model, the training data set, and the initial first intelligent model based on the structure of the first intelligent model, the terminal may start to perform model training of the first intelligent model.
  • the terminal determines that the model training mode is a supervised learning mode, and the training data set includes training samples and labels.
  • the terminal inputs the training samples into the first intelligent model in each training, and the first intelligent model outputs the inference result after processing the training samples, and the terminal uses the loss function to obtain the loss value output by the loss function based on the inference result and the label.
  • the model optimization algorithm is used to optimize the parameters of the first intelligent model, and the first intelligent model after updating parameters is obtained, and the first intelligent model after updating parameters is applied to the next model training, after multiple iterations of training , when the loss value output by the loss function is less than or equal to the first threshold, the terminal determines that the training of the first intelligent model is completed, and obtains the trained first intelligent model.
  • the terminal determines that the model training mode is an unsupervised learning mode, and the training data set includes training samples.
  • the terminal inputs the training samples into the first intelligent model during each training, and the first intelligent model outputs an inference result after processing the training samples.
  • the terminal uses the loss function to obtain the output value of the loss function.
  • the model optimization algorithm is used to optimize the parameters of the first intelligent model, and the first intelligent model with updated parameters is obtained, and the first intelligent model with updated parameters is applied to the next model training. Iterative training, when the output value of the loss function is greater than or equal to the second threshold, the terminal determines to complete the training of the first intelligent model, and obtains the trained first intelligent model.
  • the network can inform the terminal of the training strategy of the intelligent model, and the terminal executes the training of the intelligent model based on the training strategy provided by the network, so that the model obtained by the terminal training can match the model adopted by the network and meet the expected performance requirements.
  • the air interface resource overhead caused by the access network node sending the model parameters of the first intelligent model (such as the weight value, activation function and bias of each neuron, etc.) to the terminal is reduced.
  • FIG. 5C is a schematic flowchart of a communication method 500 provided by the present disclosure. It should be noted that, for the same parts in the method shown in FIG. 5 as in the method shown in FIG. 4 , reference may be made to the description in the method shown in FIG. 4 , and details are not repeated here for brevity.
  • the access network node sends first information to the terminal, where the first information is used to indicate a training policy of the first intelligent model.
  • the terminal trains the first intelligent model according to the training policy.
  • the terminal may obtain the structure of the first intelligent model to determine the first intelligent model, and use the training data set to train the first intelligent model by using the training policy.
  • reference may be made to S402, which will not be repeated here.
  • the access network node sends fourth information to the terminal, where the fourth information is used to indicate test information, and the test information is used to test performance of the first smart model.
  • the terminal receives the fourth information from the access network node, and after receiving the fourth information, the terminal tests the trained first intelligent model according to the fourth information.
  • test information includes one or more of the following:
  • Test data information For evaluation methods or performance evaluation parameters.
  • the measurement information includes one or two of test data information, performance evaluation methods, and performance evaluation parameters, such as including test data information and performance evaluation methods, other items not included, such as performance evaluation parameters, may be It is stipulated in the agreement or determined by other means, and shall not be restricted. Or in the following example, the terminal does not need to know the performance evaluation mode and/or performance evaluation parameters, and the access network node performs the performance evaluation.
  • the test data information is used to indicate one or more of the test data, the type of the test data, the label data or the type of the label data.
  • the performance evaluation method can be a method of calculating the loss value between the inference data and the label data, and the performance evaluation parameter can be parameters such as the threshold value of the evaluation loss value, but the disclosure is not limited thereto, and the performance evaluation method can also be an evaluation function, etc. .
  • the terminal sends fifth information and/or sixth information to the access network node, where the fifth information is used to indicate a test result of the first intelligent model, and where the sixth information is used to indicate reasoning data.
  • the reasoning data is obtained by reasoning the test data of the first intelligent model, and the test information indicated by the fourth information includes the test data.
  • the terminal may send fifth information to the access network node, where the fifth information may indicate a test result of the first smart model.
  • the test result may be that the performance of the first intelligent model meets or fails.
  • the test information sent by the access network node to the terminal may include test data, verification data, and a verification threshold, and the terminal may use the test data as input to the first intelligent model after training, and obtain the first intelligent model based on the test data.
  • Inference gets inference data.
  • the terminal calculates a loss value between the inference data and the verification data.
  • the terminal may use the loss function used when training the first intelligent model to calculate the loss value between the inference data and the verification data. But the present application is not limited thereto.
  • the terminal compares the loss value with the verification threshold value.
  • the terminal can determine that the trained first intelligent model meets the standard and can be applied to actual communication, and pass the fifth information Notify the access network node. Alternatively, if the loss value is greater than the verification threshold, the terminal may notify the access network node that the trained first intelligent model does not meet the standard through the fifth information.
  • the test result may be the loss value between the inference data and the verification data.
  • the access network node indicates the test data and the verification data through the fourth information, and the fourth information also indicates that the performance evaluation method is calculation reasoning The loss value between data and check data.
  • the terminal uses the trained first intelligent model to infer to obtain the inference data, calculates the loss value between the inference data and the label data, and notifies the access network node of the loss value through the fifth information, and the network Based on the loss value, it is determined whether the performance of the first intelligent model trained by the terminal meets the standard.
  • it may be an access network node or a third-party node in the network to judge whether the standard is met, and the third-party node can obtain the loss value through the forwarding of the access network node.
  • the terminal may send sixth information to the access network node, where the sixth information is used to indicate the inference data of the first intelligent model.
  • the fourth information indicates test data
  • the terminal uses the trained first intelligent model to infer the test data to obtain inference data output by the first intelligent model, and the terminal sends the obtained inference data to the access network node.
  • a network node (such as an access network node or a third-party node in the network) judges whether the performance of the trained first intelligent model meets the standard.
  • the terminal may send the fifth information and the sixth information to the access network node.
  • the terminal not only sends the test result to the access network node, but also sends inference data to the access network node. smart model. But the present disclosure is not limited thereto.
  • the access network node sends seventh information to the terminal, where the seventh information is used to indicate the updated training policy of the first intelligent model, and/or is used to indicate the updated training strategy of the first intelligent model. structure.
  • the terminal receives the seventh information from the access network node.
  • the terminal trains the first smart model again based on the seventh information.
  • the access network node after receiving the fifth information and/or sixth information from the terminal, the access network node determines that the performance of the first intelligent model trained by the terminal is not up to standard, and the access network node may send the second intelligent model to the terminal. Seventh, the information notifies the first intelligent model of the updated training strategy and/or the updated structure of the first intelligent model.
  • the access network node determines that the first intelligent model needs to be updated based on the communication performance, and the access network node may send the first intelligent model to the terminal. Seven information.
  • the access network node can send the seventh information to the terminal, so that the terminal can perform model training on the first intelligent model again , making the retrained first intelligent model adapt to the changed environment.
  • the manner in which the seventh information indicates the training strategy may be the same as the manner in which the first information indicates the training strategy, and the manner in which the seventh information indicates the model structure may be the same as the manner in which the second information indicates the model structure.
  • the seventh information may be specifically used to indicate at least one change amount of the training strategy, and/or at least one change amount of the structure of the first intelligent model.
  • the seventh information indicates at least one variation of the training strategy, and after receiving the seventh information, the terminal determines an updated training strategy based on the variation indicated by the seventh information. For example, if the access network node adjusts the parameters of the optimization algorithm (for example, the learning rate is changed from 0.1 indicated by the first information to 0.01), only the parameters of the optimization algorithm in the training strategy change, and the other training strategies do not change, then the seventh information can indicate The seventh information indicates that the updated learning rate is 0.01.
  • the first information indicates that the amount of batch-processed data is N, that is, each time N training data is taken for gradient descent update, and the access network node adjusts the batch-processed data amount to M, then the access network can be grounded through
  • the present disclosure is not limited thereto.
  • the seventh information may also indicate at least one variation of the model structure, and the specific manner is similar to that of indicating at least one variation of the training strategy, which may be implemented with reference to the above description, and will not be repeated here.
  • the access network node can send test information to the terminal, and the terminal tests the trained first intelligent model based on the test information, and the terminal or the access network
  • the node can judge whether the performance of the first intelligent model after training is up to standard based on the test results and/or inference data, and can apply the first intelligent model in actual communication when the performance is up to standard, so as to improve wireless communication performance (such as communication Reliability and other performance) purposes.
  • Fig. 6 is another schematic flowchart of the communication method provided by the present disclosure. It should be noted that the same parts in the method shown in FIG. 6 and other methods may be mutually referenced or implemented in combination, and for the sake of brevity, details are not described here.
  • the access network node sends second information to the terminal, the second information is used to indicate the structure of the first intelligent model, and the second information includes the network layer structure information of the first intelligent model, the dimension of input data or the dimension of output data One or more structural information in a dimension.
  • the terminal receives the second information from the access network node.
  • the network layer structure information of the first intelligent model may include but not limited to one or more of the following:
  • the number of neural network layers contained in the first intelligent model the type of neural network layers, the use of neural network layers, the cascading relationship between neural network layers, the dimension of input data of neural network layers or the output of neural network layers Dimensions of the data.
  • the terminal determines a first smart model according to the second information.
  • the terminal may generate the first smart model with the structure based on the structure of the first smart model indicated by the second information.
  • the terminal performs model training on the first smart model.
  • the terminal After determining the first smart model in S602, the terminal trains the first smart model.
  • the access network node notifies the terminal of the structure of the first intelligent model, and the terminal generates the first intelligent model with the structure based on the instruction of the access network node, so that the structure of the first intelligent model used by the terminal can meet the requirements of the access network. Node requirements, so that the first intelligent model can be used to improve the performance of wireless communication.
  • the terminal executes the training of the first intelligent model, it can be applied in communication.
  • the air interface resource overhead caused by the model parameters (such as the weight value, activation function and bias of each neuron) notified by the access network node to the first intelligent model can be reduced.
  • the terminal may train the first intelligent model by using a predefined training policy agreed upon by the terminal and the access network node, or the access network node may notify the terminal of the training policy of the first intelligent model through the first information.
  • the training data set used by the terminal to train the first intelligent model may be pre-stored by the terminal or received from an access network node.
  • each network element may include a hardware structure and/or a software module, and realize the above functions in the form of a hardware structure, a software module, or a hardware structure plus a software module. Whether one of the above-mentioned functions is executed in the form of a hardware structure, a software module, or a hardware structure plus a software module depends on the specific application and design constraints of the technical solution.
  • Fig. 7 is a schematic block diagram of a communication device provided by the present disclosure.
  • the communication device 700 may include a transceiver unit 720 .
  • the communication device 700 may correspond to the terminal device in the above method, or a chip configured in (or used in) the terminal device, or other devices, modules, circuit or unit etc.
  • the communications apparatus 700 may include a unit for executing the method performed by the terminal device in the methods shown in FIG. 4 , FIG. 5C , and FIG. 6 .
  • each unit and the above-mentioned other operations and/or functions in the communication device 700 are for realizing the corresponding processes of the methods shown in FIG. 4 , FIG. 5C , and FIG. 6 .
  • the communication device 700 may further include a processing unit 710, and the processing unit 710 may be configured to process instructions or data to implement corresponding operations.
  • the transceiver unit 720 in the communication device 700 may be an input/output interface or circuit of the chip, and the processing in the communication device 700 Unit 710 may be a processor in a chip.
  • the communication device 700 may further include a storage unit 730, which may be used to store instructions or data, and the processing unit 710 may execute the instructions or data stored in the storage unit, so that the communication device realizes corresponding operations .
  • a storage unit 730 which may be used to store instructions or data
  • the processing unit 710 may execute the instructions or data stored in the storage unit, so that the communication device realizes corresponding operations .
  • the transceiver unit 720 in the communication device 700 can be implemented through a communication interface (such as a transceiver or an input/output interface), for example, it can correspond to the transceiver 810 in the terminal device 800 shown in FIG. 8 .
  • the processing unit 710 in the communication apparatus 700 may be implemented by at least one processor, for example, may correspond to the processor 820 in the terminal device 800 shown in FIG. 8 .
  • the processing unit 710 in the communication device 700 may also be implemented by at least one logic circuit.
  • the storage unit 730 in the communication device 700 may correspond to the memory in the terminal device 800 shown in FIG. 8 .
  • the communication device 700 may correspond to the access network node in the above method, for example, a chip configured (or used) in the access network node, or other devices capable of implementing access A device, module, circuit or unit, etc. of a method of a network node.
  • the communication device 700 may include a unit for performing the method performed by the access network node in the methods shown in FIG. 4 , FIG. 5C , and FIG. 6 .
  • each unit and the above-mentioned other operations and/or functions in the communication device 700 are for realizing the corresponding processes of the methods shown in FIG. 4 , FIG. 5C , and FIG. 6 .
  • the communication device 700 may further include a processing unit 710, and the processing unit 710 may be configured to process instructions or data to implement corresponding operations.
  • the transceiver unit 720 in the communication device 700 may be an input/output interface or circuit of the chip, and the communication device 700
  • the processing unit 710 may be a processor in a chip.
  • the communication device 700 may further include a storage unit 730, which may be used to store instructions or data, and the processing unit 710 may execute the instructions or data stored in the storage unit, so that the communication device realizes corresponding operations .
  • a storage unit 730 which may be used to store instructions or data
  • the processing unit 710 may execute the instructions or data stored in the storage unit, so that the communication device realizes corresponding operations .
  • the transceiver unit 720 in the communication device 700 can be implemented through a communication interface (such as a transceiver or an input/output interface), for example, it can correspond to the Transceiver 910 in network device 900 .
  • the processing unit 910 in the communication device 900 can be implemented by at least one processor, for example, it can correspond to the processor 920 in the network device 900 shown in FIG. circuit implementation.
  • the storage unit 730 in the communication device 700 may correspond to the memory in the network device 900 shown in FIG. 9 .
  • FIG. 8 is a schematic structural diagram of a terminal device 800 provided in the present disclosure.
  • the terminal device 800 can be applied to the system shown in FIG. 1 to perform the functions of the terminal device in the above method.
  • the terminal device 800 includes a processor 820 and a transceiver 810 .
  • the terminal device 800 further includes a memory.
  • the processor 820, the transceiver 810, and the memory may communicate with each other through an internal connection path, and transmit control signals and/or data signals.
  • the memory is used to store computer programs, and the processor 820 is used to execute the computer programs in the memory to control the transceiver 810 to send and receive signals.
  • the above-mentioned processor 820 can be used to execute the actions described in the previous method implemented inside the terminal device, and the transceiver 810 can be used to execute the actions described in the previous method that the terminal device sends to or receives from the network device.
  • the transceiver 810 can be used to execute the actions described in the previous method that the terminal device sends to or receives from the network device.
  • the terminal device 800 may further include a power supply, configured to provide power to various devices or circuits in the terminal device.
  • FIG. 9 is a schematic structural diagram of a network device 900 provided in the present disclosure.
  • the network device 900 may be applied to the system shown in FIG. 1 to perform the functions of the network device in the above method.
  • the network device 900 includes a processor 920 and a transceiver 910 .
  • the network device 900 further includes a memory.
  • the processor 920, the transceiver 910, and the memory may communicate with each other through an internal connection path, and transmit control and/or data signals.
  • the memory is used to store computer programs, and the processor 920 is used to execute the computer programs in the memory to control the transceiver 910 to send and receive signals.
  • the above-mentioned processor 920 can be used to execute the actions described in the previous method implemented by the network device, and the transceiver 910 can be used to perform the actions described in the previous method sent by the network device to the network device or received from the network device.
  • the transceiver 910 can be used to perform the actions described in the previous method sent by the network device to the network device or received from the network device.
  • the foregoing network device 900 may further include a power supply, configured to provide power to various devices or circuits in the network device.
  • the processor and the memory may be combined into a processing device, and the processor is used to execute the program code stored in the memory to realize the above functions.
  • the memory may also be integrated in the processor, or be independent of the processor.
  • the processor may correspond to the processing unit in FIG. 7 .
  • the transceiver may correspond to the transceiver unit in FIG. 7 .
  • the transceiver 810 may include a receiver (or called a receiver, a receiving circuit) and a transmitter (or called a transmitter, a transmitting circuit). Among them, the receiver is used to receive signals, and the transmitter is used to transmit signals.
  • a processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component, and may realize or execute the present disclosure
  • a general purpose processor may be a microprocessor or any conventional processor or the like. The steps combined with the method of the present disclosure may be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.
  • the memory may be a non-volatile memory, such as a hard disk (hard disk drive, HDD) or a solid-state drive (solid-state drive, SSD), etc., or a volatile memory (volatile memory), such as random access Memory (random-access memory, RAM).
  • a memory is, but is not limited to, any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer.
  • the memory in the present application may also be a circuit or any other device capable of implementing a storage function for storing program instructions and/or data.
  • the present disclosure also provides a processing device, including a processor and a (communication) interface; the processor is configured to perform any of the above methods.
  • the above processing device may be one or more chips.
  • the processing device may be a field programmable gate array (field programmable gate array, FPGA), an application specific integrated circuit (ASIC), or a system chip (system on chip, SoC). It can be a central processor unit (CPU), a network processor (network processor, NP), a digital signal processing circuit (digital signal processor, DSP), or a microcontroller (micro controller unit) , MCU), can also be a programmable controller (programmable logic device, PLD) or other integrated chips.
  • CPU central processor unit
  • NP network processor
  • DSP digital signal processor
  • microcontroller micro controller unit
  • PLD programmable logic device
  • the present disclosure also provides a computer program product, the computer program product comprising: computer program code, when the computer program code is executed by one or more processors, causing an apparatus including the processor to execute The method shown in Fig. 4, Fig. 5C and Fig. 6 .
  • the technical solution provided by the present disclosure may be fully or partially realized by software, hardware, firmware or any combination thereof.
  • software When implemented using software, it may be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer instructions.
  • the processes or functions according to the present invention will be generated in whole or in part.
  • the above computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, and the computer-readable storage medium may be any available medium that can be accessed by a computer or contain One or more data storage devices such as servers and data centers that can be integrated with media.
  • the available medium may be a magnetic medium (for example, a floppy disk, a hard disk, or a magnetic tape), an optical medium (for example, a digital video disc (digital video disc, DVD)), or a semiconductor medium.
  • the present disclosure also provides a computer-readable storage medium, the computer-readable storage medium stores program code, and when the program code is executed by one or more processors, the The device executes the methods shown in FIG. 4 , FIG. 5C , and FIG. 6 .
  • the present disclosure also provides a system, which includes the aforementioned one or more first devices.
  • the system can also further include the aforementioned one or more third devices.
  • the disclosed systems, devices and methods may be implemented in other ways.
  • the devices described above are only schematic.
  • the division of the units is only a logical function division.
  • there may be other division methods for example, multiple units or components can be combined or integrated.
  • to another system or some features may be ignored, or not implemented.
  • the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of this solution.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Databases & Information Systems (AREA)
  • Electromagnetism (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本公开提供了一种通信方法和通信装置，该方法包括：通信装置接收第一信息，所述第一信息用于指示第一智能模型的训练策略；通信装置根据所述训练策略，对所述第一智能模型进行模型训练。能够在减小空口资源开销的情况下，得到满足通信性能需求的智能模型。

