WO2020191686A1 - 一种基于神经网络的功率分配方法及装置 - Google Patents

一种基于神经网络的功率分配方法及装置 Download PDF

Info

Publication number
WO2020191686A1
WO2020191686A1 PCT/CN2019/079957 CN2019079957W WO2020191686A1 WO 2020191686 A1 WO2020191686 A1 WO 2020191686A1 CN 2019079957 W CN2019079957 W CN 2019079957W WO 2020191686 A1 WO2020191686 A1 WO 2020191686A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
network model
layer
output
value
Prior art date
Application number
PCT/CN2019/079957
Other languages
English (en)
French (fr)
Inventor
黄鸿基
胡慧
刘劲楠
杨帆
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Priority to PCT/CN2019/079957 priority Critical patent/WO2020191686A1/zh
Priority to CN201980094522.3A priority patent/CN113615277B/zh
Publication of WO2020191686A1 publication Critical patent/WO2020191686A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W72/00—Local resource management
    • H04W72/04—Wireless resource allocation
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Definitions

  • This application relates to the field of wireless communication technology, and in particular to a method and device for power distribution based on neural networks.
  • the multiple-input multiple-output (MIMO) communication system can greatly increase the capacity and communication rate of the next-generation communication system, and is considered to be the development direction of future wireless communication technology.
  • Power allocation is an important technology of the MIMO communication system.
  • the transmission power is allocated reasonably at the signal transmitting end, which can make the MIMO communication system obtain the maximum system capacity without increasing the additional transmission power and bandwidth consumption, and ensure the communication of the MIMO communication system effectiveness.
  • MIMO communication systems With the emergence of massive MIMO and non-orthogonal frequency division multiple access (multiple-input-multiple-output non-orthogonal multiple access, MIMO-NOMA) and other technologies, MIMO communication systems The complexity of the MIMO communication system continues to increase, and the existing power allocation method cannot fully learn the spatial characteristics and channel characteristics of the complex MIMO communication system, resulting in an unsatisfactory power allocation effect. Therefore, it is necessary to find a solution that can achieve better power distribution.
  • MIMO-NOMA multiple-input-multiple-output non-orthogonal multiple access
  • the embodiments of the present application provide a neural network-based power distribution method and device, which are used to solve the current technical problem that the information pushed to the user according to the key characters selected by the user is not accurate enough.
  • this application provides a neural network-based power distribution method.
  • the steps shown in the method can be executed by a computer device.
  • the computer device may input input parameters to the input layer of the neural network model, where the input parameters include the channel vector characteristics of each of the multiple antennas in the communication system for each of the multiple users.
  • the input parameters are extracted, the characteristics of the communication system are extracted, the characteristics are fitted through multiple iterations based on the neural network model, and the fitting results are transferred to the output layer of the neural network model, from the The output layer obtains the transmission power allocated to each user, where the transmission power is determined according to the fitting result.
  • the characteristics of the communication system are extracted according to the input parameters, and the transmission power allocated to each user is determined based on the characteristics.
  • a more optimized power allocation scheme can be obtained to realize the transmission power of the antenna in the MIMO communication system. optimization.
  • the activation function of the output layer of the neural network model may be a Maxout function.
  • a plurality of hidden layers of the neural network model may also be included between the input layer and the output layer. Calculating through these hidden layers can increase the amount of calculation of the neural network model and increase the accuracy of the result.
  • the activation function of the output layer of the neural network model can be expressed by the following formula:
  • f Maxout represents the operation result of the activation function
  • x represents the output value set of the hidden layer adjacent to the output layer
  • the x includes X output values
  • X is a positive integer
  • w i represents the i-th
  • Sigmoid() represents the Sigmoid function.
  • the activation function of the output layer of the neural network model can also be expressed by the following formula:
  • y represents the operation result of the activation function
  • P max is a preset value
  • x represents the output value set of the hidden layer adjacent to the output layer
  • the x includes X output values
  • X is a positive integer .
  • the activation functions of the input layer and the multiple hidden layers are both linear rectification functions.
  • the loss function of the neural network model includes a first penalty term and/or a second penalty term, and the loss function is used for offline training of the neural network model, wherein the first The penalty term is used to constrain the transmission power to be greater than the target power value, and the second penalty term is used to constrain the transmission rate allocated to each user to be not less than the minimum value of the target transmission rate.
  • the weights and deviations in the neural network model can better reflect the characteristics of the communication system and optimize the output results.
  • the value of the coefficient of the first penalty term may be [0, 1]
  • the value of the coefficient of the second penalty term may be [0, 1].
  • the loss function of the neural network model can be expressed by the following formula:
  • L represents the calculation result of the loss function
  • N represents the number of training samples of the neural network model
  • M represents the number of antennas
  • K represents the number of users
  • R sum represents the number of users allocated to each user.
  • R min represents the minimum value of the transmission rate
  • R m,k represents the transmission rate allocated by the m-th antenna to the k-th user
  • p m represents the transmission power of the m-th antenna
  • ⁇ represents the coefficient of the first penalty term
  • ⁇ represents the coefficient of the second penalty term
  • the R sum can be expressed by the following formula:
  • ⁇ 2 is the variance of the additive white Gaussian noise, It is the letter dry ratio. Since the power allocation factor ⁇ for different users is not set to the same value, the power allocation result can be flexibly adjusted according to the factor, such as increasing the transmission power for certain users, so as to improve the communication quality for the users.
  • the value range of ⁇ i,j,l may be [0,1].
  • embodiments of the present application provide a computer device, which may have the function of the method involved in the above-mentioned first aspect or any one of the possible designs of the first aspect, so as to be used to execute the above-mentioned first aspect or Any one of the possibilities of the first aspect relates to the method shown.
  • This function can be realized by hardware, or by hardware executing corresponding software.
  • the hardware or software includes one or more modules corresponding to the above-mentioned functions.
  • the computer device may include an input module, a feature extraction module, a fitting module, and an output module.
  • the input module may be used to input input parameters to the input layer of the neural network model, and the input parameters include the channel vector characteristics of each of the multiple antennas in the communication system for each of the multiple users.
  • the feature extraction module can be used to extract features of the communication system according to the input parameters.
  • the fitting module can be used to fit the features through multiple iterations based on the neural network model, and transfer the fitting results to the output layer of the neural network model.
  • the output module may be used to obtain the transmission power allocated to each user from the output layer, wherein the transmission power is determined according to the fitting result.
  • the activation function of the output layer of the neural network model may be a Maxout function.
  • a plurality of hidden layers of the neural network model may also be included between the input layer and the output layer.
  • the activation function of the output layer of the neural network model can be expressed by the following formula:
  • f Maxout represents the operation result of the activation function
  • x represents the output value set of the hidden layer adjacent to the output layer
  • the x includes X output values
  • X is a positive integer
  • w i represents the i-th
  • Sigmoid() represents the Sigmoid function.
  • the activation function of the output layer of the neural network model can also be expressed by the following formula:
  • y represents the operation result of the activation function
  • P max is a preset value
  • x represents the output value set of the hidden layer adjacent to the output layer
  • the x includes X output values
  • X is a positive integer .
  • the activation functions of the input layer and the multiple hidden layers are both linear rectification functions.
  • the loss function of the neural network model includes a first penalty term and/or a second penalty term, and the loss function is used for offline training of the neural network model, wherein the first The penalty term is used to constrain the transmission power to be greater than the target power value, and the second penalty term is used to constrain that the transmission rate allocated to each user is not less than the minimum value of the target transmission rate, so that the weights in the neural network model can be made after training
  • the sum deviation can better reflect the characteristics of the communication system and optimize the output results.
  • the value of the coefficient of the first penalty term may be [0, 1]
  • the value of the coefficient of the second penalty term may be [0, 1].
  • the loss function of the neural network model can be expressed by the following formula:
  • L represents the calculation result of the loss function
  • N represents the number of training samples of the neural network model
  • M represents the number of antennas
  • K represents the number of users
  • R sum represents the number of users allocated to each user.
  • R min represents the minimum value of the transmission rate
  • R m,k represents the transmission rate allocated by the m-th antenna to the k-th user
  • p m represents the transmission power of the m-th antenna
  • ⁇ represents the coefficient of the first penalty term
  • ⁇ represents the coefficient of the second penalty term
  • the R sum can be expressed by the following formula:
  • ⁇ 2 is the variance of the additive white Gaussian noise, It is the letter dry ratio.
  • the value range of ⁇ i,j,l may be [0,1].
  • an embodiment of the present application provides a computer-readable storage medium, including program instructions.
  • the program instructions When the program instructions are used on a computer, the computer realizes the first aspect or any possible design of the first aspect.
  • the embodiments of the present application provide a computer program product, which when running on a computer, enables the computer to implement the functions involved in the first aspect or any possible design of the first aspect.
  • an embodiment of the present application provides a system, which includes the second aspect or any possible design of the second aspect.
  • an embodiment of the present application provides a chip, which may be coupled with a memory, and used to read and execute programs or instructions stored in the memory to implement any of the first aspect of the first aspect or the second aspect.
  • the functions involved in the possible design are included in the possible design.
  • FIG. 1 is a schematic diagram of a neural network model applicable to an embodiment of this application
  • FIG. 2 is a schematic structural diagram of a computer device provided by an embodiment of this application.
  • FIG. 3 is a schematic flowchart of a management distribution method provided by an embodiment of the application.
  • FIG. 4 is a schematic structural diagram of a neural network model provided by an embodiment of the application.
  • FIG. 5 is a schematic structural diagram of another neural network model provided by an embodiment of the application.
  • FIG. 6 is a schematic structural diagram of another computer device provided by an embodiment of this application.
  • At least one refers to one or more, and “multiple” refers to two or more.
  • And/or describes the association relationship of the associated objects, indicating that there can be three relationships, for example, A and/or B, which can mean: A alone exists, A and B exist at the same time, and B exists alone, where A, B can be singular or plural.
  • the character “/” generally indicates that the associated objects are in an “or” relationship.
  • the following at least one (item) or similar expressions refers to any combination of these items, including any combination of a single item (a) or a plurality of items (a).
  • At least one of a, b, or c can mean: a, b, c, a and b, a and c, b and c, or a, b and c, where a, b, c It can be single or multiple.
  • [a, b] means that the value is in the range a (including a) to b (including b), and a is less than or equal to b.
  • Neural networks is a complex network system formed by a large number of simple nodes (or called processing units, neurons or neuron nodes, etc.) widely connected to each other. It reflects the functions of the human brain Many basic characteristics of The function and characteristics of the neuron can be imitated through the mathematical model, so that a neural network can be constructed based on the mathematical model of the neuron (also referred to as a neural network model, or a network model in this application).
