Scalable and fast waveform learning in multi-user communication system

Abstract

Various example embodiments relate to an apparatus and method for supporting scalable and fast waveform learning in a multi-user communication system. A network device may generate a first training signal using a transmitter neural network having a first set of weight values trained for at least one first terminal device, and transmit the first training signal to at least one second terminal device. The network device may update the first set of weight values based on loss and / or gradient values received from the at least one second terminal device to obtain a second set of weight values and calculate a first difference between the first set of weight values and the second set of weight values. In response to the first difference being less than or equal to a first threshold, the network device may select the first set of weight values for the transmitter neural network.

Description

Technical Field

[0001] Various example embodiments described herein relate generally to communication technology, and more particularly to apparatus and methods for supporting scalable and fast waveform learning in a multi-user communication system. Background Art

[0002] Certain abbreviations that may be found in the specification and / or drawings are defined as follows:

[0003] AI Artificial Intelligence

[0004] BCE Binary Cross Entropy

[0005] BPSK Binary Phase Shift Keying

[0006] BS Base Station

[0007] E2E End to End

[0008] GS Geometry Shaping

[0009] I / Q In-Phase / Quadrature

[0010] LLR log-likelihood ratio

[0011] MAP Maximum A Posteriori

[0012] MCS modulation and coding scheme

[0013] ML Machine Learning

[0014] MU Multi-User

[0015] NN Neural Network

[0016] OFDM Orthogonal Frequency Division Multiple Access

[0017] QAM Quadrature Amplitude Modulation

[0018] QPSK Quadrature Phase Shift Keying

[0019] ResNet residual network

[0020] Rx Receiver

[0021] Rx-NN Receiver Neural Network

[0022] TMM Trainable Modulation Mapper

[0023] Tx Transmitter

[0024] Tx-NN Transmitter Neural Network

[0025] UE User Equipment

[0026] Conventional signal processing design in wireless communication systems involves the optimization of multiple individual modules at transmitters such as base stations (BS) and receivers such as user equipment (UE). However, optimization of each module does not necessarily lead to global optimization of the entire communication system. With the development of artificial intelligence (AI) and machine learning (ML) technologies, end-to-end (E2E) learning has shown great promise in improving the overall performance of communication systems because it can jointly train transmitters and receivers. For 5G and future radio systems that are envisioned to operate at very high carrier frequencies and be as energy efficient as possible, E2E learning will play an important role in the overall optimization of the system. Summary of the invention

[0027] A brief summary of each exemplary embodiment is provided below to provide a basic understanding of some aspects of various embodiments. It should be noted that this summary is not intended to identify the key features of each basic element or define the scope of each embodiment, and its only purpose is to introduce some concepts in a simplified form as a prelude to the more detailed description provided below.

[0028] In a first aspect, an example embodiment of a network device is provided. The network device may include at least one processor and at least one memory including computer program code. The at least one memory and the computer program code may be configured to, with the at least one processor, cause the network device to at least: generate a first training signal from first training bits using a transmitter neural network, the transmitter neural network having a first set of weight values ​​trained for at least one first terminal device; and transmit the first training signal to at least one second terminal device for training a receiver neural network at the at least one second terminal device. The network device may update the first set of weight values ​​of the transmitter neural network based on loss and / or gradient values ​​received from the at least one second terminal device to obtain a second set of weight values ​​for the transmitter neural network, and calculate a first difference between the first set of weight values ​​and the second set of weight values ​​for the transmitter neural network. In response to the first difference being less than or equal to a first threshold, the network device may select the first set of weights for the transmitter neural network.

[0029] In a second aspect, an example embodiment of a terminal device is provided. The terminal device may include at least one processor and at least one memory including computer program code. The at least one memory and the computer program code may be configured to utilize the at least one processor to cause the terminal device to at least receive a first training signal from a network device and train a receiver neural network using the first training signal until the receiver neural network converges. The receiver neural network is configured to recover a first training bit from the first training signal. The terminal device may, in response to convergence of the receiver neural network, send a first loss and / or a corresponding gradient value calculated based on the recovered first training bit to the network device.

[0030] Example embodiments of methods, apparatuses, and computer program products are also provided. Such example embodiments generally correspond to the example embodiments described above, and for convenience, a repeated description thereof is omitted here.

[0031] Other features and advantages of example embodiments of the present disclosure will also be apparent from the following description of specific embodiments when read in conjunction with the accompanying drawings, which illustrate, by way of example, the principles of example embodiments of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Some example embodiments will now be described, by way of non-limiting examples, with reference to the accompanying drawings.

[0033] Figure 1 is a schematic diagram illustrating conventional digital signal processing (DSP) blocks in a communication system.

[0034] Figure 2 is a schematic diagram showing a neural network based DSP block in a communication system.

[0035] Figure 3 is a schematic diagram illustrating an E2E learning framework for a communication system.

[0036] Figure 4 is a schematic diagram illustrating a scalable E2E learning framework for a multi-user communication system according to an example embodiment of the present disclosure.

[0037] Figure 5 is a schematic flow chart illustrating an E2E learning process according to an example embodiment of the present disclosure.

[0038] Fig. 6A is a schematic block diagram illustrating an example implementation of a transmitter neural network according to an example embodiment of the present disclosure.

[0039] Figure 6B is a schematic block diagram illustrating an example implementation of a receiver neural network according to an example embodiment of the present disclosure.

[0040] Figure 7 is a schematic message sequence chart illustrating the first stage of an E2E learning process according to an example embodiment of the present disclosure.

[0041] Figure 8 is a schematic message sequence chart illustrating the second stage of the E2E learning process according to an example embodiment of the present disclosure.

[0042] Fig. 9 is a schematic block diagram illustrating a device according to an example embodiment of the present disclosure.

[0043] Fig.10 is a schematic block diagram illustrating a device according to an example embodiment of the present disclosure.

[0044] Fig.11 is a diagram showing that the loss decreases as the number of iterations increases.

[0045] Fig.12 is a schematic block diagram illustrating devices in a communication system in which example embodiments of the present disclosure may be implemented.

[0046] Throughout the drawings, the same or similar reference numerals denote the same or similar elements, and repeated description of the same elements will be omitted. DETAILED DESCRIPTION

[0047] Below, some example embodiments are described in detail with reference to the accompanying drawings. The following description includes specific details intended to provide a thorough understanding of the various concepts. However, it will be apparent to those skilled in the art that these concepts can be practiced without these specific details. In some cases, well-known circuits, techniques, and components are shown in block diagram form to avoid obscuring the concepts and features described.

[0048] As used herein, the term "network equipment" refers to any suitable equipment or entity that can provide a cell or coverage, through which a terminal device can access a network or receive a service. Network equipment may generally be referred to as a base transceiver station (BTS), a base station (BS), or some other suitable term. The term "base station" or "base transceiver station" used herein may represent a node B (NodeB or NB), an evolved node B (eNodeB or eNB), a next generation node B (gNB), or a next generation enhanced node B (ng-eNB). A base station may be embodied as a macro base station, a relay node, or a low power node such as a micro base station or a femto base station. A base station may also include or may be referred to as a RAN (radio access network) node, and may be composed of several distributed network units such as a central unit (CU), one or more distributed units (DU), one or more remote radio heads (RRH), or a remote radio unit (RRU). The number and functionality of these distributed units depends on the selected decomposed RAN architecture.

[0049] As used herein, the term "terminal device" or "user equipment" (UE) refers to any device or entity that can communicate wirelessly with network devices or each other. Examples of terminal devices may include mobile phones, mobile terminals, mobile stations (MS), subscriber stations, portable subscriber stations, access terminals, personal digital assistants (PDAs), computers, wearable devices, vehicle-mounted communication devices, machine type communication (MTC) devices, D2D communication devices, V2X communication devices, sensors, etc. The term "terminal device" can be used interchangeably with UE, user terminal, mobile terminal, mobile station, or wireless device.

[0050] Conventional communication systems break down signal processing into a chain of multiple independent blocks at the transmitter (eg, base station) and receiver (eg, user equipment (UE)), respectively, with each block individually optimized for different functions. Figure 1 FIG. 1 is a block diagram of a conventional communication system 100A. Figure 1As shown in , a transmitter (e.g., a base station 110) may include a coding block 112, a constellation modulation block 114, a resource mapping block 116, and an OFDM modulation block 118. The coding block 112 may perform source coding and channel coding on the source bits. For example, the coding block 112 may compress the source bits and add redundancy in a controlled manner to cope with the negative effects of the communication medium. The constellation modulation block 114 may modulate a block of binary bits into a complex-valued modulation symbol corresponding to a constellation point. The length (number of bits) of each bit block may depend on the modulation algorithm used in the constellation modulation block 114. The resource mapping block 116 may map the modulation symbols to resource elements (REs) in a time-frequency grid. The same modulation symbol may be mapped to different resource elements and thus undergo different wireless transmissions. The OFDM modulation block 118 may modulate the frequency domain OFDM subcarriers into a time domain OFDM signal based on OFDM parameters such as a fast Fourier transform (FFT) length, a cyclic prefix (CP) length, a guard period (GP), etc. The OFDM signal is then transmitted by a transmitter antenna over an air channel 101 to a receiver, such as a user equipment (UE) 120 .