Description

通信方法和通信装置
本申请要求于2021年12月02日提交中国国家知识产权局、申请号为202111462667.X、申请名称为“通信方法和通信装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本公开涉及通信领域，并且更具体地，涉及一种通信方法和通信装置。
背景技术
在无线通信网络中，例如在移动通信网络中，网络支持的业务越来越多样，因此需要满足的需求越来越多样。例如，网络需要能够支持超高速率、超低时延和/或超大连接。该特点使得网络规划、网络配置和/或资源调度越来越复杂。此外，由于网络的功能越来越强大，例如支持的频谱越来越高、支持高阶多入多出(multiple input multiple output，MIMO)技术、支持波束赋形、和/或支持波束管理等新技术，使得网络节能成为了热门研究课题。这些新需求、新场景和新特性给网络规划、运维和高效运营带来了前所未有的挑战。为了迎接该挑战，可以将人工智能技术引入无线通信网络中，从而实现网络智能化。基于此，如何在网络中有效地实现人工智能是一个值得研究的问题。
发明内容
本公开提供了一种通信方法和通信装置，能够在减小空口资源开销的情况下，得到满足通信性能需求的智能模型。
第一方面，提供了一种通信方法，该方法可以由终端或配置于(或用于)终端的模块(如芯片)执行。
该方法包括：接收第一信息，该第一信息用于指示第一智能模型的训练策略；根据该训练策略，对该第一智能模型进行模型训练。
根据上述方案，网络可以通知终端智能模型的训练策略，由终端基于网络提供的训练策略执行智能模型的训练，减小了接入网节点向终端发送智能模型的模型参数(如每个神经元的权重值、激活函数和偏置等)带来的空口资源开销，能够在尽量不影响业务数据的传输效率的情况下，使得终端训练得到的模型能够与网络采用的模型相匹配，且满足应用于实际通信中的性能需求。
结合第一方面，在第一方面的某些实施方式中，该训练策略包括以下一项或多项：
模型训练方式、损失函数信息、模型初始化方式、模型优化算法的类型、或优化算法的参数。
可选地，上述模型训练方式可以是但不限于监督学习、无监督学习或强化学习中的一种训练方式。
可选地，上述损失函数信息可以指示训练第一智能模型采用的损失函数，该损失函数可以是但不限于交叉熵损失函数或平方损失函数。或者该损失函数还可以是一个机器学习智能模型。
可选地，上述模型初始化方式可以是第一智能模型的每个神经元的权值的初始化方式。一个示例中，第一信息可以指示该模型初始化方式为每个神经元的初始权重为预设范围内的随机值。可选地，该第一信息包括该预设范围的起始值和终止值。另一个示例中，第一信息可以指示神经元的初始权重均为0。又一个示例中，第一信息可以指示神经元的初始权重采用预设函数生成。又一个示例中，第一信息可以指示预定义多种初始化方式其中一种初始化方式的标识信息。
可选地，上述该模型优化算法可以是自适应动量估计(adaptive momentum estimation，ADAM)算法、随机梯度下降算法(stochastic gradient descent，SGD)或批量梯度下降算法(batch gradient descent，BGD)。
可选地，上述优化算法的参数可以包括但不限于以下一项或多项：
学习率、迭代次数、或批量处理的数据量。
结合第一方面，在第一方面的某些实施方式中，该方法还包括：获取第二信息，该第二信息用于指示第一智能模型的结构。
根据上述方案，由接入网节点通知终端第一智能模型的结构，终端基于接入网节点的指示生成具有该结构的初始第一智能模型，并基于指示的训练策略训练该第一智能模型，使得终端训练得到的第一智能模型能够与网络设备采用的模型相匹配，满足接入网节点的性能需求，从而能够使用第一智能模型提高无线通信的性能。
结合第一方面，在第一方面的某些实施方式中，该第二信息具体用于指示该第一智能模型的以下一种或多种结构信息：
网络层结构信息、输入数据的维度或输出数据的维度。
根据上述方案，接入网节点可以通过第二信息通知终端网络层结构信息(如神经网络层数、类型等)、输入数据的维度、输出数据的维度中的一项或多项，这些结构信息的信息量远小于神经网络中每个神经元的权重值、激活函数和偏置等的信息量，能够减小资源占用率的情况小，使得终端训练得到满足性能需求的第一智能模型。
结合第一方面，在第一方面的某些实施方式中，该网络层结构信息包括以下一项或多项：
该第一智能模型包含的神经网络层数、该神经网络层的类型、该神经网络层的使用方式、神经网络层之间的级联关系、该神经网络层的输入数据的维度或该神经网络层的输出数据的维度。
结合第一方面，在第一方面的某些实施方式中，该方法还包括：接收第三信息，该第三信息用于指示训练数据集合的信息，该训练数据集合用于训练该第一智能模型。
可选地，该训练数据集合可以包括训练样本，或者训练样本和标签。
可选地，训练数据集合的信息可以指示训练样本的类型，或者训练样本的类型和标签的类型。可选地，训练数据集合的信息可以指示第一智能模型的用途。可选地，该第三信息可以包括训练数据集合。
根据上述方案，接入网节点还可以通知终端其训练第一智能模型时使用的训练数据集合。使得终端训练得到的第一智能模型能够更接近接入网节点的需求，或者说更匹配接入网节点的智能模型。
结合第一方面，在第一方面的某些实施方式中，该方法还包括：发送能力信息，该能力信息用于指示运行智能模型的能力。
根据上述方案，终端可以通过能力信息向接入网节点终端设备运行智能模型的能力，以便接入网节点可以基于终端的能力向终端指示满足终端能力需求的训练策略和/或模型结构。
结合第一方面，在第一方面的某些实施方式中，该能力信息具体用于指示以下一项或多项能力：是否支持运行智能模型、支持运行的智能模型的类型、数据处理能力或存储能力。
结合第一方面，在第一方面的某些实施方式中，该方法还包括：接收第四信息，该第四信息用于指示测试信息，该测试信息用于测试该第一智能模型的性能。
根据上述方案，接入网节点还可以向终端发送测试信息，使得终端能够基于测试信息对第一智能模型进行测试。
结合第一方面，在第一方面的某些实施方式中，该测试信息包括以下一项或多项：
测试数据信息、性能评估方式或性能评估参数。
结合第一方面，在第一方面的某些实施方式中，该方法还包括：发送第五信息，该第五信息用于指示该第一智能模型的测试结果。
根据上述方案，终端基于测试信息测试第一智能模型后，确定训练后的第一智能模型是否满足性能需求，并通过第五信息通知接入网节点测试结果。
结合第一方面，在第一方面的某些实施方式中，发送第六信息，该第六信息用于指示推理数据，该推理数据是该第一智能模型推理测试数据得到的，该测试信息包括该测试数据。
根据上述方案，终端可以推理测试数据得到推理数据，并通过第六信息将推理数据发送给接入网节点，使得接入网节点能够基于推理数据确定终端训练后的第一智能模型是否满足性能需求。
结合第一方面，在第一方面的某些实施方式中，该方法还包括：接收第七信息，该第七信息用于指示该第一智能模型的更新后的训练策略，和/或，用于指示该第一智能模型的更新后的结构。
根据上述方案，若接入网节点确定终端训练的第一智能模型不满足性能需求，则接入网节点通过第七信息通知终端更新后的训练策略和/或更新后的模型结构，终端可以基于第七信息再次训练第一智能模型。
结合第一方面，在第一方面的某些实施方式中，该第七信息具体用于指示该训练策略的至少一个变换量，和/或该第一智能模型的结构的至少一个变化量。
根据上述方案，第七信息具体可以指示训练策略的至少一个变化量和/或模型结构的至少一个变化量，能够减小信息开销。
第二方面，提供了一种通信方法，该方法可以由接入网节点或配置于(或用于)接入网节点的模块(如芯片)执行。
该方法包括：发送第一信息，该第一信息用于指示第一智能模型的训练策略。
关于该训练策略与第一方面中介绍的训练策略相同，为了简要，在此不再赘述。
结合第二方面，在第二方面的某些实施方式中，该方法还包括：发送第二信息，该第二信息用于指示第一智能模型的结构。
关于第二信息具体指示的结构信息与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第二方面，在第二方面的某些实施方式中，该方法还包括：发送第三信息，该第三信息用于指示用于训练数据集合的信息，该训练数据集合用于训练该第一智能模型。
结合第二方面，在第二方面的某些实施方式中，该方法还包括：接收能力信息，该能力信息用于指示运行智能模型的能力。
根据该方法，可以根据该能力信息，确定该训练策略和/或该第一智能模型的结构。
关于能力信息具体指示的能力与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第二方面，在第二方面的某些实施方式中，该根据该能力信息，确定该训练策略和/或该第一智能模型的结构，包括：根据该能力信息，确定该第一智能模型的结构；和/或，根据该训练策略训练该第一智能模型。
结合第二方面，在第二方面的某些实施方式中，该方法还包括：
发送第四信息，该第四信息用于指示测试该第一智能模型性能的测试信息。
关于测试信息包含的内容与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第二方面，在第二方面的某些实施方式中，该方法还包括：
接收第五信息，该第五信息用于指示该第一智能模型的测试结果；和/或，
接收第六信息，该第六信息用于指示推理数据，该推理数据是该第一智能模型推理测试数据得到的，该测试信息包括该测试数据。
结合第二方面，在第二方面的某些实施方式中，该第七信息具体用于指示该训练策略的至少一个参数变换量，和/或该第一智能模型的结构的至少一个参数变化量。
第三方面，提供了一种通信方法，该方法可以由终端或配置于(或用于)终端的模块(如芯片)执行，以下以该方法由终端执行为例进行说明。
该方法包括：接收第二信息，该第二信息用于指示第一智能模型的结构，该第二信息包括该第一智能模型的网络层结构信息、输入数据的维度或输出数据的维度中的一种或多种结构信息；根据该第二信息，确定该第一智能模型的结构；对该第一智能模型进行模型训练。
关于该网络层结构信息与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第三方面，在第三方面的某些实施方式中，该方法还包括：获取第一信息，该第一信息用于指示第一智能模型的训练策略；以及，该对该第一智能模型进行模型训练，包括：根据该训练策略，对该第一智能模型进行模型训练。
关于该训练策略与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第三方面，在第三方面的某些实施方式中，该方法还包括：接收第三信息，该第三信息用于指示训练数据集合的信息，该训练数据集合用于训练该第一智能模型。
结合第三方面，在第三方面的某些实施方式中，该方法还包括：发送能力信息，该能力信息用于指示运行智能模型的能力。
关于该能力信息与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第三方面，在第三方面的某些实施方式中，该方法还包括：接收第四信息，该第四信息用于指示测试信息，该测试信息用于测试该第一智能模型的性能。
关于该测试信息与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第三方面，在第三方面的某些实施方式中，该方法还包括：发送第五信息，该第五信息用于指示该第一智能模型的测试结果；和/或，发送第六信息，该第六信息用于指示推理数据，该推理数据是该第一智能模型推理测试数据得到的，该测试信息包括该测试 数据。
结合第三方面，在第三方面的某些实施方式中，该方法还包括：接收第七信息，该第七信息用于指示该第一智能模型的更新后的训练策略，和/或，用于指示该第一智能模型的更新后的结构。
结合第三方面，在第三方面的某些实施方式中，该第七信息具体用于指示该训练策略的至少一个变换量，和/或该第一智能模型的结构的至少一个变化量。
第四方面，提供了一种通信方法，该方法可以由接入网节点或配置于(或用于)接入网节点的模块(如芯片)执行。
该方法包括：发送第二信息，该第二信息用于指示第一智能模型的结构，该第二信息包括该第一智能模型的网络层结构信息、输入数据的维度或输出数据的维度中的一种或多种结构信息。
关于该网络层结构信息与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第四方面，在第四方面的某些实施方式中，该方法还包括：发送第一信息，该第一信息用于指示第一智能模型的训练策略。
关于该训练策略与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第四方面，在第四方面的某些实施方式中，该方法还包括：发送第三信息，该第三信息用于指示训练数据集合的信息，该训练数据集合用于训练该第一智能模型。
结合第四方面，在第四方面的某些实施方式中，该方法还包括：接收能力信息，该能力信息用于指示运行智能模型的能力。
关于该能力信息与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第四方面，在第四方面的某些实施方式中，该方法还包括：发送第四信息，该第四信息用于指示测试信息，该测试信息用于测试该第一智能模型的性能。
关于该测试信息与第一方面中介绍的相同，为了简要，在此不再赘述。
结合第四方面，在第四方面的某些实施方式中，该方法还包括：接收第五信息，该第五信息用于指示该第一智能模型的测试结果；和/或，接收第六信息，该第六信息用于指示推理数据，该推理数据是该第一智能模型推理测试数据得到的，该测试信息包括该测试数据。
结合第四方面，在第四方面的某些实施方式中，该方法还包括：发送第七信息，该第七信息用于指示该第一智能模型的更新后的训练策略，和/或，用于指示该第一智能模型的更新后的结构。
结合第四方面，在第四方面的某些实施方式中，该第七信息具体用于指示该训练策略的至少一个变换量，和/或该第一智能模型的结构的至少一个变化量。
第五方面，提供了一种通信装置，一种设计中，该装置可以包括执行第一方面中所描述的方法/操作/步骤/动作所一一对应的模块，该模块可以是硬件电路，也可是软件，也可以是硬件电路结合软件实现。一种设计中，该装置包括：收发单元，用于接收第一信息，该第一信息用于指示第一智能模型的训练策略；处理单元，用于根据该训练策略，对该第一智能模型进行模型训练。
第六方面，提供了一种通信装置，一种设计中，该装置可以包括执行第二方面中所描述的方法/操作/步骤/动作所一一对应的模块，该模块可以是硬件电路，也可是软件，也可以是硬件电路结合软件实现。一种设计中，该装置包括：处理单元，用于确定该第一智能 模型的训练策略；收发单元，用于发送第一信息，该第一信息用于指示第一智能模型的训练策略。
第七方面，提供了一种通信装置，一种设计中，该装置可以包括执行第三方面中所描述的方法/操作/步骤/动作所一一对应的模块，该模块可以是硬件电路，也可是软件，也可以是硬件电路结合软件实现。一种设计中，该装置包括：收发单元，用于接收第二信息，该第二信息用于指示第一智能模型的结构，该第二信息包括该第一智能模型的网络层结构信息、输入数据的维度或输出数据的维度中的一种或多种结构信息；处理单元，用于根据该第二信息，确定该第一智能模型的结构；以及该处理单元还用于对该第一智能模型进行模型训练。
第八方面，提供了一种通信装置，一种设计中，该装置可以包括执行第四方面中所描述的方法/操作/步骤/动作所一一对应的模块，该模块可以是硬件电路，也可是软件，也可以是硬件电路结合软件实现。一种设计中，该装置包括：处理单元，用于确定第一智能模型的结构；收发单元，用于发送第二信息，该第二信息用于指示第一智能模型的结构，该第二信息包括该第一智能模型的网络层结构信息、输入数据的维度或输出数据的维度中的一种或多种结构信息。
第九方面，提供了一种通信装置，包括处理器。该处理器可以实现上述第一方面或第三方面以及第一方面或第三方面中任一种可能实现方式中的方法。可选地，该通信装置还包括存储器，该处理器与该存储器耦合，可用于执行存储器中的指令，以实现上述第一方面或第三方面以及第一方面或第三方面中任一种可能实现方式中的方法。可选地，该通信装置还包括通信接口，处理器与通信接口耦合。本公开中，通信接口可以是收发器、管脚、电路、总线、模块或其它类型的通信接口，不予限制。
在一种实现方式中，该通信装置为终端。当该通信装置为终端时，该通信接口可以是收发器，或，输入/输出接口。
在另一种实现方式中，该通信装置为配置于终端中的芯片。当该通信装置为配置于终端中的芯片时，该通信接口可以是输入/输出接口。