  • a neural network model may include a multi-layer structure, such as an input layer, which is the first layer of the neural network, and an output layer, which is the last layer of the neural network. Between the input layer and the output layer, a hidden layer may also be included.
  • DNN deep neural networks
  • the nodes in the neural network are shown in Figure 1 (a).
  • the a1 ⁇ an on the left side of the circle in the figure are the input data of the node, w1 ⁇ wn are the weight of the node, and b is the bias of the node.
  • the circle in the figure represents the internal calculation of the node, that is, a specific output function f, also known as the activation function or excitation function.
  • Each node can have multiple output data z, but the value is the same.
  • the neural network can be composed of input layer, hidden layer, and output layer.
  • the input layer is used to input the input data of the neural network;
  • the output layer is used to output the output data of the neural network;
  • the hidden layer is composed of many node connections between the input layer and the output layer, and is used to process the input data.
  • the hidden layer can include one or more layers. The number of hidden layers and the number of nodes in the neural network are directly related to the complexity of the problems actually solved by the neural network, the number of nodes in the input layer and the number of nodes in the output layer.
  • Convolutional layer used in neural network models to extract local features of input data through convolution operations, for example, to obtain various local feature maps for input images. It should be understood that the convolutional layer involved in the embodiment of the present application is not limited to the convolutional layer in a convolutional neural network, and may also be a convolutional layer in other types of neural network models.
  • Each node of the fully connected layer (FC) is connected to each node of the upper layer.
  • the fully connected layer can be used to synthesize the local features extracted from the upper layer according to the weight. For example, Various local feature maps are combined into a complete image again.
  • Each output of the fully connected layer can be seen as each node of the previous layer multiplied by a weight coefficient, and finally a deviation is added.
  • the node in the neural network accepts the output value of the node of the previous layer as the input value of the node. For example, the input layer node will directly transfer the input attribute value to the node of the next layer (hidden layer or output layer).
  • the activation function also called the activation function.
  • the loss function of the neural network is used to describe the deviation between the predicted value of the neural network and the true value.
  • the loss function is a non-negative function. The smaller the loss function, the better the robustness of the neural network model.
  • the good performance of a neural network means that the most suitable weight has been found to minimize the loss function.
  • the process of finding the most suitable weight is the process of neural network learning.
  • Fig. 2 shows a structural diagram of a possible computer device to which the method provided in the embodiment of the present application is applicable.
  • the computer device includes: a processor 210, a memory 220, a communication module 230, an input unit 240, a display unit 250, a power supply 260 and other components.
  • a processor 210 the computer device includes: a processor 210, a memory 220, a communication module 230, an input unit 240, a display unit 250, a power supply 260 and other components.
  • the structure of the computer device shown in FIG. 2 does not constitute a limitation on the computer device.
  • the computer device provided in the embodiment of the present application may include more or less components than those shown in the figure, or Combining certain components, or different component arrangements.
  • the communication module 230 may be connected to other devices through a wireless connection or a physical connection to realize data transmission and reception of the computer device.
  • the communication module 230 may include any one or a combination of a radio frequency (RF) circuit, a wireless fidelity (wireless fidelity, WiFi) module, a communication interface, a Bluetooth module, etc. This embodiment of the application does not do this. limited.
  • RF radio frequency
  • the memory 220 can be used to store program instructions and data.
  • the processor 210 executes various functional applications and data processing of the computer device by running the program instructions stored in the memory 220.
  • the program instructions include program instructions that enable the processor 210 to execute the power allocation method provided in the following embodiments of the present application.
  • the memory 220 may mainly include a program storage area and a data storage area.
  • the storage program area can store operating systems, various application programs, and program instructions, etc.;
  • the storage data area can store various data such as neural networks.
  • the memory 220 may include a high-speed random access memory, and may also include a non-volatile memory, such as a magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.
  • the input unit 240 may be used to receive information such as data or operation instructions input by the user.
  • the input unit 240 may include input devices such as a touch panel, function keys, a physical keyboard, a mouse, a camera, and a monitor.
  • the display unit 250 can implement human-computer interaction, and is used to display information input by the user, information provided to the user, and other content through a user interface.
  • the display unit 250 may include a display panel 251.
  • the display panel 251 may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), etc.
  • the touch panel can cover the display panel 251, and when the touch panel detects a touch event on or near it, it is transmitted to the processor 210 To determine the type of touch event to perform the corresponding operation.
  • the processor 210 is the control center of the computer device, and uses various interfaces and lines to connect the above components.
  • the processor 210 may execute the program instructions stored in the memory 220 and call the data stored in the memory 220 to complete various functions of the computer device and realize the neural network accuracy provided by the embodiments of the present application. Adjustment method.
  • the processor 210 may include one or more processing units.
  • the processing unit may include a hardware device capable of floating-point operations, such as a CPU and/or GPU.
  • the processing unit can process and output the data input to the neural network.
  • the processing unit reads the data of the neural network from the storage data area of the memory 220, quantifies the data input to the neural network, and can also perform the neural network Processing such as forward propagation and back propagation.
  • the computer device also includes a power source 260 (such as a battery) for powering various components.
  • a power source 260 such as a battery
  • the power supply 260 may be logically connected to the processor 210 through a power management system, so that functions such as charging and discharging the computer device can be realized through the power management system.
  • the computer device may also include components such as a camera, a sensor, and an audio collector, which will not be repeated here.
  • the embodiment of the present application provides a neural network-based power distribution method, which can be applied to the computer device shown in FIG. 2 and/or the neural network shown in FIG. 1(b). This method can be executed by the processor in the computer device as shown in FIG. 2. Referring to Figure 3, the process of the method may include:
  • S101 Input an input parameter to the input layer of the neural network model, where the input parameter includes the channel vector characteristics of each antenna of the multiple antennas in the communication system for each of the multiple users;
  • S104 Obtain the transmission power allocated to each user from the output layer, where the transmission power is determined according to the fitting result.
  • the neural network model can be used to extract the characteristics of the communication system based on the channel vector characteristics of the MIMO communication system, and the neural network model can be further used to fit the characteristics through multiple iterations, and the allocation for each user can be determined according to the fitting results. Since the characteristics of the communication system are considered in the power allocation, the transmission power allocation result can be optimized.
  • the communication systems involved in S101 include but are not limited to traditional MIMO, massive MIMO, MIMO-NOMA communication systems or other MIMO systems.
  • the method shown in FIG. 3 can be used for the transmission power allocation of the above-mentioned communication system to optimize the power allocation scheme.
  • p m is defined as the transmit power of the mth antenna (m is a positive integer) in the communication system involved in S101 above
  • R sum represents the total transmission rate of the communication system (that is, the sum of the transmission rates allocated to each user in the communication system)
  • R m,k represents the transmission rate allocated by the mth antenna to the kth user
  • P represents the communication system
  • R min represents the minimum value of the transmission rate allocated by the mth antenna to the kth user.
  • ⁇ i,j,l can be preset values, and the value range is [0,1].
  • R m,k may be determined according to the signal interference noise ratio (SINR) of the mth antenna for the kth user. Specifically, R m,k can be determined by the following formula:
  • ⁇ m,k represents the signal-to-dry ratio of the m-th antenna to the k-th user.
  • ⁇ m,k can be determined according to the following formula:
  • ⁇ k represents the power allocation factor of the k-th user
  • ⁇ l represents the power allocation factor of the l- th user other than the k-th user
  • ⁇ 2 is the variance of additive white Gaussian noise.
  • Both ⁇ k and ⁇ l are preset values, and the value range of ⁇ k and ⁇ l is [0,1].
  • the user's power allocation factor can be determined according to channel state information (CSI) for the user.
  • CSI channel state information
  • the condition shown in item C1 is used to constrain the sum of the transmission power of all antennas in the communication system to be no greater than the total transmission power of the communication system.
  • the condition shown in item C2 is used to restrict the transmission rate allocated by any antenna to any user not less than the minimum value of the transmission power.
  • the condition shown in item C3 is used to constrain the transmit power of any antenna to be no less than 0.
  • the condition shown in item C4 is used to restrict the value range of each power allocation factor to [0,1].
  • the neural network model involved in the embodiments of the present application may include an input layer, an output layer, and multiple hidden layers located between the input layer and the output layer, where the multiple hidden layers may include at least one convolution Layer and/or a fully connected layer.
  • the output layer of the neural network model may be a convolutional layer.
  • the activation function of the output layer of the neural network model involved in the embodiment of the present application is a Maxout function.
  • the activation function of the output layer of the neural network model involved in the embodiment of the present application is a Maxout function.
  • f Maxout represents the operation result of the activation function
  • x represents the output value set of the hidden layer adjacent to the output layer
  • x can contain X output values
  • x can be a matrix composed of X output values
  • X can be the same as the number of nodes in the adjacent hidden layer
  • w i represents the weight of the i-th said output value
  • b i represents the deviation of the i-th said output value
  • i 1, 2...X
  • x T represents the transposition of x
  • Sigmoid() represents the Sigmoid function.
  • the activation function of the output layer is also expressed by the following formula:
  • P max can represent the maximum transmit power of the base station
  • P max can be a preset value
  • x represents the output value set of the hidden layer adjacent to the output layer
  • the x Contains X output values, X is a positive integer.
  • a linear rectification (rectified linear unit, ReLU) function can be used as its activation function.
  • the activation function of the input layer and any one of the multiple hidden layers can be expressed by the following formula:
  • x0 represents the set of output values of the previous layer adjacent to the layer
  • x0 can be a matrix composed of X0 output values
  • X0 is a positive integer
  • X0 can be the same as the number of nodes in the adjacent hidden layer .
  • a neural network model may include multiple convolutional layers and multiple fully connected layers, where the input layer is a fully connected layer, and the input layer and the output layer can be connected in sequence. It includes two convolutional layers and four fully connected layers. The input layer is adjacent to a convolutional layer, and the output layer is adjacent to a fully connected layer.
  • the number of nodes in the input layer of the above neural network model is the same as the number of input parameters, and is used to map the input parameters to the characteristics of the antenna.
  • the input layer may correspond to M (M is a positive integer) antennas respectively
  • M M is a positive integer
  • the input parameters are mapped to M features (that is, the number of nodes included in the input layer is M), and then the M features are passed to the next layer.
