WO2022237865A1 - 一种数据处理方法及装置 - Google Patents

一种数据处理方法及装置 Download PDF

Info

Publication number
WO2022237865A1
WO2022237865A1 PCT/CN2022/092404 CN2022092404W WO2022237865A1 WO 2022237865 A1 WO2022237865 A1 WO 2022237865A1 CN 2022092404 W CN2022092404 W CN 2022092404W WO 2022237865 A1 WO2022237865 A1 WO 2022237865A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
machine learning
learning model
model
optimization problem
Prior art date
Application number
PCT/CN2022/092404
Other languages
English (en)
French (fr)
Inventor
郭佳
杨晨阳
王坚
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 WO2022237865A1 publication Critical patent/WO2022237865A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • 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
    • 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
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means

Definitions

  • the present application relates to the technical field of data processing, and in particular to a data processing method and device.
  • AI artificial intelligence
  • the training data is input into the fully connected neural network, the parameters of the fully connected neural network are trained, and the trained fully connected neural network can solve the optimization problem. But for different optimization problems, the same fully connected neural network structure is used to solve them. Therefore, in order to be able to adapt to the solution of different optimization problems, a large amount of training data is often used, resulting in higher training complexity.
  • Embodiments of the present application provide a data processing method and device for reducing the amount of training data in a machine learning process and reducing training complexity.
  • the embodiment of the present application provides a data processing method, the method includes: obtaining first data, the first data is determined according to the optimization goal of the optimization problem and the second data; inputting the first data into the first machine learning In the model, the first inference result is obtained.
  • the second data is used as the environment data of the optimization problem
  • the first data is obtained according to the optimization goal of the optimization problem and the second data, that is, the first data takes into account the influence of the environment data on the optimization goal of the optimization problem
  • the second One data is input into the first machine learning model, which means that the first machine learning model has known the impact of environmental data on the optimization objective of the optimization problem, so it is not necessary to use a large amount of training data for training to learn how the environmental data affects the optimization problem
  • the optimization goal which can reduce the amount of training data in the training process and reduce the training complexity.
  • the method does not depend on iterative algorithms, and can solve optimization problems with or without iterative algorithms, so it can be applied to the solution of more optimization problems.
  • the second data can also be input into the first function of the second machine learning model to determine the first data
  • the first function is determined according to the optimization objective of the optimization problem
  • the first The second machine learning model also includes the first machine learning model.
  • the first function can be used to determine the first data based on the second data, which is equivalent to the influence of the known environmental data on the optimization objective of the optimization problem in the first function, which can be obtained without training.
  • the second machine learning model is the tth one of T cascaded second machine learning models, T is a positive integer, and t is less than or equal to T.
  • the first data when the first data is determined according to the optimization goal of the optimization problem and the second data, the first data can be determined according to the optimization goal of the optimization problem, the second data and the second reasoning The result is determined, and the second reasoning result is the initial reasoning result or the reasoning result output by the t-1th second machine learning model.
  • the first function also takes into account the influence of the reasoning results of the previous level when determining the first data, so that the calculated results of the first data can be more accurate, thereby further improving the accuracy of the solution results of the optimization problem.
  • the first machine learning model determines the first reasoning result, it also takes into account the influence of the second reasoning result and/or the second data, which can further improve the accuracy of the solution result of the optimization problem.
  • the first data when the first data is determined according to the optimization goal of the optimization problem and the second data, the first data can be determined according to the optimization goal of the optimization problem, the second data, and the The constraints of the optimization problem are determined.
  • the first function also takes into account the influence of the constraints of the optimization problem on the optimization objective of the optimization problem when determining the first data, so that the calculated results of the first data can be more accurate, thereby further improving the solution result of the optimization problem accuracy.
  • the second machine learning model is the Tth of T cascaded second machine learning models
  • the Tth The parameters of the second machine learning model are adjusted, so as to train the second machine learning model.
  • the second machine learning model further includes a dimensionality reduction model, and the dimensionality reduction model is used to perform dimensionality reduction processing on the second data.
  • Key information in the second data can be extracted during the dimensionality reduction process, which can reduce the calculation amount of the second machine learning model and reduce the calculation complexity.
  • the second data before the first data is input into the first machine learning model, the second data can also be sorted, which can reduce the amount of training data required for training and reduce the complexity of training. Spend.
  • the second data includes channel data and/or communication scene information.
  • the communication scenario information may include indoor environment information and/or outdoor environment information, or the communication scenario information may also include urban dense environment information and/or suburban environment information.
  • the embodiment of the present application provides a data processing model
  • the data processing model includes a T-level second machine learning model
  • the second machine learning model includes the first function and the first machine learning model.
  • the first function is used to determine the first data according to the optimization objective of the optimization problem and the second data, and input the first data into the first machine learning model.
  • the first machine learning model is used to solve the optimization problem.
  • the first machine learning model can be used to output a first reasoning result according to the first data, and the first reasoning result is a solution to the optimization problem.
  • the input of the data processing model may be the second data, and the output may be the first reasoning result.
  • the second data can be used as the input of the second machine learning model of each level.
  • the first inference result output by the T-th level second machine learning model is the output of the data processing model.
  • the optional first reasoning result output by the non-T-level second machine learning model is the input of the next-level second machine learning model.
  • the first function can also be used to determine the first data according to the optimization objective of the optimization problem, the second data and the second reasoning result. If the first function belongs to the second machine learning model of the first level, the second inference result is the initial inference result; if the first function does not belong to the second machine learning model of the first level, the second inference result is the second machine learning of the previous level The first inference result output by the model.
  • the first function can also be used to determine the first data according to the optimization objective of the optimization problem, the second data and the constraints.
  • the first function can also be used to determine the first data according to the optimization objective of the optimization problem, the second data, the constraints and the second reasoning result.
  • the first machine learning model can be used to output a first reasoning result according to the first data and the second data.
  • the first machine learning model can be used to output the first reasoning result according to the first data and the second reasoning result.
  • the first machine learning model can be used to output the first reasoning result according to the first data, the second reasoning result and the second data.
  • the above-mentioned second data may be data obtained after dimensionality reduction processing and/or sorting processing are performed on the second data.
  • the second machine learning model may also include a dimensionality reduction subnetwork, the input of the dimensionality reduction subnetwork is the second data, and the output of the dimensionality reduction subnetwork is obtained after dimensionality reduction processing of the second data data.
  • the output of the optional dimensionality reduction sub-network can be used as the input of the first function, and/or can be used as the input of the first machine learning model.
  • the embodiment of the present application provides a data processing device, and the device may have the function of realizing the above-mentioned first aspect or any possible design of the first aspect.
  • the functions of the above-mentioned data processing device may be realized by hardware, and may also be realized by executing corresponding software by hardware, and the hardware or software includes one or more modules corresponding to the above-mentioned functions.
  • the device may include: an acquisition unit and a processing unit.
  • the obtaining unit is used to obtain the first data, and the first data is determined according to the optimization goal of the optimization problem and the second data;
  • the processing unit is used to input the first data into the first machine learning model to obtain the first inference As a result, the first machine learning model is used to solve the optimization problem.
  • the acquisition unit is specifically configured to input the second data into the first function of the second machine learning model to determine the first data, and the first function is determined according to the optimization objective of the optimization problem , the second machine learning model further includes the first machine learning model.
  • the second machine learning model is the tth one of T cascaded second machine learning models, T is a positive integer, and t is less than or equal to T.
  • the first data is determined according to the optimization objective of the optimization problem, the second data and the second reasoning result, and the second reasoning result is the initial reasoning result or the t-1th An inference result output by the second machine learning model.
  • the processing unit is specifically configured to input at least one of the second inference result and the second data, as well as the first data, into the first machine learning model.
  • the first data is determined according to an optimization objective of the optimization problem, the second data, and constraints of the optimization problem.
  • the device may further include: an adjustment unit configured to, when the second machine learning model is the Tth one of T cascaded second machine learning models, according to the first As a result of reasoning, adjust the parameters of the T second machine learning models.
  • the second machine learning model further includes a dimensionality reduction model, and the dimensionality reduction model is used to perform dimensionality reduction processing on the second data.
  • the device may further include: a sorting unit, configured to sort the second data.
  • the second data includes channel data and/or communication scene information.
  • the embodiment of the present application provides a data processing device, and the device may have the function of realizing the above-mentioned first aspect or any possible design of the first aspect.
  • the structure of the device includes at least one processor, and may also include at least one memory. At least one processor is coupled with at least one memory, and can be used to execute computer program instructions stored in the memory, so that the device executes the method in the above first aspect or any possible design of the first aspect.
  • the device further includes a communication interface, and the processor is coupled to the communication interface.
  • the communication interface may be a transceiver or an input/output interface; when the device is a chip included in the server, the communication interface may be an input/output interface of the chip.
  • the transceiver may be a transceiver circuit, and the input/output interface may be an input/output circuit.
  • the embodiment of the present application provides a chip system, including: a processor, the processor is coupled with a memory, and the memory is used to store programs or instructions, and when the programs or instructions are executed by the processor , so that the system-on-a-chip implements the above-mentioned first aspect or the method in any possible design of the first aspect.
  • the chip system further includes an interface circuit for receiving code instructions and transmitting them to the processor.
  • processors in the chip system, and the processors may be implemented by hardware or by software.
  • the processor may be a logic circuit, an integrated circuit, or the like.
  • the processor may be a general-purpose processor implemented by reading software codes stored in a memory.
  • the memory can be integrated with the processor, or can be set separately from the processor, which is not limited in this application.
  • the memory can be a non-transitory processor, such as a read-only memory ROM, which can be integrated with the processor on the same chip, or can be respectively arranged on different chips.
  • the setting method of the processor is not specifically limited.
  • the embodiment of the present application provides a data processing device, including a processor and an interface circuit; the interface circuit is configured to receive code instructions and transmit them to the processor; the processor is configured to run the code instructions To implement the method in the above first aspect or any possible design of the first aspect.
  • the embodiment of the present application provides a readable storage medium, on which a computer program or instruction is stored, and when the computer program or instruction is executed, the computer executes any one of the above-mentioned first aspect or the first aspect. possible design approach.
  • the embodiment of the present application provides a computer program product, which enables the computer to execute the method in the above-mentioned first aspect or any possible design of the first aspect when the computer reads and executes the computer program product.
  • Figure 1 is a schematic diagram of the architecture of a fully connected neural network
  • Fig. 2 is a schematic diagram of loss function optimization
  • Fig. 3 is a schematic diagram of gradient backpropagation
  • FIG. 4 is a schematic diagram of training based on a fully connected neural network
  • FIG. 5 is a schematic diagram of a data processing flow provided by an embodiment of the present application.
  • FIG. 6 is a schematic diagram of a first machine learning model provided by an embodiment of the present application.
  • FIG. 7 is a schematic diagram of a first machine learning model provided by an embodiment of the present application.
  • FIG. 8 is a schematic diagram of a first machine learning model provided by an embodiment of the present application.
  • FIG. 9 is a schematic diagram of a second machine learning model provided by an embodiment of the present application.
  • FIG. 10 is a schematic diagram of a second machine learning model provided in the embodiment of the present application.
  • Fig. 11 is a schematic diagram of a second machine learning model provided by the embodiment of the present application.
  • FIG. 12 is a schematic diagram of a second machine learning model provided by the embodiment of the present application.
  • FIG. 13 is a schematic structural diagram of a communication system provided by an embodiment of the present application.
  • FIG. 14 is a schematic diagram of a simulation of the amount of required training data provided by the embodiment of the present application.
  • FIG. 15 is a schematic structural diagram of a data processing device provided in an embodiment of the present application.
  • FIG. 16 is a schematic structural diagram of a data processing device provided by an embodiment of the present application.
  • the present application provides a data processing method and device, aiming at reducing the number of training samples and training complexity in the process of machine learning.
  • the method and the device are based on the same technical conception. Since the principle of solving the problem of the method and the device is similar, the implementation of the device and the method can be referred to each other, and the repetition will not be repeated.
  • optimization problems different application scenarios have corresponding optimization problems.
  • the application scenarios may include data classification, data regression, data clustering, spam filtering, web page retrieval, natural language processing, image recognition, speech recognition, and wireless communication.
  • different optimization problems can be set according to different requirements, which is not limited here.
  • the application scenario includes wireless communication
  • the optimization problem may include power allocation, network optimization, mobility management, channel coding, or channel prediction.
  • the optimization goal can also be set according to different needs.
  • different optimization objectives may be set, for example, the optimization objective of power allocation may include throughput maximization and/or delay minimization, and the like.
  • a constraint condition of the optimization problem may also be set, and the constraint condition is used to constrain/limit the solution of the optimization problem, that is, the solution of the optimization problem satisfies the constraint condition.
  • a solution to an optimization problem may include one or more (approximately) optimal solutions, or may include one or more suboptimal solutions.
  • Machine learning refers to a technology that does not rely on explicit instruction codes, but solves optimization problems based on algorithms/models used in reasoning.
  • the commonly used machine learning includes (with) supervised learning and unsupervised learning (or unsupervised learning).
  • a training set is provided, and the training data in the training set is labeled with the correct label (generally the optimal solution to the optimization problem).
  • the mapping relationship between training data and labels can be learned through the training set, and the mapping relationship is applicable to data outside the training set.
  • mapping relationship In unsupervised learning, training data is provided. Through the training data and the optimization target, the mapping relationship between the training data and the solution of the optimization problem can be learned. The mapping relationship can be used to solve the optimization problem and obtain the solution of the optimization problem.
  • Training data can also be referred to as training samples, or sample data.
  • the machine learning model can use the input data to output the inference result, which is the solution of the inferred optimization problem, and realizes the solution of the optimization problem.
  • the machine learning model satisfies constraints when solving the optimization problem.
  • Machine learning models may include mathematical models and/or neural network models and the like.
  • the mathematical model can be an algorithm, such as an algorithm for classification (such as K-proximity algorithm), an algorithm for regression (such as linear regression algorithm), and an algorithm for clustering (such as K-means algorithm).
  • an algorithm for classification such as K-proximity algorithm
  • an algorithm for regression such as linear regression algorithm
  • an algorithm for clustering such as K-means algorithm
  • the neural network (neural networks, NN) model is usually a model based on the neural network architecture, and the neural network includes deep neural network (deep neural network, DNN), fully connected neural network (fully connected neural network, FCNN), convolutional neural network ( convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GAN), etc.
  • the neural network model can achieve better results and performance when solving large-scale optimization problems.
  • MLP multilayer perceptron
  • An MLP includes an input layer, an output layer, and a hidden layer.
  • the number of input layer, output layer and hidden layer can be one or more, and each (neural network) layer includes a plurality of neurons.
  • MLP includes an input layer (including neurons x1, x2, x3 and x4), multiple hidden layers (as shown in the middle part of Figure 1) and an output layer (including neurons y1, y2, y3, y4, y5 and y6), the neurons of two adjacent layers are connected in pairs.
  • f the activation function
  • w the weight matrix
  • b the bias vector
  • x Neurons in the previous layer connected to neurons in the next layer.
  • it is obtained by weighting each neuron x connected to it in the previous layer, and using an activation function to process the weighted results.
  • a neural network can be understood as a mapping relationship from input data (such as the above training data) to output data (such as the above labels).
  • the process of obtaining this mapping relationship based on (randomly initialized) w and b and training data can be called the training (or learning) process of the neural network.
  • the horizontal axis is the value of the neural network parameters (such as w and/or b), and the vertical axis is the output of the loss function Value, the curve is the mapping relationship between the neural network parameters and the loss function.