可选地，该收发器可以为收发电路。可选地，该输入/输出接口可以为输入/输出电路。
第十方面，提供了一种通信装置，包括处理器。该处理器可以实现上述第二方面或第四方面以及第二方面或第四方面中任一种可能实现方式中的方法。可选地，该通信装置还包括存储器，该处理器与该存储器耦合，可用于执行存储器中的指令，以实现上述第二方面或第四方面以及第二方面或第四方面中任一种可能实现方式中的方法。可选地，该通信装置还包括通信接口，处理器与通信接口耦合。
在一种实现方式中，该通信装置为接入网节点。当该通信装置为接入网节点时，该通信接口可以是收发器，或，输入/输出接口。
在另一种实现方式中，该通信装置为配置于接入网节点中的芯片。当该通信装置为配置于第一接入网节点中的芯片时，该通信接口可以是输入/输出接口。
可选地，该收发器可以为收发电路。可选地，该输入/输出接口可以为输入/输出电路。
第十一方面，提供了一种处理器，包括：输入电路、输出电路和处理电路。该处理电路用于通过该输入电路接收信号，并通过该输出电路发射信号，使得该处理器执行第一方面至第四方面以及第一方面至第四方面中任一种可能实现方式中的方法。
在具体实现过程中，上述处理器可以为一个或多个芯片，输入电路可以为输入管脚， 输出电路可以为输出管脚，处理电路可以为晶体管、门电路、触发器和各种逻辑电路等。输入电路所接收的输入的信号可以是由例如但不限于接收器接收并输入的，输出电路所输出的信号可以是例如但不限于输出给发射器并由发射器发射的，且输入电路和输出电路可以是同一电路，该电路在不同的时刻分别用作输入电路和输出电路。本公开对处理器及各种电路的具体实现方式不做限定。
第十二方面，提供了一种计算机程序产品，该计算机程序产品包括：计算机程序(也可以称为代码，或指令)，当该计算机程序被运行时，使得计算机执行上述第一方面至第四方面以及第一方面至第四方面中任一种可能实现方式中的方法。
第十三方面，提供了一种计算机可读存储介质，该计算机可读存储介质存储有计算机程序(也可以称为代码，或指令)当其在计算机上运行时，使得计算机执行上述第一方面至第四方面以及第一方面至第四方面中任一种可能实现方式中的方法。
第十四方面，提供了一种通信系统，包括前述的至少一个用于实现终端的方法的装置和至少一个用于实现接入网节点的方法的装置。
附图说明
图1是本公开提供的通信系统的一个示意图；
图2A是本公开提供的AI在通信系统中的应用框架的一个示例图；
图2B至图2E是本公开提供的网络架构的示意图；
图2F是本公开提供的神经网络的层关系的一个示意图；
图3是本公开提供的智能模型的训练过程的一个示意图；
图4是本公开提供的通信方法的一个示意性流程图；
图5A是本公开提供的不采用dropout方式时神经元的输入输出关系的示意图；
图5B是本公开提供的采用dropout方式时神经元的输入输出关系的示意图；
图5C是本公开提供的通信方法的另一个示意性流程图；
图6是本公开提供的通信方法的另一个示意性流程图；
图7是本公开提供的通信装置的一例的示意性框图；
图8是本公开提供的终端设备的一例的示意性结构图；
图9是本公开提供的网络设备的一例的示意性结构图。
具体实施方式
本公开中，至少一个(项)还可以描述为一个(项)或多个(项)，多个(项)可以是两个(项)、三个(项)、四个(项)或者更多个(项)，不予限制。“/”可以表示前后关联的对象是一种“或”的关系，例如，A/B可以表示A或B；“和/或”可以用于描述关联对象存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况，其中A，B可以是单数或者复数。为了便于描述本公开的技术方案，可以采用“第一”、“第二”、“A”、或“B”等字样对功能相同或相似的技术特征进行区分。该“第一”、“第二”、“A”、或“B”等字样并不对数量和执行次序进行限定。并且，“第一”、“第二”、“A”、或“B”等字样也并不限定一定不同。“示例性的”或者“例如”等词用于表示例子、例证或说明，被描述为“示例性的”或者“例如”的任何设计方案不应被解释为比其它设计方案更优选或更具优势。使用“示例性的” 或者“例如”等词旨在以具体方式呈现相关概念，便于理解。
图1是本公开能够应用的通信系统1000的架构示意图。如图1所示，该通信系统包括无线接入网(radio access network，RAN)100和核心网(core network，CN)200。可选的，通信系统1000还可以包括互联网300。其中，无线接入网100可以包括至少一个接入网节点(如图1中的110a和110b)，还可以包括至少一个终端(如图1中的120a-120j)。终端通过无线通信的方式与接入网节点相连。接入网节点通过无线通信或有线通信的方式与核心网连接。核心网设备与接入网节点可以是独立的不同的物理设备；或者可以是集成了核心网设备的功能与接入网节点的功能的同一个物理设备；或者可以是其他可能的情况，例如一个物理设备上可以集成接入网节点的功能和部分核心网设备的功能，另一个物理设备实现核心网设备的其余部分功能。本公开不限制核心网设备和接入网节点的物理存在形式。终端和终端之间可以通过有线或无线的方式相互连接。接入网节点和接入网节点之间可以通过有线或无线的方式相互连接。图1只是示意图，该通信系统中还可以包括其它网络设备，如还可以包括无线中继设备和无线回传设备等。
接入网节点可以是接入网设备，如基站(base station)、节点B(Node B)、演进型节点B(evolved NodeB，eNodeB或eNB)、发送接收点(transmission reception point，TRP)、第五代(5th generation，5G)移动通信系统中的下一代节点B(next generation NodeB，gNB)、开放无线接入网(open radio access network，O-RAN或open RAN)中的接入网节点、第六代(6th generation，6G)移动通信系统中的下一代基站、未来移动通信系统中的基站或无线保真(wireless fidelity，WiFi)系统中的接入节点等。或者，接入网节点可以是完成基站部分功能的模块或单元，例如，可以是集中式单元(central unit，CU)、分布式单元(distributed unit，DU)、集中单元控制面(CU control plane，CU-CP)模块、或集中单元用户面(CU user plane，CU-UP)模块。接入网节点可以是宏基站(如图1中的110a)，也可以是微基站或室内站(如图1中的110b)，还可以是中继节点或施主节点等。本公开中对接入网节点所采用的具体技术和具体设备形态不做限定。其中，5G系统还可以被称为新无线(new radio，NR)系统。
在本公开中，用于实现接入网节点的功能的装置可以是接入网节点；也可以是能够支持接入网节点实现该功能的装置，例如芯片系统、硬件电路、软件模块、或硬件电路加软件模块，该装置可以被安装在接入网节点中或可以与接入网节点匹配使用。在本公开中，芯片系统可以由芯片构成，也可以包括芯片和其他分立器件。为了便于描述，下文以用于实现接入网节点的功能的装置是接入网节点，并可选的以接入网节点是基站为例，描述本公开提供的技术方案。
终端也可以称为终端设备、用户设备(user equipment，UE)、移动台、移动终端等。终端可以广泛应用于各种场景进行通信。该场景例如包括但不限于以下至少一个场景：增强移动宽带(enhanced mobile broadband，eMBB)、超高可靠性超低时延通信(ultra-reliable low-latency communication，URLLC)、大规机器类型通信(massive machine-type communications，mMTC)、设备到设备(device-to-device，D2D)、车物(vehicle to everything，V2X)、机器类型通信(machine-type communication，MTC)、物联网(internet of things，IOT)、虚拟现实、增强现实、工业控制、自动驾驶、远程医疗、智能电网、智能家具、智能办公、智能穿戴、智能交通、或智慧城市等。终端可以是手机、平板电脑、带无线收发功能的电脑、可穿戴设备、车辆、无人机、直升机、飞机、轮船、机器人、机械臂、或智 能家居设备等。本公开对终端所采用的具体技术和具体设备形态不做限定。
在本公开中，用于实现终端的功能的装置可以是终端；也可以是能够支持终端实现该功能的装置，例如芯片系统、硬件电路、软件模块、或硬件电路加软件模块，该装置可以被安装在终端中或可以与终端匹配使用。为了便于描述，下文以用于实现终端的功能的装置是终端为例，并可选的以终端是UE为例，描述本公开提供的技术方案。
基站和/或终端可以是固定位置的，也可以是可移动的。基站和/或终端可以部署在陆地上，包括室内或室外、手持或车载；或者可以部署在水面上；或者可以部署在空中的飞机、气球和人造卫星上。本公开对基站和终端所处的环境/场景不做限定。基站和终端可以部署在相同的或不同的环境/场景，例如，基站和终端同时部署在陆地上；或者，基站部署在陆地上，终端部署在水面上等，不再一一举例。
基站和终端的角色可以是相对的。例如，图1中的直升机或无人机120i可以被配置成移动基站，对于那些通过120i接入到无线接入网100的终端120j来说，终端120i是基站；但对于基站110a来说，120i可以是终端，即110a与120i之间可以是通过无线空口协议进行通信的。或者，110a与120i之间是通过基站与基站之间的接口协议进行通信的，此时，相对于110a来说，120i也是基站。因此，基站和终端都可以统一称为通信装置(或通信设备)，图1中的110a和110b可以称为具有基站功能的通信装置，图1中的120a-120j可以称为具有终端功能的通信装置。
可选的，接入网节点和终端之间的协议层结构可以包括人工智能(artificial intelligence，AI)层，用于传输AI功能相关的数据。
在本公开中，可以在前述图1所示的通信系统中引入独立的网元(如称为AI实体、AI网元、AI节点、或AI设备等)来实现AI相关的操作。AI网元可以和基站直接连接，或者可以通过第三方网元和基站实现间接连接。可选的，第三方网元可以是接入和移动性管理功能移动管理功能(access and mobility management function，AMF)网元或用户面功能(user plane function，UPF)网元等核心网网元。或者，可以在通信系统中的网元内配置AI实体(或者，可以称为AI模块或其他名称)来实现AI相关的操作。可选的，该配置有AI实体的网元可以是基站、核心网设备、或网管(operation,administration and maintenance，OAM)等。其中，OAM用于操作、管理和/或维护核心网设备(核心网设备的网管)，和/或，用于操作、管理和/或维护接入网节点(接入网节点的网管)。
可选的，为了匹配支持AI，终端或终端芯片中可以集成AI实体。
可选的，本公开中，AI实体还可以称为其他名称，如AI模块或AI单元等，主要用于实现AI功能(或称为AI相关的操作)，本公开不限制其具体名称。
本公开中，AI模型是实现AI功能的具体方法。AI模型表征了模型的输入和输出之间的映射关系。AI模型可以是神经网络、线性回归模型、决策树模型、聚类SVD模型或者其他机器学习模型。其中，AI模型可以简称为智能模型、模型或其他名称，不予限制。AI相关的操作可以包括以下至少一项：数据收集、模型训练、模型信息发布、模型测试(或称为模型校验)、模型推断(或称为模型推理、推理、或预测等)、或推理结果发布等。
如图2A所示为AI在通信系统中的第一种应用框架的示例图。在图2A中，数据源(data source)用于存储训练数据和推理数据。模型训练节点(model trainning host)通过对数据源提供的训练数据(training data)进行分析或训练，得到AI模型，且将 AI模型部署在模型推理节点(model inference host)中。其中，AI模型表征了模型的输入和输出之间的映射关系。通过模型训练节点学习得到AI模型，相当于由模型训练节点利用训练数据学习得到模型的输入和输出之间的映射关系。模型推理节点使用AI模型，基于数据源提供的推理数据进行推理，得到推理结果。该方法还可以描述为：模型推理节点将推理数据输入到AI模型，通过AI模型得到输出，该输出即为推理结果。该推理结果可以指示：由执行对象使用(执行)的配置参数、和/或由执行对象执行的操作。推理结果可以由执行(actor)实体统一规划，并发送给一个或多个执行对象(例如，网元)去执行。可选的，模型推理节点可以将其推理结果反馈给模型训练节点，该过程可以称为模型反馈，所反馈的参数用于模型训练节点更新AI模型，并将更新后的AI模型部署在模型推理节点中。可选的，执行对象可以将其收集到的网络参数反馈给数据源，该过程可以称为表现反馈，所反馈的参数可以作为训练数据或推理数据。
在本公开中，图2A所示的应用框架可以部署在图1中所示的网元。例如，图2A的应用框架可以部署在图1的终端设备、接入网设备、核心网设备、或独立部署的AI网元(未示出)中的至少一项。例如，AI网元(可看做模型训练节点)可对终端设备和/或接入网设备提供的训练数据(training data)进行分析或训练，得到一个模型。终端设备、接入网设备、或核心网设备中的至少一项(可看做模型推理节点)可以使用该模型和推理数据进行推理，得到模型的输出。其中，推理数据可以是由终端设备和/或接入网设备提供的。该模型的输入包括推理数据，该模型的输出即为该模型所对应的推理结果。终端设备、接入网设备、或核心网设备中的至少一项(可看做执行对象)可以根据推理数据和/或推理结果进行相应的操作。其中，模型推理节点和执行对象可以相同，也可以不同，不予限制。
下面结合图2B～2E对本公开提供的通信方案能够应用的网络架构进行介绍。
如图2B所示，第一种可能的实现中，接入网设备中包括近实时接入网智能控制(RAN intelligent controller，RIC)模块，用于进行模型学习和推理。例如，近实时RIC可以用于训练AI模型，利用该AI模型进行推理。例如，近实时RIC可以从CU、DU或RU中的至少一项获得网络侧和/或终端侧的信息，该信息可以作为训练数据或者推理数据。可选的，近实时RIC可以将推理结果递交至CU、DU或RU中的至少一项。可选的，CU和DU之间可以交互推理结果。可选的，DU和RU之间可以交互推理结果，例如近实时RIC将推理结果递交至DU，由DU转发给RU。
如图2B所示，第二种可能的实现中，接入网之外包括非实时RIC(可选的，非实时RIC可以位于OAM中或者核心网设备中)，用于进行模型学习和推理。例如，非实时RIC用于训练AI模型，利用该模型进行推理。例如，非实时RIC可以从CU、DU或RU中的至少一项获得网络侧和/或终端侧的信息，该信息可以作为训练数据或者推理数据，该推理结果可以被递交至CU、DU或RU中的至少一项。可选的，CU和DU之间可以交互推理结果。可选的，DU和RU之间可以交互推理结果，例如非实时RIC将推理结果递交至DU，由DU转发给RU。
如图2B所示，第三种可能的实现中，接入网设备中包括近实时RIC，接入网设备之外包括非实时RIC(可选的，非实时RIC可以位于OAM中或者核心网设备中)。同上述第二种可能的实现，非实时RIC可以用于进行模型学习和推理；和/或，同上 述第一种可能的实现，近实时RIC可以用于进行模型学习和推理；和/或，近实时RIC可以从非实时RIC获得AI模型信息，并从CU、DU或RU中的至少一项获得网络侧和/或终端侧的信息，利用该信息和该AI模型信息得到推理结果，可选的，近实时RIC可以将推理结果递交至CU、DU或RU中的至少一项，可选的，CU和DU之间可以交互推理结果，可选的，DU和RU之间可以交互推理结果，例如近实时RIC将推理结果递交至DU，由DU转发给RU。例如，近实时RIC用于训练模型A，利用模型A进行推理。例如，非实时RIC用于训练模型B，利用模型B进行推理。例如，非实时RIC用于训练模型C，将模型C递交给近实时RIC，近实时RIC利用模型C进行推理。
图2C所示为本公开提供的方法能够应用的一种网络架构的示例图。相对图2B，图2B中将CU分离为了CU-CP和CU-UP。
图2D所示为本公开提供的方法能够应用的一种网络架构的示例图。如图2D所示，可选的，接入网设备中包括一个或多个AI实体，该AI实体的功能类似上述近实时RIC。可选的，OAM中包括一个或多个AI实体，该AI实体的功能类似上述非实时RIC。可选的，核心网设备中包括一个或多个AI实体，该AI实体的功能类似上述非实时RIC。当OAM和核心网设备中都包括AI实体时，他们各自的AI实体所训练得到的模型不同，和/或用于进行推理的模型不同。
