COMMUNICATION METHOD AND COMMUNICATION DEVICE
By providing a training policy to terminals for intelligent model training, the method addresses the challenge of network resource overhead in AI-integrated wireless communication, ensuring efficient and aligned model deployment without excessive air interface usage.
Patent Information
- Application Number
- JP2024532663
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-12-02
- Filing Date
- 2022-12-02
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2042-12-02
AI Technical Summary
The integration of artificial intelligence in wireless communication networks poses challenges due to complex network requirements such as ultra-high speeds, low latency, and high-volume connections, leading to increased complexity in network planning and resource management, with existing methods incurring significant air interface resource overhead when deploying intelligent models.
A communication method where an access network node provides a training policy to a terminal for training an intelligent model, reducing the need to transmit model parameters by allowing the terminal to train a model that aligns with network requirements, using techniques like supervised, unsupervised, or reinforcement learning, along with specific configurations and datasets, thereby minimizing resource overhead.
This approach enables the terminal to train an intelligent model consistent with network standards, reducing air interface resource usage and maintaining communication efficiency while meeting performance requirements.
Smart Images

Figure 2024542716000001_ABST
Abstract
Description
[Technical field]
[0001] The present disclosure relates to the field of communications, and more particularly, to a communication method and device. [Background technology]
[0002] This application claims priority to Chinese Patent Application No. 202111462667.X, filed with the China State Intellectual Property Office on December 2, 2021, and entitled “COMMUNICATION METHOD AND COMMUNICATION APPARATUS,” which is incorporated herein by reference in its entirety.
[0003] In wireless communication networks (e.g., mobile communication networks, etc.), the services supported by the network are becoming more and more diverse, and therefore the requirements that need to be met are also becoming more and more diverse. For example, the network needs to be capable of supporting ultra-high speed, ultra-low latency, and / or a large number of connections, which makes network planning, network configuration, and / or resource scheduling more and more complex. Furthermore, as networks have more and more powerful capabilities, e.g., supporting higher and higher spectrum, and new technologies such as high-order multiple-input multiple-output (MIMO) technology, beamforming, and / or beam management, energy saving in the network has become a hot research topic. These new requirements, scenarios, and capabilities bring unprecedented challenges to network planning, operation and maintenance, and efficient operation. To address such challenges, artificial intelligence techniques may be introduced into wireless communication networks to implement network intelligence. In light of this, how to effectively implement artificial intelligence in networks is a topic worthy of research. Summary of the Invention
[0004] The present disclosure provides a communication method and a communication apparatus to obtain an intelligent model that meets communication performance requirements while reducing the overhead of air interface resources.
[0005] According to a first aspect, there is provided a communication method, the method being capable of being executed by a terminal or a module (such as a chip) located in (or used in) the terminal.
[0006] The method includes: receiving first information, where the first information indicates a training policy for a first intelligent model, and performing model training on the first intelligent model according to the training policy.
[0007] Based on the above solution, the network can inform the terminal of the training policy of the intelligent model, and the terminal performs training on the intelligent model according to the training policy provided by the network, which reduces the air interface resource overhead generated when the access network node transmits the model parameters of the intelligent model (such as the weight, activation function, and offset of each neuron) to the terminal, so that the model obtained through the training by the terminal is consistent with the model used by the network, and can meet the performance requirements of actual communication while minimizing the impact on the transmission efficiency of service data.
[0008] In relation to the first aspect, in some implementations of the first aspect, the training policy 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.
[0009] Optionally, the model training approach may be one of, but is not limited to, supervised learning, unsupervised learning, or reinforcement learning.
[0010] Optionally, the loss function information may indicate a loss function used to train the first intelligent model, which may be, but is not limited to, a cross-entropy loss function or a quadratic loss function. Alternatively, the loss function may be a machine learning intelligent model.
[0011] Optionally, the model initialization technique may be an initialization technique for weights of each neuron of the first intelligent model. In one example, the first information may indicate that the model initialization technique is that the initial weights of each neuron are random values within a preset range. Optionally, the first information includes a start value and an end value within the preset range. In another example, the first information may indicate that the initial weights of the neurons are 0. In yet another example, the first information may indicate that the initial weights of the neurons are generated by using a preset function. In yet another example, the first information may indicate an identification of one of a plurality of predefined initialization techniques.
[0012] Optionally, the type of model optimization algorithm may be an adaptive momentum estimation (ADAM) algorithm, a stochastic gradient descent (SGD) algorithm, or a batch gradient descent (BGD) algorithm.
[0013] Optionally, the parameters of the optimization algorithm may include, but are not limited to, one or more of the following: The learning rate, number of iterations, or amount of data processed into a batch.
[0014] In relation to the first aspect, in some implementations of the first aspect, the method further includes: obtaining second information, the second information indicating a configuration of the first intelligent model.
[0015] Based on the above solution, the access network node informs the terminal of the configuration of the first intelligent model, and the terminal generates an initial first intelligent model having the configuration according to the instruction of the access network node, and trains the first intelligent model according to the instruction training policy, so that the first intelligent model obtained through the training by the terminal is consistent with the model used by the network equipment and can meet the performance requirements of the access network node. In this way, the first intelligent model can be used to improve wireless communication performance.
[0016] In relation to the first aspect, in some implementations of the first aspect, the second information specifically indicates one or more of the following configuration information of the first intelligent model: Network layer configuration information, input data dimensions, or output data dimensions It is.
[0017] Based on the above-mentioned solution, the access network node may inform the terminal of one or more of the configuration information of the network layer (such as the number and type of neural network layers), the dimension of the input data, and the dimension of the output data by using the second information. The information amount of the configuration information is much less than the information amount of the weight, activation function, offset, and the like of each neuron in the neural network, so that the resource usage can be reduced, and the terminal obtains the first intelligent model that meets the performance requirement through training.
[0018] In relation to the first aspect, in some implementations of the first aspect, the network layer configuration information includes one or more of the following: The number of neural network layers included in the first intelligent model, the type of the neural network layers, the method of using the neural network layers, the cascade relationship between the neural network layers, the dimension of the input data of the neural network layers, or the dimension of the output data of the neural network layers. It is.
[0019] In relation to the first aspect, in some implementations of the first aspect, the method further includes receiving third information, the third information indicating information about a training dataset, the training dataset being used to train the first intelligent model.
[0020] Optionally, the training dataset may include training samples, or may include training samples and labels.
[0021] Optionally, the information about the training dataset may indicate a type of training sample, or a type of training sample and a type of label. Optionally, the information about the training dataset may indicate use of the first intelligent model. Optionally, the third information may include the training dataset.
[0022] Based on the above-mentioned solution, the access network node can further inform the terminal of the training data set used by the terminal to train the first intelligent model. In this way, the first intelligent model obtained through training by the terminal can be closer to the requirements of the access network node, in other words, can be better matched with the intelligent model of the access network node.
[0023] In relation to the first aspect, in some implementations of the first aspect, the method further includes: transmitting capability information, the capability information indicating a capability for running the intelligent model.
[0024] Based on the above-mentioned solution, the terminal may indicate to the access network node the terminal's capability for running an intelligent model by using the capability information, so that the access network node may indicate to the terminal, based on the terminal's capabilities, a training policy and / or model configuration that meets the terminal's capability requirements.
[0025] In relation to the first aspect, in some implementations of the first aspect, the capability information specifically indicates one or more of the following capabilities: whether or not the system supports running an intelligent model, the type of intelligent model that can be run, data processing, or data storage capabilities. It is.
[0026] In relation to the first aspect, in some implementations of the first aspect, the method further includes: receiving fourth information, the fourth information indicating test information, the test information being used to test performance of the first intelligent model.
[0027] Based on the above solution, the access network node may further send test information to the terminal, so that the terminal can test the first intelligent model based on the test information.
[0028] In relation to the first aspect, in some implementations of the first aspect, the test information includes one or more of the following: Test data information, performance evaluation method, or performance evaluation parameters It is.
[0029] In relation to the first aspect, in some implementations of the first aspect, the method further includes: transmitting fifth information, the fifth information indicating a test result of the first intelligent model.
[0030] Based on the above-mentioned solution, after testing the first intelligent model based on the testing information, the terminal judges whether the trained first intelligent model meets the performance requirements, and notifies the access network node of the testing result by using the fifth information.
[0031] In relation to the first aspect, in some implementations of the first aspect, the method further includes: transmitting sixth information, the sixth information indicating inference data, the inference data being obtained through inference based on the test data by the first intelligent model, and the test information including the test data.
[0032] Based on the above-mentioned solution, the terminal can infer the test data to obtain inference data, and send the inference data to the access network node by using the sixth information, so that the access network node can determine whether the first intelligent model trained by the terminal meets the performance requirements based on the inference data.
[0033] In relation to the first aspect, in some implementations of the first aspect, the method further includes receiving seventh information, where the seventh information indicates an updated training policy of the first intelligent model and / or indicates an updated configuration of the first intelligent model.
[0034] Based on the above-mentioned solution, if the access network node determines that the first intelligent model trained by the terminal does not meet a performance requirement, the access network node may notify the terminal of an updated training policy and / or an updated model configuration by using the seventh information, and the terminal may re-train the first intelligent model based on the seventh information.
[0035] In relation to the first aspect, in some implementations of the first aspect, the seventh information specifically indicates at least one variation in the training policy and / or at least one variation in the configuration of the first intelligent model.
[0036] Based on the above-mentioned solution, the seventh information may specifically indicate at least one variation in the training policy and / or at least one variation in the model configuration, thereby reducing information overhead.
[0037] According to a second aspect, there is provided a communication method, the method being capable of being performed by an access network node or a module (e.g. a chip, etc.) located in (or used in) an access network node.
[0038] The method includes: transmitting first information, the first information indicating a training policy for a first intelligent model.
[0039] The training policy is the same as the training policy described in the first embodiment, and for the sake of brevity, the details will not be described again here.
[0040] In relation to the second aspect, in some implementations of the second aspect, the method further includes: transmitting second information, the second information indicating a configuration of the first intelligent model.
[0041] The configuration information specifically indicated by the second information is the same as that described in the first aspect, and for the sake of brevity, the details will not be described again herein.
[0042] In relation to the second aspect, in some implementations of the second aspect, the method further includes: transmitting third information, the third information indicating information about a training dataset, the training dataset being used to train the first intelligent model.
[0043] In relation to the second aspect, in some implementations of the second aspect, the method further includes: receiving capability information, the capability information indicating a capability to run the intelligent model.
[0044] According to the method, a training policy and / or configuration of the first intelligent model may be determined based on the capability information.
[0045] The capabilities specifically indicated by the capability information are the same as those described in the first embodiment, and for the sake of brevity, the details will not be described again herein.
[0046] In relation to the second aspect, in some implementations of the second aspect, determining a training policy and / or a configuration of the first intelligent model based on the capability information further includes: determining a configuration of the first intelligent model based on the capability information; and / or training the first intelligent model according to the training policy.
[0047] In relation to the second aspect, in some implementations of the second aspect, the method further includes: Transmitting fourth information, the fourth information indicating test information for testing performance of the first intelligent model.
[0048] The content contained in the test information is the same as that described in the first aspect, and for the sake of brevity, the details will not be described again here.
[0049] In relation to the second aspect, in some implementations of the second aspect, the method further includes: receiving fifth information, the fifth information indicating a test result of the first intelligent model; and / or A step of receiving sixth information, the sixth information indicating inference data, the inference data being obtained through inference based on the test data by the first intelligent model, and the test information including the test data.
[0050] In relation to the second aspect, in some implementations of the second aspect, the seventh information specifically indicates a variation of at least one parameter in the training policy and / or a variation of at least one parameter in the configuration of the first intelligent model.
[0051] According to a third aspect, a communication method is provided. The method can be performed by a terminal or a module (e.g., a chip, etc.) disposed in (or used in) the terminal. The following will use an example in which the method is performed by the terminal to describe the method.
[0052] The method includes: receiving second information, where the second information indicates a configuration of a first intelligent model, and the second information includes one or more of the following configuration information of the first intelligent model: configuration information of a network layer, a dimension of input data, or a dimension of output data; determining a configuration of the first intelligent model based on the second information; and performing model training on the first intelligent model.
[0053] The configuration information of the network layer is the same as that described in the first embodiment, and for the sake of brevity, the details will not be described again here.
[0054] In relation to the third aspect, in some implementations of the third aspect, the method further includes: obtaining first information, the first information indicating a training policy for the first intelligent model, and performing model training on the first intelligent model includes performing model training on the first intelligent model according to the training policy.
[0055] The training policy is the same as that described in the first embodiment, and for the sake of brevity, the details will not be described again here.
[0056] In relation to the third aspect, in some implementations of the third aspect, the method further includes receiving third information, the third information indicating information regarding a training dataset, the training dataset being used to train the first intelligent model.
[0057] In relation to the third aspect, in some implementations of the third aspect, the method further includes: transmitting capability information, the capability information indicating a capability to run the intelligent model.
[0058] The capability information is the same as that described in the first embodiment, and for the sake of brevity, the details will not be described again here.
[0059] In relation to the third aspect, in some implementations of the third aspect, the method further includes receiving fourth information, the fourth information indicating test information, the test information being used to test performance of the first intelligent model.
[0060] The test information is the same as that described in the first embodiment, and for the sake of brevity, the details will not be repeated here.
[0061] In relation to the third aspect, in some implementations of the third aspect, the method further includes: transmitting fifth information, where the fifth information indicates a test result of the first intelligent model; and / or transmitting sixth information, where the sixth information indicates inference data, where the inference data is obtained through inference based on the test data by the first intelligent model, where the test information includes the test data.
[0062] In relation to the third aspect, in some implementations of the third aspect, the method further includes receiving seventh information, the seventh information indicating an updated training policy of the first intelligent model and / or indicating an updated configuration of the first intelligent model.
[0063] In relation to the third aspect, in some implementations of the third aspect, the seventh information specifically indicates at least one variation in the training policy and / or at least one variation in the configuration of the first intelligent model.
[0064] According to a fourth aspect, there is provided a communication method, the method being capable of being performed by an access network node or a module (e.g. a chip, etc.) located in (or used in) the access network node.
[0065] The method includes: sending second information, where the second information indicates a configuration of the first intelligent model, and the second information includes one or more of the following configuration information of the first intelligent model: configuration information of a network layer, a dimension of input data, or a dimension of output data.
[0066] The configuration information of the network layer is the same as that described in the first embodiment, and for the sake of brevity, the details will not be described again here.
[0067] In relation to the fourth aspect, in some implementations of the fourth aspect, the method further includes: transmitting first information, the first information indicating a training policy for the first intelligent model.
[0068] The training policy is the same as that described in the first embodiment, and for the sake of brevity, the details will not be described again here.
[0069] In relation to the fourth aspect, in some implementations of the fourth aspect, the method further includes: transmitting third information, the third information indicating information about a training dataset, the training dataset being used to train the first intelligent model.
[0070] In relation to the fourth aspect, in some implementations of the fourth aspect, the method further includes: receiving capability information, the capability information indicating a capability for running the intelligent model.
[0071] The capability information is the same as that described in the first embodiment, and for the sake of brevity, the details will not be described again here.
[0072] In relation to the fourth aspect, in some implementations of the fourth aspect, the method further includes: transmitting fourth information, the fourth information indicating test information, the test information being used to test performance of the first intelligent model.
[0073] The test information is the same as that described in the first embodiment, and for the sake of brevity, the details will not be repeated here.
[0074] In relation to the fourth aspect, in some implementations of the fourth aspect, the method further includes receiving fifth information, the fifth information indicating a test result of the first intelligent model, and / or receiving sixth information, the sixth information indicating inference data, the inference data being obtained through inference based on the test data by the first intelligent model, and the test information including the test data.
[0075] In relation to the fourth aspect, in some implementations of the fourth aspect, the method further includes: transmitting seventh information, where the seventh information indicates an updated training policy of the first intelligent model and / or indicates an updated configuration of the first intelligent model.
[0076] In relation to the fourth aspect, in some implementations of the fourth aspect, the seventh information specifically indicates at least one variation in the training policy and / or at least one variation in the configuration of the first intelligent model.
[0077] According to a fifth aspect, a communication apparatus is provided. In design, the apparatus may include corresponding modules for performing the methods / operations / procedures / actions described in the first aspect. These modules may be hardware circuits, software, or may be implemented by using a combination of hardware circuits and software. In design, the apparatus includes: a transceiver unit configured to receive first information, the first information indicating a training policy of a first intelligent model; and a processing unit configured to perform model training on the first intelligent model according to the training policy.
[0078] According to a sixth aspect, a communication device is provided. In design, the device may include corresponding modules for performing the methods / operations / procedures / actions described in the second aspect. These modules may be hardware circuits, software, or may be implemented by using a combination of hardware circuits and software. In design, the device includes: a processing unit configured to determine a training policy of a first intelligent model; and a transceiver unit configured to transmit first information, the first information indicating the training policy of the first intelligent model.
[0079] According to a seventh aspect, a communication device is provided. In design, the device may include corresponding modules for performing the methods / operations / procedures / actions described in the third aspect. These modules may be hardware circuits, software, or may be implemented by using a combination of hardware circuits and software. In design, the device includes: a transceiver unit configured to receive second information, the second information indicating a configuration of a first intelligent model, the second information including one or more of the following configuration information of the first intelligent model: configuration information of a network layer, a dimension of input data, or a dimension of output data; and a processing unit configured to determine a configuration of the first intelligent model based on the second information. The processing unit is further configured to perform model training on the first intelligent model.
[0080] According to an eighth aspect, there is provided a communication apparatus. In design, the apparatus may include corresponding modules for performing the methods / operations / procedures / actions described in the fourth aspect. These modules may be hardware circuits, software, or may be implemented by using a combination of hardware circuits and software. In design, the apparatus includes: a processing unit configured to determine a configuration of a first intelligent model; and a transceiver unit configured to receive second information, the second information indicating a configuration of the first intelligent model, the second information including one or more of the following configuration information of the first intelligent model: network layer configuration information, input data dimension, or output data dimension.
[0081] According to a ninth aspect, a communication device is provided. The device includes a processor. The processor may implement the method according to any one of the first aspect or the third aspect and possible implementations in the first aspect or the third aspect. Optionally, the communication device further includes a memory. The processor may be coupled to the memory and configured to execute instructions in the memory to implement the method according to any one of the first aspect or the third aspect and possible implementations in the first aspect or the third aspect. Optionally, the communication device further includes a communication interface. The processor is coupled to the communication interface. In the present disclosure, the communication interface may be a transceiver, a pin, a circuit, a bus, a module, or another type of communication interface. This is not limited.
[0082] In an implementation, the communications device is a terminal. If the communications device is a terminal, the communications interface may be a transceiver or an input / output interface.
[0083] In another implementation, the communication device is a chip located in a terminal. When the communication device is a chip located in a terminal, the communication interface may be an input / output interface.
[0084] Optionally, the transceiver may be a transceiver circuit. Optionally, the input / output interface may be an input / output circuit.
[0085] According to a tenth aspect, there is provided a communication device. The device includes a processor. The processor may implement a method according to any one of the second aspect or the fourth aspect and possible implementations in the second aspect or the fourth aspect. Optionally, the communication device further includes a memory. The processor may be coupled to the memory and configured to execute instructions in the memory to implement a method according to any one of the second aspect or the fourth aspect and possible implementations in the second aspect or the fourth aspect. Optionally, the communication device further includes a communication interface. The processor is coupled to the communication interface.
[0086] In an implementation, the communication device is an access network node. When the communication device is an access network node, the communication interface may be a transceiver or an input / output interface.
[0087] In another implementation, the communication device is a chip located in an access network node. If the communication device is a chip located in a first access network node, the communication interface may be an input / output interface.
[0088] Optionally, the transceiver may be a transceiver circuit. Optionally, the input / output interface may be an input / output circuit.
[0089] According to an eleventh aspect, there is provided a processor, the processor including an input circuit, an output circuit, and a processing circuit, the processing circuit configured to receive signals via the input circuit and send signals via the output circuit to enable the processor to perform a method according to any one of the first to fourth aspects and possible implementations thereof.
[0090] In a particular implementation, the processor may be one or more chips, the input circuits may be input pins, the output circuits may be output pins, and the processing circuits may be transistors, gate circuits, triggers, any logic circuits, or the like. The input signals received by the input circuits may be, for example, but not limited to, received and input by a receiver. The signals output by the output circuits may be, for example, but not limited to, output to a transmitter and transmitted by the transmitter. Also, the input circuits and the output circuits may be the same circuit, which is used as an input circuit and an output circuit at different times. The particular implementation of the processor and the various circuits is not limited by this disclosure.
[0091] According to a twelfth aspect, there is provided a computer program product, the computer program product including a computer program (sometimes called code or instructions) which, when executed, enables a computer to carry out a method according to any one of the first to fourth aspects and possible implementations of the first to fourth aspects.
[0092] According to a thirteenth aspect, there is provided a computer readable storage medium. The computer readable storage medium stores a computer program (also called code or instructions). When the computer program is run on a computer, the computer is enabled to carry out a method according to any one of the first to fourth aspects and possible implementations of the first to fourth aspects.
[0093] According to a fourteenth aspect, there is provided a communication system comprising at least one of the above-mentioned devices configured to implement the terminal method and at least one of the above-mentioned devices configured to implement the access network node method. [Brief description of the drawings]
[0094] [Figure 1] 1 is a diagram of a communication system in accordance with the present disclosure. [Figure 2A] FIG. 1 illustrates an example application framework for AI in a communication system according to the present disclosure. [Figure 2B] FIG. 1 illustrates a network architecture in accordance with the present disclosure. [Figure 2C] FIG. 1 illustrates a network architecture in accordance with the present disclosure. [Figure 2D] FIG. 1 illustrates a network architecture in accordance with the present disclosure. [Figure 2E] FIG. 1 illustrates a network architecture in accordance with the present disclosure. [Figure 2F] FIG. 2 illustrates the relationships between neural network layers in accordance with the present disclosure. [Diagram 3] FIG. 1 illustrates a process for training an intelligent model in accordance with the present disclosure. [Figure 4] 1 is a schematic flow chart illustrating a communication method according to the present disclosure. [Figure 5A] FIG. 1 illustrates the input / output relationship of a neuron without using the dropout approach, in accordance with the present disclosure. [Figure 5B] FIG. 1 illustrates the input-output relationship of a neuron using a dropout approach in accordance with the present disclosure. [Figure 5C] 4 is another schematic flow chart illustrating a communication method according to the present disclosure. [Figure 6] 4 is another schematic flow chart illustrating a communication method according to the present disclosure. [Figure 7] 1 is a schematic block diagram of an example of a communication device according to the present disclosure. [Figure 8] FIG. 2 is a diagram illustrating a configuration of an example of a terminal device according to the present disclosure. [Figure 9] FIG. 1 is a schematic diagram illustrating a configuration of an example of a network device according to the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0095] In the present disclosure, "at least one" may be further described as "one or more", and "more" may be two, three, or four or more. This is not limited. " / " may represent an "or" relationship between related objects. For example, A / B may represent A or B. "And / or" may be used to represent the existence of three relationships between related objects. For example, A and / or B may represent the following three cases: only A exists, both A and B exist, and only B exists, and A and B may be singular or plural. To facilitate the description of the technical solutions in the present disclosure, terms such as "first", "second", "A", or "B" may be used to distinguish technical features with the same or similar functions. Terms such as "first", "second", "A", or "B" do not limit the quantity and execution order. Furthermore, terms such as "first", "second", "A", or "B" do not limit the clearly different ones. Terms such as "example" or "for example" are used to indicate an example, illustration, or explanation. Any design solution described as an "example" or "for example" should not be described as preferred or advantageous over another design solution. Terms such as "example" and "for example" are used to present related concepts in a specific embodiment for ease of understanding.
[0096] FIG. 1 is a diagram illustrating an architecture of a communication system 1000 to which the present disclosure can be applied. As shown in FIG. 1, the communication system includes a radio access network (RAN) 100 and a core network (CN) 200. Optionally, the communication system 1000 may further include an Internet 300. The radio access network 100 may include at least one access network node (e.g., 110a and 110b in FIG. 1), and may further include at least one terminal (e.g., 120a to 120j in FIG. 1). The terminal is connected to the access network node in a wireless communication manner. The access network node is connected to the core network in a wireless communication manner or a wired communication manner. The core network equipment and the access network node may be separate physical devices independent of each other, or may be the same physical device that integrates the functions of the core network equipment and the access network node, or may be another possible example. For example, one physical device may integrate the functions of an access network node and a part of the functions of a core network device, and the other physical device implements the remaining part of the functions of the core network device. The physical existence forms of the core network device and the access network node are not limited in this disclosure. A terminal may be connected to another terminal in a wired manner or in a wireless manner. An access network node may be connected to another access network node in a wired manner or in a wireless manner. Figure 1 is merely a diagram. The communication system may further include other network devices, for example, wireless relay devices, wireless backhaul devices, and the like.
[0097] The access network node may be an access network device, for example, a base station, a NodeB, an evolved NodeB (eNodeB or eNB), a transmission reception point (TRP), a next generation NodeB (gNB) of a 5th generation (5G) mobile communication system, an access network node of an open radio access network (O-RAN or open RAN), a next generation base station of a 6th generation (6G) mobile communication system, a base station of a future mobile communication system, or an access node of a wireless fidelity (Wi-Fi) system. Alternatively, the access network node may be a module or unit that complements part of the functions of a base station, for example, a central unit (CU), a distributed unit (DU), a central unit control plane (CU-CP) module, or a central unit user plane (CU-UP) module. The access network node may be a macro base station (e.g., 110a in FIG. 1), a micro base station or an indoor base station (e.g., 110b in FIG. 1), or a relay node, a donor node, or the like. The specific technology and the specific device form used by the access network node are not limited by this disclosure. The 5G system may also be referred to as a new radio (NR) system.
[0098] In the present disclosure, the device configured to implement the functions of the access network node may be an access network node, or may be a device capable of supporting the access network node in implementing the functions, such as a chip system, a hardware circuit, a software module, or a hardware circuit + software module. The device may be installed in the access network node, or may be used together with the access network node in a matching manner. In the present disclosure, the chip system may include a chip, or may include a chip and another discrete component. For ease of explanation, the following describes the technical solution provided in the present disclosure by using an example in which the device configured to implement the functions of the access network node is an access network node, and optionally, the access network node is a base station.
[0099] The terminal may also be referred to as terminal equipment, user equipment (UE), mobile station, mobile terminal, or the like. The terminal may be widely used in various scenarios for communication. For example, the scenarios include, but are not limited to, at least one of the following scenarios: enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), massive machine-type communication (mMTC), device-to-device (D2D), vehicle-to-everything (V2X), machine-type communication (MTC), internet of things (IOT), virtual reality, augmented reality, industrial control, autonomous driving, telemedicine, smart grid, smart furniture, smart office, smart wearable, intelligent transportation, smart city, or the like. The terminal may be a mobile phone, a tablet computer, a computer with wireless transceiver capabilities, a wearable device, a vehicle, an unmanned aerial vehicle, a helicopter, an airplane, a ship, a robot, a robotic arm, a smart home device, or the like. The specific technology used by the terminal and the specific device form factor are not limited by this disclosure.
[0100] In the present disclosure, the device configured to implement the functions of the terminal may be a terminal, or may be a device that can support the terminal device in implementing the functions, such as a chip system, a hardware circuit, a software module, or a hardware circuit + software module, etc. This device may be incorporated into the terminal or used together with the terminal in a matching manner. For ease of explanation, the following describes the technical solutions provided in the present disclosure by using an example in which the device configured to implement the functions of the terminal is a terminal, and optionally by using an example in which the terminal is a UE.
[0101] The base station and / or terminal may be at a fixed location or may be mobile. The base station and / or terminal may be deployed on land, indoors, or outdoors, or may be hand-held or vehicle-mounted, deployed on water, or deployed on an airplane, balloon, or airborne satellite. The environment / scenario for the base station and terminal is not limited in this disclosure. The base station and terminal may be deployed in the same environment / scenario or in different environments / scenarios. For example, the base station and terminal are both deployed on land. Alternatively, the base station is deployed on land and the terminal is deployed on water. These examples are not provided one by one.
[0102] The roles of the base station and the terminal may be relative. For example, the helicopter or unmanned aerial vehicle 120i in FIG. 1 may be configured as a mobile base station. For the terminal 120j that accesses the wireless access network 100 via 120i, the terminal 120i may be a base station, while for the base station 110a, 120i may be a terminal. In other words, 110a and 120i may communicate with each other according to a wireless air interface protocol. Alternatively, 110a and 120i may communicate with each other according to an interface protocol between base stations. In this case, 120i may also be a base station for 110a. Therefore, the base station and the terminal may be collectively referred to as a communication device (or communication equipment). 110a and 110b in FIG. 1 may be referred to as a communication device having a base station function, and 120a to 120j in FIG. 1 may be referred to as a communication device having a terminal function.
[0103] Optionally, the protocol layer configuration between the access network node and the terminal may include an artificial intelligence (AI) layer used for transmitting data related to AI functions.
[0104] In the present disclosure, an independent network element (e.g., referred to as an AI entity, AI network element, AI node, or AI device, etc.) may be introduced into the communication system shown in FIG. 1 to implement AI-related operations. The AI network element may be directly connected to the base station or indirectly connected to the base station via a third-party network element. Optionally, the third-party network element may be a core network element, such as an access and mobility management function (AMF) network element or a user plane function (UPF) network element. Alternatively, an AI entity (sometimes referred to as an AI module or another name) may be located in a network element in the communication system to implement AI-related operations. Optionally, the network element in which the AI entity is located may be a base station, a core network equipment, an operations, administration, and maintenance (OAM) device, or the like. OAM is configured for an operations, administration, and maintenance system of the core network device and / or for an operations, administration, and maintenance system of the access network node.
[0105] Optionally, to match and support the AI, the AI entity may be integrated into a terminal or a terminal chip.
[0106] Optionally, in this disclosure, an AI entity may also be referred to by another name, such as an AI module or an AI unit, and is primarily configured to implement an AI function (or an AI-related operation). The particular name of an AI entity is not limited in this disclosure.
[0107] In this disclosure, an AI model is a specific method for implementing an AI function. An AI model represents a mapping relationship between the input and output of a model. An AI model may be a neural network, a linear regression model, a decision tree model, an SVD clustering model, or another machine learning model. An AI model may be referred to as an intelligent model, a model, or another name for short. This is not limited. An AI-related operation may include at least one of the following: data collection, model training, model information release, model testing (or called model checking), model inference (or called model inference, inference, prediction, or the like), inference result release, or the like.
[0108] FIG. 2A is a diagram illustrating an example of a first application framework of AI in a communication system. In FIG. 2A, a data source is used to store training data and inference data. A model training host analyzes or trains the training data provided by the data source to obtain an AI model, and deploys the AI model to a model inference host. The AI model represents a mapping relationship between the input and output of the model. Obtaining an AI model through learning by the model training host is equivalent to obtaining a mapping relationship between the input and output of the model through learning by the model training host by using the training data. The model inference host uses the AI model to perform inference based on the inference data provided by the data source, and obtains an inference result. And the method can be described as follows: That is, the model inference host inputs inference data into the AI model, and obtains an output by using the AI model, and the output becomes an inference result. The inference result may indicate the configuration parameters used (operated) by the execution object, and / or the operation performed by the execution object. Also, the inference results may be uniformly planned by an actor entity and sent to one or more execution objects (e.g., network elements, etc.) for execution. Optionally, the model inference host may feed back the inference results of the model inference host to the model training host. This process may be referred to as model feedback. The feedback parameters are used by the model training host to update the AI model and deploy the updated AI model to the model inference host. Optionally, the execution object may feed back the network parameters collected by the execution object to a data source. This process may be referred to as performance feedback. The fed back parameters may be used as training data or inference data.
[0109] In the present disclosure, the application framework shown in FIG. 2A may be deployed in the network element shown in FIG. 1. For example, the application framework in FIG. 2A may be deployed in at least one of the terminal equipment, the access network equipment, the core network equipment, or the independently deployed AI network element (not shown) in FIG. 1. For example, the AI network element (which may be considered as a model training host) may perform analysis or training on training data provided by the terminal equipment and / or the access network equipment to obtain a model. At least one of the terminal equipment, the access network equipment, or the core network equipment (which may be considered as a model inference host) may perform inference by using the model and the inference data to obtain an output of the model. The inference data may be provided by the terminal equipment and / or the access network equipment. The input of the model includes the inference data, and the output of the model is an inference result corresponding to the model. At least one of the terminal equipment, the access network equipment, or the core network equipment (which may be considered as an execution object) may perform a corresponding operation based on the inference data and / or the inference result. The model inference host and the execution object may be the same or different. This is not limited.
[0110] 2B to 2E, the following describes a network architecture to which the communication solutions provided in the present disclosure can be applied.
[0111] As shown in FIG. 2B, in a first possible implementation, the access network device includes a near-real-time access network intelligent controller (RAN intelligent controller, RIC) module configured to perform model learning and inference. For example, the near-real-time RIC may be configured to train an AI model and use the AI model for inference. For example, the near-real-time RIC may obtain information on the network side and / or terminal side from at least one of the CU, DU, or RU. The information may be used as training data or may be used as inference data. Optionally, the near-real-time RIC may submit the inference result to at least one of the CU, DU, or RU. Optionally, the CU and the DU may exchange the inference result. Optionally, the DU and the RU may exchange the inference result. For example, the near-real-time RIC submits the inference result to the DU, and the DU forwards the inference result to the RU.
[0112] As shown in FIG. 2B, in a second possible implementation, a non-real-time RIC (optionally, the non-real-time RIC may be located in an OAM or core network device) is outside the access network and is configured to perform model learning and inference. For example, the non-real-time RIC is configured to train an AI model and use the model for inference. For example, the non-real-time RIC may obtain information on the network side and / or terminal side from at least one of a CU, a DU, or a RU. This information may be used as training data or may be used as inference data, and the inference result may be submitted to at least one of a CU, a DU, or a RU. Optionally, the CU and the DU may exchange the inference result. Optionally, the DU and the RU may exchange the inference result. For example, the non-real-time RIC submits the inference result to the DU, and the DU forwards the inference result to the RU.
[0113] As shown in FIG. 2B, in a third possible implementation, the access network equipment includes a near-real-time RIC, and a non-real-time RIC (optionally, the non-real-time RIC may be located in an OAM or core network equipment) is outside the access network. As in the second possible implementation, the non-real-time RIC may be configured to perform model learning and inference. Additionally / alternatively, as in the first possible implementation, the near-real-time RIC may be configured to perform model learning and inference. Additionally / alternatively, the near-real-time RIC may obtain AI model information from the non-real-time RIC, obtain information on the network side and / or terminal side from at least one of the CU, DU, or RU, and obtain an inference result by using the information and the AI model information. Optionally, the near-real-time RIC may submit the inference result to at least one of the CU, DU, or RU. Optionally, the CU and the DU may exchange the inference result. Optionally, the DU and the RU may exchange the inference result. For example, the near-real-time RIC submits the inference result to the DU, and the DU submits the inference result to the RU. For example, the near real-time RIC may be configured to train model A and use model A for inference. For example, the non-real-time RIC may be configured to train model B and use model B for inference. For example, the non-real-time RIC may be configured to train model C and submit information about model C to the near real-time RIC, which uses model C for inference.
[0114] 2C is a diagram showing an example of a network architecture to which the method provided in the present disclosure can be applied. Compared with FIG. 2B, in FIG. 2C, the CU is divided into a CU-CP and a CU-UP.
[0115] 2D is a diagram illustrating an example of a network architecture to which the method provided in the present disclosure can be applied. As shown in FIG. 2D, optionally, the access network equipment includes one or more AI entities, and the function of the AI entities is similar to the function of the near real-time RIC described above. Optionally, the OAM includes one or more AI entities, and the function of the AI entities is similar to the function of the non-real-time RIC described above. Optionally, the core network equipment includes one or more AI entities, and the function of the AI entities is similar to the function of the non-real-time RIC described above. When the OAM and the core network equipment each include an AI entity, the models obtained through training by the AI entities of the OAM and the core network equipment are different from each other and / or the models used for inference are different from each other.
[0116] In the present disclosure, model differences include at least one of the following differences: model configuration parameters (e.g., at least one of the number of model layers, model width, inter-layer connectivity, neuron weights, neuron activation functions, or offsets in activation functions), model input parameters (e.g., input parameter types and / or input parameter dimensions), or model output parameters (e.g., output parameter types and / or output parameter dimensions).
[0117] FIG. 2E is a diagram illustrating an example of a network architecture to which the method provided in the present disclosure can be applied. Compared with FIG. 2D, the access network equipment in FIG. 2E is divided into a CU and a DU. Optionally, the CU may include an AI entity, and the function of the AI entity is similar to that of the near-real-time RIC. Optionally, the DU may include an AI entity, and the function of the AI entity is similar to that of the near-real-time RIC. When the CU and the DU include an AI entity, respectively, the models obtained through training by the AI entities of the CU and the DU are different, and / or the models used for inference are different. Optionally, the CU in FIG. 2E may be further divided into a CU-CP and a CU-UP. Optionally, one or more AI models may be deployed in the CU-CP. Optionally, one or more AI models may be deployed in the CU-UP.
[0118] Optionally, as described above, in Fig. 2D or Fig. 2E, the OAM of the access network equipment and the OAM of the core network equipment may be deployed separately. In Fig. 2D or Fig. 2E, the core network equipment may include, for example, one or more modules including
[0119] In the present disclosure, a parameter or multiple parameters may be obtained through inference by using one model. The learning processes of different models may be located on different devices or nodes, or may be located on the same device or node. The inference processes of different models may be located on different devices or nodes, or may be located on the same device or node.
[0120] The network architecture and service scenarios described in this disclosure are intended to more clearly explain the technical solutions in this disclosure, and do not constitute limitations on the technical solutions provided in this disclosure. Those skilled in the art may understand that with the evolution of network architectures and the emergence of new service scenarios, the technical solutions provided in this disclosure can also be applied to similar technical problems.
[0121] Before describing the method in this disclosure, some related knowledge about artificial intelligence will be briefly described first. Artificial intelligence can give a machine human intelligence. For example, a machine can simulate some intelligent human behavior by using computer software and hardware. Machine learning methods or other methods can be used to implement artificial intelligence. This is not limited.
[0122] A neural network (NN) is a specific implementation of machine learning. According to the general approximation theorem, a neural network can theoretically approximate any continuous function, which allows the neural network to learn any mapping. Therefore, a neural network can accurately perform abstract modeling for complex high-dimensional problems. In other words, an intelligent model can be implemented by using a neural network.
[0123] The concept of a neural network comes from the neuronal organization of the brain. Each neuron performs a weighted sum operation on its inputs and outputs the result of the weighted sum through an activation function. If the input of a neuron is x=[x 0 ,...,x n ], and the weights (or weights) corresponding to the inputs are w=[w 0 ,...,w n] and the offset of the weighted sum is assumed to be b. The form of the activation function can be varied. If the activation function of a neuron is y = f(z) = max(0,z), then the output of the neuron is
[0124]
number
[0125] As another example, if a neuron's activation function is y=f(z)=z, then the output of the neuron is
[0126]
number
[0127] Element x in the neuron's input x i , element w of weight w i , or the offset b, can take on a variety of possible values, such as a decimal number, an integer (including 0, a positive integer, a negative integer, or the like), or a complex number. The activation functions of different neurons in a neural network may be the same or different.
[0128] A neural network generally includes a multi-layer configuration, and each layer may include one or more neurons. Increasing the depth and / or width of a neural network can improve the representation capability of the neural network and provide a more powerful information extraction and abstract modeling capability for complex systems. The depth of a neural network indicates the number of layers included in the neural network, and the number of neurons included in each layer is sometimes called the layer width. Figure 2F is a diagram showing the layer relationship of a neural network.
[0129] In the implementation, the neural network includes an input layer and an output layer. After performing neuron processing on the received input, the input layer of the neural network transfers the result to the output layer, which obtains the output result of the neural network.
[0130] In another implementation, the neural network includes an input layer, a hidden layer, and an output layer. The input layer of the neural network performs neuron processing on the received input, and then transfers the result to the intermediate hidden layer. The hidden layer then transfers the calculation result to the output layer or the adjacent hidden layer. Finally, the output layer obtains the output result of the neural network. A neural network may include one or more hidden layers connected in sequence. This is not limited.
[0131] In the training process of an AI model, a loss function may be defined. The loss function describes the difference between the output value of the AI model and an ideal target value. The specific form of the loss function is not limited in this disclosure. The training process of an AI model is a process of adjusting some or all of the parameters of the AI model so that the value of the loss function is less than a threshold or meets a target requirement. For example, the AI model is a neural network. In the training process of a neural network, one or more of the following parameters may be adjusted so that the difference between the output of the neural network and the ideal target value is as small as possible: the number of layers of the neural network, the width of the neural network, the connection relationship between the layers, the weights of the neurons, the activation function of the neurons, or the offset of the activation function.
[0132] FIG. 3 is a diagram illustrating an example of a training process of an intelligent model. As shown in FIG. 3, the intelligent model is a neural network model, and the neural network f θWe will explain using an example where (·) may include a network with four convolutional layers and a network with four fully connected layers. The AI entity may obtain training data from a base dataset (e.g., a set of channel data). The training data may include training samples and labels. When used as input, a sample x is fed to a neural network f θ (·), and then the inference result f θ (x) is output. The error between the inference result and the label is obtained through calculation by using a loss function. The AI entity may optimize the network parameters θ by using a backpropagation optimization algorithm (sometimes called a model optimization algorithm) based on the error obtained by using the loss function. The neural network is trained by using a large amount of training data, such that the training of the neural network is completed after the difference between the output of the neural network and the label is less than a preset value.
[0133] It should be noted that in the training process shown in FIG. 3, a supervised learning training approach is used, in other words, model training is implemented based on samples and labels by using a loss function. The intelligent model training process may alternatively use unsupervised learning. The internal patterns of the samples are learned by using an algorithm to complete the training of the intelligent model based on the samples. The intelligent model training process may alternatively use reinforcement learning. An excitation signal fed back by the environment is obtained through interaction with the environment to learn a problem-solving policy and optimize the model. The model training method and model type are not limited in this disclosure.
[0134] The trained AI model can perform inference tasks. After real data is input into the AI model for processing, corresponding inference results are obtained. Optionally, a parameter or multiple parameters can be obtained through inference by using an AI model.
[0135] Applying artificial intelligence in wireless communication systems can greatly improve the performance of the communication system. In most scenarios, the network side and the terminal need to use matching artificial intelligence models to improve wireless communication performance. For example, the terminal may perform compression encoding on the uplink information by using a compression encoder model, and then transmit the encoded uplink information to the network side, and the network side decodes the decoded uplink information by using a matching decoder model to obtain the uplink information transmitted by the terminal. The application of artificial intelligence in the field of wireless communication is usually associated with complex nonlinear function fitting. The scale of the intelligent model is usually large, for example, the number of model layers is large, and the number of model parameters is large. For example, a convolutional neural network (CNN) of an encoder model implementing CSI feedback compression may include 15 neural network layers.
[0136] To meet the performance requirements, the intelligent model used by the terminal is usually specified by the network. The network side may select one model from a number of predefined models and notify the terminal of the model's identifier, so that the terminal can determine the model to be used based on the identifier. However, there are various types of terminals, the channel environment is complex and subject to change, and the number of predefined models is limited. Therefore, good communication performance cannot be ensured in all environments. In order to ensure communication performance, the network side may alternatively provide the terminal with an intelligent model. However, the number of parameters of the intelligent model is large, and a large amount of air interface resources need to be occupied, which affects the transmission efficiency of service data. When the terminal downloads a model from the network, the terminal needs to download the configuration parameters of the model, the weights of each neuron, the activation function, the offset, and the like. The total amount of parameters reaches hundreds of thousands of floating-point numbers. In this case, the overhead of the air interface resources is huge, which reduces the air interface resources of the service data and reduces the transmission efficiency of the service data. In response to this problem, the present disclosure proposes the following. That is, the network may inform the terminal of the training policy of the intelligent model, and the terminal trains the intelligent model according to the training policy provided by the network, so that the model obtained through training by the terminal can be consistent with the model used by the network, thereby meeting the expected performance requirements.
[0137] The following describes the communication method provided in the present disclosure with reference to the accompanying drawings. The interaction between a terminal and an access network node is used as an example in this specification, but it should be noted that the present application is not limited thereto. The communication method provided in the present disclosure can be used between any two nodes. One node obtains a training policy and / or a model configuration of an intelligent model from the other node, and performs model training according to the training policy and / or based on the model configuration. In the present disclosure, as described above, one or more modules (such as RU, DU, CU, CU-CP, CU-UP, or near real-time RIC) in the access network node may implement the corresponding method or operation of the access network node.
[0138] FIG. 4 is a schematic flow chart illustrating a communication method 400 according to the present disclosure.
[0139] S401: An access network node sends first information to a terminal, where the first information indicates a training policy of a first intelligent model.
[0140] In response, the terminal receives first information from the access network node. The terminal determines a training policy for the first intelligent model based on the first information.
[0141] In this disclosure, by way of example and not limitation, training policies may include 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 It is.
[0142] Below, the above-mentioned training policies will be explained one by one.
[0143] (1) Model training method
[0144] The first information may indicate a training method for the terminal to train the first intelligent model, and the terminal may train the first intelligent model in the model training method indicated by the first information, so that the first intelligent model obtained through training can be consistent with the intelligent model on the network side, thus improving wireless communication performance.
[0145] For example, the first information may instruct the terminal to train the first intelligent model in one of supervised learning, unsupervised learning, or reinforcement learning, although the present disclosure is not limited in this respect.
[0146] (2) Loss function information
[0147] The first information may include loss function information. After receiving the first information, the terminal may obtain, based on the loss function information, a loss function used to train the first intelligent model.
[0148] The access network node may inform the terminal of the loss function to be used by using the loss function information, which may indicate the loss function directly or the loss function information may indicate the identity of one of a number of predefined loss functions.
[0149] In one example, the terminal may train the first intelligent model in a supervised learning manner. For example, the supervised learning training manner may be predefined, or the access network node may indicate that the model training manner is a supervised learning training manner by using the first information. For example, a training objective of the supervised learning model training manner is to allow the value of the loss function to reach a minimum value. In a particular implementation, the model training may be completed when the value of the loss function obtained through training is equal to or less than a first threshold. Optionally, the first threshold may be predefined or indicated by the loss function information.
[0150] For example, the loss function may be the cross-entropy loss function:
[0151]
number
[0152] It is.
[0153] log(A) represents the calculation of the logarithm of A, and Σ n B n is the B based on the range of values of n n Represents the sum of.
[0154] Alternatively, the loss function can be a quadratic loss function:
[0155]
number
[0156] It is.
[0157] x n is the sample, and y n is the label of the nth sample, and f θ (x n ) is the input data x n This is the output of the model for
[0158] In another example, the terminal may train the first intelligent model in an unsupervised learning manner. For unsupervised learning, the loss function may be a function used to evaluate model performance. For example, the training objective of the unsupervised learning model training method is to allow the value of the loss function to reach a maximum value. In a specific implementation, when the value of the loss function obtained through training is equal to or greater than a second threshold, the model training may be completed. Optionally, the second threshold may be predefined or may be indicated by the loss function information.
[0159] For example, if the first intelligent model is used to infer the transmit power based on the channel response, the input of the first intelligent model is the channel response h n and the inference result of the first intelligent model is the transmission power P n In this case, the loss function may be a function for calculating the data throughput. For example, the loss function may be as follows:
[0160]
number
[0161] log(A) represents the calculation of the logarithm of A, and N 0 is the noise power.
[0162] The loss function may be the above-mentioned formula. Optionally, the loss function may alternatively be a machine learning intelligent model. The complexity and parameter amount of the loss function model are much smaller than the complexity and parameter amount of the first intelligent model. The loss function information may include configuration information and / or parameter information of the loss function, so that the terminal can obtain the loss function model after receiving the first information, and train the first intelligent model based on the loss function, thereby improving the model training effect of the terminal.
[0163] For example, if an unsupervised learning method is used, the loss function may be a non-linear performance measurement model for system throughput, signal bit error rate, signal to interference plus noise ratio, and so on.
[0164] (3) Model initialization method
[0165] The first information may indicate a model initialization technique, and the terminal may initialize the first intelligent model based on the model initialization technique indicated by the first information.
[0166] In one example, the first intelligent model may be a neural network model, and the first information may indicate how to initialize the weights of the neurons in the first intelligent model. For example, the access network node may indicate to the terminal to randomly select values for the weights of each neuron, or indicate that the weights of all neurons are 0, or that the weights of the neurons are generated by using a preset function, or the like.
[0167] For example, the access network node may use the first information to provide the terminal with a neuron weight of [z min ,z max For example, the first information may indicate that a value within the range z min and z max After receiving the first information, the terminal may indicate for each neuron "z min ,z max " and use that value as the initial value for the neuron's weight.
[0168] As another example, the first information indicates one of the identification information of a plurality of predefined initialization methods. For example, the first information includes a two-bit instruction field to indicate an initialization method of the weight of a neuron. The two-bit instruction "00" indicates that the initial value of the weight of the neuron is selected randomly. Optionally, the terminal may select a weight within a preset value range for each neuron, or if the first information indicates a method, the first information further includes an end value within the value range, and the terminal selects a weight within the value range indicated by the first information for the neuron. The two-bit instruction "01" indicates that the initial values of all neurons are set to "0". The two-bit instruction "10" indicates that an initial value sequence is generated by using a preset function, and each value in the sequence is used as the initial value of one neuron.
[0169] (4) Type of model optimization algorithm
[0170] In the example shown in Fig. 3, in the model training, the training device may obtain an optimization parameter of the intelligent model based on the output value of the loss function by using a model optimization algorithm. After updating the intelligent model based on the optimization parameter, the training device performs next model training on the intelligent model whose parameter has been updated. The access network node may inform the terminal of the optimization algorithm of the first intelligent model by using the first information.
[0171] For example, the first information may indicate a type of optimization algorithm among multiple predefined types of model optimization algorithms. When training the first intelligent model, the terminal may perform model optimization by using the model optimization algorithm indicated by the first information. By way of example and not limitation, the model optimization algorithm may include, but is not limited to, one or more of the following algorithms: The adaptive momentum estimation (ADAM) algorithm, the stochastic gradient descent (SGD) algorithm, or the batch gradient descent (BGD) algorithm It is.
[0172] Optionally, the communication protocol may predefine multiple types of model optimization algorithms, and parameters of each type of model optimization algorithm may be predefined. The terminal may determine the type of optimization algorithm and the predefined parameters of the optimization algorithm based on the identification information of the type of model optimization algorithm indicated by the first information. During training of the first intelligent model, the terminal updates the weight of the model by using the model optimization algorithm. Alternatively, the parameters of the optimization algorithm may be indicated by the first information. For details, please refer to the following description.
[0173] (5) Optimization algorithm parameters
[0174] The first information may indicate parameters of the model optimization algorithm. The type of the model optimization algorithm may be indicated by the access network node by using the first information, or the type of the model optimization algorithm may be pre-configured.
[0175] By way of example and not limitation, the parameters of the model optimization algorithm may include, but are not limited to, one or more of the following parameters: The learning rate, number of iterations, or amount of data processed in a batch It is.
[0176] The learning rate is a parameter representing a gradient descent step in model training. The learning rate may be a fixed value, and the access network node may inform the terminal of the value of the learning rate by using the first information. For example, the learning rate may be set to 0.001. Alternatively, the learning rate may be set to a value that gradually decreases with the iteration step. The first information may indicate the learning rate and the decreasing value of the learning rate obtained through each iteration. For example, the first information may indicate that the initial value is 0.1, and the value decreases by 0.01 after each iteration. The number of iterations is the number of times to iterate the training dataset. The amount of data processed into a batch is the amount of training data selected from the training dataset during each model training and used for gradient descent update.
[0177] According to the above description, the training policy of the first intelligent model indicated by the first information may include, but is not limited to, one or more of the above-mentioned policies (1) to (5). In this way, the terminal can train the first intelligent model according to the training policy to obtain an intelligent model that is compatible with the network.
[0178] It should be noted that if the access network node informs the terminal that one or more of the training policies (1) to (5) are not indicated, the terminal may determine the training policies not indicated by the access network node, including but not limited to, in the following manner:
[0179] In implementation, the terminal may be pre-configured with one or more of the above-mentioned training policies. If the access network node does not indicate one or more of the policies, the terminal performs model training according to the pre-configured training policies. The pre-configured training policies may be defined in the protocol, or the pre-configured training policies may be configured by the manufacturer based on the terminal manufacturing implementation.
[0180] In another implementation, the terminal may determine one or more training policies not indicated by the access network node based on one or more training policies indicated by the access network node. In other words, there may be an association relationship between multiple training policies. By indicating a training policy, the access network node implicitly indicates another training policy related to the training policy. The association relationship may be agreed upon in a protocol or may be notified to the terminal in advance by the access network node. This is not limited.
[0181] For example, the access network node may indicate a model training approach by using the first information, but without indicating the loss function information, the terminal determines the loss function based on the model training approach indicated by the access network node. If the first information indicates that the model training approach is a supervised learning approach, 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 quadratic loss function. However, the present application is not limited thereto.
[0182] Optionally, the access network node may further send second information to the terminal, where the second information indicates a configuration of the first intelligent model.
[0183] In addition to informing the terminal of the training policy of the first intelligent model, the access network node further informs the terminal of the configuration of the first intelligent model by using the second information. After receiving the second information, the terminal generates an initial first intelligent model (or referred to as a pre-trained first intelligent model) having its configuration based on the second information, and trains the first intelligent model by using the training policy indicated by the first information, so that the trained first intelligent model can be consistent with the model used by the access network node.
[0184] The access network node may indicate an identification information of one configuration among the predefined intelligent model configurations and notify the terminal of the intelligent model using the configuration, or the access network node may indicate the configuration parameters of the first intelligent model.
[0185] By way of example and not limitation, the second information may specifically indicate one or more of the following configuration information of the first intelligent model: Network layer configuration information, input data dimensions, or output data dimensions It is.
[0186] For example, the second information may indicate the dimension of the input data of the first intelligent model, i.e., the dimension of the training sample. When the first intelligent model is a neural network model and includes an input layer and an output layer, the dimension of the input data is the dimension of the input data in the input layer of the first intelligent model.
[0187] As another example, the second information may indicate the dimension of the output data of the first intelligent model, i.e., the dimension of the inference data output by the first intelligent model. When the first intelligent model is a neural network model, the dimension of the output data is the dimension of the output data at the output layer of the first intelligent model.
[0188] The second information may include network layer configuration information of the first intelligent model, where the network layer configuration information may include, but is not limited to, one or more of the following:
[0189] The number of neural network layers included in the first intelligent model, the types of neural network layers, the manner in which the neural network layers are used, the cascade relationship between the neural network layers, the dimensions of the input data of the neural network layers, or the dimensions of the output data of the neural network layers.
[0190] For example, 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, for example, a CNN neural network layer, an FNN neural network layer, or a dimension transformation neural network layer.
[0191] If the network layer configuration information indicates that the type of the neural network layer in the first intelligent model is CNN, the CNN may be implemented by a filter. The access network node may further inform the terminal of the type of filter for implementing the CNN by using the network layer configuration information. For example, the type of filter may be the dimension of the filter. For example, the network layer configuration information may indicate that the dimension of the filter is 3×3, 5×5, 3×3×3, or the like.
[0192] The dimension conversion neural network layer is used to convert data dimensions. For example, the dimension conversion neural network layer can be disposed between two neural network layers. The dimension conversion neural network layer is used to do the following: convert the dimension of the input data (i.e., the output data of the previous neural network layer, etc.) to obtain the dimension of the input data of the next neural network layer, and output the data obtained through the dimension conversion to the next neural network layer. For example, it is required that the dimension of the output data of the neural network layer A is 3×3×10, and the dimension of the input data of the neural network layer B is 30×3. In this case, the dimension conversion neural network layer is set between the two neural network layers, and the 3×3×10 data output by the neural network layer A is converted into 30×3 data, and the 30×3 data is output to the neural network layer B. In this way, the data obtained by the dimension conversion can be continuously processed by the neural network layer B.
[0193] The access network node may further inform the terminal of the manner in which the neural network layer is used by using the second information, for example, may inform the terminal whether one or more neural network layers in the neural network use a temporary dropout manner, which means that during model training, one or more neurons in the neural network layer may be randomly selected to skip the current training, and one or more neurons may return to the model training in the next training.
[0194] For example, when the CNN neural network layer does not use the dropout method, the input-output relationship of the neurons in the neural network layer may be shown in FIG. 5A. When the CNN neural network layer uses the dropout method, the input-output relationship of the neurons in the neural network layer may be shown in FIG. 5B. The terminal may randomly select multiple neurons in the CNN neural network layer and disengage the neurons from model training, and the output of the neurons is set to 0, and the weights of the neurons disengaged from model training are not updated in the current training, and the weights are still used when the neurons are engaged in model training in the next training.
[0195] The access network node may further inform 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. Additionally / alternatively, the access network node may further inform the terminal of the cascade relationship between the neural network layers. For example, the cascade relationship between the neural network layers may be the arrangement order of the neural network layers in the first intelligent model.
[0196] Table 1 shows an example of a neural network configuration. The access network node may inform the terminal of the neural network configuration shown in Table 1. However, the present disclosure is not limited thereto. For example, the second information indicates that the quantity of the neural network layers in the first intelligent model is six layers, and informs the type of each neural network layer. For example, the type of the neural network layer includes an input layer, a CNN layer, a shape transformation layer, an FNN layer, and an output layer. The second information may further indicate a cascade relationship between the neural network layers. For example, the second information may indicate that the arrangement order of the six neural network layers may be an input layer, a CNN layer, a CNN layer, a shape transformation layer, an FNN layer, and an output layer. In addition, the second information may further indicate the input dimension and the output dimension of each neural network layer as shown in Table 1, N T is the number of antennas, and N sc is the quantity of subcarriers. The second information may further inform the terminal that the filters implementing the two CNN layers, specifically the second layer and the third layer, are 3×3 filters, and that the second, third, fourth, and fifth neural network layers may use a dropout technique. However, the present disclosure is not limited thereto. A part of the configuration of the first intelligent model may be predefined, and another part may be instructed by the access network node by using the second information. For example, the access network node may inform the terminal of some or all of the configuration parameters shown in Table 1. N / A shown in Table 1 indicates that the parameter is not applicable (or not configured) in this row.
[0197] [Table 1]
[0198] The terminal may generate an initial first intelligent model based on the configuration of the first intelligent model shown in Table 1 and indicated by the second information. The terminal may first determine that the first intelligent model includes six neural network layers based on the second information, and indicate the type, input dimension, and output dimension of each of the six neural network layers based on the second information to generate six neural network layers that meet the type and dimension requirements, namely, an input layer, a CNN layer, a CNN layer, a shape transformation layer, an FNN layer, and an output layer. The terminal implements two CNN layers based on the indication of the second information by using a 3×3 filter. After obtaining the six neural network layers, the terminal determines the arrangement order of the six neural network layers based on the cascade relationship indicated by the second information to obtain the initial first intelligent model. In the process of training the first intelligent model, the neurons in the second, third, fourth, and fifth neural network layers are randomly selected and disengaged from the current training in each training in a dropout manner.
[0199] Optionally, the access network node sends third information to the terminal, where the third information indicates information about a training dataset, where the training dataset is used to train the first intelligent model.
[0200] In one example, the third information may indicate identification information of a preset training dataset, so that after receiving the third information, the terminal trains the first intelligent model by using the training dataset corresponding to the identification information.
[0201] For example, the protocol may predefine multiple training data sets, and each training data set corresponds to one identification information. The access network node may indicate one identification information of the multiple training data sets by using third information. The terminal determines, based on the identification information, to train the first intelligent model by using the training data set corresponding to the identification information.
[0202] In another example, the information about the training dataset may indicate the type of training samples, or the type of training samples and the type of labels.
[0203] For example, the training method of the first intelligent model may be an unsupervised learning method, and the third information may indicate a type of the training sample. For example, the first intelligent model may estimate a transmission power based on input channel data. In this case, the third information indicates that the type of the training sample is channel data, and the terminal may train the first intelligent model by using the channel data. The channel data may be obtained through measurement by the terminal based on a reference signal transmitted by an access network node, or may be pre-configured in the terminal. However, the present application is not limited thereto.
[0204] As another example, the training method of the first intelligent model is an unsupervised learning method. The third information may indicate a type of training sample and a type of label. The terminal may determine a training dataset based on the type of training sample and the type of label. The data included in the training dataset may be data collected during communication between the terminal and the access network node, or may be a corresponding type of data preconfigured in the terminal. For example, the first intelligent model implements a channel coding function of data, in other words, outputs coded data through inference based on input pre-coded data. The third information may indicate that the type of training sample is data before channel coding, and the type of label is data after channel coding.
[0205] In another example, the information about the training dataset may further indicate use of the first intelligent model. For example, the information about the training dataset indicates that the first intelligent model is used for compression encoding, decoding, or inferring transmission power. The terminal determines the corresponding training dataset based on the use of the first intelligent model.
[0206] For example, the third information may indicate that the use of the first intelligent model is compression coding, and the training method of the first intelligent model is supervised learning. In this case, the terminal may determine, based on the use of the first intelligent model, that the type of training sample used to train the first intelligent model is information data before compression coding, and the label is compressed data after compression coding, and determine a training dataset of the first intelligent model including the training sample and the label. The data included in the training dataset may be data collected during communication between the terminal and the access network node, or may be a corresponding type of data pre-configured in the terminal. However, the present application is not limited thereto.
[0207] In another example, the third information may include a training dataset. For example, the training dataset may include training samples, or training samples and labels. In other words, the terminal may obtain the training dataset from the network.
[0208] For example, the training method of the first intelligent model is a supervised learning method, and 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.
[0209] The terminal may train the first intelligent model while downloading training data in the training dataset in a narrow resource bandwidth, thereby reducing resource usage and not affecting the transmission speed of service data.
[0210] In the present disclosure, the access network node may determine one or more of a training policy, a model configuration, or a training dataset of the first intelligent model based on the capability information of the terminal and / or based on environmental data (such as channel data) between the access network node and the terminal. Alternatively, the access network node may obtain a training policy and / or a model configuration of the first intelligent model from a third party node in the network, and then forward the training policy and / or the model configuration to the terminal. In other words, the information (including but not limited to the first information, the second information, and the third information) transmitted 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 the third party node and then forwarded to the terminal. For example, the third party node may be a node having an AI function, or may be a node configured with an AI entity in the network. For example, the third party node may be a non-real-time RIC, which may be included in the OAM, and the third party node may actually be the OAM. Optionally, the access network node receives capability information from the terminal, which indicates the capability of the terminal to run the intelligent model.
[0211] By way of example and not limitation, the capability information may specifically indicate one or more of the following capabilities of the terminal: Whether it supports the running of intelligent models, data processing capabilities, storage capabilities, the types of intelligent models that can be run, processor power consumption, or battery capacity It is.
[0212] For example, the capability information may indicate whether the terminal supports running an intelligent model. After receiving the capability information, if the access network node determines based on the capability information that the terminal supports the intelligent model, the access network node sends first information to the terminal to inform the terminal of a policy for training the first intelligent model.
[0213] As another example, the access network node needs to inform the terminal of the configuration of the first intelligent model by using the second information. The capability information may indicate 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 can be supported by the terminal, or the operator library information may indicate a basic intelligent model computing unit that is not supported by the terminal. The access network node may check the operator library information of the model supported by the terminal based on the capability information to determine the configuration of the first intelligent model.
[0214] For example, the intelligent model computing unit may include, but is not limited to, one or more of the following: One-dimensional CNN unit (conv1d), two-dimensional CNN unit (conv2d), three-dimensional CNN unit (conv3d), pooling layer, pooling operation, or sigmoid activation function It is.
[0215] As another example, the capability information indicates the data processing capability of the terminal. For example, the capability information may indicate a type of processor, an operating speed of the processor, an amount of data that can be processed by the processor, and the like. For example, the type of processor may include a type of graphics processing unit (GPU) and / or a type of central processing unit (CPU). The access network node may determine one or more of a training policy, a model configuration, or a training dataset of the first intelligent model based on the data processing capability of the terminal.
[0216] In implementation, the access network node may obtain an intelligent model that satisfies the terminal capability and / or the channel condition through training based on the capability information of the terminal and / or based on the environment data (such as channel data) between the access network node and the terminal. The access network node informs the terminal of a training policy for training the intelligent model, so that the terminal trains the first intelligent model by using the training policy, and the first intelligent model obtained through training by the terminal is as similar as possible to the intelligent model obtained through training by the access network node. In this way, in the process of communication with the access network node, the first intelligent model applied by the terminal can be consistent with the intelligent model used by the access network node, thereby improving the communication performance.
[0217] For example, the first intelligent model is a compression encoding model, and the access network node may obtain a compression decoding model used by the access network node and a compression encoding model used by the terminal that are consistent with each other through training based on the capability information and channel data of the terminal. The access network node informs the terminal of a training policy used by the access network node to train the compression encoding model. For example, the access network node may inform the terminal of a training policy used by the access network node that the initialization method used by the access network node uses [z min ,z max ], the model training method is a supervised learning method, the loss function is a cross-entropy loss function, and the optimization algorithm is a stochastic gradient descent algorithm. After obtaining the training policy of the access network node, the terminal trains the compression coding model according to the training policy, so that the compression coding model obtained through training by the terminal is as similar as possible to the compression coding model obtained through training by the access network node. In this way, the compression coding model used by the terminal in communication is consistent with the compression decoding model used by the access network node in communication.
[0218] Optionally, the access network node may further inform the terminal of the configuration of the intelligent model as the configuration of the first intelligent model, and / or the access network node may further indicate to the terminal information regarding a training dataset used by the access network node to train the intelligent model.
[0219] In another implementation, a third party node in a network may obtain capability information of a terminal from an access network node, and / or obtain environmental data between the access network node and the terminal, and may obtain an intelligent model applied by the access network node and an intelligent model applied by the terminal that are consistent with each other based on the obtained information through training. The third party node transmits the intelligent model applied by the access network node to the access network node, and informs the access network node of a training policy used by the terminal to train the intelligent model, and then the access network node forwards the training policy to the terminal by using the first information. Optionally, the third party node may further inform the terminal of information regarding the configuration of the intelligent model and / or the training dataset by using the access network node.
[0220] S402: The terminal performs model training on the first intelligent model according to the training policy.
[0221] The terminal may determine a training policy of the first intelligent model, such as a model training methodology, a loss function, a model initialization methodology, a type of a model optimization algorithm, and parameters of the optimization algorithm. A part or all of the training policy may be determined by the terminal based on the first information from the access network node. A training policy that is not instructed by the access network node may be obtained by the terminal based on pre-configured information or an associated training policy instructed by the access network node.
[0222] Optionally, the terminal may obtain a configuration of the first intelligent model, and determine the first intelligent model based on the configuration.
[0223] In an implementation, the terminal may receive second information from an access network node, the second information indicating a configuration of the first intelligent model, and the terminal determines an initial first intelligent model (or referred to as a pre-trained first intelligent model) based on the second information.
[0224] For a specific implementation of the terminal obtaining the initial first intelligent model based on the second information, please refer to the above description, and for the sake of brevity, the details will not be described again in this specification.
[0225] In another implementation, the terminal obtains a configuration of the first intelligent model from the pre-configured information and determines an initial first intelligent model. Optionally, the terminal may receive third information from the access network node, where the third information indicates information about the training dataset.
[0226] The terminal obtains a training dataset based on the third information, and performs model training on the first intelligent model based on the training dataset by using the training policy indicated by the first information. However, the present disclosure is not limited thereto. The terminal may further perform model training on the first intelligent model based on a dataset stored in the terminal.
[0227] After obtaining the training policy and training dataset of the first intelligent model and obtaining an initial first intelligent model based on the configuration of the first intelligent model, the terminal may start performing model training on the first intelligent model.
[0228] For example, the terminal determines according to the training policy that the model training method is a supervised learning method, and the training dataset includes a training sample and a label. The terminal inputs the training sample into the first intelligent model in each training. The first intelligent model processes the training sample, and then outputs an inference result. The terminal obtains a loss value output by the loss function based on the inference result and the label by using the loss function. Then, the parameters of the first intelligent model are optimized based on the loss value by using a model optimization algorithm to obtain a first intelligent model with updated parameters. The first intelligent model with updated parameters is applied to the next model training. After multiple iterations of training, if the loss value output by the loss function is less than or equal to a first threshold, the terminal determines that the training of the first intelligent model is completed, and a trained first intelligent model is obtained.
[0229] As another example, the terminal determines, according to the training policy, that the model training method is an unsupervised learning method, and the training data set includes a training sample. The terminal inputs the training sample into the first intelligent model in each training. The first intelligent model processes the training sample, and then outputs an inference result. The terminal obtains an output value of the loss function based on the inference result by using a loss function. Then, the parameters of the first intelligent model are optimized based on the output value of the loss function by using a model optimization algorithm to obtain a first intelligent model with updated parameters. The first intelligent model with updated parameters is applied to the next model training. After multiple iterations of training, if the output value of the loss function is equal to or greater than a second threshold, the terminal determines that the training of the first intelligent model is completed, and a trained first intelligent model is obtained. Based on the above solution, the network can inform the terminal of the training policy of the intelligent model, and the terminal trains the intelligent model according to the training policy provided by the network, so that the model obtained through the training by the terminal can be consistent with the model used by the network, thereby meeting expected performance requirements. This reduces the air interface resource overhead incurred when the access network node transmits model parameters of the first intelligent model (e.g., weights, activation functions, and offsets for each neuron) to the terminal.
[0230] Fig. 5C is a schematic flow chart showing a communication method 500 according to the present disclosure. It should be noted that for the parts in the method shown in Fig. 5C that are the same as those in the method shown in Fig. 4, please refer to the description in the method shown in Fig. 4. For the sake of brevity, the details will not be described again in this specification.
[0231] S501: An access network node sends first information to a terminal, where the first information indicates a training policy of a first intelligent model.
[0232] For specific implementation, please refer to step S401, and the details will not be described again in this specification.
[0233] S502: The terminal trains a first intelligent model according to the training policy.
[0234] The terminal may obtain the configuration of the first intelligent model, determine the first intelligent model, and train the first intelligent model by using the training data set and the training policy. For specific implementation, refer to step S402. Details are not described again in this specification.
[0235] S503: The access network node sends fourth information to the terminal, where the fourth information indicates test information, and the test information is used to test performance of the first intelligent model.
[0236] In response, the terminal receives fourth information from the access network node. After receiving the fourth information, the terminal tests the trained first intelligent model according to the fourth information.
[0237] By way of example and not limitation, the test information may include one or more of the following: Test data information, performance evaluation method, or performance evaluation parameters It is.
[0238] Optionally, when the test information includes one or two of the test data information, the performance evaluation method, and the performance evaluation parameters, for example, when the test information includes the test data information and the performance evaluation method, the remaining items not included, for example, the performance evaluation parameters, may be agreed upon in the protocol or may be determined in another manner. This is not limited. Alternatively, in the following example, the terminal does not need to check the performance evaluation method and / or the performance evaluation parameters, and the access network node performs the performance evaluation.
[0239] The test data information indicates one or more of the test data, a type of the test data, the label data, or a type of the label data. The performance evaluation technique may be a technique for calculating a loss value between the inference data and the label data. The performance evaluation parameter may be a threshold for evaluating the loss value or another parameter, although the present disclosure is not limited thereto. The performance evaluation technique may alternatively be an evaluation function, or the like.
[0240] S504: The terminal sends fifth information and / or sixth information to the access network node, where the fifth information indicates a test result of the first intelligent model, and the sixth information indicates inference data.
[0241] The inference data is obtained by inferring the test data by the first intelligent model, and the test information indicated by the fourth information includes the test data.
[0242] In an implementation, the terminal may send fifth information to the access network node, and the fifth information may indicate a test result of the first intelligent model.
[0243] For example, the test result may be that the performance of the first intelligent model meets or does not meet the requirements. The test information sent by the access network node to the terminal may include test data, check data, and a check threshold. The terminal may use the test data as an input of the trained first intelligent model to obtain inference data obtained by inferring the test data by the first intelligent model. The terminal calculates a loss value between the inference data and the check data. For example, the terminal may obtain a loss value between the inference data and the check data through calculation by using a loss function used when training the first intelligent model. However, the present application is not limited thereto. The terminal compares the loss value with the check threshold. If the loss value is less than the check threshold, the terminal may determine that the trained first intelligent model meets the requirements and can be applied to actual communication, and then notify the access network node by using the fifth information. Alternatively, if the loss value exceeds the check threshold, the terminal may notify the access network node that the trained first intelligent model does not meet the requirements by using the fifth information.
[0244] As another example, the test result may be a loss value between the inference data and the check data. For example, in addition to indicating the test data and the check data by the access network node by using the fourth information, the fourth information further indicates that the performance evaluation method is to calculate a loss value between the inference data and the check data. The terminal obtains inference data through inference by using the trained first intelligent model based on the test data indicated by the fourth information, calculates a loss value between the inference data and the label data, and notifies the access network node of the loss value by using the fifth information. The network determines whether the performance of the first intelligent model obtained through training by the terminal meets the requirements based on the loss value. For example, the access network node or a third party node in the network may determine whether the performance meets the requirements, and the third party node may obtain the loss value through forwarding by the access network node.
[0245] In another implementation, the terminal may send sixth information to the access network node, where the sixth information indicates the inference data of the first intelligent model.
[0246] For example, the fourth information indicates test data, and 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 an access network node. The network node (such as an access network node or a third party node in a network) determines whether the performance of the trained first intelligent model meets the requirements according to the inference data.
[0247] In another implementation, the terminal may transmit the fifth information and the sixth information to the access network node.
[0248] In other words, the terminal transmits the test result to the access network node and transmits the inference data to the access network node. The access network node can refer to the test result and the inference data to determine whether to use the first intelligent model obtained through training by the terminal in actual communication. However, the present disclosure is not limited thereto.
[0249] Optionally, the access network node sends seventh information to the terminal, where the seventh information indicates an updated training policy of the first intelligent model and / or indicates an updated configuration of the first intelligent model.
[0250] In response, the terminal receives seventh information from the access network node. The terminal re-trains the first intelligent model based on the seventh information.
[0251] In an implementation, after receiving the fifth information and / or the sixth information from the terminal, the access network node determines that the performance of the first intelligent model obtained through training by the terminal does not meet the requirements, in which case, the access network node may send the seventh information to the terminal to inform the terminal of an updated training policy of the first intelligent model and / or an updated configuration of the first intelligent model.
[0252] In another implementation, when the terminal uses the trained first intelligent model for communication, the access network node may determine that the first intelligent model needs to be updated based on the communication performance, and the access network node may send the seventh information to the terminal.
[0253] For example, after the terminal applies the model for a certain period of time, a situation occurs, for example, the channel environment may change, so that the performance of the first intelligent model no longer meets the requirements. The access network node may send the seventh information to the terminal, and then the terminal performs model training again on the first intelligent model, so that the first intelligent model obtained through retraining is adapted to the changed environment.
[0254] The seventh information may indicate the training policy in the same manner as the first information indicates the training policy, the seventh information may indicate the model configuration in the same manner as the second information indicates the model configuration, or the seventh information may specifically indicate at least one variation of the training policy and / or at least one variation of the configuration of the first intelligent model.
[0255] For example, the seventh information indicates at least one modification in the training policy. After receiving the seventh information, the terminal determines an updated training policy based on the modification indicated by the seventh information. For example, the access network node adjusts a parameter of an optimization algorithm (e.g., the learning rate is changed from 0.1 indicated by the first information to 0.01), and in the training policy, only the parameter of the optimization algorithm is changed, and the remaining training policy is not changed. In this case, the seventh information may indicate that the updated learning rate is 0.01. As another example, the first information indicates that the amount of data processed into a batch is N, in other words, N training data are selected each time for gradient descent update, and the access network node adjusts the amount of data processed into a batch to be M. In this case, the access network node may use the seventh information to indicate that the amount of data processed into a batch is adjusted to be M, or the seventh information may indicate that the amount of data processed into a batch is increased by Q, where Q=MN. After receiving the seventh information, the terminal increases the amount of data processed into a batch by Q to N+Q, so that M training data are selected for gradient descent update every time, although the present disclosure is not limited thereto.
[0256] Additionally / alternatively, the seventh information may indicate at least one variation in the model configuration. The specific manner is similar to the manner of indicating at least one variation in the training policy. For implementation, please refer to the above description. Details will not be described again in this specification.
[0257] Based on the above solution, after the terminal performs model training according to the training policy instructed by the access network node, the access network node can send test information to the terminal, and the terminal tests the trained first intelligent model according to the test information, and the terminal or the access network node can determine whether the performance of the trained first intelligent model meets the requirements according to the test result and / or inference data. If the performance meets the requirements, the first intelligent model can be applied to actual communication to achieve the purpose of improving wireless communication performance (such as communication reliability).
[0258] 6 is another schematic flow chart showing a communication method according to the present disclosure. It should be noted that the parts in the method shown in FIG. 6 and the same parts in another method may be mutually referenced or implemented in combination. For the sake of brevity, the details will not be described again in this specification.
[0259] S601: The access network node sends second information to the terminal, where the second information indicates a configuration of a first intelligent model, and the second information includes one or more of the following configuration information of the first intelligent model: configuration information of a network layer, a dimension of input data, or a dimension of output data.
[0260] In response, the terminal receives second information from the access network node.
[0261] The configuration information of the network layer in the first intelligent model may include, but is not limited to, one or more of the following:
[0262] The number of neural network layers included in the first intelligent model, the types of neural network layers, the manner in which the neural network layers are used, the cascade relationship between the neural network layers, the dimensions of the input data of the neural network layers, or the dimensions of the output data of the neural network layers.
[0263] S602: The terminal determines a first intelligent model based on the second information.
[0264] After receiving the second information, the terminal may generate, based on the configuration of the first intelligent model indicated by the second information, a first intelligent model having the configuration.
[0265] S603: The terminal performs model training on the first intelligent model.
[0266] After determining the first intelligent model in step S602, the terminal trains the first intelligent model.
[0267] Based on the above-mentioned solution, the access network node informs the terminal of the configuration of the first intelligent model, and the terminal generates a first intelligent model having the configuration based on the instruction of the access network node, so that the configuration of the first intelligent model used by the terminal can meet the requirements of the access network node. In this way, the first intelligent model can be used to improve wireless communication performance. After training the first intelligent model, the terminal can apply the first intelligent model to communication. This can reduce the air interface resource overhead that occurs when the access network node informs the terminal of the model parameters of the first intelligent model (such as the weights, activation functions, and offsets of each neuron, etc.).
[0268] 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 inform the terminal of the training policy of the first intelligent model by using the first information.
[0269] The training dataset used by the terminal to train the first intelligent model may be pre-stored by the terminal or may be received from an access network node.
[0270] The method provided in the present disclosure is described in detail above with reference to Figures 4 to 6. The following accompanying drawings will describe the communication device and communication equipment provided in the present disclosure. To implement the above-mentioned functions in the method provided in the present disclosure, each network element includes a hardware configuration and / or a software module, and may implement the above-mentioned functions in the form of a hardware configuration, a software module, or a combination of a hardware configuration and a software module. Whether a function in the above-mentioned functions is performed by using a hardware configuration, a software module, or a combination of a hardware configuration and a software module depends on the design constraints of a specific application and technical solution.
[0271] 7 is a schematic block diagram illustrating a communication device according to the present disclosure. As shown in FIG. 7, the communication device 700 may include a transceiver unit 720.
[0272] In a possible design, the communications device 700 may correspond to a terminal device in the methods described above, or a chip disposed in (or used in) the terminal device, or another device, module, circuit, unit, or the like capable of implementing the methods performed by the terminal device.
[0273] It should be understood that the communication device 700 may include units configured to perform the methods performed by the terminal equipment in the methods shown in Figures 4, 5C, and 6. Furthermore, the units in the communication device 700 and other operations and / or functions described above are intended to implement corresponding processes of the methods shown in Figures 4, 5C, and 6, respectively.
[0274] Optionally, the communication device 700 may further include a processing unit 710. The processing unit 710 may be configured to process instructions or data and perform corresponding operations.
[0275] It should be further understood that if the communication device 700 is a chip disposed in (or used in) a terminal device, the transceiver unit 720 in the communication device 700 may be an input / output interface or circuitry in the chip, and the processing unit 710 in the communication device 700 may be a processor in the chip.
[0276] Optionally, the communication device 700 may further include a storage unit 730. The storage unit 730 may be configured to store instructions or data. The processing unit 710 may execute the instructions or data stored in the storage unit to enable the communication device to implement corresponding operations.
[0277] It should be understood that the transceiver unit 720 in the communication device 700 may be implemented via a communication interface (e.g., a transceiver or an input / output interface, etc.) and may correspond, for example, to the transceiver 810 in the terminal device 800 shown in FIG. 8. The processing unit 710 in the communication device 700 may be implemented by using at least one processor and may correspond, for example, to the processor 820 in the terminal device 800 shown in FIG. 8. Alternatively, the processing unit 710 in the communication device 700 may be implemented by using at least one logic circuit. The storage unit 730 in the communication device 700 may correspond to a memory in the terminal device 800 shown in FIG. 8.
[0278] It should be further understood that the specific processes by which the units perform the corresponding procedures described above are described in detail in the methods described above, and for the sake of brevity, the details will not be described again herein.
[0279] In another possible design, the communications device 700 may correspond to an access network node in the methods described above, or a chip disposed in (or used in) an access network node, or another apparatus, module, circuit, unit, or the like capable of implementing the methods performed by the access network node.
[0280] It should be understood that the communication device 700 may include units configured to perform the methods performed by the access network nodes in the methods illustrated in Figures 4, 5C, and 6. Furthermore, the units in the communication device 700 and other operations and / or functions described above are intended to implement corresponding processes of the methods illustrated in Figures 4, 5C, and 6, respectively.
[0281] Optionally, the communication device 700 may further include a processing unit 710. The processing unit 710 may be configured to process instructions or data and perform corresponding operations.
[0282] It should be further appreciated that if the communications device 700 is a chip located (or used) in an access network node, the transceiver unit 720 in the communications device 700 may be an input / output interface or circuitry in the chip, and the processing unit 710 in the communications device 700 may be a processor in the chip.
[0283] Optionally, the communication device 700 may further include a storage unit 730. The storage unit 730 may be configured to store instructions or data. The processing unit 710 may execute the instructions or data stored in the storage unit to enable the communication device to implement corresponding operations.
[0284] It should be understood that if the communication device 700 is an access network node, the transceiver unit 720 in the communication device 700 may be implemented via a communication interface (e.g., a transceiver or an input / output interface, etc.) and may correspond, for example, to the transceiver 910 in the network device 900 shown in FIG. 9. The processing unit 710 in the communication device 700 may be implemented by using at least one processor and may correspond, for example, to the processor 920 in the network device 900 shown in FIG. 9. Alternatively, the processing unit 710 in the communication device 700 may be implemented by using at least one logic circuit. The storage unit 730 in the communication device 700 may correspond to a memory in the network device 900 shown in FIG. 9.
[0285] It should be further understood that the specific processes by which the units perform the corresponding procedures described above are described in detail in the methods described above, and for the sake of brevity, the details will not be described again herein.
[0286] FIG. 8 is a diagram showing a configuration of a terminal device 800 according to the present disclosure. The terminal device 800 may be used in the system shown in FIG. 1 to perform the functions of the terminal device in the above-mentioned method. As shown in the figure, the terminal device 800 includes a processor 820 and a transceiver 810. Optionally, the terminal device 800 further includes a memory. The processor 820, the transceiver 810, and the memory may communicate with each other via an internal connection path to perform transmission of control signals and / or data signals. The memory is configured to store a computer program, and the processor 820 is configured to execute the computer program in the memory to control the transceiver 810 to receive and transmit signals.
[0287] The processor 820 may be configured to perform operations implemented within the terminal device and described in the above-mentioned method. The transceiver 810 may be configured to perform operations transmitted or received by the terminal device to or from the network device and described in the above-mentioned method. For details, please refer to the description of the above-mentioned method. For details, the present specification will not be repeated.
[0288] Optionally, the terminal device 800 may further include a power source configured to provide power to various components or circuits within the terminal device.
[0289] FIG. 9 is a diagram illustrating a configuration of a network device 900 according to the present disclosure. The network device 900 may be used in the system illustrated in FIG. 1 to perform the functions of the network device in the above-described method. As illustrated in the figure, the network device 900 includes a processor 920 and a transceiver 910. Optionally, the network device 900 further includes a memory. The processor 920, the transceiver 910, and the memory may communicate with each other via an internal connection path to perform transmission of control signals and / or data signals. The memory is configured to store a computer program, and the processor 920 is configured to execute the computer program in the memory to control the transceiver 910 to receive and transmit signals.
[0290] The processor 920 may be configured to perform operations implemented within the network device and described in the above-mentioned method. The transceiver 910 may be configured to perform operations transmitted or received by the network device to or from the terminal device and described in the above-mentioned method. For details, please refer to the description of the above-mentioned method. For details, the present specification will not be described again.
[0291] Optionally, the network device 900 may further include a power supply configured to provide power to various components or circuits within the network device.
[0292] In the terminal device shown in FIG. 8 and the network device shown in FIG. 9, the processor and the memory may be combined into one processing unit, and the processor is configured to execute the program code stored in the memory to implement the above-mentioned functions. In a particular implementation, the memory may be integrated into the processor or may be separate from 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 referred to as a receiving machine or a receiving circuit) and a transmitter (or referred to as a transmitting machine or a transmitting circuit). The receiver is configured to receive a signal, and the transmitter is configured to transmit a signal.
[0293] In this disclosure, 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 that may implement or execute the methods, procedures, and logic block diagrams in this disclosure. A general purpose processor may be a microprocessor, any conventional processor, or the like. The procedures of the methods in this disclosure may be executed directly by a hardware processor, or may be executed by a combination of hardware and software modules in a processor.
[0294] In the present disclosure, the memory may be a non-volatile memory, such as a hard disk drive (HDD) or a solid-state drive (SSD), or a volatile memory, such as a random access memory (RAM). The memory is, but is not limited to, any other medium capable of transmitting or storing expected program code in the form of instructions or data structures and accessible by a computer. Alternatively, the memory in the present application may be a circuit or any other device capable of implementing a storage function and configured to store program instructions and / or data.
[0295] The present disclosure further provides a processing device comprising a processor and a (communication) interface, the processor being configured to perform a method according to any one of the methods described above.
[0296] It should be understood that a processing device may be one or more chips, For example, a processing device may be a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a system on chip (SoC), a central processing unit (CPU), a network processor (NP), a digital signal processor (DSP), a microcontroller unit (MCU), a programmable logic device (PLD), or another integrated chip.
[0297] According to the method provided in the present disclosure, the present disclosure further provides a computer program product, which includes computer program code, which, when executed by one or more processors, enables an apparatus including the processors to perform the methods shown in Figures 4, 5C, and 6.
[0298] The technical solutions provided in this disclosure may be implemented in whole or in part by using software, hardware, firmware, or any combination thereof. When software is used to implement the technical solutions, the technical solutions 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. When the computer program instructions are loaded and executed on a computer, the process or function according to the present invention is generated in whole or in part. The 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. The computer-readable storage medium may be any usable medium that is accessible by a computer, or may be a data storage device that integrates one or more usable media, such as a server or a data center. The usable medium may be a magnetic medium (e.g., a floppy disk, a hard disk, or a magnetic tape, etc.), an optical medium (e.g., a digital video disc (DVD), etc.), a semiconductor medium, or the like.
[0299] According to the method provided in the present disclosure, the present disclosure further provides a computer-readable storage medium. The computer-readable storage medium stores program code. When the program code is executed by one or more processors, the device including the processors is capable of performing the methods shown in Figures 4, 5C, and 6.
[0300] In accordance with the methods provided herein, the present disclosure further provides a system including one or more of the devices described above.
[0301] In some embodiments provided in the present disclosure, it should be understood that the disclosed system, device, and method may be implemented in other ways. For example, the described device is merely an example. For example, the division into units is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined with another system or integrated into another system, or some functions may be ignored or not performed. Furthermore, the shown or described mutual couplings, or direct couplings, or communication connections may be implemented by using some interfaces. Indirect couplings or communication connections between devices or units may be implemented in electronic, mechanical, or other forms.
[0302] The units described as separate parts may or may not be physically separate, and the parts shown as units may or may not be physical units, located in one location or distributed across multiple network units. Some or all of the units may be selected based on actual requirements to achieve the objectives of the solution.
[0303] The above description is merely a specific implementation of the present disclosure, but is not intended to limit the scope of protection of the present disclosure. Any variation or replacement that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present disclosure shall be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure shall be subject to the scope of protection of the claims.
Claims
1. 1. A communication method comprising: receiving first information, the first information indicating a training policy for a first intelligent model; performing model training on the first intelligent model according to the training policy; A method comprising:
2. The training policy includes: Model training method, loss function information, model initialization method, type of model optimization algorithm, or parameters of the optimization algorithm The method of claim 1 , comprising one or more of:
3. the model optimization algorithm is an adaptive moment estimation algorithm, a stochastic gradient descent algorithm, or a batch gradient descent algorithm; and / or The method of claim 2 , wherein parameters of the optimization algorithm include one or more of a learning rate, a number of iterations, or an amount of data processed in a batch.
4. The method comprises: obtaining second information, the second information indicative of a configuration of the first intelligent model; The method of claim 1 , further comprising:
5. The second information is the following configuration information of the first intelligent model: Network layer configuration information, input data dimensions, or output data dimensions The method of claim 4, further comprising one or more of:
6. The network layer configuration information includes: The number of neural network layers configured in the first intelligent model, the type of the neural network layers, the usage manner of the neural network layers, the cascade relationship between the neural network layers, the dimension of the input data of the neural network layers, or the dimension of the output data of the neural network layers. The method of claim 5 , comprising one or more of:
7. The method comprises: receiving third information, the third information indicating information regarding a training dataset, the training dataset being used to train the first intelligent model; The method of claim 1 , further comprising:
8. The method of claim 7 , wherein the training dataset includes training samples or includes training samples and labels.
9. The method comprises: transmitting capability information, the capability information indicating a capability for running the intelligent model; The method of claim 1 , further comprising:
10. The capability information includes the following capabilities: Whether the operation of the intelligent model is supported, the type of the intelligent model that can be operated, data processing capability, or storage capability The method of claim 9, further comprising one or more of:
11. The method comprises: receiving fourth information, the fourth information indicating test information, the test information being used to test performance of the first intelligent model; The method of claim 1 , further comprising:
12. The test information comprises: Test data information, performance evaluation method, or performance evaluation parameters The method of claim 11 , comprising one or more of:
13. The method transmitting fifth information, the fifth information indicating a test result of the first intelligent model; and / or transmitting sixth information, the sixth information indicating inference data, the inference data being obtained by inferring test data for the first intelligent model, and the test information including the test data; Further comprising:
13. The method according to claim 11 or 12.
14. The method comprises: receiving seventh information, the seventh information indicating an updated training policy for the first intelligent model and / or indicating an updated configuration for the first intelligent model; The method of claim 13 further comprising:
15. The method of claim 14 , wherein the seventh information specifically indicates at least one variation in a training policy of the first intelligent model and / or at least one variation in a configuration of the first intelligent model.
16. 1. A communication method comprising: Transmitting first information, the first information indicating a training policy for the first intelligent model. A method comprising:
17. The training policy includes: Model training method, loss function information, model initialization method, type of model optimization algorithm, or parameters of the optimization algorithm 17. The method of claim 16, comprising one or more of:
18. the type of model optimization algorithm is an adaptive moment estimation algorithm, a stochastic gradient descent algorithm, or a batch gradient descent algorithm; and / or The parameters of the optimization algorithm include one or more of a learning rate, a number of iterations, or an amount of data processed in a batch.
20. The method of claim 17.
19. The method comprises: transmitting second information, the second information indicating a configuration of the first intelligent model; 19. The method of any one of claims 16 to 18, further comprising:
20. The second information includes the following configuration information of the first intelligent model: Network layer configuration information, input data dimensions, or output data dimensions 20. The method of claim 19, specifically indicating one or more of:
21. The network layer configuration information includes: The number of neural network layers configured in the first intelligent model, the type of the neural network layers, the usage manner of the neural network layers, the cascade relationship between the neural network layers, the dimension of the input data of the neural network layers, or the dimension of the output data of the neural network layers.
21. The method of claim 20, comprising one or more of:
22. The method comprises: transmitting third information, the third information indicating information regarding a training dataset, the training dataset being used to train the first intelligent model; 22. The method of any one of claims 16 to 21, further comprising:
23. 23. The method of claim 22, wherein the training dataset comprises training samples, or comprises training samples and labels.
24. The method comprises: receiving capability information, the capability information indicating a capability for running an intelligent model; 24. The method of any one of claims 16 to 23, further comprising:
25. The capability information includes the following capabilities: Whether the operation of the intelligent model is supported, the type of the intelligent model that can be operated, data processing capability, or storage capability 25. The method of claim 24, wherein the method further comprises one or more of:
26. The method comprises: Sending fourth information, the fourth information indicating test information for testing a performance of the first intelligent model.
26. The method of any one of claims 16 to 25, further comprising:
27. The test information comprises: Test data information, performance evaluation method, or performance evaluation parameters 27. The method of claim 26, comprising one or more of:
28. The method comprises: receiving fifth information, the fifth information being indicative of a test result of the first intelligent model; and / or receiving sixth information, the sixth information being indicative of inference data, the inference data being obtained by inference of test data by the first intelligent model, and the test information including the test data; 28. The method of claim 26 or 27, further comprising:
29. The method comprises: transmitting seventh information, the seventh information indicating an updated training policy of the first intelligent model and / or indicating an updated configuration of the first intelligent model.
30. The method of claim 28, further comprising:
30. 30. The method of claim 29, wherein the seventh information specifically indicates a modification of at least one parameter in the training policy and / or a modification of at least one parameter in the configuration of the first intelligent model.
31. 1. A communication method comprising: receiving second information, the second information indicating a configuration of a first intelligent model, the second information including configuration information of the first intelligent model, i.e., one or more of network layer configuration information, input data dimension, or output data dimension; determining a configuration of the first intelligent model based on the second information; performing model training on the first intelligent model; A method comprising:
32. The method comprises: Obtaining first information, the first information indicating a training policy for the first intelligent model. Further equipped with The step of performing model training on the first intelligent model includes: performing model training on the first intelligent model according to the training policy. Including, 32. The method of claim 31 .
33. The training policy includes: Model training method, loss function information, model initialization method, type of model optimization algorithm, or parameters of the optimization algorithm 33. The method of claim 32, comprising one or more of:
34. The method comprises: receiving third information, the third information indicating information regarding a training dataset, the training dataset being used to train the first intelligent model; 34. The method of any one of claims 31 to 33, further comprising:
35. 35. The method of claim 34, wherein the training dataset comprises training samples, or comprises training samples and labels.
36. The method comprises: transmitting capability information, the capability information indicating a capability for running the intelligent model; 36. The method of any one of claims 31 to 35, further comprising:
37. The capability information includes the following capabilities: Whether the operation of the intelligent model is supported, the type of the intelligent model that can be operated, data processing capability, or storage capability 37. The method of claim 36, wherein the method further comprises one or more of:
38. The method comprises: receiving fourth information, the fourth information indicating test information, the test information being used to test performance of the first intelligent model; 38. The method of any one of claims 31 to 37, further comprising:
39. The test information comprises: Test data information, performance evaluation method, or performance evaluation parameters 40. The method of claim 38, comprising one or more of:
40. The method comprises: transmitting fifth information, the fifth information indicating a test result of the first intelligent model; and / or transmitting sixth information, the sixth information indicating inference data, the inference data being obtained by inferring test data by the first intelligent model, and the test information including the test data; 40. The method of claim 38 or 39, further comprising:
41. The method comprises: receiving seventh information, the seventh information indicating an updated training policy of the first intelligent model and / or indicating an updated configuration of the first intelligent model; 41. The method of any one of claims 31 to 40, further comprising:
42. 42. The method of claim 41, wherein the seventh information specifically indicates at least one variation in a training policy of the first intelligent model and / or at least one variation in a configuration of the first intelligent model.
43. A communication method sending second information, the second information indicating a configuration of a first intelligent model, the second information including one or more of the following configuration information of the first intelligent model: configuration information of a network layer, a dimension of input data, or a dimension of output data; A method comprising:
44. The method comprises: Transmitting first information, the first information indicating a training policy for the first intelligent model.
44. The method of claim 43, further comprising:
45. The training policy includes: Model training method, loss function information, model initialization method, type of model optimization algorithm, or parameters of the optimization algorithm 45. The method of claim 44, comprising one or more of:
46. The method comprises: transmitting third information, the third information indicating information regarding a training dataset, the training dataset being used to train the first intelligent model; 46. The method of any one of claims 43 to 45, further comprising:
47. 47. The method of claim 46, wherein the training dataset comprises training samples, or comprises training samples and labels.
48. The method comprises: receiving capability information, the capability information indicating a capability for running an intelligent model; 48. The method of any one of claims 43 to 47, further comprising:
49. The capability information includes the following capabilities: Whether the operation of the intelligent model is supported, the type of the intelligent model that can be operated, data processing capability, or storage capability 44. The method of claim 43, wherein the method further comprises one or more of:
50. The method comprises: sending fourth information, the fourth information indicating test information, the test information being used to test performance of the first intelligent model; 50. The method of any one of claims 43 to 49, further comprising:
51. The test information comprises: Test data information, performance evaluation method, or performance evaluation parameters 51. The method of claim 50, comprising one or more of:
52. The method comprises: receiving fifth information, the fifth information being indicative of a test result of the first intelligent model; and / or receiving sixth information, the sixth information indicative of inference data, the inference data being obtained by inferring test data for the first intelligent model, and the test information including the test data; 52. The method of claim 50 or 51, further comprising:
53. The method comprises: transmitting seventh information, the seventh information indicating an updated training policy of the first intelligent model and / or indicating an updated configuration of the first intelligent model.
53. The method of any one of claims 43 to 52, further comprising:
54. 54. The method of claim 53, wherein the seventh information specifically indicates at least one variation in the training policy and / or at least one variation in a configuration of the first intelligent model.
55. A communications device, the communications device being configured to implement a method according to any one of claims 1 to 15 or claims 31 to 42.
56. A communications device comprising a processor and a memory, the memory coupled to the processor, the processor configured to perform a method according to any one of claims 1 to 15 or claims 31 to 42.
57. A communications device comprising a processor and a communications interface, the processor being operable via the communications interface to execute a method according to any one of claims 1 to 15 or claims 31 to 42.
58. A communications device, the communications device being configured to implement a method according to any one of claims 16 to 30 or claims 43 to 54.
59. A communications device comprising a processor and a memory, said memory coupled to said processor, said processor configured to perform a method according to any one of claims 16 to 30 or claims 43 to 54.
60. A communications device comprising a processor and a communications interface, the processor executing a method according to any one of claims 16 to 30 or claims 43 to 54 via the communications interface.
61. A communication system comprising a communication device according to any one of claims 55 to 57 and a communication device according to any one of claims 58 to 60.
62. 55. A computer readable storage medium storing instructions which, when executed on a computer, enable the computer to carry out a method according to any one of claims 1 to 54.
63. 55. A computer program product comprising instructions which, when executed on a computer, enable the computer to carry out a method according to any one of claims 1 to 54.
Citation Information
Patent Citations
Methods and apparatus for generating models
CN107766940B
Methods, devices, systems and servers for quality assessment of care services
CN107944770B
Ability imparting data generation device
CN108701264A
Auditable outsourcing machine learning service method for providing integrity verification
CN112488227A
Learning control method and computer system
JP2019219741A