  • the initial neural network parameters (the starting point shown in Figure 2) are set, and the training is performed based on the initial neural network parameters.
  • the loss function is used to calculate the error between the output result of the neural network and the labeling result of the training data , that is, the loss function is used to evaluate the output of the neural network, and the error is backpropagated, and the parameters of the neural network are iteratively optimized by the method of gradient descent until the output value of the loss function reaches the minimum value, at which point the optimal point is reached.
  • the process of gradient descent can be expressed as: ⁇ is the neural network parameter to be optimized (such as w and/or b), L is the loss function, and ⁇ is the learning rate, which is used to control the step size of gradient descent.
  • the process of backpropagation can be realized by using the chain rule for partial derivatives, that is, the gradient of the parameters of the previous layer can be recursively calculated from the gradient of the parameters of the latter layer, as shown in Figure 3, the parameters of the latter layer are The parameters of the previous layer are The formula can be expressed as: Where w ij is the weight of neuron j connected to neuron i, and s i is the weighted sum of inputs on neuron i.
  • the machine learning model involved in the embodiments of the present application may be a model obtained through machine learning, or a model in machine learning.
  • the machine learning model is a neural network model
  • the model obtained through machine learning may refer to a trained model
  • the model in machine learning may refer to a model to be trained.
  • the model obtained through machine learning can determine the solution of the optimization problem.
  • Electronic equipment also called equipment or nodes, including terminal equipment and/or network equipment.
  • Terminal equipment also called user equipment (UE) is a device with a wireless transceiver function, which can be accessed through an access network device (or also called an access network) in a radio access network (radio access network, Ingress device) communicates with one or more core network (core network, CN) devices (or may also be referred to as core devices).
  • UE user equipment
  • CN core network
  • User equipment may also be called an access terminal, terminal, subscriber unit, subscriber station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, user agent, or user device, among others.
  • User equipment can be deployed on land, including indoor or outdoor, handheld or vehicle-mounted; it can also be deployed on water (such as ships, etc.); it can also be deployed in the air (such as on aircraft, balloons, and satellites, etc.).
  • the user equipment can be a cellular phone, a cordless phone, a session initiation protocol (SIP) phone, a smart phone, a mobile phone, a wireless local loop (WLL) Station, personal digital assistant (PDA), etc.
  • SIP session initiation protocol
  • WLL wireless local loop
  • PDA personal digital assistant
  • the user equipment can also be a handheld device with wireless communication function, a computing device or other devices connected to a wireless modem, a vehicle device, a wearable device, a drone device or a terminal in the Internet of Things, the Internet of Vehicles, the fifth generation Mobile communication (5th-generation, 5G) network and any form of terminal in the future network, relay user equipment or terminal in the future evolved PLMN, etc.
  • the relay user equipment may be, for example, a 5G residential gateway (residential gateway, RG).
  • the user equipment can be a virtual reality (virtual reality, VR) terminal, an augmented reality (augmented reality, AR) terminal, a wireless terminal in industrial control (industrial control), a wireless terminal in self driving (self driving), telemedicine Wireless terminals in remote medical, wireless terminals in smart grid, wireless terminals in transportation safety, wireless terminals in smart city, and smart home wireless terminals, etc.
  • the embodiment of the present application does not limit the type or category of the terminal device.
  • the device for realizing the function of the terminal device may be the terminal device, or may be a device capable of supporting the terminal device to realize the function, such as a chip system, and the device may be installed in the terminal device.
  • the system-on-a-chip may be composed of chips, or may include chips and other discrete devices.
  • Network equipment refers to equipment that can provide wireless access functions for terminal equipment.
  • the network device may support at least one wireless communication technology, such as long term evolution (long term evolution, LTE), NR, and the like.
  • LTE long term evolution
  • NR NR
  • a network device may include an access network device (also referred to as an access node or a node).
  • the network equipment includes but is not limited to: a next-generation node B (generation nodeB, gNB), an evolved node B (evolved node B, eNB), a radio network controller (radio network controller, RNC), and a node in a 5G network.
  • B node B, NB
  • home base station for example, home evolved node B, or home node B, HNB
  • baseband unit baseband unit, BBU
  • sending and receiving point transmitting and receiving point, TRP
  • transmitting point transmitting point
  • TP mobile switching center
  • small station small station
  • micro station etc.
  • the network device may also be a wireless controller, a centralized unit (centralized unit, CU), and/or a distributed unit (distributed unit, DU) in a cloud radio access network (cloud radio access network, CRAN) scenario, or the network device may For relay stations, access points, vehicle-mounted devices, terminals, wearable devices, device-to-device (Device-to-Device, D2D), vehicle-to-everything (V2X), machine-to-machine (machine-to- -machine, M2M) communication, Internet of Things (Internet of Things) communication equipment that assumes the base station function, or network equipment in future mobile communication, or network equipment in the future evolution of public land mobile network (PLMN) Wait.
  • a wireless controller a centralized unit (centralized unit, CU), and/or a distributed unit (distributed unit, DU) in a cloud radio access network (cloud radio access network, CRAN) scenario
  • the network device may For relay stations, access points, vehicle-mounted devices, terminal
  • the network device may include a core network (CN) device, and the core network device includes, for example, an access and mobility management function (access and mobility management function, AMF) and the like.
  • CN core network
  • AMF access and mobility management function
  • a plurality referred to in this application refers to two or more than two.
  • the word "exemplary” is used as an example, illustration or explanation. Any embodiment or design described herein as “example” is not to be construed as preferred or advantageous over other embodiments or designs. Rather, the use of the word example is intended to present concepts in a concrete manner.
  • AI When solving optimization problems, if the problem scale of the optimization problem is relatively large and the solution complexity is relatively high, it can be realized through AI.
  • the neural network in AI has the characteristics of high offline training complexity and low online computing complexity, so offline training and online computing can be performed on optimization problems based on neural networks.
  • the training data is input into a fully connected neural network (FCNN), and the parameters of the FCNN can be trained using supervised learning or unsupervised learning.
  • FCNN can infer the power allocation strategy of environmental data, according to the power allocation strategy and target power
  • the difference in the allocation strategy determines the loss function, and trains the parameters of the FCNN;
  • the FCNN can infer the power allocation strategy of the environmental data, and determine the loss function according to the system performance corresponding to the power allocation strategy and the target system performance. Train the parameters of the FCNN.
  • FCNN FCNN structure is used to solve different optimization problems. Therefore, in order to be able to adapt to the solution of different optimization problems, a large amount of training data is often used, resulting in higher training complexity. And when the environmental data changes, if the generalization ability of FCNN is exceeded, FCNN needs to be retrained to adapt to the new environmental data, which will also lead to higher training complexity.
  • an embodiment of the present application provides a data processing method.
  • the second data is used as the environment data of the optimization problem
  • the first data is obtained according to the optimization goal of the optimization problem and the second data, that is, the first data takes into account the influence of the environment data on the optimization goal of the optimization problem
  • the second One data is input into the first machine learning model, which means that the first machine learning model has known the impact of environmental data on the optimization objective of the optimization problem, so it is not necessary to use a large amount of training data for training to learn how the environmental data affects the optimization problem
  • the optimization goal which can reduce the amount of training data in the training process and reduce the training complexity.
  • FIG. 5 shows a possible data processing method provided by the embodiment of the present application.
  • the data processing method can be applied to a data processing device, and the data processing device can be an electronic device or be located in the electronic device.
  • the data processing method includes the following steps:
  • S502 Input the first data into the first machine learning model to obtain a first inference result.
  • the first data is determined according to the optimization objective of the optimization problem and the second data.
  • the first data may be determined according to the optimization objective of the optimization problem and the second data, so as to obtain the first data.
  • the first data may be directly obtained from the storage medium, that is, the first data may be determined in advance according to the optimization objective of the optimization problem and the second data, and then the first data may be stored in the storage medium.
  • the second reasoning result is the initial reasoning result or the previous level reasoning result (if any).
  • the optimization problem is power allocation
  • the optimization objective is throughput maximization and/or delay minimization, which can optionally be achieved by solving a power allocation strategy.
  • the second data may include channel data and/or communication scene information and the like.
  • the channel data is related to the channel between the transceiver pair (referring to a pair of transmitter and receiver), wherein the quality of the channel between the transceiver pair will affect the power allocation.
  • the communication scene information may include indoor scene information and/or outdoor scene information, or may further include urban dense area information and/or suburban information.
  • the occlusion between the transceiver pairs in the indoor scene and the outdoor scene are different, which may affect the channel quality between the transceiver pairs, and the number of transceiver pairs included in the urban dense area and the suburban area is different, the transceiver The channels between the pairs are subject to different interference conditions, which may affect the channel quality between the transceiver pairs, thereby affecting the power allocation.
  • the constraint condition may be related to the communication scene information, for example, the maximum transmit power of the transmitter in the indoor scene may be different from the maximum transmit power of the transmitter in the outdoor scene.
  • the communication scene information can be input in the form of an image, for example, the input image is an image containing map information, and the location of the terminal and the location of the network device are marked in the image, so it can be analyzed that the communication scene is a dense urban environment or a suburban area Environment; if the input image is an image of an indoor environment, the image includes images of furniture and household appliances, so it can be analyzed that the communication scene is an indoor environment.
  • the second data may include road images collected by cameras.
  • road images can contain images of lanes, traffic lights, and vehicles, so that it can be identified whether the vehicle has changed lanes illegally or ran a red light.
  • the first data F can be determined through a first function.
  • T(H) can be pre-designed in the first function, that is, the impact of the known environmental data (including the second data) on the optimization objective of the optimization problem in the first function, so that there is no need to train the first machine learning model to learn how the environmental data affects
  • the optimization objective of an optimization problem can reduce the amount of training data required.
  • C(H,p) can also be pre-designed in the first function, so that it is not necessary to train the constraints of the optimization problem, and the amount of required training data can be reduced. That is to say, T(H) and C(H,p) can be pre-modeled according to the experience of relevant personnel.
  • T(H) and C(H,p) can also be obtained through training, and there is no limitation here.
  • the first machine learning model can be used to solve an optimization problem.
  • the optimization problem and optimization goal can be determined according to the application scenario and actual needs, and there is no limitation here.
  • the optimization problem is power allocation
  • a power allocation strategy may be determined for K objects, where K is a positive integer.
  • the first machine learning model may be a machine learning model to be trained, and correspondingly, the first data may be data participating in training, and the second data may be environment data participating in training.
  • the first machine learning model may be a trained machine learning model, and correspondingly, the first data may be data to be inferred, and the second data may be environment data to be inferred.
  • the first machine learning model is a machine learning model to be trained
  • the first data takes into account the impact of environmental data on the optimization objective of the optimization problem, which is equivalent to the fact that the first machine learning model has Know the impact of environmental data on the optimization goal of the optimization problem, so there is no need to use a large amount of training data for training to learn how the environmental data affects the optimization goal of the optimization problem, so the amount of training data in the training process can be reduced, and the training complexity can be reduced .
  • the first machine learning model may include a mathematical model, or the first machine learning model may include a neural network model, or the first machine learning model may include a mathematical model and a neural network model. Wherein the neural network model includes one or more neural network layers.
  • the second data may also be processed to further reduce the amount of training data and the complexity of training.
  • dimension reduction processing and sorting processing may be performed on the second data.
  • the dimensionality reduction process can be realized by a dimensionality reduction model, which can be a neural network model (such as a dimensionality reduction sub-network), or other dimensionality reduction methods (such as a principal component analysis (PCA) model, or an autoencoder (autoEncoder)).
  • PCA principal component analysis
  • autoEncoder autoEncoder
  • the sorting process can use the property of permutation invariance to sort the second data in a set order (such as from small to large or from large to small).
  • At least one of the second inference result and the second data, and the first data may also be input into the first machine learning model, which can improve the accuracy of inference of the first machine learning model.
  • the data processing device includes T-level first machine learning models, that is, includes T cascaded first machine learning models, or T first machine learning models for iteration (a total of T iterations) , T is a positive integer.
  • T-level first machine learning models can be directly cascaded, and the first machine learning models at each level can obtain the first data.
  • the first function is cascaded before the first machine learning model, that is, the first function can be cascaded between two first machine learning models.
  • a dimensionality reduction model is cascaded before the first machine learning model and/or the first function.
  • the first machine learning model is the t-th one of the T cascaded first machine learning models, and t is less than or equal to T.
  • the input of the first machine learning model includes first data.
  • the input of the first machine learning model includes first data and second data.
  • the input of the first machine learning model is the second reasoning result, the second data and the first data.
  • the input of the T-level first machine learning model is the first data, the second data (optional) and the initial reasoning result (optional) ), the output of the T-level first machine learning model is the final reasoning result.
  • the first data input by the first first machine learning model can be based on the optimization goal of the optimization problem, the second data, and the constraints Conditions (optional) and initial inference results (optional) are determined.
  • the second reasoning result is an initial reasoning result
  • the initial reasoning result may be determined based on experience of relevant personnel, or may be randomly generated.
  • the 1st first machine learning model outputs the first inference result, which can be used to determine the first data input by the (t+1)th first machine learning model, and optionally as the (t+1)th An input to a machine learning model.
  • the first machine learning model is the t-th first machine learning model (at this time, t is greater than 1, and t is less than T)
  • the first data input by the t-th first machine learning model can be based on the optimization objective of the optimization problem
  • the second data, constraints (optional) and (t-1)th level reasoning results (optional) are determined.
  • the (t-1)th level inference result is the first inference result output by the (t-1)th first machine learning model
  • the (t-1)th first machine learning model is related to the t-th first
  • the previous level of the machine learning model cascade is the first machine learning model.
  • the first inference result output by the tth first machine learning model can be used to determine the first data input by the (t+1)th first machine learning model, and optionally as the (t+1)th An input to the first machine learning model.
  • the first data input by the T first machine learning model can be optimized according to The optimization objective of the problem, the second data, the constraints (optional) and the (T-1) level inference results (optional) are determined.
  • the first inference result output by the T-th first machine learning model may be used as the final inference result.
  • parameters of the T first machine learning models may be adjusted according to the final inference result.
  • the T first machine learning parameters can be adjusted according to the error between the final inference result and the expected result (ie label) of the first data.
  • the parameters of the T first machine learning can be adjusted according to the final reasoning result and the optimization goal.
  • the data processing device includes T-level second machine learning models, that is, includes T cascaded second machine learning models, or iterates T second machine learning models (a total of T iterations ), T is a positive integer.
  • the second machine learning model includes a concatenation of the first function and the first machine learning model.
  • the second machine learning model includes a dimensionality reduction model, and the dimensionality reduction model is cascaded before the first machine learning model and/or the first function.
  • the second machine learning model is the t-th one of the T cascaded second machine learning models, and t is less than or equal to T.
  • the input of the first machine learning model of the second machine learning model is the first data.
  • the input of the first machine learning model of the second machine learning model is the first data and the second data.
  • the input of the first machine learning model of the second machine learning model is the second reasoning result, the second data and the first data.
  • the first function is a bijective function about the second data and the first data, That is, the second data has a one-to-one relationship with the first data; the input of the optional first function also includes the second reasoning result, then the second data, the second reasoning result (optional) and the first data have a one-to-one relationship One-to-one correspondence; b, for any two sets of inputs (second data, optional second inference results), the first data obtained by the first function are the same, then when the second machine learning model converges, according to the arbitrary two The same is true for the first inference result determined by the set (second data, optional second inference result) input.
  • the second machine learning model also includes a dimensionality reduction subnetwork, and the dimensionality reduction subnetwork can input the data after dimensionality reduction processing on the second data into the first machine learning model.
  • the data after performing dimension reduction processing on the second data may also be input into the first function.
  • the input of the T-level second machine learning model is the first data, the second data (optional) and the initial reasoning result (optional) ), the output of the T-level second machine learning model is the final reasoning result.
  • the first second machine learning model inputs the second data
  • the first function can be based on the optimization objective of the optimization problem, the first function For the second data, the constraint condition (optional) and the initial reasoning result (optional) determine the first data.