本公开中，模型不同包括以下至少一项不同：模型的结构参数(例如模型的层数、模型的宽度、层间的连接关系、神经元的权值、神经元的激活函数、或激活函数中的偏置中的至少一项)、模型的输入参数(例如输入参数的类型和/或输入参数的维度)、或模型的输出参数(例如输出参数的类型和/或输出参数的维度)。
图2E所示为本公开提供的方法能够应用的一种网络架构的示例图。相对图2D，图2E中的接入网设备分离为CU和DU。可选的，CU中可以包括AI实体，该AI实体的功能类似上述近实时RIC。可选的，DU中可以包括AI实体，该AI实体的功能类似上述近实时RIC。当CU和DU中都包括AI实体时，他们各自的AI实体所训练得到的模型不同，和/或，用于进行推理的模型不同。可选的，还可以进一步将图2E中的CU拆分为CU-CP和CU-UP。可选的，CU-CP中可以部署有一个或多个AI模型。可选的，CU-UP中可以部署有一个或多个AI模型。
可选的，如前文所述，图2D或图2E中，接入网设备的OAM和核心网设备的OAM可以分开独立部署。图2D或图2E中，核心网设备可以包括一个或多个模块，例如包括
本公开中，一个模型可以推理得到一个参数，或者推理得到多个参数。不同模型的学习过程可以部署在不同的设备或节点中，也可以部署在相同的设备或节点中。不同模型的推理过程可以部署在不同的设备或节点中，也可以部署在相同的设备或节点中。
本公开描述的网络架构以及业务场景是为了更加清楚的说明本公开的技术方案，并不构成对于本公开提供的技术方案的限定，本领域普通技术人员可知，随着网络架构的演变和新业务场景的出现，本公开提供的技术方案对于类似的技术问题，同样适用。
在描述本公开的方法之前，先简单介绍一些关于人工智能的相关知识。人工智能，可以让机器具有人类的智能，例如可以让机器应用计算机的软硬件来模拟人类某些智能行为。为了实现人工智能，可以采取机器学习方法或其他方法，不予限制。
神经网络(neural network,NN)是机器学习的一种具体实现方式。根据通用近似定理，神经网络理论上可以逼近任意连续函数，从而使得神经网络具备学习任意映射的能力。因此神经网络可以对复杂的高维度问题进行准确地抽象建模。也就是说，智能模型可以通过神经网络实现。
神经网络的思想来源于大脑组织的神经元结构。每个神经元都对其输入值做加权求和运算，将加权求和结果通过一个激活函数产生输出。假设神经元的输入为x＝[x 0,...,x n]，与输入对应的权重值(或称为权值)为w＝[w 0,...,w n]，加权求和的偏置为b。激活函数的形式可以多样化。假设一个神经元的激活函数为：y＝f(z)＝max(0,z)，则该神经元的输出为：
Figure PCTCN2022136138-appb-000001
再例如一个神经元的激活函数为：y＝f(z)＝z，则该神经元的输出为：
Figure PCTCN2022136138-appb-000002
神经元的输入x中的元素x i、权重值w的元素w i、或偏置b可以为小数、整数(包括0、正整数或负整数等)、或复数等各种可能的取值。神经网络中不同神经元的激活函数可以相同或不同。
神经网络一般包括多层结构，每层可包括一个或多个神经元。增加神经网络的深度和/或宽度可以提高该神经网络的表达能力，为复杂系统提供更强大的信息提取和抽象建模能力。其中，神经网络的深度指神经网络包括的层数，每层包括的神经元个数可以称为该层的宽度。如图2F所示，为神经网络的层关系示意图。
一种实现中，神经网络包括输入层和输出层。神经网络的输入层将接收到的输入经过神经元处理后，将结果传递给输出层，由输出层得到神经网络的输出结果。
另一种实现中，神经网络包括输入层、隐藏层和输出层。神经网络的输入层将接收到的输入经过神经元处理后，将结果传递给中间的隐藏层，隐藏层再将计算结果传递给输出层或者相邻的隐藏层，最后由输出层得到神经网络的输出结果。一个神经网络可以包括一层或多层依次连接的隐藏层，不予限制。
AI模型的训练过程中，可以定义损失函数。损失函数描述了AI模型的输出值和理想目标值之间的差距或差异，本公开不限制损失函数的具体形式。AI模型的训练过程就是通过调整AI模型的部分或全部参数使得损失函数的值小于门限值或者满足目标需求的过程。如AI模型为神经网络，训练过程中可以调整以下一项或多项参数：神经网络的层数、神经网络的宽度、层间的连接关系、神经元的权重值、神经元的激活函数、或激活函数中的偏置，使得神经网络的输出与理想目标值之间的差异尽可能小。
图3为智能模型的训练过程的一个示例图。图3所示以智能模型为一个神经网络模型为例，且以该神经网络f θ(·)可以包括4层卷积层网络，4层全连接层网络为例进行介绍。AI实体可以从基础数据集(如信道数据的集合)中得到训练数据，训练数据可以包括训练样本和标签，样本x作为输入经过神经网络f θ(·)处理后输出推理结果f θ(x)，损失函数计算得到推理结果与标签之间的误差，AI实体可以基于损失函数得到的误差，采用反向传播优化算法(可以称为模型优化算法)，优化网络参数θ。通过大量训练数据对神经网络进行训练，使得神经网络的输出与标签之间的差异小于预设值后完成神经网络的训练。
需要说明的是，图3所示的训练过程采用了监督学习的训练方式，即基于样本与标签，利用损失函数实现模型的训练。智能模型的训练过程也可以采用无监督学习，利用算法学 习样本的内在模式，实现基于样本完成智能模型的训练。智能模型的训练过程还可以采用强化学习，通过与环境进行交互获取环境反馈的激励信号，从而学习解决问题的策略，实现模型的优化。本公开对于模型的训练方法和模型的类型不予限制。
训练之后的AI模型可以执行推理任务。实际数据输入AI模型进行处理后，得到对应的推理结果。可选地，一个AI模型可以推理得到一个参数，或者推理得到多个参数。
人工智能应用在无线通信系统中能够显著提升通信系统的性能。在大部分场景中，网络侧与终端需要使用相匹配的人工智能模型来提升无线通信性能。例如，终端可以采用压缩编码器模型对上行信息进行压缩编码后发送给网络侧，网络侧采用相匹配的解码器模型解码得到终端发送的上行信息。人工智能应用在无线通信领域中往往涉及较复杂的非线性函数拟合。智能模型的规模往往较大，例如模型层数较多、模型参数个数较多。例如，实现CSI反馈压缩的编码器模型的卷积神经网络(convolutional neural network，CNN)可能包括15层的神经网络层。
为了满足性能需求，终端使用的智能模型通常由网络指定。网络侧可以从预定义的多个模型中选择一个模型，并通知终端该模型的标识，使得终端设备能够基于标识确定采用的模型。然而终端种类多样且信道环境复杂多变，预定义的模型数量有限，无法保证在各种环境下均能得到良好的通信性能。为了保证通信性能，网络侧也可以将智能模型下发给终端，但智能模型的参数量庞大，需要占用大量空口资源，将影响业务数据的传输效率。若终端从网络下载该模型，需要下载该模型的结构参数、每个神经元的权重值、激活函数、和偏置等，总参数量高达数十万浮点数。此时，空口资源开销巨大、使得业务数据的空口资源减少，降低了业务数据的传输效率。针对该问题，本公开提出网络可以通知终端智能模型的训练策略，由终端基于网络提供的训练策略执行智能模型的训练，使得终端训练得到的模型能够与网络采用的模型相匹配，达到预期的性能需求。
下面将结合附图，对本公开提供的通信方法进行说明。需要说明的是，本文以终端和接入网节点之间的交互为例，但本申请不限于此，本公开提供的通信方法可以用于任意两个节点之间，由一个节点从另一个节点获取智能模型的训练策略和/或模型结构，并基于训练策略和/或模型结构执行模型训练。本公开中，如前文所述，可以由接入网节点的一个或多个模块(如RU、DU、CU、CU-CP、CU-UP或近实时RIC)实现接入网节点的相应方法或操作。
图4是本公开提供的通信方法400的一个示意性流程图。
S401，接入网节点向终端发送第一信息，该第一信息用于指示第一智能模型的训练策略。
相应地，终端接收来自该接入网节点的该第一信息。终端根据该第一信息，确定第一智能模型的训练策略。
本公开中，作为示例非限定，该训练策略包括以下一项或多项：
模型训练方式、损失函数信息、模型初始化方式、模型优化算法的类型、或优化算法的参数。
下面对上述各项训练策略分别进行介绍。
(1)模型训练方式
该第一信息可以指示终端训练第一智能模型的训练方式，终端可以采用第一信息指示的模型训练方式训练第一智能模型。以使训练得到的第一智能模型能够匹配网络侧的智能 模型。达到提升无线通信性能的效果。
例如，第一信息可以指示终端采用监督学习、无监督学习或强化学习中的一种方式训练第一智能模型。但本公开不限于此。
(2)损失函数信息
该第一信息可以包括损失函数信息。终端接收到第一信息后可以基于该损失函数信息得到用于训练第一智能模型的损失函数。
接入网节点可以通过损失函数信息通知终端采用的损失函数。该损失函数信息可以直接指示该损失函数，或者该损失函数信息可以指示预定义的多个损失函数中的一个损失函数的标识信息。
一个示例中，终端可以采用监督学习的方式训练该第一智能模型，例如，可以是预定义采用该监督学习的训练方式，或者接入网节点通过第一信息指示模型训练方式为监督学习的训练方式。例如，监督学习的模型训练方式的训练目标是使损失函数的值达到最小值，具体实施中，可以在训练得到该损失函数的值小于或等于第一阈值的情况下完成模型训练。可选地，该第一阈值可以是预定义的或该损失函数信息指示的。
例如，该损失函数可以是如下交叉熵损失函数：
Σ-[y n×log(f θ(x n))+(1-y n)×log(1-f θ(x n))]
其中，log(A)表示计算A的对数，Σ nB n表示基于n的取值范围对B n求和。
或者，该损失函数可以是如下平方损失函数：
Σ n(y n-y θ(x n)) 2
其中，x n为样本，y n为第n个样本的标签，f θ(x n)为模型针对输入数据x n的输出结果。
另一个示例中，终端可以采用无监督学习的方式训练该第一智能模型，对于无监督学习，损失函数可以是用于评价模型性能的一个函数。例如，无监督学习的模型训练方式的训练目标是使损失函数的值达到最大值，具体实施中，可以在训练得到该损失函数的值大于或等于第二阈值的情况下完成模型训练。可选地，该第二阈值可以是预定义的或该损失函数信息指示的。
例如，第一智能模型用于基于信道响应推理发射功率，则第一智能模型的输入为信道响应h n，第一智能模型的推理结果为发射功率P n，则损失函数可以是计算数据吞吐量的函数，如损失函数可以如下：
Figure PCTCN2022136138-appb-000003
其中，log(A)表示计算A的对数，N 0为噪声功率。
损失函数可以是上述数学表达式，可选地，损失函数也可以是一个机器学习智能模型。损失函数模型的复杂度、参数量远小于第一智能模型的复杂度、参数量，该损失函数信息可以包括损失函数的结构信息和/或参数信息，使得终端接收到该第一信息后可以得到损失函数模型，基于该损失函数进行第一智能模型的训练，能够提高终端的模型训练的效果。
例如，采用无监督学习的训练方式时，该损失函数可以是系统吞吐量、信号误比特率、信干噪比等非线性的性能度量模型。
(3)模型初始化方式
该第一信息可以指示模型初始化方式，终端可以基于第一信息指示的模型初始化方式初始化第一智能模型。
一个示例中，第一智能模型可以是神经网络模型，第一信息可以指示第一智能模型中神经元权重的初始化方式。如接入网节点可以指示终端每个神经元的权重随机取值、或神经元的权重值均为0，或神经元的权重值采用预设函数生成等方式。
例如，接入网节点可以通过第一信息指示终端在[z min，z max]区间内随机取值作为神经元的权重，如该第一信息可以指示z min和z max。终端接收到该第一信息后，在[z min，z max]区间内为每个神经元选择一个值，作为该神经元的权重初始值。
再例如，第一信息指示预定义多种初始化方式其中一种初始化方式的标识信息，如通过第一信息包括2比特指示域，用于指示神经元权重的初始化方式，该2比特指示“00”表示神经元的权重初始值采用随机取值的方式，可选地，终端可以在预设取值区间内为每个神经元选择一个权重值，或者该第一信息指示该方式时第一信息还包括取值区间的端点值，终端在第一信息指示的取值区间内为神经元的权重值；该2比特指示“01”表示神经元的初始值均设置为0；该2比特指示“10”表示采用预设函数，生成初始值序列，序列中的每个值作为一个神经元的初始值。
(4)模型优化算法的类型
如图3示例中，在模型训练中，训练装置可以基于损失函数的输出值采用模型优化算法得到智能模型的优化参数，基于该优化参数更新智能模型后，采用该更新参数后的智能模型进行下一次模型训练。接入网节点可以通过第一信息通知终端第一智能模型的优化算法。
例如，第一信息可以指示在预定义的多种类型的模型优化算法中指示一种类型的优化算法，终端可以在训练第一智能模型时，采用第一信息指示的模型优化算法进行模型优化。作为示例非限定，模型优化算法可以包括但不限于以下一种或多种算法：
自适应动量估计(adaptive momentum estimation，ADAM)算法、随机梯度下降算法(stochastic gradient descent，SGD)或批量梯度下降算法(batch gradient descent，BGD)。
可选地，通信协议可以预定义多种类型的模型优化算法，每种类型的模型优化算法的参数也可以是预定义的，终端可以基于第一信息指示的模型优化算法的类型的标识信息，确定该优化算法的类型，以及预定义的该优化算法的参数。终端在第一智能模型的训练中，采用该模型优化算法进行模型权重的更新。或者，优化算法的参数也可以是第一信息指示的，可以参考下文中的描述。
(5)优化算法的参数
第一信息可以指示模型优化算法的参数，其中，该模型优化算法的类型可以是接入网节点可以通过第一信息指示的，或者模型优化算法的类型可以是预配置的。
作为示例非限定，模型优化算法的参数可以包括但不限于以下一种或多种参数：
学习率、迭代次数、批量处理的数据量。
其中，学习率为模型训练中表征梯度下降步长的参数。学习率可以为一个定值，接入网节点可以通过第一信息通知终端，该学习率的值。例如，学习率可以设置为0.001。或者，学习率也可以设置为随迭代步骤逐渐减小的值，第一信息可以指示学习率和每次迭代学习率减小的值，如第一信息可以指示初始值为0.1，每迭代一次减小0.01。迭代次数为指训练数据集被遍历的次数。批量处理的数据量是指每次模型训练时在训练数据集中选取的、用于梯度下降更新的训练数据的数据量。
以上介绍了第一信息指示的第一智能模型的训练策略可以包括但不限于如上(1)至 (5)中的一种或多种策略。使得终端设备可以基于该训练策略，训练第一智能模型，得到与网络相匹配的智能模型。
需要说明的是，若接入网节点通知终端未指示上述训练策略(1)至(5)中的一种或多种策略，终端可以采用但不限于以下方式确定接入网节点未指示的训练策略。
一种实施方式中，终端可以预配置有上述训练策略中的一种或多种策略，在接入网节点未指示其中一种或多种策略的情况下，终端基于预配置的训练策略进行模型训练。该预配置的训练策略可以是协议定义的，或者该预配置的训练策略可以是基于终端生产实现由厂商配置的。
另一种实施方式中，终端可以基于接入网节点指示的一种或多种训练策略，确定接入网节点未指示的一种或多种训练策略。也就是说，多种训练策略之间可以具有关联关系，接入网节点通过指示一种训练策略，隐式地指示了与该训练策略相关的其他训练策略。该关联关系可以是协议约定的，或者是接入网节点预先通知终端的，不予限制。
例如，接入网节点可以通过第一信息指示模型训练方式而不指示损失函数信息，终端基于接入网节点指示的模型训练方式确定损失函数。如第一信息指示模型训练方式为监督学习的方式，终端可以确定损失函数为交叉熵损失函数，或者终端可以确定损失函数为平方损失函数。但本申请不限于此。
可选地，接入网节点还可以向终端发送第二信息，该第二信息用于指示该第一智能模型的结构。
接入网节点通知终端该第一智能模型的训练策略以外，还通过第二信息通知终端第一智能模型的结构。终端接收到该第二信息后，基于第二信息生成具有该结构的初始的第一智能模型(或者称为训练前的第一智能模型)，并采用第一信息指示的训练策略对该第一智能模型进行训练，使得训练后的第一智能模型能够与接入网节点采用的模型相匹配。
接入网节点可以在预定义的多种智能模型的结构中指示一种结构的标识信息，通知终端采用该结构的智能模型。或者，接入网节点可以指示第一智能模型的结构参数。