  • the number of antennas can be 64.
  • the first convolutional layer adjacent to the input layer can be used to map the M features passed by the input layer into 64 features.
  • the size of the convolution kernel can be 7*7.
  • the stride of the convolutional layer The parameter can be configured to 2.
  • the step parameter can be used to indicate the number of grids that the convolution kernel jumps each time when scanning in the input feature map composed of multiple input features.
  • the number of grids to be jumped can be controlled by the compensation parameter to reduce the convolution kernel. Repeated calculations during scanning to improve scanning efficiency.
  • the convolutional layer adjacent to the first convolution kernel (hereinafter referred to as the second convolutional layer) can be used to map the 64 features passed by the first convolutional layer into 32 features.
  • the size can be 3*3, and the step length parameter can be configured as 2.
  • the number of nodes in the four fully connected layers after the second convolutional layer can be configured to 220, 85, 80, and 64 in turn.
  • the four fully connected layers can map the input features of each layer to 220, 85, 80, and 64 features.
  • the output layer of the neural network model can contain K (K is a positive integer) nodes, its convolution kernel size can be 3*3, and its step parameter can be configured to 2. Therefore, the output layer can fully connect the previous one
  • K is the number of users in the communication system
  • the K output results are the transmission power allocated to each user. For example, the number of users can be 32.
  • another neural network model provided by the embodiment of the present application further includes six fully connected layers between the input layer and the output layer.
  • the structure of the neural network model has a lower complexity, and a reasonable distribution of the transmission power can be realized with less calculation.
  • the number of nodes in the input layer of the above neural network model is the same as the number of input parameters.
  • the input layer can map input parameters corresponding to M (M is a positive integer) antennas into M features (ie, input The number of nodes contained in a layer is M), and then the M features are transferred to the next layer.
  • M is a positive integer
  • the number of antennas can be 64.
  • the number of nodes of the four fully connected layers after the input layer can be configured as 256, 220, 128, 85, 80, and 64 in turn. These fully connected layers can map the input features of each layer to 220, 85, 80, and 64 in turn feature.
  • the output layer of the neural network model can contain K (K is a positive integer) nodes, its convolution kernel size can be 3*3, and its step parameter can be configured to 2. Therefore, the output layer can fully connect the previous one
  • K is the number of users in the communication system, and the K output results are the transmit power allocated to each user.
  • the neural network model involved in S101 may be trained offline.
  • a loss function including the first penalty term and/or the second penalty term can be used, and the loss function can be used for offline training of the neural network model.
  • the first penalty term can be used to constrain the transmission power to be greater than the target power value
  • the second penalty term can be used to constrain the transmission rate allocated to each user to be not less than the minimum value of the target transmission rate.
  • the loss function can be expressed by the following formula:
  • L represents the calculation result of the loss function
  • N represents the number of training samples of the neural network model
  • M represents the number of antennas in the communication system
  • K represents the number of users in the communication system
  • R sum represents each user
  • R min represents the minimum value of the transmission rate
  • R m,k represents the transmission rate allocated by the m-th antenna to the k-th user
  • p m represents the transmission power of the m-th antenna
  • ⁇ represents the value of the first penalty term Coefficient
  • ⁇ represents the coefficient of the second penalty term
  • Indicates the first penalty item Indicates the second penalty item.
  • the range of ⁇ is [0,1].
  • the value range of ⁇ is [0,1].
  • R sum can be expressed by the following formula:
  • ⁇ 2 represents the variance of additive white Gaussian noise, Drying ratio for the letter, ⁇ k represents the k-th user power allocation factor, ⁇ l represents other than k-th user's l th user power allocation factor.
  • the training samples can be input to the neural network model shown in Figure 4 or Figure 5, and the weights and deviations of the neural network can be updated based on Equation 7.
  • the training algorithm used here may be a stochastic gradient descent algorithm. After many iterations, the output of the neural network model can be obtained.
  • the computer device may include hardware structures and/or software modules corresponding to each function.
  • the computer device may have the structure shown in FIG. 2.
  • this application can be implemented in the form of hardware, computer software or a combination of hardware and computer software. Whether a certain function is executed by hardware or computer software-driven hardware depends on the specific application and design constraint conditions of the technical solution.
  • the computer device provided in the embodiment of the present application may have a structure as shown in FIG. 6.
  • a computer device 600 provided by an embodiment of the present application may have an input module 601, a feature extraction module 602, a fitting module 603, and an output module 604.
  • the computer device 600 can be used to execute the method provided in the embodiments of the present application to optimize the power allocation scheme of the antenna of the MIMO communication system.
  • the input module 601 can be used to input input parameters into the input layer of the neural network model, where the input parameters include the channel vector characteristics of each of the multiple antennas in the communication system for each of the multiple users.
  • the feature extraction module 602 can be used to extract features of the communication system according to the input parameters;
  • the fitting module 603 can be used to fit the features through multiple iterations based on the neural network model, and The fitting result is passed to the output layer of the neural network model;
  • the output module 604 can be used to obtain the transmission power allocated to each user from the output layer, wherein the transmission power is determined according to the fitting result .
  • the activation function of the output layer is a Maxout function.
  • multiple hidden layers of the neural network model are included between the input layer and the output layer.
  • the activation function of the output layer can be expressed by the following formula:
  • f Maxout represents the operation result of the activation function
  • x represents the output value set of the hidden layer adjacent to the output layer
  • the x includes X output values
  • X is a positive integer
  • w i represents the i-th
  • Sigmoid() represents the Sigmoid function.
  • y represents the operation result of the activation function
  • P max is a preset value
  • x represents the output value set of the hidden layer adjacent to the output layer
  • the x includes X output values
  • X is a positive integer .
  • the activation functions of the input layer and the multiple hidden layers can be set as linear rectification functions.
  • the loss function of the neural network model includes a first penalty term and/or a second penalty term, wherein the loss function is used for offline training of the neural network model.
  • the first penalty term is used to constrain the transmission power to be greater than a target power value.
  • the second penalty term is used to constrain the transmission rate allocated to each user to be not less than the minimum value of the target transmission rate.
  • the value of the coefficient of the first penalty term may be [0, 1]; the value of the coefficient of the second penalty term may be [0, 1].
  • the loss function of the neural network model is expressed by the following formula:
  • L represents the calculation result of the loss function
  • N represents the number of training samples of the neural network model
  • M represents the number of antennas
  • K represents the number of users
  • R sum represents the number of users allocated to each user.
  • R min represents the minimum value of the transmission rate
  • R m,k represents the transmission rate allocated by the m-th antenna to the k-th user
  • p m represents the transmission power of the m-th antenna
  • ⁇ represents the coefficient of the first penalty term
  • ⁇ represents the coefficient of the second penalty term
  • the R sum can be expressed by the following formula:
  • ⁇ 2 is the variance of the additive white Gaussian noise, It is the letter dry ratio.
  • the value range of ⁇ i,j,l mentioned above can be [0,1].
  • FIG. 6 only shows a modular division method of the computer device 600, and this application does not limit the computer device 600 to have other module division methods.
  • the computer device 600 can be modularized into a processing unit and a storage unit.
  • the processing unit may have the functions of the input module 601, the feature extraction module 602, the fitting module 603, and the output module 604.
  • the storage unit may be used to store the application programs, instructions, and corresponding data required by the processing unit to perform the above functions, thereby processing
  • the unit and the storage unit cooperate with each other to enable the computer device 600 to implement the functions of the power distribution method provided in the embodiment of the present application.
  • the storage unit can also be used to store the above neural network model, and obtain the neural network model when performing the above operation based on the data network model.
  • the storage unit may store a neural network model that has not been trained offline, or the storage unit may also be used to store a neural network model that has undergone offline training.
  • the processing unit can also be used for offline training of the above neural network model.
  • the method provided by the embodiment of the present application can also be implemented by a computer device as shown in FIG. 2.
  • the processor 210 can be used to execute the steps performed by the input module 601, the feature extraction module 602, the fitting module 603, and the output module 604 above.
  • the memory 220 can also store a neural network model that has not been trained offline, or the memory 220 can also be used to store a neural network model that has been trained offline.
  • the processor 210 may also be used to perform offline training on the neural network model stored in the memory 202.
  • processors or processing unit described in this application may be a central processing unit (CPU), or a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or Other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or any conventional processor.
  • the memory or storage unit may include a read-only memory and a random access memory, and provide instructions and data to the processor.
  • the memory may also be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.
  • the non-volatile memory can be read-only memory (ROM), programmable read-only memory (programmable ROM, PROM), erasable programmable read-only memory (erasable PROM, EPROM), and electronic Erase programmable read-only memory (electrically EPROM, EEPROM) or flash memory.
  • the volatile memory may be random access memory (RAM), which is used as an external cache.
  • RAM random access memory
  • SRAM static random access memory
  • DRAM dynamic random access memory
  • SDRAM synchronous dynamic random access memory
  • Double data rate synchronous dynamic random access memory double data date SDRAM, DDR SDRAM
  • enhanced SDRAM enhanced synchronous dynamic random access memory
  • SLDRAM synchronous connection dynamic random access memory
  • direct rambus RAM direct rambus RAM
  • the communication module may be a circuit, a device, a communication interface, a bus, a software module, a wireless transceiver, or any other component that can implement information/data transceiving.
  • the embodiment of the present application also provides a computer storage medium on which some instructions are stored. When these instructions are called for execution, the computer can execute any of the above method embodiments and method embodiments. The steps performed by a computer device in one possible implementation.
  • the readable storage medium is not limited, for example, it may be RAM (random-access memory, random access memory), ROM (read-only memory, read-only memory), etc.
  • the embodiment of the present application also provides a computer program product.
  • the computer program product When the computer program product is run by a computer, the computer can make the computer execute any of the above method embodiments and method embodiments. The steps performed by the computer device in the implementation mode.
  • the embodiment of the present application also provides a computer system.
  • the communication system may include the computer device provided in the embodiment of the present application, or include the computer device and other necessary devices, such as input input. Devices, etc.
  • an embodiment of the present application also provides a chip.
  • the chip may include a processor, and the processor may be coupled with the memory.
  • the chip can be used in a computer device to implement the functions involved in any one of the foregoing method embodiments and method embodiments.
  • the embodiment of the present application also provides a chip system.
  • the chip system may include the above-mentioned chip, or may include a chip and other discrete devices.
  • the chip system may include a chip, a memory, and a communication module.

Landscapes

  • Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Artificial Intelligence (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Automation & Control Theory (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

一种基于神经网络的功率分配方法及装置。根据该方法，可通过神经网络模型基于MIMO通信系统的信道向量特征，提取通信系统的特征，进一步基于神经网络模型通过多次迭代对该特征进行拟合，根据拟合结果确定针对每个用户分配的发送功率，由于在功率分配中考虑了通信系统的特征，可以优化发送功率的分配结果。

Description

一种基于神经网络的功率分配方法及装置 技术领域
本申请涉及无线通信技术领域，尤其涉及一种基于神经网络的功率分配方法及装置。
背景技术
多输入输出(multiple-input multiple-output，MIMO)通信系统可以极大地提高下一代通信系统的容量和通信速率，被认为是未来无线通信技术的发展方向。功率分配是MIMO通信系统的重要技术，在信号发射端合理的分配发射功率，能够在不增加额外发射功率和带宽消耗的情况下，令MIMO通信系统获得最大的系统容量，保证MIMO通信系统的通信效率。
随着大规模多输入输出技术(massive MIMO)以及非正交频分多址多输入多输出(multiple-input-multiple-output non-orthogonal multiple access，MIMO-NOMA)等技术的出现，MIMO通信系统的复杂度不断提升，现有功率分配方法由于无法充分学习复杂的MIMO通信系统的空间特征和信道特性导致功率分配效果不够理想。因此，需要寻找一种既能够实现更优的功率分配方案。
发明内容
本申请实施例提供一种基于神经网络的功率分配方法及装置，用于解决目前根据用户选取的关键字符向用户推送的信息不够精确的技术问题。
第一方面，本申请提供一种基于神经网络的功率分配方法。该方法所示步骤可由计算机装置执行。具体的，计算机装置可将输入参数输入至神经网络模型的输入层，所述输入参数包括通信系统中多个天线中的每个天线对于多个用户中的每个用户的信道向量特征，根据所述输入参数，提取所述通信系统的特征，基于所述神经网络模型，通过多次迭代对所述特征进行拟合，并将拟合结果传递到所述神经网络模型的输出层，从所述输出层获取为每个所述用户分配的发送功率，其中，所述发送功率根据所述拟合结果确定。
采用以上方法，通过神经网络模型，根据输入参数提取通信系统的特征，基于该特征确定为每个用户分配的发送功率，能够获得更为优化的功率分配方案，实现MIMO通信系统中天线发送功率的优化。
在一种可能的设计中，该神经网络模型的输出层的激活函数可以是Maxout函数。
在一种可能的设计中，所述输入层与所述输出层之间还可包括所述神经网络模型的多个隐含层。通过这些隐含层进行运算，可提高神经网络模型的运算量，提高结果的运算精度。
在一种可能的设计中，该神经网络模型的输出层的激活函数可以通过以下公式表示：
f Maxout＝Sigmoid(x Tw i+b i)；
其中，f Maxout表示所述激活函数的运算结果，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数，w i表示第i个所述输出值的权重，b i表示第i个所述输出值的偏差，i＝1、2……、X，x T表示x的转置，Sigmoid()表示Sigmoid函数。
在一种可能的设计中，该神经网络模型的输出层的激活函数，还可通过以下公式表示：
y＝min(max(x,0),P max)；
其中，y表示所述激活函数的运算结果，P max为预设值，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数。
在一种可能的设计中，所述输入层和所述多个隐含层的激活函数均为线性整流函数。
在一种可能的设计中，所述神经网络模型的损失函数包含第一惩罚项和/或第二惩罚项，所述损失函数用于所述神经网络模型的离线训练，其中，所述第一惩罚项用于约束所述发送功率大于目标功率值，所述第二惩罚项用于约束为每个用户分配的发送速率不小于目标发送速率的最小值。从而经过离线训练，能够使得神经网络模型中权重和偏差更能够体现通信系统的特征，优化输出结果。
在一种可能的设计中，所述第一惩罚项的系数的取值可以为[0,1]，和/或，所述第二惩罚项的系数的取值可以为[0,1]。
在一种可能的设计中，所述神经网络模型的损失函数可通过以下公式表示：
Figure PCTCN2019079957-appb-000001
其中，L表示所述损失函数的运算结果，N表示所述神经网络模型的训练样本的数量，M表示所述天线的数量，K表示所述用户的数量，R sum表示为每个用户分配的发送速率的总和，β i,j,l为功率分配因子，i＝1、2、……、N，j＝1、2、……、N，l＝1、2、……、N，R min表示所述发送速率的最小值，R m,k表示第m个天线为第k个用户分配的发送速率，p m表示第m个天线的发送功率，τ表示所述第一惩罚项的系数，ρ表示所述第二惩罚项的系数，
Figure PCTCN2019079957-appb-000002
表示所述第一惩罚项，
Figure PCTCN2019079957-appb-000003
表示所述第二惩罚项。由于功率分配因子并非设置为一个数值，可根据该因子灵活调整功率分配结果，如提高针对某些用户的发送功率，以针对用户提高通信质量。
在一种可能的设计中，所述R sum可以通过以下公式表示：
Figure PCTCN2019079957-appb-000004
其中，δ 2为加性高斯白噪声的方差，
Figure PCTCN2019079957-appb-000005
为信干燥比。由于针对不同用户的功率分配因子β并非设置为一个相同的数值，可根据该因子灵活调整功率分配结果，如提高针对某些用户的发送功率，以针对用户提高通信质量。
在一种可能的设计中，所述β i,j,l的取值范围可以为[0,1]。
第二方面，本申请实施例提供一种计算机装置，该计算机装置可具有上述第一方面或第一方面的任意一种可能的设计中所涉及方法的功能，从而可用于执行上述第一方面或第一方面的任意一种可能的涉及所示的方法。该功能可通过硬件实现，也可以通过硬件执行相应的软件实现。该硬件或软件包括一个或多个与上述功能相对应的模块。
在一种可能的设计中，该计算机装置可包括输入模块、特征提取模块、拟合模块以及输出模块。其中，输入模块可用于将输入参数输入至神经网络模型的输入层，所述输入参数包括通信系统中多个天线中的每个天线对于多个用户中的每个用户的信道向量特征。特征提取模块可用于根据所述输入参数，提取所述通信系统的特征。拟合模块可用于基于所述神经网络模型，通过多次迭代对所述特征进行拟合，并将拟合结果传递到所述神经网络模型的输出层。输出模块可用于从所述输出层获取为每个所述用户分配的发送功率，其中，所述发送功率根据所述拟合结果确定。
在一种可能的设计中，该神经网络模型的输出层的激活函数可以是Maxout函数。
在一种可能的设计中，所述输入层与所述输出层之间还可包括所述神经网络模型的多个隐含层。
在一种可能的设计中，该神经网络模型的输出层的激活函数可以通过以下公式表示：
f Maxout＝Sigmoid(x Tw i+b i)；
其中，f Maxout表示所述激活函数的运算结果，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数，w i表示第i个所述输出值的权重，b i表示第i个所述输出值的偏差，i＝1、2……、X，x T表示x的转置，Sigmoid()表示Sigmoid函数。
在一种可能的设计中，该神经网络模型的输出层的激活函数，还可通过以下公式表示：
y＝min(max(x,0),P max)；
其中，y表示所述激活函数的运算结果，P max为预设值，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数。
在一种可能的设计中，所述输入层和所述多个隐含层的激活函数均为线性整流函数。
在一种可能的设计中，所述神经网络模型的损失函数包含第一惩罚项和/或第二惩罚项，所述损失函数用于所述神经网络模型的离线训练，其中，所述第一惩罚项用于约束所述发送功率大于目标功率值，所述第二惩罚项用于约束为每个用户分配的发送速率不小于目标发送速率的最小值，从而经过训练能够使得神经网络模型中权重和偏差更能够体现通信系统的特征，优化输出结果。
在一种可能的设计中，所述第一惩罚项的系数的取值可以为[0,1]，和/或，所述第二惩罚项的系数的取值可以为[0,1]。
在一种可能的设计中，所述神经网络模型的损失函数可通过以下公式表示：
Figure PCTCN2019079957-appb-000006
其中，L表示所述损失函数的运算结果，N表示所述神经网络模型的训练样本的数量，M表示所述天线的数量，K表示所述用户的数量，R sum表示为每个用户分配的发送速率的总和，β i,j,l为功率分配因子，i＝1、2、……、N，j＝1、2、……、N，l＝1、2、……、N，R min表示所述发送速率的最小值，R m,k表示第m个天线为第k个用户分配的发送速率，p m表示第m个天线的发送功率，τ表示所述第一惩罚项的系数，ρ表示所述第二惩罚项的系数，
Figure PCTCN2019079957-appb-000007
表示所述第一惩罚项，
Figure PCTCN2019079957-appb-000008
表示所述第二惩罚项。
在一种可能的设计中，所述R sum可以通过以下公式表示：
Figure PCTCN2019079957-appb-000009
其中，δ 2为加性高斯白噪声的方差，
Figure PCTCN2019079957-appb-000010
为信干燥比。
在一种可能的设计中，所述β i,j,l的取值范围可以为[0,1]。
第三方面，本申请实施例提供一种计算机可读存储介质，包括程序指令，当所述程序指令在计算机上运用时，使得计算机实现第一方面或第一方面的任意可能的设计所涉及的功能。
第四方面，本申请实施例提供一种计算机程序产品，当其在计算机上运行时，使得计算机实现第一方面或第一方面的任意可能的设计所涉及的功能。
第五方面，本申请实施例提供一种系统，该系统包括第二方面或第二方面的任意可能的设计所涉及的装置。
第六方面，本申请实施例提供一种芯片，该芯片可以与存储器耦合，用于读取并执行该存储器中存储的程序或指令，以实现上述第一方面或第二方面第一方面的任意可能的设计所涉及的功能。
以上第二方面至第六方面任意可能的设计的有益效果，可参照第一方面及第一方面中任意可能的设计的有益效果。
附图说明
图1为本申请实施例适用的一种神经网络模型的示意图；
图2为本申请实施例提供的一种计算机装置的结构示意图；
图3为本申请实施例提供的一种管理分配方法的流程示意图；
图4为本申请实施例提供的一种神经网络模型的结构示意图；
图5为本申请实施例提供的另一种神经网络模型的结构示意图；
图6为本申请实施例提供的另一种计算机装置的结构示意图。
具体实施方式
应理解，本申请实施例中“至少一个”是指一个或者多个，“多个”是指两个或两个以上。“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的情况，其中A、B可以是单数或者复数。字符“/”一般表示前后关联对象是一种“或”的关系。“以下至少一(项)个”或其类似表达，是指的这些项中的任意组合，包括单项(个)或复数项(个)的任意组合。例如，a、b或c中的至少一项(个)，可以表示：a，b，c，a和b，a和c，b和c，或a、b和c，其中a、b、c可以是单个，也可以是多个。[a,b]表示取值在范围a(含a)到b(含b)之间，a小于等于b。
下面，对本申请实施例涉及的技术术语进行介绍：
1、神经网络(neural networks，NN)，是由大量、简单的节点(或称为处理单元、神经元或神经元节点等)广泛地互相连接而形成的复杂网络系统，它反映了人脑功能的许多基本特征。通过数学模型可模仿神经元的功能和特性，从而以神经元的数学模型为基础可构建神经网络(在本申请中也称为神经网络模型，或网络模型等)。神经网络模型可包含多层结构，如包含输入层(input layer)，即神经网络的第一层，和输出层(output layer)，即神经网络的最后一层。在输入层和输出层之间，还可包括隐含层(hidden layer)。具备多个隐含层的神经网络也可被称为深度神经网络(deep neural networks，DNN)。
神经网络中的节点如图1中的(a)图所示，图中的圆圈左侧的a1～an为该节点的输入数据，w1～wn为该节点的权值，b为该节点的偏置。图中的圆圈代表该节点的内部计算，即一个特定的输出函数f，又称为激活函数或激励函数。每个节点可以有多个输出数据z，但是值相同。
如图1中的(b)图所示，神经网络可由输入层、隐含层、输出层组成。其中，输入层用于输入神经网络的输入数据；输出层用于输出神经网络的输出数据；而隐含层由输入层和输出层之间众多节点连接组成，用于对输入数据进行运算处理。其中，隐含层可以包含一层或多层结构。神经网络中隐含层的层数、节点数，与该神经网络实际解决的问题的复杂程度、输入层的节点以及输出层的节点的个数有着直接关系。
2、卷积层(convolutional layer)，在神经网络模型中用于通过卷积运算提取输入数据的局部特征，例如针对输入图像获取各类局部特征图。应理解，本申请实施例所涉及的卷积层，不限于卷积神经网络中的卷积层，还可以是其他类型神经网络模型中的卷积层。
3、全连接层(fully connected layers，FC)的每一个节点都与上一层的每一个节点连接，全连接层可用于将上一层提取到的局部特征根据权值进行综合，例如，将各类局部特征图再次组合为完整的图像。全连接层的每一个输出都可以看成前一层的每一个结点乘以一个权重系数，最后加上一个偏差得到。
4、激活函数。神经网络中的节点接受上一层节点的输出值作为本节点的输入值，如，输入层节点会将输入属性值直接传递给下一层(隐含层或输出层)的节点，在多层神经网络中，上一层节点的输出和下一层节点的输入之间具有一个函数关系，这个函数称为激活函数(又称激励函数)。
5、神经网络的损失函数，用于描述神经网络的预测值和真实值之间的偏差。损失函数是一个非负函数。损失函数越小，则神经网络模型的鲁棒性越好。一个神经网络有好的性能意味着找到了最适合的权值，使损失函数最小，这个寻找最合适的权值的过程就是神经网络学习的过程。
下面结合附图对本申请实施例进行具体说明。
图2示出了本申请实施例提供的方法适用的一种可能的计算机装置的结构图。参阅图2所示，所述计算机装置中包括：处理器210、存储器220、通信模块230、输入单元240、显示单元250、电源260等部件。本领域技术人员可以理解，图2中示出的计算机装置的结构并不构成对计算机装置的一种限定，本申请实施例提供的计算机装置可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件布置。
下面结合图2对计算机装置的各个构成部件进行具体的介绍：
所述通信模块230可以通过无线连接或物理连接的方式连接其他设备，实现计算机装置的数据发送和接收。可选的，所述通信模块230可以包含射频(radio frequency，RF)电路、无线保真(wireless fidelity，WiFi)模块、通信接口，蓝牙模块等任一项或组合，本申请实施例对此不作限定。
所述存储器220可用于存储程序指令和数据。所述处理器210通过运行存储在所述存储器220的程序指令，从而执行计算机装置的各种功能应用以及数据处理。其中，所述程序指令中存在可使所述处理器210执行本申请以下实施例提供的功率分配方法的程序指令。
可选的，所述存储器220可以主要包括存储程序区和存储数据区。其中，存储程序区可存储操作系统、各种应用程序，以及程序指令等；存储数据区可存储神经网络等各种数据。此外，所述存储器220可以包括高速随机存取存储器，还可以包括非易失性存储器，例如磁盘存储器件、闪存器件、或其他易失性固态存储器件。
所述输入单元240可用于接收用户输入的数据或操作指令等信息。可选的，所述输入单元240可包括触控面板、功能键、物理键盘、鼠标、摄像头、监控器等输入设备。
所述显示单元250可以实现人机交互，用于通过用户界面显示由用户输入的信息，提供给用户的信息等内容。其中，所述显示单元250可以包括显示面板251。可选的，所述显示面板251可以采用液晶显示屏(liquid crystal display，LCD)、有机发光二极管(organic light-emitting diode，OLED)等形式来配置。
进一步的，当输入单元中包含触控面板时，该触控面板可覆盖所述显示面板251，当 所述触控面板检测到在其上或附近的触摸事件后，传送给所述处理器210以确定触摸事件的类型从而执行相应的操作。
所述处理器210是计算机装置的控制中心，利用各种接口和线路连接以上各个部件。所述处理器210可以通过执行存储在所述存储器220内的程序指令，以及调用存储在所述存储器220内的数据，以完成计算机装置的各种功能，实现本申请实施例提供的神经网络精度调整方法。
可选的，所述处理器210可包括一个或多个处理单元。具体的，所述处理单元可以包括CPU和/或GPU等能够基于浮点运算的硬件设备。该处理单元能够对输神经网络的数据进行处理后输出。在所述计算机装置的处理器210实现神经网络精度调整方法时，该处理单元从存储器220的存储数据区读取神经网络的数据，并对输入神经网络的数据进行量化，还可以对神经网络进行前向传播和反向传播等处理。
所述计算机装置还包括用于给各个部件供电的电源260(比如电池)。可选的，所述电源260可以通过电源管理系统与所述处理器210逻辑相连，从而通过电源管理系统实现对所述计算机装置的充电、放电等功能。
尽管未示出，所述计算机装置还可以包括摄像头、传感器、音频采集器等部件，在此不再赘述。
本申请实施例提供了一种基于神经网络的功率分配方法，该方法可适用于如图2所示的计算机装置和/或图1(b)所示的神经网络。该方法可以由如图2所示的计算机装置中的处理器执行。参阅图3所示，该方法的流程可包括：
S101：将输入参数输入至神经网络模型的输入层，该输入参数包括通信系统中多个天线中的每个天线对于多个用户中的每个用户的信道向量特征；
S102：根据该输入参数，提取该通信系统的特征；
S103：基于该神经网络模型，通过多次迭代对该特征进行拟合，并将拟合结果传递到该神经网络模型的输出层；
S104：从所述输出层获取为每个用户分配的发送功率，其中，该发送功率根据该拟合结果确定。
采用以上方法，可通过神经网络模型基于MIMO通信系统的信道向量特征，提取通信系统的特征，进一步基于神经网络模型通过多次迭代对该特征进行拟合，根据拟合结果确定针对每个用户分配的发送功率，由于在功率分配中考虑了通信系统的特征，可以优化发送功率的分配结果。
应理解，S101所涉及的通信系统，包括但不限于传统MIMO、大规模MIMO、MIMO-NOMA通信系统或其他MIMO系统。如图3所示方法可用于进行上述通信系统的发送功率分配，以优化功率分配方案。
若定义p m为以上S101所涉及的通信系统中第m(m为正整数)个天线的发送功率，
Figure PCTCN2019079957-appb-000011
为该通信系统中第m个天线针对第k(k为正整数)个用户的第i个信道向量特征，以及
Figure PCTCN2019079957-appb-000012
为针对第k个用户的第i个功率向量样本，可组成神经网络训练的样本集
Figure PCTCN2019079957-appb-000013
本申请实施例中，为了最大化该通信系统的总发送速率，可确定以下公式所示的优化问题：
Figure PCTCN2019079957-appb-000014
使得满足：
Figure PCTCN2019079957-appb-000015
Figure PCTCN2019079957-appb-000016
Figure PCTCN2019079957-appb-000017
Figure PCTCN2019079957-appb-000018
其中，R sum表示该通信系统的总发送速率(即该通信系统中为每个用户分配的发送速率的总和)，R m,k表示第m个天线为第k个用户分配的发送速率，β i,j为功率分配因子，其中，i＝1、2、……、N，j＝1、2、……、N，i可表示样本数量，j可表示迭代次数，P表示该通信系统的总发射功率，R min表示第m个天线为第k个用户分配的发送速率的最小值。β i,j,l可以是预设的值，其取值范围为[0,1]。
在实施中，可根据第m个天线针对第k个用户的信干燥比(signal interference noise ratio，SINR)确定R m,k。具体的，可通过以下公式确定R m,k：
R m,k＝log 2(1+γ m,k)，(公式二)
其中，γ m,k表示第m个天线针对第k个用户的信干燥比。
例如，可根据以下公式确定γ m,k：
Figure PCTCN2019079957-appb-000019
其中，β k表示第k个用户的功率分配因子，β l表示第k个用户以外的第l个用户的功率分配因子，δ 2为加性高斯白噪声的方差。β k以及β l均为预设的值，β k以及β l的取值范围为[0,1]。在实施中，用户的功率分配因子可根据针对用户的信道状态信息(channel state information，CSI)确定。
另外，以上优化问题中，C1项所示条件用于约束该通信系统中所有天线的发送功率的总和，不大于该通信系统总的总发射功率。C2项所示条件用于约束任意一个天线为任意一个用户分配的发送速率不小于发送功率的最小值。C3项所示条件用于约束任意一个天线的发送功率不小于0。C4项所示条件用于约束每个功率分配因子的取值范围为[0,1]。基于以上问题，寻求功率分配因子和每一个天线最优的发射功率，可获得最优的功率分配方案。
示例性的，本申请实施例所涉及的神经网络模型可包括输入层以及输出层以及位于输入层与输出层之间的多个隐含层，其中，多个隐含层可包括至少一个卷积层和/或一个全连接层。其中，该神经网络模型的输出层可以是卷积层。
在一种可能的实现方式中，本申请实施例所涉及的神经网络模型的输出层的激活函数为Maxout函数。例如，将如下公式作为输出层的激活函数：
f Maxout＝Sigmoid(x Tw... i+b i)，(公式四)
其中，f Maxout表示该激活函数的运算结果，x表示与该输出层相邻的隐含层的输出值集合，x可包含X个输出值，x可以是由X个输出值组成的矩阵，X为正整数，且X可以与该相邻的隐含层的节点数量相同，w i表示第i个所述输出值的权重，b i表示第i个所述输出值的偏差，i＝1、2……、X，x T表示x的转置，Sigmoid()表示Sigmoid函数。
在另外的实现方式中，输出层的激活函数还通过以下公式表示：
y＝min(max(x,0),P max)，(公式五)
其中，y表示该激活函数的运算结果，P max可表示基站的最大发射功率，P max可以是预设值，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数。
另外，对于本申请实施例所涉及的输入层以及多个隐含层，均可将线性整流(rectified linear unit，ReLU)函数作为其激活函数。具体的，输入层以及多个隐含层中任一层的激活函数，可由以下公式表示：
f(x0)＝max(0,x0)，(公式六)
其中，x0表示与该层相邻的前一层的输出值集合，x0可以是由X0个输出值组成的矩阵，X0为正整数，且X0可以与该相邻的隐含层的节点数量相同。
如图4所示，本申请实施例提供的一种神经网络模型可包括多个卷积层以及多个全连接层，其中，输入层为全连接层，输入层与输出层之间依次还可包括两个卷积层以及四个全连接层，输入层与一个卷积层相邻，输出层与一个全连接层相邻。
示例性的，以上神经网络模型的输入层的节点数量与输入参数的数量相同，用于将输入参数映射为天线的特征，例如，输入层可将M(M为正整数)个天线分别对应的输入参数映射为M个特征(即，输入层包含的节点数量为M)，之后将该M个特征传递至下一层，如，天线数量可以为64。与输入层相邻的第一个卷积层可用于将输入层传递的M个特征映射为64个特征，其卷积核的尺寸可以是7*7，该卷积层的步长(stride)参数可配置为2，步长参数可用于指示卷积核在多个输入特征组成的输入特征图中扫描时每次跳跃的格数，通过补偿参数控制跳跃的格数，可减少卷积核进行扫描时进行的重复计算，以提升扫描效率。与该第一个卷积核相邻的卷积层(以下称为第二个卷积层)可用于将第一个卷积层传递的64个特征映射为32个特征，其卷积核的尺寸可以是3*3，其步长参数可配置为2。第二个卷积层之后的四个全连接层的节点数量依次可配置为220、85、80以及64，该四个全连接层依次可将各层的输入特征映射为220、85、80以及64个特征。该神经网络模型的输出层可包含K(K为正整数)个节点，其卷积核尺寸可以是3*3，其步长参数可配置为2，因此，该输出层可将前一个全连接层传递的64个特征映射为K个输出结果，其中K为该通信系统中用户的数量，该K个输出结果即为每个用户分配的发送功率，如，用户数量可以为32。
如图5所示，本申请实施例提供的另一种神经网络模型在输入层与输出层之间依次还包括六个全连接层。该神经网络模型的结构相比于图4所示神经网络模型的结构的复杂度较低，可通过较少的计算量实现发送功率的合理分配。
示例性的，以上神经网络模型的输入层的节点数量与输入参数的数量相同，例如，输入层可将M(M为正整数)个天线分别对应的输入参数映射为M个特征(即，输入层包含的节点数量为M)，之后将该M个特征传递至下一层，如，天线数量可以为64。
输入层之后的四个全连接层的节点数量依次可配置为256、220、128、85、80以及64，这些全连接层依次可将各层的输入特征映射为220、85、80以及64个特征。该神经网络模型的输出层可包含K(K为正整数)个节点，其卷积核尺寸可以是3*3，其步长参数可配置为2，因此，该输出层可将前一个全连接层传递的64个特征映射为K个输出结果，其中K为该通信系统中用户的数量，该K个输出结果即为每个用户分配的发送功率。
在本申请实施例中，S101所涉及的神经网络模型可经过离线训练。在进行神经网络模型的离线训练时，可采用包含第一惩罚项和/或第二惩罚项的损失函数，该损失函数可用于 进行该神经网络模型的离线训练。其中，第一惩罚项可用于约束所述发送功率大于目标功率值，所述第二惩罚项可用于约束为每个用户分配的发送速率不小于目标发送速率的最小值。
具体的，损失函数可通过以下公式表示：
Figure PCTCN2019079957-appb-000020
其中，L表示该损失函数的运算结果，N表示该神经网络模型的训练样本的数量，M表示该通信系统中天线的数量，K表示该通信系统中用户的数量，R sum表示为每个用户分配的发送速率的总和，β i,j,l为功率分配因子，i＝1、2、……、N，j＝1、2、……、N，l＝1、2、……、N，R min表示该发送速率的最小值，R m,k表示第m个天线为第k个用户分配的发送速率，p m表示第m个天线的发送功率，τ表示所述第一惩罚项的系数，ρ表示该第二惩罚项的系数，
Figure PCTCN2019079957-appb-000021
表示该第一惩罚项，
Figure PCTCN2019079957-appb-000022
表示该第二惩罚项。其中，τ的取值范围均为[0,1]。和/或，ρ的取值范围为[0,1]。
示例性的，R sum可以通过以下公式表示：
Figure PCTCN2019079957-appb-000023
其中，δ 2表示加性高斯白噪声的方差，
Figure PCTCN2019079957-appb-000024
为信干燥比，β k表示第k个用户的功率分配因子，β l表示第k个用户以外的第l个用户的功率分配因子。
在进行离线训练时，可将训练样本输入如图4或图5所示的神经网络模型，基于公式七对神经网络的权重和偏差进行更新。这里采用的训练算法可以是随机梯度下降算法。在经过多次迭代后，可得到神经网络模型的输出。
以上主要从本申请实施例所提供的计算机装置能够执行的操作以及本申请实施例所涉及的神经网络的架构的角度，对本申请实施例提供的方案进行了介绍。可以理解的是，该计算机装置为了实现上述功能，可包含执行各个功能相应的硬件结构和/或软件模块。例如，该计算机装置可具有如图2所示结构。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、计算机软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。
当通过软件模块来实现时，本申请实施例提供的计算机装置可具有如图6所示的结构。如图6所示，本申请实施例提供的一种计算机装置600，可具有输入模块601、特征提取模块602、拟合模块603以及输出模块604。该计算机装置600可用于执行本申请实施例所提供的方法，以优化MIMO通信系统天线的功率分配方案。
具体的，该输入模块601可用于将输入参数输入至神经网络模型的输入层，所述输入参数包括通信系统中多个天线中的每个天线对于多个用户中的每个用户的信道向量特征；该特征提取模块602可用于根据所述输入参数，提取所述通信系统的特征；该拟合模块603可用于基于所述神经网络模型，通过多次迭代对所述特征进行拟合，并将拟合结果传递到所述神经网络模型的输出层；该输出模块604可用于从所述输出层获取为每个所述用户分配的发送功率，其中，所述发送功率根据所述拟合结果确定。
可选的，所述输出层的激活函数为Maxout函数。
在一种可能的实现方式中，所述输入层与所述输出层之间包括所述神经网络模型的多个隐含层。
所述输出层的激活函数，可通过以下公式表示：
f Maxout＝Sigmoid(x Tw i+b i)；
其中，f Maxout表示所述激活函数的运算结果，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数，w i表示第i个所述输出值的权重，b i表示第i个所述输出值的偏差，i＝1、2……、X，x T表示x的转置，Sigmoid()表示Sigmoid函数。
或者，所述激活函数通过以下公式表示：
y＝min(max(x,0),P max)；
其中，y表示所述激活函数的运算结果，P max为预设值，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数。
另外，所述输入层和所述多个隐含层的激活函数均可设置为线性整流函数。
示例性的，所述神经网络模型的损失函数包含第一惩罚项和/或第二惩罚项，其中，该损失函数用于所述神经网络模型的离线训练。所述第一惩罚项用于约束所述发送功率大于目标功率值。所述第二惩罚项用于约束为每个用户分配的发送速率不小于目标发送速率的最小值。在实施中，所述第一惩罚项的系数的取值可以为[0,1]；所述第二惩罚项的系数的取值可以为[0,1]。
示例性的，所述神经网络模型的损失函数通过以下公式表示：
Figure PCTCN2019079957-appb-000025
其中，L表示所述损失函数的运算结果，N表示所述神经网络模型的训练样本的数量，M表示所述天线的数量，K表示所述用户的数量，R sum表示为每个用户分配的发送速率的总和，β i,j,l为功率分配因子，i＝1、2、……、N，j＝1、2、……、N，l＝1、2、……、N，R min表示所述发送速率的最小值，R m,k表示第m个天线为第k个用户分配的发送速率，p m表示第m个天线的发送功率，τ表示所述第一惩罚项的系数，ρ表示所述第二惩罚项的系数，
Figure PCTCN2019079957-appb-000026
表示所述第一惩罚项，
Figure PCTCN2019079957-appb-000027
表示所述第二惩罚项。
所述R sum可以通过以下公式表示：
Figure PCTCN2019079957-appb-000028
其中，δ 2为加性高斯白噪声的方差，
Figure PCTCN2019079957-appb-000029
为信干燥比。
以上所述β i,j,l的取值范围可以是[0,1]。
应理解，图6仅示出了计算机装置600的一种模块化的划分方式，本申请并不限制计算机装置600具有其他模块划分方式，例如，计算机装置600可模块化为处理单元、存储单元，其中，处理单元可具有上述输入模块601、特征提取模块602、拟合模块603以及输出模块604的功能，存储单元可用于存储处理单元执行上述功能所需的应用程序、指令和相应数据，从而处理单元与存储单元相互配合，令计算机装置600实现本申请实施例提供的功率分配方法所具有的功能。存储单元还可用于存储以上神经网络模型，并在执行以上基于数据网络模型的操作时，获取该神经网络模型。示例性的，存储单元可存储未经离线训练的神经网络模型，或者，存储单元也可用于存储经过离线训练的神经网络模型。可选地，处理单元还可用于对以上神经网络模型进行离线训练。
另外，本申请实施例提供的方法也可由如图2所示的计算机装置实现。应理解，在实现该方法时，处理器210可用于执行以上由输入模块601、特征提取模块602、拟合模块 603以及输出模块604执行的步骤。另外，存储器220还可存储未经离线训练的神经网络模型，或者，存储器220也可用于存储经过离线训练的神经网络模型。处理器210还可用于对存储器202所存储的神经网络模型进行离线训练。
应理解，本申请所述处理器或处理单元可以是中央处理单元(CPU)，或者通用处理器、数字信号处理器(DSP)、专用集成电路(ASIC)、现场可编程门阵列(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。其中，通用处理器可以是微处理器或者是任何常规的处理器等。
存储器或存储单元可以包括只读存储器和随机存取存储器，并向处理器提供指令和数据。存储器还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(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 date 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)。
通信模块，可以是电路、器件、通信接口、总线、软件模块、无线收发器或者其它任意可以实现信息/数据收发的组件。
基于与上述实施例相同构思，本申请实施例还提供了一种计算机存储介质，其上存储有一些指令，这些指令被调用执行时，可以使得计算机执行上述方法实施例、方法实施例的任意一种可能的实现方式中由计算机装置所执行的步骤。本申请实施例中，对可读存储介质不做限定，例如，可以是RAM(random-access memory，随机存取存储器)、ROM(read-only memory，只读存储器)等。
基于与上述方法实施例相同构思，本申请实施例还提供了一种计算机程序产品，当所述计算机程序产品被计算机运行时，可以使得计算机执行上述方法实施例、方法实施例的任意一种可能的实现方式中由计算机装置所执行的步骤。
基于与上述方法实施例相同构思，本申请实施例还提供了一种计算机系统，该通信系统可包括本申请实施例提供的计算机装置，或者，包括该计算机装置以及其他必要的装置，如输入输装置等。
基于与上述方法实施例相同构思，本申请实施例还提供了一种芯片。该芯片可包括处理器，处理器可以与存储器耦合。该芯片可用于计算机装置实现上述方法实施例、方法实施例的任意一种可能的设计中所涉及的功能。
另外，本申请实施例还提供了一种芯片系统。该芯片系统可包括上述芯片，也可以包含芯片和其他分立器件，例如，芯片系统可包含芯片、存储器以及通信模块。

Claims (22)

  1. 一种基于神经网络的功率分配方法，其特征在于，包括：
    将输入参数输入至神经网络模型的输入层，所述输入参数包括通信系统中多个天线中的每个天线对于多个用户中的每个用户的信道向量特征；
    根据所述输入参数，提取所述通信系统的特征；
    基于所述神经网络模型，通过多次迭代对所述特征进行拟合，并将拟合结果传递到所述神经网络模型的输出层；
    从所述输出层获取为每个所述用户分配的发送功率，其中，所述发送功率根据所述拟合结果确定。
  2. 如权利要求1所述的方法，其特征在于，所述输出层的激活函数为Maxout函数。
  3. 如权利要求1所述的方法，其特征在于，所述输入层与所述输出层之间包括所述神经网络模型的多个隐含层。
  4. 如权利要求3所述的方法，其特征在于，所述激活函数通过以下公式表示：
    f Maxout＝Sigmoid(x Tw i+b i)；
    其中，f Maxout表示所述激活函数的运算结果，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数，w i表示第i个所述输出值的权重，b i表示第i个所述输出值的偏差，i＝1、2……、X，x T表示x的转置，Sigmoid()表示Sigmoid函数。
  5. 如权利要求3所述的方法，其特征在于，所述激活函数通过以下公式表示：
    y＝min(max(x,0),P max)；
    其中，y表示所述激活函数的运算结果，P max为预设值，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数。
  6. 如权利要求3所述的方法，其特征在于，所述输入层和所述多个隐含层的激活函数均为线性整流函数。
  7. 如权利要求1-6任一所述的方法，其特征在于，
    所述神经网络模型的损失函数包含第一惩罚项和/或第二惩罚项，所述损失函数用于所述神经网络模型的离线训练；
    所述第一惩罚项用于约束所述发送功率大于目标功率值；
    所述第二惩罚项用于约束为每个用户分配的发送速率不小于目标发送速率的最小值。
  8. 如权利要求7所述的方法，其特征在于，
    所述第一惩罚项的系数的取值为[0,1]；和/或，
    所述第二惩罚项的系数的取值为[0,1]。
  9. 如权利要求7或8所述的方法，其特征在于，所述神经网络模型的损失函数通过以下公式表示：
    Figure PCTCN2019079957-appb-100001
    其中，L表示所述损失函数的运算结果，N表示所述神经网络模型的训练样本的数量，M表示所述天线的数量，K表示所述用户的数量，R sum表示为每个用户分配的发送速率的总和，β i,j,l为功率分配因子，i＝1、2、……、N，j＝1、2、……、N，l＝1、2、……、N，R min表示所述发送速率的最小值，R m,k表示第m个天线为第k个用户分配的发送速率，p m表示第m个天线的发送功率，τ表示所述第一惩罚项的系数，ρ表示所述第二惩罚项的系数，
    Figure PCTCN2019079957-appb-100002
    表示所述第一惩罚项，
    Figure PCTCN2019079957-appb-100003
    表示所述第二惩罚项。
  10. 如权利要求9所述的方法，其特征在于，所述R sum通过以下公式表示：
    Figure PCTCN2019079957-appb-100004
    其中，δ 2为加性高斯白噪声的方差，
    Figure PCTCN2019079957-appb-100005
    为信干燥比。
  11. 如权利要求9或10所述的方法，其特征在于，所述β i,j,l的取值范围为[0,1]。
  12. 一种计算机装置，其特征在于，包括：
    输入模块，用于将输入参数输入至神经网络模型的输入层，所述输入参数包括通信系统中多个天线中的每个天线对于多个用户中的每个用户的信道向量特征；
    特征提取模块，用于根据所述输入参数，提取所述通信系统的特征；
    拟合模块，用于基于所述神经网络模型，通过多次迭代对所述特征进行拟合，并将拟合结果传递到所述神经网络模型的输出层；
    输出模块，用于从所述输出层获取为每个所述用户分配的发送功率，其中，所述发送功率根据所述拟合结果确定。
  13. 如权利要求12所述的装置，其特征在于，所述输出层的激活函数为Maxout函数。
  14. 如权利要求12所述的装置，其特征在于，所述输入层与所述输出层之间包括所述神经网络模型的多个隐含层。
  15. 如权利要求14所述的装置，其特征在于，所述激活函数通过以下公式表示：
    f Maxout＝Sigmoid(x Tw i+b i)；
    其中，f Maxout表示所述激活函数的运算结果，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数，w i表示第i个所述输出值的权重，b i表示第i个所述输出值的偏差，i＝1、2……、X，x T表示x的转置，Sigmoid()表示Sigmoid函数。
  16. 如权利要求14所述的装置，其特征在于，所述激活函数通过以下公式表示：
    y＝min(max(x,0),P max)；
    其中，y表示所述激活函数的运算结果，P max为预设值，x表示与所述输出层相邻的隐含层的输出值集合，所述x包含X个输出值，X为正整数。
  17. 如权利要求14所述的装置，其特征在于，所述输入层和所述多个隐含层的激活函数均为线性整流函数。
  18. 如权利要求12-17任一所述的装置，其特征在于，
    所述神经网络模型的损失函数包含第一惩罚项和/或第二惩罚项，所述损失函数用于所述神经网络模型的离线训练；
    所述第一惩罚项用于约束所述发送功率大于目标功率值；
    所述第二惩罚项用于约束为每个用户分配的发送速率不小于目标发送速率的最小值。
  19. 如权利要求18所述的装置，其特征在于，
    所述第一惩罚项的系数的取值为[0,1]；和/或，
    所述第二惩罚项的系数的取值为[0,1]。
  20. 如权利要求18或19所述的装置，其特征在于，所述神经网络模型的损失函数通过以下公式表示：
    Figure PCTCN2019079957-appb-100006
    其中，L表示所述损失函数的运算结果，N表示所述神经网络模型的训练样本的数量，M表示所述天线的数量，K表示所述用户的数量，R sum表示为每个用户分配的发送速率的总和，β i,j,l为功率分配因子，i＝1、2、……、N，j＝1、2、……、N，l＝1、2、……、N，R min表示所述发送速率的最小值，R m,k表示第m个天线为第k个用户分配的发送速率，p m表示第m个天线的发送功率，τ表示所述第一惩罚项的系数，ρ表示所述第二惩罚项的系数，
    Figure PCTCN2019079957-appb-100007
    表示所述第一惩罚项，
    Figure PCTCN2019079957-appb-100008
    表示所述第二惩罚项。
  21. 如权利要求20所述的装置，其特征在于，所述R sum通过以下公式表示：
    Figure PCTCN2019079957-appb-100009
    其中，δ 2为加性高斯白噪声的方差，
    Figure PCTCN2019079957-appb-100010
    为信干燥比。
  22. 如权利要求20或21所述的装置，其特征在于，所述β i,j,l的取值范围为[0,1]。
PCT/CN2019/079957 2019-03-27 2019-03-27 一种基于神经网络的功率分配方法及装置 WO2020191686A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2019/079957 WO2020191686A1 (zh) 2019-03-27 2019-03-27 一种基于神经网络的功率分配方法及装置
CN201980094522.3A CN113615277B (zh) 2019-03-27 2019-03-27 一种基于神经网络的功率分配方法及装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/079957 WO2020191686A1 (zh) 2019-03-27 2019-03-27 一种基于神经网络的功率分配方法及装置

Publications (1)

Publication Number Publication Date
WO2020191686A1 true WO2020191686A1 (zh) 2020-10-01

Family

ID=72608738

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/079957 WO2020191686A1 (zh) 2019-03-27 2019-03-27 一种基于神经网络的功率分配方法及装置

Country Status (2)

Country Link
CN (1) CN113615277B (zh)
WO (1) WO2020191686A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113919253A (zh) * 2021-10-08 2022-01-11 西安电子科技大学 硅通孔阵列峰值温度和参数的优化方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8595162B2 (en) * 2011-08-22 2013-11-26 King Fahd University Of Petroleum And Minerals Robust controller for nonlinear MIMO systems
CN108462517A (zh) * 2018-03-06 2018-08-28 东南大学 一种基于机器学习的mimo链路自适应传输方法
WO2018193360A1 (en) * 2017-04-17 2018-10-25 Cerebras Systems Inc. Task synchronization for accelerated deep learning
CN109474980A (zh) * 2018-12-14 2019-03-15 北京科技大学 一种基于深度增强学习的无线网络资源分配方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8595162B2 (en) * 2011-08-22 2013-11-26 King Fahd University Of Petroleum And Minerals Robust controller for nonlinear MIMO systems
WO2018193360A1 (en) * 2017-04-17 2018-10-25 Cerebras Systems Inc. Task synchronization for accelerated deep learning
CN108462517A (zh) * 2018-03-06 2018-08-28 东南大学 一种基于机器学习的mimo链路自适应传输方法
CN109474980A (zh) * 2018-12-14 2019-03-15 北京科技大学 一种基于深度增强学习的无线网络资源分配方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113919253A (zh) * 2021-10-08 2022-01-11 西安电子科技大学 硅通孔阵列峰值温度和参数的优化方法及系统
CN113919253B (zh) * 2021-10-08 2023-08-11 西安电子科技大学 硅通孔阵列峰值温度和参数的优化方法及系统

Also Published As

Publication number Publication date
CN113615277A (zh) 2021-11-05
CN113615277B (zh) 2023-03-24

Similar Documents

Publication Publication Date Title
CN113950066B (zh) 移动边缘环境下单服务器部分计算卸载方法、系统、设备
WO2021259090A1 (zh) 联邦学习的方法、装置和芯片
Agrawal et al. Spider monkey optimization: a survey
Yang et al. Energy-efficient processing and robust wireless cooperative transmission for edge inference
Bi et al. Distributed extreme learning machine with kernels based on mapreduce
CN108416436A (zh) 使用多核心处理模块进行神经网络划分的方法及其系统
CN113902021A (zh) 一种高能效的聚类联邦边缘学习策略生成方法和装置
Li et al. Task offloading scheme based on improved contract net protocol and beetle antennae search algorithm in fog computing networks
CN116187483A (zh) 模型训练方法、装置、设备、介质和程序产品
CN116187430A (zh) 一种联邦学习方法及相关装置
WO2020191686A1 (zh) 一种基于神经网络的功率分配方法及装置
CN116468114A (zh) 一种联邦学习方法及相关装置
Lee et al. Data distribution-aware online client selection algorithm for federated learning in heterogeneous networks
KR20220009682A (ko) 분산 기계 학습 방법 및 시스템
Hu et al. Dynamic task offloading in MEC-enabled IoT networks: A hybrid DDPG-D3QN approach
Yan et al. Deep reinforcement learning based offloading for mobile edge computing with general task graph
Chen et al. Optimal sequential relay-remote selection and computation offloading in mobile edge computing
Tong et al. Multi-type task offloading for wireless Internet of Things by federated deep reinforcement learning
Cui et al. Multi-Agent Reinforcement Learning Based Cooperative Multitype Task Offloading Strategy for Internet of Vehicles in B5G/6G Network
US20210150371A1 (en) Automatic multi-objective hardware optimization for processing of deep learning networks
Ning et al. A multi-objective decomposition-based ant colony optimisation algorithm with negative pheromone
Lu et al. A drl-based decentralized computation offloading method: An example of an intelligent manufacturing scenario
CN109937410B (zh) 核心调度方法和终端
Wang Multimodal emotion recognition algorithm based on edge network emotion element compensation and data fusion
WO2020199914A1 (zh) 训练神经网络的方法和装置

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: 19921977

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: 19921977

Country of ref document: EP

Kind code of ref document: A1