WO2023185890A1 - 一种数据处理方法及相关装置 - Google Patents

一种数据处理方法及相关装置 Download PDF

Info

Publication number
WO2023185890A1
WO2023185890A1 PCT/CN2023/084530 CN2023084530W WO2023185890A1 WO 2023185890 A1 WO2023185890 A1 WO 2023185890A1 CN 2023084530 W CN2023084530 W CN 2023084530W WO 2023185890 A1 WO2023185890 A1 WO 2023185890A1
Authority
WO
WIPO (PCT)
Prior art keywords
terminals
processing module
dimensional data
network device
communication
Prior art date
Application number
PCT/CN2023/084530
Other languages
English (en)
French (fr)
Inventor
吴佳骏
孙乘坚
杨晨阳
王坚
李榕
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Publication of WO2023185890A1 publication Critical patent/WO2023185890A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the present application relates to the field of neural networks, and in particular, to a data processing method and related devices.
  • AI artificial intelligence
  • the neural network can be trained under the condition that the number of devices is a preset number, and a neural network can be obtained to solve the problem to be solved.
  • the relationship between the output and input of the problem to be solved changes as the number of devices changes.
  • the neural network without retraining is still used, the accuracy of the solution to the problem to be solved will be lower, resulting in a greater performance loss in the system. Therefore, it is often necessary to retrain the neural network or pre-train multiple models to adapt to different numbers of devices, which results in high storage and processing overhead of the system.
  • Embodiments of the present application provide a data processing method and related devices, which can ensure the accuracy of solutions to problems to be solved with low system overhead.
  • embodiments of the present application provide a data processing method, which can be applied to a first device.
  • the method includes obtaining K-dimensional data, inputting the K-dimensional data into a first machine learning model, and obtaining a solution to the problem to be solved.
  • the first machine learning model includes a first processing module and a second processing module.
  • the second processing module is determined based on the constraints of the problem to be solved.
  • the second processing module is used to perform dimensional generalization on K-dimensional data.
  • the first processing module is trained based on m-dimensional data.
  • the value of m has nothing to do with the value of K.
  • K and m are positive integers.
  • a second processing module when solving the problem to be solved, is used to perform dimensional generalization on K-dimensional data, thereby enhancing the dimensional generalization ability of K-dimensional data.
  • a second processing module obtained by pre-training based on m dimensions is used. The first processing module does not need to retrain the first processing module due to changes in the dimensions of the input data, and can ensure the accuracy of the solution to the problem to be solved with low system overhead.
  • the first device may obtain K-dimensional data through channel estimation.
  • the K-dimensional data is obtained by performing channel estimation on the terminal, and the first device receives the K-dimensional data from the terminal.
  • the problem to be solved does not have dimensional generalization properties.
  • the first processing module obtained by training based on m-dimensional data is generalized to K-dimensional data. Then, if the problem to be solved is solved based on the first processing module, the accuracy of the obtained solution to the problem to be solved is low. Therefore, solving the problem to be solved based on the first processing module obtained by pre-training and the second processing module that performs dimensional generalization on K-dimensional data can improve the dimensional generalization ability of K-dimensional data and obtain a more accurate solution.
  • K-dimensional data is input into the first machine learning model to obtain a solution to the problem to be solved, including It includes: inputting K-dimensional data into the first processing module to obtain K first intermediate solutions; inputting K first intermediate solutions into the second processing module to obtain the solution to the problem to be solved.
  • K-dimensional data is input into the pre-trained first processing module to obtain K first intermediate solutions, and then the K first intermediate solutions are dimensionalized through the second processing module that performs dimensional generalization.
  • Generalization thereby achieving dimensional generalization of K-dimensional data, which can make the solution more accurate.
  • the problem to be solved is the power of communication signals sent by the network device to K terminals when the total bandwidth used by a network device to communicate with K terminals is minimized.
  • the constraints of the problem to be solved are: The total power including the communication signals sent by the network device to the K terminals is within the first range.
  • the problem to be solved is the power allocation problem of the network device sending communication signals to K terminals when a network device performs downlink communication with K terminals.
  • the problem to be solved is the power of K terminals sending communication signals to the network device when the total bandwidth used by K terminals to communicate with a network device is minimized.
  • the constraints of the problem to be solved are The total power including the K terminals sending communication signals to the network device is within the third range. That is to say, the problem to be solved is the power allocation problem of K terminals sending communication signals to the network device when K terminals communicate with a network device in uplink.
  • the second processing module when the problem to be solved is the power allocation problem of uplink or downlink communication between the K terminals and a network device, the second processing module includes a normalized exponential function activation layer. That is to say, when the first device solves the power allocation problem, it performs dimensional generalization on the K-dimensional data through the normalized exponential function activation layer.
  • the problem to be solved is the bandwidth used by the network device to communicate with K terminals when the total bandwidth used by a network device to communicate with K terminals is minimized.
  • the problem to be solved is The constraint includes that the service quality when the network device communicates with each of the K terminals is within the second range.
  • the problem to be solved is the bandwidth allocation problem used when a network device performs uplink communication or downlink communication with K terminals.
  • the above-mentioned second processing module when the problem to be solved is the bandwidth allocation problem used by a network device to perform uplink communication or downlink communication with K terminals, the above-mentioned second processing module includes an activation layer and a scaling factor layer.
  • the kth scaling factor in the scaling factor layer is obtained by inputting the kth data and K into the scaling factor calculation module, where k is a positive integer less than or equal to K. It can be seen that when solving this bandwidth allocation problem, the K-dimensional data is dimensionally generalized through the activation layer and the scaling factor layer.
  • the K first intermediate solutions are input into the second processing module to obtain the bandwidth allocation problem to be solved.
  • Solving the solution to the problem includes: inputting K first intermediate solutions into the activation layer to obtain K second intermediate solutions; inputting K second intermediate solutions into the scaling factor layer to obtain the solution to the problem to be solved.
  • the K first intermediate solutions are then dimensionally generalized based on the activation layer and the scaling factor layer to obtain the solution to the bandwidth allocation problem.
  • the K-dimensional data is the channel gain between the network device and K terminals.
  • the K channel gains may be obtained by the first device through channel estimation.
  • the above-mentioned first processing module is a permutation equivariant neural network or a graph neural network with permutation equivariant characteristics. This method allows the first device to use the first processing module when solving a problem to be solved that has permutation equivariant characteristics.
  • this application also provides a communication device.
  • the communication device has the ability to implement part or all of the functions of the first device described in the first aspect.
  • the functions of the communication device may have the functions of some or all of the embodiments of the first device described in the first aspect of this application, or may have the functions of independently implementing any of the embodiments of this application.
  • the functions described can be implemented by hardware, or can be implemented by hardware executing corresponding software.
  • the hardware or software includes one or more units or modules corresponding to the above functions.
  • the structure of the communication device may include a processing unit and a communication unit, and the processing unit is configured to support the communication device to perform corresponding functions in the above method.
  • the communication unit is used to support communication between the communication device and other communication devices.
  • the communication device may further include a storage unit coupled to the processing unit and the communication unit, which stores necessary program instructions and data for the communication device.
  • the communication device includes: a processing unit and a communication unit, the communication unit is used to send and receive data/signaling;
  • Processing unit used to obtain K-dimensional data
  • a processing unit also configured to input the K-dimensional data into a first machine learning model to obtain a solution to the problem to be solved;
  • the first machine learning model includes a first processing module and a second processing module;
  • the second processing module is determined based on the constraints of the problem to be solved; the second processing module is used to perform dimensional generalization on the K-dimensional data; the first processing module is trained based on m-dimensional data Obtained; the value of m has nothing to do with the value of K; the K and m are positive integers.
  • the communication unit may be a transceiver or a communication interface
  • the storage unit may be a memory
  • the processing unit may be a processor
  • the communication device includes: a processor and a transceiver, and the transceiver is used to transmit and receive data/signaling;
  • a processor also configured to input the K-dimensional data into a first machine learning model to obtain a solution to the problem to be solved;
  • the first machine learning model includes a first processing module and a second processing module;
  • the second processing module is determined based on the constraints of the problem to be solved; the second processing module is used to perform dimensional generalization on the K-dimensional data; the first processing module is trained based on m-dimensional data Obtained; the value of m has nothing to do with the value of K; the K and m are positive integers.
  • the communication device is a chip or a chip system.
  • the processing unit can also be embodied as a processing circuit or a logic circuit; the transceiver unit can be an input/output interface, interface circuit, output circuit, input circuit, pin or related circuit on the chip or chip system.
  • the processor may be used to perform, for example, but not limited to, baseband related processing
  • the transceiver may be used to perform, for example, but not limited to, radio frequency transceiver.
  • the above-mentioned devices may be arranged on separate chips, or at least part or all of them may be arranged on the same chip.
  • processors can be further divided into analog baseband processors and digital baseband processors.
  • the analog baseband processor can be integrated with the transceiver on the same chip, and the digital baseband processor can be set on an independent chip. With the continuous development of integrated circuit technology, more and more devices can be integrated on the same chip.
  • the digital baseband processor can be integrated with a variety of application processors (such as but not limited to graphics processors, multimedia processors, etc.) on the same chip.
  • application processors such as but not limited to graphics processors, multimedia processors, etc.
  • SoC system on a chip
  • the embodiments of this application do not limit the implementation form of the above devices.
  • this application also provides a processor for executing the various methods mentioned above.
  • the process of sending the above information and receiving the above information in the above method can be understood as the process of the processor outputting the above information, and the process of the processor receiving the input above information.
  • the processor When outputting the above information, the processor outputs the above information to the transceiver for transmission by the transceiver. After the above information is output by the processor, it may also need to be Other processing is performed before reaching the transceiver.
  • the processor receives the above information input, the transceiver receives the above information and inputs it into the processor. Furthermore, after the transceiver receives the above information, the above information may need to undergo other processing before being input to the processor.
  • processor output and reception, input operations rather than the transmitting and receiving operations performed directly by RF circuits and antennas.
  • the above-mentioned processor may be a processor specifically designed to perform these methods, or may be a processor that executes computer instructions in a memory to perform these methods, such as a general-purpose processor.
  • the above-mentioned memory can be a non-transitory memory, such as a read-only memory (Read Only Memory, ROM), which can be integrated on the same chip with the processor, or can be separately provided on different chips.
  • ROM Read Only Memory
  • this application also provides a communication system, which includes one or more network devices and one or more terminal devices.
  • the system may also include other devices that interact with network devices and terminal devices.
  • the present application provides a computer-readable storage medium for storing instructions that, when executed by a computer, implement the method described in any one of the above first aspects.
  • the present application also provides a computer program product including instructions that, when run on a computer, implement the method described in any one of the above first aspects.
  • the application provides a chip system.
  • the chip system includes a processor and an interface.
  • the interface is used to obtain a program or instructions.
  • the processor is used to call the program or instructions to implement or support the first device.
  • Implement the functions involved in the first aspect For example, at least one of the data and information involved in the above method is determined or processed.
  • the chip system further includes a memory, and the memory is used to store necessary program instructions and data for the terminal.
  • the chip system may be composed of chips, or may include chips and other discrete devices.
  • Figure 1 is a schematic system structure diagram of a communication system provided by an embodiment of the present application.
  • Figure 2 is a schematic structural diagram of a fully connected neural network provided by an embodiment of the present application.
  • Figure 3 is a schematic diagram of a neural network training method provided by an embodiment of the present application.
  • Figure 4 is a schematic diagram of gradient reverse transmission provided by an embodiment of the present application.
  • Figure 5 is a schematic diagram of a dimensional generalization feature provided by an embodiment of the present application.
  • Figure 6 is a schematic structural diagram of a deep neural network provided by an embodiment of the present application.
  • Figure 7 is a schematic flowchart of a data processing method provided by an embodiment of the present application.
  • Figure 8 is a schematic structural diagram of a first machine learning model provided by an embodiment of the present application.
  • FIG. 9 is a schematic flowchart of another data processing method provided by an embodiment of the present application.
  • Figure 10 is a schematic structural diagram of another first machine learning model provided by an embodiment of the present application.
  • FIG 11 is a schematic flowchart of yet another data processing method provided by an embodiment of the present application.
  • Figure 12 is a schematic structural diagram of yet another first machine learning model provided by an embodiment of the present application.
  • Figure 13 is a schematic structural diagram of a scaling factor calculation module provided by an embodiment of the present application.
  • Figure 14 is a schematic structural diagram of a communication device provided by an embodiment of the present application.
  • Figure 15 is a schematic structural diagram of another communication device provided by an embodiment of the present application.
  • Figure 16 is a schematic structural diagram of a chip provided by an embodiment of the present application.
  • Embodiments of the present application can be applied to fifth generation mobile communication (5G) systems, satellite communications, short-range and other wireless communication systems.
  • the system architecture is shown in Figure 1.
  • a wireless communication system may include one or more network devices and one or more terminal devices.
  • Wireless communication systems can also perform point-to-point communication, such as communication between multiple terminal devices.
  • the wireless communication systems mentioned in the embodiments of this application include but are not limited to: narrowband-internet of things (NB-IoT), long term evolution (LTE), 5G mobile communications
  • NB-IoT narrowband-internet of things
  • LTE long term evolution
  • 5G mobile communications The three major application scenarios of the system: enhanced mobile broadband (eMBB), ultra-reliable low latency communication (URLLC) and massive machine type of communication (mMTC), wireless security True (wireless fidelity, WiFi) system, or mobile communication system after 5G, etc.
  • eMBB enhanced mobile broadband
  • URLLC ultra-reliable low latency communication
  • mMTC massive machine type of communication
  • wireless security True wireless fidelity, WiFi
  • mobile communication system after 5G etc.
  • the network device is a device with wireless transceiver function, used to communicate with the terminal device, and can be an evolved base station (evolved Node B, eNB or eNodeB) in LTE; or a base station in the 5G network or Base stations in future evolved public land mobile networks (PLMN), broadband network gateways (BNG), aggregation switches or non-3rd generation partnership project (3GPP) interfaces Enter equipment, etc.
  • the network equipment in the embodiments of this application may include various forms of base stations, such as: macro base stations, micro base stations (also called small stations), relay stations, access points, future equipment that implements base station functions, and WiFi systems.
  • Access nodes transmitting and receiving point (TRP), transmitting point (TP), mobile switching center and device-to-device (D2D), vehicle outreach (vehicle- to-everything (V2X), machine-to-machine (M2M) communication, equipment that assumes base station functions, etc., the embodiments of the present application do not specifically limit this.
  • Network equipment can communicate and interact with core network equipment and provide communication services to terminal equipment.
  • the core network equipment is, for example, equipment in the 5G network core network (core network, CN).
  • core network As a bearer network, the core network provides an interface to the data network, provides terminals with communication connections, authentication, management, policy control, and carries data services.
  • the terminal devices involved in the embodiments of this application may include various handheld devices with wireless communication functions, vehicle-mounted devices, wearable devices, computing devices, or other processing devices connected to wireless modems. Terminal devices may also be called terminals. Terminal equipment can also refer to user equipment (UE), access terminal, subscriber unit, user agent, cellular phone, smart phone, wireless data card, personal digital assistant ( Personal digital assistant (PDA) computers, tablet computers, wireless modems, handsets, laptop computers, machine type communication (MTC) terminals, and high-altitude aircraft Communication equipment, wearable devices, drones, robots, terminals in device-to-device (D2D) communication, terminals in vehicle to everything (V2X), virtual reality, VR) terminal equipment, augmented reality (AR) terminal equipment, wireless terminals in industrial control, wireless terminals in self-driving, remote medicine Wireless terminals in medical, wireless terminals in smart grid, wireless terminals in transportation safety, wireless terminals in smart city, and wireless terminals in smart home Or terminal equipment in future communication networks, etc., this application does not limit
  • the terminal device may also have AI processing capabilities
  • the network device may also have AI processing capabilities
  • the terminal device can have neural network training capabilities, reasoning capabilities, etc.
  • the first device in the embodiment of this application may be the above-mentioned network device or the above-mentioned terminal device, that is, it may be the network device or the terminal device that executes the data processing method proposed in the embodiment of this application.
  • the first device is a device other than network equipment and terminal equipment.
  • the first device executes the data processing method proposed in the embodiment of this application, obtains a solution to the problem to be solved, and then sends the solution to the network device or terminal device. solution to the problem.
  • the optimal in the embodiments of this application refers to the optimal under certain conditions.
  • the "optimal resource allocation strategy" in the embodiment of this application refers to the "optimal resource allocation strategy under certain conditions.”
  • power allocation problem is equivalent to “power allocation strategy”
  • bandwidth allocation problem is equivalent to “bandwidth allocation strategy”.
  • Fully connected neural network is also called multilayer perceptron (MLP).
  • MLP multilayer perceptron
  • an MLP includes an input layer, an output layer, and multiple hidden layers, and each layer includes multiple nodes, which are called neurons. Among them, neurons in two adjacent layers are connected in pairs.
  • the output h of the neuron in the next layer is the weighted sum of all the neurons x in the previous layer connected to it and passes through the activation function.
  • a neural network can be understood as a mapping relationship from an input data set to an output data set.
  • the training of neural network refers to the process of using existing data to obtain the mapping relationship of the above formula (2) from random w and b.
  • the specific method of training the neural network is to use the loss function (loss function) to evaluate the output results of the neural network, and reversely transmit the error, and iteratively optimize w and b through the gradient descent method until the loss function reaches the minimum value.
  • the gradient descent process can be expressed as:
  • ⁇ is the parameter to be optimized, such as ⁇ is w and b, and L is the loss function.
  • eta is the learning efficiency, which is used to control the step size of gradient descent.
  • the process of reverse transmission utilizes the chain rule of partial derivatives, that is, the gradient of the previous layer parameters can be calculated recursively from the gradient of the subsequent layer parameters.
  • the gradient of the weight w ij between neuron j and neuron i in Figure 4 can be expressed as:
  • PE Permutation equivalence
  • Neural networks with permutation equivalence characteristics include permutation equivalence neural network (PENN) and graph neural network (GNN). Then, if a multivariable function represents a strategy that is independent of the order in which objects are arranged, the strategy is called a permutation equivariant strategy.
  • PNN permutation equivalence neural network
  • NNN graph neural network
  • the dimensional generalization property refers to the property that when the number of variables K changes, the relationship between the output y k and the input x k does not change. It is understandable that when the number (dimension) K of input data in the problem to be solved changes, the relationship between the output y k and the input x k does not change, then the problem to be solved can be said to have dimensional generalization characteristics.
  • Wireless resource allocation based on optimization problems refers to how to achieve optimal resource allocation to multiple terminals under certain constraints. For example, for a wireless resource allocation problem that contains K terminals and has permutation equivariant characteristics, the model of the wireless resource allocation problem is:
  • the optimal resource allocation strategy P* in this resource allocation problem can be regarded as a function of the environmental state h.
  • the optimal strategy of each terminal in P* will also change, so the resource allocation strategy has the above-mentioned substitution equivariant characteristics.
  • a deep neural network can fit any function with arbitrary accuracy. Therefore, network devices can adopt policy neural networks Fit the functional relationship between the optimal resource allocation strategy and the environmental state. They are the weight parameters and bias parameters of the policy neural network respectively.
  • W l represents the connection weight between layer l-1 and layer l
  • b l represents the bias of layer l
  • h represents the environment state.
  • the 0th layer is the input layer
  • the l+1th layer is the output layer.
  • Network equipment can use supervised learning, unsupervised learning and other learning methods to train the weight parameters and bias parameters of the policy neural network.
  • mobile terminals will access or leave the network from time to time, which causes the number of terminals considered by network equipment to be constantly changing when allocating resources. Furthermore, when network equipment uses an AI model to solve wireless resource allocation problems, the amount of input data to the AI model will continue to change.
  • multiple AI models applicable to different dimensions can be pre-trained to address the wireless resource allocation problem that changes in this dimension.
  • an AI model with appropriate dimensions is selected for solution. or,
  • the AI model suitable for the number of terminals is retrained, and the real-time trained AI model is used to solve the wireless resource allocation problem.
  • it is to pre-train multiple AI models or to retrain the AI models on demand each time it will bring significant overhead. For example, pre-training multiple AI models will require larger storage resources to store these AI models, which will incur larger storage overhead. For another example, training a new AI model on demand each time will bring large delay overhead and computing overhead.
  • the above wireless resource allocation problem has the characteristics of replacement and equivariance, and the replacement environment state elements in , the strategy obtained by solving the above formula (5) The elements in will also undergo the same displacement.
  • Figure 6 shows the structure of a DNN with replacement equivariant characteristics.
  • the k-th vector of the l-th hidden layer It can be obtained based on the activation function, that is, the l-th layer forward inference of DNN is:
  • k is the k-th vector of the l-th hidden layer
  • U l and V l are the weight parameter submatrices of the l-th hidden layer
  • c l is the bias sub-vector of the l-th hidden layer
  • ⁇ ( ⁇ ) represents the activation function , for example, ⁇ ( ⁇ ) can be a softmax function.
  • k is greater than or equal to 1 and less than or equal to K.
  • K is the number of terminals and is a positive integer.
  • the weight parameter matrix W l and bias vector b l of the lth hidden layer of DNN can be obtained with the following structure:
  • the weight parameter matrix W l is composed of two sub-matrices U l and V l according to certain rules, that is, the diagonal of W l is U l and the off-diagonal is
  • the bias vector b l is composed of K bias sub-vectors c l . Therefore, when the number of terminals K in the above wireless resource allocation problem changes, the submatrices U l and V l can be adjusted, as well as the number of vectors c l , to form a new dimension neural network suitable for the number of terminals, and use the new dimension neural network
  • the network solves wireless resource allocation strategies without retraining the neural network.
  • this wireless resource allocation problem does not have the above-mentioned dimensional generalization characteristics. Therefore, the submatrices U l and V l obtained by training under a certain value of K, as well as c l , cannot obtain the optimal value when the number of terminals is not equal to K.
  • the optimal output y k means that the optimal resource allocation strategy cannot be obtained, and the accuracy of the obtained resource allocation strategy is low, which will bring greater losses to the system.
  • the embodiment of this application provides a data processing method 100.
  • the first device acquires K-dimensional data.
  • the first device inputs the K-dimensional data into the first machine learning model to obtain a solution to the problem to be solved.
  • the first machine learning model includes a first processing module and a second processing module.
  • the second processing module is determined based on the constraints of the problem to be solved.
  • the second processing module is used to perform dimensional generalization on K-dimensional data.
  • the first processing module is trained based on m-dimensional data.
  • the value of m has nothing to do with the value of K.
  • the first device uses a second processing module that performs dimensional generalization on K-dimensional data, thereby enhancing the dimensional generalization ability of K-dimensional data.
  • it uses m-dimensional pre-training
  • the first processing module obtained through training does not need to be retrained due to changes in the dimensions of the input data, and the accuracy of the solution to the problem to be solved can be ensured with low system overhead.
  • the embodiment of the present application further assumes that the problem to be solved is when the total bandwidth used by a network device to communicate with K terminals is minimized, the power of the communication signal sent by the network device to K terminals, and the constraints of the problem to be solved include the above transmission Taking the total power used by the communication signal in the first range as an example, a data processing method 200 is proposed.
  • the first device acquires K-dimensional data, inputs the K-dimensional data into the first processing module, obtains K first intermediate solutions, and then inputs the K first intermediate solutions into the second processing module to obtain the solution to be solved. solution to the problem.
  • K-dimensional data is the channel gain between the network device and K terminals.
  • the power allocation problem when the network equipment sends communication signals to K terminals does not have dimensional generalization characteristics, and the first device cannot generalize the first processing module to K-dimensional data. Therefore, the first device passes the first intermediate solution output by the first processing module through the normalized indication layer to achieve dimensional generalization of K-dimensional data.
  • the power of the communication signal sent by the network device to the K terminals can be solved more accurately.
  • a data processing method 300 is proposed when the condition includes that the service quality when the network device communicates with each of the K terminals is within the second range.
  • the first device acquires K-dimensional data; the first device inputs the K-dimensional data into the first processing module to obtain K first intermediate solutions; the first device inputs the K first intermediate solutions into the activation layer, K second intermediate solutions are obtained; the first device inputs the K second intermediate solutions into the scaling factor layer to obtain a solution to the problem to be solved.
  • K-dimensional data is the channel gain between the network device and K terminals.
  • the bandwidth allocation problem when the network device communicates with K terminals does not have dimensional generalization characteristics, and the first device cannot generalize the first processing module to K-dimensional data. Therefore, the first device passes the first intermediate solution output by the first processing module through the activation layer and the scaling factor layer to achieve dimensional generalization of the K-dimensional data, thereby improving the accuracy of the solution to the bandwidth allocation problem, that is, improving the network The accuracy of the bandwidth used by the device to communicate with K terminals.
  • FIG. 7 is a schematic flowchart of the data processing method 100.
  • the data processing method 100 is explained from the perspective of the first device.
  • the data processing method 100 includes but is not limited to the following steps:
  • the first device acquires K-dimensional data.
  • K is a positive integer.
  • This K-dimensional data is used to solve the problem to be solved.
  • the problem to be solved may be a problem in a wireless communication system.
  • the problem to be solved is how to allocate resources to K terminals, or how to perform channel prediction on K channels, etc. Therefore, the K-dimensional data is determined based on the problem to be solved, that is, the K-dimensional data is different when the problem to be solved is different.
  • the first device may obtain K-dimensional data through channel estimation, or may obtain K-dimensional data through other methods such as receiving from other devices.
  • the embodiments of the present application do not limit the implementation manner in which the first device obtains K-dimensional data.
  • the first device inputs K-dimensional data into the first machine learning model to obtain the solution to the problem to be solved.
  • the first machine learning model includes a first processing module and a second processing module.
  • the second processing module is based on the constraints of the problem to be solved. The conditions are determined.
  • the second processing module is used to perform dimensional generalization on K-dimensional data.
  • the first processing module is obtained by training based on m-dimensional data. The value of m has nothing to do with the value of K.
  • m is a positive integer.
  • the value of m has nothing to do with the value of K, indicating that the first processing module obtained by pre-training has nothing to do with K.
  • the k-th output of the first processing module is related to the k-th input and has nothing to do with the number K of inputs. That is, the first processing module may have permutation equivariant characteristics.
  • the second processing module is determined based on the constraints of the problem to be solved. Therefore, when the problems to be solved are different, the second processing modules are different, or when the constraints of the same problem to be solved are different, the second processing modules are also different. .
  • the second processing module is used to perform dimensional generalization on K-dimensional data, thereby improving the dimensional generalization ability of K-dimensional data, thereby improving the accuracy of the solution to the problem to be solved.
  • the problem to be solved does not have dimensional generalization properties.
  • the first processing module obtained by training based on m-dimensional data cannot be generalized to K-dimensional data. Then, if the problem to be solved is solved based on the first processing module, the accuracy of the obtained solution to the problem to be solved is low. Therefore, the first device solves the problem to be solved based on the first processing module obtained through m-dimensional data training and the second processing module that performs dimensional generalization on K-dimensional data, which can improve the dimensional generalization ability of K-dimensional data, thereby obtaining a more comprehensive Accurate solution.
  • the first processing module is a permutation equivariant neural network or a graph neural network with permutation equivariant characteristics.
  • the first processing module can also be other neural networks with permutation equivariant characteristics. This method allows the first device to use the first processing module when solving a problem to be solved that has permutation equivariant characteristics.
  • the first device inputs K-dimensional data into the first machine learning model. Before obtaining the solution to the problem to be solved, it can also determine whether the problem to be solved has based on the relationship between the problem to be solved and K. Dimensional generalization properties. When the first device determines that the problem to be solved does not have dimensional generalization characteristics, it inputs K-dimensional data into the first machine learning model to obtain a solution to the problem to be solved; when the first device determines that the problem to be solved has dimensional generalization characteristics, The K-dimensional data is input into the second machine learning model to obtain a solution to the problem to be solved.
  • the second machine learning model includes a first processing module.
  • the first processing module obtained by training based on m-dimensional data is then generalized to K-dimensional data. Therefore, when the first device solves the problem to be solved, there is no need to go through the second processing module.
  • the problem to be solved can be directly solved in the first processing module to obtain more accurate values.
  • the preset value is predefined, for example, the preset value is 100, 200, etc.
  • the function characterizing the problem to be solved when the function characterizing the problem to be solved is independent of the number K of variables, the problem to be solved has dimensional generalization properties; when the function characterizing the problem to be solved is related to the number K of variables, the problem to be solved does not have dimensional generalization properties. chemical characteristics. That is to say, as the number K of variables changes, when the function characterizing the problem to be solved does not change, the problem to be solved has dimensional generalization properties; as the number K of variables changes, the function characterizing the problem to be solved does not change. If the function changes, the problem to be solved does not have dimensional generalization properties.
  • the K-dimensional data is input into the first machine learning model to obtain a solution to the problem to be solved, including: inputting the K-dimensional data into the first machine learning model.
  • the first processing module obtains the first intermediate solution; then the K first intermediate solutions are input into the second processing module to obtain the solution to the problem to be solved.
  • the K-dimensional data is input into the pre-trained first processing module to obtain K first intermediate solutions, and then the K first intermediate solutions are processed through the second processing module that performs dimensional generalization.
  • the solution is dimensionally generalized, thereby achieving dimensional generalization of K-dimensional data, which in turn makes the solution more accurate.
  • the structure of the first machine learning model may be as shown in Figure 8.
  • the PENN is trained based on m-dimensional data.
  • the second processing module is used to perform dimensional generalization on K-dimensional data.
  • the first device inputs K-dimensional data into PENN, and then inputs the output of PENN (K first intermediate solutions) into the second processing module, and the second processing module outputs the solution to the problem to be solved.
  • the second processing module performs dimensional generalization on the K first intermediate solutions, thereby improving the dimensional generalization ability of the K-dimensional data, thereby improving the accuracy of the solution to the problem to be solved.
  • the first device when solving the problem to be solved, not only uses the first processing module obtained by pre-training, but also uses the second processing module that performs dimensional generalization on K-dimensional data, thereby enhancing The dimensional generalization ability of K-dimensional data improves the accuracy of solutions to problems to be solved.
  • this method enables the first machine learning model obtained by training in low dimensions by the first device to be directly applied to communication problems in high dimensions without retraining the learning model, thereby saving costs.
  • the embodiment of the present application also assumes that the problem to be solved is when the total bandwidth used by a network device to communicate with K terminals is minimized, the power of the communication signal sent by the network device to K terminals, and the constraints of the problem to be solved include the network device
  • a data processing method 200 is proposed.
  • FIG. 9 is a schematic flowchart of the data processing method 200.
  • the data processing method 200 can be applied to the first device.
  • the data processing method 200 includes but is not limited to the following steps:
  • the problem to be solved is the power of the communication signal sent by the network device to K terminals when the total bandwidth used by a network device to communicate with K terminals is minimized.
  • the constraint condition of the problem to be solved includes that the total power of communication signals sent by the network device to the K terminals is within the first range.
  • the constraints also include that the K terminals are served by the same single-antenna network device, the bandwidth used by the network device to communicate with the K terminals is equal, and the service quality of each of the K terminals is within the second range.
  • K is a positive integer.
  • the problem to be solved is that a single antenna network device serves K terminals, the network device uses the same bandwidth to communicate with each terminal, the total power used by the network device to send communication signals to K terminals is limited, and the service quality of each terminal is limited.
  • the power allocation problem of the network device sending communication signals to K terminals is a problem.
  • the problem to be solved is the power allocation problem of the network device sending communication signals to K terminals when a network device performs downlink communication with K terminals under specific conditions.
  • the above K-dimensional data is the channel gain between the network device and K terminals.
  • the first device may be a network device, or the first device may be a device other than a network device.
  • the terminal device performs channel estimation to obtain K-dimensional data, and then the terminal device sends the K-dimensional data to the network device, so that the first device K-dimensional data is obtained at the terminal device.
  • the first device performs channel estimation to obtain K-dimensional data.
  • the first device when the first device is a device other than a terminal device and a network device, the first device also obtains K-dimensional data from the terminal device, and the K-dimensional data is obtained by performing channel estimation on the terminal device.
  • the model of the above problem to be solved can be expressed as:
  • B is the bandwidth used by the network device to communicate with each terminal
  • P k is the transmission power of the network device to send data to the k-th terminal
  • g [g 1 , g 2 ,..., g K ] is the channel gain between the network equipment and K terminals, that is, K-dimensional data
  • N 0 is the noise power
  • s k is the receiving throughput of the k-th terminal
  • s 0 is the minimum throughput requirement of the terminal
  • P max is the network The maximum value of the total power sent by the device.
  • N 0 is obtained by noise estimation by the kth terminal equipment, and then the kth terminal equipment feeds back to the first device.
  • s 0 is determined based on the communication requirements of the k-th terminal device.
  • P max is specified by the communication system, such as network equipment.
  • the total power of the above K terminals within the first range means that the total power of the K terminals is less than or equal to P max ;
  • the service quality of each of the above K terminals within the second range means that each The receiving throughput of the terminal is greater than or equal to s 0 .
  • the first device determines whether the power allocation problem has dimensional generalization characteristics. s k in the above formula (10) is a joint convex function about P k and B. Therefore, this power allocation problem is a convex problem.
  • the first device obtains the global optimal solution to the power allocation problem based on the Karush-Kuhn-Tucker (KKT) condition:
  • the first machine learning model includes a first processing module and a second processing module.
  • the first processing module is obtained by training based on m data, and the value of m has nothing to do with the value of K.
  • the second processing module is used to perform dimensional generalization on K-dimensional data. Therefore, when the first device solves the above power allocation problem, it can perform dimensional generalization on the K-dimensional data through the second processing module to obtain a more accurate solution.
  • ⁇ s is a normalized exponential function, for example, softmax function
  • the first processing module in the first machine learning model can be PENN or GNN, that is, PENN or GNN training is used
  • the second processing module of the first machine learning model may be a normalized exponential activation layer.
  • the normalized exponential activation layer may include a softmax function, and the softmax function is used to perform dimensional generalization on K-dimensional data.
  • the above gradient reverse transfer method can be used to train PENN, GNN, and the normalized exponential function, and the training process will not be described in detail.
  • inputting K-dimensional data into the first processing module and obtaining K first intermediate solutions means to convert [g 1 , g 2 ,..., g K ] is input into the PENN, and the K first intermediate solutions output by the PENN are obtained, which is
  • the PENN is obtained by the first device based on the scale channel gain training between the network equipment and m terminals.
  • the value of m has nothing to do with the value of K.
  • the scale channel gain can be a small-scale channel gain or a large-scale channel gain. Channel gain.
  • the small-scale channel gain table collects the impact of the channel's own weakening on signal transmission when the originating point is relatively close, for example, the impact of the channel's multipath weakening and Doppler weakening on signal transmission.
  • the large-scale channel gain table captures the impact of path loss on signal transmission when the origin is far away.
  • the power allocation problem in the embodiment of the present application does not have dimensional generalization characteristics. Therefore, the PENN obtained by training on m-dimensional data cannot be generalized to K-dimensional data. Therefore, the K first intermediate solutions obtained by inputting K-dimensional data into PENN have relatively good performance. Due to the low accuracy, the K first intermediate solutions cannot be directly used as the power allocation strategy for the network device to send communication signals to K terminals.
  • the second processing module includes a normalized exponential activation layer.
  • obtaining the solution to the problem to be solved means to Enter the normalized exponential activation layer, which is As input to the softmax function, get the output Should is the solution to the problem to be solved, that is, the power allocation strategy for the network device to send communication signals to K terminals.
  • the normalized exponential activation layer can perform dimensional generalization on K-dimensional data, so the PENN output Input to the normalized index activation layer, the output result of PENN can be generalized to dimension K, thereby improving the dimensional generalization ability of K-dimensional data, and then using the output of the normalized index activation layer as the solution to the power allocation problem , which can improve the accuracy of the solution.
  • the first device when the first device is a network device, after obtaining the solution to the problem to be solved, the first device sends communication signals to K terminals based on the solution (ie, the power of sending communication signals to K terminals).
  • the first device when the first device is a terminal device, or a device other than the network device and the terminal device, after obtaining the solution to the problem to be solved, the first device sends the solution to the problem to be solved to the network device to The network device is caused to send communication signals to K terminals based on the solution (ie, the power of sending communication signals to K terminals).
  • the problem to be solved is when the total bandwidth used by a network device to communicate with K terminals is minimized, the power of the communication signal sent by the network device to K terminals, and the constraints of the problem to be solved include:
  • the K-dimensional data is input into the first processing module obtained by pre-training, and then the first intermediate solution output by the first processing module is input into the normalized index.
  • the second processing module of the layer obtains the solution to the problem to be solved.
  • This power allocation problem does not have dimensional generalization properties, and the first processing module cannot be generalized to K-dimensional data. Therefore, the first intermediate solution output by the first processing module is then passed through the normalized exponential activation layer to achieve dimensional generalization of K-dimensional data, so that when the total bandwidth is minimized, the solution power is more accurate.
  • the above data processing method 200 takes a network device performing downlink transmission with K single-antenna terminal devices as an example to solve the power distribution of communication signals sent by the network device to K terminals.
  • the problem to be solved is to minimize the total bandwidth used by a network device to communicate with K terminals.
  • the power of communication signals sent by K terminals to the network device, and the constraints of the problem to be solved include when the total power of communication signals sent by K terminals to the network device is in the third range
  • the implementation is the same as in the above-mentioned data processing method 200 The implementation is similar and will not be described again here.
  • FIG. 11 is a schematic flowchart of the data processing method 300.
  • the data processing method 300 can be applied to the first device.
  • the data processing method 300 includes but is not limited to the following steps:
  • the problem to be solved is the power and bandwidth used by the network device to send communication signals to K terminals when the total bandwidth used by a network device to communicate with K terminals is minimized.
  • the constraints of the problem to be solved include that the service quality when the network device communicates with each of the K terminals is within the second range.
  • the constraints also include that K terminals are served by the same network device equipped with N antennas, and the bandwidths used by the K terminals to communicate with the network device are not equal, and the total power of the network device sending communication signals to the K terminals within the first range.
  • the problem to be solved is that a multi-antenna network device serves K terminals, each terminal may use a different bandwidth, the total power of the network device to send communication signals to K terminals is limited, and the service quality of each terminal is limited, With the goal of minimizing the total bandwidth of communication signals sent by network equipment to K terminals, the bandwidth allocation problem and power allocation problem of network equipment sending communication signals to K antenna terminals are solved.
  • the problem to be solved is the bandwidth allocation problem and power allocation problem when a network device performs downlink communication with K terminals under certain constraints.
  • the K-dimensional data is the channel gain between the network device and K terminals.
  • the first device may be a network device, or the first device may be a device other than a network device.
  • the terminal performs channel estimation to obtain K-dimensional data, and then the terminal sends the K-dimensional data to the network device, so that the first device obtains the K-dimensional data from the terminal.
  • the first device performs channel estimation to obtain K-dimensional data.
  • the first device when the first device is a device other than a terminal and a network device, the first device also obtains K-dimensional data from the terminal, and the K-dimensional data is obtained by performing channel estimation on the terminal.
  • B k is the bandwidth used by the k-th terminal
  • P k is the transmission power of the network device to send data to the k-th terminal
  • ⁇ is the quality of service (QoS) index
  • S E is the target throughput, which is determined based on the service requirements of the terminal device.
  • P max is the maximum value of the total power sent by the network device, which is specified by the system.
  • the large-scale channel gain between k terminals represents the channel fading.
  • the network equipment determines that the optimal solution to the problem satisfies the following conditions:
  • the network device solves the power allocation problem and bandwidth allocation in equation (14), it determines whether the problem to be solved has dimensional generalization characteristics. According to formula (15), it can be seen that since the upper limit of the total power P max of K terminals is fixed, the optimal power allocation strategy Decreases as K increases. Since the power P k of each terminal decreases as K increases, the required bandwidth needs to increase to meet the throughput requirements of the terminal. Therefore, the optimal bandwidth allocation strategy B k * increases as K increases. To sum up, the power allocation strategy in the embodiment of this application and bandwidth allocation strategies None of them satisfy the dimensional generalization properties.
  • the first machine learning model including the first processing module and the second processing module can be used to solve the problem.
  • the first processing module is trained based on m-dimensional data.
  • the second processing module is determined based on the constraints of the problem to be solved, and is used to perform dimensional generalization on K-dimensional data. Therefore, by using the first machine learning model to solve the power allocation and bandwidth allocation for K terminals, a more accurate solution can be obtained.
  • the constraints for solving the power allocation strategy in the embodiment of the present application are the same as the constraints in the above-mentioned data processing method 300. Therefore, the first machine model in the above-mentioned data processing method 300 can be used to solve the power of communication signals sent by the network device to K terminals. Allocation, more accurate allocation results can be obtained. Therefore, the data processing method proposed in the embodiment of this application is suitable for solving the bandwidth allocation when the network device sends communication signals to K terminals.
  • the quality of service based on the communication between the network device and each of the K terminals is within the second range.
  • the constraints determine that the first machine learning model includes a first processing module and a second processing module as shown in Figure 12.
  • the first processing module includes PENN or GNN, or other neural networks with permutation equivariant characteristics.
  • obtaining K first intermediate solutions means inputting K-dimensional data into the first processing module, that is, inputting [ ⁇ 1 , ⁇ 2 ,..., ⁇ K ] into PENN to obtain K first intermediate solutions.
  • the intermediate solution, the K first intermediate solutions are
  • the K second intermediate solutions are obtained by inputting K first intermediate solutions into the activation layer.
  • the second processing module includes an activation layer and a scaling factor layer.
  • the kth scaling factor in the scaling factor layer is obtained by inputting the kth dimension data and K into the scaling factor calculation module.
  • k is a positive integer less than or equal to K.
  • the scaling factor calculation module can be shown in Figure 13.
  • the scaling factor calculation module includes a fully connected neural network (fully connected neural network, FNN) and an activation layer.
  • the activation layer can also be a softplus function.
  • the network device inputs the k-th data ⁇ k and K, and then inputs the output result of FNN into the softplus function to obtain the k-th scaling factor. Therefore, the first device can obtain K scaling factors based on the scaling factor calculation module, K-dimensional data and K, and the K scaling factors are used for the scaling factor layer in the second processing module.
  • the scaling factor layer is used to scale the second intermediate solution output by the activation layer.
  • the activation layer includes the softplus function, so obtaining K second intermediate solutions can mean dividing the K general Enter the softplus function and obtain K second intermediate solutions as
  • bandwidth allocation problem and power allocation problem of formula (14) in the embodiment of this application are equivalent to the following principal dual optimization problem:
  • L( ⁇ ,P k ,B k , ⁇ k ) represents the Lagrangian function
  • ⁇ k is the Lagrange multiplier
  • IE( ⁇ ) is the expectation operation.
  • the problem to be solved in the above formula (14) can be solved through three policy neural networks and one FNN.
  • the three policy neural networks can be expressed as It is the neural network shown in Figure 10 above, including PENN and a normalized exponential activation layer with the softmax function as the activation function, used for the output power allocation strategy. It is the neural network shown in Figure 12 above, including PENN, an activation layer with softplus as the activation function, and a scaling layer, which is used to output the bandwidth allocation policy to be scaled. It includes PENN and an activation layer with softplus as the activation function, which is used to output Lagrange multipliers.
  • FNN is Used to output the scaling factor of the bandwidth allocation strategy to be scaled, ⁇ k is the large-scale channel gain between the network device and the k-th terminal, and ⁇ v is the parameter to be trained in the FNN.
  • the first device uses a supervised learning method to train the parameter ⁇ v , that is, according to Use bisection method to solve get training labels Define the loss function as Should is the output of FNN. Then use the loss function to find the gradient of the FNN parameter ⁇ v to obtain the gradient of each parameter, and perform reverse transfer to update the parameter ⁇ v to realize the training of FNN, that is, the solution is obtained
  • the first device uses an unsupervised learning method to train the parameter ⁇ v , that is, the objective function of FNN training is defined as the target throughput S E and the actual throughput The difference between them is then minimized by adjusting the parameter ⁇ v of FNN to obtain
  • ⁇ P , ⁇ B , and ⁇ ⁇ are three neural networks respectively. learning efficiency.
  • the power allocation strategy and bandwidth allocation strategy in equation (14) can be solved based on the trained three policy neural networks and one FNN.
  • the first device adopts That is, the first machine learning model shown in Figure 10 solves the power allocation problem in formula (14); the first device uses That is, the first machine learning model shown in Figure 12 solves the bandwidth allocation problem in formula (14).
  • the scaling factor of the scaling layer in Figure 12 is obtained based on the scaling factor calculation module shown in Figure 13, and the FNN in the scaling factor calculation module is the FNN obtained by the above training.
  • the above training process of the neural network can be performed offline, and therefore is not limited to execution by the first device or other equipment.
  • the first device when the first device is a network device, after obtaining a solution to the problem to be solved, the first device sends communication signals to K terminals based on the solution.
  • the first device when the first device is a terminal device, or a device other than the network device and the terminal device, after obtaining the solution to the problem to be solved, the first device sends the solution to the problem to be solved to the network device to The network device is caused to send communication signals to K terminals based on the solution.
  • the constraints include that the service quality of each terminal among the K terminals is within the second range, minimizing the total bandwidth of the K terminals is the goal, and the network device sends communication signals to the K terminals.
  • the network device sends communication signals to the K terminals.
  • first input the acquired K-dimensional data into the first processing module to obtain K first intermediate solutions then input the K first intermediate solutions into the activation layer, obtain K second intermediate solutions, and then The K second intermediate solutions are input to the scaling factor layer, and the scaling factor layer is used to scale the K second intermediate solutions to obtain the bandwidth of K terminals.
  • the K-dimensional data is dimensionally generalized through the activation layer and the scaling factor layer, which can improve the dimensional generalization ability of the K-dimensional data, thereby improving the accuracy of bandwidth allocation when the network device sends communication signals to K terminals.
  • the above data processing method 300 takes a network device performing downlink transmission with K single-antenna terminal devices as an example to solve the bandwidth allocation and power allocation when the network device sends communication signals to K terminals.
  • a network device performs uplink transmission with K terminals and solves the bandwidth allocation problem and power allocation problem of K terminals sending communication signals to the network device, that is, the problem to be solved is the network device used to communicate with K terminals.
  • the total bandwidth is minimized, the power and bandwidth used by K terminals to send communication signals to the network device, and the constraints include the quality of service when the network device communicates with each of the K terminals is within the second range.
  • the implementation is similar to the implementation in the above-mentioned data processing method 300 and will not be described again here.
  • the embodiment of the present application aims at minimizing the total bandwidth of the K single-antenna terminals when the network equipment constraints include that the service quality of each single-antenna terminal among the K single-antenna terminals is within the second range, and solves the problem of K single-antenna terminals.
  • the network availability and bandwidth requirements of the communication system are simulated using the data processing method 100 proposed in the embodiment of the present application and some currently adopted solution methods.
  • the simulation parameters of the communication system are: the distance d k between the k-th terminal and the network equipment is between 50 and 250 meters, the path loss model of the k-th terminal is 35.3+37.6log(d k ), and the network equipment
  • K is the number of single-antenna terminals
  • t is the learning time
  • Pre-trained fully connected neural network When supervised learning is used for training, and a sample is expressed as ⁇ k is obtained from the randomly generated terminal position through the path model, K is randomly generated from 10 to 200, and a total of 1000 pre-training samples are generated. Train a neural network When , unsupervised learning is used, a sample is represented as (g, ⁇ , K), and the small-scale channel g is generated by Rayleigh distribution.
  • the training set, validation set and test set are generated as shown in Table 2:
  • the training set refers to the value of K when the network device trains the first machine learning model.
  • the verification set refers to the value of K when the network device verifies the first trained machine learning model.
  • the test set refers to the value of K when the network device uses a larger value than the training set to test the first machine learning model.
  • the value of K in the training set is [10,11,...,30], which means that K in the training set is any value in the set [10,11,...,30].
  • the values of K in the validation set and test set are also similar.
  • the number of samples generated for each K refers to the number of samples generated for each value of K. For example, in the training set, when the value of K is 10, the number of samples generated by the value of K is 10; when the value of K is 30, the number of samples generated by the value of K is also 10, so the total number of samples generated by the value of K is 10. for 210.
  • the number of samples generated by the K value is 100; when the K value is 50, the number of samples generated by the K value is 100; the value of K is When the K value is 100, the number of samples generated by the K value is 100; when the K value is 200, the number of samples generated by the K value is also 100. Therefore, in the test set, the total number of samples generated by the K value is 400.
  • the performance indicators simulated by the embodiments of this application are network availability and bandwidth requirements.
  • network availability ⁇ K can be expressed as:
  • ⁇ K is the set of all samples with dimension K in the test set
  • is the number of elements in the set
  • I( ⁇ ) is the indicator function.
  • the network availability index refers to the ratio of the number of terminals in the network that can meet the packet loss rate requirements to the total number of terminals. The larger the network availability index, the better the system performance.
  • P-PENN is the first machine learning model in the data processing method 300 proposed in the embodiment of this application.
  • FNN is currently a pre-trained fully connected neural network.
  • the first machine learning model in the data processing method 300 is trained in a scenario where K is 10 to 30 terminals, when used for other more terminal numbers, that is, the number of terminals is At 50, 100, and 200, network availability ⁇ K is all 1, and bandwidth requirements Smaller than other options. That is to say, for a large number of terminals, the first machine learning model in the data processing method 300 can meet the packet loss rate requirement for all terminals, and the bandwidth requirement is relatively small.
  • the network device uses M-PENN 10 , although the bandwidth requirement of the terminal is less than when the first machine learning model in the data processing method 300 is used, in the scenario of multiple terminals, its network availability approaches 0, that is, the network availability is extremely high. Poor and cannot be applied to actual scenarios.
  • the network device uses M-PENN 200 , the network availability can be guaranteed to be 1 in the scenario of multiple terminals, but the bandwidth requirement is greater than the bandwidth requirement when using the first machine learning model in the data processing method 300, so more data needs to be used. Many resources.
  • the network device adopts FNN more bandwidth is used than when the first machine learning model in the data processing method 300 is adopted, but the network availability is still not guaranteed to be 1.
  • the network device adopts the data processing method 300, when the first machine model pre-trained with a small number of terminals is applied to a scenario with a large number of terminals, it can still achieve relatively small bandwidth requirements. , ensuring network availability. That is to say, when the network device uses the data processing method 300 to solve resource allocation, the resource allocation results obtained are more accurate and can ensure system performance.
  • the first device may include a hardware structure and/or a software module to implement the above functions in the form of a hardware structure, a software module, or a hardware structure plus a software module. Whether one of the above functions is performed as a hardware structure, a software module, or a hardware structure plus a software module depends on the specific application and design constraints of the technical solution.
  • an embodiment of the present application provides a communication device 1400.
  • the communication device 1400 may be a component of the first device (eg, integrated circuit, chip, etc.), or may be a component of the first device (eg, integrated circuit, chip, etc.).
  • the communication device 1400 may also be other communication units, used to implement the methods in the method embodiments of the present application.
  • the communication device 1400 may include: a communication unit 1401 and a processing unit 1402.
  • a storage unit 1403 may also be included.
  • one or more units as shown in Figure 14 may be implemented by one or more processors, or by one or more processors and memories; or by one or more processors and a transceiver; or may be implemented by one or more processors, memories, and transceivers, which are not limited in the embodiments of the present application.
  • the processor, memory, and transceiver can be set separately or integrated.
  • the communication device 1400 has the function of implementing the first device described in the embodiment of this application.
  • the communication device 1400 includes modules or units or means (means) corresponding to the first device executing the steps involved in the first device described in the embodiments of this application.
  • the functions, units or means (means) can be implemented through software, Or it can be implemented through hardware, it can also be implemented through hardware to execute corresponding software, or it can also be implemented through a combination of software and hardware.
  • a communication device 1400 may include: a processing unit 1402 and a communication unit 1401, the communication unit 1401 being used to transmit and receive data/signaling;
  • Processing unit 1402 used to obtain K-dimensional data
  • the processing unit 1402 is also used to input the K-dimensional data into a first machine learning model to obtain a solution to the problem to be solved;
  • the first machine learning model includes a first processing module and a second processing module;
  • the second processing module is determined based on the constraints of the problem to be solved; the second processing module is used to perform dimensional generalization on the K-dimensional data; the first processing module is trained based on m-dimensional data Obtained; the value of m has nothing to do with the value of K; the K and m are positive integers.
  • the problem to be solved does not have dimensional generalization properties.
  • the processing unit 1402 inputs the K-dimensional data into the first machine learning model to obtain a solution to the problem to be solved, specifically for:
  • the problem to be solved is the power of the communication signal sent by the network device to the K terminals when the total bandwidth used by a network device to communicate with K terminals is minimized;
  • the constraint condition includes that the total power of communication signals sent by the network device to the K terminals is within a first range.
  • the problem to be solved is the power of communication signals sent by the K terminals to the network device when the total bandwidth used by K terminals to communicate with a network device is minimized;
  • the constraint condition includes that the total power of communication signals sent by the K terminals to the network device is within a third range.
  • the second processing module includes a normalized exponential function activation layer.
  • the problem to be solved is when the total bandwidth used by a network device to communicate with K terminals is minimized, the bandwidth used by the network device to communicate with the K terminals is ;
  • the constraint condition includes that the service quality when the network device communicates with each of the K terminals is within the second range.
  • the second processing module includes an activation layer and a scaling factor layer; the kth scaling factor in the scaling factor layer is the input of the kth data and the K into the scaling factor calculation module. Obtained; the k is a positive integer less than or equal to K.
  • the processing unit 1402 inputs the K first intermediate solutions into the second processing module to obtain the solution to the problem to be solved, specifically for:
  • the K second intermediate solutions are input into the scaling factor layer to obtain the solution to the problem to be solved.
  • the K-dimensional data is channel gains between the network device and the K terminals.
  • the first processing module is a permutation equivariant neural network or a graph neural network with permutation equivariant characteristics.
  • FIG. 15 is a schematic structural diagram of the communication device 1500.
  • the communication device 1500 may be a first device, or may be a chip, chip system, or processor that supports the first device to implement the above method.
  • the device can be used to implement the method described in the above method embodiment. For details, please refer to the description in the above method embodiment.
  • the communication device 1500 may include one or more processors 1501.
  • the processor 1501 may be a general-purpose processor or a special-purpose processor.
  • it can be a baseband processor, a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components or a central processing unit (Central Processing Unit, CPU).
  • the baseband processor can be used to process communication protocols and communication data
  • the central processor can be used to process communication devices (such as base stations, baseband chips, terminals, terminal chips, distributed units (DU) or centralized units (centralized units)). unit, CU), etc.) to control, execute software programs, and process data of software programs.
  • DU distributed units
  • centralized units centralized units
  • the communication device 1500 may include one or more memories 1502, on which instructions 1504 may be stored, and the instructions may be executed on the processor 1501, causing the communication device 1500 to perform the above method. Methods described in the Examples.
  • the memory 1502 may also store data.
  • the processor 1501 and the memory 1502 can be provided separately or integrated together.
  • the memory 1502 may include, but is not limited to, non-volatile memories such as hard disk drive (HDD) or solid-state drive (SSD), random access memory (Random Access Memory, RAM), erasable programmable memory, etc.
  • non-volatile memories such as hard disk drive (HDD) or solid-state drive (SSD), random access memory (Random Access Memory, RAM), erasable programmable memory, etc.
  • Read-only memory Erasable Programmable ROM, EPROM
  • ROM or portable read-only memory Compact Disc Read-Only Memory, CD-ROM
  • the communication device 1500 may also include a transceiver 1505 and an antenna 1506.
  • the transceiver 1505 may be called a transceiver unit, a transceiver, a transceiver circuit, etc., and is used to implement transceiver functions.
  • the transceiver 1505 may include a receiver and a transmitter.
  • the receiver may be called a receiver or a receiving circuit, etc., used to implement the receiving function;
  • the transmitter may be called a transmitter, a transmitting circuit, etc., used to implement the transmitting function.
  • the communication device 1500 is the first device: the processor 1501 is used to execute S101 and S102 in the above-mentioned data processing method 100, and is used to execute S201, S202 and S203 in the data processing method 200, and is used to execute the data processing method. S301, S302, S303, S304 in 300.
  • the processor 1501 may include a transceiver for implementing receiving and transmitting functions.
  • the transceiver may be a transceiver circuit, an interface, or an interface circuit.
  • the transceiver circuits, interfaces or interface circuits used to implement the receiving and transmitting functions can be separate or integrated together.
  • the above-mentioned transceiver circuit, interface or interface circuit can be used for reading and writing codes/data, or the above-mentioned transceiver circuit, interface or interface circuit can be used for signal transmission or transfer.
  • the processor 1501 can store instructions 1503, and the instructions 1503 are run on the processor 1501, which can cause the communication device 1500 to execute the method described in the above method embodiment.
  • the instructions 1503 may be fixed in the processor 1501, in which case the processor 1501 may be implemented by hardware.
  • the communication device 1500 may include a circuit, and the circuit may implement the sending or receiving or communication functions in the foregoing method embodiments.
  • the processor and transceiver described in the embodiments of this application can be implemented in integrated circuits (ICs), analog ICs, radio frequency integrated circuits (RFICs), mixed signal ICs, application specific integrated circuits (application specific integrated circuits). , ASIC), printed circuit board (PCB), electronic equipment, etc.
  • the processor and transceiver can also be manufactured using various IC process technologies, such as complementary metal oxide semiconductor (CMOS), N-type metal-oxide-semiconductor (NMOS), P-type Metal oxide semiconductor (positive channel metal oxide semiconductor (PMOS), bipolar junction transistor (BJT), bipolar CMOS (BiCMOS), silicon germanium (SiGe), gallium arsenide (GaAs), etc.
  • CMOS complementary metal oxide semiconductor
  • NMOS N-type metal-oxide-semiconductor
  • PMOS positive channel metal oxide semiconductor
  • BJT bipolar junction transistor
  • BiCMOS bipolar CMOS
  • SiGe silicon germanium
  • GaAs gallium arsenide
  • the communication device described in the above embodiments may be the first device, but the scope of the communication device described in the embodiments of the present application is not limited thereto, and the structure of the communication device may not be limited by FIG. 15 .
  • the communication device may be a stand-alone device or may be part of a larger device.
  • the communication device may be:
  • the IC collection may also include a storage component for storing data and instructions;
  • ASIC such as modem (modulator)
  • the communication device may be a chip or a chip system
  • the chip 1600 shown in FIG. 16 includes a processor 1601 and an interface 1602.
  • the number of processors 1601 may be one or more, and the number of interfaces 1602 may be multiple.
  • the processor 1601 may be a logic circuit, and the interface 1602 may be an input-output interface, an input interface or an output interface.
  • the chip 1600 may also include memory 1603 .
  • the interface 1602 is used for output or reception.
  • the processor 1601 is used to obtain K-dimensional data
  • the processor 1601 is also used to input the K-dimensional data into a first machine learning model to obtain a solution to the problem to be solved;
  • the first machine learning model includes a first processing module and a second processing module;
  • the second processing module is determined based on the constraints of the problem to be solved; the second processing module is used to perform dimensional generalization on the K-dimensional data; the first processing module is trained based on m-dimensional data Obtained; the value of m has nothing to do with the value of K; the K and m are positive integers.
  • the communication device 1500 and the chip 1600 can also perform the implementation described above for the communication device 1400.
  • the various illustrative logical blocks and steps listed in the embodiments of this application can be implemented by electronic hardware, computer software, or a combination of both. Whether such functionality is implemented in hardware or software depends on the specific application and overall system design requirements. Those skilled in the art can use various methods to implement the described functions for each specific application, but such implementation should not be understood as exceeding the protection scope of the embodiments of the present application.
  • This application also provides a computer-readable storage medium for storing computer software instructions. When the instructions are executed by a communication device, the functions of any of the above method embodiments are implemented.
  • This application also provides a computer program product for storing computer software instructions. When the instructions are executed by a communication device, the functions of any of the above method embodiments are implemented.
  • This application also provides a computer program that, when run on a computer, implements the functions of any of the above method embodiments.
  • This application also provides a communication system, which includes one or more network devices and one or more terminal devices.
  • the system may also include other devices that interact with network devices and terminal devices in the solution provided by this application.
  • the above embodiments may be implemented in whole or in part by software, hardware, firmware, or any combination thereof.
  • software When implemented using software, it may be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on the computer, the processes or functions described in the embodiments of the present application are generated in whole or in part.
  • the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.
  • the computer instructions may be stored in or transmitted from one computer-readable storage medium to another, e.g., the computer instructions may be transferred from a website, computer, server, or data center Transmission to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) means.
  • the computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains one or more available media integrated.
  • the available media may be magnetic media (eg, floppy disk, hard disk, tape), optical media (eg, high-density digital video disc (DVD)), or semiconductor media (eg, SSD), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本申请提供了一种数据处理方法及相关装置。该方法包括：获取K维数据，将该K维数据输入第一机器学习模型，获得待求解问题的解。其中，第一机器学习模型包括第一处理模块和第二处理模块。第二处理模块是基于待求解问题的约束条件确定的。第二处理模块用于对K维数据进行维度泛化。第一处理模块是根据m维数据训练获得的。m的取值与K的取值无关，K、m为正整数。在求解待求解问题时，采用了对K维数据进行维度泛化的第二处理模块，从而可增强K维数据的维度泛化能力，另外采用了根据m维预先训练获得的第一处理模块，无需因为输入数据的维度的变化而重新训练第一处理模块，可以以较低的系统开销，保证待求解问题的解的准确性。

Description

一种数据处理方法及相关装置
本申请要求于2022年3月31日提交中国国家知识产权局、申请号为202210336391.9、申请名称为“一种数据处理方法及相关装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及神经网络领域，尤其涉及一种数据处理方法及相关装置。
背景技术
目前，人工智能(artificial intelligence，AI)技术已应用于无线通信系统中的网络层相关问题(如网络优化、移动性管理、资源分配等)和物理层相关问题(如信道编译码、信道预测、接收机等)等方面。
当AI技术应用于无线通信系统中时，可在设备个数为预设数的条件下，训练神经网络，获得求解待求解问题的神经网络。一些情况下，待求解问题的输出与输入之间的关系会随着设备数量的变化而发生变化。当设备的数量发生变化时，若仍采用未经重新训练的神经网络，使得待求解问题的解的准确性较低，从而使得系统产生的较大性能损失。因此，往往需要重新训练神经网络，或预先训练多个模型以适配不同的设备数量，存在系统的存储和处理开销大的问题。
发明内容
本申请实施例提供了一种数据处理方法及相关装置，可以以较低的系统开销，保证待求解问题的解的准确性。
第一方面，本申请实施例提供一种数据处理方法，可以应用于第一装置。该方法包括，获取K维数据，将该K维数据输入第一机器学习模型，获得待求解问题的解。其中，第一机器学习模型包括第一处理模块和第二处理模块。第二处理模块是基于待求解问题的约束条件确定的。第二处理模块用于对K维数据进行维度泛化。第一处理模块是根据m维数据训练获得的。m的取值与K的取值无关，K、m为正整数。
本申请实施例中，在求解待求解问题时，采用了对K维数据进行维度泛化的第二处理模块，从而可增强K维数据的维度泛化能力，另外采用了根据m维预先训练获得的第一处理模块，无需因为输入数据的维度的变化而重新训练第一处理模块，可以以较低的系统开销，保证待求解问题的解的准确性。
一种可选的实施方式中，第一装置可通过信道估计获取K维数据。可选的，K维数据由终端进行信道估计获得，第一装置再从该终端处接收K维数据。
一种可选的实施方式中，上述待求解问题不具有维度泛化特性。待求解问题不具有维度泛化特性时，根据m维数据训练获得的第一处理模块泛化到K维数据。那么，若基于第一处理模块求解待求解问题，则获得的待求解问题的解的准确性较低。因此，基于预先训练获得的第一处理模块和对K维数据进行维度泛化的第二处理模块，求解待求解问题，可提高K维数据的维度泛化能力，从而获得较为准确的解。
一种可选的实施方式中，将K维数据输入第一机器学习模型，获得待求解问题的解，包 括：将K维数据输入第一处理模块，获得K个第一中间解；将K个第一中间解输入第二处理模块，获得待求解问题的解。
可见，在求解待求解问题时，将K维数据输入预先训练的第一处理模块，获得K个第一中间解，再通过进行维度泛化的第二处理模块对K个第一中间解进行维度泛化，从而实现对K维数据的维度泛化，进而可使得求解的解更加准确。
一种可选的实施方式中，待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，网络设备向K个终端发送通信信号的功率，该待求解问题的约束条件包括网络设备向K个终端发送通信信号的总功率在第一范围内。也就是说，待求解问题是一个网络设备与K个终端进行下行通信时，网络设备向K个终端发送通信信号的功率分配问题。
一种可选的实施方式中，待求解问题是K个终端与一个网络设备进行通信所使用的总带宽最小化时，K个终端向网络设备发送通信信号的功率，该待求解问题的约束条件包括K个终端向网络设备发送通信信号的总功率在第三范围内。也就是说，待求解问题是K个终端与一个网络设备进行上行通信时，K个终端向网络设备发送通信信号的功率分配问题。
一种可选的实施方式中，待求解问题是上述K个终端与一个网络设备进行上行或下行通信的功率分配问题时，上述第二处理模块包括归一化指数函数激活层。也就是说，第一装置求解该功率分配问题时，通过归一化指数函数激活层对K维数据进行维度泛化。
另一种可选的实施方式中，待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，网络设备与K个终端进行通信所使用的带宽，该待求解问题的约束条件包括网络设备与K个终端中每个终端进行通信时的服务质量在第二范围内。也就是说，待求解问题为一个网络设备与K个终端进行上行通信或下行通信时，所使用的带宽分配问题。
一种可选的实施方式中，待求解问题为一个网络设备与K个终端进行上行通信或下行通信，所使用的带宽分配问题时，上述第二处理模块包括激活层和缩放因子层。缩放因子层中的第k个缩放因子是将第k个数据和K输入缩放因子计算模块获得的，k为小于或等于K的正整数。可见，求解该带宽分配问题时，通过激活层和缩放因子层对K维数据进行维度泛化。
一种可选的实施方式中，待求解问题为一个网络设备与K个终端进行上行通信或下行通信，所使用的带宽分配问题时，将K个第一中间解输入第二处理模块，获得待求解问题的解，包括：将K个第一中间解输入激活层，获得K个第二中间解；将K个第二中间解输入缩放因子层，获得待求解问题的解。
可见，基于预先训练的第一处理模块获得K个第一中间解后，再依次基于激活层和缩放因子层对K个第一中间解进行维度泛化，获得带宽分配问题的解。
一种可选的实施方式中，待求解问题是上述功率分配问题或带宽分配问题时，K维数据是网络设备与K个终端之间的信道增益。该K个信道增益可以是第一装置通过信道估计获得的。
一种可选的实施方式中，上述第一处理模块是具有置换等变特性的置换等变神经网络或图神经网络。该方式可使得第一装置求解具有置换等变特性的待求解问题时，可采用该第一处理模块。
第二方面，本申请还提供一种通信装置。该通信装置具有实现上述第一方面所述的第一装置的部分或全部功能。比如，该通信装置的功能可具备本申请中第一方面所述的第一装置的部分或全部实施例中的功能，也可以具备单独实施本申请中的任一个实施例的功能。所述功能可以通过硬件实现，也可以通过硬件执行相应的软件实现。所述硬件或软件包括一个或多个与上述功能相对应的单元或模块。
在一种可能的设计中，该通信装置的结构中可包括处理单元和通信单元，所述处理单元被配置为支持通信装置执行上述方法中相应的功能。所述通信单元用于支持该通信装置与其他通信装置之间的通信。所述通信装置还可以包括存储单元，所述存储单元用于与处理单元和通信单元耦合，其保存通信装置必要的程序指令和数据。
一种实施方式中，所述通信装置包括：处理单元和通信单元，通信单元用于进行数据/信令收发；
处理单元，用于获取K维数据；
处理单元，还用于将所述K维数据输入第一机器学习模型，获得待求解问题的解；所述第一机器学习模型包括第一处理模块和第二处理模块；
所述第二处理模块是基于所述待求解问题的约束条件确定的；所述第二处理模块用于对所述K维数据进行维度泛化；所述第一处理模块是根据m维数据训练获得的；所述m的取值与所述K的取值无关；所述K、所述m为正整数。
另外，该方面中，通信装置其他可选的实施方式可参见上述第一方面的相关内容，此处不再详述。
作为示例，通信单元可以为收发器或通信接口，存储单元可以为存储器，处理单元可以为处理器。
一种实施方式中，所述通信装置包括：处理器和收发器，收发器用于进行数据/信令收发；
处理器，用于获取K维数据；
处理器，还用于将所述K维数据输入第一机器学习模型，获得待求解问题的解；所述第一机器学习模型包括第一处理模块和第二处理模块；
所述第二处理模块是基于所述待求解问题的约束条件确定的；所述第二处理模块用于对所述K维数据进行维度泛化；所述第一处理模块是根据m维数据训练获得的；所述m的取值与所述K的取值无关；所述K、所述m为正整数。
另外，该方面中，上行通信装置其他可选的实施方式可参见上述第一方面的相关内容，此处不再详述。
另一种实施方式中，该通信装置为芯片或芯片系统。所述处理单元也可以体现为处理电路或逻辑电路；所述收发单元可以是该芯片或芯片系统上的输入/输出接口、接口电路、输出电路、输入电路、管脚或相关电路等。
在实现过程中，处理器可用于进行，例如但不限于，基带相关处理，收发器可用于进行，例如但不限于，射频收发。上述器件可以分别设置在彼此独立的芯片上，也可以至少部分的或者全部的设置在同一块芯片上。例如，处理器可以进一步划分为模拟基带处理器和数字基带处理器。其中，模拟基带处理器可以与收发器集成在同一块芯片上，数字基带处理器可以设置在独立的芯片上。随着集成电路技术的不断发展，可以在同一块芯片上集成的器件越来越多。例如，数字基带处理器可以与多种应用处理器(例如但不限于图形处理器，多媒体处理器等)集成在同一块芯片之上。这样的芯片可以称为系统芯片(system on a chip，SoC)。将各个器件独立设置在不同的芯片上，还是整合设置在一个或者多个芯片上，往往取决于产品设计的需要。本申请实施例对上述器件的实现形式不做限定。
第三方面，本申请还提供一种处理器，用于执行上述各种方法。在执行这些方法的过程中，上述方法中有关发送上述信息和接收上述信息的过程，可以理解为由处理器输出上述信息的过程，以及处理器接收输入的上述信息的过程。在输出上述信息时，处理器将该上述信息输出给收发器，以便由收发器进行发射。该上述信息在由处理器输出之后，还可能需要进 行其他的处理，然后才到达收发器。类似的，处理器接收输入的上述信息时，收发器接收该上述信息，并将其输入处理器。更进一步的，在收发器收到该上述信息之后，该上述信息可能需要进行其他的处理，然后才输入处理器。
对于处理器所涉及的发送和接收等操作，如果没有特殊说明，或者，如果未与其在相关描述中的实际作用或者内在逻辑相抵触，则均可以更加一般性的理解为处理器输出和接收、输入等操作，而不是直接由射频电路和天线所进行的发送和接收操作。
在实现过程中，上述处理器可以是专门用于执行这些方法的处理器，也可以是执行存储器中的计算机指令来执行这些方法的处理器，例如通用处理器。上述存储器可以为非瞬时性(non-transitory)存储器，例如只读存储器(Read Only Memory，ROM)，其可以与处理器集成在同一块芯片上，也可以分别设置在不同的芯片上，本申请实施例对存储器的类型以及存储器与处理器的设置方式不做限定。
第四方面，本申请还提供了一种通信系统，该系统包括一个或多个网络设备，以及一个或多个终端设备。在另一种可能的设计中，该系统还可以包括与网络设备、终端设备进行交互的其他设备。
第五方面，本申请提供了一种计算机可读存储介质，用于储存指令，当所述指令被计算机运行时，实现上述第一方面任一项所述的方法。
第六方面，本申请还提供了一种包括指令的计算机程序产品，当其在计算机上运行时，实现上述第一方面任一项所述的方法。
第七方面，本申请提供了一种芯片系统，该芯片系统包括处理器和接口，所述接口用于获取程序或指令，所述处理器用于调用所述程序或指令以实现或者支持第一装置实现第一方面所涉及的功能。例如，确定或处理上述方法中所涉及的数据和信息中的至少一种。在一种可能的设计中，所述芯片系统还包括存储器，所述存储器，用于保存终端必要的程序指令和数据。该芯片系统，可以由芯片构成，也可以包括芯片和其他分立器件。
附图说明
图1是本申请实施例提供的一种通信系统的系统结构示意图；
图2是本申请实施例提供的一种全连接神经网络的结构示意图；
图3是本申请实施例提供的一种神经网络的训练方式示意图；
图4是本申请实施例提供的一种梯度反向传输的示意图；
图5是本申请实施例提供的一种维度泛化特性的示意图；
图6是本申请实施例提供的一种深度神经网络的结构示意图；
图7是本申请实施例提供的一种数据处理方法的流程示意图；
图8是本申请实施例提供的一种第一机器学习模型的结构示意图；
图9是本申请实施例提供的另一种数据处理方法的流程示意图；
图10是本申请实施例提供的另一种第一机器学习模型的结构示意图；
图11是本申请实施例提供的又一种数据处理方法的流程示意图；
图12是本申请实施例提供的又一种第一机器学习模型的结构示意图；
图13是本申请实施例提供的一种缩放因子计算模块的结构示意图；
图14是本申请实施例提供的一种通信装置的结构示意图；
图15是本申请实施例提供的另一种通信装置的结构示意图；
图16是本申请实施例提供的一种芯片的结构示意图。
具体实施方式
下面结合本申请实施例中的附图对本申请实施例中的技术方案进行清楚、完整的描述。
一.通信系统。
为了更好的理解本申请实施例公开的数据处理方法，对本申请实施例适用的通信系统进行描述。
本申请实施例可应用于第五代移动通信(5th generation mobile communication，5G)系统、卫星通信及短距等无线通信系统中，系统架构如图1所示。无线通信系统可以包括一个或多个网络设备，以及一个或多个终端设备。无线通信系统也可以进行点对点通信，如多个终端设备之间互相通信。
可理解的，本申请实施例提及的无线通信系统包括但不限于：窄带物联网系统(narrow band-internet of things，NB-IoT)、长期演进系统(long term evolution，LTE)，5G移动通信系统的三大应用场景：增强移动宽带(enhanced mobile broadband，eMBB)、超可靠低时延通信(ultra reliable low latency communication，URLLC)和海量机器类通信(massive machine type of communication，mMTC)，无线保真(wireless fidelity，WiFi)系统，或者5G之后的移动通信系统等。
本申请实施例中，网络设备是具有无线收发功能的设备，用于与终端设备进行通信，可以是LTE中的演进型基站(evolved Node B，eNB或eNodeB)；或者是5G网络中的基站或者未来演进的公共陆地移动网络(public land mobile network，PLMN)中的基站，宽带网络业务网关(broadband network gateway，BNG)，汇聚交换机或者非第三代合作伙伴项目(3rd generation partnership project，3GPP)接入设备等。可选的，本申请实施例中的网络设备可以包括各种形式的基站，例如：宏基站、微基站(也称为小站)、中继站、接入点、未来实现基站功能的设备、WiFi系统中的接入节点，传输接收点(transmitting and receiving point，TRP)、发射点(transmitting point，TP)、移动交换中心以及设备到设备(device-to-device，D2D)、车辆外联(vehicle-to-everything，V2X)、机器到机器(machine-to-machine，M2M)通信中承担基站功能的设备等，本申请实施例对此不作具体限定。
网络设备可以和核心网设备进行通信交互，向终端设备提供通信服务。核心网设备例如为5G网络核心网(core network，CN)中的设备。核心网作为承载网络提供到数据网络的接口，为终端提供通信连接、认证、管理、策略控制以及对数据业务完成承载等。
本申请实施例所涉及到的终端设备可以包括各种具有无线通信功能的手持设备、车载设备、可穿戴设备、计算设备或连接到无线调制解调器的其它处理设备。终端设备也可称为终端。终端设备也可以指用户设备(user equipment，UE)、接入终端、用户单元(subscriber unit)、用户代理、蜂窝电话(cellular phone)、智能手机(smart phone)、无线数据卡、个人数字助理(personal digital assistant，PDA)电脑、平板型电脑、无线调制解调器(modem)、手持设备(handset)、膝上型电脑(laptop computer)、机器类型通信(machine type communication，MTC)终端、高空飞机上搭载的通信设备、可穿戴设备、无人机、机器人、设备到设备通信(device-to-device，D2D)中的终端、车到一切(vehicle to everything，V2X)中的终端、虚拟现实(virtual reality，VR)终端设备、增强现实(augmented reality，AR)终端设备、工业控制(industrial control)中的无线终端、无人驾驶(self driving)中的无线终端、远程医疗(remote  medical)中的无线终端、智能电网(smart grid)中的无线终端、运输安全(transportation safety)中的无线终端、智慧城市(smart city)中的无线终端、智慧家庭(smart home)中的无线终端或者未来通信网络中的终端设备等，本申请不作限制。
本申请实施例中，终端设备还可以具备AI处理能力，网络设备也可以具备AI处理能力。例如，终端设备可以具备神经网络的训练能力、推理能力等。
本申请实施例中的第一装置可以是上述网络设备，也可以是上述终端设备，即可以是网络设备或终端设备执行本申请实施例提出的数据处理方法。可选的，第一装置是除网络设备和终端设备之外的设备。当第一装置是除网络设备和终端设备之外的设备时，第一装置执行本申请实施例所提出的数据处理方法，获得待求解问题的解，然后向网络设备或终端设备发送该待求解问题的解。
本申请实施例中的最优是指在一定条件下的最优。例如，本申请实施例中“最优的资源分配策略”是指“在一定条件下，最优的资源分配策略”。
本申请实施例中，“功率分配问题”等价于“功率分配策略”，“带宽分配问题”等价于“带宽分配策略”。
本申请公开的实施例将围绕包括多个设备、组件、模块等的系统来呈现本申请的各个方面、实施例或特征。应当理解和明白的是，各个系统可以包括另外的设备、组件、模块等，并且/或者可以并不包括结合附图讨论的所有设备、组件、模块等。此外，还可以使用这些方案的组合。
二.相关概念。
为了更好的理解本申请实施例公开的数据处理方法，对本申请实施例涉及的相关概念进行简单的介绍。
1.全连接神经网络、神经网络的训练。
全连接神经网络又叫多层感知机(multilayer perceptron，MLP)。如图2所述，一个MLP包含一个输入层，一个输出层，及多个隐藏层，且每层包括多个节点，该节点称为神经元。其中，相邻两层的神经元间两两相连。
对于相邻两层的神经元而言，下一层的神经元的输出h为所有与之相连的上一层神经元x的加权和并经过激活函数。下一层的神经元的输出h用矩阵可以表示为：
h＝f(wx+b)              (1)
其中w为权重矩阵，b为偏置向量，f为激活函数。则神经网络的输出可以递归表达为：
y＝fn(wnfn-1(...)+bn)           (2)
也就是说，可将神经网络理解为一个从输入数据集合到输出数据集合的映射关系。
神经网络的训练是指用已有数据从随机的w和b得到上述公式(2)的映射关系的过程。如图3所示，神经网络的训练的具体方式为采用损失函数(loss function)对神经网络的输出结果进行评价，并将误差反向传输，通过梯度下降的方法迭代优化w和b直到损失函数达到最小值。
其中，梯度下降的过程可以表示为：
其中，θ为待优化参数，如θ为w和b，L为损失函数。η为学习效率，用于控制梯度下降的步长。
反向传输的过程利用到求偏导的链式法则，即前一层参数的梯度可以由后一层参数的梯度递推计算得到。如图4所示，图4中神经元j与神经元i之间权重wij的梯度可以表示为：
其中，si是神经元i上的输入加权和。从上述公式(4)可以看出神经元j与神经元i之间权重wij的梯度需要根据神经元i上的梯度确定。
2.置换等变(permutation equivalence，PE)。
给定X＝[x1,x2,...,xk]，如果方程Y＝f(X)＝[y1,y2,...,yk]，对于任意的置换π(k)＝pk，均满足则该函数f对于X是置换等变的。也就是说，当函数f的变量位置发生变化时，该函数f的输出也随着变量位置的变化而变化，则该函数f是置换等变的。
具有置换等变特性的神经网络包括置换等变网络(permutation equivalence neural network，PENN)和图神经网络(graph neural network，GNN)。那么，如果一个多变量函数表示与对象排列顺序无关的策略，则称该策略为置换等变策略。
3.维度泛化特性。
如图5中的左图所示，维度泛化特性是指变量个数K发生变化时，输出yk与输入xk之间的关系不变的特性。可理解的，待求解问题中的输入数据的个数(维度)K发生变化时，输出yk与输入xk之间的关系不发生变化，则可称该待求解问题具有维度泛化特性。
而如图5中的右图所示，随着变量个数K的变化，输出yk与输入xk之间的关系发生了改变，则该右图所表征的待求解问题不具有维度泛化特性。
4.基于优化问题的无线资源分配。
基于优化问题的无线资源分配是指：在一定约束条件下，如何实现对多个终端的最优资源分配。示例性的，对于一个包含K个终端且具有置换等变特性的无线资源分配问题，其无线资源分配问题的模型为：
其中，是资源分配策略，是环境状态，pk和hk分别表示终端k的资源分配策略和环境状态，C(·)表示约束。该资源分配问题中的最优资源分配策略P*可以看作是环境状态h的函数。当h中各终端的环境状态发生位置变换时，P*中各终端的最优策略也会发生变化，从而该资源分配策略具有上述置换等变特性。
5.基于人工智能(artificial intelligence，AI)技术的无线资源分配。
采用AI技术求解上述无线资源分配问题时，为降低复杂度和减少神经网络在线决策的时间，可使用深度学习方法进行求解。可理解的，由万能近似定理(universal approximation theorem)可知，给定足够多的隐藏层，深度神经网络(deep neural network，DNN)可以以任意精度拟合任意函数。因此，网络设备可采用策略神经网络拟合最优资源分配策略与环境状态之间的函数关系。分别是策略神经网络的权重参数和偏置参数。其中，Wl表示第l-1层和第l层的连接权重，bl表示第l层的偏置，h表示环境状态。第0层为输入层，第l+1层为输出层。网络设备可采用监督学习、非监督学习等学习方法对策略神经网络的权重参数和偏置参数进行训练。
在无线通信系统中，移动终端会时不时的接入或离开网络，从而导致网络设备在进行资源分配时，考虑的终端数一直在发生改变。进而，网络设备使用AI模型解决无线资源分配问题时，该AI模型的输入数据的数量会不断发生变化。
目前，针对该维度变化的无线资源分配问题，可预先训练多个适用不同维度的AI模型。具体求解资源分配时，根据实际计算时终端的数量，选择合适维度的AI模型进行求解。或者， 在每次终端数量发生变化时，重新训练适合该终端数量的AI模型，采用实时训练的AI模型求解无线资源分配问题。然而，无论是预先训练多个AI模型，还是每次按需重新训练AI模型，都会带来较大的开销。例如，预先训练多个AI模型，会需要较大的存储资源存储这些AI模型，即会带来较大的存储开销。再例如，每次按需训练新的AI模型，会带来较大的延时开销和计算开销。
另外，上述无线资源分配问题具有置换等变特性，置换环境状态中的元素，求解上述公式(5)得到的策略中的元素也会经历相同的置换。
图6为具有置换等变特性的DNN的结构。如图6所示，当使用DNN实现函数f时，第l个隐藏层的第k个向量可基于激活函数获得，即DNN的第l层前向推理为：
其中，为第l个隐藏层的第k个向量，Ul和Vl是第l个隐藏层的权重参数子矩阵，cl是第l个隐藏层的偏置子向量，σ(·)代表激活函数，比如，σ(·)可以是softmax函数。k大于或等于1，且小于或等于K。K为终端数，且为正整数。
基于公式(6)可获得DNN的第l个隐藏层的权重参数矩阵Wl和偏置向量bl具有如下结构：

可以看出，权重参数矩阵Wl由Ul和Vl两个子矩阵按照一定规律组合而成，即Wl的对角线是Ul，非对角线上是偏置向量bl则由K个偏置子向量的cl组成。因此，上述无线资源分配问题中的终端数K发生变化时，可调整子矩阵Ul和Vl，以及向量cl的数量，组成适合该终端数的新维度神经网络，并采用该新维度神经网络求解无线资源分配策略，而无需重新训练神经网络。
然而，该无线资源分配问题不具有上述维度泛化特性，因此在某个K的取值下训练获得的子矩阵Ul和Vl，以及cl，当终端数不等于K时，无法获得最优的输出yk，即无法获得最优的资源分配策略，获得的资源分配策略的准确性较低，会给系统带来较大的损失。
同理，当无线通信系统中的其他待求解问题不具有维度泛化特性，且设备数量发生变化时，若仍基于已训练的神经网络求解待求解问题，也会降低待求解问题的解的准确性。
本申请实施例提供了一种数据处理方法100。数据处理方法100中，第一装置获取K维数据。第一装置将该K维数据输入第一机器学习模型，获得待求解问题的解。其中，第一机器学习模型包括第一处理模块和第二处理模块。第二处理模块是基于待求解问题的约束条件确定的。第二处理模块用于对K维数据进行维度泛化。第一处理模块是根据m维数据训练获得的。m的取值与K的取值无关。第一装置在求解待求解问题时，采用了对K维数据进行维度泛化的第二处理模块，从而可增强K维数据的维度泛化能力，另外采用了根据m维预先训 练获得的第一处理模块，无需因为输入数据的维度的变化而重新训练第一处理模块，可以以较低的系统开销，保证待求解问题的解的准确性。
本申请实施例还以待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，网络设备向K个终端发送通信信号的功率，以及待求解问题的约束条件包括上述发送通信信号所用的总功率在第一范围时为例，提出一种数据处理方法200。该数据处理方法200中，第一装置获取K维数据，将K维数据输入第一处理模块，获得K个第一中间解，再将K个第一中间解输入第二处理模块，获得待求解问题的解。其中，K维数据是网络设备与K个终端之间的信道增益。该申请实施例中，网络设备向K个终端发送通信信号时的功率分配问题不具有维度泛化特性，第一装置无法将第一处理模块泛化到K维数据。因此，第一装置将第一处理模块输出的第一中间解再经过归一化指示层，实现对K维数据的维度泛化。从而可使得网络设备与K个终端进行通信所使用的总带宽最小化时，求解的网络设备向K个终端发送通信信号的功率更加准确。
本申请实施例还以待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，该网络设备与K个终端进行通信所使用的带宽和功率，以及待求解问题的约束条件包括网络设备与K个终端中每个终端进行通信时的服务质量在第二范围内时为例，提出一种数据处理方法300。该数据处理方法300中，第一装置获取K维数据；第一装置将K维数据输入第一处理模块，获得K个第一中间解；第一装置将K个第一中间解输入激活层，获得K个第二中间解；第一装置将K个第二中间解输入缩放因子层，获得待求解问题的解。其中，K维数据是网络设备与K个终端之间的信道增益。该申请实施例中，网络设备与K个终端进行通信时的带宽分配问题也不具有维度泛化特性，第一装置无法将第一处理模块泛化到K维数据。因此，第一装置将第一处理模块输出的第一中间解再经过激活层和缩放因子层，实现对K维数据的维度泛化，从而可提高带宽分配问题的解的准确性，即提高网络设备与K个终端进行通信时所采用的带宽的准确性。
三.数据处理方法。
本申请实施例提出一种数据处理方法100，图7是该数据处理方法100的流程示意图。该数据处理方法100从第一装置的角度进行阐述。该数据处理方法100包括但不限于以下步骤：
S101.第一装置获取K维数据。
其中，K为正整数。该K维数据用于求解待求解问题。该待求解问题可以是无线通信系统中的问题。例如，待求解问题是如何对K个终端进行资源分配，或者是对K个信道进行信道预测，等等。因此，K维数据是基于待求解问题确定的，即待求解问题不同时，该K维数据不同。
第一装置可通过信道估计获取K维数据，或者还可通过从其他设备处接收等其他方式获取K维数据。本申请实施例不限定第一装置获取K维数据的实施方式。
S102.第一装置将K维数据输入第一机器学习模型，获得待求解问题的解，第一机器学习模型包括第一处理模块和第二处理模块，第二处理模块是基于待求解问题的约束条件确定的，第二处理模块用于对K维数据进行维度泛化，第一处理模块是根据m维数据训练获得的，m的取值与K的取值无关。
其中，m为正整数。m的取值与K的取值无关，表明预先训练获得的第一处理模块与K无关。第一处理模块的第k个输出与第k个输入有关，与输入的数量K无关。也就是说，第一处理模块可以具有置换等变特性。
第二处理模块是基于待求解问题的约束条件确定的，从而待求解问题不相同时，第二处理模块不相同，或者同一个待求解问题的约束条件不相同时，第二处理模块也不相同。
另外，第二处理模块是用于对K维数据进行维度泛化，从而可提高K维数据的维度泛化能力，进而可提高待求解问题的解的准确性。
一种可选的实施方式中，待求解问题不具有维度泛化特性。待求解问题不具有维度泛化特性时，无法将根据m维数据训练获得的第一处理模块泛化到K维数据。那么，若基于第一处理模块求解待求解问题，则获得的待求解问题的解的准确性较低。因此，第一装置基于m维数据训练获得的第一处理模块和对K维数据进行维度泛化的第二处理模块，求解待求解问题，可提高K维数据的维度泛化能力，从而获得较为准确的解。
一种可选的实施方式中，第一处理模块是具有置换等变特性的置换等变神经网络或图神经网络。可选的，第一处理模块还可以是具有置换等变特性的其他神经网络。该方式可使得第一装置求解具有置换等变特性的待求解问题时，可采用该第一处理模块。
一种可选的实施方式中，第一装置将K维数据输入第一机器学习模型，获得待求解问题的解之前，还可基于待求解问题与K之间的关系，确定待求解问题是否具有维度泛化特性。第一装置在确定待求解问题不具有维度泛化特性时，将K维数据输入第一机器学习模型中，获得待求解问题的解；第一装置在确定待求解问题具有维度泛化特性时，将K维数据输入第二机器学习模型中，获得待求解问题的解，第二机器学习模型包括第一处理模块。
也就是说，待求解问题具有维度泛化特性时，基于m维数据训练获得的第一处理模块进而泛化到K维数据，因此第一装置求解该待求解问题时，无需通过第二处理模块对K维数据进行维度泛化，可直接采用在第一处理模求解待求解问题，获得较为准确的值。
以下介绍待求解问题是否具有维度泛化特性的维度泛化条件：
本申请实施例中，待求解问题具有置换等变特性，因此对于一个PE策略y＝F(X)，其对任意的置换矩阵Π满足Πy＝F(ΠX)。其中，由Kolmogorov-Arnold表示定理可以得到该策略的第k个输出可表示为其中，φ(·)表示一个函数。令[xj]j≠k，表示从X中移除了xk的向量，将这个向量按照降序(或者升序)排列，并记排列后的向量为因为的求和结果不受其每一项的排序影响，所以当K→∞且xj独立同分布时，可得到的方差趋于0，且均值依赖于K。因此，当K→∞时，趋于一个与K相关的函数，比如，因此yk可表示为：
可见，yk在K大于预设值时趋近为一个关于xk和K的函数，即PE策略y＝F(X)在高维度下具备可分解性。该预设值是预定义的，比如，预设值为100、200等。
从而，待求解问题的维度泛化特性的判断条件可表示为：
当K,K'→∞时，如果一个PE策略y＝F(X)对任意K≠K'满足：yk＝f(xk,K)＝f(xk,K')，那么在维度K时训练好的神经网络可以泛化到维度，该PE策略的待求解问题具有维度泛化特性，否则该PE策略的待求解问题不具有维度泛化特性。
可见，表征待求解问题的函数与变量的个数K无关时，该待求解问题具有维度泛化特性；表征待求解问题的函数与变量的个数K相关时，该待求解问题不具有维度泛化特性。也就是说，随着变量的个数K的变化，表征待求解问题的函数不发生变化时，该待求解问题具有维度泛化特性；随着变量的个数K的变化，表征待求解问题的函数发生改变，则该待求解问题不具有维度泛化特性。
一种可选的实施方式中，第一装置确定待求解问题不具有维度泛化特性时，将K维数据输入第一机器学习模型，获得待求解问题的解，包括：将K维数据输入第一处理模块，获得第一中间解；然后将K个第一中间解输入第二处理模块，获得待求解问题的解。
可见，第一装置在求解待求解问题时，将K维数据输入预先训练的第一处理模块，获得K个第一中间解，再通过进行维度泛化的第二处理模块对K个第一中间解进行维度泛化，从而实现对K维数据的维度泛化，进而可使得求解的解更加准确。
示例性的，第一处理模块为PENN时，第一机器学习模型的结构可如图8所示。该PENN是基于m维数据训练获得的。该第二处理模块用于对K维数据进行维度泛化。第一装置将K维数据输入PENN，再将PENN的输出(K个第一中间解)输入第二处理模块，第二处理模块输出待求解问题的解。第二处理模块对K个第一中间解进行了维度泛化，从而可提升K维数据的维度泛化能力，进而可提高待求解问题的解的准确性。
可见，本申请实施例中，第一装置在求解待求解问题时，不仅采用了预先训练获得的第一处理模块，还采用了对K维数据进行维度泛化的第二处理模块，从而可增强K维数据的维度泛化能力，提高待求解问题的解的准确性。
另外，该方法使得第一装置在低维度训练获得的第一机器学习模型可直接应用于高维度下的通信问题，而无需重新训练学习模型，从而可节省开销。
本申请实施例还以待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，网络设备向K个终端发送通信信号的功率，以及待求解问题的约束条件包括网络设备向K个终端发送通信信号的总功率在第一范围为例，提出一种数据处理方法200，图9是该数据处理方法200的流程示意图。该数据处理方法200可应用于第一装置中。该数据处理方法200包括但不限于以下步骤：
S201.获取K维数据。
本申请实施例中，待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，网络设备向K个终端发送通信信号的功率。该待求解问题的约束条件包括网络设备向K个终端发送通信信号的总功率在第一范围内。该约束条件还包括K个终端由同一个单天线网络设备服务，该网络设备与K个终端进行通信所采用的带宽相等，K个终端中每个终端的服务质量在第二范围内。K为正整数。
可见，待求解问题为单天线网络设备服务K个终端，网络设备和每个终端通信采用相同的带宽，网络设备向K个终端发送通信信号所用的总功率受限，以及每个终端的服务质量受限时，以最小化网络设备和K个单终端进行通信所用的总带宽为目标，网络设备向K个终端发送通信信号的功率分配问题。也就是说，待求解问题是特定条件约束下，一个网络设备与K个终端进行下行通信时，网络设备向K个终端发送通信信号的功率分配问题。
本申请实施例中，上述K维数据是网络设备与K个终端之间的信道增益。第一装置可以是网络设备，或者第一装置是除网络设备之外的其他设备。第一装置是网络设备时，终端设备进行信道估计，获得K维数据，终端设备再向网络设备发送该K维数据，从而第一装置从 终端设备处获取K维数据。可选的，第一装置是终端设备时，第一装置进行信道估计获得K维数据。可选的，第一装置是除终端设备和网络设备之外的设备时，第一装置也从终端设备处获取K维数据，该K维数据是终端设备进行信道估计获得的。
S202.将K维数据输入第一处理模块，获得K个第一中间解。
一种可能的实现中，上述待求解问题的模型可表示为：
其中，B为网络设备与每个终端进行通信时所使用的带宽，Pk为网络设备向第k个终端发送数据的发射功率，g＝[g1,g2,...,gK]为网络设备和K个终端之间的信道增益，也即为K维数据，N0为噪声功率，sk为第k个终端的接收吞吐，s0为终端的最小吞吐需求，Pmax为网络设备发送总功率的最大值。N0是第k个终端设备进行噪声估计获得的，然后第k个终端设备反馈给第一装置。s0是基于第k个终端设备的通信需求确定的。Pmax是通信系统指定的，比如是网络设备指定的。
因此，上述K个终端的总功率在第一范围内是指，K个终端的总功率小于或等于Pmax；上述K个终端中每个终端的服务质量在第二范围内是指，每个终端的接收吞吐大于或等于s0。
一种可选的实施方式中，第一装置求解上述功率分配问题之前，确定该功率分配问题是否具有维度泛化特性。上述公式(10)中的sk是关于Pk和B的联合凸函数。因此，该功率分配问题是凸问题。第一装置根据卡罗需-库恩-塔克条件(Karush-Kuhn-Tucker，KKT)条件，获得该功率分配问题的全局最优解为：
当K足够大时，该功率分配问题的最优解可被近似为：
其中，IE(·)为求期望运算。从公式(12)可以看出，最优的功率分配与终端数K有关。因此，公式(12)中的功率分配问题不符合维度泛化条件。那么，第一机器学习模型包括第一处理模块和第二处理模块，第一处理模块是根据m为数据训练获得的，m的取值与K的取值无关。第二处理模块用于对K维数据进行维度泛化。从而第一装置求解上述功率分配问题时，可通过第二处理模块对K维数据进行维度泛化，获得较为准确的解。
公式(12)的一种可能的变形为：
其中，σs是归一化指数函数，例如，softmax函数，因此，如图10所示，第一机器学习模型中的第一处理模块可以为PENN或GNN，即采用PENN或GNN训练 第一机器学习模型的第二处理模块可以为归一化指数激活层，该归一化指数激活层可包括softmax函数，该softmax函数用于对K维数据进行维度泛化。
本申请实施例中，可采用上述梯度反向传递方式对PENN、GNN，以及归一化指数函数进行训练，其训练过程不再详述。
示例性的，如图10所示，第一处理模块为PENN时，将K维数据输入第一处理模块，获得K个第一中间解是指，将[g1,g2,...,gK]输入该PENN中，获得PENN输出的K个第一中间解，即为该PENN是第一装置根据网络设备和m个终端之间的尺度信道增益训练获得的，m的取值与K的取值无关，该尺度信道增益可以为小尺度信道增益，也可以为大尺度信道增益。
本申请实施例中，小尺度信道增益表征收发端距离较近时信道自身衰弱对信号传输的影响，例如，信道的多径衰弱和多普勒衰弱对信号传输的影响。大尺度信道增益表征收发端距离较远时，路径损耗对信号传输的影响。
本申请实施例中的功率分配问题不具有维度泛化特性，因此在m维数据训练获得的PENN无法泛化到K维数据，从而将K维数据输入PENN获得的K个第一中间解具有较低的准确性，无法直接将K个第一中间解作为网络设备向K个终端发送通信信号的功率分配策略。
S203.获取待求解问题的解，待求解问题的解是将K个第一中间解输入第二处理模块获得的。
一种可能的实现中，第二处理模块包括归一化指数激活层。如图10所示，获取待求解问题的解是指，将输入归一化指数激活层，即将作为softmax函数的输入，获得输出的该为待求解问题的解，也即为网络设备向K个终端发送通信信号的功率分配策略。
归一化指数激活层可对K维数据进行维度泛化，因此将PENN输出的输入到该归一化指数激活层，可将PENN的输出结果泛化到维度K，从而可提高K维数据的维度泛化能力，进而将归一化指数激活层的输出作为功率分配问题的解，可提高该解的准确性。
本申请实施例中，第一装置是网络设备时，第一装置获得待求解问题的解后，基于该解(即向K个终端发送通信信号的功率)向K个终端发送通信信号。可选的，第一装置是终端设备，或者而是除网络设备和终端设备之外的设备时，第一装置获得待求解问题的解后，向网络设备发送该将待求解问题的解，以使得网络设备基于该解(即向K个终端发送通信信号的功率)，向K个终端发送通信信号。
可见，本申请实施例中，待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，网络设备向K个终端发送通信信号的功率，以及待求解问题的约束条件包括网络设备向K个终端发送通信信号的总功率在第一范围时，将K维数据输入预先训练获得的第一处理模块，再将第一处理模块输出的第一中间解输入包括归一化指数层的第二处理模块，获得待求解问题的解。该功率分配问题不具有维度泛化特性，无法将第一处理模块泛化到K维数据。因此，将第一处理模块输出的第一中间解再经过归一化指数激活层，实现对K维数据的维度泛化，从而使得总带宽最小化时，求解的功率更加准确。
可见，上述数据处理方法200是以一个网络设备与K单天线个终端设备进行下行传输为例，求解网络设备向K个终端发送通信信号的功率分配。当一个网络设备与K个终端进行上行传输，求解K个终端向该网络设备发送通信信号的功率分配问题时，即待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，K个终端向网络设备发送通信信号的功率，以及待求解问题的约束条件包括K个终端向网络设备发送通信信号的总功率在第三范围时，其实施方式与上述数据处理方法200中的实施方式类似，在此不再赘述。
本申请实施例还以待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，网络设备向K个终端发送通信信号所使用的带宽和功率，以及约束条件包括网络设备与K个终端中每个终端进行通信时的服务质量在第二范围内为例，提出一种数据处理方法300，图11是该数据处理方法300的流程示意图。该数据处理方法300可应用于第一装置。该数据处理方法300包括但不限于以下步骤：
S301.获取K维数据。
本申请实施例中，待求解问题是一个网络设备与K个终端的进行通信所使用的总带宽最小化时，网络设备向K个终端发送通信信号所使用的功率和带宽。该待求解问题的约束条件包括网络设备与K个终端中每个终端进行通信时的服务质量在第二范围内。该约束条件还包括K个终端由同一个配备有N根天线的网络设备服务，且K个终端与网络设备进行通信时所采用的带宽不相等，网络设备向K个终端发送通信信号的总功率在第一范围内。
可见，待求解问题为多天线网络设备服务K个终端，每个终端可能采用不同的带宽，网络设备向K个终端发送通信信号的总功率受限，以及每个终端的服务质量受限时，以最小化网络设备向K个终端发送通信信号的总带宽为目标，网络设备向K个天线终端发送通信信号的带宽分配问题和功率分配问题。也就是说，待求解问题为一定约束条件下，一个网络设备与K个终端进行下行通信时的带宽分配问题和功率分配问题。
本申请实施例中，K维数据是网络设备与K个终端之间的信道增益。第一装置可以是网络设备，或者第一装置是除网络设备之外的其他设备。第一装置是网络设备时，终端进行信道估计，获得K维数据，终端再向网络设备发送该K维数据，从而第一装置从终端处获取K维数据。可选的，第一装置是终端时，第一装置进行信道估计获得K维数据。可选的，第一装置是除终端和网络设备之外的设备时，第一装置也从终端处获取K维数据，该K维数据是终端进行信道估计获得的。
S302.将K维数据输入第一处理模块，获得K个第一中间解。
本申请实施例中，上述待求解问题的模型可表示为：
其中，Bk为第k个终端使用的带宽，Pk为网络设备向第k个终端发送数据的发射功率，且为第k个终端的有效吞吐，θ是服务质量(quality of service，QoS)指数。SE为目标吞吐，是根据终端设备的服务需求确定的。Pmax为网络设备发送总功率的最大值，是系统指定的。sk是第k个终端的可达吞吐，是基第k个终端的需求确定的，且可近似为其中，τ是传输时间，μ是数据包大小，α＝[α1,α2,...,αK]是网络设备与K个终端之间的大尺度增益，g＝[g1,g2,...,gK]为网络设备和K个终端之间的小尺度信道增益，N0为噪声功率，是第k个终端的解码错误概率，是高斯Q函数的反函数。是根据网络设备与第k个终端之间的信道条件确定的，该信道条件是指网络设备与第 k个终端之间的大尺度信道增益，该大尺度信道增益表征信道衰落。
网络设备基于KKT条件，确定该问题的最优解满足如下条件：
一种可选的实施方式中，网络设备求解公式(14)中的功率分配问题和带宽分配之前，确定该待求解问题是否具有维度泛化特性。根据公式(15)可知，由于K个终端的总功率Pmax的上限是固定的，因此最优功率分配策略随着K的增加而减少。又由于每个终端的功率Pk随着K的增加而减少，从而为满足终端的吞吐需求，所需带宽需增加，因此最优带宽分配策略Bk *随着K的增加而增加。综上所述，本申请实施例中的功率分配策略和带宽分配策略均不满足维度泛化特性。
那么，针对功率分配问题和带宽分配问题，可以采用包括第一处理模块和第二处理模块的第一机器学习模型进行求解。第一处理模块是根据m维数据训练获得的。第二处理模块是基于待求解问题的约束条件确定的，且第二处理模块用于对K维数据进行维度泛化。从而，采用该第一机器学习模型求解对K个终端的功率分配和带宽分配，可获得较为准确的解。
本申请实施例中求解功率分配策略的约束条件和上述数据处理方法300中的约束条件相同，因此可采用上述数据处理方法300中的第一机器模型求解网络设备向K个终端发送通信信号的功率分配，可获得较为准确的分配结果。从而，本申请实施例提出的数据处理方法适用于求解网络设备向K个终端发送通信信号时的带宽分配。
一种可选的实施方式中，求解网络设备向该K个单天终端发送通信信号所采用的带宽时，基于网络设备与该K个终端中每个终端进行通信的服务质量在第二范围内的约束条件，确定第一机器学习模型包括如图12所示的第一处理模块和第二处理模块。第一处理模块包括PENN或GNN，或其他具有置换等变特性的神经网络。
如图12所示，获取K个第一中间解是指，将K维数据输入第一处理模块，即将[α1,α2,...,αK]输入PENN中，获得K个第一中间解，K个第一中间解为
S303.获取K个第二中间解，K个第二中间解是将K个第一中间解输入激活层获得的。
第二处理模块包括激活层和缩放因子层。其中，激活层可以是softplus函数，softplus函数可表示为ζ(x)＝log(1+ex)，x为第一处理模块输出的第一中间解。
缩放因子层中的第k个缩放因子是将第k维数据和K输入缩放因子计算模块获得的。k为小于或等于K的正整数。缩放因子计算模块可如图13所示，该缩放因子计算模块包括全连接神经网络(fully connected neural network，FNN)和激活层，该激活层也可以为softplus函数。网络设备将第k个数据αk和K输入，再将FNN的输出结果输入softplus函数，可获得第k个缩放因子从而，第一装置可基于该缩放因子计算模块、K维数据和K，获得K个缩放因子，K个缩放因子用于第二处理模块中的缩放因子层。缩放因子层用于对激活层输出的第二中间解进行缩放。
如图12所示，激活层包括softplus函数，因此获取K个第二中间解可以指，将K个将输入softplus函数，获得K个第二中间解为
S304.获取待求解问题的解，待求解问题的解是将K个第二中间解输入缩放因子层获得的。
如图12所示，可以通过将输入缩放因子层，获得网络设备向K个终端发送通信信号时的带宽分配，即以获取待求解问题的解是指。
一种可能的实现中，本申请实施例中公式(14)的带宽分配问题和功率分配问题等价于如下主对偶优化问题：
其中，L(α,Pk,Bk,λk)表示拉格朗日函数，λk是拉格朗日乘子，IE(·)为求期望运算。
可以通过三个策略神经网络和一个FNN求解上述公式(14)的待求解问题。三个·策略神经网络可以分别表示为是上述图10所示的神经网络，包括PENN和以softmax函数为激活函数的归一化指数激活层，用于输出功率分配策略。为上述图12所示的神经网络，包括PENN、以softplus为激活函数的激活层和缩放层，用于输出待缩放的带宽分配策略。包括PENN和以softplus为激活函数的激活层，用于输出拉格朗日乘子。FNN为用于输出待缩放带宽分配策略的缩放因子，αk为网络设备与第k个终端之间的大尺度信道增益，θv为FNN中待训练的参数。
由上述公式(15)可知，求解公式(14)中的带宽分配问题和功率分配问题的条件与Pk和Bk均有关，即Pk和Bk是相互影响的。又由公式(16)可知，求解带宽分配问题和功率分配问题的神经网络的损失函数不仅与输入α相关，还与Pk、Bk、λk相关，即Pk和Bk之间可通过λk建立联系，从而需联合训练以获得求解上述功率分配问题的和求解上述带宽分配问题的可见，是用于建立和之间的关系。
以下介绍上述多个神经网络的联合训练方法：
S1，对全连接神经网络的参数θv进行预训练，其训练步骤如下：
S11，假设K个终端等分总功率，即
S12，基于训练参数θv。
一种可选的实施方式中，第一装置使用监督学习方法训练参数θv，即根据使用二分法求解得到训练标签定义损失函数为该为FNN的输出。再将损失函数对FNN参数θv求梯度，获得各参数的梯度，并进行反向传递，更新参数θv，实现FNN的训练，即求解获得
另一种可选的实施方式中，第一装置使用非监督学习方法训练参数θv，即将FNN训练的目标函数定义为目标吞吐SE和实际吞吐之间的差值，再通过调整FNN的参数θv最小化这一差值，从而获得
S2，联合训练三个策略神经网络，即联合训练的参数θP、θB、θλ，其训练过程如下：
S21，随机初始化θP、θB、θλ。
S22，将αk、K输入到全连接神经网络获得缩放因子对于的输出进行缩放，得到
S23，利用反向梯度传递和随机梯度下降法，更新参数θP、θB、θλ，即：


其中，γP、γB、γλ分别为三个神经网络的学习效率。
S3，固定以的输出为功率分配结果，训练更新全连接神经网络其训练方式和上述S1类似，不再赘述。
S4，重复上述S2和S3，直到达到收敛条件，比如迭代次数达到预设次数，或网络性能达到预设要求时，停止对全连接神经网络的训练。
从而，可基于训练的三个策略神经网络和一个FNN求解公式(14)中的功率分配策略和带宽分配策略。其中，第一装置采用即图10所的第一机器学习模型，求解公式(14)中的功率分配问题；第一装置采用即图12所示的第一机器学习模型求解公式(14)中的带宽分配问题。其中，图12中缩放层的缩放因子是基于图13所示的缩放因子计算模块获得的，该缩放因子计算模块中的FNN即为上述训练获得的FNN。
上述神经网络的训练过程可以离线进行，因此不限于第一装置或其它设备执行。
本申请实施例中，第一装置是网络设备时，第一装置获得待求解问题的解后，基于该解向K个终端发送通信信号。可选的，第一装置是终端设备，或者而是除网络设备和终端设备之外的设备时，第一装置获得待求解问题的解后，向网络设备发送该将待求解问题的解，以使得网络设备基于该解向K个终端发送通信信号。
可见，本申请实施例中，在约束条件包括K个终端中每个终端的服务质量在第二范围内，以K个终端的总带宽最小化为目标，求解网络设备向K个终端发送通信信号的带宽分配时，先将获取的K维数据输入第一处理模块中，获得K个第一中间解，再将K个第一中间解输入激活层中，获得K个第二中间解，然后将K个第二中间解输入缩放因子层，采用缩放因子层对K个第二中间解进行缩放，获得K个终端的带宽。通过激活层和缩放因子层对K维数据进行了维度泛化，从而可提高K维数据的维度泛化能力，进而提高网络设备向对K个终端发送通信信号时带宽分配的准确性。
可见，上述数据处理方法300是以一个网络设备与K单天线个终端设备进行下行传输为例，求解网络设备向K个终端发送通信信号时的带宽分配和功率分配。当一个网络设备与K个终端进行上行传输，求解K个终端向该网络设备发送通信信号的带宽分配问题和功率分配问题时，即待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，K个终端向网络设备发送通信信号所使用的功率和带宽，以及约束条件包括网络设备与K个终端中每个终端进行通信时的服务质量在第二范围内时，其实施方式与上述数据处理方法300中的实施方式类似，在此不再赘述。
本申请实施例以网络设备在约束条件包括K个单天线终端中每个单天线终端的服务质量在第二范围内时，以K个单天线终端的总带宽最小化为目标，求解K个单天线终端的功率和带宽分配为例，仿真了采用本申请实施例提出了数据处理方法100，以及目前所采用的一些求解方法时，通信系统的网络可用性和带宽需求两个性能。
其中，通信系统的仿真参数为：第k个终端与网络设备之间的距离dk在50至250米之间，第k个终端的路径损耗模型为35.3+37.6log(dk)，网络设备最大发射功率Pmax＝20W，网络设备天线数Nt＝8，噪声功率谱密度N0＝-173dbm/Hz，每帧内的传输时长τ＝0.05ms，包大小u＝20字节，数据的解码错误概率εk＝5e-6，数据的目标最大丢包概率εmax＝1e-5。
三个PENN和FNN的超参设置如表1所示：
表1
其中，K为单天线终端数，t为学习时间。
预训练全连接神经网络时，采用监督学习进行训练，一个样本表示为αk是由随机产生的终端位置经过路径模型获得的，K是随机从10至200之间产生的，共产生1000个预训练样本。训练神经网络时，采用无监督学习，一个样本表示为(g,α,K)，小尺度信道g由瑞利分布产生。
训练集、验证集和测试集产生方式如表2所示：
表2
其中，训练集是指网络设备训练第一机器学习模型时，K的取值。验证集是指网络设备验证已训练好的第一机器学习模型时，K的取值。测试集是指网络设备采用比训练集更大的取值测试第一机器学习模型时，K的取值。
训练集中K的取值为[10,11,…,30]，表示训练集中的K为该[10,11,…,30]集合中的任一值。验证集和测试集中K的取值也类似。每个K所产生的样本数是指K为每个值时所产生的样本数。例如，在训练集中，K取值为10时，该K值产生的样本为10个；K取值为30时，该K值所产生的样本也为10个，从而K值产生的总样本数为210个。再例如，在测试集中，K取值为25时，该K值所产生的样本数为100个；K取值为50时，该K值所产生的样本数为100个；K的取值为100时，该K值所产生样本数为100个；K的取值为200时，该K值所产生的样本数也为100个。从而在测试集中，K值所产生的总样本数为400个。
本申请实施例所仿真的性能指标为网络可用性和带宽需求。其中，网络可用性ΑK可表示为：
其中，ΝK为测试集中维度为K的所有样本集合，|ΝK|是该集合中元素的数量，I(·)为示性函数。网络可用性指标是指网络中能够满足丢包率需求的终端数占总终端数的比例，网络可用性指标越大，表示系统性能越好。
带宽需求可表示为：
其中，为第k个终端的带宽分配。该带宽需求表示所有终端所使用的总带宽。
本申请实施例基于上述仿真条件，获得仿真结果如表3所示：
表3
其中，P-PENN是本申请实施例所提出的数据处理方法300中的第一机器学习模型。M-PENN10为目前预先训练的PENN，且是在K＝10个终端场景下训练得到的模型。M-PENN200为目前预先训练的PENN，且是在K＝200个终端场景下训练得到的模型。FNN为目前预先训练的全连接神经网络。
从表3可以看出，数据处理方法300中的第一机器学习模型虽然是在K为10至30个终端场景下进行训练获得的，但当用于其他更多终端数时，即终端数为50、100、200时，网络可用性ΑK均为1，带宽需求相对于其他方案而言较小。也就是说，数据处理方法300中的第一机器学习模型对于较多终端数而言，所有终端均可满足丢包率需求，且带宽需求也相对较小。
网络设备采用M-PENN10时，终端的带宽需求虽然小于采用数据处理方法300中的第一机器学习模型时，但在多个终端的场景下，其网络可用性趋近于0，即网络可用性极差，无法应用于实际场景。网络设备采用M-PENN200时，在多个终端的场景下，可保证网络可用性为1，但是带宽需求大于采用数据处理方法300中的第一机器学习模型时的带宽需求，因此需要用到更多的资源。网络设备采用FNN时，使用了比采用数据处理方法300中的第一机器学习模型时更多的带宽，但仍然无法保证网络可用性为1。
因此，综上所述，网络设备采用数据处理方法300时，将预先在少量终端数训练获得的第一机器模型应用于终端数较大的场景下时，仍可在相对较小的带宽需求下，保障网络的可用性。也就是说，网络设备采用数据处理方法300求解资源分配时，获得的资源分配结果的准确性较高，可保障系统性能。
四.装置实施例。
为了实现上述本申请实施例提供的方法中的各功能，第一装置可以包括硬件结构和/或软件模块，以硬件结构、软件模块、或硬件结构加软件模块的形式来实现上述各功能。上述各功能中的某个功能以硬件结构、软件模块、还是硬件结构加软件模块的方式来执行，取决于技术方案的特定应用和设计约束条件。
如图14所示，本申请实施例提供了一种通信装置1400。该通信装置1400可以是第一装置的部件(例如，集成电路，芯片等等)，也可以是第一装置的部件(例如，集成电路，芯片等等)。该通信装置1400也可以是其他通信单元，用于实现本申请方法实施例中的方法。该通信装置1400可以包括：通信单元1401和处理单元1402。可选的，还可以包括存储单元1403。
在一种可能的设计中，如图14中的一个或者多个单元可能由一个或者多个处理器来实现，或者由一个或者多个处理器和存储器来实现；或者由一个或多个处理器和收发器实现；或者由一个或者多个处理器、存储器和收发器实现，本申请实施例对此不作限定。所述处理器、存储器、收发器可以单独设置，也可以集成。
所述通信装置1400具备实现本申请实施例描述的第一装置的功能。比如，所述通信装置1400包括第一装置执行本申请实施例描述的第一装置涉及步骤所对应的模块或单元或手段(means)，所述功能或单元或手段(means)可以通过软件实现，或者通过硬件实现，也可以通过硬件执行相应的软件实现，还可以通过软件和硬件结合的方式实现。详细可进一步参考前述对应方法实施例中的相应描述。
在一种可能的设计中，一种通信装置1400可包括：处理单元1402和通信单元1401，通信单元1401用于进行数据/信令收发；
处理单元1402，用于获取K维数据；
处理单元1402，还用于将所述K维数据输入第一机器学习模型，获得待求解问题的解；所述第一机器学习模型包括第一处理模块和第二处理模块；
所述第二处理模块是基于所述待求解问题的约束条件确定的；所述第二处理模块用于对所述K维数据进行维度泛化；所述第一处理模块是根据m维数据训练获得的；所述m的取值与所述K的取值无关；所述K、所述m为正整数。
一种可选的实现方式中，所述待求解问题不具有维度泛化特性。
一种可选的实现方式中，所述处理单元1402将所述K维数据输入第一机器学习模型，获得待求解问题的解，具体用于：
将所述K维数据输入所述第一处理模块，获得K个第一中间解；
将所述K个第一中间解输入所述第二处理模块，获得所述待求解问题的解。
一种可选的实现方式中，所述待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，所述网络设备向所述K个终端发送通信信号的功率；所述约束条件包括所述网络设备向所述K个终端发送通信信号的总功率在第一范围内。
另一种可选的实现方式中，所述待求解问题是K个终端与一个网络设备进行通信所使用的总带宽最小化时，所述K个终端向所述网络设备发送通信信号的功率；所述约束条件包括所述K个终端向所述网络设备发送通信信号的总功率在第三范围内。
一种可选的实现方式中，所述第二处理模块包括归一化指数函数激活层。
另一种可选的实现方式中，所述待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，所述网络设备与所述K个终端进行通信所使用的带宽；所述约束条件包括所述网络设备与所述K个终端中每个终端进行通信时的服务质量在第二范围内。
一种可选的实现方式中，所述第二处理模块包括激活层和缩放因子层；所述缩放因子层中的第k个缩放因子是将第k个数据和所述K输入缩放因子计算模块获得的；所述k为小于或等于K的正整数。
一种可选的实现方式中，所述处理单元1402将所述K个第一中间解输入所述第二处理模块，获得所述待求解问题的解，具体用于：
将所述K个第一中间解输入所述激活层，获得K个第二中间解；
将所述K个第二中间解输入所述缩放因子层，获得所述待求解问题的解。
一种可选的实现方式中，所述K维数据是所述网络设备与所述K个终端之间的信道增益。
一种可选的实现方式中，所述第一处理模块是具有置换等变特性的置换等变神经网络或图神经网络。
本申请实施例和上述所示方法实施例基于同一构思，其带来的技术效果也相同，具体原理请参照上述所示实施例的描述，不再赘述。
本申请实施例还提供一种通信装置1500，图15为通信装置1500的结构示意图。所述通信装置1500可以是第一装置，也可以是支持第一装置实现上述方法的芯片、芯片系统、或处理器等。该装置可用于实现上述方法实施例中描述的方法，具体可以参见上述方法实施例中的说明。
所述通信装置1500可以包括一个或多个处理器1501。所述处理器1501可以是通用处理器或者专用处理器等。例如可以是基带处理器、数字信号处理器、专用集成电路、现场可编程门阵列或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件或中央处理器(Central Processing Unit，CPU)。基带处理器可以用于对通信协议以及通信数据进行处理，中央处理器可以用于对通信装置(如，基站、基带芯片，终端、终端芯片，分布单元(distributed unit，DU)或集中单元(centralized unit，CU)等)进行控制，执行软件程序，处理软件程序的数据。
可选的，所述通信装置1500中可以包括一个或多个存储器1502，其上可以存有指令1504，所述指令可在所述处理器1501上被运行，使得所述通信装置1500执行上述方法实施例中描述的方法。可选的，所述存储器1502中还可以存储有数据。所述处理器1501和存储器1502可以单独设置，也可以集成在一起。
存储器1502可包括但不限于硬盘(hard disk drive，HDD)或固态硬盘(solid-state drive，SSD)等非易失性存储器，随机存储记忆体(Random Access Memory，RAM)、可擦除可编程只读存储器(Erasable Programmable ROM，EPROM)、ROM或便携式只读存储器(Compact Disc Read-Only Memory，CD-ROM)等等。
可选的，所述通信装置1500还可以包括收发器1505、天线1506。所述收发器1505可以称为收发单元、收发机、或收发电路等，用于实现收发功能。收发器1505可以包括接收器和发送器，接收器可以称为接收机或接收电路等，用于实现接收功能；发送器可以称为发送机或发送电路等，用于实现发送功能。
所述通信装置1500为第一装置：处理器1501用于执行上述数据处理方法100中的S101、S102，以及用于执行数据处理方法200中的S201、S202、S203，以及用于执行数据处理方法300中的S301、S302、S303、S304。
另一种可能的设计中，处理器1501中可以包括用于实现接收和发送功能的收发器。例如该收发器可以是收发电路，或者是接口，或者是接口电路。用于实现接收和发送功能的收发电路、接口或接口电路可以是分开的，也可以集成在一起。上述收发电路、接口或接口电路可以用于代码/数据的读写，或者，上述收发电路、接口或接口电路可以用于信号的传输或传递。
又一种可能的设计中，可选的，处理器1501可以存有指令1503，指令1503在处理器1501上运行，可使得所述通信装置1500执行上述方法实施例中描述的方法。指令1503可能固化在处理器1501中，该种情况下，处理器1501可能由硬件实现。
又一种可能的设计中，通信装置1500可以包括电路，所述电路可以实现前述方法实施例中发送或接收或者通信的功能。本申请实施例中描述的处理器和收发器可实现在集成电路(integrated circuit，IC)、模拟IC、射频集成电路(radio frequencyintegrated circuit，RFIC)、混合信号IC、专用集成电路(application specific integrated circuit，ASIC)、印刷电路板(printed circuit board，PCB)、电子设备等上。该处理器和收发器也可以用各种IC工艺技术来制造，例如互补金属氧化物半导体(complementary metal oxide semiconductor，CMOS)、N型金属氧化物半导体(nMetal-oxide-semiconductor，NMOS)、P型金属氧化物半导体(positive channel  metal oxide semiconductor，PMOS)、双极结型晶体管(Bipolar Junction Transistor，BJT)、双极CMOS(BiCMOS)、硅锗(SiGe)、砷化镓(GaAs)等。
以上实施例描述中的通信装置可以是第一装置，但本申请实施例中描述的通信装置的范围并不限于此，而且通信装置的结构可以不受图15的限制。通信装置可以是独立的设备或者可以是较大设备的一部分。例如所述通信装置可以是：
(1)独立的集成电路IC，或芯片，或，芯片系统或子系统；
(2)具有一个或多个IC的集合，可选的，该IC集合也可以包括用于存储数据，指令的存储部件；
(3)ASIC，例如调制解调器(modulator)；
(4)可嵌入在其他设备内的模块；
对于通信装置可以是芯片或芯片系统的情况，可参见图16所示的芯片的结构示意图。图16所示的芯片1600包括处理器1601和接口1602。其中，处理器1601的数量可以是一个或多个，接口1602的数量可以是多个。该处理器1601可以是逻辑电路，该接口1602可以是输入输出接口、输入接口或输出接口。所述芯片1600还可包括存储器1603。
一种设计中，对于芯片用于实现本申请实施例中第二通信装置的功能的情况：接口1602用于进行输出或接收。
所述处理器1601，用于获取K维数据；
所述处理器1601，还用于将所述K维数据输入第一机器学习模型，获得待求解问题的解；所述第一机器学习模型包括第一处理模块和第二处理模块；
所述第二处理模块是基于所述待求解问题的约束条件确定的；所述第二处理模块用于对所述K维数据进行维度泛化；所述第一处理模块是根据m维数据训练获得的；所述m的取值与所述K的取值无关；所述K、所述m为正整数。
本申请实施例中通信装置1500、芯片1600还可执行上述通信装置1400所述的实现方式。本领域技术人员还可以了解到本申请实施例列出的各种说明性逻辑块(illustrative logical block)和步骤(step)可以通过电子硬件、电脑软件，或两者的结合进行实现。这样的功能是通过硬件还是软件来实现取决于特定的应用和整个系统的设计要求。本领域技术人员可以对于每种特定的应用，可以使用各种方法实现所述的功能，但这种实现不应被理解为超出本申请实施例保护的范围。
本申请实施例和上述数据处理方法100至数据处理方法300所示方法实施例基于同一构思，其带来的技术效果也相同，具体原理请参照上述数据处理方法100至数据处理方法300所示实施例的描述，不再赘述。
本申请还提供了一种计算机可读存储介质，用于储存计算机软件指令，当所述指令被通信装置执行时，实现上述任一方法实施例的功能。
本申请还提供了一种计算机程序产品，用于储存计算机软件指令，当所述指令被通信装置执行时，实现上述任一方法实施例的功能。
本申请还提供了一种计算机程序，当其在计算机上运行时，实现上述任一方法实施例的功能。
本申请还提供了一种通信系统，该系统包括一个或多个网络设备，以及一个或多个终端设备。在另一种可能的设计中，该系统还可以包括本申请提供的方案中与网络设备、终端设备进行交互的其他设备。
上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机指令时，全部或部分地产生按照本申请实施例所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线(digital subscriber line，DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质(例如，软盘、硬盘、磁带)、光介质(例如，高密度数字视频光盘(digital video disc，DVD))、或者半导体介质(例如，SSD)等。
以上所述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以所述权利要求的保护范围为准。

Claims (28)

  1. 一种数据处理方法，其特征在于，所述方法包括：
    获取K维数据；
    将所述K维数据输入第一机器学习模型，获得待求解问题的解；所述第一机器学习模型包括第一处理模块和第二处理模块；
    所述第二处理模块是基于所述待求解问题的约束条件确定的；所述第二处理模块用于对所述K维数据进行维度泛化；所述第一处理模块是根据m维数据训练获得的；所述m的取值与所述K的取值无关；所述K、所述m为正整数。
  2. 根据权利要求1所述的方法，其特征在于，所述获取K维数据，包括：
    通过信道估计获取K维数据；或者，
    接收来自终端的K维数据，所述K维数据是所述终端进行信道估计获得的。
  3. 根据权利要求1或2所述的方法，其特征在于，所述待求解问题不具有维度泛化特性。
  4. 根据权利要求1至3任一项所述的方法，其特征在于，所述将所述K维数据输入第一机器学习模型，获得待求解问题的解，包括：
    将所述K维数据输入所述第一处理模块，获得K个第一中间解；
    将所述K个第一中间解输入所述第二处理模块，获得所述待求解问题的解。
  5. 根据权利要求1至4任一项所述的方法，其特征在于，所述待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，所述网络设备向所述K个终端发送通信信号的功率；
    所述约束条件包括所述网络设备向所述K个终端发送通信信号的总功率在第一范围内。
  6. 根据权利要求1至4任一项所述的方法，其特征在于，所述待求解问题是K个终端与一个网络设备进行通信所使用的总带宽最小化时，所述K个终端向所述网络设备发送通信信号的功率；
    所述约束条件包括所述K个终端向所述网络设备发送通信信号的总功率在第三范围内。
  7. 根据权利要求5或6所述的方法，其特征在于，所述第二处理模块包括归一化指数函数激活层。
  8. 根据权利要求1至4任一项所述的方法，其特征在于，所述待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，所述网络设备与所述K个终端进行通信所使用的带宽；
    所述约束条件包括所述网络设备与所述K个终端中每个终端进行通信时的服务质量在第二范围内。
  9. 根据权利要求8所述的方法，其特征在于，所述第二处理模块包括激活层和缩放因子 层；
    所述缩放因子层中的第k个缩放因子是将第k个数据和所述K输入缩放因子计算模块获得的；所述k为小于或等于K的正整数。
  10. 根据权利要求9所述的方法，其特征在于，所述将所述K个第一中间解输入所述第二处理模块，获得所述待求解问题的解，包括：
    将所述K个第一中间解输入所述激活层，获得K个第二中间解；
    将所述K个第二中间解输入所述缩放因子层，获得所述待求解问题的解。
  11. 根据权利要求5至10任一项所述的方法，其特征在于，所述K维数据是所述网络设备与所述K个终端之间的信道增益。
  12. 根据权利要求1至11任一项所述的方法，其特征在于，所述第一处理模块是具有置换等变特性的置换等变神经网络或图神经网络。
  13. 一种通信装置，其特征在于，所述装置包括通信单元和处理单元，通信单元用于进行数据/信令收发；
    处理单元，用于获取K维数据；
    所述处理单元，还用于将所述K维数据输入第一机器学习模型，获得待求解问题的解；所述第一机器学习模型包括第一处理模块和第二处理模块；
    所述第二处理模块是基于所述待求解问题的约束条件确定的；所述第二处理模块用于对所述K维数据进行维度泛化；所述第一处理模块是根据m维数据训练获得的；所述m的取值与所述K的取值无关；所述K、所述m为正整数。
  14. 根据权利要求13所述的装置，其特征在于，所述处理单元获取K维数据，具体用于：
    通过信道估计获取K维数据；或者，
    接收来自终端的K维数据，所述K维数据是所述终端进行信道估计获得的。
  15. 根据权利要求13或14所述的装置，其特征在于，所述待求解问题不具有维度泛化特性。
  16. 根据权利要求13至15任一项所述的装置，其特征在于，所述处理单元将所述K维数据输入第一机器学习模型，获得待求解问题的解，具体用于：
    将所述K维度数据输入所述第一处理模块，获得K个第一中间解；
    将所述K个第一中间解输入所述第二处理模块，获得所述待求解问题的解。
  17. 根据权利要求13至16任一项所述的装置，其特征在于，所述待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，所述网络设备向所述K个终端发送通信信号的功率；
    所述约束条件包括所述网络设备向所述K个终端发送通信信号的总功率在第一范围内。
  18. 根据权利要求13至16任一项所述的装置，其特征在于，所述待求解问题是K个终端与一个网络设备进行通信所使用的总带宽最小化时，所述K个终端向所述网络设备发送通信信号的功率；
    所述约束条件包括所述K个终端向所述网络设备发送通信信号的总功率在第三范围内。
  19. 根据权利要求17或18所述的装置，其特征在于，所述第二处理模块包括归一化指数函数激活层。
  20. 根据权利要求13至16任一项所述的装置，其特征在于，所述待求解问题是一个网络设备与K个终端进行通信所使用的总带宽最小化时，所述网络设备与所述K个终端进行通信所使用的带宽；
    所述约束条件包括所述网络设备与所述K个终端中每个终端进行通信时的服务质量在第二范围内。
  21. 根据权利要求20所述的装置，其特征在于，所述第二处理模块包括激活层和缩放因子层；
    所述缩放因子层中的第k个缩放因子是将第k个数据和所述K输入缩放因子计算模块获得的；所述k为小于或等于K的正整数。
  22. 根据权利要求21所述的装置，其特征在于，所述处理单元将所述K个第一中间解输入所述第二处理模块，获得所述待求解问题的解，具体用于：
    将所述K个第一中间解输入所述激活层，获得K个第二中间解；
    将所述K个第二中间解输入所述缩放因子层，获得所述待求解问题的解。
  23. 根据权利要求17至22任一项所述的装置，其特征在于，所述K维数据是所述网络设备与所述K个终端之间的信道增益。
  24. 根据权利要求13至23任一项所述的装置，其特征在于，所述第一处理模块是具有置换等变特性的置换等变神经网络或图神经网络。
  25. 一种通信装置，其特征在于，包括处理器和收发器，所述收发器用于与其它通信装置进行通信；所述处理器用于运行程序，以使得所述通信装置实现权利要求1至12任一项所述的方法。
  26. 根据权利要求25所述的装置，其特征在于，还包括存储器，用于存储所述程序。
  27. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质用于存储有指令，当其在计算机上运行时，使得权利要求1至12任一项所述的方法被执行。
  28. 一种包含指令的计算机程序产品，其特征在于，当其在计算机上运行时，使得权利要求1至12任一项所述的方法被执行。
PCT/CN2023/084530 2022-03-31 2023-03-28 一种数据处理方法及相关装置 WO2023185890A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210336391.9A CN116933855A (zh) 2022-03-31 2022-03-31 一种数据处理方法及相关装置
CN202210336391.9 2022-03-31

Publications (1)

Publication Number Publication Date
WO2023185890A1 true WO2023185890A1 (zh) 2023-10-05

Family

ID=88199212

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/084530 WO2023185890A1 (zh) 2022-03-31 2023-03-28 一种数据处理方法及相关装置

Country Status (2)

Country Link
CN (1) CN116933855A (zh)
WO (1) WO2023185890A1 (zh)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111915060A (zh) * 2020-06-30 2020-11-10 华为技术有限公司 组合优化任务的处理方法以及处理装置
CN113965233A (zh) * 2021-10-19 2022-01-21 东南大学 一种基于深度学习的多用户宽带毫米波通信资源分配方法及系统
KR20220013906A (ko) * 2020-07-27 2022-02-04 한국전자통신연구원 심층 학습 기반의 빔포밍 방법 및 이를 위한 장치

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111915060A (zh) * 2020-06-30 2020-11-10 华为技术有限公司 组合优化任务的处理方法以及处理装置
KR20220013906A (ko) * 2020-07-27 2022-02-04 한국전자통신연구원 심층 학습 기반의 빔포밍 방법 및 이를 위한 장치
CN113965233A (zh) * 2021-10-19 2022-01-21 东南大学 一种基于深度学习的多用户宽带毫米波通信资源分配方法及系统

Also Published As

Publication number Publication date
CN116933855A (zh) 2023-10-24

Similar Documents

Publication Publication Date Title
WO2020187004A1 (zh) 通信系统中的调度方法、装置及存储介质
US20230342593A1 (en) Neural network training method and related apparatus
WO2023036268A1 (zh) 一种通信方法及装置
WO2023020502A1 (zh) 数据处理方法及装置
CN116848828A (zh) 机器学习模型分布
US11456834B2 (en) Adaptive demodulation reference signal (DMRS)
WO2023185890A1 (zh) 一种数据处理方法及相关装置
WO2021114968A1 (zh) 一种调度方法及装置
US10872045B1 (en) Method and apparatus for symbol detection
CN112888076B (zh) 一种调度方法及装置
WO2023197950A1 (zh) 一种通信方法及相关装置
WO2024008004A1 (zh) 一种通信方法及装置
WO2022262687A1 (zh) 一种数据处理方法及装置
WO2022237865A1 (zh) 一种数据处理方法及装置
US20230259742A1 (en) Communication method, apparatus, and system
WO2024045888A1 (zh) 一种处理装置及控制方法
US20240049188A1 (en) Gradient transmission method and related apparatus
WO2022218386A1 (zh) 分布式学习方法及装置
WO2024046419A1 (zh) 一种通信方法及装置
WO2023179577A1 (zh) 一种通信方法及相关装置
CN117155805B (zh) 一种基于统计机器学习的设备管理方法及装置
WO2024051789A1 (zh) 一种波束管理方法
WO2023036280A1 (zh) 一种模型测试方法及装置
WO2024044881A1 (zh) 一种数据处理方法、训练方法及相关装置
WO2024046215A1 (zh) 一种通信方法及装置

Legal Events

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

Ref document number: 23778255

Country of ref document: EP

Kind code of ref document: A1