作为示例非限定，该第二信息具体可以用于指示该第一智能模型的以下一种或多种结构信息：
网络层结构信息、输入数据的维度或输出数据的维度。
例如，第二信息可以指示第一智能模型的输入数据的维度，即一个训练样本的维度。如第一智能模型为神经网络模型，包括输入层和输出层，则该输入数据的维度即为该第一智能模型的输入层的输入数据的维度。
再例如，该第二信息可以指示第一智能模型的输出数据的维度，即第一智能模型输出的推理数据的维度。如第一智能模型为神经网络模型，则该输出数据的维度即为该第一智能模型的输出层的输出数据的维度。
第二信息可以包括该第一智能模型的网络层结构信息，该网络层结构信息可以包括但不限于以下一项或多项：
第一智能模型包含的神经网络层的层数、神经网络层的类型、神经网络层的使用方式、神经网络层之间的级联关系、神经网络层的输入数据的维度或神经网络层的输出数据的维度。
例如，接入网节点可以指示第一智能模型包含的神经网络层的层数，以及可以指示每个神经网络层的类型，如CNN、FNN、维度变换神经网络层等。
若上述网络层结构信息指示了该第一智能模型中的一个神经网络层的类型为CNN时，CNN可以由滤波器实现。接入网节点还可以通过该网络层结构信息通知终端实现该CNN的滤波器(Filter)的类型。如Filter的类型可以是滤波器的维度。例如，网络层结构信息可以指示滤波器的维度为3×3、5×5或者3×3×3等。
维度变换神经网络层用于数据维度的变换，例如，维度变换神经网络层可以位于两个神经网络层之间，该维度变换神经网络层用于将输入数据(即前一个神经网络的输出数据)的维度变换后得到一个神经网络层的输入数据的维度，将该变换维度后的数据输出给后一个神经网络层。例如，该神经网络层A的输出数据的维度为3×3×10，而神经网络层B的输入数据的维度要求为30×3。则可以在该两个神经网络之间设置一个维度变换神经网络，将神经网络层A的输出的3×3×10的数据变换为30×3的数据，将该数据输出至神经网络层B。使得数据经过维度变换后能够由神经网络层B继续处理。
接入网节点还可以通过第二信息通知终端神经网络层的使用方式，如可以通知终端神经网络中的一个或多个神经网络层是否使用暂时丢弃(dropout)的方式。dropout方式是指在模型训练时，可以随机地选择神经网络层中的一个或多个神经元不参与本次训练，而在下一次训练中可以恢复该一个或多个神经元参与模型训练。
例如，一个CNN神经网络层不采用dropout方式时该神经网络层中的神经元的输入输出关系可以如图5A所示。若采用dropout方式，该神经网络层中的神经元的输入输出关系可以如图5B所示。终端可以随机地选择该CNN神经网络层中的多个神经元不参与模型训练，该神经元的输出为置0，且不参与模型训练的神经元在本次训练中不更细神经元的权重，在下一次参与的模型训练中继续使用该权重。
接入网节点还可以通知终端第一智能模型中的神经网络层的输入数据的维度和/或输出数据的维度。和/或，接入网节点还可以通知终端神经网络层之间的级联关系，例如，该神经网络层之间的级联关系可以是神经网络层在第一智能模型中的排列顺序。
表1示出了一个神经网络结构的示例，接入网节点可以通知终端如表1所示的神经网络结构，但本公开不限于此。如第二信息指示该第一智能模型的神经网络层数为6，并通知每个神经网络层的类型，如神经网络层的类型包括输入层、CNN、形状变换、FNN以及输出层，第二信息还可以指示神经网络层间的级联关系，如第二信息可以指示该6个神经网络层的排列顺序可以依次为输入层、CNN、CNN、形状变换、FNN以及输出层。以及第二信息还可以指示每层神经网络层的输入维度和输出维度如表2所示，其中，N T为天线数，N sc为子载波数。第二信息还可以通知终端实现第2层、第3层CNN层的滤波器为3×3的滤波器，以及，其中2、3、4、5层神经网络层可以采用dropout的使用方式。但本公开不限于此，第一智能模型的结构可以部分为预定义的部分由接入网节点通过第二信息指示的。例如，接入网节点可以通知终端如表1所示的部分或全部结构参数。其中，表1所示的N/A表示该行不适用(或不配置)该参数。
表1
Figure PCTCN2022136138-appb-000004
Figure PCTCN2022136138-appb-000005
终端可以基于第二信息指示的表1所示的第一智能模型的结构，生成初始的第一智能模型。终端首先基于第二信息可以确定该第一智能模型包括6个神经网络层，基于第二信息指示该6个神经网络层中每个神经网络层的类型、输入维度和输出维度，生成满足类型和维度要求的6个神经网络层，即输入层、CNN层、CNN层、形状变换层、FNN层以及输出层。其中，终端基于第二信息的指示通过3×3的滤波器实现其中的两个CNN层。终端得到该6个神经网络层后，基于第二信息指示的级联关系确定该6个神经网络层的排列顺序，得到初始的该第一智能模型。在训练该第一智能模型的过程中，对2、3、4、5层神经网络层采用dropout的使用方式在每次训练中随机选择网络层中的神经元不参与本次训练。
可选地，接入网节点向终端发送第三信息，该第三信息用于指示训练数据集合的信息，该训练数据集合用于训练该第一智能模型。
一个示例中，第三信息可以指示预设训练数据集合的标识信息，使得终端接收到第三信息后采用该标识信息对应的训练数据集合训练该第一智能模型。
例如，协议可以预定义多个训练数据集合，每个训练数据集合对应一个标识信息，接入网节点可以通过第三信息指示该多个训练数据集合中的一个训练数据集合的标识信息，终端根据该标识信息确定采用该标识信息对应的训练数据集合训练第一智能模型。
另一个示例中，训练数据集合的信息可以指示训练样本的类型，或者训练样本的类型和标签的类型。
例如，第一智能模型的训练方式为无监督学习的方式，该第三信息可以指示训练样本的类型，如第一智能模型可以基于输入的信道数据推理发射功率。则该第三信息可以指示训练样本的类型为信道数据，终端可以根据采用信道数据对第一智能模型进行训练。该信道数据可以是终端基于接入网节点发送的参考信号测量得到的，或者可以是预配置在终端的。但本申请不限于此。
再例如，第一智能模型的训练方式为无监督学习的方式，该第三信息可以指示训练样本的类型和标签的类型，终端可以基于训练样本的类型和标签的类型确定训练数据集合，该训练数据集合中包含的数据可以是终端与接入网节点通信中采集到的数据，或者可以是预配置在终端的相应类型的数据。例如，第一智能模型实现数据的信道编码功能，即基于输入的编码前数据推理输出编码后的数据，第三信息可以指示训练样本的类型为信道编码前的数据，标签的类型为信道编码后的数据。
另一个示例中，训练数据集合的信息还可以指示第一智能模型的用途，如训练数据集合的信息指示第一智能模型用于压缩编码、译码或发射功率。终端根据第一智能模型的用途，确定相应的训练数据集合。
例如，第三信息可以指示第一智能模型的用途为压缩编码，且第一智能模型的训练方 式为监督学习，则终端基于该第一智能模型的用途，可以确定用于训练第一智能模型的训练样本的类型为压缩编码前的信息数据，以及标签为压缩编码后的压缩数据，从而确定包括训练样本和标签的第一智能模型的训练数据集合。该训练数据集合中包含的数据可以是终端与接入网节点通信中采集到的数据，或者可以是预配置在终端的相应类型的数据。但本申请不限于此。
另一个示例中，第三信息可以包括训练数据集合，如该训练数据集合可以包括训练样本、或者训练样本和标签，也就是说终端可以从网络获取训练数据集合。
例如，第一智能模型的训练方式为监督学习的方式，该第三信息中的训练数据集合包括训练样本和标签。第一智能模型的训练方式为无监督学习的方式，该第三信息中的训练数据集合包括训练样本而不包括标签。
终端可以以小的资源带宽边下载该训练数据集合中的训练数据边训练第一智能模型，能够减小资源占用率，避免影响业务数据的传输速率。
在本公开中，接入网节点可以基于终端的能力信息，和/或，基于接入网节点与终端之间的环境数据(如信道数据等)，确定该第一智能模型的训练策略、模型结构或训练数据集合中的一项或多项。或者，接入网节点可以从网络中的第三方节点获取该第一智能模型的训练策略和/或模型结构后，转发给终端。也就是说，接入网节点向终端发送的上述信息(包括但不限于第一信息、第二信息、第三信息)可以是接入网节点生成的，也可以是接入网节点从第三方节点获取到并转发给终端的。例如，该第三方节点可以是网络中具有AI功能或配置有AI实体的节点，如第三方节点可以是非实时RIC，非实时RIC可以包括在OAM中，该第三方节点实际可以是OAM。可选的，接入网节点接收来自该终端的能力信息，该能力信息用于指示终端运行智能模型的能力。
作为示例非限定，该能力信息具体用于指示终端的以下一项或多项能力：
是否支持运行智能模型、数据处理能力、存储能力、支持运行的智能模型的类型、处理器的功耗或电池容量。
例如，该能力信息可以指示终端是否支持运行智能模型，接入网节点接收到该能力信息后，基于该能力信息确定终端支持智能模型的情况下，接入网节点向终端发送该第一信息，以通知终端训练第一智能模型的策略。
再例如，接入网节点需要通过第二信息通知终端第一智能模型的结构，该能力信息可以指示终端支持的机器学习模型的算子库信息。该算子库信息可以指示终端能够支持的基本智能模型运算单元的集合，或者，该算子库信息可以指示终端不支持的基本智能模型运算单元。接入网节点可以该能力信息获知终端支持的模型算子库信息，从而确定第一智能模型的结构。
比如，智能模型运算单元可以包括但不限于以下一项或多项：
一维CNN单元(conv1d)、二维CNN单元(conv2d)、三维CNN单元(conv3d)、池化(pooling)层、池化操作或s形(sigmoid)激活函数。
再例如，能力信息指示终端的数据处理能力，例如，能力信息可以指示处理器的类型、处理器的运算速度、处理器能够处理的数据量大小等。如处理器的类型可以包括图形处理器(graphics processing unit，GPU)的类型和/或中央处理器(central processing unit，CPU)的类型。接入网节点可以基于终端的数据处理能力确定第一智能模型的训练策略、模型结构或训练数据集合中的一项或多项。
一种实施方式中，接入网节点可以基于终端的能力信息，和/或，基于接入网节点与终端之间的环境数据(如信道数据等)，训练得到符合终端能力和/或信道条件的智能模型。接入网节点将训练该智能模型的训练策略通知终端，以便终端采用该训练策略训练第一智能模型，使得终端训练后的第一智能模型与接入网节点训练得到的该智能模型尽可能的相同，从而是终端与接入网节点在通信过程中应用的第一智能模型能够与接入网节点使用的智能模型相匹配，达到提升通信性能的效果。
例如，第一智能模型为压缩编码模型，接入网节点可以基于终端的能力信息和信道数据，训练得到相互匹配的接入网节点使用的压缩解码模型和终端使用的压缩编码模型。接入网节点通知终端接入网节点训练该压缩编码模型所使用的训练策略，如接入网节点通知终端接入网节点采用的初始化方式为在[z min，z max]区间内随机取值作为神经元的权重、模型训练方式为监督学习的方式，损失函数为交叉熵损失函数，优化算法为随机梯度下降算法。终端获取到接入网节点的训练策略后基于该训练策略训练压缩编码模型，使得终端训练后的压缩编码模型与接入网节点训练得到的压缩编码模型尽可能的相同，从而实现终端在通信中使用的压缩编码模型与接入网节点在通信中使用的压缩解码模型相匹配。
可选地，接入网节点还可以将该智能模型的结构作为第一智能模型的结构通知终端，和/或，接入网节点还可以将向终端指示接入网节点训练该智能模型的训练数据集合的信息。
另一种实施方式中，可以由网络中的第三方节点从接入网节点获取终端的能力信息，和/或，获取接入网节点与终端之间的环境数据，基于获取到的信息训练得到相匹配的接入网节点应用的智能模型和终端应用的智能模型，并由第三方节点将该接入网节点应用的智能模型发送给接入网节点，以及通知接入网节点终端训练智能模型的训练策略，由接入网节点通过第一信息转发给终端。可选地，第三方节点还可以通过接入网节点通知终端智能模型的结构和/或训练数据集合信息。
S402，终端根据该训练策略对第一智能模型进行模型训练。
终端可以确定第一智能模型的训练策略，如模型训练方式、损失函数、模型初始化方式、模型优化算法的类型以及优化算法的参数等，训练策略中的部分或全部可以是终端基于来自接入网节点的第一信息确定的，接入网节点未指示的训练策略可以是终端基于预配置的信息或基于接入网节点指示的相关联的训练策略得到的。
可选地，终端可以获取第一智能模型的结构，基于该结构确定该第一智能模型。
一种实施方式中，终端可以接收来自接入网节点的第二信息，第二信息用于指示该第一智能模型的结构，终端根据该第二信息确定初始的第一智能模型(或者称为训练前的第一智能模型)。
具体终端设备基于第二信息得到初始的第一智能模型的实施方式可以参考前文中的描述，为了简要，在此不再赘述。
另一种实施方式中，终端从预配置信息中获取第一智能模型的结构，确定初始的第一智能模型。可选地，终端可以接收到来自接入网节点的第三信息，该第三信息用于指示训练数据集合的信息，
终端根据第三信息获取训练数据集合，并基于的该训练数据集合采用第一信息指示的训练策略对该第一智能模型进行模型训练。但本公开不限于此，终端还可以基于终端存储的数据集合对第一智能模型进行模型训练。
终端获取到第一智能模型的训练策略、训练数据集合以及基于第一智能模型的结构得到初始的第一智能模型后，终端可以开始执行第一智能模型的模型训练。
例如，终端基于训练策略确定模型训练方式为监督学习的方式，该训练数据集合中包括训练样本和标签。终端在每次训练中将训练样本输入第一智能模型，第一智能模型对训练样本的处理后输出推理结果，终端基于推理结果和标签，利用损失函数得到损失函数输出的损失值。再基于该损失值采用模型优化算法优化第一智能模型的参数，得到更新参数后的第一智能模型，将该更新参数后的第一智能模型应用于下一次模型训练中，经过多次迭代训练，当损失函数输出的损失值小于或等于第一阈值时，终端确定完成该第一智能模型的训练，得到训练后的该第一智能模型。
再例如，终端基于训练策略确定模型训练方式为无监督学习的方式，该训练数据集合中包括训练样本。终端在每次训练中将训练样本输入第一智能模型，第一智能模型对训练样本的处理后输出推理结果，终端基于推理结果，利用损失函数得到损失函数的输出值。再基于该损失函数的输出值采用模型优化算法优化第一智能模型的参数，得到更新参数后的第一智能模型，将该更新参数后的第一智能模型应用于下一次模型训练中，经过多次迭代训练，当损失函数的输出值大于或等于第二阈值时，终端确定完成该第一智能模型的训练，得到训练后的该第一智能模型。根据上述方案，网络可以通知终端智能模型的训练策略，由终端基于网络提供的训练策略执行智能模型的训练，使得终端训练得到的模型能够与网络采用的模型相匹配，达到预期的性能需求。减小了接入网节点向终端发送第一智能模型的模型参数(如每个神经元的权重值、激活函数和偏置等)带来的空口资源开销。
图5C是本公开公开提供的通信方法500的一个示意性流程图。需要说明的是，图5所示方法中与图4所示方法中相同的部分，可以参考图4所示方法中的描述，为了简要，在此不再赘述。
S501，接入网节点向终端发送第一信息，该第一信息用于指示第一智能模型的训练策略。
具体实现可参考S401，此处不再赘述。
S502，终端根据该训练策略，对第一智能模型进行训练。
终端可以获取第一智能模型的结构确定第一智能模型，利用训练数据集合，采用该训练策略对第一智能模型进行训练。具体实现可参考S402，此处不再赘述。
S503，接入网节点向终端发送第四信息，该第四信息用于指示测试信息，该测试信息用于测试该第一智能模型的性能。
相应地，终端接收来自接入网节点的第四信息，终端接收到第四信息后根据第四信息对训练后的第一智能模型进行测试。
作为示例非限定，该测试信息包括以下一项或多项：
测试数据信息、性能评估方式或性能评估参数。
可选地，当测量信息包括测试数据信息、性能评估方式和性能评估参数中的一项或两项时，例如包括测试数据信息和性能评估方式，未包括的其他项，例如性能评估参数，可以是协议约定的，或者通过其他方式确定的，不予限制。或者如下示例，终端无需获知性能评估方式和/或性能评估参数，由接入网节点进行性能评估。
其中，测试数据信息用于指示测试数据、测试数据的类型、标签数据或标签数据的类型中的一项或多项。性能评估方式可以是计算推理数据与标签数据之间的损失值的方式，性能评估参数可以是评估损失值的门限值等参数，但本公开不限于此，性能评估方式也可以是评估函数等。