[0051] The above operations are reversed at the UE 120 to reconstruct the source bits sent by the transmitter 110. For example, the OFDM demodulation block 128 can demodulate the time domain OFDM samples and output subcarriers based on the OFDM parameters. The channel estimation block 127 can perform channel estimation using the input data and the reference subcarriers and send channel state information (CSI) back to the transmitter 110. The equalization block 126 can equalize the OFDM data based on the channel estimation. For example, the equalization block 126 can use zero-forcing (ZF) or minimum mean square error (MMSE) algorithms to perform channel equalization in the frequency domain. The constellation demodulation block 124 can receive the equalized complex data symbols and demodulate each complex data symbol into a set of log-likelihood ratio (LLR) values, which indicate the probability that the binary bit is "0" or "1". The LLR values ​​are fed to the decoding block 122 that recovers the source bits from the LLR values.

[0052] When a signal is transmitted at the base station 110 and received at the UE 120, noise is added to the signal due to hardware damage at the base station 110 and the UE 120. In addition, the air channel 101 also distorts and attenuates the signal. Therefore, it is necessary to optimize the above-mentioned signal processing blocks at the base station 110 and the UE 120 to ensure reliable signal transmission while maximizing the air interface capacity and efficiency of the system. However, the individual optimization of the signal processing block does not necessarily optimize the entire communication system. In this regard, the joint design of each communication block is conducive to the overall performance improvement of the communication system. Inspired by the joint design paradigm and AI / ML technology, it has been proposed to use neural networks to replace one or more signal processing blocks in the transmitter and receiver.

[0053] Figure 2 A DSP block based on a neural network in a communication system 100B is shown. Compared to the communication system 100A, the constellation modulation block 114 is replaced by a geometric shaping (GS) block 113 (also known as a trainable modulation mapper (TMM) block) in the base station 110, and the constellation demodulation block 124, the equalization block 126, and the channel estimation block 127 are replaced by a deep learning reception (DeepRx) block 123 in the UE 120. The GS / TMM block 113 may be implemented as or execute a transmitter neural network (Tx-NN) that maps a block of binary bits to a complex-valued modulation symbol having a real component and an imaginary component. The DeepRx block 123 may be implemented as or execute a receiver neural network (Rx-NN) that receives an OFDM symbol (i.e., a time-frequency grid symbol) from an OFDM demodulation block 128 and recovers the transmitted binary bit from the received OFDM symbol. In one example, the Rx-NN may generate an LLR value indicating the probability that each recovered bit is "0" or "1".

[0054] It should be understood that the GS / TMM block 113 and the DeepRx block 123 are described herein as examples, and that the Tx-NN and Rx-NN may be implemented to implement more or less functionality. For example, the Tx-NN may be implemented to further enable at least a portion of the encoding, resource mapping, and OFDM modulation functionality, and the Rx-NN may be implemented to further enable at least a portion of the OFDM demodulation and decoding functionality.

[0055] Figure 3 is a schematic diagram showing an E2E learning framework for the communication system 100B. Figure 3 In the example shown, the overall performance of the communication system 100B is improved by jointly training the Tx-NN and the Rx-NN as an autoencoder, rather than optimizing individual blocks at the base station 110 and the UE 120. An autoencoder is a neural network consisting of an encoder that learns a (latent) representation of a given data and a decoder that reconstructs the input data from the encoded data. In this setting, the Tx-NN at the base station 110 can be considered as an encoder, and the Rx-NN at the UE 120 can be considered as a decoder.

[0056] In E2E learning, a plurality of blocks of training bits are first obtained, for example, from a uniform distribution (UD). The blocks may have a block size / length Q (the number of bits in the block), which is determined by the constellation modulation scheme employed at the Tx-NN. For example, for the constellation modulation scheme 16QAM, the block size Q is 4, and every 4 training bits form a bit block, which will be mapped to a complex-valued modulation symbol having a real component and an imaginary component. The real and imaginary components may define points called constellation points in a two-dimensional coordinate system having a horizontal axis representing the real component and a vertical axis representing the imaginary component. There are a total of 2 Qblock types, corresponding to 2 Q constellation points. 2 Q The constellation points may refer to a constellation pattern corresponding to a constellation modulation scheme. The total number of bit blocks may depend on, for example, the bandwidth, the modulation scheme, and related configurations for communication between the base station 110 and the UE 120. It should be understood that Tx-NN may also employ other modulation algorithms, such as BPSK, QPSK, 8QAM, 64QAM, and 256QAM, and the size of the bit block will change accordingly.

[0057] Prior to training, the Tx-NN and Rx-NN may be initialized with random states. For example, for a 16QAM Tx-NN, 16 constellation points are randomly generated as initial weight values ​​for the Tx-NN, and the real and imaginary coordinate values ​​of these constellation points are represented by a pair of real data values ​​(i.e., 16 pairs of real data values). The Rx-NN may also be initialized with a set of randomly generated weight values. The Tx-NN receives a training bit block as input data and maps the bit block to a complex-valued modulation symbol (i.e., a pair of real data values) corresponding to these constellation points. The modulation symbol may be further processed, for example, by resource mapping and OFDM modulation, and then transmitted to the UE 120 on an OFDM channel. Figure 3 A transmitter hardware (Tx HW) block is shown, which represents the various hardware implemented at the base station 110 for processing and transmitting signals.

[0058] UE 120 receives the signal and constructs input data for Rx-NN. For example, UE 120 may demodulate the received signal into OFDM symbols (i.e., time-frequency grid symbols) and concatenate the OFDM symbols with additional information such as noise information, positioning coding levels. The OFDM symbols may then be fed to Rx-NN to recover the transmitted bits. Rx-NN may generate LLR values ​​to indicate the probability distribution of possible bits recovered from the OFDM symbols. Figure 3 Also shown is a receiver hardware (Rx HW) block, which represents the various hardware implemented at the UE 120 for receiving and processing signals.

[0059] The LLR value may be compared with the transmitted bits (i.e., training bits) to calculate the training loss, for which a predetermined loss function may be used. UE 120 may receive an indication of the transmitted bits from base station 110 in advance. The weight values ​​of Tx-NN and Rx-NN may be updated based on the loss gradient using a back propagation (BP) algorithm. The above operations (except for the initialization of the neural network) may be iteratively performed until the training loss reaches a minimum value and no longer decreases.

[0060] The above-described E2E learning framework gives the communication system 100B the ability to jointly optimize components to achieve the best overall performance, rather than the suboptimal single-block-based signal processing design in conventional systems. By minimizing the loss function at the UE 120, the Tx-NN and Rx-NN are mutually adaptively optimized, thereby exploring the data transmission capabilities for jointly optimizing the transmission rate under complex air channel conditions and hardware constraints between terminals. In addition, the system is able to learn and combat hardware impairments and some limitations such as limited quantization, nonlinearity, and receiver IQ imbalance in an E2E manner under certain constraints. Such a design is also flexible, low in complexity, and can be adapted to specific applications and scenarios.

[0061] The E2E learning framework also has some disadvantages. For example, the current E2E learning framework trains the transmitter and receiver in a 1:1 (one transmitter to one receiver) manner, which lacks scalability. Intuitively, in a multi-user (MU) communication system where the base station 110 serves K UEs 120 in a cell, the base station 110 needs to maintain K different Tx-NNs to coordinate with the K UEs 12, and must perform K times (or more times for new UEs) of 1:1 E2E learning. This is time-consuming and increases the complexity of the system.

[0062] Considering another scenario where a common Tx-NN is trained at a base station 110 to serve multiple UEs 120a and a new UE 120b joins the system, the trained common Tx-NN may not be suitable for the new UE 120b because the new UE 120b may have different hardware impairments and channel conditions from the multiple initial UEs 120a. If the common Tx-NN is further trained to coordinate with the new UE 120b, it will affect the waveform that tends to the multiple initial UEs 120a. An intuitive way to adjust the common Tx-NN for all UEs 120a, 120b is to train the common Tx-NN and Rx-NN at UEs 120a, 120a (collectively referred to as UE 120) from scratch. Since E2E learning based on an autoencoder is time-consuming, E2E trained from scratch requires a long training time and a large number of iterations. During training, frequent data exchanges and loss reports between the base station 110 and the UE 120 will result in a large amount of data traffic, which is not ideal for air training.

[0063] Hereinafter, various exemplary embodiments of devices, methods, and computer program products that support a scalable and fast E2E learning process will be described in detail with reference to the accompanying drawings. In some exemplary embodiments, a public Tx-NN trained for one or more initial UEs can be reused in a supervised learning manner to train a newly joined UE, which can be expected to reduce data exchange and speed up training. Since the initial UE is not trained during the training of the new UE, the service provided to the initial UE may be interrupted. When the training of the new UE is completed, a criterion may be applied to determine whether the public Tx-NN is applicable to the new UE. If the public Tx-NN is not applicable to the new UE, the previously trained public Tx-NN and Rx-NN can be used as the starting neural network to train the base station and all UEs including the initial UE and the new UE. Compared with training from scratch, the proposed method can greatly shorten the training time and reduce the data exchange between the base station and the UE. As the number of UEs in the system increases, the public Tx-NN becomes more universal, and the fluctuations caused by the newly joined UE to the initial UE are attenuated, which means that the stability of this method grows with the expansion of the system scale.

[0064] Figure 4 A scalable E2E learning framework for a multi-user communication system according to an example embodiment of the present disclosure is shown. Figure 4 First, a common Tx-NN is trained at the base station 110 to communicate with multiple UEs 120a (UE 1 To UE K ) at a series of Rx-NN (Rx-NN 1 To Rx-NN K ) coordination. The common Tx-NN is trained to compensate for impairments caused by common Tx-side hardware such as nonlinear power amplifiers, and the Rx-NN is responsible for its individual channel and Rx-side issues. A series of Rx-NNs (Rx-NN 1 To Rx-NN K ) can be trained simultaneously or by using the scalable E2E learning process proposed in the present disclosure, the details of which will be described below. K+1 ) joins the system, the pre-trained public Tx-NN can be used to accelerate the training of new UEs. K+1 120b Rx-NN K+1 The proposed E2E learning framework can be extended to 1:K learning and 1:(K+1) learning, which improves the training efficiency. The details of the proposed scalable E2E learning framework are described below.

[0065] Figure 5A scalable E2E learning process 200 according to an example embodiment of the present disclosure is shown. The process 200 may be performed by a transmitter (e.g., the base station 110 described above) and multiple receivers (e.g., the UE 120 described above). The base station 110 and the UE 120 may include multiple devices, modules, components, or elements for performing the operations in the process 200, and these devices, modules, or elements may be implemented in various ways including, but not limited to, software, hardware, firmware, or any combination thereof.

[0066] At the beginning of the process 200, it is assumed that a common Tx-NN (K) has been well trained for K active UEs 120a at the base station 110, where K is an integer greater than or equal to 1. The K UEs 120a also have well-trained Rx-NN (Rx-N N1 To Rx-N NK ). The K UEs 120a may be trained simultaneously or by using the scalable E2E learning process 200 described herein. For example, first, one UE may be trained using a conventional 1:1 E2E learning process, and then one or more new UEs may be trained by the proposed scalable E2E learning method 200. The process 200 may be performed more than once until the K UEs 120 are trained.

[0067] At 210, when a new UE (ie, UE K+1 When a new UE 120a joins the system, the base station 110 may start training the new UE by loading / reusing the common Tx-NN(K) that has been trained for the K initial UEs 120a. At this time, Tx-NN(K) has a first set of weight values.

[0068] At 212, the base station 110 may use the common Tx-NN(K) to generate a training signal from the training bits. It is noteworthy that the training signal is generated for the new UE 120b, rather than for the trained K initial UEs 120a. In the first stage of the process 200, the K initial UEs 120a do not need to be trained. The training bits may be a pseudo-random sequence including uniformly distributed (UD) bits, and the UD bits may be provided as input data to the common Tx-NN(K).

[0069] Fig. 6A An example Tx-NN implemented at the base station 110 (e.g., in the GS / TMM block 113) according to an example embodiment of the present disclosure is shown. The Tx-NN may have a mapping layer 131, a convolution layer 133, and a subsequent real-to-complex (R2C) conversion layer 135. The mapping layer 131 may receive training bits (e.g., UD bits) as input data and map each training bit block to a one-hot vector. The bit block may have a block size Q (the number of bits in the block) and a total of 2 Qdifferent blocks of bits. A one-hot vector is 2 Q dimension vector, where only one element is 1 and the rest (2 Q -1) element is 0, and there are 2 Q different one-hot vectors. Therefore, 2 Q Different bit blocks can be mapped to 2 Q The mapping layer 131 receives the training bits and maps each bit block to a corresponding one-hot vector.

[0070] The convolutional layer 133 may include 2 Q trainable weights w i (i=1,...,2 Q , and each weight w i The number Q comprises trainable coordinate pairs (i.e., a pair of real data values) representing constellation points. The number Q is determined by the constellation modulation scheme employed at the base station 110. For example, for a constellation modulation scheme 16QAM using a constellation pattern including 16 constellation points, the number Q is 4, and the convolutional layer 133 comprises 16 coordinate pairs. At step 212, the 16 coordinate pairs of the convolutional layer 133 have been trained for K UEs 120a. The base station 110 may also employ other constellation modulation schemes, such as BPSK, QPSK, 8QAM, 64QAM, and 256QAM, and the number Q may change accordingly. For example, the number Q may be 1 for BPSK and 2 for QPSK. The convolutional layer 133 performs the one-hot vector and 2 Q Weight w i The convolution operation is performed between them and the 2 corresponding to the element "1" in the one-hot vector is output. Q Weight w i One of the output weights w i Can consist of a pair of real coordinate values.

[0071] The R2C conversion layer 135 may convert the pair of real coordinate values ​​into a complex-valued modulation symbol as output data of the common Tx-NN. The complex-valued modulation symbol has a real (in-phase) component and an imaginary (quadrature) component, which correspond to the output weights w i Therefore, Tx-NN can modulate the training bit block into a complex-valued modulation symbol. It should be understood that Figure 5 A only shows an example of Tx-NN, and Tx-NN may also use other neural network models. For example, in an example embodiment, Tx-NN may also include a normalization layer after the R2C conversion layer 135 to ensure that certain physical constraints on the Tx signal are met.

[0072] Return to reference Figure 5At step 212, after mapping the training bit block to the complex-valued modulation symbol using Tx-NN(K), the base station 110 may perform further processing on the complex-valued modulation symbol including, for example, resource mapping and OFDM modulation, thereby generating a signal to be transmitted to the UE. K+1 120b training signal. The total number of bit blocks and corresponding modulation symbols may depend on the bandwidth, modulation scheme, and related configuration. In some example embodiments, at least a portion of further processing such as resource mapping and OFDM modulation may also be performed in Tx-NN(K).

[0073] At 214, base station 110 may send a K+1 At step 214, the base station 110 does not need to transmit a training signal to the K initial UEs 120a, because it is not necessary to train the K initial UEs 120a at this stage.

[0074] At 216, UE K+1 120b may use the training signal received from the base station 110 to train the UE K+1 Rx-NN implemented at 120a K+1 (eg in DeepRx block 123).

[0075] Figure 6B An example Rx-NN implemented at the DeepRx block 123 of the UE 120 according to an example embodiment of the present disclosure is shown. In the example shown, the Rx-NN may be implemented as a residual neural network including a complex to real (C2R) conversion layer 141 and a plurality of residual network (ResNet) blocks 143. Figure 6B Three ResNet blocks 143a, 143b, and 143c are shown, but the Rx-NN may include more or fewer ResNet blocks. Although not shown, each of the ResNet blocks 143 may include one or more convolutional layers.

[0076] Before training, UE K+1 120b Rx-NN K+1 In an example embodiment, the UE K+1 120b can initialize Rx-NN with a set of random weight values K+1 In another example embodiment, the UE K+1 120b may initialize the Rx-NN with a set of initial weight values ​​received from the base station 110 K+1 For example, if the base station 110 finds that a UE (one of the K initial UEs 120a or another UE) is already in communication with the UE K+1 120b has trained its Rx-NN under substantially the same or similar conditions, for example, the UE hasK+1 120b has the same or similar capabilities and / or is located with the UE K+1 120b, the base station 110 may assume that the Rx-NN trained at the UE is at least partially applicable to the newly joined UE. K+1 120b. Then, the base station 110 may request the trained Rx-NN weight value from the UE and send the trained weight value to the UE. K+1 120b for Rx-NN K+1 Initialization.

[0077] As mentioned above about Figure 2 As mentioned above, UE K+1 120b can demodulate the received training signal into OFDM symbols (i.e., time-frequency grid symbols) in OFDM demodulation block 128 and provide Rx-NN K+1 Construct input data. In one example, UE K+1 120b can concatenate the OFDM symbols with additional information such as noise information, positioning code, etc. Then, the OFDM symbols can be fed to the Rx-NN K+1 . Rx-NN K+1 The C2R conversion layer 141 may convert a complex-valued OFDM symbol into a two-dimensional real-valued tensor having a shape of N(2+Na), where N rows correspond to the number of received OFDM symbols in a batch, 2 columns correspond to recovered constellation points (coordinate pairs) obtained by extracting the real and imaginary components of the complex-valued OFDM symbol, and Na columns correspond to additional information such as noise and positioning information concatenated to the OFDM symbol. For an OFDM system in which a multi-carrier waveform is used for a multipath channel, noise information (e.g., channel state information) helps to recover the transmitted data bits.

[0078] The N×(2+Na) tensors are processed in multiple ResNet blocks 143, and the last ResNet module 143c may output a two-dimensional real-valued tensor of shape N×Q, which corresponds to the log-likelihood ratio (LLR) value of the recovered training bit. The LLR value indicates the probability distribution of possible bits recovered from the OFDM codeword. It should be understood that Figure 5 B only shows an example of Rx-NN, and Rx-NN may also use other neural network models. For example, in an example embodiment, Rx-NN may also include a maximum activation layer after ResNet block 143, which selects the possible bit with the highest probability as the output data of Rx-NN.

[0079] Return to reference Figure 5 , at step 216, when Rx-NN K+1 When the LLR value of the recovered training bit is obtained, the UE K+1120b may compare the LLR value with the transmitted training bits and calculate the training loss based on a predetermined loss function. K+1 120b may receive an indication of the training bits from the base station 110 in advance via a control channel. In one example, a binary cross entropy (BCE) loss function may be used to calculate the training loss. BCE loss It can be calculated as follows:

[0080]

[0081] Where S is the number of training batches, l is the index of the associated batch, and n D is the index set of modulation symbols carrying training bits, b k,i is the i-th bit in the k-th modulation symbol, Q k,i (b k,i |y) is the b given the received codeword y k,i is the approximate maximum a posteriori (MAP) probability of , which can be calculated from the associated LLRs, and E is the expectation function.

[0082] Based on the calculated training loss UE K+1 120b can update Rx-NN K+1 The trainable weights are such that the training loss In one example, Rx-NN K+1 The trainable weights w j It can be updated as follows:

[0083]

[0084] where γ is the learning rate, Based on the training loss The calculated weight w j gradient.

[0085] It should be understood that transmitting / receiving training signals, calculating training losses, and updating Rx-NN K+1 The step of trainable weights can be iterated multiple times until the training loss reaches a minimum and no longer decreases, which means that Rx-NN K+1 Converged (ie, well trained). Since the Tx-NN (K) trained for the K initial UEs 120a is used to supervise the UE K+1 Rx-NN at 120b K+ 1 training, so it is expected to accelerate the training. K+1 During the iterative training, UE K+1120b does not need to report the training loss to the base station 110, and the overhead on the air interface can be reduced. In addition, when the base station 110 assigns the weight value trained under similar conditions to the UE K+1 120b Rx-NN K+1 The training time can be further reduced.

[0086] At 218, when Rx-NN K+1 When the training is completed, UE K+1 120b may report its training loss or equivalent descent gradient to base station 110. Base station 110 may update the trainable weights w of Tx-NN(K) based on the training loss or equivalent gradient i It should be noted that the weight w i Each consists of a coordinate pair (x i ,y i ), and each coordinate value x i ,y i It can be updated in a manner similar to the above formula (2). After the update, a second set of weight values, which can be referred to as temp-Tx-NN(K), is obtained.

[0087] At 220, the base station 110 may calculate the difference δ between the first set of weight values ​​(i.e., constellation points) of Tx-NN(K) and the second set of weight values ​​(i.e., constellation points) of temp Tx-NN(K). (K,temp K) In one example, the difference δ (K,temp K) It can be calculated as the normalized mean square error (NMSE) between Tx-NN(K) and temp Tx-NN(K) at each constellation point. Other metrics can also be used for the difference δ (K,temp K) Calculation.

[0088] At 222, the base station 110 may determine the difference δ between Tx-NN(K) and Temp-Tx-NN(K). (K,temp K) Is it less than or equal to a predetermined threshold? (K,temp K) If the value is less than or equal to the threshold, the base station 110 knows that Tx-NN(K) (i.e., the first set of weight values ​​trained for the K initial UEs 120a) is also applicable to the newly joined UEs K+1 120b. At 224, the base station 110 may select Tx-NN(K) as the common transmitter neural network Tx-NN(K+1) for the (K+1) UEs 120 (including 120a and 120b). The process 200 ends at 242. It is worth noting that when the newly joined UE K+1When UE 120b operates under the same or similar conditions as one of the K initial UEs 120a, it is likely that the difference δ between Tx-NN(K) and Temp-Tx-NN(K) (K,temp K) Small, and Tx-NN(K) is suitable for newly added UEs K+1 120b. For new UE K+1 The training of 120b will be completed quickly, and no further training is required for the common Tx-NN at the base station 110 and the Rx-NN at each initial UE 120a. Compared with training the system from scratch, the proposed method can significantly shorten the training time and reduce the number of newly added UEs. K+1 Data exchange for 120b training.

[0089] If the difference δ between Tx-NN(K) and Temp-Tx-NN(K) at 222 (K,temp K) If the value of Tx-NN(K) is greater than the threshold, the base station 110 may determine that Tx-NN(K) (i.e., the first set of weight values ​​trained for the K initial UEs 120a) is not suitable for the newly joined UE. K+1 120b. Then, at 226, the base station 110 may select Temp-Tx-NN(K) (i.e., the second set of weight values ​​obtained by updating the first set of weight values ​​at 218) as the common transmitter neural network Tx-NN(K+1) for the (K+1) UEs 120. In response to the change of the Tx-side neural network from Tx-NN(K) to Tx-NN(K+1), the K initial UEs 120a and the new UEs 120a and 120b may be connected to the base station 110. K+1 The Rx-NN at 120b will need to be further trained in the second stage.

[0090] At 228, base station 110 may generate a training signal from the training bits using Tx-NN(K+1) obtained at 226. Step 228 may be similar to step 212, except that a common Tx-NN(K+1) is used for all UEs 120a and 120b (collectively, UE 120), not just for the new UE. K+1 120b generates a training signal. The training bits used at 228 may be the same as or different from the training bits used at 212. In one example, the training bits of each UE may be determined by the corresponding configuration of each UE. The details of the training signal generation operation may refer to the description of step 212 above, and repeated descriptions are omitted here.

[0091] At 230, the base station 110 may transmit a training signal to each UE 120.

[0092] At 232, the UEs 120 may use the received training signals to train their respective Rx-NNs. K (k=1, ..., k+1). As described above with respect to step 216, UE 120 may construct input data based on the received training signal and feed the input data to the Rx-NN K , by converting the K The output LLR value is compared with the training bit indicated by the base station 110 to calculate the training loss Based on the loss Update Rx-NN K The details of such operation can refer to the above description of step 216, and for convenience, repeated description is omitted here.

[0093] At 234, UE 120 may report the loss to base station 110 and / or an equivalent descent gradient, and the base station 110 may be based on the received training loss Or gradient to update the trainable weight w of Tx-NN(k+1) i In an example embodiment, base station 110 may combine the received training losses as follows or the equivalent gradient:

[0094]

[0095] Where i is the index of the UE, is the loss of the i-th UE at iteration t, K+1 is the total number of UEs, α i (t-1) is the attention factor, which can be calculated based on the loss calculated in the previous iteration as follows:

[0096]

[0097] As can be seen from formula (4), the focus factor α i (t-1) is calculated as a weighted value of the sum of the loss received from a certain UE and the loss received from UE 120 in the last iteration before the current iteration. The equivalent gradient can be combined in a similar manner. Then, the base station 110 can update Tx-NN(K+1) based on the combined loss or the equivalent gradient according to formula (2), and repeated descriptions are omitted here. The updated Tx-NN(K+1) has a third set of weight values, which is referred to as Temp-Tx-NN(K+1) for convenience.

[0098] In an example embodiment, UE 120 may update Rx-NN every time K Report the loss to the base station 110 In this way, the Rx-NN at UE 120K The Tx-NN(K+1) at the base station 110 and the Tx-NN(K+1) at the UE 120 can be regarded as an autoencoder and trained synchronously. Since the pre-trained neural network is used as the starting neural network for training at the base station 110 and the UE 120, it is expected that the training time can be reduced.

[0099] It should be understood that the principle of synchronously training the base station 110 and the UE 120 by combining the training losses received at the base station 110 is also applicable to other scenarios, such as initial training of a common Tx-NN at the base station 110 before step 210, and training a series of Rx-NNs (Rx-NNs) at the K UEs 120a. 1 To Rx-NN K ) for initial training.

[0100] At 236, the base station 110 may calculate the difference δ between the second set of weight values ​​of Tx-NN(K+1) and the third set of weight values ​​of Temp-Tx-NN(K+1). (K+1,temp K+1) In one example, the difference δ (K+1,temp K+1) can be calculated as the constellation point between Tx-NN(K+1) and Temp-Tx-NN(K+1) (i.e., weight w i ) is the normalized mean square error (NMSE).

[0101] At 238, the base station 110 may determine the difference δ between Tx-NN(K+1) and Temp-Tx-NN(K+1). (K+1,temp K+1) Is it less than or equal to a predetermined threshold? (K+1,temp K+1) If the difference δ is less than or equal to the threshold, the base station 110 may determine that any one of Tx-NN(K+1) and Temp-Tx-NN(K+1) is suitable for the UE 120. The base station 110 may use Tx-NN(K+1) or tempTx-NN(K+1) as a common transmitter neural network for the UE 120, and the process 200 ends at 242. (K+1,temp K+1) If the difference δ is greater than the threshold, the base station 110 may use Temp-Tx-NN(K+1) as Tx-NN(K+1) at 240 and iterate steps 228-238 until the difference δ is greater than the threshold. (K+1,temp K+1) becomes less than or equal to the threshold.

[0102] As mentioned above Figure 4 As mentioned above, when the new UE K+1 120b is used for new UE when joining the system K+1The training of 120b can be performed in two stages. In the first stage, only the new UE K+1 120b, and the common Tx-NN at the base station 110 is used to train the new UE in a supervised manner K+1 120b. It can accelerate UE K+1 120b training and reduce air interface overhead. K+1 If the training loss received by 120b causes the common Tx-NN at the base station 110 to become unsuitable for the previously trained UE 120a, a second stage of training may be performed on the entire system including the base station 110 and the UEs 120a, 120b. In the second stage of training, the previously trained neural networks at the base station 110 and the UE 120 are used as the starting neural networks. Compared to training the system from scratch, the proposed method can shorten the training time and reduce data exchange by utilizing the pre-trained neural networks for the first and second stage training. It can be expected that as the number of UEs in the system increases, the common Tx-NN becomes more versatile and the fluctuations caused by the newly added UEs to the initial UEs will be attenuated, which means that as the system scales up, the stability of this method will increase and the training time required for new UEs will be further reduced.

[0103] Although only one new UE 120b is described in the above example embodiments, it should be understood that the example embodiments are also applicable when multiple new UEs join the system. Multiple new UEs may be trained simultaneously in a manner similar to the training of the new UE 120b as described above. During the simultaneous training of multiple new UEs, the loss reports from the multiple new UEs will be combined at the base station 110, and the base station 110 will update its common Tx-NN based on the combined losses or equivalent gradients.

[0104] Figure 7 is a schematic message sequence chart illustrating a first stage 300A of an E2E learning process according to an example embodiment of the present disclosure. Figure 7 The process 300A shown in FIG. 1 may be performed by a base station (such as the base station 110 described above) and a UE (such as the newly added UE described above). K+1 120b) is executed. Figure 5 Some steps in process 300A are discussed and will be described here in a simplified manner.

[0105] refer to Figure 7 At 310, for example, when UE K+1 When UE 120b initially connects to a cell supported by base station 110, K+1 120a may send a training request to the base station 110. The training request may include a K+1120b trains the receiver neural network Rx-NN K+1 , such as UE capabilities, positioning information, hardware and software configuration, Rx-NN model information, etc. In one example, UE K+1 120b may also request the base station 110 to provide the UE with a training request. K+1 Rx-NN at 120a K+1 Assign initial weight values.

[0106] At 312, base station 110 may respond to UE with an acknowledgement message. K+1 120b. The confirmation message may include a K+1 The training configuration may include one or more of an indication of a modulation and coding scheme (MCS), subcarrier information, and an indication of training bits (e.g., a random seed for generating a pseudo-random sequence). The base station 110 may determine the training configuration for the UE based on the training-related information received at 310. K+1 120b training configuration.

[0107] In an example embodiment, the training configuration may also include a K+1 120b Rx-NN K+1 For example, if another UE is already in communication with UE K+1 120b is trained under substantially the same or similar conditions, for example, another user equipment has the same K+1 120b, the base station 110 may obtain the trained Rx-NN weight value from another user equipment and send the obtained weight value to the UE. K+1 120a. It can greatly speed up Rx-NN K+1 training.

[0108] At 314, UE K+1 120b can initialize its Rx-NN K+1 If UE K+1 120b has received the initial weight value assigned by the base station 110, then UE K+1 120a can use the received weight value to initialize Rx-NN K+1 If the base station 110 does not send K+1 120b assigns weight values, then UE K+1 120a can initialize Rx-NN from a random state K+1 .

[0109] At 316, UE K+1 120b may indicate to the base station 110 that the UE K+1120a is ready for training. After receiving the ready indication, the base station 110 may start sending the training signal to the UE. K+1 120b transmits a training signal.

[0110] At 318, the base station 110 may generate a training signal from the training bits (eg, UD bits) using a transmitter neural network Tx-NN(K) having a first set of weight values ​​that have been optimized for the UE. K+1 As described above, generating a training signal may include mapping a block of training bits to complex-valued modulation symbols using Tx-NN(K), and further processing the complex-valued modulation symbols by, for example, resource mapping and OFDM modulation.

[0111] At 320, base station 110 may send a K+1 120b transmits a training signal.

[0112] At 322, UE K+1 120b may use the training signal received from base station 110 to train its Rx-NN K+1 As described above, train Rx-NN K+1 This may include constructing input data from the received training signal and feeding the input data to the Rx-NN K+1 To recover the training bits sent from base station 110, by comparing the Rx-NN K+1 The output LLR and the transmitted training bits indicated by the base station 110 are used to calculate the training loss and update the Rx-NN K+1 The weight value of Rx-NN is used to minimize the training loss. Steps 320 and 322 may be performed iteratively until Rx-NN K+1 Convergence means that the training loss reaches a minimum and does not decrease any more.

[0113] At 324, UE K+1 120b may respond to Rx-NN K+1 The training loss and / or equivalent gradient are sent to the base station 110 based on the convergence of

[0114] At 326, the base station 110 may update the first set of weight values ​​for Tx-NN(K) based on the received losses and / or gradients to obtain a second set of weight values ​​referred to as temp Tx-NN(K).

[0115] At 328, the base station 110 may calculate the difference δ between the first set of weight values ​​of Tx-NN(K) and the second set of weight values ​​of Temp-Tx-NN(K). (K,temp K) In one example, the difference δ (K,temp K)It can be calculated as the normalized mean square error (NMSE) of each constellation point between Tx-NN(K) and temp Tx-NN(K).

[0116] At 330, if the difference δ (K,temp K) is less than or equal to a predetermined threshold, the base station 110 may select Tx-NN(K) having a first set of weight values ​​as a common transmitter neural network Tx-NN(K+1) for transmitting the K initial UEs 120a and the newly added UEs K+1 120b (collectively referred to as UE 120) transmits a signal. If the difference δ (K,temp K) If the value is greater than the threshold, the base station 110 may select temp Tx-NN(K) with the second set of weight values ​​as Tx-NN(K+1) and trigger the second stage 300B to further train the base station 110 and the UE 120 .

[0117] Figure 8 3 is a schematic message sequence diagram illustrating the second stage 300B of the E2E learning process according to an exemplary embodiment of the present disclosure. Unlike process 300A, process 300B may be performed by base station 110 and all K+1 UEs 120. Figure 5 Some steps in process 300B are discussed and will be described here in a simplified manner.

[0118] refer to Figure 8 At 332, base station 110 may transmit a training request to UE 120. In an example embodiment, the training request may include a training configuration for the second phase of training. For example, the training configuration may include one or more of an indication of a modulation and coding scheme, subcarrier information, and an indication of training bits. In another example embodiment, the training request may not include a training configuration, and UE 120 may reuse a training configuration previously received from base station 110.

[0119] At 334, UE 120 may respond to base station 110 with an acknowledgement message. Upon receiving the acknowledgement message, base station 110 may begin transmitting training signals to UE 120 for the second phase of training.

[0120] At 336, the base station 110 may generate a training signal from the training bits (e.g., UD bits) using a common transmitter neural network Tx-NN(K+1) having a second set of weight values. As described above, generating the training signal may include mapping the blocks of training bits to complex-valued modulation symbols using Tx-NN(K+1), and further processing the complex-valued modulation symbols by, for example, resource mapping and OFDM modulation. The training signal is only for the newly joined UE. K+1Unlike the first stage training prepared for the new UE 120b, the base station 110 prepares training signals for the initial UE 120a and the new UE 120b at step 336. Depending on the configuration of the UE 120, the base station 110 may use different training bits to generate training signals for each UE 120.

[0121] At 338 , base station 110 may transmit a training signal to UE 120 .

[0122] At 340, UE 120 may train their Rx-NN using the training signal received from base station 110. K (k=1, ..., k+1). As described above, train Rx-NN K This may include constructing input data from the received training signal and feeding the input data to the Rx-NN K To recover the training bits transmitted from base station 110, by comparing the K The output LLR and the transmitted training bits indicated by the base station 110 are used to calculate the training loss and update the Rx-NN K The weight values ​​of 1 and 2 are used to minimize the training loss. Each UE 120 can train its Rx-NN separately. K .

[0123] At 342, UE 120 may report the training loss and / or equivalent gradient to base station 110. UE 120 may report the training loss and / or equivalent gradient to base station 110 in each iteration.

[0124] At 344, base station 110 may combine the training loss and / or equivalent gradient reports from UE 120. For example, base station 110 may combine the reports based on equations (3)-(4) above.

[0125] At 346, the base station 110 may update the second set of weight values ​​of Tx-NN(K+1) based on the combined loss or equivalent gradient to obtain a third set of weight values ​​that may be referred to as temp-Tx-NN(K+1). As described above, the base station 110 may update Tx-NN(K+1) based on formula (2).

[0126] At 348, the base station 110 may calculate the difference δ between Tx-NN(K+1) and Temp-Tx-NN(K+1). (K+1,temp K+1) In one example, the difference δ (K+1,temp K+1) It can be calculated as the normalized mean square error (NMSE) of each constellation point between Tx-NN(K+1) and Temp-Tx-NN(K+1).

[0127] At 350, the base station 110 may determine the difference δ between Tx-NN(K+1) and Temp-Tx-NN(K+1). (K+1,temp K+1) Is it less than or equal to a predetermined threshold? (K+1,temp K+1) If the difference δ is less than or equal to the threshold, the base station 110 may determine that any one of Tx-NN(K+1) and Temp-Tx-NN(K+1) is suitable for the UE 120. The base station may use Tx-NN(K+1) or temp-Tx-NN(K+1) as the common Tx-NN for the UE 120. (K+1,temp K+1) If the difference δ is greater than the threshold, the base station 110 may determine that both Tx-NN(K+1) and Temp-Tx-NN(K+1) are not suitable for the UE 120. At 352, the base station 110 may replace the second set of weight values ​​of Tx-NN(K+1) with the second set of weight values ​​of Temp-Tx-NN(K+1), and iteratively perform steps from 336 to 350 until the difference δ is greater than the threshold. (K+1,temp K+1) becomes less than or equal to the threshold.

[0128] At 354, in response to the difference δ (K+1,temp K+1) If the training completion message is less than or equal to the threshold, the base station 110 may send a training completion message to the UE 120. In response to the training completion message, the UE 120 may terminate the training process and use the trained Rx-NN K Communication with base station 110 is initiated.

[0129] Fig. 9 4 is a schematic block diagram showing a device 400 according to an example embodiment of the present disclosure. The device 400 may be implemented at a network device (e.g., a base station 110) to perform the above-described scalable E2E waveform learning process. Since the scalable E2E waveform learning process has been discussed in detail above, each block of the device 400 will be briefly described here, and the details thereof may refer to the above description.

[0130] refer to Fig. 9 , the apparatus 400 may include: a first device 410 for generating a first training signal from a first training bit using a transmitter neural network, the transmitter neural network having a first set of weight values ​​trained for at least one first UE 120a; and a second device 420 for transmitting the first training signal to at least one second UE 120b for training a receiver neural network at at least one of the second UEs 120a.

[0131] The apparatus 400 may further include a third means 430 for updating a first set of weight values ​​of the transmitter neural network based on loss and / or gradient values ​​received from at least one second UE 120b to obtain a second set of weight values ​​for the transmitter neural network.

[0132] The apparatus 400 may further include a fourth means 440 for calculating a first difference between the first set of weight values ​​and the second set of weight values ​​for the transmitter neural network.

[0133] The apparatus 400 may further include a fifth means 450 for selecting the first set of weight values ​​for the transmitter neural network in response to the first difference being less than or equal to a first threshold.

[0134] In an example embodiment, optionally, the apparatus 400 may further include a sixth means 460 for generating a second training signal from the second training bits using a transmitter neural network having a second set of weight values ​​in response to the first difference being greater than the first threshold.

[0135] In an example embodiment, optionally, the apparatus 400 may further include a seventh means 470 for transmitting a second training signal to at least one first UE 120a and at least one second UE 120b for training corresponding receiver neural networks at the at least one first UE 120a and at least one second UE 120b.

[0136] In an example embodiment, optionally, the device 400 may also include an eighth device 480 for updating a second set of weight values ​​of the transmitter neural network based on loss or gradient values ​​received from at least one first UE 120a and at least one second UE 120b to obtain a third set of weight values ​​for the transmitter neural network.

[0137] In an example embodiment, the eighth device 480 may include: a first sub-device 482 for combining loss or gradient values ​​received from at least one first UE 120a and at least one second UE 120b; and a second sub-device 484 for updating a second set of weight values ​​of the transmitter neural network based on the combined loss or gradient values.

[0138] In an example embodiment, optionally, the device 400 may also include a ninth means 490 for causing the sixth means 460, the seventh means 470 and the eighth means 480 to iteratively perform their operations until a second difference between the second set of weight values ​​and the third set of weight values ​​is less than or equal to a second threshold.

[0139] In an example embodiment, optionally, the apparatus 400 may further include a tenth means 402 for sending a training configuration to the at least one second UE 120b before generating the first training signal. The training configuration may include at least one of an indication of a modulation and coding scheme, subcarrier information, an indication of training bits, and an initial weight value for a receiver neural network of the at least one second UE 120b.

[0140] In an example embodiment, initial weight values ​​for a receiver neural network for at least one second UE 120b are received from one of the at least one first UE 120a or from another UE in which the receiver neural network was trained under substantially the same or similar conditions as the at least one second UE 120b.

[0141] In an example embodiment, the first set of weight values ​​and the second set of weight values ​​of the transmitter neural network include a predetermined number of coordinate pairs representing constellation points, and the transmitter neural network is configured to map the training bit blocks to complex-valued modulation symbols corresponding to the constellation points.

[0142] In an example embodiment, the apparatus 400 may also include additional means for performing operations related to the base station 110 as described above.

[0143] Fig.10 is a schematic block diagram showing a device 500 according to an example embodiment of the present disclosure. The device 500 may be implemented at a terminal device such as UE 120 b to perform the above-described scalable E2E waveform learning process. Since the scalable E2E waveform learning process has been discussed in detail above, each block of the device 500 will be briefly described here, and the details thereof may refer to the above description.

[0144] refer to Fig.10 , the apparatus 500 may include: a first device 510 for receiving a first training signal from the base station 110; and a second device 520 for training a receiver neural network using the first training signal until the receiver neural network converges. The receiver neural network may be configured to recover the first training bit from the first training signal.

[0145] The apparatus 500 may further include a third means 530 for sending the first loss and / or the corresponding gradient value calculated based on the recovered first training bits to the base station 110 in response to the convergence of the receiver neural network.

[0146] In an exemplary embodiment, optionally, the apparatus 500 may further include a fourth means 540 for receiving a second training signal from the base station 110, and a fifth means 550 for training the receiver neural network using the second training signal. The receiver neural network may be configured to recover the second training bits from the second training signal.

[0147] In an exemplary embodiment, optionally, the apparatus 500 may further include a sixth means 560 for sending the second loss calculated according to the restored second training bits and / or the corresponding gradient value to the base station 110 .

[0148] In an example embodiment, optionally, the apparatus 500 may further include a seventh means 570 for receiving a training completion message from the base station 110 .

[0149] In an example embodiment, the fourth device 540 , the fifth device 550 , and the sixth device 560 may iteratively perform their operations until the seventh device 570 receives a training completion message from the base station 110 .

[0150] In an example embodiment, optionally, the apparatus 500 may further include an eighth means 502 for receiving a training configuration from the base station 110 before the first means 510 receives the first training signal. The training configuration may include at least one of an indication of a modulation and coding scheme, subcarrier information, an indication of a first training bit, and an initial weight value for a receiver neural network.

[0151] In an example embodiment, optionally, the apparatus 500 may further include a ninth means 504 for initializing the receiver neural network with the initial weight values ​​received from the base station 110 or with random weight values ​​before the second means 520 trains the receiver neural network.

[0152] In an example embodiment, the receiver neural network may be configured to receive as input data OFDMA symbols demodulated from a received training signal and output a probability distribution of possible bits recovered from the OFDMA symbols.

[0153] In an example embodiment, the apparatus 500 may also include additional means for performing operations as described above in relation to the UE 120a and / or the UE 120b.

[0154] The exemplary embodiments have been described in the context of an OFDM system. It should be appreciated that the principles of the exemplary embodiments are also applicable to communication systems employing waveforms other than OFDM.

[0155] Fig.11 The performance evaluation of the proposed scalable E2E learning scheme compared with conventional E2E learning is shown. The performance evaluation is simulated in an environment where a base station serves 3 to 4 UEs. The base station is deployed with a common Tx-NN for constellation mapping, while each UE is deployed with a separate Rx-NN. The Tx-NN and Rx-NN have Fig. 6A and 6BThe neural network structure is shown in . The neural network is first trained under the 1-BS:3-UE E2E learning framework, and then when the fourth UE joins the system, the proposed scalable E2E learning and conventional E2E learning from scratch are performed to study their performance.

[0156] The simulation configuration is as follows. The modulation scheme used for Tx-NN is 16QAM, using 4 bits per channel. Therefore, the block size Q=4, and there are a total of 2 Q = 16 bit blocks. OFDM modulation is applied. The subcarrier size in each OFDM symbol is 72. The number of OFDM symbols in a time slot is 14. Therefore, there are a total of 72×14 bit blocks in one time slot. On the transmitter side, the nonlinearity of the power amplifier is also modeled as a hardware defect. An additive white Gaussian noise (AWGN) channel is used as the channel model.

[0157] For IQ imbalance at each receiver, the "I" dimension gain and associated angular rotation may vary with some uniform random bias value. For example, 0.85 / 0.09 means the "I" dimension gain is 0.85, the uniform random bias value is 0.01, and the associated angular rotation is -0.09π, the uniform random error value is 0.01π.

[0158] To construct a common shared Tx-NN with 3 initial UEs, the Tx-NN and Rx-NN are trained with 800 iterations. In all trainings, the signal-to-noise ratio (SNR) of each OFDM frame is randomly uniformly selected from 0 to 25 dB. The training batch size is 100 frames.

[0159] The simulation results show that by utilizing the common shared Tx-NN as the starting neural network, the changes of the constellation mapping coordinate pairs before and after the fourth UE joins the system are negligible, so it has little impact on other UEs. Fig.11 The proposed method speeds up the training speed for new UEs. Compared with conventional training from scratch, the proposed method can greatly reduce the number of iterations. The effort of hundreds of iterations for optimizing the fourth UE can be saved. Therefore, online training in such a system can reduce training time, power consumption and data exchange.

[0160] Fig.12 6 is a schematic block diagram illustrating devices in a communication system 600 in which example embodiments of the present disclosure may be implemented. Fig.12 As shown, the communication system 600 may include: a network device 610, which may be implemented as the above-mentioned base station 110; and a terminal device 620, which may be implemented as one of the above-mentioned UEs 120. Fig.12Only one terminal device 620 is shown, but it should be understood that the communication system 600 may include multiple terminal devices 620 wirelessly connected to the network device 610.

[0161] refer to Fig. 9 , the network device 610 may include one or more processors 611, one or more memories 612, one or more transceivers 613, and one or more network interfaces 617 interconnected by one or more buses 614. The one or more buses 614 may be address, data, or control buses, and may include any interconnection mechanism, such as a series of lines on a motherboard or integrated circuit, optical fiber, optical or other optical communication devices, etc. Each of the one or more transceivers 613 may include a receiver and a transmitter connected to a plurality of antenna elements 616. In one example, the network device 610 may include a plurality of transceivers 613 for a plurality of baseband boards, and each transceiver 613 is connected to a set of antenna elements 616 for sending and receiving radio signals. The network device 610 may operate as a base station for a terminal device 620 and communicate wirelessly with the terminal device 620 through the antenna elements 616. Using the plurality of antenna elements 616, the network device 610 may support MIMO technology including, for example, beamforming. One or more network interfaces 617 may provide a wired or wireless communication link through which the network device 610 can communicate with other network devices, entities, or functions. The one or more memories 612 may include computer program code 615. The one or more memories 612 and computer program code 615 may be configured to, when executed by the one or more processors 611, cause the network device 610 to perform processes and operations associated with the base station 110, as described above.

[0162] The terminal device 620 may include one or more processors 621, one or more memories 622, and one or more transceivers 623 interconnected by one or more buses 624. The one or more buses 624 may be address, data, or control buses, and may include any interconnection mechanism, such as a series of lines on a motherboard or integrated circuit, optical fiber, optical or other optical communication devices, etc. Each of the one or more transceivers 623 may include a receiver and a transmitter connected to a plurality of antenna elements 626. The terminal device 620 may communicate wirelessly with the network device 610 via the antenna elements 626. The one or more memories 622 may include computer program code 625. The one or more memories 622 and the computer program code 625 may be configured to, when executed by the one or more processors 621, cause the terminal device 620 to perform processing and operations related to the UE 120, as described above.

[0163] The one or more processors 611, 621 may be any suitable type suitable for the local technical network, and may include one or more of the following: a general-purpose processor, a special-purpose processor, a microprocessor, a digital signal processor (DSP), one or more processors in a processor-based multi-core processor architecture, and a special-purpose processor, such as a processor developed based on a field programmable gate array (FPGA) and an application-specific integrated circuit (ASIC). The one or more processors 611, 621 may be configured to control other elements of the terminal / network device and operate in conjunction with them to implement the above process.

[0164] The one or more memories 612, 622 may include at least one storage medium in various forms, such as volatile memory and / or non-volatile memory. Volatile memory may include, but is not limited to, for example, random access memory (RAM) or cache. Non-volatile memory may include, but is not limited to, for example, read-only memory (ROM), hard disk, flash memory, etc. In addition, the one or more memories 612, 622 may include, but are not limited to, electrical, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices or devices, or any combination of the above.

[0165] The network device 610 may be implemented as a single network node, or decomposed / distributed across two or more network nodes such as a central unit (CU), a distributed unit (DU), a remote radio head (RRH) using different functional division architectures and different interfaces.

[0166] It should be understood that the blocks in the drawings can be implemented in various ways including software, hardware, firmware or any combination thereof. In some example embodiments, one or more blocks may be implemented using software and / or firmware, such as machine executable instructions stored in a storage medium. In addition to or in lieu of machine executable instructions, some or all of the blocks in the drawings may be implemented at least in part by one or more hardware logic components. For example, but not limited to, exemplary types of hardware logic components that may be used include field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0167] Some example embodiments also provide computer program codes or instructions that, when executed by one or more processors, enable a device or apparatus to perform the above-described processes. The computer program code for executing the processes of each example embodiment may be written in any combination of one or more programming languages. These program codes may be provided to one or more processors or controllers of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the program code is executed by a processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code may be executed entirely on the machine, partially on the machine as a separate software package, partially on the machine, partially on a remote machine, or entirely on a remote computer or server.

[0168] Some example embodiments also provide a computer program product or computer readable medium in which computer program code or instructions are stored. A computer readable medium may be any tangible medium that may contain or store for use or in combination with an instruction execution system, device or apparatus. A machine readable medium may be a machine readable signal medium or a machine readable storage medium. A machine readable medium may include, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared or semiconductor system, device or device, or any suitable combination thereof. More specific examples of machine readable storage media will include an electrical connection with one or more circuits, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0169] In addition, although each operation is described in a specific order, this should not be understood as requiring such operations to be performed in the specific order or sequential order shown, or requiring the execution of all the operations shown to obtain the desired result. In some cases, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these details should not be interpreted as limitations on the scope of the present disclosure, but rather as descriptions of features specific to a particular embodiment. Certain features described in the context of a separate embodiment may also be implemented in combination in a single embodiment. On the contrary, the various features described in the context of a single embodiment may also be implemented in multiple embodiments individually or in any suitable sub-combination.

[0170] Although the subject matter has been described in language specific to structural features and / or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts discussed above are disclosed as examples of implementing the claims.

Claims

1. A network device, comprising: at least one processor; as well as at least one memory including computer program code, the at least one memory and the computer program code being configured to, with the at least one processor, cause the network device to at least: generating a first training signal from first training bits using a transmitter neural network having a first set of weight values ​​trained for at least one first terminal device; transmitting the first training signal to at least one second terminal device to train a receiver neural network at the at least one second terminal device; updating the first set of weight values ​​of the transmitter neural network based on the loss and / or gradient values ​​received from the at least one second terminal device to obtain a second set of weight values ​​for the transmitter neural network; calculating a first difference between the first set of weight values ​​and the second set of weight values ​​for the transmitter neural network; as well as In response to the first difference being less than or equal to a first threshold, the first set of weight values ​​for the transmitter neural network is selected.

2. The network device according to claim 1, wherein: The at least one memory and the computer program code are further configured to, with the at least one processor, cause the network device to at least: In response to the first difference being greater than the first threshold, generating a second training signal from second training bits using the transmitter neural network having the second set of weight values; Transmitting the second training signal to the at least one first terminal device and the at least one second terminal device to train respective receiving neural networks at the at least one first terminal device and the at least one second terminal device; updating the second set of weight values ​​of the transmitter neural network based on the loss or gradient values ​​received from the at least one first terminal device and the at least one second terminal device to obtain a third set of weight values ​​for the transmitter neural network; The steps of generating the second training signal, transmitting the second training data, and updating the second set of weight values ​​are iteratively performed until a second difference between the second set of weight values ​​and the third set of weight values ​​is less than or equal to a second threshold.

3. The network device according to claim 1, wherein: The at least one memory and the computer program code are further configured to, with the at least one processor, cause the network device to at least: Prior to generating the first training signal, a training configuration is transmitted to the at least one second terminal device, the training configuration comprising an indication of a modulation and coding scheme, subcarrier information, an indication of the training bits, and at least one of initial weight values ​​of the receiver neural network for the at least one second terminal device.

4. The network device according to claim 3, wherein: The initial weight values ​​for the receiver neural network for the at least one second terminal device are received from one of the at least one first terminal device or from another terminal device in which the receiver neural network was trained under substantially the same or similar conditions as the at least one second terminal device.

5. The network device according to claim 2, wherein: Updating the second set of weight values ​​for the transmitter neural network includes: combining the loss or gradient values ​​received from the at least one first terminal device and the at least one second terminal device; and The second set of weight values ​​of the transmitter neural network is updated based on the combined loss or gradient value.

6. The network device according to claim 1, wherein: The first set of weight values ​​and the second set of weight values ​​of the transmitter neural network include a predetermined number of data pairs, and the transmitter neural network is configured to map training bit blocks to complex-valued modulation codewords having real components and imaginary components corresponding to a pair of data values ​​selected from the predetermined number of data pairs.

7. A terminal device, comprising: at least one processor; as well as at least one memory including computer program code, the at least one memory and the computer program code being configured to, with the at least one processor, cause the terminal device to at least: receiving a first training signal from a network device; training a receiver neural network using the first training signal until the receiver neural network converges, the receiver neural network being configured to recover first training bits from the first training signal; as well as In response to the convergence of the receiver neural network, a first loss calculated based on the recovered first training bits and / or a corresponding gradient value are sent to the network device.

8. The terminal device according to claim 7, wherein: The at least one memory and the computer program code are further configured to, with the at least one processor, cause the terminal device to at least: receiving a second training signal from the network device; training the receiver neural network using the second training signal, the receiver neural network being configured to recover second training bits from the second training signal; as well as sending a second loss calculated according to the restored second training bit and / or a corresponding gradient value to the network device, Therein, the steps of receiving the second training signal, training the receiver neural network and sending the second loss and / or the corresponding gradient value are performed iteratively.

9. The terminal device according to claim 8, wherein: The steps of receiving the second training signal, training the receiver neural network, and sending the second loss and / or the corresponding gradient value are iteratively performed until the terminal device receives a training completion message from the network device.

10. The terminal device according to claim 7, wherein: The at least one memory and the computer program code are further configured to, with the at least one processor, cause the terminal device to at least: Prior to receiving the first training signal, a training configuration is received from the network device, the training configuration comprising at least one of an indication of a modulation and coding scheme, subcarrier information, an indication of the first training bits, and an initial weight value for the receiver neural network.

11. The terminal device according to claim 7, wherein: The at least one memory and the computer program code are further configured to, with the at least one processor, cause the terminal device to at least: The receiver neural network is initialized with initial weight values ​​received from the network device or with random weight values ​​before training the receiver neural network.

12. The terminal device according to claim 7, wherein: The receiver neural network is configured to receive as input data OFDMA symbols demodulated from the received training signal and to output a probability distribution of possible bits recovered from the OFDMA symbols.

13. A method implemented at a network device, comprising: generating a first training signal from first training bits using a transmitter neural network having a first set of weight values ​​trained for at least one first terminal device; transmitting the first training signal to at least one second terminal device for training a receiver neural network at the at least one second terminal device; updating the first set of weight values ​​of the transmitter neural network based on the loss and / or gradient values ​​received from the at least one second terminal device to obtain a second set of weight values ​​for the transmitter neural network; calculating a first difference between the first set of weight values ​​and the second set of weight values ​​for the transmitter neural network; as well as In response to the first difference being less than or equal to a first threshold, the first set of weight values ​​for the transmitter neural network is selected.

14. The method of claim 13, further comprising: In response to the first difference being greater than the first threshold, generating a second training signal from second training bits using the transmitter neural network having the second set of weight values; Transmitting the second training signal to the at least one first terminal device and the at least one second terminal device for training corresponding receiver neural networks at the at least one first terminal device and the one second terminal device; updating the second set of weight values ​​of the transmitter neural network based on the loss or gradient values ​​received from the at least one first terminal device and the at least one second terminal device to obtain a third set of weight values ​​for the transmitter neural network; The steps of generating the second training signal, transmitting the second training signal, and updating the second set of weight values ​​are iteratively performed until a second difference between the second set of weight values ​​and the third set of weight values ​​is less than or equal to a second threshold.

15. The method of claim 13, further comprising: Prior to generating the first training signal, a training configuration is sent to the second terminal device, the training configuration comprising an indication of a modulation and coding scheme, subcarrier information, an indication of the training bits, and at least one of an initial weight value of the receiver neural network for the second terminal device.

16. The method of claim 15, wherein: The initial weight values ​​for the receiver neural network for the second terminal device are received from one of the at least one first terminal device or from another terminal device in which the receiver neural network was trained under substantially the same or similar conditions as the second terminal device.

17. The method of claim 14, wherein: Updating the second set of weight values ​​for the transmitter neural network includes: combining the loss or gradient values ​​received from the at least one first terminal device and the at least one second terminal device; and The second set of weight values ​​of the transmitter neural network is updated based on the combined loss or gradient value.

18. The method of claim 13, wherein: The first set of weight values ​​and the second set of weight values ​​of the transmitter neural network include a predetermined number of data pairs, and the transmitter neural network is configured to map a training bit block to a complex-valued modulation codeword having a real component and an imaginary component corresponding to a pair of data values ​​selected from the predetermined number of data pairs.

19. A method implemented at a terminal device, comprising: receiving a first training signal from a network device; training a receiver neural network using the first training signal until the receiver neural network converges, the receiver neural network being configured to recover first training bits from the first training signal; as well as In response to the convergence of the receiver neural network, a first loss calculated based on the recovered first training bits and / or a corresponding gradient value are sent to the network device.

20. The method of claim 19, further comprising: receiving a second training signal from the network device; training the receiver neural network using the second training signal, the receiver neural network being configured to recover second training bits from the second training signal; as well as sending a second loss calculated according to the restored second training bit and / or a corresponding gradient value to the network device, Therein, the steps of receiving the second training signal, training the receiver neural network and sending the second loss and / or the corresponding gradient value are performed iteratively.

21. The method of claim 20, wherein: The steps of receiving the second training signal, training the receiver neural network, and sending the second loss and / or the corresponding gradient value are iteratively performed until the terminal device receives a training completion message from the network device.

22. The method of claim 19, further comprising: A training configuration is received from the network device prior to receiving the first training signal, the training configuration comprising at least one of an indication of a modulation and coding scheme, subcarrier information, an indication of the first training bits, and an initial weight value for the receiver neural network.

23. The method of claim 19, further comprising: The receiver neural network is initialized with initial weight values ​​received from the network device or with random weight values ​​before training the receiver neural network.

24. The method of claim 19, wherein: The receiver neural network is configured to receive as input data OFDMA symbols demodulated from the received training signal and to output a probability distribution of possible bits recovered from the OFDMA symbols.

25. An apparatus comprising: means for generating, at a network device, a first training signal from first training bits using a transmitter neural network having a first set of weight values ​​trained for at least one first terminal device; means for transmitting said first training signal from said network device to at least one second terminal device to train a receiver neural network at said at least one second terminal device; means for updating the first set of weight values ​​of the transmitter neural network based on the loss and / or gradient values ​​received from the at least one second terminal device to obtain a second set of weight values ​​for the transmitter neural network; means for calculating a first difference between said first set of weight values ​​and said second set of weight values ​​for said transmitter neural network; as well as means for selecting the first set of weight values ​​for the transmitter neural network in response to the first difference being less than or equal to a first threshold.

26. An apparatus comprising: means for receiving, at a terminal device, a first training signal from a network device; means for training a receiver neural network using the first training signal until the receiver neural network converges, the receiver neural network being configured to recover first training bits from the first training signal; as well as Means for sending, in response to convergence of the receiver neural network, a first loss and / or a corresponding gradient value calculated based on the recovered first training bits from the terminal device to the network device.

27. A computer program product, embodied in at least one computer-readable medium, and comprising instructions that, when executed by at least one processor of a network device, cause the network device to perform at least the following operations: generating a first training signal from first training bits using a transmitter neural network having a first set of weight values ​​trained for at least one first terminal device; transmitting the first training signal to at least one second terminal device for training a receiver neural network at the at least one first terminal device; updating the first set of weight values ​​of the transmitter neural network based on the loss and / or gradient values ​​received from the at least one second terminal device to obtain a second set of weight values ​​for the transmitter neural network; calculating a first difference between the first set of weight values ​​and the second set of weight values ​​for the transmitter neural network; as well as In response to the first difference being less than or equal to a first threshold, the first set of weight values ​​for the transmitter neural network is selected.

28. A computer program product, embodied in at least one computer-readable medium, and comprising instructions that, when executed by at least one processor of a terminal device, cause the terminal device to perform at least the following operations: receiving a first training signal from a network device; training a receiver neural network using the first training signal until the receiver neural network converges, the receiver neural network being configured to recover first training bits from the first training signal; and In response to the convergence of the receiver neural network, a first loss calculated based on the recovered first training bits and / or a corresponding gradient value are sent to the network device.

Images