  • the second reasoning result is the initial reasoning result.
  • the first inference result output by the first second machine learning model can be used to determine the first data input by the (t+1)th second machine learning model, and optionally as the (t+1)th One input to the two machine learning models.
  • the second machine learning model is the t-th second machine learning model (at this time, t is greater than 1, and t is less than T), and the t-th first machine learning model inputs the second data
  • the first function can be optimized according to the optimization problem
  • the target, the second data, the constraints (optional) and the initial inference result (optional) determine the first data.
  • the second inference result is the first inference result output by the (t-1)th second machine learning model.
  • the first inference result output by the tth second machine learning model can be used to determine the first data input by the (t+1)th second machine learning model, and optionally as the (t+1)th An input to the second machine learning model.
  • the first function can be The first data is determined according to the optimization objective of the optimization problem, the second data, constraint conditions (optional) and initial reasoning results (optional).
  • the first inference result output by the T-th second machine learning model may be used as the final inference result.
  • the parameters of the T second machine learning models may be adjusted according to the final inference result.
  • all or some parameters of the T second machine learning models may be adjusted according to the error between the final inference result and the expected result (ie label) of the second data.
  • all or some parameters of the T second machine learning models may be adjusted according to the final reasoning result and optimization goal.
  • the second machine learning model also includes a dimensionality reduction subnetwork
  • the parameters of the first machine learning model and the dimensionality reduction subnetwork can be trained at the same time, or the parameters of the dimensionality reduction subnetwork are fixed, while the parameters of the first machine learning model to train.
  • the parameters of the dimensionality reduction subnetwork may be parameters after the dimensionality reduction subnetwork training is completed.
  • the embodiment of the present application includes a T1-level first machine learning model and a T2-level second machine learning model, where T1 is a positive integer and T2 is a positive integer, and T1 and T2 may be the same or different.
  • first machine learning model and the second machine learning model can refer to the above methods, which will not be repeated here.
  • the second data is used as the environment data of the optimization problem
  • the first data is obtained according to the optimization goal of the optimization problem and the second data, that is, the first data takes into account the optimization goal of the environment data for the optimization problem
  • inputting the first data into the first machine learning model means that the first machine learning model knows the impact of the environmental data on the optimization objective of the optimization problem, so there is no need to use a large amount of training data for training to learn How the environmental data affects the optimization objective of the optimization problem, so that the amount of training data in the training process can be reduced and the training complexity can be reduced.
  • the embodiments of the present application do not depend on iterative algorithms, and can solve optimization problems with iterative algorithms or without iterative algorithms, so it is applicable to solving more optimization problems.
  • the data processing method provided in the embodiment of the present application is applicable to a wireless communication scenario.
  • Wireless communication scenarios can be applied to various communication systems, including terrestrial communication systems, systems where terrestrial communication systems and satellite communication systems are integrated, or short-distance communication systems.
  • the mobile communication system may be a 4G communication system (for example, a long term evolution (long term evolution, LTE) system), a 5G communication system (for example, a new radio (new radio, NR) system), and a future mobile communication system.
  • FIG. 13 is a schematic diagram of a possible communication system architecture provided by an embodiment of the present application.
  • the communication system includes a network device 1310 and two terminal devices 1320, where the network device can provide communication services to the two terminal devices, and the two terminal devices The devices communicate separately, and two terminal devices can communicate with each other. It can be understood that the number of network devices in the communication system may be one or more, and the number of terminal devices may be one or more, and other possible types of network devices and terminal devices are not limited.
  • K objects such as K users or network devices or transceiver peers
  • K is a positive integer
  • k is less than or equal to K
  • T k (H, p) is a function related to the kth object
  • the aggregate function can be obtained by a summation algorithm, or by a minimum value algorithm, and of course can also be obtained by other algorithms, which are not limited here.
  • C k (H,p) is the constraint condition of the kth object, is the feasible region of the constraints of the kth object.
  • the optimization problem is power allocation
  • the optimization goal is throughput maximization
  • the environment data H of the optimization problem includes channel data, and the channel data may be a vector matrix composed of channels between transceiver pairs, where the electronic device can obtain the local channels of the electronic device and the channels between neighboring devices.
  • the electronic device may be a terminal device or a network device.
  • T k (H,p) can be expressed as:
  • C k (H,p) may be the maximum transmit power P max allowed by the kth transmitter.
  • the mapping relationship between the optimal power allocation strategy and H can be learned by training the neural network model, such as the mapping between the optimal power allocation strategy and H
  • the set value can be 20, 30, or less or more, and there is no limit to the set value here.
  • the neural network model includes a first function and a first machine learning model.
  • the neural network model can represent the mapping relationship between p and H, and learn the mapping relationship between the optimal power allocation strategy p * and H through training.
  • the mapping relationship between the input and output of the neural network model can be expressed as: ⁇ is the parameter to be trained of the neural network model, such as the weight matrix and/or bias vector mentioned above.
  • the constraints in the neural network model may include the user data rate upper limit obtained according to the Shannon formula, and the structure of the neural network model may be called a data rate-based knowledge neural network (DRNN).
  • DRNN data rate-based knowledge neural network
  • the first function can output the vector F according to the optimization goal of the optimization problem, the environmental data H (ie the above-mentioned second data), the vector p (ie the above-mentioned second reasoning result) and C k (H, p) (ie the above-mentioned constraint conditions) (i.e. the first data mentioned above).
  • the first machine learning model is used to determine the power allocation strategy, so the first machine learning model may also be called a policy allocation model, and the policy allocation model may be realized based on a neural network.
  • the mapping relationship between the input and output of the policy allocation model can be expressed as: If the input of the policy allocation model includes vector F and environmental data H, the mapping relationship between the input and output of the policy allocation model can be expressed as: If the input of the policy allocation model includes vector F, vector p and environmental data H, the mapping relationship between the input and output of the policy allocation model can be expressed as:
  • the neural network model can be designed as an iterative structure, that is, the neural network model includes a multi-level second machine learning model, and each level of the second machine learning model includes the first function and the second machine learning model A machine learning model.
  • the iterative update value p (t) of the output power allocation strategy p of the t-th level second machine learning model (as the second reasoning result above), that is, each level of the second machine learning model updates the power allocation strategy once.
  • the p (t) output by the second machine learning model at level t is the input of the second machine learning model at level (t+1), and the output of the second machine learning model at level (t+1) is the same as that of the second machine learning model at level t
  • the relationship of the output of the machine learning model can be expressed as: or or
  • the iterative function between the (t+1)th level second machine learning model and the tth level second machine learning model can be expressed as: or or or If H is processed by a dimensionality reduction sub-network, the iterative function can be expressed as or or is the output of the dimensionality reduction sub-network, is the parameter of the dimensionality reduction sub-network. Among them, H can also undergo sorting processing.
  • the input of the first-level second machine learning model may include an initial power allocation strategy p (0) .
  • the second machine learning model at the T-th level outputs p (T) .
  • the training of the neural network model can adopt supervised learning or unsupervised learning.
  • the training set includes N tr training data (that is, the environmental data participating in the training) H [i] , and each input training data H [i] corresponds to the output [i] means corresponding to the i-th training data.
  • the training goal of the neural network can be to minimize the loss function, which measures the actual output and expected output of the neural network model for the i-th training data gap between.
  • the purpose of training the neural network model is to minimize the loss function by adjusting the parameters of the neural network.
  • the neural network model parameters that minimize the loss function can be expressed as: in is the actual output and expected output of the i-th training data
  • the gap is the loss function.
  • the actual output and expected output of the i-th training data The gap can be expressed by the square error, then the loss function can be mean squared error (mean squared error, MSE).
  • the neural network model can deduce an (approximately) optimal solution based on the acquired environmental data H, and then send the (approximately) optimal solution to other electronic devices that communicate wirelessly with the electronic device.
  • the network device may send the inferred optimal power allocation strategy to other network devices and terminal devices within the communication range of the network device.
  • the amount of training data required by the data processing method provided by the embodiment of the present application will be described below in combination with simulation results.
  • Table 1 also shows the amount of training data required by the data processing method provided by the embodiment of the present application and the amount of training data required by the FCNN.
  • the amount of training data required for training by the data processing method provided by the embodiment of the present application is much smaller than that of FCNN, which can reduce the complexity of training.
  • the embodiment of the present application further provides a data processing device, which can be used to implement the method described in the above method embodiment.
  • the device may include corresponding hardware structures and/or software modules for performing various functions.
  • the present application can be implemented in the form of hardware or a combination of hardware and computer software in combination with the units and algorithm steps of each example described in the embodiments disclosed herein. Whether a certain function is executed by hardware or computer software drives hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the present application.
  • the data processing apparatus 1500 may include an acquisition unit 1501 and a processing unit 1502 .
  • the data processing apparatus 1500 may be an electronic device, or may be a chip provided in the electronic device.
  • the obtaining unit 1501 is configured to obtain first data, the first data is determined according to the optimization objective of the optimization problem and the second data;
  • the processing unit 1502 is configured to input the first data into the first machine learning model, and output a first reasoning result.
  • the obtaining unit 1501 is specifically configured to input the second data into a first function of the second machine learning model to determine the first data, the first function is determined according to the optimization objective of the optimization problem,
  • the second machine learning model also includes the first machine learning model.
  • the second machine learning model is the t-th one of T cascaded second machine learning models, T is a positive integer, and t is less than or equal to T.
  • the first data is determined according to the optimization objective of the optimization problem, the second data and the second reasoning result, and the second reasoning result is the initial reasoning result or the t-1th second The inference results output by the machine learning model.
  • the processing unit 1502 is specifically configured to input at least one of the second inference result and the second data, as well as the first data, into the first machine learning model.
  • the first data is determined according to an optimization objective of the optimization problem, the second data, and constraints of the optimization problem.
  • the data processing apparatus 1500 may further include: an adjustment unit 1503 configured to, when the second machine learning model is the Tth one of T cascaded second machine learning models, according to the An inference result, adjusting the parameters of the T second machine learning models.
  • the second machine learning model further includes a dimensionality reduction model, and the dimensionality reduction model is used to perform dimensionality reduction processing on the second data.
  • the data processing apparatus 1500 may further include: a sorting unit 1504, configured to sort the second data.
  • the second data includes channel data and/or communication scene information.
  • the division of units in the embodiment of the present application is schematic, and is only a logical function division, and there may be other division methods in actual implementation.
  • Each functional unit in the embodiment of the present application may be integrated into one processing module, each unit may exist separately physically, or two or more units may be integrated into one unit.
  • the above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.
  • the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.
  • the technical solution of the present application is essentially or part of the contribution to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including several instructions to make a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor (processor) execute all or part of the steps of the methods described in the various embodiments of the present application.
  • the aforementioned storage media include various media that can store program codes.
  • the data processing device 1600 includes at least one processor 1602 , and optionally, at least one communication interface 1604 . Further, the data processing device 1600 may further include a memory 1606, and the memory 1606 is used to store computer programs or instructions.
  • the memory 1606 can be a memory inside the processor, or a memory outside the processor. In the case that each unit module described in FIG. 15 is implemented by software, the software or program codes required by the processor 1602 to perform corresponding actions are stored in the memory 1606 .
  • the processor 1602 is configured to execute programs or instructions in the memory 1606 to implement the steps shown in FIG. 5 in the above-mentioned embodiments.
  • the communication interface 1604 is used to enable communication between the device and other devices.
  • the memory 1606, the processor 1602 and the communication interface 1604 are connected to each other through a bus 1608, and the bus 1608 may be a peripheral component interconnect standard (peripheral component interconnect, PCI) bus or extended industry standard architecture (extended industry standard architecture, EISA) bus, etc.
  • PCI peripheral component interconnect
  • EISA extended industry standard architecture
  • the bus can be divided into address bus, data bus, control bus and so on. For ease of representation, only one thick line is used in FIG. 16 , but it does not mean that there is only one bus or one type of bus.
  • the embodiment of the present application also provides a chip system, including: a processor, the processor is coupled with a memory, and the memory is used to store programs or instructions, and when the programs or instructions are executed by the processor, the The chip system implements the methods in the foregoing method embodiments.
  • processors in the chip system there may be one or more processors in the chip system.
  • the processor can be realized by hardware or by software.
  • the processor may be a logic circuit, an integrated circuit, or the like.
  • the processor may be a general-purpose processor implemented by reading software codes stored in a memory.
  • the memory can be integrated with the processor, or can be set separately from the processor, which is not limited in this application.
  • the memory can be a non-transitory processor, such as a read-only memory ROM, which can be integrated with the processor on the same chip, or can be respectively arranged on different chips.
  • the setting method of the processor is not specifically limited.
  • the chip system may be a field programmable gate array (field programmable gate array, FPGA), an application specific integrated circuit (ASIC), or a system on chip (SoC), It can also be a central processing unit (central processor unit, CPU), it can also be a network processor (network processor, NP), it can also be a digital signal processing circuit (digital signal processor, DSP), it can also be a microcontroller (micro controller unit, MCU), and can also be a programmable logic device (programmable logic device, PLD) or other integrated chips.
  • FPGA field programmable gate array
  • ASIC application specific integrated circuit
  • SoC system on chip
  • each step in the foregoing method embodiments may be implemented by an integrated logic circuit of hardware in a processor or instructions in the form of software.
  • the method steps disclosed in connection with the embodiments of the present application may be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.
  • the embodiment of the present application also provides a computer-readable storage medium, the computer-readable instruction is stored in the computer-readable medium, and when the computer reads and executes the computer-readable instruction, the computer is made to execute the method in the above-mentioned embodiment. method.
  • the embodiments of the present application also provide a computer program product, which enables the computer to execute the methods in the above method embodiments when the computer reads and executes the computer program product.
  • processors mentioned in the embodiments of the present application may be a CPU, or other general-purpose processors, DSP, ASIC, FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and the like.
  • a general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like.
  • the memory mentioned in the embodiments of the present application may be a volatile memory or a nonvolatile memory, or may include both volatile and nonvolatile memories.
  • the non-volatile memory can be read-only memory (read-only memory, ROM), programmable read-only memory (programmable ROM, PROM), erasable programmable read-only memory (erasable PROM, EPROM), electrically programmable Erases programmable read-only memory (electrically EPROM, EEPROM) or flash memory.
  • Volatile memory can be random access memory (RAM), which acts as external cache memory.
  • RAM random access memory
  • SRAM static random access memory
  • DRAM dynamic random access memory
  • DRAM synchronous dynamic random access memory
  • SDRAM double data rate synchronous dynamic random access memory
  • double data rate SDRAM double data rate SDRAM
  • DDR SDRAM enhanced synchronous dynamic random access memory
  • ESDRAM enhanced synchronous dynamic random access memory
  • serial link DRAM SLDRAM
  • direct memory bus random access memory direct rambus RAM, DR RAM
  • the processor is a general-purpose processor, DSP, ASIC, FPGA or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components
  • the memory storage module
  • sequence numbers of the above-mentioned processes do not mean the order of execution, and the execution order of the processes should be determined by their functions and internal logic, and should not be used in the embodiments of the present application.
  • the implementation process constitutes any limitation.
  • the disclosed systems, devices and methods may be implemented in other ways.
  • the device embodiments described above are schematic.
  • the division of the units is only a logical function division.
  • multiple units or components can be combined or can be Integrate into another system, or some features may be ignored, or not implemented.
  • the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • each functional unit in each embodiment of the present application may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit.
  • the functions described above are realized in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium.
  • the technical solution of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application.
  • the aforementioned storage medium includes: various media capable of storing program codes such as U disk, mobile hard disk, ROM, RAM, magnetic disk or optical disk.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Neurology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本申请实施例涉及一种数据处理方法及装置，该方法包括：获取第一数据，所述第一数据根据优化问题的优化目标和第二数据确定；将第一数据输入第一机器学习模型中，得到第一推理结果，可以减少机器学习过程中的训练数据的数量，降低训练复杂度。

Description

一种数据处理方法及装置
相关申请的交叉引用
本申请要求在2021年05月14日提交中国国家知识产权局、申请号为202110530279.4、申请名称为“一种数据处理方法及装置”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
本申请涉及数据处理技术领域，尤其涉及一种数据处理方法及装置。
背景技术
在求解优化问题时，如果优化问题的问题规模比较大，求解复杂度比较高，可以通过人工智能(artificial intelligence，AI)进行离线训练和在线计算。
在离线训练时，将训练数据输入全连接神经网络，对全连接神经网络的参数进行训练，训练完成的全连接神经网络可以实现优化问题的求解。但是对于不同的优化问题，都采用相同的全连接神经网络结构进行求解。因此为了能够适应不同的优化问题的求解，往往使用大量训练数据，导致训练复杂度更高。
发明内容
本申请实施例提供一种数据处理方法及装置，用以减少机器学习过程中的训练数据的数量，降低训练复杂度。
第一方面，本申请实施例提供一种数据处理方法，该方法包括：获取第一数据，所述第一数据根据优化问题的优化目标和第二数据确定；将第一数据输入第一机器学习模型中，得到第一推理结果。
该方法适用于训练过程和计算过程。在该方法中，第二数据作为优化问题的环境数据，第一数据根据优化问题的优化目标和第二数据得到，即第一数据考虑到了环境数据对优化问题的优化目标的影响，而将第一数据输入到第一机器学习模型，相当于第一机器学习模型已知了环境数据对于优化问题的优化目标的影响，因此无需再使用大量训练数据进行训练，以学习到环境数据如何影响优化问题的优化目标，从而可以减少训练过程中训练数据的数量，降低训练复杂度。并且该方法不依赖于迭代算法，可以求解存在迭代算法的优化问题，也可以求解不存在迭代算法的优化问题，因此可以适用更多优化问题的求解。
在一种可能的设计中，还可以将所述第二数据输入第二机器学习模型的第一函数，确定第一数据，所述第一函数根据所述优化问题的优化目标确定，所述第二机器学习模型还包括所述第一机器学习模型。第一函数可以用于根据第二数据确定第一数据，相当于第一函数中已知环境数据对于优化问题的优化目标的影响，不需要进行训练就可以得到。
在一种可能的设计中，所述第二机器学习模型为T个级联的第二机器学习模型中的第t个，T为正整数，t小于或等于T。通过多级第二机器学习模型的迭代，可以进一步提高优化问题的求解结果的准确性。
在一种可能的设计中，所述第一数据根据优化问题的优化目标和第二数据确定时，所述第一数据可以根据所述优化问题的优化目标，所述第二数据和第二推理结果确定，所述第二推理结果为初始推理结果或第t-1个第二机器学习模型输出的推理结果。第一函数在确定第一数据时还考虑到前一级推理结果的影响，因此可以使得计算出的第一数据的结果更加准确，从而进一步提高优化问题的求解结果的准确性。
在一种可能的设计中，将第一数据输入第一机器学习模型中时，可以将所述第二推理结果和所述第二数据中的至少一个，以及所述第一数据，输入第一机器学习模型中。第一机器学习模型在确定第一推理结果时，还考虑到第二推理结果和/或第二数据的影响，可以进一步提高优化问题的求解结果的准确性。
在一种可能的设计中，所述第一数据根据优化问题的优化目标和第二数据确定时，所述第一数据可以根据所述优化问题的优化目标，所述第二数据，和所述优化问题的约束条件确定。第一函数在确定第一数据时还考虑到所述优化问题的约束条件对优化问题的优化目标的影响，因此可以使得计算出的第一数据的结果更加准确，从而进一步提高优化问题的求解结果的准确性。
在一种可能的设计中，在所述第二机器学习模型为T个级联的第二机器学习模型中的第T个时，还可以根据所述第一推理结果，对所述T个第二机器学习模型的参数进行调整，从而对第二机器学习模型进行训练。
在一种可能的设计中，所述第二机器学习模型还包括降维模型，所述降维模型用于对所述第二数据进行降维处理。降维过程中可以提取出第二数据中的关键信息，可以减少第二机器学习模型的计算量，降低计算复杂度。
在一种可能的设计中，所述将第一数据输入第一机器学习模型中之前，还可以对所述第二数据进行排序处理，可以减少训练时所需的训练数据的数量，降低训练复杂度。
在一种可能的设计中，所述第二数据包括信道数据和/或通信场景信息。在优化问题应用于无线通信场景，通信场景信息可以包括室内环境信息和/或室外环境信息，或者通信场景信息也可以包括市区密集环境信息和/或郊区环境信息等。
第二方面，本申请实施例提供一种数据处理模型，数据处理模型包括T级第二机器学习模型，第二机器学习模型包括第一函数和第一机器学习模型。第一函数用于根据优化问题的优化目标和第二数据确定第一数据，将第一数据输入第一机器学习模型。第一机器学习模型用于求解优化问题，第一机器学习模型可以用于根据第一数据，输出第一推理结果，第一推理结果为优化问题的解。
其中数据处理模型的输入可以为第二数据，输出可以为第一推理结果。其中第二数据可以作为每级第二机器学习模型的输入。第T级第二机器学习模型输出的第一推理结果为数据处理模型的输出。可选的非第T级第二机器学习模型输出的第一推理结果为下一级第二机器学习模型的输入。
在一种可能的设计中，第一函数还可以用于根据优化问题的优化目标，第二数据和第二推理结果确定第一数据。若第一函数属于第1级第二机器学习模型，第二推理结果为初始推理结果，若第一函数不属于第1级第二机器学习模型，第二推理结果为前一级第二机器学习模型输出的第一推理结果。
在一种可能的设计中，第一函数还可以用于根据优化问题的优化目标，第二数据和约束条件确定第一数据。
在一种可能的设计中，第一函数还可以用于根据优化问题的优化目标，第二数据，约束条件和第二推理结果确定第一数据。
在一种可能的设计中，第一机器学习模型可以用于根据第一数据和第二数据，输出第一推理结果。
在一种可能的设计中，第一机器学习模型可以用于根据第一数据和第二推理结果，输出第一推理结果。
在一种可能的设计中，第一机器学习模型可以用于根据第一数据，第二推理结果和第二数据，输出第一推理结果。
在一种可能的设计中，上述第二数据可以为对第二数据经过降维处理和/或排序处理后得到的数据。
在一种可能的设计中，第二机器学习模型还可以包括降维子网络，降维子网络的输入为第二数据，降维子网络的输出为对第二数据经过降维处理后得到的数据。可选的降维子网络的输出可以作为第一函数的输入，和/或可以作为第一机器学习模型的输入。
第三方面，本申请实施例提供一种数据处理装置，该装置可具有实现上述第一方面或第一方面的任一种可能的设计中的功能。上述数据处理装置的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模块。
示例性的，该装置可以包括：获取单元和处理单元。其中，获取单元，用于获取第一数据，所述第一数据根据优化问题的优化目标和第二数据确定；处理单元，用于将第一数据输入第一机器学习模型中，得到第一推理结果，所述第一机器学习模型用于求解所述优化问题。
在一种可能的设计中，获取单元，具体用于将所述第二数据输入第二机器学习模型的第一函数，确定第一数据，所述第一函数根据所述优化问题的优化目标确定，所述第二机器学习模型还包括所述第一机器学习模型。
在一种可能的设计中，所述第二机器学习模型为T个级联的第二机器学习模型中的第t个，T为正整数，t小于或等于T。
在一种可能的设计中，所述第一数据根据所述优化问题的优化目标，所述第二数据和第二推理结果确定，所述第二推理结果为初始推理结果或第t-1个第二机器学习模型输出的推理结果。
在一种可能的设计中，所述处理单元，具体用于将所述第二推理结果和所述第二数据中的至少一个，以及所述第一数据，输入第一机器学习模型中。
在一种可能的设计中，所述第一数据根据所述优化问题的优化目标，所述第二数据，和所述优化问题的约束条件确定。
在一种可能的设计中，该装置还可以包括：调整单元，用于在所述第二机器学习模型为T个级联的第二机器学习模型中的第T个时，根据所述第一推理结果，对所述T个第二机器学习模型的参数进行调整。
在一种可能的设计中，所述第二机器学习模型还包括降维模型，所述降维模型用于对所述第二数据进行降维处理。
在一种可能的设计中，所述装置还可以包括：排序单元，用于对所述第二数据进行排序处理。
在一种可能的设计中，第二数据包括信道数据和/或通信场景信息。
第四方面，本申请实施例提供一种数据处理装置，该装置可具有实现上述第一方面或第一方面的任一种可能的设计中的功能。
该装置的结构中包括至少一个处理器，还可以包括至少一个存储器。至少一个处理器与至少一个存储器耦合，可用于执行存储器中存储的计算机程序指令，以使装置执行上述第一方面或第一方面的任一种可能的设计中的方法。可选地，该装置还包括通信接口，处理器与通信接口耦合。当装置为服务器时，该通信接口可以是收发器或输入/输出接口；当该装置为服务器中包含的芯片时，该通信接口可以是芯片的输入/输出接口。可选地，收发器可以为收发电路，输入/输出接口可以是输入/输出电路。
第五方面，本申请实施例提供一种芯片系统，包括：处理器，所述处理器与存储器耦合，所述存储器用于存储程序或指令，当所述程序或指令被所述处理器执行时，使得该芯片系统实现上述第一方面或第一方面的任一种可能的设计中的方法。
可选地，该芯片系统还包括接口电路，该接口电路用于接收代码指令并传输至所述处理器。
可选地，该芯片系统中的处理器可以为一个或多个，该处理器可以通过硬件实现也可以通过软件实现。当通过硬件实现时，该处理器可以是逻辑电路、集成电路等。当通过软件实现时，该处理器可以是一个通用处理器，通过读取存储器中存储的软件代码来实现。
可选地，该芯片系统中的存储器也可以为一个或多个。该存储器可以与处理器集成在一起，也可以和处理器分离设置，本申请并不限定。示例性的，存储器可以是非瞬时性处理器，例如只读存储器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，本申请对存储器的类型，以及存储器与处理器的设置方式不作具体限定。
第六方面，本申请实施例提供一种数据处理装置，包括处理器和接口电路；所述接口电路，用于接收代码指令并传输至所述处理器；所述处理器用于运行所述代码指令以执行上述第一方面或第一方面的任一种可能的设计中的方法。
第七方面，本申请实施例提供一种可读存储介质，其上存储有计算机程序或指令，当该计算机程序或指令被执行时，使得计算机执行上述第一方面或第一方面的任一种可能的设计中的方法。
第八方面，本申请实施例提供一种计算机程序产品，当计算机读取并执行所述计算机程序产品时，使得计算机执行上述第一方面或第一方面的任一种可能的设计中的方法。
上述第二方面至第八方面中任一方面及其任一方面中任意一种可能的设计可以达到的技术效果，请参照上述第一方面及其第一方面中相应设计可以带来的技术效果描述，这里不再重复赘述。
附图说明
图1为一种全连接神经网络的架构示意图；
图2为一种损失函数优化示意图；
图3为一种梯度反向传播示意图；
图4为一种基于全连接神经网络的训练示意图；
图5为本申请实施例提供的一种数据处理流程示意图；
图6为本申请实施例提供的一种第一机器学习模型的示意图；
图7为本申请实施例提供的一种第一机器学习模型的示意图；
图8为本申请实施例提供的一种第一机器学习模型的示意图；
图9为本申请实施例提供的一种第二机器学习模型的示意图；
图10为本申请实施例提供的一种第二机器学习模型的示意图；
图11为本申请实施例提供的一种第二机器学习模型的示意图；
图12为本申请实施例提供的一种第二机器学习模型的示意图；
图13为本申请实施例提供的一种通信系统的架构示意图；
图14为本申请实施例提供的所需训练数据的数量的仿真示意图；
图15为本申请实施例提供的一种数据处理装置的结构示意图；
图16为本申请实施例提供的一种数据处理装置的结构示意图。
具体实施方式
本申请提供一种数据处理方法及装置，旨在减少机器学习过程中的训练样本的数量以及降低训练复杂度。其中，方法和装置是基于同一技术构思的，由于方法及装置解决问题的原理相似，因此装置与方法的实施可以相互参见，重复之处不再赘述。
以下对本申请实施例的部分用语进行解释说明，以便于本领域技术人员理解。
1)优化问题，不同的应用场景有对应的优化问题。可选的，应用场景可以包括数据分类、数据回归、数据聚类、垃圾邮件过滤、网页检索、自然语言处理、图像识别、语音识别以及无线通信等。在不同的应用场景中，根据不同需求可以设置不同的优化问题，在此不做限制。例如应用场景包括无线通信，优化问题可以包括功率分配、网络优化、移动性管理、信道编译码、或信道预测等。
一般设置有优化问题的优化目标，优化目标也可以根据不同需求进行设置。对于相同的优化问题，可以设置不同的优化目标，例如功率分配的优化目标可以包括吞吐最大化和/或时延最小化等。
可选的，还可以设置优化问题的约束条件，约束条件用于约束/限制优化问题的解，即优化问题的解满足约束条件。
一个优化问题的解可能包括一个或多个(近似)最优解，或者可能包括一个或多个次优解。
2)机器学习(machine learning，ML)，指不依赖明确的指令代码，而是基于推理所使用的算法/模型求解优化问题的技术。其中常用的机器学习包括(有)监督学习和无监督学习(或者非监督学习)。
监督学习中提供有训练集，训练集中的训练数据标记有正确的标签(一般为优化问题的最优解)。通过训练集可以学习到训练数据与标签的映射关系，且该映射关系适用于训练集之外的数据。
无监督学习中提供有训练数据，通过训练数据与优化目标，可以学习到训练数据与优化问题的解的映射关系。该映射关系可以用于求解优化问题，得到优化问题的解。
训练数据也可以称为训练样本，或样本数据。
3)机器学习模型，可以利用输入数据输出推理结果，该推理结果为推理到的优化问题的解，实现优化问题的求解。可选的，机器学习模型在求解优化问题时满足约束条件。机器学习模型可以包括数学模型和/或神经网络模型等。
其中数学模型可以为算法，例如用于实现分类的算法(如K-临近算法)，用实现回归的算法(如线性回归算法)，以及用于实现聚类的算法(如K-均值算法)。
神经网络(neural networks，NN)模型通常为基于神经网络架构的模型，神经网络包括深度神经网络(deep neural network，DNN)、全连接神经网络(fully connected neural network，FCNN)、卷积神经网络(convolutional neural networks，CNN)、循环神经网络(recurrent neural networks，RNN)、生成对抗网络(generative adversarial networks，GAN)等。神经网络模型在解决规模较大的优化问题时，可以达到更优的效果和性能。
以全连接神经网络为例进行说明，全连接神经网络也称为多层感知机(multilayer perceptron，MLP)，一个MLP包括输入层、输出层和隐含层。输入层、输出层和隐含层的数量均可以为一个或多个，每个(神经网络)层包括多个神经元。如图1所示，MLP包括一个输入层(包括神经元x1、x2、x3和x4)、多个隐含层(如图1所示中间部分)和一个输出层(包括神经元y1、y2、y3、y4、y5和y6)，相邻两层的神经元间两两相连。
其中，对于相邻两层的神经元，下一层的神经元的输出h满足：h＝f(wx+b)，其中f为激活函数，w为权重矩阵，b为偏置向量，x为与该下一层的神经元连接的上一层神经元。一般的，在确定下一层的神经元的输出h时，通过将与之连接的每个上一层神经元x进行加权，并采用激活函数对加权结果处理后得到。
神经网络的输出可以递归表达为y＝f n(w nf n-1(…)+b n)，n为当前计算的神经网络层的层数，y为当前计算的神经网络层的输出。
综上，可以将神经网络理解为从输入数据(如上述训练数据)到输出数据(如上述标签)的映射关系。基于(随机初始化的)w和b和训练数据得到这个映射关系的过程，就可以称为神经网络的训练(或学习)过程。
训练过程中对损失函数(或损失函数中的神经网络参数)优化的过程如图2所示，横轴为神经网络参数(如w和/或b)的取值，纵轴为损失函数的输出值，曲线为神经网络参数和损失函数的映射关系。一般的，设置有初始神经网络参数(如图2所示的起始点)，基于初始神经网络参数进行训练，在训练过程中，采用损失函数计算神经网络的输出结果和训练数据的标注结果的误差，即采用损失函数对神经网络的输出结果进行评价，并将误差反向传播，通过梯度下降的方法来迭代优化神经网络参数，直至损失函数的输出值达到最小值，此时达到最优点。
其中，梯度下降的过程可以表示为：
Figure PCTCN2022092404-appb-000001
θ为待优化的神经网络参数(如w和/或b)，L为损失函数，η为学习率，用于控制梯度下降的步长。
反向传播的过程可以利用求偏导的链式法则实现，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3所示，后一层参数为
Figure PCTCN2022092404-appb-000002
前一层参数为
Figure PCTCN2022092404-appb-000003
公式可以表示为：
Figure PCTCN2022092404-appb-000004
其中w ij为神经元j连接神经元i的权重，s i为神经元i上的输入加权和。
除特别说明外，本申请实施例所涉及的机器学习模型可以为经过机器学习得到的模型，或者机器学习中的模型。在机器学习模型为神经网络模型时，经过机器学习得到的模型可以指训练完成的模型，机器学习中的模型可以指待训练的模型。经过机器学习得到的模型 可以确定优化问题的解。
4)电子设备，也称设备或节点，包括终端设备和/或网络设备。
终端设备，也称用户设备(user equipment，UE)，是一种具有无线收发功能的设备，可以经无线接入网(radio access network，RAN)中的接入网设备(或者也可以称为接入设备)与一个或多个核心网(core network，CN)设备(或者也可以称为核心设备)进行通信。
用户设备也可称为接入终端、终端、用户单元、用户站、移动站、移动台、远方站、远程终端、移动设备、用户终端、用户代理或用户装置等。用户设备可以部署在陆地上，包括室内或室外、手持或车载；也可以部署在水面上(如轮船等)；还可以部署在空中(例如飞机、气球和卫星上等)。用户设备可以是蜂窝电话(cellular phone)、无绳电话、会话启动协议(session initiation protocol，SIP)电话、智能电话(smart phone)、手机(mobile phone)、无线本地环路(wireless local loop，WLL)站、个人数字处理(personal digital assistant，PDA)等。或者，用户设备还可以是具有无线通信功能的手持设备、计算设备或连接到无线调制解调器的其它设备、车载设备、可穿戴设备、无人机设备或物联网、车联网中的终端、第五代移动通信(5th-generation，5G)网络以及未来网络中的任意形态的终端、中继用户设备或者未来演进的PLMN中的终端等。其中，中继用户设备例如可以是5G家庭网关(residential gateway，RG)。例如用户设备可以是虚拟现实(virtual reality，VR)终端、增强现实(augmented reality，AR)终端、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)中的无线终端等。本申请实施例对终端设备的类型或种类等并不限定。
本申请实施例中，用于实现终端设备的功能的装置可以是终端设备，也可以是能够支持终端设备实现该功能的装置，例如芯片系统，该装置可以被安装在终端设备中。本申请实施例中，芯片系统可以由芯片构成，也可以包括芯片和其他分立器件。本申请实施例提供的技术方案中，以用于实现终端的功能的装置是终端设备为例，描述本申请实施例提供的技术方案。
网络设备，指可以为终端设备提供无线接入功能的设备。其中，网络设备可以支持至少一种无线通信技术，例如长期演进(long term evolution，LTE)、NR等。
例如网络设备可以包括接入网设备(也称接入节点或节点)。示例的，网络设备包括但不限于：5G网络中的下一代节点B(generation nodeB，gNB)、演进型节点B(evolved node B，eNB)、无线网络控制器(radio network controller，RNC)、节点B(node B，NB)、家庭基站(例如，home evolved node B、或home node B，HNB)、基带单元(baseband unit，BBU)、收发点(transmitting and receiving point，TRP)、发射点(transmitting point，TP)、移动交换中心、小站、微型站等。网络设备还可以是云无线接入网络(cloud radio access network，CRAN)场景下的无线控制器、集中单元(centralized unit，CU)、和/或分布单元(distributed unit，DU)，或者网络设备可以为中继站、接入点、车载设备、终端、可穿戴设备、在设备到设备(Device-to-Device，D2D)、车辆外联(vehicle-to-everything，V2X)、机器到机器(machine-to-machine，M2M)通信、物联网(Internet of Things)通信中承担基站功能的设备、或者未来移动通信中的网络设备或者未来演进的公共移动陆地网络(public land  mobile network，PLMN)中的网络设备等。
又如，网络设备可以包括核心网(CN)设备，核心网设备例如包括接入和移动管理功能(access and mobility management function，AMF)等。
本申请中的“和/或”，描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。字符“/”一般表示前后关联对象是一种“或”的关系。
本申请中所涉及的多个，是指两个或两个以上。
另外，需要理解的是，在本申请的描述中，“第一”、“第二”等词汇，仅用于区分描述的目的，而不能理解为指示或暗示相对重要性，也不能理解为指示或暗示顺序。
在本申请实施例中，“示例的”一词用于表示作例子、例证或说明。本申请中被描述为“示例”的任何实施例或设计方案不应被解释为比其他实施例或设计方案更优选或更具优势。确切而言，使用示例的一词旨在以具体方式呈现概念。
在求解优化问题时，如果优化问题的问题规模比较大，求解复杂度比较高，可以通过AI实现。AI中的神经网络具有离线训练复杂度高，在线计算复杂度低的特点，因此可以基于神经网络对优化问题进行离线训练和在线计算。
在离线训练时，将训练数据输入全连接神经网络(FCNN)，可以使用有监督学习或非监督学习的方法对FCNN的参数进行训练。如图4所示，以优化问题为功率分配为例，在使用监督学习时，训练集中包括环境数据和目标功率分配策略，FCNN可以推理出环境数据的功率分配策略，根据功率分配策略和目标功率分配策略的差别确定损失函数，对FCNN的参数进行训练；在使用无监督学习时，FCNN可以推理出环境数据的功率分配策略，根据功率分配策略对应的系统性能，与目标系统性能确定损失函数，对FCNN的参数进行训练。
但是这种方式中，对于不同的优化问题，都采用相同的FCNN结构进行求解。因此为了能够适应不同的优化问题的求解，往往使用大量训练数据，导致训练复杂度更高。并且当环境数据发生改变时，如果超出了FCNN的泛化能力，FCNN需要重新进行训练来适应新的环境数据，也会导致训练复杂度更高。
鉴于此，本申请实施例提供的一种数据处理方法。在该方法中，第二数据作为优化问题的环境数据，第一数据根据优化问题的优化目标和第二数据得到，即第一数据考虑到了环境数据对优化问题的优化目标的影响，而将第一数据输入到第一机器学习模型，相当于第一机器学习模型已知了环境数据对于优化问题的优化目标的影响，因此无需再使用大量训练数据进行训练，以学习到环境数据如何影响优化问题的优化目标，从而可以降低训练过程中训练数据的数量，降低训练复杂度。
图5为本申请实施例提供的一种可能的数据处理方法，该数据处理方法可以应用于数据处理装置，该数据处理装置可以为电子设备，或位于电子设备中。该数据处理方法包括以下步骤：
S501：获取第一数据。
S502：将第一数据输入第一机器学习模型中，得到第一推理结果。
其中第一数据根据优化问题的优化目标和第二数据确定。在S501中，可以根据优化 问题的优化目标和第二数据，确定第一数据，从而获取到第一数据。或者在S501，可以在存储介质中直接获取到第一数据，也就是说可以预先根据优化问题的优化目标和第二数据确定第一数据，然后将第一数据存储在存储介质中。
举例来说，第一数据可以表示为向量F＝[T(H)]，其中向量F为第一数据，H为第二数据，T(H)为根据优化问题的优化目标确定的函数。可选的，第一数据还可以根据第二推理结果和约束条件中的至少一个，以及优化问题的优化目标和第二数据确定，例如第一数据可以表示为F＝[T(H,p),C(H,p)]，其中p为第二推理结果，C(H,p)为优化问题的约束条件。其中第二推理结果为初始推理结果或前一级推理结果(如果有的话)。
例如优化问题为功率分配，优化目标为吞吐最大化和/或时延最小化，可选的可以通过求解功率分配策略实现该优化目标。第二数据可以包括信道数据和/或通信场景信息等。信道数据与收发机对(指一对发射机和接收机)之间的信道有关，其中收发机对之间的信道质量的好坏会影响功率分配。通信场景信息可以包括室内场景信息和/或室外场景信息等，或者还可以包括市区密集区域信息和/或郊区信息等。可能的，室内场景和室外场景中收发机对之间的遮挡情况不同，可能影响收发机对之间的信道质量，以及市区密集区域和郊区区域中包括的收发机对的数量不同，收发机对之间的信道受到的干扰情况不同，可能影响收发机对之间的信道质量，从而影响功率分配。并且可能的，约束条件可以与通信场景信息有关，例如室内场景中发射机的最大发射功率，与室外场景中发射机的最大发射功率可以不同。通信场景信息可以以图像的形式进行输入，例如输入的图像为包含地图信息的图像，该图像中标记有终端的位置和网络设备的位置，由此可以分析出通信场景为市区密集环境或者郊区环境；又如输入的图像为室内环境的图像，该图像中包含家具和家用电器的图像，由此可以分析出通信场景为室内环境。
又如优化问题为图像识别，优化目标为违章车辆识别，第二数据可以包括摄像头采集到的道路图像。例如道路图像中可以包含车道，交通灯和车辆的图像，由此可以识别出车辆是否有违章变道或闯红灯等。
可选的，可以通过第一函数可以确定第一数据F。第一函数中可以预先设计有T(H)，即第一函数中已知环境数据(包括第二数据)对优化问题的优化目标的影响，这样无需训练第一机器学习模型学习环境数据如何影响优化问题的优化目标，可以减少所需的训练数据的数量。第一函数中还可以预先设计有C(H,p)，这样可以不必须训练优化问题的约束条件，可以减少所需的训练数据的数量。也就是说T(H)和C(H,p)可以根据相关人员的经验预先建模。当然，T(H)和C(H,p)也可以通过训练得到，这里不做限制。
第一机器学习模型可以用于求解优化问题。其中优化问题和优化目标可以根据应用场景和实际需求确定，这里不做限制。例如优化问题为功率分配，可以针对K个对象确定功率分配策略，K为正整数。
第一机器学习模型可以为待训练的机器学习模型，对应的，第一数据可以为参与训练的数据，第二数据可以为参与训练的环境数据。或者第一机器学习模型可以为训练完成的机器学习模型，对应的，第一数据可以为待推理的数据，第二数据可以为待推理的环境数据。
若第一机器学习模型为待训练的机器学习模型，在该第一机器学习模型的训练过程中，第一数据考虑到了环境数据对优化问题的优化目标的影响，相当于第一机器学习模型已知环境数据对优化问题的优化目标的影响，因此无需再使用大量训练数据进行训练，以学习 到环境数据如何影响优化问题的优化目标，因此可以减少训练过程中训练数据的数量，降低训练复杂度。
第一机器学习模型可以包括数学模型，或者第一机器学习模型可以包括神经网络模型，或者第一机器学习模型可以包括数学模型和神经网络模型。其中神经网络模型中包括一个或多个神经网络层。
可选的，还可以对第二数据进行处理，进一步减少训练数据的数量以及降低训练的复杂度。例如可以对第二数据进行降维处理和排序处理。其中降维处理可以通过降维模型实现，降维模型可以为神经网络模型(如降维子网络)，或者可以其它降维方式(如主成分分析(principal component analysis，PCA)模型，或自编码器(autoEncoder))。排序处理可以利用置换不变性的特性，按照设定顺序(如从小到大或从大到小)将第二数据进行排序。
可选的，还可以将第二推理结果和第二数据中的至少一个，以及第一数据输入到第一机器学习模型中，可以提高第一机器学习模型推理的准确性。
一种可能的方式中，数据处理装置包括T级第一机器学习模型，即包括T个级联的第一机器学习模型，或者说T个第一机器学习模型进行迭代(共进行T次迭代)，T为正整数。可选的，T级第一机器学习模型之间可以直接级联，每级第一机器学习模型可以获取到第一数据。或者在第一机器学习模型之前级联有第一函数，即第一函数可以级联在两级第一机器学习模型之间。可选的，第一机器学习模型和/或第一函数之前级联有降维模型。
第一机器学习模型为T个级联的第一机器学习模型中的第t个，t小于或等于T。
如图6所示，第一机器学习模型的输入包括第一数据。如图7所示，第一机器学习模型的输入包括第一数据和第二数据。如图8所示，第一机器学习模型的输入为第二推理结果，第二数据和第一数据。
其中将级联得到的该T级第一机器学习模型看作一个整体时，该T级第一机器学习模型的输入为第一数据，第二数据(可选)和初始推理结果(可选的)，该T级第一机器学习模型的输出为最终的推理结果。
如果第一机器学习模型为第1个第一机器学习模型(此时t＝1)，第1个第一机器学习模型输入的第一数据，可以根据优化问题的优化目标，第二数据，约束条件(可选)和初始推理结果(可选)确定。在第1个第一机器学习模型的输入包括第二推理结果时，该第二推理结果为初始推理结果，初始推理结果可以根据相关人员的经验确定，或者可以为随机生成得到。第1个第一机器学习模型输出第一推理结果，可以用于确定第(t+1)个第一机器学习模型所输入的第一数据，以及可选的作为第(t+1)个第一机器学习模型的一个输入。
如果第一机器学习模型为第t个第一机器学习模型(此时t大于1，且t小于T)，第t个第一机器学习模型输入的第一数据，可以根据优化问题的优化目标，第二数据，约束条件(可选)和第(t-1)级推理结果(可选)确定。第(t-1)级推理结果为第(t-1)个第一机器学习模型输出的第一推理结果，该第(t-1)个第一机器学习模型为与该第t个第一机器学习模型级联的前一级第一机器学习模型。第t个第一机器学习模型输出的第一推理结果，可以用于确定第(t+1)个第一机器学习模型所输入的第一数据，以及可选的作为第(t+1)个第一机器学习模型的一个输入。
如果第一机器学习模型为第T个第一机器学习模型(此时t＝T)，即最后一级第一机器 学习模型，第T个第一机器学习模型输入的第一数据，可以根据优化问题的优化目标，第二数据，约束条件(可选)和第(T-1)级推理结果(可选)确定。第T个第一机器学习模型输出的第一推理结果可以作为最终的推理结果。在训练过程中，可以根据该最终的推理结果，对T个第一机器学习模型的参数进行调整。在使用监督学习方式时，可以根据最终的推理结果和第一数据的期望结果(即标签)之间的误差，对T个第一机器学习的参数进行调整。在使用非监督学习方式时，可以根据最终的推理结果和优化目标，对T个第一机器学习的参数进行调整。
另一种可能的方式中，数据处理装置包括T级第二机器学习模型，即包括T个级联的第二机器学习模型，或者说T个第二机器学习模型进行迭代(共进行T次迭代)，T为正整数。第二机器学习模型包括级联的第一函数和第一机器学习模型。可选的，第二机器学习模型包括降维模型，第一机器学习模型和/或第一函数之前级联有降维模型。
第二机器学习模型为T个级联的第二机器学习模型中的第t个，t小于或等于T。如图9所示，第二机器学习模型的第一机器学习模型的输入为第一数据。如图10所示，第二机器学习模型的第一机器学习模型的输入为第一数据和第二数据。如图11所示，第二机器学习模型的第一机器学习模型的输入为第二推理结果，第二数据和第一数据。可选的，满足以下条件中的至少一个时，采用图9所示的结构可以保证优化问题的推理结果的最优：a，第一函数为关于第二数据和第一数据的双射函数，即第二数据与第一数据具有一一对应的关系；可选的第一函数的输入还包括第二推理结果，则第二数据，第二推理结果(可选的)与第一数据具有一一对应的关系；b，对于任意两组(第二数据，可选的第二推理结果)输入，第一函数得到的第一数据相同，则在第二机器学习模型收敛时，根据该任意两组(第二数据，可选的第二推理结果)输入确定的第一推理结果也相同。
如图12所示，为一种可能的第二机器学习模型，第二机器学习模型还包括降维子网络，降维子网络可以对将对第二数据进行降维处理之后的数据输入第一机器学习模型。可选的，对第二数据进行降维处理之后的数据也可以输入第一函数中。
其中将级联得到的该T级第二机器学习模型看作一个整体时，该T级第二机器学习模型的输入为第一数据，第二数据(可选)和初始推理结果(可选的)，该T级第二机器学习模型的输出为最终的推理结果。
如果第二机器学习模型为第1个第二机器学习模型中的(此时t＝1)，第1个第二机器学习模型输入第二数据，第一函数可以根据优化问题的优化目标，第二数据，约束条件(可选)和初始推理结果(可选)确定第一数据。在第1个第二机器学习模型的输入包括第二推理结果时，该第二推理结果为初始推理结果。第1个第二机器学习模型输出第一推理结果，可以用于确定第(t+1)个第二机器学习模型所输入的第一数据，以及可选的作为第(t+1)个第二机器学习模型的一个输入。
如果第二机器学习模型为第t个第二机器学习模型(此时t大于1，且t小于T)，第t个第一机器学习模型输入第二数据，第一函数可以根据优化问题的优化目标，第二数据，约束条件(可选)和初始推理结果(可选)确定第一数据。在第t个第二机器学习模型的输入包括第二推理结果时，该第二推理结果为第(t-1)个第二机器学习模型输出的第一推理结果。第t个第二机器学习模型输出的第一推理结果，可以用于确定第(t+1)个第二机器学习模型所输入的第一数据，以及可选的作为第(t+1)个第二机器学习模型的一个输入。
如果第二机器学习模型为第T个第二机器学习模型(此时t＝T)，即最后一级第二机器 学习模型，第T个第二机器学习模型输入第二数据，第一函数可以根据优化问题的优化目标，第二数据，约束条件(可选)和初始推理结果(可选)确定第一数据。第T个第二机器学习模型输出的第一推理结果可以作为最终的推理结果。在训练过程中，可以根据该最终的推理结果，对T个第二机器学习模型的参数进行调整。在使用监督学习方式时，可以根据最终的推理结果和第二数据的期望结果(即标签)之间的误差，对T个第二机器学习模型的全部或部分参数进行调整。在使用非监督学习方式时，可以根据最终的推理结果和优化目标，对T个第二机器学习模型的全部或部分参数进行调整。如果第二机器学习模型还包括降维子网络，可以对第一机器学习模型的参数和降维子网络的参数同时训练，或者降维子网络的参数固定，而对第一机器学习模型的参数进行训练。可选的，降维子网络的参数可以为降维子网络训练完成后的参数。
又一种可能的方式中，本申请实施例包括T1级第一机器学习模型和T2级第二机器学习模型，T1为正整数，T2为正整数，T1和T2可以相同或不同。其中第一机器学习模型和第二机器学习模型可以参见上述方式，此处不做赘述。
本申请实施例提供的数据处理方法中，第二数据作为优化问题的环境数据，第一数据根据优化问题的优化目标和第二数据得到，即第一数据考虑到了环境数据对优化问题的优化目标的影响，而将第一数据输入到第一机器学习模型，相当于第一机器学习模型已知了环境数据对于优化问题的优化目标的影响，因此无需再使用大量训练数据进行训练，以学习到环境数据如何影响优化问题的优化目标，从而可以减少训练过程中训练数据的数量，降低训练复杂度。并且本申请实施例不依赖于迭代算法，可以求解存在迭代算法的优化问题，也可以求解不存在迭代算法的优化问题，因此可以适用更多优化问题的求解。
可以理解，本申请实施例描述的网络架构以及业务场景是为了更加清楚的说明本申请实施例的技术方案，并不构成对于本申请实施例提供的技术方案的限定，本领域普通技术人员可知，随着网络架构的演变和新业务场景的出现，本申请实施例提供的技术方案对于类似的技术问题，同样适用。
一种可能的方式中，本申请实施例提供的数据处理方法适用于无线通信场景。无线通信场景可以应用于各种通信系统，包括地面通信系统，以及地面通信系统与卫星通信系统相融合的系统，或者短距通信系统。其中移动通信系统可以为4G通信系统(例如，长期演进(long term evolution，LTE)系统)，5G通信系统(例如，新无线(new radio，NR)系统)，及未来的移动通信系统等。
图13为本申请实施例提供的一种可能通信系统的架构示意图，该通信系统包括一个网络设备1310和两个终端设备1320，其中网络设备可以向两个终端设备提供通信服务，与两个终端设备分别通信，两个终端设备之间可以互相通信。可以理解，通信系统中网络设备的数量可以为一个或多个，终端设备的数量可以为一个或多个，并且对网络设备和终端设备的其它可能的形式类型不做限定。
在无线通信场景下，考虑包括K个对象(如K个用户或网络设备或收发机对等)的优化问题可以满足以下公式：
Figure PCTCN2022092404-appb-000005
其中H为环境数据，p为推理结果，K为正整数，k小于或等于K，T k(H,p)为与第k个对象有关的函数，
Figure PCTCN2022092404-appb-000006
为K个对象有关的函数的聚合函数。该聚合函数可以通过求和算法得到，或通过求最小值算法得到，当 然也可以通过其它算法得到，这里不做限制。考虑包括K个对象的约束条件可以满足以下公式：
Figure PCTCN2022092404-appb-000007
C k(H,p)是第k个对象的约束条件，
Figure PCTCN2022092404-appb-000008
是第k个对象的约束条件的可行域。
下面以优化问题为功率分配，优化目标为吞吐最大化进行说明。
该优化问题的环境数据H包括信道数据，信道数据可以是由收发机对之间的信道构成的向量矩阵，其中电子设备可以获取到的电子设备本地和邻居设备之间的信道。该电子设备可以为终端设备或网络设备等。
T k(H,p)可以表示为：
Figure PCTCN2022092404-appb-000009
上述优化问题的公式
Figure PCTCN2022092404-appb-000010
可以表示为：
Figure PCTCN2022092404-appb-000011
其中h kj为第j个发射机与第k个接收机之间的信道，j＝1,2,…,K，k＝1,2,…,K，
Figure PCTCN2022092404-appb-000012
是噪声方差，p k是第k个发射机的发射功率，p为功率控制的功率分配策略，p＝[p 1,p 2,…,p k]是由K个发射机的发射功率构成的向量，环境数据H是所有发射机与接收机之间的信道矩阵，可以表示为
Figure PCTCN2022092404-appb-000013
其中
Figure PCTCN2022092404-appb-000014
可以根据发射机发射的导频信号的功率与接收机接收到的导频信号的功率之间的差值估计得到。C k(H,p)可以为第k发射机允许的最大发射功率P max。
例如K大于设定数值时，认为功率分配问题的规模较大，可以通过训练神经网络模型来学习最优功率分配策略和H之间的映射关系，例如最优功率分配策略和H之间的映射关系f可以采用函数p *＝f(H)表示，
Figure PCTCN2022092404-appb-000015
p *为最优解，即最优功率分配策略，则可以训练神经网络模型拟合映射关系f。例如设定数值可以为20，30，或者更少或更多，在此对设定数值不做限制。
如图11所示，神经网络模型包括第一函数和第一机器学习模型。该神经网络模型可以表示p和H之间的映射关系，并通过训练，学习到最优功率分配策略p *和H之间的映射关系。该神经网络模型的输入和输出的映射关系可以表示为：为
Figure PCTCN2022092404-appb-000016
θ为神经网络模型待训练的参数，如上述权重矩阵和/或偏置向量。
Figure PCTCN2022092404-appb-000017
为神经网络拟合的函数族，与神经网络的结构有关。神经网络模型中的约束条件可以包括根据香农公式得到的用户数据率上限，该神经网络模型的结构可以称为基于数据率的神经网络(data rate-based knowledge neural network，DRNN)。
第一函数可以根据优化问题的优化目标，环境数据H(即上述第二数据)，向量p(即上述第二推理结果)和C k(H,p)(即上述约束条件)，输出向量F(即上述第一数据)。向量F满足以下公式：F(H,p)＝[T 1(H,p),T 2(H,p),…,T K(H,p),C 1(H,p),C 2(H,p),…,C K(H,p)]。优化问题为功率分配，优化目标为吞吐最大化时，
Figure PCTCN2022092404-appb-000018
C k(H,p)＝p k，p k为第k个发射机的发射功率，
Figure PCTCN2022092404-appb-000019
第一机器学习模型用于确定功率分配策略，因此第一机器学习模型也可以称为策略分配模型，该策略分配模型可以为基于神经网络实现。如果策略分配模型的输入包括向量F，该策略分配模型的输入和输出的映射关系可以表示为：
Figure PCTCN2022092404-appb-000020
如果策略分配模型的输入包括向量F和环境数据H，该策略分配模型的输入和输出的映射关系可以表示为：
Figure PCTCN2022092404-appb-000021
如果策略分配模型的输入包括向量F，向量p和环境数据H，该策略分配模型的输入和输出的映射关系可以表示为：
Figure PCTCN2022092404-appb-000022
由于功率分配策略p在求解完成之前是未知的，因此可以将神经网络模型设计为迭代结构，即神经网络模型包括多级第二机器学习模型，每级第二机器学习模型包括第一函数和第一机器学习模型。第t级第二机器学习模型的输出功率分配策略p的迭代更新值p (t)(如上述第二推理结果)，即每级第二机器学习模型进行一次功率分配策略的更新。第t级第二机器学习模型输出的p (t)为第(t+1)级第二机器学习模型的输入，第(t+1)级第二机器学习模型的输出与第t级第二机器学习模型的输出的关系可以表示为：
Figure PCTCN2022092404-appb-000023
或
Figure PCTCN2022092404-appb-000024
或
Figure PCTCN2022092404-appb-000025
第(t+1)级第二机器学习模型与第t级第二机器学习模型之间的迭代函数可以表示为:
Figure PCTCN2022092404-appb-000026
或
Figure PCTCN2022092404-appb-000027
或
Figure PCTCN2022092404-appb-000028
Figure PCTCN2022092404-appb-000029
如果H经过降维子网络进行处理，则该迭代函数可以表示为
Figure PCTCN2022092404-appb-000030
或
Figure PCTCN2022092404-appb-000031
或
Figure PCTCN2022092404-appb-000032
为降维子网络的输出，
Figure PCTCN2022092404-appb-000033
为降维子网络的参数。其中H还可以经过排序处理，排序处理过程中，可以按照信号信道(矩阵H的对角元)的绝对值从大到小排序，矩阵H中元素的顺序发生变化，对K个对象进行功率控制时的顺序也相应变化，但不影响功率控制的值。第1级第二机器学习模型的输入可以包括初始的功率分配策略p (0)。功率分配策略经过T次迭代更新后，第T级第二机器学习模型输出p (T)。
该神经网络模型的训练可以采用监督学习或无监督学习的方式。以监督学习为例，训练集包括N tr个训练数据(即参与训练的环境数据)H [i]，每个输入的训练数据H [i]对应输出
Figure PCTCN2022092404-appb-000034
[i]表示对应第i个训练数据。神经网络的训练目标可以为最小化损失函数，损失函数衡量神经网络模型针对第i个训练数据实际的输出和期望输出
Figure PCTCN2022092404-appb-000035
之间的差距。训练神经网络模型的目的是通过调整神经网络的参数，使得损失函数最小化。使得损失函数最小化的神经网络模型参数可以表示为：
Figure PCTCN2022092404-appb-000036
其中
Figure PCTCN2022092404-appb-000037
为第i个训练数据实际的输出和期望输出
Figure PCTCN2022092404-appb-000038
的差距，即损失函数。例如第i个训练数据实际的输出和期望输出
Figure PCTCN2022092404-appb-000039
的差距可以通过平方误差表示，则损失函数可以 为均方误差(mean squared error，MSE)。
在神经网络模型训练完成后，神经网络模型可以基于获取到的环境数据H，推理出(近似)最优解，然后将(近似)最优解发送给与电子设备进行无线通信的其它电子设备。例如网络设备可以将推理到的最优功率分配策略发送给位于该网络设备通信范围的其它网络设备和终端设备。
下面结合仿真结果，对本申请实施例提供的数据处理方法所需的训练数据的数量进行说明。如图14所示，在收发机对的数量K＝10时，所需的训练数据的数量少于100个，DRNN收敛吞吐量达到最优性能的97.5％时，在收发机对的数量K＝20时，所需的训练数量的数据少于2000个，DRNN收敛吞吐量达到最优性能的97.5％时，所需的训练数量的数据少于4000个，在收发机对的数量K＝30时，DRNN收敛吞吐量达到最优性能的97.5％时。因此本申请实施例提供的数据处理方法可以达到较好的性能，实现优化问题的求解。
另外，表1中还示出了本申请实施例提供的数据处理方法所需的训练数据的数量与FCNN所需的训练数据的数量。
在收发机对的数量K＝10时，吞吐量达到最优性能的95％时，FCNN需要100000个训练数据，本申请实施例提供的DRNN需要20000个，经过信道数据排序处理的FCNN需要5000，而本申请实施例经过信道数据排序处理的DRNN仅需要20个。
在收发机对的数量K＝30时，吞吐量达到最优性能的85％时，FCNN需要1000000个训练数据，本申请实施例提供的DRNN需要10000个，经过信道数据排序的FCNN需要2000，而本申请实施例经过信道数据排序的DRNN仅需要10个。
表1
Figure PCTCN2022092404-appb-000040
可见，本申请实施例提供数据处理方法在训练时所需的训练数据数量远小于FCNN，可以降低训练的复杂度。
基于与上述数据处理方法的同一技术构思，本申请实施例还提供了一种数据处理装置，可用于实现上述方法实施例中描述的方法。可以理解的是，为了实现上述功能，装置可以包括执行各个功能相应的硬件结构和/或软件模块。本领域技术人员应该很容易意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，本申请能够以硬件或硬件和计算机软件的结合形式来实现。某个功能究竟以硬件还是计算机软件驱动硬件的方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
数据处理装置的一种可能的表现形式如图15所示，该数据处理装置1500可以以软件的形式存在。数据处理装置1500可以包括获取单元1501和处理单元1502。
该数据处理装置1500可以为电子设备，或者可以为设置在电子设备中的芯片。在一个实施例中，获取单元1501，用于获取第一数据，所述第一数据根据优化问题的优化目标和第二数据确定；
处理单元1502，用于将第一数据输入第一机器学习模型中，输出第一推理结果。
在一个实现方式中，获取单元1501，具体用于将所述第二数据输入第二机器学习模型的第一函数，确定第一数据，所述第一函数根据所述优化问题的优化目标确定，所述第二机器学习模型还包括所述第一机器学习模型。
在一个实现方式中，所述第二机器学习模型为T个级联的第二机器学习模型中的第t个，T为正整数，t小于或等于T。
在一个实现方式中，所述第一数据根据所述优化问题的优化目标，所述第二数据和第二推理结果确定，所述第二推理结果为初始推理结果或第t-1个第二机器学习模型输出的推理结果。
在一个实现方式中，处理单元1502，具体用于将所述第二推理结果和所述第二数据中的至少一个，以及所述第一数据，输入第一机器学习模型中。
在一个实现方式中，所述第一数据根据所述优化问题的优化目标，所述第二数据，和所述优化问题的约束条件确定。
在一个实现方式中，数据处理装置1500还可以包括：调整单元1503，用于在所述第二机器学习模型为T个级联的第二机器学习模型中的第T个时，根据所述第一推理结果，对所述T个第二机器学习模型的参数进行调整。
在一个实现方式中，所述第二机器学习模型还包括降维模型，所述降维模型用于对所述第二数据进行降维处理。
在一个实现方式中，数据处理装置1500还可以包括：排序单元1504，用于对所述第二数据进行排序处理。
在一个实现方式中，所述第二数据包括信道数据和/或通信场景信息。
本申请实施例中对单元的划分是示意性的，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式。在本申请的实施例中的各功能单元可以集成在一个处理模块中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
所述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)或处理器(processor)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质包括各种可以存储程序代码的介质。
数据处理装置的另一种可能的表现形式如图16所示，该数据处理装置1600包括至少一个处理器1602，可选的，还可以包括至少一个通信接口1604。进一步地，该数据处理装置1600中还可以包括存储器1606，所述存储器1606用于存储计算机程序或指令。所述存储器1606既可以是处理器内的存储器，也可以是处理器之外的存储器。在图15中所描述的各单元模块为通过软件实现的情况下，处理器1602执行相应动作所需的软件或程序代码存储在存储器1606中。所述处理器1602用于执行存储器1606中的程序或指令，以实现上述实施例中图5所示的步骤。通信接口1604用于实现该装置与其他装置之间的通信。
在存储器1606置于处理器之外的情况下，所述存储器1606、处理器1602和通信接口1604通过总线1608相互连接，所述总线1608可以是外设部件互连标准(peripheral  component interconnect，PCI)总线或扩展工业标准结构(extended industry standard architecture，EISA)总线等。应理解，总线可以分为地址总线、数据总线、控制总线等。为便于表示，图16中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
需要说明的是，该数据处理装置1600中的各个模块的操作和/或功能分别为了实现上述方法实施例，为了简洁，在此不再赘述。
本申请实施例还提供一种芯片系统，包括：处理器，所述处理器与存储器耦合，所述存储器用于存储程序或指令，当所述程序或指令被所述处理器执行时，使得该芯片系统实现上述方法实施例中的方法。
可选地，该芯片系统中的处理器可以为一个或多个。该处理器可以通过硬件实现也可以通过软件实现。当通过硬件实现时，该处理器可以是逻辑电路、集成电路等。当通过软件实现时，该处理器可以是一个通用处理器，通过读取存储器中存储的软件代码来实现。
可选地，该芯片系统中的存储器也可以为一个或多个。该存储器可以与处理器集成在一起，也可以和处理器分离设置，本申请并不限定。示例性的，存储器可以是非瞬时性处理器，例如只读存储器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，本申请对存储器的类型，以及存储器与处理器的设置方式不作具体限定。
示例性的，该芯片系统可以是现场可编程门阵列(field programmable gate array，FPGA)，可以是专用集成芯片(application specific integrated circuit，ASIC)，还可以是系统芯片(system on chip，SoC)，还可以是中央处理器(central processor unit，CPU)，还可以是网络处理器(network processor，NP)，还可以是数字信号处理电路(digital signal processor，DSP)，还可以是微控制器(micro controller unit，MCU)，还可以是可编程控制器(programmable logic device，PLD)或其他集成芯片。
应理解，上述方法实施例中的各步骤可以通过处理器中的硬件的集成逻辑电路或者软件形式的指令完成。结合本申请实施例所公开的方法步骤可以直接体现为硬件处理器执行完成，或者用处理器中的硬件及软件模块组合执行完成。
本申请实施例还提供一种计算机可读存储介质，所述计算机存储介质中存储有计算机可读指令，当计算机读取并执行所述计算机可读指令时，使得计算机执行上述方法实施例中的方法。
本申请实施例还提供一种计算机程序产品，当计算机读取并执行所述计算机程序产品时，使得计算机执行上述方法实施例中的方法。
应理解，本申请实施例中提及的处理器可以是CPU，还可以是其他通用处理器、DSP、ASIC、FPGA或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
还应理解，本申请实施例中提及的存储器可以是易失性存储器或非易失性存储器，或可包括易失性和非易失性存储器两者。其中，非易失性存储器可以是只读存储器(read-only memory，ROM)、可编程只读存储器(programmable ROM，PROM)、可擦除可编程只读存储器(erasable PROM，EPROM)、电可擦除可编程只读存储器(electrically EPROM，EEPROM)或闪存。易失性存储器可以是随机存取存储器(random access memory，RAM)，其用作外部高速缓存。通过示例性但不是限制性说明，许多形式的RAM可用，例如静态随机存取存储器(static RAM，SRAM)、动态随机存取存储器(dynamic RAM，DRAM)、同步动态随机存取存储器(synchronous DRAM，SDRAM)、双倍数据速率同步动态随机存 取存储器(double data rate SDRAM，DDR SDRAM)、增强型同步动态随机存取存储器(enhanced SDRAM，ESDRAM)、同步连接动态随机存取存储器(synchlink DRAM，SLDRAM)和直接内存总线随机存取存储器(direct rambus RAM，DR RAM)。
需要说明的是，当处理器为通用处理器、DSP、ASIC、FPGA或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件时，存储器(存储模块)集成在处理器中。
应注意，本文描述的存储器旨在包括但不限于这些和任意其它适合类型的存储器。
应理解，在本申请的各种实施例中，上述各过程的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的系统、装置和单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请所提供的几个实施例中，应该理解到，所揭露的系统、装置和方法，可以通过其它的方式实现。例如，以上所描述的装置实施例是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或单元的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。
所述功能如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、ROM、RAM、磁碟或者光盘等各种可以存储程序代码的介质。
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所述以权利要求的保护范围为准。

Claims (33)

  1. 一种数据处理方法，其特征在于，包括：
    获取第一数据，所述第一数据根据优化问题的优化目标和第二数据确定；
    将所述第一数据输入第一机器学习模型中，得到第一推理结果，所述第一机器学习模型用于求解所述优化问题。
  2. 如权利要求1所述的方法，其特征在于，所述方法还包括：
    将所述第二数据输入第二机器学习模型的第一函数，确定第一数据，所述第一函数根据所述优化问题的优化目标确定，所述第二机器学习模型还包括所述第一机器学习模型。
  3. 如权利要求2所述的方法，其特征在于，所述第二机器学习模型为T个级联的第二机器学习模型中的第t个，T为正整数，t小于或等于T。
  4. 如权利要求3所述的方法，其特征在于，所述第一数据根据优化问题的优化目标和第二数据确定，包括：
    所述第一数据根据所述优化问题的优化目标，所述第二数据和第二推理结果确定，所述第二推理结果为初始推理结果或第t-1个第二机器学习模型输出的推理结果。
  5. 如权利要求4所述的方法，其特征在于，所述将第一数据输入第一机器学习模型中，包括：
    将所述第二推理结果和所述第二数据中的至少一个，以及所述第一数据，输入第一机器学习模型中。
  6. 如权利要求1-5任一项所述的方法，其特征在于，所述第一数据根据优化问题的优化目标和第二数据确定，包括：
    所述第一数据根据所述优化问题的优化目标，所述第二数据，和所述优化问题的约束条件确定。
  7. 如权利要求3-6任一项所述的方法，其特征在于，所述方法还包括：
    在所述第二机器学习模型为T个级联的第二机器学习模型中的第T个时，根据所述第一推理结果，对所述T个第二机器学习模型的参数进行调整。
  8. 如权利要求2-7任一项所述的方法，其特征在于，所述第二机器学习模型还包括降维模型，所述降维模型用于对所述第二数据进行降维处理。
  9. 如权利要求1-8任一项所述的方法，其特征在于，所述将第一数据输入第一机器学习模型中之前，所述方法还包括：
    对所述第二数据进行排序处理。
  10. 如权利要求1-9任一项所述的方法，其特征在于，所述第二数据包括信道数据和/或通信场景信息。
  11. 一种数据处理装置，其特征在于，包括：
    获取单元，用于获取第一数据，所述第一数据根据优化问题的优化目标和第二数据确定；
    处理单元，用于将第一数据输入第一机器学习模型中，输出第一推理结果，所述第一机器学习模型用于求解所述优化问题。
  12. 如权利要求11所述的装置，其特征在于，所述获取单元，具体用于将所述第二数据输入第二机器学习模型的第一函数，确定第一数据，所述第一函数根据所述优化问题的 优化目标确定，所述第二机器学习模型还包括所述第一机器学习模型。
  13. 如权利要求12所述的装置，其特征在于，所述第二机器学习模型为T个级联的第二机器学习模型中的第t个，T为正整数，t小于或等于T。
  14. 如权利要求13所述的装置，其特征在于，所述第一数据根据所述优化问题的优化目标，所述第二数据和第二推理结果确定，所述第二推理结果为初始推理结果或第t-1个第二机器学习模型输出的推理结果。
  15. 如权利要求14所述的装置，其特征在于，所述处理单元，具体用于将所述第二推理结果和所述第二数据中的至少一个，以及所述第一数据，输入第一机器学习模型中。
  16. 如权利要求11-15任一项所述的装置，其特征在于，所述第一数据根据所述优化问题的优化目标，所述第二数据，和所述优化问题的约束条件确定。
  17. 如权利要求13-16任一项所述的装置，其特征在于，所述装置还包括：
    调整单元，用于在所述第二机器学习模型为T个级联的第二机器学习模型中的第T个时，根据所述第一推理结果，对所述T个第二机器学习模型的参数进行调整。
  18. 如权利要求12-17任一项所述的装置，其特征在于，所述第二机器学习模型还包括降维模型，所述降维模型用于对所述第二数据进行降维处理。
  19. 如权利要求11-18任一项所述的装置，其特征在于，所述装置还包括：
    排序单元，用于对所述第二数据进行排序处理。
  20. 如权利要求11-19任一项所述的装置，其特征在于，所述第二数据包括信道数据和/或通信场景信息。
  21. 一种数据处理装置，其特征在于，包括处理器；
    所述处理器，用于读取存储器中存储的计算机程序或指令，并执行所述计算机程序或指令，以使所述数据处理装置执行如权利要求1-10中任一项所述的方法。
  22. 如权利要求21所述的装置，其特征在于，还包括所述存储器。
  23. 一种数据处理装置，其特征在于，包括处理器和接口电路；
    所述接口电路，用于接收代码指令并传输至所述处理器；
    所述处理器用于运行所述代码指令以执行如权利要求1-10中任一项所述的方法。
  24. 一种可读存储介质，其特征在于，存储有计算机程序指令，当所述指令被执行时，使如权利要求1-10中任一项所述的方法被实现。
  25. 一种计算机程序产品，其特征在于，当其在计算机上运行时，使得权利要求1-10中任一项所述的方法被执行。
  26. 一种数据处理模型，其特征在于，所述数据处理模型包括T级第二机器学习模型，所述第二机器学习模型包括第一函数和第一机器学习模型，T为正整数；
    所述第一函数用于根据优化问题的优化目标和第二数据确定第一数据，将所述第一数据输入所述第一机器学习模型；
    所述第一机器学习模型用于根据第一数据，输出第一推理结果，第一推理结果为所述优化问题的解。
  27. 如权利要求26所述的数据处理模型，其特征在于，所述数据处理模型的输入为所述第二数据，输出为所述第一推理结果，所述第二数据作为T级中每级所述第二机器学习模型的输入。
  28. 如权利要求26或27所述的数据处理模型，其特征在于，第T级第二机器学习模型 输出的所述第一推理结果所述为数据处理模型的输出；
    非第T级第二机器学习模型输出的所述第一推理结果为下一级第二机器学习模型的输入。
  29. 如权利要求26-28任一项所述的数据处理模型，其特征在于，所述第一函数还用于根据所述优化问题的优化目标，所述第二数据和第二推理结果确定所述第一数据；或者
    所述第一函数还用于根据所述优化问题的优化目标，所述第二数据和约束条件确定所述第一数据；或者
    所述第一函数还用于根据所述优化问题的优化目标，所述第二数据，约束条件和第二推理结果确定所述第一数据；
    若所述第一函数属于第1级第二机器学习模型，所述第二推理结果为初始推理结果，若所述第一函数不属于第1级第二机器学习模型，所述第二推理结果为前一级第二机器学习模型输出的所述第一推理结果。
  30. 如权利要求29所述的数据处理模型，其特征在于，所述第一机器学习模型用于根据所述第一数据和所述第二数据，输出所述第一推理结果；或者
    所述第一机器学习模型用于根据所述第一数据和所述第二推理结果，输出所述第一推理结果；或者
    所述第一机器学习模型用于根据所述第一数据，所述第二推理结果和所述第二数据，输出所述第一推理结果。
  31. 如权利要求26-30任一项所述的数据处理模型，其特征在于，所述第二数据为经过降维处理和/或排序处理后得到的数据。
  32. 如权利要求26-31任一项所述的数据处理模型，其特征在于，所述第二机器学习模型还包括降维子网络，所述降维子网络的输入为所述第二数据，所述降维子网络的输出为对所述第二数据经过降维处理后得到的数据。
  33. 如权利要求32所述的数据处理模型，其特征在于，所述降维子网络的输出作为所述第一函数的输入，和/或作为所述第一机器学习模型的输入。
PCT/CN2022/092404 2021-05-14 2022-05-12 一种数据处理方法及装置 WO2022237865A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110530279.4A CN115345271A (zh) 2021-05-14 2021-05-14 一种数据处理方法及装置
CN202110530279.4 2021-05-14

Publications (1)

Publication Number Publication Date
WO2022237865A1 true WO2022237865A1 (zh) 2022-11-17

Family

ID=83977789

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/092404 WO2022237865A1 (zh) 2021-05-14 2022-05-12 一种数据处理方法及装置

Country Status (2)

Country Link
CN (1) CN115345271A (zh)
WO (1) WO2022237865A1 (zh)

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103761426A (zh) * 2014-01-02 2014-04-30 中国科学院数学与系统科学研究院 一种在高维数据中快速识别特征组合的方法及系统
CN108848068A (zh) * 2018-05-29 2018-11-20 上海海事大学 基于深度信念网络-支持向量数据描述的apt攻击检测方法
CN109118473A (zh) * 2018-07-03 2019-01-01 深圳大学 基于神经网络的角点检测方法、存储介质与图像处理系统
CN109558899A (zh) * 2018-11-13 2019-04-02 中国石油天然气股份有限公司 数据降维方法及装置
US20200250529A1 (en) * 2019-01-31 2020-08-06 Hitachi, Ltd. Arithmetic device
CN112465029A (zh) * 2020-11-27 2021-03-09 北京三快在线科技有限公司 一种实例追踪的方法及装置
US10963802B1 (en) * 2019-12-19 2021-03-30 Sas Institute Inc. Distributed decision variable tuning system for machine learning
CN113361047A (zh) * 2021-08-10 2021-09-07 浙江大学 面向多层感知结构的叶轮高维优化及选型方法

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103761426A (zh) * 2014-01-02 2014-04-30 中国科学院数学与系统科学研究院 一种在高维数据中快速识别特征组合的方法及系统
CN108848068A (zh) * 2018-05-29 2018-11-20 上海海事大学 基于深度信念网络-支持向量数据描述的apt攻击检测方法
CN109118473A (zh) * 2018-07-03 2019-01-01 深圳大学 基于神经网络的角点检测方法、存储介质与图像处理系统
CN109558899A (zh) * 2018-11-13 2019-04-02 中国石油天然气股份有限公司 数据降维方法及装置
US20200250529A1 (en) * 2019-01-31 2020-08-06 Hitachi, Ltd. Arithmetic device
US10963802B1 (en) * 2019-12-19 2021-03-30 Sas Institute Inc. Distributed decision variable tuning system for machine learning
CN112465029A (zh) * 2020-11-27 2021-03-09 北京三快在线科技有限公司 一种实例追踪的方法及装置
CN113361047A (zh) * 2021-08-10 2021-09-07 浙江大学 面向多层感知结构的叶轮高维优化及选型方法

Also Published As

Publication number Publication date
CN115345271A (zh) 2022-11-15

Similar Documents

Publication Publication Date Title
US20230019669A1 (en) Systems and methods for enhanced feedback for cascaded federated machine learning
US20230342593A1 (en) Neural network training method and related apparatus
WO2021259305A1 (zh) 一种多任务学习方法及装置
US11863495B2 (en) Signaling for a channel state information reference signal (CSI-RS)
US11616582B2 (en) Neural network-based spatial inter-cell interference learning
WO2023020502A1 (zh) 数据处理方法及装置
US11929853B2 (en) Data-driven probabilistic modeling of wireless channels using conditional variational auto-encoders
US11456834B2 (en) Adaptive demodulation reference signal (DMRS)
CN116419257A (zh) 一种通信方法及装置
WO2022237865A1 (zh) 一种数据处理方法及装置
WO2023036268A1 (zh) 一种通信方法及装置
TW202201935A (zh) 用於無線通道估計和追蹤的神經網路增強
TW202308343A (zh) 經由基於模型的神經網路的通道特徵提取
US11792877B2 (en) Indication triggering transmission of known data for training artificial neural networks
US20220188594A1 (en) Neural network computation for eigen value and eigen vector decomposition of matrices
WO2023236986A1 (zh) 一种通信方法及通信装置
WO2023226650A1 (zh) 一种模型训练的方法和装置
WO2023185890A1 (zh) 一种数据处理方法及相关装置
WO2024016974A1 (zh) 信息传输的方法与通信装置
WO2024067245A1 (zh) 模型匹配的方法和通信装置
US20230259742A1 (en) Communication method, apparatus, and system
WO2023230818A1 (zh) 一种波束管理方法及装置、用户设备、网络设备
WO2024000344A1 (zh) 一种模型训练方法及相关装置
WO2022257662A1 (zh) 一种应用人工智能的通信方法及通信装置
WO2023197950A1 (zh) 一种通信方法及相关装置

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: 22806827

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22806827

Country of ref document: EP

Kind code of ref document: A1