S504，终端向接入网节点发送第五信息和/或第六信息，该第五信息用于指示该第一智能模型的测试结果，该第六信息用于指示推理数据。
其中，该推理数据是该第一智能模型推理测试数据得到的，第四信息指示的测试信息包括该测试数据。
一种实施方式中，终端可以向接入网节点发送第五信息，该第五信息可以指示该第一智能模型的测试结果。
例如，该测试结果可以是第一智能模型的性能达标或不达标。接入网节点向终端发送的测试信息可以包括测试数据、校验数据和校验门限值，终端可以将测试数据作为训练后的第一智能模型的输入，得到第一智能模型基于该测试数据推理得到推理数据。终端计算推理数据与校验数据之间的损失值。例如，终端可以采用训练第一智能模型时采用的损失函数计算得到推理数据与校验数据之间的损失值。但本申请不限于此。终端比较该损失值与校验门限值的大小，若损失值小于或等于校验门限值，终端可以确定训练后的第一智能模型达标，可以应用于实际通信中，并通过第五信息通知接入网节点。或者，若损失值大于校验门限值，终端可以通过第五信息通知接入网节点训练后的第一智能模型不达标。
再例如，该测试结果可以是推理数据与校验数据之间的损失值，如接入网节点通过第四信息指示测试数据和校验数据之外，第四信息还指示性能评估方式为计算推理数据与校验数据之间的损失值。终端基于第四信息指示的测试数据，采用训练后的第一智能模型推理得到推理数据，计算推理数据与标签数据之间的损失值，通过第五信息通知接入网节点该损失值，由网络基于该损失值确定终端训练后的第一智能模型的性能是否达标。如可以是接入网节点或网络中的第三方节点判断是否达标，第三方节点可以通过接入网节点的转发获取到该损失值。
另一种实施方式中，终端可以向接入网节点发送第六信息，该第六信息用于指示第一智能模型的推理数据。
例如，第四信息指示测试数据，终端采用训练后的第一智能模型对该测试数据进行推理得到该第一智能模型输出的推理数据，终端将得到的该推理数据发送给接入网节点。由网络节点(如接入网节点或网络中的第三方节点)基于推理数据，判断训练后的第一智能模型的性能是否达标。
另一种实施方式中，终端可以向接入网节点发送第五信息和第六信息。
也就是说，终端即向接入网节点发送测试结果，又向接入网节点发送推理数据，接入网节点可以结合测试结果与推理数据，确定是否在实际通信中使用终端训练后的第一智能模型。但本公开不限于此。
可选地，接入网节点向终端发送第七信息，该第七信息用于指示该第一智能模型的更新后的训练策略，和/或，用于指示该第一智能模型的更新后的结构。
相应地，终端接收来自接入网节点的该第七信息。终端基于第七信息再次对第一智能模型进行训练。
一种实施方式中，接入网节点接收到来自终端的第五信息和/或第六信息后，确定终 端训练后的第一智能模型的性能不达标，则接入网节点可以向终端发送第七信息，通知第一智能模型更新后的训练策略，和/或第一智能模型的更新后的结构。
另一种实施方式中，在终端应用该训练后的第一智能模型进行通信的情况下，接入网节点基于通信性能，确定需要更新第一智能模型，接入网节点可以向终端发送该第七信息。
例如，终端应用模型一段时间后信道环境发生变化等情况，使得第一智能模型的性能不能满足需求，接入网节点可以向终端发送该第七信息，以便终端再次对第一智能模型进行模型训练，使得再次训练后的第一智能模型适应于变化后的环境。
第七信息指示训练策略的方式可以与第一信息指示训练策略的方式相同，以及，第七信息指示模型结构的方式可以与第二信息指示模型结构的方式相同。或者，该第七信息具体可以用于指示该训练策略的至少一个变换量，和/或该第一智能模型的结构的至少一个变化量。
例如，第七信息指示了训练策略的至少一个变化量，终端接收到该第七信息后，第七信息指示的变化量，确定更新后的训练策略。比如，接入网节点调整了优化算法参数(如学习率由第一信息指示的0.1变更为0.01)，训练策略中仅优化算法参数发生变化，其他训练策略未发生变化，则第七信息可以指示第七信息指示了更新后的学习率为0.01。再比如，第一信息指示批量处理的数据量为N，即每次取N个训练数据进行梯度下降更新，接入网节点将批量处理的数据量调整为M，则接入网接地那可以通过第七信息指示批量处理的数据量调整为M，或者，第七信息可以指示批量处理的数据量增加Q，其中Q＝M-N，终端接收到第七信息后，将批量处理的数据量增加Q，即N+Q，从而每次取M个训练数据进行梯度下降更新。但本公开不限于此。
第七信息也可以指示模型结构的至少一个变化量，具体方式与指示训练策略的至少一个变化量的方式类似，可以参考上述描述实施，在此不再赘述。
根据上述方案，终端基于接入网节点指示的训练策略进行模型训练后，接入网节点可以向终端发送测试信息，终端基于测试信息对训练后的第一智能模型进行测试，终端或接入网节点可以基于测试结果和/或推理数据判断训练后的第一智能模型的性能是否达标，能够在性能达标的情况下，在实际通信中应用第一智能模型，以达到提高无线通信性能(如通信可靠性等性能)的目的。
图6是本公开公开提供的通信方法的另一个示意性流程图。需要说明的是，图6所示方法中与其他方法中相同的部分，可以相互引用或结合实施，为了简要，在此不再赘述。
S601，接入网节点向终端发送第二信息，该第二信息用于指示第一智能模型的结构，该第二信息包括第一智能模型的网络层结构信息、输入数据的维度或输出数据的维度中的一种或多种结构信息。
相应地，终端接收来自接入网节点的该第二信息。
其中，该第一智能模型的网络层结构信息可以包括但不限于以下一项或多项：
第一智能模型包含的神经网络层的层数、神经网络层的类型、神经网络层的使用方式、神经网络层之间的级联关系、神经网络层的输入数据的维度或神经网络层的输出数据的维度。
S602，终端根据该第二信息，确定第一智能模型。
终端接收到该第二信息后，可以基于第二信息指示的第一智能模型的结构，生成具有 该结构的第一智能模型。
S603，终端对所述第一智能模型进行模型训练。
终端在S602中确定第一智能模型后，对第一智能模型进行训练。
根据上述方案，接入网节点通知终端第一智能模型的结构，终端基于接入网节点的指示生成具有该结构的第一智能模型，使得终端使用的第一智能模型的结构能够满足接入网节点的需求，从而能够使用第一智能模型提高无线通信的性能。由终端执行第一智能模型的训练后可以应用于通信中。能够减小接入网节点通知第一智能模型的模型参数(如每个神经元的权重值、激活函数和偏置等)带来的空口资源开销。
终端可以采用终端与接入网节点达成共识的预定义的训练策略对第一智能模型进行训练，或者接入网节点可以通过第一信息通知终端第一智能模型的训练策略。
终端训练第一智能模型采用的训练数据集可以是终端预存储的或者从接入网节点接收到的。
以上，结合图4至图6详细说明了本公开提供的方法。以下附图说明本公开提供的通信装置和通信设备。为了实现上述本公开提供的方法中的各功能，各网元可以包括硬件结构和/或软件模块，以硬件结构、软件模块、或硬件结构加软件模块的形式来实现上述各功能。上述各功能中的某个功能以硬件结构、软件模块、还是硬件结构加软件模块的方式来执行，取决于技术方案的特定应用和设计约束条件。
图7是本公开提供的通信装置的示意性框图。如图7所示，该通信装置700可以包括收发单元720。
在一种可能的设计中，该通信装置700可对应于上文方法中的终端设备，或者配置于(或用于)终端设备中的芯片，或者其他能够实现终端设备的方法的装置、模块、电路或单元等。
应理解，该通信装置700可以包括用于执行图4、图5C、图6所示的方法中终端设备执行的方法的单元。并且，该通信装置700中的各单元和上述其他操作和/或功能分别为了实现图4、图5C、图6所示的方法的相应流程。
可选地，通信装置700还可以包括处理单元710，该处理单元710可以用于处理指令或者数据，以实现相应的操作。
还应理解，该通信装置700为配置于(或用于)终端设备中的芯片时，该通信装置700中的收发单元720可以为芯片的输入/输出接口或电路，该通信装置700中的处理单元710可以为芯片中的处理器。
可选地，通信装置700还可以包括存储单元730，该存储单元730可以用于存储指令或者数据，处理单元710可以执行该存储单元中存储的指令或者数据，以使该通信装置实现相应的操作。
应理解，该通信装置700中的收发单元720为可通过通信接口(如收发器或输入/输出接口)实现，例如可对应于图8中示出的终端设备800中的收发器810。该通信装置700中的处理单元710可通过至少一个处理器实现，例如可对应于图8中示出的终端设备800中的处理器820。该通信装置700中的处理单元710还可以通过至少一个逻辑电路实现。该通信装置700中的存储单元730可对应于图8中示出的终端设备800中的存储器。
还应理解，各单元执行上述相应步骤的具体过程在上述方法中已经详细说明，为了简洁，在此不再赘述。
在另一种可能的设计中，该通信装置700可对应于上文方法中的接入网节点，例如，或者配置于(或用于)接入网节点中的芯片，或者其他能够实现接入网节点的方法的装置、模块、电路或单元等。
应理解，该通信装置700可以包括用于执行图4、图5C、图6所示的方法中接入网节点执行的方法的单元。并且，该通信装置700中的各单元和上述其他操作和/或功能分别为了实现图4、图5C、图6所示的方法的相应流程。
可选地，通信装置700还可以包括处理单元710，该处理单元710可以用于处理指令或者数据，以实现相应的操作。
还应理解，该通信装置700为配置于(或用于)接入网节点中的芯片时，该通信装置700中的收发单元720可以为芯片的输入/输出接口或电路，该通信装置700中的处理单元710可以为芯片中的处理器。
可选地，通信装置700还可以包括存储单元730，该存储单元730可以用于存储指令或者数据，处理单元710可以执行该存储单元中存储的指令或者数据，以使该通信装置实现相应的操作。
应理解，该通信装置700为接入网节点时，该通信装置700中的收发单元720为可通过通信接口(如收发器或输入/输出接口)实现，例如可对应于图9中示出的网络设备900中的收发器910。该通信装置900中的处理单元910可通过至少一个处理器实现，例如可对应于图9中示出的网络设备900中的处理器920，该通信装置700中的处理单元710可通过至少一个逻辑电路实现。该通信装置700中的存储单元730可对应于图9中示出的网络设备900中的存储器。
还应理解，各单元执行上述相应步骤的具体过程在上述方法中已经详细说明，为了简洁，在此不再赘述。
图8是本公开提供的终端设备800的结构示意图。该终端设备800可应用于如图1所示的系统中，执行上述方法中终端设备的功能。如图所示，该终端设备800包括处理器820和收发器810。可选地，该终端设备800还包括存储器。其中，处理器820、收发器810和存储器之间可以通过内部连接通路互相通信，传递控制信号和/或数据信号。该存储器用于存储计算机程序，该处理器820用于执行该存储器中的该计算机程序，以控制该收发器810收发信号。
上述处理器820可以用于执行前面方法中描述的由终端设备内部实现的动作，而收发器810可以用于执行前面方法中描述的终端设备向网络设备发送或从网络设备接收的动作。具体请见前面方法中的描述，此处不再赘述。
可选地，上述终端设备800还可以包括电源，用于给终端设备中的各种器件或电路提供电源。
图9是本公开提供的网络设备900的结构示意图。该网络设备900可应用于如图1所示的系统中，执行上述方法中网络设备的功能。如图所示，该网络设备900包括处理器920和收发器910。可选地，该网络设备900还包括存储器。其中，处理器920、收发器910和存储器之间可以通过内部连接通路互相通信，传递控制和/或数据信号。该存储器用于存储计算机程序，该处理器920用于执行该存储器中的该计算机程序，以控制该收发器910收发信号。
上述处理器920可以用于执行前面方法中描述的由网络设备内部实现的动作，而收发 器910可以用于执行前面方法中描述的网络设备向网络设备发送或从网络设备接收的动作。具体请见前面方法中的描述，此处不再赘述。
可选地，上述网络设备900还可以包括电源，用于给网络设备中的各种器件或电路提供电源。
图8所示的终端设备和图9所示的网络设备中，处理器可以和存储器可以合成一个处理装置，处理器用于执行存储器中存储的程序代码来实现上述功能。具体实现时，该存储器也可以集成在处理器中，或者独立于处理器。该处理器可以与图7中的处理单元对应。收发器可以与图7中的收发单元对应。收发器810可以包括接收器(或称接收机、接收电路)和发射器(或称发射机、发射电路)。其中，接收器用于接收信号，发射器用于发射信号。
本公开中，处理器可以是通用处理器、数字信号处理器、专用集成电路、现场可编程门阵列或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件，可以实现或者执行本公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者任何常规的处理器等。结合本公开的方法的步骤可以直接体现为硬件处理器执行完成，或者用处理器中的硬件及软件模块组合执行完成。
本公开中，存储器可以是非易失性存储器，比如硬盘(hard disk drive，HDD)或固态硬盘(solid-state drive，SSD)等，还可以是易失性存储器(volatile memory)，例如随机存取存储器(random-access memory，RAM)。存储器是能够用于携带或存储具有指令或数据结构形式的期望的程序代码并能够由计算机存取的任何其他介质，但不限于此。本申请中的存储器还可以是电路或者其它任意能够实现存储功能的装置，用于存储程序指令和/或数据。
本公开还提供了一种处理装置，包括处理器和(通信)接口；所述处理器用于执行上述任一方法中的方法。
应理解，上述处理装置可以是一个或多个芯片。例如，该处理装置可以是现场可编程门阵列(field programmable gate array，FPGA)，可以是专用集成芯片(application specific integrated circuit，ASIC)，还可以是系统芯片(system on chip，SoC)，还可以是中央处理器(central processor unit，CPU)，还可以是网络处理器(network processor，NP)，还可以是数字信号处理电路(digital signal processor，DSP)，还可以是微控制器(micro controller unit，MCU)，还可以是可编程控制器(programmable logic device，PLD)或其他集成芯片。
根据本公开提供的方法，本公开还提供一种计算机程序产品，该计算机程序产品包括：计算机程序代码，当该计算机程序代码由一个或多个处理器执行时，使得包括该处理器的装置执行图4、图5C、图6所示中的方法。
本公开提供的技术方案可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机程序指令时，全部或部分地产生按照本发明所述的流程或功能。上述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，该计算机可读存储介质可以是计算机可以存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质(例如，软盘、硬盘、 磁带)、光介质(例如，数字视频光盘(digital video disc，DVD))、或者半导体介质等。
根据本公开提供的方法，本公开还提供一种计算机可读存储介质，该计算机可读存储介质存储有程序代码，当该程序代码由一个或多个处理器运行时，使得包括该处理器的装置执行图4、图5C、图6所示中的方法。
根据本公开提供的方法，本公开还提供一种系统，其包括前述的一个或多个第一装置。还系统还可以进一步包括前述的一个或多个第三装置。
在本公开所提供的几个中，应该理解到，所揭露的系统、装置和方法，可以通过其它的方式实现。例如，以上所描述的装置仅仅是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或单元的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本方案的目的。
以上所述，仅为本公开的具体实施方式，但本公开的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本公开揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本公开的保护范围之内。因此，本公开的保护范围应以所述权利要求的保护范围为准。

