Machine learning-based channel state information transmission technique

Abstract

Various embodiments for a machine learning-based channel state information transmission technique are disclosed. In one embodiment, a method by which at least a first communication device transmits channel state information to a second communication device may comprise the steps of: inputting, into an encoder neural network, first channel state information including channel state information for a wireless link from the second communication device to the first communication device, so as to generate one or more latent vectors; quantizing each of the magnitude and direction of the corresponding latent vector so as to generate magnitude information and direction information for each of the one or more latent vectors; and transmitting, to the second communication device, second channel state information including magnitude information and direction information corresponding to each of the one or more latent vectors.

Description

Machine learning-based channel state information transmission technology

[0001] The present disclosure relates to a channel state information transmission technique, and more particularly, to some embodiments of a machine learning-based channel state information transmission technique.

[0002] In wireless communication systems, channel state information feedback is used in various techniques such as adaptive modulation and coding, MIMO precoding, and beamforming at the transmitter.

[0003] In order for these techniques to achieve excellent performance or increase the transmission efficiency of the entire system, it may be desirable for the information included in the channel state information feedback to be accurate and for the overhead of the channel state information feedback to be small.

[0004] For example, to achieve the performance benefits of massive MIMO, base stations require accurate information about channel state information. However, massive MIMO systems have a large channel state information dimension, which can lead to limitations such as requiring a significant amount of feedback overhead.

[0005] Therefore, an efficient channel state information transmission technique may be required.

[0006] One aspect of the present disclosure provides a method for a first communication device to transmit channel state information to a second communication device, the method comprising: inputting first channel state information, including channel state information for a wireless link from the second communication device to the first communication device, into an encoder neural network to generate at least one latent vector; quantizing a magnitude and a direction of each of the at least one latent vectors to generate magnitude information and direction information; and transmitting second channel state information, including magnitude information and direction information corresponding to each of the at least one latent vectors, to the second communication device.

[0007] In some embodiments, the step of generating the direction information may include, for each of the at least one latent vectors, selecting a codebook vector closest to the latent vector from among a plurality of codebook vectors included in a pre-provided codebook, and generating direction information of the latent vector based on the selected codebook vector.

[0008] In some embodiments, the direction information may include an identifier of the selected codebook vector.

[0009] In some embodiments, the plurality of codebook vectors may include unit norm vectors. In some embodiments, the step of generating direction information may include the steps of: normalizing each of the at least one latent vectors to become a vector with a size of 1; and, for each of the normalized latent vectors, selecting a codebook vector closest to the normalized latent vector among the plurality of codebook vectors, and generating direction information of the latent vector based on the selected codebook vector.

[0010] In some embodiments, the step of generating the magnitude information may include the step of quantizing the magnitude of each of the at least one latent vector using a μ-law quantizer.

[0011] In some embodiments, the final activation function of the encoder neural network may include Tanh.

[0012] In some embodiments, the second communication device may include a decoder neural network that obtains a reconstruction of the first channel state information based on the second channel state information. In some embodiments, the encoder neural network and the decoder neural network may be pre-trained to minimize an objective function.

[0013] In some embodiments, the step of generating the direction information may include, for each of the at least one latent vectors, selecting a codebook vector closest to the latent vector from among a plurality of codebook vectors included in a pre-provided codebook, and generating the direction information of the latent vector based on the selected codebook vector. In some embodiments, the first channel state information may include a channel matrix for the wireless link. In some embodiments, the objective function is a mathematical formula can be defined as . In the above mathematical formula, is the channel matrix, and is the restoration for the channel matrix, and z is at least one latent vector, and z qis a quantization result for at least one latent vector, sg is a stop-gradient operator, and β is a preset coefficient. In some embodiments, the encoder neural network, the decoder neural network, and the codebook may be pre-trained so that the objective function is minimized.

[0014] In some embodiments, the first communication device may include a plurality of transmit antennas, the second communication device may include at least one receive antenna, and the channel may be an OFDM-based channel. In some embodiments, the method may further include the steps of: generating a space- and frequency-based channel matrix through an estimation of the channel; and generating an angle- and delay-based channel matrix by transforming the space- and frequency-based channel matrix. In some embodiments, the step of generating at least one latent vector may include the step of generating the at least one latent vector based on the generated angle- and delay-based channel matrix.

[0015] In some embodiments, the step of generating the at least one latent vector may include the step of truncating the angle and delay-based channel matrix; and the step of inputting the truncated angle and delay channel matrix into the encoder neural network as the first channel state information.

[0016] In some embodiments, the first communication device may comprise a base station of a mobile communication system, and the second communication device may comprise a user device of the mobile communication system.

[0017] Another aspect of the present disclosure provides a method for a second communication device to receive channel state information from a first communication device, the method comprising: receiving second channel state information from the first communication device, the second channel state information including magnitude information and direction information corresponding to each of at least one latent vector; and inputting information based on the second channel state information into a decoder neural network to obtain a reconstruction of the first channel state information, the first channel state information including channel state information for a wireless link from the second communication device to the first communication device.

[0018] In some embodiments, the at least one latent vector may be generated by the first communication device inputting the first channel state information into an encoder neural network.

[0019] In some embodiments, the direction information may be generated by selecting a codebook vector closest to the potential vector from among a plurality of codebook vectors included in a codebook pre-equipped by the first communication device, and based on the selected codebook vector.

[0020] In some embodiments, the direction information may include an identifier of the selected codebook vector. In some embodiments, the step of inputting information based on the second channel state information into the decoder neural network may include the step of selecting one of a plurality of codebook vectors included in a codebook pre-equipped in the second communication device as a direction vector of the corresponding latent vector, based on an identifier included in the direction information of each of the at least one latent vector; and the step of inputting the selected direction vector into the decoder neural network.

[0021] In some embodiments, the encoder neural network and the decoder neural network may be pre-trained to minimize the objective function.

[0022] In some embodiments, the direction information may be generated by selecting a codebook vector closest to the potential vector among a plurality of codebook vectors included in a codebook pre-equipped by the first communication device and based on the selected codebook vector. In some embodiments, the first channel state information may include a channel matrix for the wireless link. In some embodiments, the objective function is a mathematical formula can be defined as . In the above mathematical formula, is the channel matrix, and is the restoration for the channel matrix, and z is at least one latent vector, and z q is a quantization result for at least one latent vector, sg is a stop-gradient operator, and β is a preset coefficient. In some embodiments, the encoder neural network, the decoder neural network, and the codebook may be pre-trained so that the objective function is minimized.

[0023] In some embodiments, the first communication device may comprise a base station of a mobile communication system, and the second communication device may comprise a user device of the mobile communication system.

[0024] Another aspect of the present disclosure provides a method for training a machine learning model, comprising at least one processor, an encoder neural network and a decoder neural network, the method comprising: inputting first channel state information, including channel state information for a wireless link from the second communication device to the first communication device, into the encoder neural network to generate at least one latent vector; quantizing a magnitude and a direction of each of the at least one latent vectors to generate magnitude information and direction information; and inputting information based on second channel state information, including magnitude information and direction information corresponding to each of the at least one latent vectors, into the decoder neural network to obtain a reconstruction of the first channel state information; and updating parameters of the encoder neural network and parameters of the decoder neural network such that an objective function is minimized.

[0025] In some embodiments, the step of generating the direction information may include, for each of the at least one latent vectors, selecting a codebook vector closest to the latent vector from among a plurality of codebook vectors included in the codebook, and generating the direction information of the latent vector based on the selected codebook vector. In some embodiments, the step of updating may include updating the parameters of the encoder neural network, the parameters of the decoder neural network, and the codebook.

[0026] In some embodiments, the objective function is a mathematical formula can be defined as . In the above mathematical formula, is the channel matrix, and is the restoration for the channel matrix, and z is at least one latent vector, and z qis a quantization result for at least one latent vector, sg is a stop-gradient operator, and β is a preset coefficient.

[0027] In some embodiments, the method may further include a step of normalizing a plurality of codebook vectors included in the codebook after the codebook is updated.

[0028] In some embodiments, the codebook may be configured in a nested codebook structure.

[0029] In some embodiments, the updating step may include using soft gradient passing during backpropagation of the direction information.

[0030] In some embodiments, the first communication device may comprise a base station of a mobile communication system, and the second communication device may comprise a user device of the mobile communication system.

[0031] Another aspect of the present disclosure provides a communication device, comprising: a processor; one or more hardware-based transceivers; and a computer-readable storage medium comprising instructions, wherein the instructions, in response to execution by the processor, cause the communication device to perform at least one of the embodiments of the method of the present disclosure.

[0032] Another aspect of the present disclosure provides a non-transitory recording medium storing instructions readable by a processor of an electronic device, the instructions causing the processor to perform embodiments of the present disclosure.

[0033] This Summary is provided to introduce a selection of concepts in a simplified form that are further explained in the Detailed Description below. This Summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. Furthermore, the claimed subject matter is not limited to implementations that solve any or all of the problems mentioned in any portion of this specification. In addition to the exemplary aspects, embodiments, and features described above, additional aspects, embodiments, and features will become apparent by reference to the following Detailed Description and Drawings.

[0034] Some embodiments of the present disclosure may have the following advantages. However, this does not mean that all embodiments must include all of the advantages, and the scope of the present invention should not be construed as being limited thereby.

[0035] According to some embodiments, accurate channel state information can be provided even with little feedback overhead.

[0036] According to some embodiments, accurate channel state information can be provided even with low computational complexity.

[0037] FIG. 1 illustrates a system for describing some embodiments of channel state information feedback of the present disclosure.

[0038] FIG. 2 is a flowchart illustrating some embodiments of channel state information feedback of the present disclosure.

[0039] FIG. 3 is a block diagram illustrating a quantization process according to some embodiments.

[0040] Figure 4 is a conceptual diagram illustrating some embodiments of size quantization.

[0041] Figure 5 is a table showing the performance of transformation techniques used for size quantization.

[0042] Figure 6 illustrates the NMSE performance of each CSI feedback technique according to feedback overhead.

[0043] Figure 7 illustrates the NMSE performance of each CSI feedback technique at a given complexity.

[0044] Figure 8 illustrates the performance of a CIS feedback technique using overlapping codebooks and a CSI feedback technique using separate codebooks.

[0045] FIG. 9 is a block diagram illustrating the internal configuration of a computing system (or computing device) according to one embodiment of the present disclosure.

[0046] The description of the present invention is merely an example for structural and functional explanation, and therefore, the scope of the present invention should not be construed as being limited by the embodiments described in the text. That is, since the embodiments can be modified in various ways and can take various forms, the scope of the present invention should be understood to include equivalents that can realize the technical idea. In addition, the purposes or effects presented in the present invention do not mean that a specific embodiment must include all of them or only such effects, and therefore, the scope of the present invention should not be construed as being limited thereby.

[0047] Meanwhile, the meanings of the terms described in this disclosure should be understood as follows.

[0048] Terms such as "first" and "second" are intended to distinguish one component from another, and the scope of the rights should not be limited by these terms. For example, the first component may be referred to as the second component, and similarly, the second component may also be referred to as the first component.

[0049] When a component is said to be "connected" to another component, it should be understood that while it may be directly connected to that other component, there may also be other components intervening. Conversely, when a component is said to be "directly connected" to another component, it should be understood that there are no other intervening components. Similarly, other expressions describing relationships between components, such as "between" and "directly between," or "adjacent to" and "directly adjacent to," should be interpreted similarly.

[0050] Singular expressions should be understood to include plural expressions unless the context clearly indicates otherwise, and terms such as "comprises" or "have" should be understood to specify the presence of a feature, number, step, operation, component, part or combination thereof, but not to exclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts or combinations thereof.

[0051] For each step, the identifiers (e.g., a, b, c, etc.) are used for convenience of explanation and do not describe the order of the steps. The steps may occur in a different order than stated unless the context clearly dictates a specific order. That is, the steps may occur in the same order as stated, may be performed substantially simultaneously, or may be performed in the opposite order.

[0052] FIG. 1 illustrates a system for describing some embodiments of channel state information feedback of the present disclosure.

[0053] In some embodiments, the system (100) may include a first communication device (110) and a second communication device (120), as illustrated in FIG. 1. In some other embodiments, the system (100) may include a plurality of first communication devices (110) and second communication devices (120). In some other embodiments, the system (100) may include a plurality of first communication devices (110) and a plurality of second communication devices (120).

[0054] In some embodiments, as illustrated in FIG. 1, the second communication device (120) comprises at least one (e.g., a plurality of N as illustrated in FIG. 1) t ) and the first communication device (120) includes at least one (e.g., N having a value of 1 as illustrated in FIG. 1) transmitting antenna. r ) may include a receiving antenna. In some other embodiments, the first communication device (120) may include a plurality of receiving antennas.

[0055] In some embodiments, the system (100) may be a cellular MIMO system that includes each first communication device (110) and each second communication device (120) as user equipment and a base station, respectively. In some other embodiments, the system (100) may be a MIMO system other than a cellular MIMO system. For example, the system (100) may be a Wi-Fi communication system that includes each first communication device (110) and each second communication device (120) as a wireless station and a wireless access point, respectively.

[0056] In some embodiments, each first communication device (110) and each second communication device (120) may communicate with each other using an OFDM modulation technique. In some other embodiments, each first communication device (110) and each second communication device (120) may use a modulation technique other than the OFDM modulation technique (e.g., CDMA).

[0057] In some embodiments, as illustrated in FIG. 1, the first communication device (110) may obtain channel state information of a channel (e.g., a wireless link from the second communication device to the first communication device, as illustrated in FIG. 1) and transmit the obtained channel state information to the second communication device (120). In some embodiments, the channel state information may include information about each channel component of the wireless link. For example, in the case of an OFDM-based system, as illustrated in FIG. 1, each channel component (h) may be specified by a path connecting each transmit antenna and each receive antenna (a single receive antenna in the case of FIG. 1) and each subcarrier (the nth subcarrier in FIG. 1).

[0058] Below, for convenience, N t (N t A base station having a number of transmitting antennas (i.e., N is a natural number greater than or equal to 2) and a single receiving antenna (i.e., N r =1) includes a plurality of user equipments (User Equipment), N c Embodiments of the present disclosure will be described primarily with reference to a single-cell MIMO system using OFDM including multiple subcarriers. Those skilled in the art will appreciate that this description will allow them to readily apply, implement, or reproduce the techniques of the present disclosure to other systems.

[0059] FIG. 2 is a flowchart illustrating some embodiments of channel state information feedback of the present disclosure.

[0060] In some embodiments, each block illustrated in FIG. 2 may be performed by the first communication device (110), the second communication device (120), the system (100), and / or equipment of a person operating the system (100) illustrated in FIG. 1.

[0061] In some embodiments, as illustrated in FIG. 1, a machine learning model (e.g., a Deep Learning Model) may be trained (S210). For example, a processor included in the first communication device (110), the second communication device (120), the system (100), or other devices may train the machine learning model.

[0062] In some embodiments, the machine learning model may include an encoder neural network and a decoder neural network. In other embodiments, the machine learning model may include an encoder neural network, a decoder neural network, and a codebook, as described below.

[0063] In some embodiments, in S210, the processor may input first channel state information into an encoder neural network to generate at least one latent vector (S212), perform quantization on the at least one latent vector to generate second channel state information (S214), input information based on the second channel state information into a decoder neural network to obtain a reconstruction of the first channel state information (S216), and train a machine learning model such that an objective function is minimized (S218).

[0064] In some embodiments, the first channel state information may include channel state information for a wireless link from the second communication device (120) to the first communication device (110). In some embodiments, the first channel state information may be secured in advance as a training data item in S210 (i.e., unlike S220 described below) and then used in the training of the present disclosure.

[0065] In some embodiments, the processor may generate magnitude information and direction information by quantizing the magnitude and direction of each of at least one latent vectors in S214, respectively.

[0066] In some embodiments, the second channel state information may include magnitude information and direction information corresponding to each of the at least one latent vectors.

[0067] In some embodiments, the processor may update the parameters of the encoder neural network and the parameters of the decoder neural network such that the objective function is minimized in S218.

[0068] In some embodiments, the processor may select, for each of the at least one latent vectors in S214, a codebook vector closest to the latent vector from among a plurality of codebook vectors included in the codebook, generate direction information of the latent vector based on the selected codebook vector, and update the parameters of the encoder neural network, the parameters of the decoder neural network, and the codebook so that the objective function is minimized in S218.

[0069] In some embodiments, in S210, the processor may perform an operation of normalizing a plurality of codebook vectors included in the codebook after the codebook is updated.

[0070] In some embodiments, the codebook may be structured as a nested codebook structure.

[0071] In some embodiments, the processor may use soft gradient passing when backpropagating the directional information in S218.

[0072] In some embodiments, as illustrated in FIG. 2, the first communication device (110) may acquire first channel status information (S220). For example, the first communication device (110) may acquire first channel status information based on a signal (e.g., a pilot signal) transmitted from the second communication device (120).

[0073] In some embodiments, the first communication device (110) may use a space- and frequency-based channel matrix as the first channel state information through channel estimation. In some other embodiments, the first communication device (110) may transform the space- and frequency-based channel matrix to generate an angle- and delay-based channel matrix, and use the generated angle- and delay-based channel matrix as the first channel state information. In some other embodiments, the first communication device (110) may truncate the generated angle- and delay-based channel matrix, and use the truncated angle- and delay-based channel matrix as the first channel state information.

[0074] In some embodiments, as illustrated in FIG. 2, the first communication device (110) may input first channel state information into an encoder neural network (e.g., the encoder neural network trained in S210) to generate at least one latent vector (S230).

[0075] In some embodiments, as illustrated in FIG. 2, the first communication device (110) may generate magnitude information and direction information by quantizing the magnitude and direction of each of at least one potential vectors (S240).

[0076] In some embodiments, as illustrated in FIG. 2, the first communication device (110) may transmit second channel state information including size information and direction information corresponding to at least one potential vector to the second communication device (S250).

[0077] In some embodiments, the first communication device (110) may, for each of at least one latent vectors in S240, select a codebook vector closest to the latent vector from among a plurality of codebook vectors included in a pre-provided codebook (e.g., a codebook trained in S210), and generate direction information of the latent vector based on the selected codebook vector. In some embodiments, the direction information of the latent vector may include an identifier of the selected codebook vector.

[0078] In some embodiments, the plurality of codebook vectors may include unit norm vectors. In some embodiments, the first communication device (110) may normalize each of the at least one latent vectors to a vector having a size of 1 in S240, select a codebook vector closest to the normalized latent vector among the plurality of codebook vectors for each of the normalized latent vectors, and generate direction information of the latent vector based on the selected codebook vector.

[0079] In some embodiments, the first communication device (110) may quantize the magnitude of each of at least one latent vector using a μ-law quantizer at S240.

[0080] In some embodiments, the final activation function of the encoder neural network may include Tanh.

[0081] In some embodiments, as illustrated in FIG. 2, the second communication device (120) may receive second channel state information from the first communication device (110) including size information and direction information corresponding to at least one potential vector (S260).

[0082] In some embodiments, as illustrated in FIG. 2, the second communication device (120) may input information based on the second channel state information into a decoder neural network (e.g., the decoder neural network trained in S210) to obtain a restoration of the first channel state information (S270).

[0083] In some embodiments, the second communication device (120) may select one of a plurality of codebook vectors included in a codebook (e.g., a codebook trained in S210) pre-equipped in the second communication device as a direction vector of the corresponding latent vector based on an identifier included in the direction information of each of at least one potential vectors in S270, and input the selected direction vector into the decoder neural network.

[0084] In some embodiments, the channel matrix H sf H in the spatial and frequency domains sf ∈ can be formed. Here, is size N c × N t It means a complex matrix.

[0085] In some embodiments, such a channel matrix H sf This disclosure can be directly input into machine learning and used in a channel state information (CSI) feedback scheme.

[0086] In some other embodiments, the channel matrix H sfChannel matrices of different types (e.g., angle- and delay-based matrices) can be directly input into the machine learning of the present disclosure and used in the CSI feedback scheme. For example, in massive-MIMO, when the number of scatter clusters is much smaller than the number of base station transmit antennas, the channel matrix H sf can be transformed into a sparse matrix in the angular-delay domain.

[0087] In some embodiments, the channel matrix H sf Silver, as shown in the mathematical expression 1 below, is a two-dimensional discrete Fourier transform (DFT), and the channel matrix H of the angular delay region is ad can be converted to

[0088]

[0089] Here, F d and F a are each of size N c × N c and N t × N t is a DFT matrix with .

[0090] In some embodiments, the channel matrix H ad All components included in can be directly input into the machine learning of the present disclosure and used in the CSI feedback method.

[0091] In some other embodiments, the channel matrix H ad Only some of the components included in can be directly input into the machine learning of the present disclosure and used in the CSI feedback scheme. For example, since multipath delay may exist only in a limited time interval, in this case, the channel matrix H in the angular delay domain ad Among the ingredients, from the front ( <N c) only the delayed components can contain significant or meaningful values. In this case, H ad By applying truncation to H, ad of Truncated channel matrix representing the top rows of the dog This can be generated. This truncated channel matrix provides real input values ​​to the network model. can be expressed as a matrix.

[0092] In some embodiments, this truncated channel matrix , may be compressed and quantized for transmission from the user device.

[0093] In some embodiments, this truncated channel matrix is the entire potential vector can be compressed into . This entire latent vector can be expressed by mathematical expression 2, and at least one latent vector (e.g., In this case, it can contain N latent vectors, each latent vector having dimension D.

[0094]

[0095] Here, represents the encoder neural network.

[0096] In some embodiments, the entire latent vector z is quantized, so that z is located in the embedding space. q can be created.

[0097] In some embodiments, the base station receives a quantized vector z from the user device. q If you receive feedback, z q Based on the channel matrix (e.g., ) for reconstruction can be derived.

[0098] In some embodiments, the process of restoring the channel matrix from the quantized vector can be expressed by mathematical expression 3.

[0099]

[0100] In some embodiments, the system may perform generation and quantization of latent vectors corresponding to a channel matrix through a machine learning model; and restoration of the channel matrix. For example, the system may perform a paper using a machine learning model. <A. van den Oord, O. Vinyals, and K. Kavukcuoglu, "Neural discrete representation learning," Adv. Neural Inf. Process. Syst., Dec. 2017, pp.6306-6315.> The VQ-VAE (Vector Quantized Variational Autoencoder) technique used for image data compression transmission can be applied to channel state information transmission.

[0101] In some embodiments, the system comprises an encoder and a decoder of general autoencoder models and a quantization module positioned between them (i.e., the encoder and the decoder), wherein the quantization module can use a codebook. In some embodiments, the codebook can be trained.

[0102] For convenience in this specification, the codebook is called Expressed as a codebook There are K D-dimensional codebook vectors (i.e., ) to describe some embodiments of the present disclosure.

[0103] In some embodiments, the quantization module generates N latent vectors (i.e., where the dimension of the entire latent vector is N×D). ) may need to perform quantization within a single codebook. Here, is [z (i-1)D+1 , … , z iD ] can be expressed as.

[0104] In some embodiments, each of the latent vectors may be quantized to its closest vector among the vectors in the codebook, as in Equation 4.

[0105]

[0106] In mathematical equation 4, is the latent vector represents the quantized result vector.

[0107] In some embodiments, the objective function of the machine learning model can be expressed by Equation 5.

[0108]

[0109] In Equation 5, sg(·) represents a stop-gradient operator that ignores the gradient descent computation as a constant.

[0110] In Equation 5, the first and second terms represent the reconstruction loss and codebook loss, respectively, and the third term represents the commitment loss, which is regularized by a coefficient β set to a value between 0.1 and 2.0.

[0111] In some embodiments, after the quantization errors corresponding to the second and third terms of Equation 5 are calculated, gradient correction for the decoder input is performed. q ←z+ sg(z q -z) can be used to ensure that the reconstruction loss term is not affected by quantization errors.

[0112] In some embodiments, by configuring the loss structure as above, the system can heuristically update the codebook, encoder, and decoder.

[0113] In some embodiments, the system may perform quantization by separating the spatial characteristics of the latent vector in size and direction when performing exhaustive quantization search of VQ-VAE in CSI feedback through VQ-VAE as described above. For example, the system may perform quantization by separating the spatial characteristics of the latent vector in size and direction as in Equation 6. i ) for size quantization (Q) mag ) and direction quantization (Q dir ) to perform the quantized latent vector (z q,i ) can be created.

[0114]

[0115] In mathematical expression 6 am.

[0116] For example, each latent vector (z i ) The size information and direction information obtained through size quantization and direction quantization are respectively B mag Bit and B dir When composed of bits, the quantized latent vector (z q,i ) is the number of bits B used to transmit B mag +B dir Here, B mag and B dir can have a preset natural number value.

[0117] FIG. 3 is a block diagram illustrating a quantization process according to some embodiments.

[0118] In some embodiments, the system may be configured to use a function f of Equation 7 quan The magnitude of each latent vector can be quantized using a uniform quantizer represented by (x).

[0119]

[0120] In some embodiments, the system may use Tanh as the final activation function of the encoder. In this case, there are N latent vectors z i Each one is can have a limited size.

[0121] In some embodiments, if the magnitude of each latent vector does not have a value in the range [0, 1], the system may perform preprocessing on each latent vector so that the magnitude of each latent vector has a value in the range [0, 1] before applying uniform quantization as illustrated in Equation 7.

[0122] For example, the system considers each potential vector as Normalization can be performed by dividing by .

[0123] As another example, the system can quantize the magnitude of each latent vector using the transform function exemplified by Equations 8 to 10. For example, the system can perform quantization by applying the transform function to the magnitude of each latent vector and then applying a uniform quantizer. For example, when Equations 8 and 9 are used as the transform functions, the magnitude quantization is performed by and can be expressed as

[0124] The conversion function according to one embodiment is defined by mathematical expression 8. It can be a cumulative distribution function, and this cumulative distribution function can be obtained through training data.

[0125]

[0126] According to another embodiment, the transformation function may be a μ-law transformation function defined by Equation 9. For example, a μ-law quantizer can be effectively applied to a system in which the magnitude of the latent vector tends to be concentrated on values ​​close to 0. In Equation 9, μ may be a preset positive constant.

[0127]

[0128] According to another embodiment, the transformation function may be a clipped μ-law transformation function defined by Equation 10. In Equation 10, A may be a preset positive constant.

[0129]

[0130] The magnitude value of each latent vector that has passed through mathematical expression 10 can have the range [0, A].

[0131] In some embodiments, when the system uses a clipped μ-law transform function, a uniform quantizer expressed by Equation 11 may be used instead of Equation 7.

[0132]

[0133] In some embodiments, the system may utilize soft gradient passing. For example, soft gradient passing may be applied to systems where the round function itself suffers from a zero gradient problem (e.g., systems using Equation 7 or Equation 11). In some embodiments, the system may replace the round function with a step-wise tanh function only when performing backpropagation. can be achieved. For example, in the case of mathematical expression 11, can be expressed by mathematical formula 12.

[0134]

[0135] Figure 4 is a conceptual diagram illustrating some embodiments of size quantization.

[0136] In Fig. 4, The identity function, μ-law quantization function, and curve according to soft gradient passing are shown. According to the embodiment of Fig. 4, the quantization result is 3 bits (i.e., Bmag =3) can be expressed as

[0137] In some embodiments, the system may use a trainable codebook as a directional quantizer. In some embodiments, the system may train the codebook by performing the following operations: initializing the codebook so that the codebook satisfies certain conditions; and updating the codebook using stochastic gradient descent of a loss function.

[0138] In some embodiments, the system assumes that the direction of each latent vector is uniformly distributed, and the codebook vector (i.e., ) can be initialized as a Grassmannian codebook that optimizes Equation 13 in a D-dimensional real space.

[0139]

[0140] In some embodiments, the codebook may be composed of unit-norm vectors. In some embodiments, the system may normalize each codebook vector after updating the codebook so that each codebook vector becomes a unit-norm vector. For example, since the vectors in the codebook may not have a unit norm after the gradient descent update, the system may perform normalization on the codebook vectors after the codebook update. In some embodiments, this normalization procedure may not be included in the gradient calculation.

[0141] In some embodiments, the system may use Equation 14 as the direction quantization rule.

[0142]

[0143] In Equation 14, d() represents the distance between two norm vectors (c1, c2). For example, d() can be measured as the sine of the angle between the two norm vectors and can be defined as Equation 15.

[0144]

[0145] In some embodiments, the system can perform gradient correction using Equation 16 after calculating the quantization error, similar to VQ-VAE. For example, this can prevent the reconstruction loss term from being affected by the quantization error.

[0146]

[0147] In some embodiments, the system may extend some of the embodiments described above to a multi-rate codebook design.

[0148] In some embodiments, the system may be sized by the user depending on the situation. The codebook can be constructed in a nested manner, which is formalized as follows: (j=1, 2, ...,L), is the number of bits representing the codebook vector of the j-th codebook.

[0149] In some embodiments, the system may train L overlapping codebooks step by step. In some embodiments, the system may train the first codebook can be initialized and trained in the same manner as described above. In some embodiments, the system may be configured to have K j - K j-1 The j-1th codebook of dog vectors Add randomly to the j(>1)th codebook can be initialized and trained according to the loss function of Equation 17.

[0150]

[0151] In mathematical expression 17, is a codebook is the entire latent vector quantized into , And, is a hyperparameter that controls the loss portion of each quantization order in the nested codebook.

[0152] We will compare the simulation results by referring to some embodiments that use size quantization and direction quantization as Shape-Gain technique and the compared technique as original VQ-VAE technique. Here, Shape-Gain technique refers to CSI feedback technique used for direction quantization and size quantization, respectively, and original VQ-VAE technique is described in the above-mentioned paper. <A. van den Oord, O. Vinyals, and K. Kavukcuoglu, "Neural discrete representation learning," Adv. Neural Inf. Process. Syst., Dec. 2017, pp.6306-6315.> It refers to a CSI feedback technique based on .

[0153] The channel data set in the simulation was generated from COST2100, an outdoor semi-urban scenario at 300 MHz. This data set was applied to several CSI models. The base station is located in the space-frequency domain. t = 32 and N c = 1024 was equipped with a uniform linear array. After conversion to the angle-delay domain, the resulting CSI image By cutting it with You can get a matrix. This The matrix is ​​a 2 x 32 x 32 matrix of real numbers, which can be provided as a real number input. The data set includes a training data set, a validation data set, and a test data set, with sizes of 100,000, 30,000, and 20,000, respectively. In addition, the batch size is set to 200.

[0154] The simulation parameters are A=0.16, B mag =4, D=16, μ=255, β=0.25, and τ=8. The Adam optimizer with a learning rate of 0.001 was used as the optimizer, and the models were trained for 1000 epochs. In particular, the nested codebooks were trained for 200 epochs at each step.

[0155] In analyzing the performance of each technique, the original channel matrix and restored channel matrix The normalized mean squared error (NMSE), exemplified by Equation 18, was used as the difference between the two.

[0156]

[0157] Figure 5 is a table showing the performance of transformation techniques used for size quantization. The performance in the table in Figure 5 represents the NMSE performance when the dimension M of the entire latent vector is 4096.

[0158] Referring to Figure 5, it can be seen that the clipped μ-law transform technique has the best NMSE performance overall, even though the distribution transform technique is optimal in terms of quantization perplexity.

[0159] Figures 6 to 8 below show simulation results when the clipped μ-law transform technique is used for size quantization.

[0160] Figure 6 illustrates the NMSE performance of each CSI feedback technique according to feedback overhead.

[0161] In Fig. 6, the ScalarQ technique represents a CSI feedback technique that does not use vector quantization. Since the ScalarQ technique has relatively large feedback overhead compared to techniques that use VQ (i.e., the original VQ-VAE technique, Shape-Gain technique) when the same latent dimension M is given, in the simulation in Fig. 6, M of the ScalarQ technique is set so that it can have a similar level of feedback overhead.

[0162] Referring to Figure 6, it can be seen that the Shape-Gain technique is the best.

[0163] To compare the complexity of the quantization module in the original VQ-VAE technique and the Shape-Gain technique, the number of multiplications in the quantization process can be considered. B mag and B dir When the number of bits representing the magnitude information and direction information of each latent vector in the Shape-Gain technique is It can be seen that it has the order of . On the other hand, the complexity of the quantization process of the original VQ-VAE technique is It can be seen that it has the order of . Here, B = B mag +B dir am.

[0164] Figure 7 illustrates the NMSE performance of each CSI feedback technique at a given complexity.

[0165] Referring to Figure 7, it can be seen that the Shape-Gain technique has a lower complexity for achieving the same NMSE performance.

[0166] Figure 8 illustrates the performance of a CIS feedback technique using overlapping codebooks and a CSI feedback technique using separate codebooks.

[0167] In the simulation of Fig. 8, M=4096, L=5, { } = {4, 5,..., 8} were used as parameters.

[0168] Referring to Figure 8, it can be seen that techniques using nested codebooks sometimes outperform techniques using separate codebooks. This suggests that the nested codebook training technique effectively positions the directional codebook vectors in the latent space.

[0169] FIG. 9 is a block diagram illustrating the internal configuration of a computing system (or computing device) according to one embodiment of the present disclosure. In FIG. 9, the system (900) is described as a single physical device, but depending on the embodiment, the system (900) may be implemented in a form in which multiple devices are linked together.

[0170] The system (900) may include a memory (910), a processor (920), a communication module (930), and an input / output interface (940) as illustrated in FIG. 9. The memory (910) may be a computer-readable storage medium, and may include a random access memory (RAM), a read only memory (ROM), and a permanent mass storage device such as a disk drive. Here, the ROM and the permanent mass storage device may be separated from the memory (910) and included as a separate permanent storage device. In addition, the memory (910) may store an operating system and at least one program code (for example, a computer program stored on a storage medium included in the system (900) to control the system (900) so as to perform a method according to embodiments of the present disclosure). These software components may be loaded from a computer-readable storage medium separate from the memory (910). Such separate computer-readable recording media may include computer-readable recording media such as floppy drives, disks, tapes, DVD / CD-ROM drives, memory cards, etc. In other embodiments, the software components may be loaded into the memory (910) via a communication module (930) that is not a computer-readable recording medium.

[0171] The processor (920) may be configured to process instructions of a computer program by performing basic arithmetic, logic, and input / output operations. Instructions may be provided to the processor (920) by the memory (910) or the communication module (930). For example, the processor (920) may be configured to execute instructions received according to program code loaded into the memory (910). More specifically, the processor (920) may sequentially execute instructions according to the code of the computer program loaded into the memory (910) to perform channel state information feedback and / or machine learning model training methods according to embodiments of the present disclosure. The communication module (930) may provide a function for communicating with other physical devices through an actual computer network. For example, the present disclosure may be implemented in a manner in which a processor (920) of a system (900) performs a part of the process of the present embodiment, and another physical device (e.g., another computing system not shown) of the network performs the remaining process, while exchanging the processing results with the computer network through a communication module (930).

[0172] The input / output interface (940) may be a means for interfacing with an input / output device (950). For example, the input / output device (950) may include a device such as a keyboard or a mouse, and the output device may include a device such as a display or a speaker. In FIG. 7, the input / output device (950) is represented as a separate device from the system (900), but depending on the embodiment, the system (900) may be implemented such that the input / output device (950) is included in the system (900).

[0173] The devices described above may be implemented as hardware components, software components, and / or a combination of hardware components and software components. For example, the devices and components described in the embodiments may be implemented using one or more general-purpose computers or special-purpose computers, such as a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing instructions and responding to them. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing device is sometimes described as being used singly; however, those skilled in the art will appreciate that the processing device may include multiple processing elements and / or multiple types of processing elements. For example, the processing device may include multiple processors, or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are also possible.

[0174] Software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing device to perform a desired operation or may independently or collectively command the processing device. The software and / or data may be embodied in any type of machine, component, physical device, computer storage medium, or device for interpretation by the processing device or for providing instructions or data to the processing device. The software may be distributed over networked computer systems and stored or executed in a distributed manner. The software and data may be stored on one or more computer-readable recording media.

[0175] The method according to the embodiment may be implemented in the form of program commands that can be executed through various computer means and recorded on a computer-readable medium. At this time, the medium may be one that continuously stores a computer-executable program or one that temporarily stores it for execution or download. In addition, the medium may be various recording means or storage means in the form of a single or multiple hardware combinations, and is not limited to a medium directly connected to a computer system, but may also be distributed on a network. Examples of the medium may include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and those configured to store program commands, including ROM, RAM, and flash memory. In addition, examples of other media may include recording media or storage media managed by app stores that distribute applications, sites that supply or distribute various software, servers, etc.

[0176] Although the embodiments described above have been described by way of limited examples and drawings, those skilled in the art will appreciate that various modifications and variations can be made based on the above teachings. For example, appropriate results can still be achieved even if the described techniques are performed in a different order than described, and / or components of the described systems, structures, devices, circuits, etc. are combined or combined in a different manner than described, or are replaced or substituted with other components or equivalents.

Claims

1. A method in which a first communication device transmits channel status information to a second communication device, A step of inputting first channel state information - including channel state information for a wireless link from said second communication device to said first communication device - into an encoder neural network to generate at least one latent vector; For each of the at least one latent vector, a step of quantizing the magnitude and direction of the latent vector to generate magnitude information and direction information; and A method comprising the step of transmitting second channel state information including magnitude information and direction information corresponding to each of the at least one potential vector to the second communication device.

2. In the first paragraph, the step of generating the direction information is as follows: A method comprising, for each of the at least one latent vector, selecting a codebook vector closest to the latent vector from among a plurality of codebook vectors included in a pre-prepared codebook, and generating direction information of the latent vector based on the selected codebook vector.

3. In the second paragraph, the direction information is, A method comprising an identifier of the above selected codebook vector.

4. In paragraph 2, The above plurality of codebook vectors include unit norm vectors, The step of generating the above direction information is: a step of normalizing each of the at least one latent vectors to become a vector of size 1; and A method comprising the steps of: for each of the above normalized latent vectors, selecting a codebook vector closest to the corresponding normalized latent vector from among a plurality of codebook vectors; and generating direction information of the corresponding latent vector based on the selected codebook vector.

5. In the first paragraph, the step of generating the size information comprises: A method comprising the step of quantizing the magnitude of each of said at least one latent vector using a μ-law quantizer.

6. In paragraph 5, The last activation function of the above encoder neural network includes Tanh.

7. In paragraph 1, The second communication device includes a decoder neural network that obtains reconstruction of the first channel state information based on the second channel state information, A method in which the above encoder neural network and the above decoder neural network are pre-trained so that the objective function is minimized.

8. In paragraph 7, The step of generating the above direction information includes, for each of the at least one potential vector, selecting a codebook vector closest to the corresponding potential vector from among a plurality of codebook vectors included in a pre-prepared codebook, and generating direction information of the corresponding potential vector based on the selected codebook vector. The above first channel state information includes a channel matrix for the wireless link, The above objective function is defined by the following mathematical equation, Above is the channel matrix, and is the restoration for the channel matrix, z is at least one latent vector, and z q is a quantization result for at least one latent vector, sg is a stop-gradient operator, β is a preset coefficient, A method in which the above encoder neural network, the decoder neural network and the codebook are pre-trained so that the objective function is minimized.

9. In paragraph 1, The above first communication device comprises a plurality of transmitting antennas, The second communication device comprises at least one receiving antenna, The above channel is an OFDM-based channel, A step of generating a spatial and frequency-based channel matrix through estimation of the above channel; and Further comprising the step of transforming the above space and frequency based channel matrix to generate an angle and delay based channel matrix, A method wherein the step of generating at least one latent vector comprises the step of generating the at least one latent vector based on the generated angle- and delay-based channel matrix.

10. In the 9th paragraph, the step of generating at least one latent vector A step of truncating the above angle and delay based channel matrix; and A method comprising the step of inputting the truncated angle and delay channel matrix into the encoder neural network as the first channel state information.

11. In paragraph 9, The above first communication device comprises a base station of a mobile communication system, A method wherein the second communication device comprises a user device of a mobile communication system.

12. In a method for a second communication device to receive channel status information from a first communication device, A step of receiving second channel state information including size information and direction information corresponding to at least one potential vector from the first communication device; and A step of inputting information based on the second channel state information into a decoder neural network to obtain restoration of the first channel state information - including channel state information for a wireless link from the second communication device to the first communication device, A method in which at least one latent vector is generated by the first communication device inputting the first channel state information into an encoder neural network.

13. In paragraph 12, the direction information is, A method in which the first communication device selects a codebook vector closest to the potential vector among a plurality of codebook vectors included in a pre-equipped codebook, and generates a codebook vector based on the selected codebook vector.

14. In paragraph 13, The above direction information includes an identifier of the selected codebook vector, The step of inputting information based on the second channel state information into the decoder neural network is: A step of selecting one of a plurality of codebook vectors included in a codebook pre-equipped in the second communication device as a direction vector of the corresponding potential vector based on an identifier included in the direction information of each of the at least one potential vector; and A method comprising the step of inputting the selected direction vector into the decoder neural network.

15. In paragraph 12, A method in which the above encoder neural network and the above decoder neural network are pre-trained so that the objective function is minimized.

16. In paragraph 15, The above direction information is generated by selecting a codebook vector closest to the potential vector among a plurality of codebook vectors included in a codebook prepared in advance by the first communication device, and based on the selected codebook vector, The above first channel state information includes a channel matrix for the wireless link, The above objective function is defined by the following mathematical equation, Above is the channel matrix, and is the restoration for the channel matrix, z is at least one latent vector, and z q is a quantization result for at least one latent vector, sg is a stop-gradient operator, β is a preset coefficient, A method in which the above encoder neural network, the decoder neural network and the codebook are pre-trained so that the objective function is minimized.

17. In paragraph 12, The above first communication device comprises a base station of a mobile communication system, A method wherein the second communication device comprises a user device of a mobile communication system.

18. A method for training a machine learning model, wherein at least one processor comprises an encoder neural network and a decoder neural network, A step of inputting first channel state information - including channel state information for a wireless link from said second communication device to said first communication device - into said encoder neural network to generate at least one latent vector; For each of the at least one latent vector, a step of quantizing the magnitude and direction of the latent vector to generate magnitude information and direction information; and A step of inputting information based on second channel state information - including size information and direction information corresponding to each of the at least one latent vector - into the decoder neural network to obtain a restoration of the first channel state information; and A method comprising the step of updating parameters of the encoder neural network and parameters of the decoder neural network so that an objective function is minimized.

19. In Article 18, The step of generating the above direction information includes, for each of the at least one latent vector, selecting a codebook vector closest to the latent vector among a plurality of codebook vectors included in the codebook, and generating direction information of the latent vector based on the selected codebook vector. A method wherein the updating step includes updating parameters of the encoder neural network, parameters of the decoder neural network, and the codebook.

20. In clause 19, the objective function Is, It is defined by the following mathematical formula, Above is the channel matrix, and is the restoration for the channel matrix, z is at least one latent vector, and z q A method wherein sg is a quantization result for at least one latent vector, β is a stop-gradient operator, and β is a preset coefficient.

21. In paragraph 19, A method further comprising the step of normalizing a plurality of codebook vectors included in the codebook after the codebook is updated.

22. In paragraph 19, The above codebook is a method composed of a nested codebook structure.

23. In paragraph 18, A method wherein the updating step includes a step of using soft gradient passing during backpropagation of the direction information.

24. In paragraph 18, The above first communication device comprises a base station of a mobile communication system, A method wherein the second communication device comprises a user device of a mobile communication system.

Images