Receiving apparatus and method for generating parameters for demodulation

Abstract

To contribute to improving a bit error rate of the decoder output of a receiver implementing a machine-learning-based iterative demodulation algorithm.SOLUTION: A method implemented in a computer system includes performing training by a machine learning method using a loss function that considers the non-Gaussianity of the log-likelihood ratio (LLR) distribution computed from the output of the network obtained by developing an iterative algorithm for demodulation or demodulation and decoding, and generating a first set of learned parameters for the iterative algorithm.SELECTED DRAWING: Figure 10

Description

Technical Field

[0001] The present disclosure relates to a wireless communication system, and particularly to received signal processing.

Background Art

[0002] The Belief Propagation (probability propagation or belief propagation) (BP) algorithm can be used for signal detection (demodulation and decoding). For example, the BP algorithm may be used for signal detection in a Multi-Input Multi-Output (MIMO) receiver.

[0003] Signal processing using the BP algorithm can be optimized using machine learning (or deep learning). This approach unfolds the iterations of the BP algorithm into a layer-wise structure similar to a neural network, introducing a large number of trainable parameters. Such an approach is called deep unfolding. Non-Patent Documents 1 and 2 propose a method of adjusting a large number of parameters of the BP algorithm for massive MIMO detection by machine learning, thereby improving the convergence characteristics of the BP algorithm. Also, Non-Patent Document 3 proposes a method of removing inter-mode interference by using an iterative demodulation algorithm called the Alternating Direction Method of Multipliers (ADMM) in an Orbital Angular Momentum (OAM) multiplexing transmission system using orbital angular momentum, and adjusting the parameters of the algorithm by machine learning.

Prior Art Documents

Non-Patent Documents

[0004]

Non-Patent Document 1

[0005] Generally, machine-learning-based iterative demodulation algorithms (e.g., Gaussian BP algorithm) are optimized by a loss function that takes into account the mean square error (MSE) or binary cross entropy (BCE). In the MSE norm, numerous parameters of the iterative demodulation algorithm (i.e., the iterative demodulation algorithm) are adjusted to minimize the error between the output of the demodulator (i.e., the iterative demodulation algorithm) and the training data (i.e., the transmitted modulation symbols). On the other hand, in the BCE norm, the parameters are adjusted to maximize the mutual information of the log likelihood ratio (LLR) obtained from the output of the demodulator (i.e., the iterative demodulation algorithm).

[0006] However, a general wireless communication system uses forward error correction (FEC), and at the receiver, error correction processing based on the demodulator output is performed. That is, the final determination of the information bits is made based on the decoder output rather than the demodulator output. It should be noted here that minimizing the bit error rate (BER) of the demodulator output does not necessarily result in minimizing the BER of the decoder output. Error correction codes used in wireless communication systems (e.g., low-density parity-check (LDPC) codes, Reed-Solomon (RS) codes, Viterbi codes, Turbo codes) are usually optimized or designed for transmission over an additive white Gaussian noise (AWGN) channel. Therefore, the decoder can provide the maximum error correction ability when the noise component of the LLR of the demodulator output follows a Gaussian distribution.

[0007] From the above, there is a problem that learning using a loss function based on the MSE metric or BCE metric of the demodulator output cannot necessarily minimize the BER of the decoder output.

[0008] One of the objectives that the embodiments disclosed herein seek to achieve is to provide an apparatus, a method, and a program that contribute to improving the bit error rate of the decoder output of a receiver implementing a machine learning-based iterative demodulation algorithm. It should be noted that this objective is only one of the multiple objectives that the multiple embodiments disclosed herein seek to achieve. Other objectives or problems and novel features will be clarified from the description of this specification or the accompanying drawings.

Means for Solving the Problems

[0009] In a first aspect, a method implemented in a computer system includes the following steps: (a) Training a network obtained by deploying an iterative algorithm for demodulation or demodulation and decoding by a machine learning technique using a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network, and (b) Generating a first set of learned parameters of the iterative algorithm by the training.

[0010] In a second aspect, a receiving device includes a memory and at least one processor. The memory stores one or more sets of learned parameters generated by the method of the first aspect. The at least one processor is configured to execute an iterative algorithm using any one of the one or more sets on a plurality of received signals to generate a plurality of LLR vectors corresponding to a plurality of transmission symbols. Further, the at least one processor is configured to perform error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences.

[0011] In a third aspect, the receiving device includes a memory and at least one processor coupled to the memory. The at least one processor is configured to execute an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of LLR vectors corresponding to a plurality of transmitted symbols. Here, the first set of the learned parameters is a set of parameters generated by training a network obtained by deploying the iterative algorithm using a machine learning method with a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network. Further, the at least one processor is configured to perform error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences.

[0012] In a fourth aspect, the method performed by the receiving device includes: (a) executing an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of LLR vectors corresponding to a plurality of transmitted symbols; and (b) performing error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences. The first set of the learned parameters is a set of parameters generated by training a network obtained by deploying the iterative algorithm using a machine learning method with a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network.

[0013] In a fifth aspect, the program includes a set of instructions (software code) for causing a computer to perform the method according to the first aspect when loaded into the computer.

[0014] In a sixth aspect, the program includes a set of instructions (software code) for causing a computer to perform the method according to the fourth aspect when loaded into the computer.

Advantages of the Invention

[0015] According to the above aspect, it is possible to provide an apparatus, a method, and a program that contribute to improving the bit error rate of the decoder output of a receiver implementing a machine learning-based iterative demodulation algorithm.

Brief Description of the Drawings

[0016]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Modes for Carrying Out the Invention

[0017] Hereinafter, specific embodiments will be described in detail with reference to the drawings. In each drawing, the same or corresponding elements are denoted by the same reference numerals, and redundant descriptions are omitted as necessary for clarity of explanation.

[0018] The embodiments described below are mainly described for large-scale multi-user MIMO receivers. However, these embodiments may be applied to other receivers implementing machine learning-based iterative demodulation algorithms. For example, the following embodiments may be applied to receivers of point-to-point MIMO systems or receivers of Line of Sight (LOS) MIMO systems. The following embodiments may be applied to multi-user detection using iterative algorithms in Non-Orthogonal Multiple Access (NOMA) systems. The following embodiments may be applied to receivers of optical MIMO systems. The optical MIMO system can be said to be a space-division multiplexing transmission system using a multimode optical fiber or a multicore optical fiber. By way of example and not limitation, the machine learning-based iterative demodulation algorithm may be the BP algorithm described in the following embodiments. However, these embodiments may use other machine learning-based iterative demodulation algorithms (e.g., the alternating direction method of multipliers).

[0019] Also, by way of example and not limitation, the following embodiments describe a receiver architecture that determines information bits after demodulation processing by a demodulator and decoding processing by a decoder. Alternatively, in the case of a communication system that is doubly error-correction coded, a receiver architecture in which first decoding processing and second decoding processing are performed after demodulation processing may be employed. In that case, an iterative (demodulation) algorithm may perform demodulation processing and first decoding processing. In other words, the receiver architecture described in the following embodiments can be replaced with a receiver architecture in which second decoding is performed after an iterative (demodulation) algorithm for demodulation and first decoding. Further alternatively, the receiver architecture described in the following embodiments can be replaced with an iterative detection and decoding type receiver architecture. This is sometimes referred to as joint iterative demodulation and decoding. An iterative demodulation and decoding receiver feeds back information on code bits obtained by a decoder (or demodulation processing, demodulation step) as a priori information to a demodulator (or demodulation processing, demodulation step), and repeats an iterative loop including demodulation and decoding. The iterative demodulation and decoding receiver may execute a plurality of iterative algorithms (or a plurality of iterative loops). Specifically, the iterative demodulation and decoding receiver may include an outer iterative loop between demodulation (e.g., MIMO detection) and decoding (e.g., LDPC decoding, or Turbo decoding), and an inner iterative loop within the demodulation processing. Further, as will be readily understood by those skilled in the art, the inner iterative loop may include a part of the decoding processing (first decoding processing). The method for generating parameters using machine learning described in the following embodiments may be used to generate a parameter set for the inner iterative loop by machine learning, or may be used to simultaneously generate parameter sets for the outer and inner iterative loops by machine learning.

[0020] That is, the terms "iterative demodulation algorithm" and "iterative algorithm" as used herein refer to iterative algorithms that are used at least for demodulation and may be used for demodulation and (part of) decoding. Also, the term "demodulation" as used herein may be referred to as, for example, soft demodulation, signal detection, detection, soft detection, demapping, or soft demapping. The following embodiments can be widely applied to a receiver architecture in which a decoding process (or a second decoding process) is arranged after an iterative demodulation algorithm.

[0021] <First Embodiment> FIG. 1 shows a configuration example of a wireless communication system (i.e., a multiple access cellular system) according to an embodiment. Referring to FIG. 1, a base station 1 provides wireless access to a plurality of wireless terminals 2. The base station 1 may be referred to by an access point, a Transmission / Reception Point (TRP), or other names. The base station 1 may be, for example, a gNB or a gNB Distributed Unit (gNB-DU) of a 5G system. In some implementations, the wireless communication system may utilize multi-user MIMO technology for uplink transmission from a plurality of wireless terminals 2 to the base station 1. In this case, the base station 1 may receive reference signals from a plurality of wireless terminals 2, estimate the MIMO channel between the plurality of wireless terminals 2 and the base station 1 using the received reference signals, receive data signals from the plurality of wireless terminals, and detect the transmission signals using the estimated channel. That is, the base station 1 may perform MIMO detection to separate the multi-user signals of the plurality of wireless terminals 2.

[0022] FIG. 2 shows an example of a system model for uplink multi-user MIMO transmission. In FIG. 2, a plurality of transmitters 20 (20-1, ··· 20-M') of a plurality of wireless terminals 2 communicate with a receiver 10 of a base station 1 via a channel (propagation path) 30. In the example of FIG. 2, each of the M' transmitters 20 has one transmit antenna. Alternatively, each transmitter 20 may have two or more transmit antennas. The receiver 10 of the base station 1 has N' receive antennas. It is assumed that the total number M' of transmit antennas is less than or equal to the total number N' of receive antennas.

[0023] In the following description, for the sake of simplicity, it is assumed that the transmission signal from each wireless terminal 2 (user) is single-carrier transmission and the propagation path between each wireless terminal 2 and the base station 1 is flat fading. On the other hand, even in a multipath fading environment where the transmission signal from each user uses Orthogonal Frequency Division Multiplexing (OFDM) or Single Carrier-Frequency Division Multiple Access (SC-FDMA), etc., by inserting a cyclic prefix of an appropriate length into the transmission signal, it can be regarded as flat fading for each subcarrier unit. Therefore, the present embodiment may be applied to OFDM and SC-FDMA.

[0024] It is assumed that transmission signals modulated by Quadrature Amplitude Modulation (QAM) are transmitted from a total of M' transmit antennas of a plurality of wireless terminals 2 and received at the base station 1 having N' receive antennas. At this time, the complex signal model in the equivalent low-pass representation is expressed by the following equation:

Equation

[0025] Let the number of modulation symbols of QAM modulation be Q’. For example, when it is Quadrature Phase shift Keying (QPSK), Q’ is 4, and when it is 16QAM, Q’ is 16. Regarding the amplitudes of the modulation symbols, assume that the amplitudes on the I-axis and Q-axis are {+c, -c} for QPSK and {+c, -c, +3c, -3c} for 16QAM. Here, c is represented by the following formula. E s is the average signal power. Let the power of the complex noise at each receiving antenna be N0.

Number

[0026] For the sake of simplicity of explanation, the received signal model obtained by replacing the equivalent low-pass representation of complex numbers with an equivalent real-valued signal model y is expressed by the following formula:

Number

[0027] Figure 3 shows a configuration example of base station 1. Referring to Figure 3, base station 1 includes a Radio Frequency (RF) transceiver 301, a network interface 303, a processor 304, and a memory 305. The RF transceiver 301 performs analog RF signal processing to communicate with a plurality of wireless terminals 2. The RF transceiver 301 may include a plurality of transceivers. The RF transceiver 301 is coupled to an antenna array 302 and a processor 304. The RF transceiver 301 receives modulation symbol data from the processor 304, generates a transmission RF signal, and supplies the transmission RF signal to the antenna array 302. Also, the RF transceiver 301 generates a baseband reception signal based on the received RF signal received by the antenna array 302 and supplies this to the processor 304. The RF transceiver 301 may include an analog beamformer circuit for beamforming. The analog beamformer circuit includes, for example, a plurality of phase shifters and a plurality of power amplifiers.

[0028] The network interface 303 is used to communicate with network nodes (e.g., other base stations and core network nodes). The network interface 303 may include, for example, a network interface card (NIC) compliant with the IEEE 802.3 series.

[0029] The processor 304 performs digital baseband signal processing (data plane processing) and control plane processing for wireless communication. The processor 304 may include a plurality of processors. For example, the processor 304 may include a modem processor (e.g., a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), or a Digital Signal Processor (DSP)) that performs digital baseband signal processing and a protocol stack processor (e.g., a Central Processing Unit (CPU) or a Micro Processing Unit (MPU)) that performs control plane processing.

[0030] For example, the digital baseband signal processing by the processor 304 may include signal processing of the Service Data Adaptation Protocol (SDAP) layer, the Packet Data Convergence Protocol (PDCP) layer, the Radio Link Control (RLC) layer, the Medium Access Control (MAC) layer, and the Physical (PHY) layer. Also, the control plane processing by the processor 304 may include processing of Non-Access Stratum (NAS) messages, Radio Resource Control (RRC) messages, Medium Access Control (MAC) Control Elements (CEs), and Downlink Control Information (DCI).

[0031] The processor 304 may include a digital beamformer module for beamforming. The digital beamformer module may include a MIMO encoder and a precoder.

[0032] The memory 305 is composed of a combination of volatile memory and non-volatile memory. The volatile memory is, for example, Static Random Access Memory (SRAM) or Dynamic RAM (DRAM) or a combination thereof. The non-volatile memory is a mask Read Only Memory (MROM), an Electrically Erasable Programmable ROM (EEPROM), a flash memory, or a hard disk drive, or any combination thereof. The memory 305 may include storage located away from the processor 304. In this case, the processor 304 may access the memory 305 via the network interface 303 or other I / O interfaces.

[0033] Memory 305 may include a computer-readable medium storing one or more software modules (computer programs) including instruction groups and data for performing at least part of the processing by base station 1. In some implementations, processor 304 may be configured to perform at least part of the processing by base station 1 described in the above embodiments by reading and executing the software modules from memory 305.

[0034] According to this embodiment, processor 304 can cause base station 1 to perform reception signal processing (i.e., demodulation and decoding) for multi-user detection (MIMO detection). For this purpose, processor 304 may include a demodulator 470 and one or more decoders 480 (480-1, ··· 480-M) shown in FIG. 4.

[0035] Demodulator 470 includes a BP detector 400 and an LLR generator 460. BP detector 400 receives N received signals y1~y N obtained by N' receiving antennas and executes an iterative BP algorithm with a total number of iterations T to perform multi-user detection (Step 501 of Figure 5) . Thereafter, BP detector 400 provides estimates s'1 of the separated M transmitted signal (transmission symbol) components (T) ~s' M (T) and equivalent gains ω'1 (T) ~ω' M (T) to LLR generator 460. LLR generator 460 generates M LLR vectors corresponding to the M transmission symbols (Step 502 of Figure 5) . Each LLR vector indicates LLR values (LLR m,1 ···LLR m,B ) of the coded bits mapped to each transmission symbol. The number B of coded bit LLR values is equal to the number of bits included in one transmission symbol. For example, if 16QAM is used as the complex modulation symbol, the number of bits per symbol is 4, and in the equivalent real number model, B is equal to 2.

[0036] Each decoder 480 receives a corresponding LLR vector and performs error correction decoding, thereby generating a decoded bit sequence. (Step 503 of Figure 5) Each decoder 480 may be, for example, but not limited to, an LDPC decoder. Each decoder 480 may perform decoding for other types of error correction codes, such as RS codes, convolutional codes, or Turbo codes.

[0037] The BP detector 400 uses the learned parameter set 450 in the BP algorithm. (Step 501 of Figure 5) The parameter set 450 may be stored in the memory 305 of the base station 1. Techniques for improving the performance of signal processing using the BP algorithm include damping, scaling, and node selection. For example, the learned parameter set 450 may include one or any combination of a plurality of scaling factors, a plurality of damping factors, and a plurality of node selection factors.

[0038] In one example, the learned parameter set 450 includes a plurality of scaling factors and a plurality of damping factors. The BP detector 400 uses these plurality of scaling factors respectively in different iterations of the BP algorithm. Similarly, the BP detector 400 uses these plurality of damping factors respectively in different iterations of the BP algorithm. Therefore, the total number of the plurality of scaling factors and the total number of the plurality of damping factors may be equal to the total number of iterations of the BP algorithm.

[0039] In other examples, the learned parameter set 450 includes a plurality of scaling factors and a plurality of node selection factors. The BP detector 400 uses these plurality of scaling factors respectively in different iterations of the BP algorithm. Similarly, the BP detector 400 uses these plurality of node selection factors respectively in different iterations of the BP algorithm. Note that, as will be described later, the BP algorithm may use a plurality of node selection factors for each iteration. In this case, the parameter set 450 may include a set of node selection factors for each iteration.

[0040] Note that damping suppresses the oscillation of beliefs that cause poor convergence by taking a weighted average of the beliefs generated in past iterations and the beliefs generated in the current iteration as a new belief. The damping factor determines the weighting factor (factor, coefficient) of the weighted average. Scaling introduces a parameter (scaling factor) that adjusts the absolute value of the belief to gradually increase as the iteration increases, considering that the reliability of the belief at the beginning of the iteration is low. In the case of MIMO detection, node selection is used as a countermeasure against fading space correlation (correlation between receiving antennas). Specifically, in node selection, a set of receiving antenna elements is divided into a plurality of subsets. Each subset is composed of spatially separated (low-correlation) receiving antenna elements. The BP algorithm with node selection updates only the beliefs of one subset in one BP iteration and sequentially updates the beliefs of other subsets in subsequent BP iterations.

[0041] The parameter set 450 is generated by training, using a machine learning technique, a network (or graph) obtained by deploying the iterative BP algorithm of the BP detector 400. This machine learning uses a loss function that takes into account the non-Gaussianity (non-Gaussian nature) of the LLR distribution calculated from the output of the network. The non-Gaussianity of the LLR distribution represents the deviation of the LLR distribution from the Gaussian distribution. In other words, the magnitude of the non-Gaussianity of the LLR distribution indicates how much the LLR distribution deviates from the Gaussian distribution. The loss function may be defined to include a term representing the kurtosis or negentropy of the LLR distribution in order to measure the non-Gaussianity of the LLR distribution. In other words, the loss function may include a term representing the kurtosis metric or the negentropy metric of the LLR distribution.

[0042] Kurtosis measures the degree of peakedness of a distribution, which is zero only in the case of a Gaussian distribution. The kurtosis of other distributions is positive if it is super-Gaussian (more spiky than the Gaussian distribution) and negative if it is sub-Gaussian (flatter than the Gaussian distribution). Therefore, the absolute value of the kurtosis or the square of the kurtosis can be used to measure the non-Gaussianity of a distribution. Negentropy is also called (normalized) differential entropy. The negentropy of a distribution is defined as the value obtained by subtracting the entropy of the distribution from the entropy of a Gaussian distribution having the same variance as the distribution. It should be noted that an important property of the Gaussian distribution is that it has the maximum entropy among all distributions over the entire real axis. Since negentropy is always greater than zero unless the distribution is a Gaussian distribution, it can be used to measure the non-Gaussianity of a distribution.

[0043] As is known in the field of independent component analysis, the negentropy J of a distribution x can be approximated as follows:

Equation

[0044] Alternatively, the negentropy J of the distribution x standardized to have zero mean and unit variance may be approximated as follows: [Number]

[0045] In the learning process for obtaining the parameter set 450, the LLR distribution calculated from the output of the network may be standardized such that it has zero mean and unit variance. Then, the parameters in the network may be adjusted to minimize a loss function including a term representing the negentropy metric (or kurtosis metric) of the standardized LLR distribution. Note that in order to maximize the error correction ability, it is known that the mean value and variance value of the LLR distribution need to satisfy the consistency condition of 1:2. Using this, the variance of the LLR distribution may be calculated by obtaining twice the mean value of the LLR distribution.

[0046] By using the learned parameter set 450 described above, the demodulator 470 can approximate the LLR output after the demodulation process (i.e., the output of the LLR generator 460) to a Gaussian distribution. In other words, the demodulator 470 can approximate the distribution of the LLR input to the subsequent decoding process (or the second decoding process) to a Gaussian distribution and can provide an LLR suitable for error correction decoding to each decoder 480. As a result, the receiver 10 of the present embodiment can improve the BER of the decoder output.

[0047] Next, some examples of loss functions used in the learning of parameter set 450 will be described below. In the first example, the loss function may be defined as follows:

Equation

[0048] In the second example, the loss function may be defined as a weighted sum of a first term representing the negative entropy metric of the LLR distribution and a second term representing the MSE metric between the output of the network and the teacher data. Specifically, the loss function may be defined as follows:

Equation

[0049] In the third example, the loss function may be defined as a weighted sum of a first term representing the negative entropy metric of the LLR distribution and a second term representing the cross-entropy metric between the output of the network and the teacher data. Specifically, the loss function may be defined as follows:

Equation

[0050] The loss functions in the above-described second and third examples contribute to determining the parameter set 450 so as to approximate the LLR distribution to a Gaussian distribution while maintaining a small error from the true value (transmitted signal).

[0051] Subsequently, a configuration example of the BP detector 400 will be described below. FIG. 6 shows a configuration example of the BP detector 400. Referring to FIG. 6, the BP detector 400 includes N soft interference cancelers (SOFT ICs) 610-1 to 610-N, a belief generator (BG) 620, and N soft replica generators (SOFT RGs) 630-1 to 630-N. The soft interference cancelers 610-1 to 610-N receive N received signals y1 to y N respectively, obtained by N' receiving antennas. For example, the soft interference canceller 610-1 receives the received signal y1 of the first antenna (referred to as the first received signal). In addition, in order to perform the t-th iteration, the soft interference canceller 610-1 receives all the soft replicas x hat 1,1 (t-1) ~x hat 1,M (t-1) generated in the previous (t-1)-th iteration. Here, x hat means a caret over the character x. Then, the soft interference canceller 610-1 generates the received signal y tilde 1,1 (t) ~y tilde 1,M (t) after cancellation. Here, y tilde means a tilde over the character y.

[0052] The belief generator 620 reads out a plurality of damping coefficients (or a plurality of sets of node selection coefficients) included in the learned parameter set 450 described above from the memory 305. The belief generator 620 receives the received signal y tilde 1,1 (t) ~y tilde 1,M (t) after cancellation from the soft interference canceller 610-1. The belief generator 620 also receives the received signal y tilde after cancellation generated in the same manner from the other soft interference cancelers 610-n (where n is from 2 to N).n,1 (t) tilde y n,M (t) receives. Then, the belief generator 620 uses the damping factor (or set of node selection factors) for the t-th iteration and generates a belief r 1,1 (t) ~r 1,M (t) associated with the first received signal. Similarly, the belief generator 620 generates beliefs associated with the remaining second to N-th received signals.

[0053] The soft replica generator 630-1 reads a plurality of scaling factors included in the learned parameter set 450 described above from the memory 305. The soft replica generator 630-1 receives the belief r 1,1 (t) ~r 1,M (t) from the belief generator 620. Then, the soft replica generator 630-1 uses the scaling factor for the t-th iteration and generates a soft replica x-hat 1,1 (t) ~x-hat 1,M (t) and further generates a soft replica power p 1,1 (t) ~p 1,M (t)

[0054] After completion of the BP processing for the total number of iterations T, the belief generator 620 provides the estimated values s’1 (T) ~s’ M (T) of the separated M transmitted signal (transmission symbol) components and the equivalent gains ω’1 (T) ~ω’ M (T) to the LLR generator 460.

[0055] Hereinafter, the processing performed by the soft interference canceller 610, the belief generator 620, and the soft replica generator 630 will be described in more detail. In addition, the processing performed by the LLR generator 460 will be described. ​

[0056] (1) Soft interference canceller In the first iteration, the soft replica has not been generated yet. Therefore, the soft interference canceller 610 supplies the first to Nth received signals to the belief generator 620 without performing the cancellation process. In the t-th iteration after the second time, the soft interference canceller 610-n associated with the n-th received signal cancels M-1 transmitted signal components other than the m-th transmitted signal from the n-th received signal, and the received signal ŷ after cancellation n,m (t) is generated. The received signal ŷ after cancellation n,m (t) is expressed by the following formula:

Equation

[0057] (2) Belief generator The belief generator 620 generates a belief using the received signal after the cancellation process. First, the belief generator 620 performs the process expressed by the following formula using the received signal ŷ after the cancellation process for the n-th receiving antenna n,m (t) to obtain the transmitted signal component s n,m (t) in the t-th iteration.

Equation

Equation

[0058] The equivalent gain ω n,m (t) for the true transmission signal x m contained in the transmission signal component s n,m (t) is expressed by the following formula and is used for normalization when performing the scaling process:

Equation

[0059] Next, the belief generator 620 uses the transmission signal component s n,m (t) to generate the belief r n,m (t) The belief generator 620 uses either the damping process or the node selection process. First, the damping process will be described. The damping process calculates the weighted average of the transmission signal component obtained in the previous (t - 1)-th iteration and the transmission signal component obtained in the current t-th iteration using the damping coefficient η (t) as follows:

Equation

Equation

[0060] Next, node selection will be described. In node selection, s' is obtained by synthesizing the transmission signal components of each antenna in the last K iterations according to the following formula: n,m (t) :

Equation

[0061] By the node selection process, the equivalent gain included in s' n,m (t) is expressed by the following formula:

Equation

[0062] The belief generator 620 normalizes s' obtained by performing either damping or node selection with ω' n,m (t) to obtain the normalized belief r n,m (t) : n,m (t)Generate this and supply it to the soft replica generator 630. The normalized belief r n,m (t) is expressed by the following equation:

Number

[0063] (3) Soft replica generator The soft replica generator 630 scales the belief r (t) by the scaling factor a n,m (t) and calculates the soft replica x-hat n,m (t) and the soft replica power p n,m (t) according to the following equation:

Number

Number

[0064] The decision threshold s' can take any value in the set S Q’ . The set S Q’ is {0} for QPSK and {0, +2c, -2c} for 16QAM. The tanh function is the hyperbolic tangent function. These equations synthesize the belief information around the decision threshold s' to generate the soft replica x-hat n,m (t) and the soft replica power p n,m (t) .

[0065] (4) Output of the BP detector When the T-th iteration ends, the belief generator 620 outputs the estimated values s’ of the M separated transmitted signal components m (T) and the equivalent gain ω’ m (T) to the LLR generator 460. The estimated value r m (T) is expressed by the following equation:

Equation

[0066] (5)LLR Generator The LLR generator 460 generates M LLR vectors corresponding to the M transmitted symbols. Each LLR vector indicates the bit LLR values (LLR m,1 ···LLR m,B ) of the codeword mapped to each transmitted symbol. The generation of the LLR vector is not limited to a specific method, and various known methods can be used. In one example, when the transmitted signal xm is composed of B codeword bits c m (1), …, c m (n s ), c m (B), the bit LLR corresponding to the n s -th codeword bit c m (n s ) may be calculated by the following equation:

Equation

Equation

[0067] Next, the learning method of the parameter set 450 will be described below. FIG. 7 is a conceptual diagram showing deep unfolding for multi-user detection based on BP. Deep unfolding is a method of unfolding an iterative algorithm in the iterative direction, regarding the obtained processing flow graph as a deep neural network (DNN), and applying a deep learning scheme. When the BP detector 400 is unfolded in the iterative direction, the BP network shown in FIG. 7 is given. Each BP iteration corresponds to one layer of the DNN. This enables the learning of the meta-parameters embedded in the BP network. For example, as shown in FIG. 7, the learnable (trainable) parameters are the scaling coefficient a (t) and the damping coefficient η (t) of each iteration (each layer). Instead of this, the learnable (trainable) parameters may include the scaling coefficient a (t) of each iteration (each layer) and a set of node selection coefficients {η i,t-k (t)}. The learning may be performed based on the gradient method, and the meta-parameters may be adjusted together in the direction in which the loss function becomes smaller. As already described, the loss function takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network (BP network). Thereby, the meta-parameters are adjusted to reduce the non-Gaussianity of the LLR distribution. The loss function may further consider the difference between the output of the network and the teacher data (true value). The loss function may be defined as a weighted sum of a first term representing the negentropy metric of the LLR distribution and a second term representing the MSE (or BCE) metric between the output of the network and the teacher data.

[0068] FIG. 8 shows an example of a parameter set 450 obtained by deep learning. In the example of FIG. 8, a parameter set is determined for each Modulation and Coding Scheme (MCS). For example, as known for Long Term Evolution (LTE), LTE-Advanced, and 5G systems, different MCS numbers are associated with different combinations of modulation schemes (or modulation orders) and code rates.

[0069] To obtain the plurality of parameter sets shown in FIG. 8, learning is performed for each MCS. The learning for different MCSs may use different loss functions. For example, depending on the code rate of the error correction code, the magnitude of the weight value w1 or w2 of the objective function defined by the above formula (15) or (16) may be changed. Specifically, when the code rate is relatively low, that is, when the error correction ability at the decoder is relatively strong, the weight value may be determined so that the loss function preferentially considers the negative entropy metric of the LLR distribution. In contrast, when the code rate is relatively high, that is, when the error correction ability at the decoder is relatively poor, the weight value may be determined so that the loss function preferentially considers the MSE (or BCE) metric. In other words, as the code rate increases, the weight value may be changed so as to decrease the weight of the negative entropy metric of the LLR distribution and increase the weight of the MSE (or BCE) metric. Thereby, when the error correction ability is strong, the parameter set is adjusted to emphasize the decoding process, and when the error correction ability is poor, the parameter set is adjusted to emphasize the demodulation process. Thereby, the bit error rate characteristics at each code rate can be improved.

[0070] Further or alternatively, depending on the modulation method or the number of modulation levels, the magnitude of the weight w1 or w2 of the objective function defined by the above formula (15) or (16) may be changed. Specifically, when the number of modulation levels is relatively large, that is, when the Gaussianity of the LLR of the demodulator output is low, the weight value may be determined so that the loss function preferentially considers the negentropy metric of the LLR distribution. In contrast, when the number of modulation levels is relatively small, that is, when the Gaussianity of the LLR of the demodulator output is high, the weight value may be determined so that the loss function preferentially considers the MSE (or BCE) metric. In other words, as the number of modulation levels decreases, the weight value may be changed so as to decrease the weight of the negentropy metric of the LLR distribution and increase the weight of the MSE (or BCE) metric. Thereby, when the number of modulation levels is large and the Gaussianity of the LLR distribution is low, the error correction ability can be enhanced by making the LLR distribution closer to the Gaussian distribution. Conversely, when the number of modulation levels is small and the Gaussianity of the LLR distribution of the demodulator output is high, the parameter set is adjusted to emphasize the demodulation process. Thereby, the bit error rate characteristics when each modulation method is adopted can be improved.

[0071] In one example, learning using the first loss function may be performed to obtain the first parameter set, and learning using the second loss function may be performed to obtain the second parameter set. The second loss function is defined to consider the non-Gaussianity of the LLR distribution deeper than the first loss function. The first and second loss functions may be distinguished by the difference in the weight value w1 or w2 of the objective function defined by the above formula (15) or (16). The first parameter set is used when the coding rate is the first value or when the number of modulation levels is the second value. On the other hand, the second parameter set is used when the coding rate is lower than the first value or when the number of modulation levels is larger than the second value.

[0072] Further or alternatively, multiple trainings may be performed to obtain a plurality of parameter sets associated with different Signal to Noise Ratios (SNRs). Trainings regarding different SNRs may use different loss functions. For example, depending on the Signal to Noise Ratio (SNR), the magnitude of the weight w1 or w2 of the objective function defined by the above formula (15) or (16) may be changed. Specifically, when the SNR is small, the accuracy of the demodulator output is low, and the effect of the error correction ability by the decoder becomes relatively large. Therefore, the weight value may be determined so that the loss function preferentially considers the negative entropy metric of the LLR distribution. In contrast, when the SNR is large, since the accuracy of the demodulator output becomes high, the weight value may be determined so that the loss function preferentially considers the MSE (or BCE) metric. In other words, as the SNR increases, the weight value may be changed so as to decrease the weight of the negative entropy metric of the LLR distribution and increase the weight of the MSE (or BCE) metric. Thereby, when the SNR is low and the accuracy of the demodulator output is low, the error correction ability can be enhanced by making the LLR distribution closer to the Gaussian distribution. On the contrary, when the SNR is high, the parameter set is adjusted to emphasize the demodulation process. Thereby, the bit error rate characteristics at each coding rate can be improved.

[0073] In one example, training using the first loss function may be performed to obtain the first parameter set, and training using the third loss function may be performed to obtain the third parameter set. The third loss function is defined to consider the non-Gaussianity of the LLR distribution deeper than the first loss function. The first and third loss functions may be distinguished by the difference in the weight value w1 or w2 of the objective function defined by the above formula (15) or (16). The first parameter set is used when the SNR is the third value. On the other hand, the third parameter set is used when the SNR is smaller than the third value.

[0074] FIG. 9 shows an example of a training system environment. The training data set 910 includes a transmitted signal data set 912 and a received signal data set 914. The transmitted signal data set 912 may be randomly generated. The received signal data set 914 corresponds to the transmitted signal data set 912 and is generated using the transmitted signal data set 912 and a given channel matrix. The channel matrix may be randomly generated or may be based on a propagation model defined in, for example, the 3rd Generation Partnership Project (3GPP) specifications. Alternatively, the channel matrix may be generated based on measurement results in the actual environment where base station 1 is installed.

[0075] The training system 920 includes a demodulator module 922 and a learning module 924. The demodulator module 922 emulates the processor 304 or the demodulator 470 of base station 1. The demodulator module 922 can execute the same BP algorithm implemented in base station 1. The learning module 924 trains the demodulator module 922 using the training data set 910. The learning module 924 may apply one or more machine learning algorithms.

[0076] The learning module 924 uses the loss function described above, that is, a loss function that takes into account the non-Gaussianity of the LLR distribution. In one example, the learning module 924 may use an update algorithm that follows the gradient method. The update algorithm of the gradient method used may be, for example, the Adaptive moment estimation (Adam) optimizer algorithm. In addition, the learning module 924 may use mini-batch learning. The number of learning times may be set to an appropriate number considering overfitting to the training data. For updating the learning rate, a Step algorithm that gradually narrows the update width with respect to the number of learning times may be used.

[0077] The learning module 924 outputs a learned (or trained) parameter set 930 obtained by machine learning. The parameter set 930 may be stored in the memory 305 of the base station 1 as the parameter set 450.

[0078] FIG. 10 shows an example of the operation of the training system 920. In step 1001, the training system 920 receives a training data set 910. In step 1002, the training system 920 trains a network obtained by developing an iterative algorithm (e.g., BP algorithm) for demodulation (or demodulation and decoding) by machine learning using a loss function that takes into account the non-Gaussianity of the LLR distribution. In step 1003, the training system 920 stores the learned (or trained) parameter set in the memory.

[0079] The training system 920 may be a computer system as shown in FIG. 11. FIG. 11 shows a configuration example of the computer system 1100. The computer system 1100 executes a computer program including a set of instructions, thereby performing, for example, a method for the training system 920. The training system 920 may be a stand-alone computer or may include one or more networked computers. The computer system 1100 may be a server or a client or both in a server-client environment. The computer system 1100 may be a personal computer, a tablet computer, or a smartphone.

[0080] In the example of FIG. 11, computer system 1100 includes one or more processors 1110, memory 1120, and mass storage 1130, which communicate with each other via bus 1170. One or more processors 1110 may include, for example, a Central Processing Unit (CPU) or a Graphics Processing Unit (GPU) or both. Computer system 1100 may include other devices such as one or more output devices 1140, one or more input devices 1150, and one or more peripherals 1160. One or more output devices 1140 include, for example, a video display, speakers. One or more input devices 1150 include, for example, a keyboard, a mouse, a keypad, a touchpad, or a touch screen, or any combination thereof. One or more peripherals 1160 include a printer, a modem, or a network adapter, or any combination thereof.

[0081] One or both of memory 1120 and mass storage 1130 include a computer-readable medium storing one or more instruction sets. These instructions may be disposed, in part or in whole, in memory within processor 1110. These instructions, when executed on processor 1110, cause processor 1110 to perform, for example, the machine learning process described with reference to FIG. 10.

[0082] As described above, in some implementations, the processor 304 of the base station 1 executes one or more programs including a set of instructions for causing a computer to perform the algorithms described in this embodiment. In addition, the training system 920 executes one or more programs including a set of instructions for causing a computer to perform the machine learning described in this embodiment. These programs can be stored in various types of non-transitory computer readable media and supplied to a computer. The non-transitory computer readable media include various types of tangible storage media. Examples of non-transitory computer readable media include magnetic recording media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks), Compact Disc Read Only Memory (CD-ROM), CD-R, CD-R / W, semiconductor memories (e.g., mask ROM, Programmable ROM (PROM), Erasable PROM (EPROM), flash ROM, Random Access Memory (RAM)). Also, the programs may be supplied to a computer by various types of transitory computer readable media. Examples of transitory computer readable media include electrical signals, optical signals, and electromagnetic waves. The transitory computer readable media can supply the programs to a computer via wired communication paths such as electric wires and optical fibers, or wireless communication paths.

[0083] FIG. 12 shows the bit error rate (BER) performance of the receiver (demodulator 470 and decoder 480) according to the present embodiment. These are simulation results for a multi-user MIMO configuration with the number of terminals M' and the number of receiving antenna elements N', where (N', M') = (32, 32). Further, the error correction code is an LDPC code (coding rate 1 / 3, code length 1024 bits), and the modulation method is 16QAM. Graph 1210 shows the BER of the receiver using a parameter set obtained using the loss function of Equation (15) considering the negentropy metric with the weight value w1 = 0.75. In contrast, graph 1220 shows the BER of the receiver using a parameter set obtained using a loss function that considers only the MSE metric (i.e., weight value w1 = 0). Comparing these two graphs, it can be seen that a gain of 1 dB or more can be obtained in terms of the bit error rate by performing learning using the loss function considering the negentropy weight. -3 at.

[0084] <Second Embodiment> This embodiment provides a modification of the learning of the parameter set described in the first embodiment. The configuration examples of the wireless communication system and the base station according to this embodiment are the same as those in the first embodiment.

[0085] In this embodiment, the receiver 10 (demodulator 470, BP detector 400) of the base station 1 uses a second parameter set different from the first parameter set used for the demodulation process at the first transmission for the demodulation process at the retransmission. Similar to the first parameter set, the second parameter set is obtained by training a network obtained by deploying an iterative algorithm for demodulation (or demodulation and decoding) using a machine learning method. However, the second loss function used for the learning of the second parameter set is different from the first loss function used for the learning of the first parameter set. The second loss function is defined to consider the non-Gaussianity of the LLR distribution of the demodulator output more deeply than the first loss function.

[0086] When a bit error is detected in the decoder output, the receiver 10 can request retransmission from the transmitter 20. For example, the receiver 10 can use hybrid automatic repeat request (HARQ) to combine the LLR of the first transmission and the LLR of the retransmission. The use of the combined LLR brings about an improvement in error correction ability. Therefore, for the second loss function, it is preferable to increase the weight of the negentropy metric of the LLR distribution more than that of the first loss function. Thereby, the parameter set is adjusted to emphasize the decoding process during retransmission when the error correction ability is strong. This can contribute to the improvement of the bit error rate characteristic.

[0087] <Other Embodiments> As already described, the above-described embodiments may be applied to other receivers implementing a machine learning-based iterative demodulation algorithm different from the wireless receiver for multi-user MIMO.

[0088] The configuration of the demodulator 470 described in the above-described embodiments is an example, and various modifications are possible. For example, the BP detector 410 may include an LLR generator within the iterative process. In this case, the belief generator 620 of the BP detector 410 may output an LLR. A deinterleaver may be arranged between the demodulator 470 and the decoder 480.

[0089] The above-described embodiments are merely examples regarding the application of the technical idea obtained by the present inventor. That is, the technical idea is not limited to only the above-described embodiments, and it goes without saying that various changes are possible.

[0090] For example, part or all of the above-described embodiments may be described as follows in the appended claims, but are not limited thereto.

[0091] (Appended Claim 1) A method implemented in a computer system, Training a network obtained by developing an iterative algorithm for demodulation or demodulation and decoding by a machine learning method using a first loss function that takes into account the non-Gaussianity of the log likelihood ratio (LLR) distribution calculated from the output of the network, and generating a first set of learned parameters of the iterative algorithm by the training; A method comprising: (Appendix 2) The method according to Appendix 1, wherein the first loss function is defined to include a term representing the negentropy or kurtosis of the LLR distribution for measuring the non-Gaussianity of the LLR distribution. (Appendix 3) The method according to Appendix 1 or 2, wherein the first loss function is further defined to take into account the difference between the output of the network and the teacher data. (Appendix 4) The method according to Appendix 3, wherein the first loss function is defined as a weighted sum of a term representing the negentropy of the LLR distribution and a term representing the mean squared error between the output of the network and the teacher data. (Appendix 5) The method according to Appendix 3, wherein the first loss function is defined as a weighted sum of a term representing the negentropy of the LLR distribution and a term representing the cross entropy between the output of the network and the teacher data. (Appendix 6) The iterative algorithm is an iterative Belief Propagation (BP) algorithm, and the first set of learned parameters includes one or any combination of a plurality of scaling coefficients, a plurality of damping coefficients, and a plurality of node selection coefficients. The method according to any one of Appendices 1 to 5. (Appendix 7) Further comprising generating a second set of learned parameters of the iterative algorithm by training the network by a machine learning method using a second loss function. The second loss function is defined to consider the non-Gaussianity of the LLR distribution more deeply than the first loss function. The first set is used when the coding rate is a first value or when the modulation order is a second value. The second set is used when the coding rate is lower than the first value or when the modulation order is greater than the second value. The method according to any one of Appendices 1 to 6. (Appendix 8) The method further includes generating a third set of learned parameters of the iterative algorithm by training the network with a machine learning method using a third loss function. The third loss function is defined to consider the non-Gaussianity of the LLR distribution more deeply than the first loss function. The first set is used when the signal-to-noise ratio (SNR) is a third value. The third set is used when the SNR is smaller than the third value. The method according to any one of Appendices 1 to 7. (Appendix 9) The method further includes generating a second set of learned parameters of the iterative algorithm by training the network with a machine learning method using a fourth loss function. The fourth loss function is defined to consider the non-Gaussianity of the LLR distribution more deeply than the first loss function. The first set is used for the demodulation process at the first transmission, and the fourth set is used for the demodulation process at the retransmission. The method according to any one of Appendices 1 to 8. (Appendix 10) A program comprising a set of instructions that cause the computer system to perform the method according to any one of Appendices 1 to 9 when loaded into the computer system. (Appendix 11) A memory storing one or more sets of learned parameters generated by the method according to any one of Supplementary Notes 1 to 7, at least one processor, and comprising: wherein the at least one processor executes an iterative algorithm using any of the one or more sets on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmission symbols, and performs error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences. configured as such, a receiving device. (Supplementary Note 12) a memory, at least one processor coupled to the memory, and comprising: wherein the at least one processor executes an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmission symbols, and performs error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences. configured as such, wherein the first set of learned parameters is a set of parameters generated by training a network obtained by deploying the iterative algorithm using a machine learning method with a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network. a receiving device. (Supplementary Note 13) wherein the at least one processor is configured to select between the first set and a second set of learned parameters for use in the iterative algorithm according to at least one of the modulation order and the coding rate. The receiving device according to Supplementary Note 12. (Appendix 14) The second set is generated by training the network by a machine learning method using a second loss function, The second loss function is defined to consider the non-Gaussianity of the LLR distribution deeper than the first loss function, The first set is used when the coding rate is a first value or when the modulation order is a second value, The second set is used when the coding rate is lower than the first value or when the modulation order is larger than the second value, The receiving device according to Appendix 13. (Appendix 15) The at least one processor is configured to select between the first set and a third set of learned parameters for use in the iterative algorithm according to the Signal to Noise Ratio (SNR), The receiving device according to any one of Appendices 12 to 14. (Appendix 16) The third set is generated by training the network by a machine learning method using a third loss function, The third loss function is defined to consider the non-Gaussianity of the LLR distribution deeper than the first loss function, The first set is used when the SNR is a third value, The third set is used when the SNR is smaller than the third value, The receiving device according to Appendix 15. (Appendix 17) The at least one processor is configured to use the first set for the demodulation process at the first transmission and use a fourth set of learned parameters for the demodulation process at the retransmission, The receiving device according to any one of Appendices 12 to 16. (Appendix 18) The fourth set is generated by training the network by a machine learning method using a fourth loss function, The fourth loss function is defined to consider the non-Gaussianity of the LLR distribution deeper than the first loss function. The receiving device according to Supplementary Note 17. (Supplementary Note 19) Executing an iterative algorithm using the first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmission symbols, and Performing error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences, comprising The first set of the learned parameters is a parameter set generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that considers the non-Gaussianity of the LLR distribution calculated from the output of the network. A method performed by a receiving device. (Supplementary Note 20) A program for causing a processor included in a receiving device to perform a method, The method includes Executing an iterative algorithm using the first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmission symbols, and Performing error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences, comprising The first set of the learned parameters is a parameter set generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that considers the non-Gaussianity of the LLR distribution calculated from the output of the network. Program.

Explanation of Signs

[0092] 1 Base station 2 Wireless terminal 301 RF transceiver 303 Network interface 304 Processor 305 Memory 400 BP detector 450 Learned parameter set 460 LLR generator 470 Demodulator 610 Soft interference canceller 620 Belief generator 630 Soft replica generator 910 Training data set 920 Training system

Claims

1. A method implemented in a computer system, comprising: training a network obtained by deploying an iterative algorithm for demodulation or demodulation and decoding, using a machine learning technique with a first loss function that takes into account the non-Gaussianity of the log likelihood ratio (LLR) distribution calculated from the output of the network; generating, by said training, a first set of learned parameters of said iterative algorithm; and generating a second set of learned parameters of said iterative algorithm by training said network using a machine learning technique with a second loss function; wherein said second loss function is defined to take into account the non-Gaussianity of said LLR distribution more deeply than said first loss function; said first set is used when the coding rate is a first value or the modulation order is a second value; said second set is used when the coding rate is lower than said first value or the modulation order is greater than said second value; a method.

2. A method implemented in a computer system, comprising: training a network obtained by deploying an iterative algorithm for demodulation or demodulation and decoding, using a machine learning technique with a first loss function that takes into account the non-Gaussianity of the log likelihood ratio (LLR) distribution calculated from the output of the network; generating, by said training, a first set of learned parameters of said iterative algorithm; and generating a third set of learned parameters of said iterative algorithm by training said network using a machine learning technique with a third loss function; wherein said third loss function is defined to take into account the non-Gaussianity of said LLR distribution more deeply than said first loss function; said first set is used when the signal-to-noise ratio (SNR) is a third value; said third set is used when the SNR is less than said third value; a method.

3. A method implemented in a computer system, comprising: Training a network obtained by developing an iterative algorithm for demodulation or demodulation and decoding using a machine learning method with a first loss function that takes into account the non-Gaussianity of the log likelihood ratio (LLR) distribution calculated from the output of the network, generating, by said training, a first set of learned parameters of said iterative algorithm, and generating, by training said network using a machine learning method with a fourth loss function, a fourth set of learned parameters of said iterative algorithm, comprising, wherein said fourth loss function is defined to take into account the non-Gaussianity of said LLR distribution deeper than said first loss function, wherein said first set is used for demodulation processing at the first transmission in hybrid automatic repeat request (HARQ), and said fourth set is used for demodulation processing at the retransmission in said HARQ, method.

4. A program comprising a set of instructions that, when loaded into a computer system, cause the computer system to perform the method according to any one of claims 1 to 3.

5. A memory storing one or more sets of learned parameters generated by the method according to any one of claims 1 to 3, at least one processor, comprising, wherein said at least one processor executes an iterative algorithm using any of said one or more sets on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmitted symbols, performs error correction decoding using said plurality of LLR vectors to generate a plurality of decoded bit sequences, configured as receiver.

6. A memory, at least one processor coupled to said memory, comprising, wherein said at least one processor executes an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmitted symbols, performs error correction decoding using said plurality of LLR vectors to generate a plurality of decoded bit sequences, configured as, The first set of the learned parameters is a set of parameters generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network. The at least one processor is configured to select between the first set and a second set of learned parameters for use in the iterative algorithm according to at least one of the modulation order and the coding rate. A receiving device.

7. The second set is generated by training the network by a machine learning method using a second loss function. The second loss function is defined to consider the non-Gaussianity of the LLR distribution deeper than the first loss function. The first set is used when the coding rate is a first value or when the modulation order is a second value. The second set is used when the coding rate is lower than the first value or when the modulation order is greater than the second value. The receiving device according to claim 6.

8. A memory and at least one processor coupled to the memory, comprising: The at least one processor is configured to execute an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmitted symbols, perform error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences. configured as The first set of the learned parameters is a set of parameters generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network. The at least one processor is configured to select between the first set and a third set of learned parameters for use in the iterative algorithm according to the signal-to-noise ratio (SNR). A receiving device.

9. The third set is generated by training the network by a machine learning method using a third loss function, The third loss function is defined to consider the non-Gaussianity of the LLR distribution deeper than the first loss function, The first set is used when the SNR is a third value, The third set is used when the SNR is smaller than the third value, The receiving device according to claim 8.

10. A memory, At least one processor coupled to the memory, Comprising, The at least one processor, Executes an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmitted symbols, Performs error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences, Is configured as, The first set of the learned parameters is a set of parameters generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that considers the non-Gaussianity of the LLR distribution calculated from the output of the network, The at least one processor is configured to use the first set for demodulation processing at the first transmission in hybrid automatic repeat request (HARQ) and use a fourth set of learned parameters for demodulation processing at the retransmission in the HARQ, Receiving device.

11. The fourth set is generated by training the network by a machine learning method using a fourth loss function, The fourth loss function is defined to consider the non-Gaussianity of the LLR distribution deeper than the first loss function, The receiving device according to claim 10.

12. A method performed by a receiving device, Executing an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmitted symbols, and Performing error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences, Comprising, The first set of the learned parameters is a set of parameters generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network. The method further comprises selecting between the first set and a second set of learned parameters for use in the iterative algorithm according to at least one of the modulation order and the coding rate. Method. A method performed by a receiving device according to claim 13, executing an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmitted symbols, and performing error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences. comprising The first set of the learned parameters is a set of parameters generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network. The method further comprises selecting between the first set and a third set of learned parameters for use in the iterative algorithm according to the signal-to-noise ratio (SNR). Method. A method performed by a receiving device according to claim 14, executing an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmitted symbols, and performing error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences. comprising The first set of the learned parameters is a set of parameters generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network. The method further comprises using the first set for demodulation processing at the first transmission in hybrid automatic repeat request (HARQ), and using a fourth set of learned parameters for demodulation processing at retransmission in the HARQ. Method. **Claim 15** A program for causing a processor included in a receiving device to perform a method, wherein the method executes an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmission symbols, and performs error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences. The method comprises: The first set of learned parameters is a parameter set generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network. The method further comprises selecting between the first set and a second set of learned parameters for use in the iterative algorithm according to at least one of the modulation order and the coding rate. Program. **Claim 16** A program for causing a processor included in a receiving device to perform a method, wherein the method executes an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmission symbols, and performs error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences. The method comprises: The first set of learned parameters is a parameter set generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network. The method further comprises selecting, according to a Signal to Noise Ratio (SNR), between the first set and a third set of learned parameters for use in the iterative algorithm. Program. [

17. ] A program for causing a processor included in a receiving device to perform a method, wherein the method performs an iterative algorithm using a first set of learned parameters on a plurality of received signals to generate a plurality of log likelihood ratio (LLR) vectors corresponding to a plurality of transmission symbols, and performs error correction decoding using the plurality of LLR vectors to generate a plurality of decoded bit sequences, and the first set of learned parameters is a parameter set generated by training a network obtained by deploying the iterative algorithm by a machine learning method using a first loss function that takes into account the non-Gaussianity of the LLR distribution calculated from the output of the network, the method further comprises using the first set for demodulation processing at the first transmission in hybrid automatic repeat request (HARQ) and using a fourth set of learned parameters for demodulation processing at the retransmission in the HARQ. Program.

Images