Claims (63)

  1. 一种通信方法，其特征在于，包括：
    接收第一信息，所述第一信息用于指示第一智能模型的训练策略；
    根据所述训练策略，对所述第一智能模型进行模型训练。
  2. 根据权利要求1所述的方法，其特征在于，所述训练策略包括以下一项或多项：
    模型训练方式、损失函数信息、模型初始化方式、模型优化算法的类型、或优化算法的参数。
  3. 根据权利要求2所述的方法，其特征在于，所述模型优化算法为自适应动量估计算法、随机梯度下降算法或批量梯度下降算法；和/或，
    所述优化算法的参数包括学习率、迭代次数、或批量处理的数据量中的一项或多项。
  4. 根据权利要求1至3中任一项所述的方法，其特征在于，所述方法还包括：
    获取第二信息，所述第二信息用于指示第一智能模型的结构。
  5. 根据权利要求4所述的方法，其特征在于，所述第二信息具体用于指示所述第一智能模型的以下一种或多种结构信息：
    网络层结构信息、输入数据的维度或输出数据的维度。
  6. 根据权利要求5所述的方法，其特征在于，所述网络层结构信息包括以下一项或多项：
    所述第一智能模型包含的神经网络层的层数、所述神经网络层的类型、所述神经网络层的使用方式、所述神经网络层之间的级联关系、所述神经网络层的输入数据的维度或神经网络层的输出数据的维度。
  7. 根据权利要求1至6中任一项所述的方法，其特征在于，所述方法还包括：
    接收第三信息，所述第三信息用于指示训练数据集合的信息，所述训练数据集合用于训练所述第一智能模型。
  8. 根据权利要求7所述的方法，其特征在于，所述训练数据集合包括训练样本，或者包括训练样本和标签。
  9. 根据权利要求1至8中任一项所述的方法，其特征在于，所述方法还包括：
    发送能力信息，所述能力信息用于指示运行智能模型的能力。
  10. 根据权利要求9所述的方法，其特征在于，所述能力信息具体用于指示以下一项或多项能力：
    是否支持运行智能模型、支持运行的智能模型的类型、数据处理能力或存储能力。
  11. 根据权利要求1至10中任一项所述的方法，其特征在于，所述方法还包括：
    接收第四信息，所述第四信息用于指示测试信息，所述测试信息用于测试所述第一智能模型的性能。
  12. 根据权利要求11所述的方法，其特征在于，所述测试信息包括以下一项或多项：
    测试数据信息、性能评估方式或性能评估参数。
  13. 根据权利要求11或12所述的方法，其特征在于，所述方法还包括：
    发送第五信息，所述第五信息用于指示所述第一智能模型的测试结果；和/或，
    发送第六信息，所述第六信息用于指示推理数据，所述推理数据是所述第一智能模型推理测试数据得到的，所述测试信息包括所述测试数据。
  14. 根据权利要求13所述的方法，其特征在于，所述方法还包括：
    接收第七信息，所述第七信息用于指示所述第一智能模型的更新后的训练策略，和/或，用于指示所述第一智能模型的更新后的结构。
  15. 根据权利要求14所述的方法，其特征在于，所述第七信息具体用于指示所述训练策略的至少一个变换量，和/或所述第一智能模型的结构的至少一个变化量。
  16. 一种通信方法，其特征在于，包括：
    发送第一信息，所述第一信息用于指示第一智能模型的训练策略。
  17. 根据权利要求16所述的方法，其特征在于，所述训练策略包括以下一项或多项：
    模型训练方式、损失函数信息、模型初始化方式、模型优化算法的类型、或优化算法的参数。
  18. 根据权利要求17所述的方法，其特征在于，所述模型优化算法的类型为自适应动量估计算法、随机梯度下降算法或批量梯度下降算法；和/或，
    所述优化算法的参数包括学习率、迭代次数、或批量处理的数据量中的一项或多项。
  19. 根据权利要求16至18中任一项所述的方法，其特征在于，所述方法还包括：
    发送第二信息，所述第二信息用于指示第一智能模型的结构。
  20. 根据权利要求19所述的方法，其特征在于，所述第二信息具体用于指示所述第一智能模型的以下一种或多种结构信息：
    网络层结构信息、输入数据的维度、或输出数据的维度。
  21. 根据权利要求20所述的方法，其特征在于，所述网络层结构信息包括以下一项或多项：
    所述第一智能模型包含的神经网络层的层数、所述神经网络层的类型、所述神经网络层的使用方式、所述神经网络层之间的级联关系、所述神经网络层的输入数据的维度或所述神经网络层的输出数据的维度。
  22. 根据权利要求16至21中任一项所述的方法，其特征在于，所述方法还包括：
    发送第三信息，所述第三信息用于指示用于训练数据集合的信息，所述训练数据集合用于训练所述第一智能模型。
  23. 根据权利要求22所述的方法，其特征在于，所述训练数据集合包括训练样本，或者包括训练样本和标签。
  24. 根据权利要求16至23中任一项所述的方法，其特征在于，所述方法还包括：
    接收能力信息，所述能力信息用于指示运行智能模型的能力。
  25. 根据权利要求24所述的方法，其特征在于，所述能力信息用于指示以下一项或多项能力：
    是否支持运行智能模型、支持运行的智能模型的类型、数据处理能力或存储能力。
  26. 根据权利要求16至25中任一项所述的方法，其特征在于，所述方法还包括：
    发送第四信息，所述第四信息用于指示测试所述第一智能模型性能的测试信息。
  27. 根据权利要求26所述的方法，其特征在于，所述测试信息包括以下一项或多项：
    测试数据信息、性能评估方式或性能评估参数。
  28. 根据权利要求26或27所述的方法，其特征在于，所述方法还包括：
    接收第五信息，所述第五信息用于指示所述第一智能模型的测试结果；和/或，
    接收第六信息，所述第六信息用于指示推理数据，所述推理数据是所述第一智能模型 推理测试数据得到的，所述测试信息包括所述测试数据。
  29. 根据权利要求28所述的方法，其特征在于，所述方法还包括：
    发送第七信息，所述第七信息用于指示所述第一智能模型的更新后的训练策略，和/或，用于指示所述第一智能模型的更新后的结构。
  30. 根据权利要求29所述的方法，其特征在于，所述第七信息具体用于指示所述训练策略的至少一个参数变换量，和/或所述第一智能模型的结构的至少一个参数变化量。
  31. 一种通信方法，其特征在于，包括：
    接收第二信息，所述第二信息用于指示第一智能模型的结构，所述第二信息包括所述第一智能模型的网络层结构信息、输入数据的维度或输出数据的维度中的一种或多种结构信息；
    根据所述第二信息，确定所述第一智能模型的结构；
    对所述第一智能模型进行模型训练。
  32. 根据权利要求31所述的方法，其特征在于，所述方法还包括：
    获取第一信息，所述第一信息用于指示第一智能模型的训练策略；
    以及，所述对所述第一智能模型进行模型训练，包括：
    根据所述训练策略，对所述第一智能模型进行模型训练。
  33. 根据权利要求32所述的方法，其特征在于，所述训练策略包括以下一项或多项：
    模型训练方式、损失函数信息、模型初始化方式、模型优化算法的类型、或优化算法的参数。
  34. 根据权利要求31至33中任一项所述的方法，其特征在于，所述方法还包括：
    接收第三信息，所述第三信息用于指示训练数据集合的信息，所述训练数据集合用于训练所述第一智能模型。
  35. 根据权利要求34所述的方法，其特征在于，所述训练数据集合包括训练样本，或者包括训练样本和标签。
  36. 根据权利要求31至35中任一项所述的方法，其特征在于，所述方法还包括：
    发送能力信息，所述能力信息用于指示运行智能模型的能力。
  37. 根据权利要求36所述的方法，其特征在于，所述能力信息用于指示以下一项或多项能力：
    是否支持运行智能模型、支持运行的智能模型的类型、数据处理能力或存储能力。
  38. 根据权利要求31至37中任一项所述的方法，其特征在于，所述方法还包括：
    接收第四信息，所述第四信息用于指示测试信息，所述测试信息用于测试所述第一智能模型的性能。
  39. 根据权利要求38所述的方法，其特征在于，所述测试信息包括以下一项或多项：
    测试数据信息、性能评估方式或性能评估参数。
  40. 根据权利要求38或39所述的方法，其特征在于，所述方法还包括：
    发送第五信息，所述第五信息用于指示所述第一智能模型的测试结果；和/或，
    发送第六信息，所述第六信息用于指示推理数据，所述推理数据是所述第一智能模型推理测试数据得到的，所述测试信息包括所述测试数据。
  41. 根据权利要求31至40中任一项所述的方法，其特征在于，所述方法还包括：
    接收第七信息，所述第七信息用于指示所述第一智能模型的更新后的训练策略，和/ 或，用于指示所述第一智能模型的更新后的结构。
  42. 根据权利要求41所述的方法，其特征在于，所述第七信息具体用于指示所述训练策略的至少一个变换量，和/或所述第一智能模型的结构的至少一个变化量。
  43. 一种通信方法，其特征在于，包括：
    发送第二信息，所述第二信息用于指示第一智能模型的结构，所述第二信息包括所述第一智能模型的网络层结构信息、输入数据的维度或输出数据的维度中的一种或多种结构信息。
  44. 根据权利要求43所述的方法，其特征在于，所述方法还包括：
    发送第一信息，所述第一信息用于指示第一智能模型的训练策略。
  45. 根据权利要求44所述的方法，其特征在于，所述训练策略包括以下一项或多项：
    模型训练方式、损失函数信息、模型初始化方式、模型优化算法的类型、或优化算法的参数。
  46. 根据权利要求43至45中任一项所述的方法，其特征在于，所述方法还包括：
    发送第三信息，所述第三信息用于指示训练数据集合的信息，所述训练数据集合用于训练所述第一智能模型。
  47. 根据权利要求46所述的方法，其特征在于，所述训练数据集合包括训练样本，或者包括训练样本和标签。
  48. 根据权利要求43至47中任一项所述的方法，其特征在于，所述方法还包括：
    接收能力信息，所述能力信息用于指示运行智能模型的能力。
  49. 根据权利要求43所述的方法，其特征在于，所述能力信息用于指示以下一项或多项能力：
    是否支持运行智能模型、支持运行的智能模型的类型、数据处理能力或存储能力。
  50. 根据权利要求43至49中任一项所述的方法，其特征在于，所述方法还包括：
    发送第四信息，所述第四信息用于指示测试信息，所述测试信息用于测试所述第一智能模型的性能。
  51. 根据权利要求50所述的方法，其特征在于，所述测试信息包括以下一项或多项：
    测试数据信息、性能评估方式或性能评估参数。
  52. 根据权利要求50或51所述的方法，其特征在于，所述方法还包括：
    接收第五信息，所述第五信息用于指示所述第一智能模型的测试结果；和/或，
    接收第六信息，所述第六信息用于指示推理数据，所述推理数据是所述第一智能模型推理测试数据得到的，所述测试信息包括所述测试数据。
  53. 根据权利要求43至52中任一项所述的方法，其特征在于，所述方法还包括：
    发送第七信息，所述第七信息用于指示所述第一智能模型的更新后的训练策略，和/或，用于指示所述第一智能模型的更新后的结构。
  54. 根据权利要求53所述的方法，其特征在于，所述第七信息具体用于指示所述训练策略的至少一个变换量，和/或所述第一智能模型的结构的至少一个变化量。
  55. 一种通信装置，其特征在于，用于实现如权利要求1至15、31至42中任一项所述的方法。
  56. 一种通信装置，其特征在于，包括处理器和存储器，所述存储器和所述处理器耦合，所述处理器用于执行权利要求1至15中、31至42中任一项所述的方法。
  57. 一种通信装置，其特征在于，包括处理器和通信接口，所述处理器利用所述通信接口，执行权利要求1至15中、31至42中任一项所述的方法。
  58. 一种通信装置，其特征在于，用于实现如权利要求16至30、43至54中任一项所述的方法。
  59. 一种通信装置，其特征在于，包括处理器和存储器，所述存储器和所述处理器耦合，所述处理器用于执行权利要求16至30、43至54中任一项所述的方法。
  60. 一种通信装置，其特征在于，包括处理器和通信接口，所述处理器利用所述通信接口，执行权利要求16至30、43至54中任一项所述的方法。
  61. 一种通信系统，其特征在于，包括权利要求55至57中任一项所述的通信装置，和权利要求58至60中任一项所述的通信装置。
  62. 一种计算机可读存储介质，其特征在于，存储有指令，当所述指令在计算机上运行时，使得所述计算机执行如权利要求1至54中任一项所述的方法。
  63. 一种计算机程序产品，其特征在于，包括指令，当所述指令在计算机上运行时，使得计算机执行如权利要求1至54中任一项所述的方法。
PCT/CN2022/136138 2021-12-02 2022-12-02 通信方法和通信装置 WO2023098860A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
KR1020247021743A KR20240117116A (ko) 2021-12-02 2022-12-02 통신 방법 및 통신 장치
EP22900664.8A EP4429309A1 (en) 2021-12-02 2022-12-02 Communication method and communication apparatus
US18/679,849 US20240320488A1 (en) 2021-12-02 2024-05-31 Communication method and communication apparatus

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111462667.X 2021-12-02
CN202111462667.XA CN116233857A (zh) 2021-12-02 2021-12-02 通信方法和通信装置

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/679,849 Continuation US20240320488A1 (en) 2021-12-02 2024-05-31 Communication method and communication apparatus

Publications (1)

Publication Number Publication Date
WO2023098860A1 true WO2023098860A1 (zh) 2023-06-08

Family

ID=86575491

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/136138 WO2023098860A1 (zh) 2021-12-02 2022-12-02 通信方法和通信装置

Country Status (5)

Country Link
US (1) US20240320488A1 (zh)
EP (1) EP4429309A1 (zh)
KR (1) KR20240117116A (zh)
CN (1) CN116233857A (zh)
WO (1) WO2023098860A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117812604B (zh) * 2024-02-28 2024-07-09 荣耀终端有限公司 通信方法、系统及相关设备

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111612153A (zh) * 2019-02-22 2020-09-01 华为技术有限公司 训练模型的方法和装置
CN112101551A (zh) * 2020-09-25 2020-12-18 北京百度网讯科技有限公司 用于训练模型的方法、装置、设备以及存储介质
CN113128686A (zh) * 2020-01-16 2021-07-16 华为技术有限公司 模型训练方法及装置
CN113412494A (zh) * 2019-02-27 2021-09-17 华为技术有限公司 一种确定传输策略的方法及装置
CN113938232A (zh) * 2020-07-13 2022-01-14 华为技术有限公司 通信的方法及通信装置
CN114036994A (zh) * 2020-07-21 2022-02-11 中兴通讯股份有限公司 训练通信决策模型的方法、电子设备、计算机可读介质
CN114143799A (zh) * 2020-09-03 2022-03-04 华为技术有限公司 通信方法及装置
CN114679355A (zh) * 2020-12-24 2022-06-28 华为技术有限公司 通信方法和装置
CN114793453A (zh) * 2020-11-23 2022-07-26 北京小米移动软件有限公司 一种训练方法、训练装置及存储介质
CN114936117A (zh) * 2021-09-02 2022-08-23 华为技术有限公司 模型训练的方法、服务器、芯片以及系统

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111612153A (zh) * 2019-02-22 2020-09-01 华为技术有限公司 训练模型的方法和装置
CN113412494A (zh) * 2019-02-27 2021-09-17 华为技术有限公司 一种确定传输策略的方法及装置
CN113128686A (zh) * 2020-01-16 2021-07-16 华为技术有限公司 模型训练方法及装置
CN113938232A (zh) * 2020-07-13 2022-01-14 华为技术有限公司 通信的方法及通信装置
CN114036994A (zh) * 2020-07-21 2022-02-11 中兴通讯股份有限公司 训练通信决策模型的方法、电子设备、计算机可读介质
CN114143799A (zh) * 2020-09-03 2022-03-04 华为技术有限公司 通信方法及装置
CN112101551A (zh) * 2020-09-25 2020-12-18 北京百度网讯科技有限公司 用于训练模型的方法、装置、设备以及存储介质
CN114793453A (zh) * 2020-11-23 2022-07-26 北京小米移动软件有限公司 一种训练方法、训练装置及存储介质
CN114679355A (zh) * 2020-12-24 2022-06-28 华为技术有限公司 通信方法和装置
CN114936117A (zh) * 2021-09-02 2022-08-23 华为技术有限公司 模型训练的方法、服务器、芯片以及系统

Also Published As

Publication number Publication date
EP4429309A1 (en) 2024-09-11
US20240320488A1 (en) 2024-09-26
CN116233857A (zh) 2023-06-06
KR20240117116A (ko) 2024-07-31

Similar Documents

Publication Publication Date Title
Yang et al. Federated learning via over-the-air computation
US20240320488A1 (en) Communication method and communication apparatus
WO2023036268A1 (zh) 一种通信方法及装置
WO2023040971A1 (zh) 一种人工智能模型下载方法、装置及系统
US20240211810A1 (en) Model test method and apparatus
WO2024164641A1 (zh) 一种csi发送和接收方法及装置
WO2024051789A1 (zh) 一种波束管理方法
WO2024017001A1 (zh) 模型的训练方法及通信装置
WO2023126007A1 (zh) 信道信息传输方法及装置
WO2023175381A1 (en) Iterative training of collaborative distributed coded artificial intelligence model
WO2023125996A1 (zh) 一种上行预编码方法及装置
WO2023093725A1 (zh) 一种校准方法及装置
WO2023093777A1 (zh) 一种通信方法及装置
WO2024169757A1 (zh) 一种通信的方法和通信装置
WO2023179458A1 (zh) 通信方法和通信装置
US20240356595A1 (en) Uplink precoding method and apparatus
WO2023231934A1 (zh) 一种通信方法及装置
WO2024208316A1 (zh) 一种压缩信息的反馈方法及装置
WO2023231881A1 (zh) 一种模型应用方法及装置
WO2024169745A1 (zh) 一种通信方法、装置及系统
WO2024046419A1 (zh) 一种通信方法及装置
WO2023066346A1 (zh) 一种通信方法及装置
WO2023125985A1 (zh) 模型的数据处理方法及装置
US20230259742A1 (en) Communication method, apparatus, and system
WO2024131900A1 (zh) 一种通信的方法和通信装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22900664

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2024532663

Country of ref document: JP

Kind code of ref document: A

WWE Wipo information: entry into national phase

Ref document number: 2022900664

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2022900664

Country of ref document: EP

Effective date: 20240606

ENP Entry into the national phase

Ref document number: 20247021743

Country of ref document: KR

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE