Communications of channel-dependent information in a communications network
By determining an optimal permutation for codeword arrangements in wireless communication networks, the method addresses distortion and overhead challenges in CSI feedback, enhancing CSI compression and reconstruction robustness and network performance.
Patent Information
- Application Number
- PCT/EP2024/057365
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-25
- Filing Date
- 2024-03-20
- Publication Date
- 2024-11-28
AI Technical Summary
Current wireless communication networks face challenges in efficiently communicating channel-dependent information, particularly in minimizing distortion and overhead during channel state information (CSI) feedback, especially in 5G networks, where accurate CSI is crucial for signal-to-noise ratio and channel capacity but is hampered by limited feedback overhead and reconstruction loss due to sparse channel assumptions and random projections.
The solution involves determining an optimal or sub-optimal permutation for arranging codewords in a quantization codebook to minimize weighted average distortion, using methods like exhaustive search, random search, or iterative algorithms, and employing a mapping function and distance metric to evaluate and select the best permutation for robust CSI communication, which can be shared between user equipment and network entities.
This approach enhances the robustness of CSI compression and reconstruction, reducing distortion and feedback overhead, thereby improving signal quality and channel capacity while being adaptable to various communication scenarios.
Smart Images

Figure EP2024057365_28112024_PF_FP_ABST
Abstract
Description
COMMUNICATIONS OF CHANNEL-DEPENDENT INFORMATION IN A COMMUNICATIONS NETWORKTECHNICAL FIELD
[0001] Various example embodiments relate to wireless communications and to communications of channel-dependent information.BACKGROUND
[0002] This section illustrates useful background information without admission of any technique described herein representative of the state of the art.
[0003] Wireless communication networks evolve constantly and there is a constant need to study and design new functionalities for the wireless communication networks.SUMMARY
[0004] The scope of protection sought for various embodiments of present disclosure is set out by the independent claims. The embodiments and features, if any, described in this specification that do not fall under the scope of the independent claims are to be interpreted as examples useful for understanding various example embodiments.
[0005] According to a first example aspect of the present disclosure, there is provided a first apparatus comprising: at least one processor; and at least one memory including executable instructions that, when executed by the at least one processor, cause the apparatus to determine a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook; and to communicate the permutation to a second apparatus.
[0006] In some example embodiments, the first apparatus is caused to determine the permutation based on one of the following: exhaustive search, random search, an iterative algorithm.
[0007] In some example embodiments, the first apparatus is caused to determine the permutation by choosing a desired permutation based on distortion calculated for a plurality of possible permutations.
[0008] In some example embodiments, the first apparatus is caused to determine the permutation by choosing a permutation that results in minimum weighted averagedistortion among plurality of possible permutations.
[0009] In some example embodiments, the first apparatus is caused to obtain a mapping function and a distance metric; and use the mapping function and the distance metric in the determination of the permutation. The determination of the permutation may comprise calculating distortion for a plurality of possible permutations based on the mapping function and the distance metric and choosing desired permutation based on the calculated distortion. The mapping function may be defined as a function that transforms an input from one space into an output belonging to another space, wherein the distance metric is defined for being applied in the another space.
[0010] In some example embodiments, the first apparatus is caused to obtain at least one of the mapping function and the distance metric by receiving the at least one of the mapping function and the distance metric from the second apparatus.
[0011] In some example embodiments, the first apparatus is caused to obtain at least one of the mapping function and the distance metric by determining the at least one of the mapping function and the distance metric at the first apparatus.
[0012] In some example embodiments, the first apparatus is caused to determine the permutation responsive to receiving, from the second apparatus, an indication of the permutation not being determined by the second apparatus.
[0013] In some example embodiments, the first apparatus is caused to use the permutation in communications between the first apparatus and the second apparatus
[0014] In some example embodiments, the first apparatus is caused to use the permutation in at least one of the following: communicating channel state information, CSI; beamforming; radio resource management; positioning.
[0015] In some example embodiments, the first apparatus is a user equipment and the second apparatus is a network entity, or vice versa.
[0016] According to a second example aspect of the present disclosure, there is provided a second apparatus comprising at least one processor; and at least one memory including executable instructions that, when executed by the at least one processor, cause the apparatus to receive, from a first apparatus, a permutation for quantization or dequantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook.
[0017] In some example embodiments, the second apparatus is caused to determine at least one of a mapping function and a distance metric for calculating distortionfor a plurality of possible permutations; and to communicate the at least one of the mapping function and the distance metric to the first apparatus, wherein the mapping function is a function that transforms an input from one space into an output belonging to another space, wherein the distance metric is defined for being applied in the another space.
[0018] In some example embodiments, the second apparatus is caused, responsive to receiving the permutation, to determine a new permutation and communicate the new permutation to the first apparatus.
[0019] In some example embodiments, the second apparatus is caused to use the permutation in communications between the first apparatus and the second apparatus.
[0020] In some example embodiments, the second apparatus is caused to use the permutation in at least one of the following: communicating channel state information, CSI; beamforming; radio resource management; positioning.
[0021] In some example embodiments, the first apparatus is a user equipment and the second apparatus is a network entity, or vice versa.
[0022] According to a third example aspect of the present disclosure, there is provided a method, comprising: determining, at a first apparatus, a permutation for quantization or dequantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook; and communicating the permutation to a second apparatus.
[0023] According to a fourth example aspect of the present disclosure, there is provided a method, comprising: receiving, at a second apparatus from a first apparatus, a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook.
[0024] According to a fifth example aspect of the present disclosure, there is provided a computer program comprising instructions for performing at least the following: determining, at a first apparatus, a permutation for quantization or dequantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook; and communicating the permutation to a second apparatus.
[0025] According to a sixth example aspect of the present disclosure, there is provided a computer program comprising instructions for performing at least the following:receiving, at a second apparatus from a first apparatus, a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook.
[0026] According to a seventh example aspect of the present disclosure, there is provided a first apparatus comprising: means for determining a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook; and means for communicating the permutation to a second apparatus.
[0027] In some example embodiments, the means for determining the permutation are configured to determine the permutation based on one of the following: exhaustive search, random search, an iterative algorithm.
[0028] In some example embodiments, the means for determining the permutation are configured to determine the permutation by choosing a desired permutation based on distortion calculated for a plurality of possible permutations.
[0029] In some example embodiments, the means for determining the permutation are configured to determine the permutation by choosing a permutation that results in minimum weighted average distortion among plurality of possible permutations.
[0030] In some example embodiments, the first apparatus of the seventh aspect is caused to obtain a mapping function and a distance metric; and use the mapping function and the distance metric in the determination of the permutation.
[0031] In some example embodiments, the first apparatus of the seventh aspect further comprises means for obtaining a mapping function and a distance metric; and means for using the mapping function and the distance metric in the determination of the permutation. The determination of the permutation may comprise calculating distortion for a plurality of possible permutations based on the mapping function and the distance metric and choosing desired permutation based on the calculated distortion. The mapping function may be defines as a function that transforms an input from one space into an output belonging to another space, wherein the distance metric is defined for being applied in the another space.
[0032] In some example embodiments, the means for obtaining the mapping functions and the distance metric are configured to obtain at least one of the mapping function and the distance metric by receiving the at least one of the mapping function and the distance metric from the second apparatus.
[0033] In some example embodiments, the means for obtaining the mappingfunctions and the distance metric are configured to obtain at least one of the mapping function and the distance metric by determining the at least one of the mapping function and the distance metric at the first apparatus.
[0034] In some example embodiments, wherein the means for determining the permutation are configured to determine the permutation responsive to receiving, from the second apparatus, an indication of the permutation not being determined by the second apparatus.
[0035] In some example embodiments, the first apparatus of the seventh aspect further comprises means for using the permutation in communications between the first apparatus and the second apparatus.
[0036] In some example embodiments, the first apparatus of the seventh aspect further comprises means for using the permutation in at least one of the following: communicating channel state information, CSI; beamforming; radio resource management; positioning.
[0037] In some example embodiments, the first apparatus of the seventh aspect is a user equipment and the second apparatus of the seventh aspect is a network entity, or vice versa.
[0038] According to an eighth example aspect of the present disclosure, there is provided a second apparatus comprising means for receiving, from a first apparatus, a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook.
[0039] In some example embodiments, the second apparatus of the eighth aspect further comprises means for determining at least one of a mapping function and a distance metric for calculating distortion for a plurality of possible permutations; and means for communicating the at least one of the mapping function and the distance metric to the first apparatus, wherein the mapping function is a function that transforms an input from one space into an output belonging to another space, wherein the distance metric is defined for being applied in the another space.
[0040] In some example embodiments, wherein the second apparatus of the eighth aspect comprises means for determining a new permutation responsive to receiving the permutation, and means for communicating the new permutation to the first apparatus.
[0041] In some example embodiments, the second apparatus of the eighth aspect further comprises means for using the permutation in communications between the first apparatus and the second apparatus.
[0042] In some example embodiments, the second apparatus of the eighth aspect further comprises means for using the permutation in at least one of the following: communicating channel state information, CSI; beamforming; radio resource management; positioning.
[0043] In some example embodiments, the first apparatus of the eighth aspect is a user equipment and the second apparatus of the eighth aspect is a network entity, or vice versa.
[0044] The computer program of the fifth and / or the sixth example aspect may be stored in a non-transitory computer readable memory medium. The term non-transitory as used herein is a limitation of the medium itself (i.e. tangible, not a signal) as opposed to a limitation on data storage persistency (e.g. RAM vs. ROM).
[0045] Any foregoing memory medium may comprise a digital data storage such as a data disc or diskette, optical storage, magnetic storage, holographic storage, opto-magnetic storage, phase-change memory, resistive random access memory, magnetic random access memory, solid-electrolyte memory, ferroelectric random access memory, organic memory or polymer memory. The memory medium may be formed into a device without other substantial functions than storing memory or it may be formed as part of a device with other functions, including but not limited to a memory of a computer, a chip set, and a sub assembly of an electronic device.
[0046] The various means of the apparatus of the seventh and / or eighth example aspect may comprise a processor; and a memory including executable instructions that, when executed by the processor, cause the performance of the means.
[0047] Different non-binding example aspects and embodiments of the present invention have been illustrated in the foregoing. The embodiments in the foregoing are used merely to explain selected aspects or steps that may be utilized in implementations of the embodiment of the present disclosure. Some embodiments may be presented only with reference to certain example aspects. It should be appreciated that corresponding embodiments may apply to other example aspects as well.BRIEF DESCRIPTION OF THE DRAWINGS
[0048] For a more complete understanding of example embodiments of the present invention, reference is now made to the following descriptions taken in connection with the accompanying drawings in which:
[0049] Fig. 1A shows a block diagram of an example scenario;
[0050] Fig. 1 B shows an example of vector quantization;
[0051] Fig 2A shows a block diagram of another example scenario;
[0052] Figs. 2B-2C depict codebook permutation examples;
[0053] Fig. 2D is a graph depicting an example of average distortion with different codebook permutations;
[0054] Fig. 3 shows an operational graph according to an example embodiment;
[0055] Figs. 4-5 are flow diagrams of processes according to example embodiments;
[0056] Figs. 6-9 show signaling diagrams of according to example embodiments; and
[0057] Figs. 10-11 show block diagrams of apparatuses according to example embodiments.DETAILED DESCRIPTON OF THE DRAWINGS
[0058] calculating distortion for a plurality of possible permutations Various embodiments relate to communication of channel-dependent information in communications networks. The term channel-dependent information may refer to information related to the propagation channel used in wireless communications. The channel information may characterize the channel behavior in time, frequency, space for example. One possible purpose where various embodiments may be applied is communication of Channel state information (CSI) of a wireless channel. This is however only one non-limiting example. Other possible purposes include for example: communicating channel-dependent information for beamforming; radio resource management; positioning.
[0059] In the following, various example embodiments of present disclosure may be discussed in detail in the context of 5G NR technology. It is however to be noted that various example embodiments of present disclosure may be equally applied to other communication networks and other radio communication technologies (e.g. future technologies) in addition to 5G NR.
[0060] Various embodiments are discussed in the context of communications between a user equipment (UE) and a base station (BS or gNB). In a broader sense, functionalities of the various embodiments concern communications between two communication entities, at least one of which may be a network entity. The UE is to be interpreted broadly to cover any type of communication device that may perform communications over a wireless communications network. The UE may be e.g. a mobile phone, a smart phone, a laptop, a computer, an loT device. The base station may be called gNB in the context of 5G networks. In some other context, some other term may be used for corresponding network entity.
[0061] Channel state information (CSI) is needed for precoding in massive multipleinput multiple-output (MIMO) communications with frequency division duplex (FDD) schemes.This applies e.g. to 5G networks. Accurate CSI can be used by a base station to obtain higher signal-to-noise-ratio (SNR) and channel capacity. However, in FDD networks, only UE can estimate the downlink CSI. Thus, the estimated CSI needs to be shared with the BS, which introduces overhead to the network. To reduce the overhead, several vector quantization methods are proposed to generate a codebook that achieves a high compression rate (CR), i.e. , the ratio of compressed size to uncompressed size. Thus, CR is a scalar in the range of [0, 1], and lower CR means more compression. However, these methods cannot provide enough accuracy for CSI reconstruction with limited feedback overhead. Compressive sensing (CS)-based methods may be used to compress the CSI based on its spatial and temporal correlation. However, channels may not be sparse, which is a key assumption for the CS- based approach. Furthermore, CS-based methods use random projections, which can lead to significant reconstruction loss in the BS.
[0062] Considering the task of compression and reconstruction, an autoencoder (AE) structure from deep learning (DL) methods may be applied to the CSI-feedback application. For example, an AE-based architecture has been proposed to do the task using an encoder and a decoder. The encoder may generate a compressed representation of the input CSI, and the decoder may reconstruct the CSI from the compressed information. By joint training of the encoder and decoder, a more accurate reconstruction may be provided compared to the CS-based methods. A modified version of the AE-based architecture may be obtained by modifying the convolution kernel size and reconstruction blocks. Further, a structure that is based on a fully convolutional neural network has been proposed to enhance the quality of reconstructed CSI and decrease the number of trainable parameters and computational resources.
[0063] To reduce the feedback payload even more, the output of the encoder can be quantized. Fig. 1 shows a block diagram of an example scenario including quantization and de-quantization blocks. The block diagram of Fig. 1 comprises a UE 101 , a gNB 111 , and a wireless channel 103 connecting the UE 101 and the gNB 111 over the air. The UE 101 comprises an encoder 105 and a quantizer 106. The gNB comprises a decoder 115 and a dequantizer 116. In operation, the quantizer 106 outputs bits, which represent output of the encoder 105. The UE 101 communicates the bits over the wireless channel and the gNB 111 receives the bits. A de-quantizer at the gNB 111 constructs the decoder input using the received bits. Lastly, the decoder 115 processes the compressed signal.
[0064] Several quantization types can be considered for the quantization task including: (a) uniform scalar quantization, (b) non-uniform scalar quantization, and (c) vector quantization. In scalar quantization (SQ), each encoder output entry (element) is quantized based on its own value and the quantization type and properties. However, in vectorquantization (VQ), the encoder outputs are divided to some segments, and then the quantization is applied to each segment separately.
[0065] Fig. 1 B shows an example of vector quantization. An input 121 is divided into segments of three. A codebook 122 with 8 codewords is used for quantization. Quantization by the codeworks of the codebook 122 is applied to the segments of the input 121 to obtain output 123.
[0066] To measure the degradation caused by the quantization block, we may define quantization loss metric aswhere Xeand X denote, respectively, the input and output signal of the quantization block.
[0067] In SQ, Xeis a 1-dimensional signal (scalar). However, in VQ, Xeis a tridimensional signal, where m denotes the segment size (codeword length). In other words, Xedenotes a segment of the encoder output Ze.
[0068] VQ uses a codebook, C = {c0, c1, ... , c / v-1}, including N codewords, where each codeword can be selected / pointed with n bits.
[0069] For an input signal Xeto the VQ, the quantized signal can be written asX = arg min||e- c |2C[ € C
[0070] In practice, the UE may obtain the codebook and share it with the gNB in advance. Therefore, for transmitting the quantized version of Xe, the UE only needs to transmit n-bits, pointing to the index of the selected codeword.
[0071] In the context of supporting quantization and de-quantization blocks in 5G networks the following cases for the evaluation of quantization aware / non-aware training, may be considered, although other cases are not excluded:
[0072] Case 1 : Quantization non-aware training, where the float-format variables are directly passed from CSI generation part to CSI reconstruction part during the training.Fixed / pre-configured quantization method / parameters are applied for the inference phase.
[0073] Case 2: Quantization aware training, where quantization / de-quantization is involved in the training process. Case 2-1 : Fixed / pre-configured quantization method / parameters are applied during the training phase and the same quantization codebook is applied for the inference phase. Case 2-2: The quantization method / parameters are updated in together with the AI / ML models during the training. When the training is finished, the final quantization codebook is applied for the inference phase.
[0074] In CSI compression using two-sided model use case, the following AI / ML model training collaboration types may be considered, although other types are not excluded: Type 1 : Joint training of the two-sided model at a single side / entity, e.g., UE-sided or network-sided.Type 2: Joint training of the two-sided model at network side and UE side, respectively.Type 3: Separate training at network side and UE side, where the UE side CSI generation part and the network side CSI reconstruction part are trained by UE side and network side, respectively.
[0075] The joint training refers to that the generation model and reconstruction model should be trained in the same loop for forward propagation and backward propagation. Joint training could be done both at a single node or across multiple nodes (e.g., through gradient exchange between nodes).
[0076] Separate training may include sequential training starting with UE side training, or sequential training starting with network side training, or parallel training at the UE side and network side.
[0077] Let’s consider a vector quantization (VQ) including N codewords, where each codeword is index with an integer. VQ has the capabilities to adapt the quantization codewords to the distribution of the encoder outputs and / or quantizer inputs and / or quantized encoder outputs. The codewords of a VQ are designed to minimize the quantization loss, the indexing of the codewords does not change the quantization loss, and therefore the indices are assigned randomly to the VQ codewords. The VQ codewords and corresponding indices are shared between the UE and gNB.
[0078] Fig. 2A shows a block diagram of an example scenario. The block diagram of Fig. 2 comprises a CSI matrix 202, an encoder 105, a quantizer 106, a wireless channel 103, a decoder 115, a de-quantizer 116, and a reconstructed CSI matrix. In operation, the CSI matrix 202 is input to the encoder 105 to compress the CSI. The quantizer 106 finds and selects the closest codeword for the compressed CSI. The index of the selected codeword is transmitted over the wireless channel 103. The transmitted bits are received at the de-quantizer 116. The de-quantizer 116 is supposed to receive bits corresponding to the selected codeword and thereby to use the same codeword that was selected in the quantizer 106 to produce de-quantized signal. The de-quantized signal is then decoded in the decoder 115 to obtain the reconstructed CSI matrix 212. However, error in data reception may lead to that the de-quantized 116 interprets the received bits wrongly and uses different codeword from the codebook for the de-quantization task. This may result in significant performance degradation in the CSI reconstruction. Therefore, an auto-encoder using VQ is very sensitive to reception error, and the codeword indexing step needs to be done in an intelligent and robust manner.
[0079] Various embodiments of present disclosure provide a scheme via which two sides of a communication link may agree on a quantization strategy (QS), including a permutation operation to make the QS robust to errors. The proposed scheme is well suited for machine learning (ML)-based CSI-compression, but the scheme can be used also for other use cases that communicate quantized information over the air.
[0080] As codewords are members of a codebook, different arrangements of the codewords into sequences can be considered. Such sequences may be called permutations of the codewords. A permutation does not change the codewords, but it changes the assigned indices to the codewords.
[0081] Various embodiments of present disclosure provide solutions for agreeing on codebook permutation that is to be used in communication of channel-dependent information between two communicating entities. The aim is to determine an optimal permutation, but equally a sub-optimal permutation or less optimal permutation may be determined. The definition of optimal permutation may depend on the optimization task, that is, different optimization tasks may result in different optimal permutations. The optimization task may be minimizing CSI reconstruction error in the presence of feedback errors. Alternatively, the optimization task may be performance optimization such as e.g. optimization of beamforming gain (which is dependent on CSI). In yet another alternative, the objective may be minimizing performance loss when compared to unquantized feedback.
[0082] Determining the permutation that is to be used may require evaluating which permutation provides the optimal or at least sub-optimal results in view of the optimization task so that optimal or at least sub-optimal permutation can be chosen.
[0083] According to some embodiments, the optimal or suboptimal permutation may be determined based on exhaustive search, random search or based on an iterative algorithm. Further, some other mathematical algorithm or optimization method may be used for determining the permutation.
[0084] In exhaustive search, a search over all the possible permutations is done, and the permutation that provides the minimum average distortion is selected. In this way, anoptimal permutation may be determined. This option may however cause high computational complexity. For a codebook with N codewords, N! permutations should be evaluated.
[0085] In random search, M random permutations for the codewords are considered and the codeword that provides the minimum average distortion among the M permutations is selected. In this case there is need to evaluate only M permutations instead of N! permutations needed in the exhaustive search; whereby less computational resources are needed. This method may provide a sub-optimal solution, which is likely to produce good results in many cases.
[0086] In the case of using an iterative algorithm, a binary switching algorithm (BSA) to find a sub-optimal permutation for the codebook may be used. This option may use PseudoGray coding disclosed for example in Zeger et al. Pseudo-Gray Coding IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 38, NO. 12, DECEMBER 1990.
[0087] Evaluating which permutation provides the optimal or at least sub-optimal results in view of the optimization task may be based on various factors, which may include one or more of: distortion for each codeword of the codebook, weighted average distortion for all the codewords for different permutations, distance between codewords of a codebook, a mapping function f() that maps the information into desired space for evaluation. The factors that are used in the evaluation of the permutations may vary depending on the optimization task.
[0088] Various embodiments are discussed in the context of vector quantization (VQ), which may be considered to cover also scalar quantization (SQ). That is, various embodiments may be applied to both VQ and SQ.
[0089] Let’s consider a codebook, C = {c0, c1, ... , cw-1}, including N codewords. In the presence of channel noise, to measure the possible distortion in the de-quantized signal, when the codeword is transmitted, we may define distortion, etaswhere St c c is a set, including Nt codewords that the gNB may receive due to the channel noise. For example, if channel may cause 1 -bit error, St includes codewords from the codebook C that have hamming distance 1 with the codeword , f() is a mapping function that represents potential processing to map the information into the desired space for evaluation and d(f (ct)denotes the distance between the two codewords and Sj.
[0090] The weighted average distortion, e, due to channel noise can be defined aswhere wtdenotes the weights for the distortion of the i-th codeword Q o1 wt = 1)- For example, we may consider equal weighting for the distortion of the codewords (w; = - / ^), or consider weighting proportional to the frequency of usage of each codeword.
[0091] A permutation changes the set St for codeword cbwhich results in different weighted average distortion. The optimal permutation is the permutation that leads to the minimum weighted average distortion. A sub-optimal permutation is a permutation that leads to lower weighted average distortion than at least some other permutation(s).
[0092] An optimal (or sub-optimal) permutation may reduce the weighted average distortion significantly. In an example case, certain permutation in which the indices assignment is not robust to the reception-error, a 1-bit reception-error resulted in average distortion (using Euclidian distance) of e=54. An example of such a codebook is shown in Fig. 2B. The same codebook with another codeword index permutation reduced the average distortion to e=23. The codebook with such codeword index permutation is shown in Fig. 2C. Fig. 2D is a graph that depicts the cumulative density function (CDF) of the average distortion caused by 1-bit reception error with different possibilities for the codewords permutation.
[0093] Several metrics can be considered for measuring the distance between codewords of a codebook, which may result to different result for the optimal permutation with the highest robustness. The measured distance may be defined to have two factors: the mapping function f() and a distance metric. Additionally, weights wj may be included as a third factor. The purpose of the mapping function f() is to emulate the impact of each codeword on the ultimate task, e.g. reconstructed CSI, beamforming gain, etc.. The purpose of the distance metric is to determine the way to define the distance between the mapped version of two codewords.
[0094] The following alternatives can be considered for the mapping function f():Identity function: The mapping function may be defined as an identity function f(x) = x, the distance is calculated directly based on the codewords. This option may be well suited for alternatives in which the UE decides on the VQ codebook and determines the permutation that is to be used.- A non-linear function: The mapping function may be defined as a non-linear function, which reflects the effects of distortion caused by channel noise on the ultimate task, e.g., reconstructed CSI at decoder output or beamforming gain.
[0095] Further, the following is noted for certain embodiments, the impact of an errorcan be evaluated directly on the output of a de-quantizer without any further processing (in case the mapping function f() is an identity function) or based on the impacts on the ultimate task (e.g., how much the CSI-reconstruction at the gNB is affected due to the error in the indexing bits).
[0096] An option for the distance metric is “mean squared error” between the obtained and intended (if the channel would be error free) outputs of the mapping function f(). Other possibilities for the distance metric can be “cosine distance” or “squared cosine distance”, or any other distance metrics.
[0097] In more detail, the following alternatives can be considered for the distance metric:Euclidian distance, which is the most well-known and widely used distance metric.The distance of two vectors a and b can be defined as: d = || a — b\\pwhere p denotes the power to scale the distance. For example, with p=2, we consider squared Euclidian distance.Cosine distance, which measures the difference between two non-zero vectors a and b in an inner product space, i.e. , / a ■ b \pd =tow where p denotes the power to scale the distance. For example, with p=2, we consider squared cosine distance.
[0098] Also, the distance metric can be calculated on the processed version of two codewords. For example, the mapping function f() is applied on the codewords, then one of the distance metrics is used to measure the distance.
[0099] In a further embodiment, the mapping function f() can be set as the functionality / transformation for emulating the reconstructed channel response based on the input codeword. Therefore, f (ct) and f(sj) denote, respectively, the reconstructed channel responses where codewords c, and s, are the outputs of the de-quantizer at the gNB. In this case, to compare the reconstructed channel responses f(ct) and f(st), we may use squared cosine distance as the distance metric for distortion evaluation. In addition, in case the optimization task is not to minimize CSI feedback error, but to optimize performance (like beamforming gain, which is dependent on CSI), the objective might be different, not necessarily a distance to minimize (unless it is distance to optimal beam) but to maximizeperformance (e.g. beamforming gain). Alternatively, the objective may be defined to minimize performance loss when compared to unquantized feedback, for example. It is important to note that the selected objective (selection may be done by gNB or UE) may result in different optimal permutation (quantization labelling or index representation) and needs to be signaled to the other entity.
[0100] Fig. 3 shows an operational graph according to an example embodiment. Fig. 3 shows a first apparatus 301 and a second apparatus 311 that may communicate over a wireless channel. The first apparatus may be a UE and the second apparatus may be a network entity, such as a gNB or BS, or vice versa. In this case, joint training of the autoencoder is performed at the first apparatus.
[0101] At step 3.1 , the first apparatus 301 and the second apparatus 311 coordinate on which entity obtains the mapping function f() and the distance metric for determining the permutation. The aim is to determine optimal permutation, but equally a sub-optimal permutation or less optimal permutation may be determined.
[0102] At step 3.2, the first apparatus 301 trains a CSI-feedback auto-encoder, where a VQ is considered for quantization. As a result, a VQ-codebook is designed / obtained. Also, the first apparatus 301 may determine the permutation for the codeword indices. However, the first apparatus 301 may skip determining the permutation, as determining the permutation may need high computational and memory resources.
[0103] At step 3.3, after finalizing the training phase, the VQ-codebook is shared with the second apparatus 311. At the same time the first apparatus 301 shares the decoder to the second apparatus 311.
[0104] At step 3.4, the first apparatus 301 provides to the second apparatus 311 the permutation determined by the first apparatus 301 or an indication of not determining the permutation in the first apparatus 301.
[0105] At step 3.5, the second apparatus 311 optionally determines a permutation for the VQ-codebook, in any of the following cases:If the permutation for the VQ codebook is not provided by the first apparatus 301.The second apparatus 311 is able to determine a better permutation, as the permutation provided by the first apparatus 301 is not optimal.The second apparatus 311 may consider a new mapping function f() based on the decoder functionalities, which results to different permutation for the codewords in the VQ-codebook.
[0106] At step 3.6, the second apparatus 311 may confirm the permutation provided by the first apparatus 301 or provide a new permutation to the first apparatus 301.
[0107] At step 3.7, the first apparatus 301 and the second apparatus 311 ,respectively, use the permutation and its inverse for quantization and de-quantization tasks. For example, the UE uses the received permutation after quantization, and the gNB uses the inverse of the permutation before de-quantization.
[0108] Figs. 4-5 are flow diagrams of processes according to example embodiments. The processes may be implemented in one or more computer programs comprising instructions for performing the steps of the respective process.
[0109] Fig. 4 shows a process to be performed in a first apparatus which may be a UE or a network entity. The process comprises the following steps:
[0110] 410: Determining a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook.
[0111] The permutation may be determined based on one of the following: exhaustive search, random search, an iterative algorithm. The determination of the permutation may include choosing a desired permutation based on distortion calculated for a plurality of possible permutations. Additionally or alternatively, the determination of the permutation may include choosing a permutation that results in minimum weighted average distortion among plurality of possible permutations.
[0112] 411 : Optionally, obtaining a mapping function and a distance metric and using the mapping function and the distance metric in the determination of the permutation. The mapping function and the distance metric may be obtained by determining at least one of them or both at the first apparatus or by receiving at least one of them or both from a second apparatus. Instead of determining both the mapping function and the distance metric at the first apparatus or receiving both the mapping function and the distance metric from the second apparatus, it is for example possible that the mapping function is determined at the first apparatus and the distance metric is received from the second apparatus or that the mapping function is received from the second apparatus and the distance metric is determined at the first apparatus. Further possibility is to obtain at least one of the mapping function and the distance metric from elsewhere.
[0113] In an embodiment, the mapping function and the distance metric may be used for evaluating different permutations in order to choose the desired (e.g. optimal or sub- optimal) permutation. The mapping function and the distance metric may be usable for calculating distortion for a plurality of possible permutations, whereby desired permutation among a plurality of permutations may be chosen based on the calculated distortion.
[0114] In an embodiment, determining the permutation based on the mapping function and the distance metric includes determining an optimal permutation by using anexhaustive search (brute-force) approach as follows:- For the codebook with N codewords, make a list of all the possible permutations for indexing of the codewords. The total number of permutations is Nt.- For each permutation, obtain the weighted average distortion of the codewords based on the considered mapping function and distance metric.- Select the permutation that provides the minimum weighted average distortion as the optimal permutation.
[0115] 412: Communicating the determined permutation to a second apparatus.
[0116] 413: Optionally, using the permutation in communications between the first apparatus and the second apparatus. The permutation may be used in quantization or dequantization tasks. The permutation may be used in at least one of the following purposes: communicating channel state information, CSI; beamforming; radio resource management; positioning.
[0117] Fig. 5 shows a process to be performed in a second apparatus, which may be a UE or a network entity. The process comprises the following steps:
[0118] 510: Receiving, from a first apparatus, a permutation for quantization or dequantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook.
[0119] 511 : Optionally, determining a mapping function and a distance metric for determination of the permutation and communicating the mapping function and the distance metric to the first apparatus.
[0120] 512: Optionally, using the permutation in communications between the first apparatus and the second apparatus. The permutation may be used in quantization or dequantization tasks. The permutation may be used in at least one of the following purposes: communicating channel state information, CSI; beamforming; radio resource management; positioning.
[0121] Figs. 6-9 show signaling diagrams of according to example embodiments. Figs. 6-9 show signaling between a UE 101 and a gNB 111.
[0122] Fig. 6 relates to an example including a one-sided joint training scheme (type 1), where the gNB determines the permutation and indicates the determined permutation to the UE. The UE may be unable to determine the permutation.
[0123] At 6.1 , the UE 101 collects training data.
[0124] At 6.2, the UE 101 trains a full autoencoder and obtains a VQ codebook.
[0125] At 6.3, the gNB 111 provides a mapping function and a distance metric to the UE 101.
[0126] At 6.4, the UE 101 provides to the gNB 111 and indication of not determining the permutation at the UE. The indication may define that the UE is not able to provide permutation for the VQ codebook.
[0127] At 6.5, the UE 101 provides to the gNB111 the decoder model and weights obtained at step 6.2.
[0128] At 6.6, the UE 101 provides to the gNB111 the VQ codebook obtained at step 6.2.
[0129] At 6.7, the gNB 111 determines permutation for the VQ codebook.
[0130] At 6.8, the gNB 111 provides the permutation to the UE 101.
[0131] At 6.9, the UE 101 estimates the CSI.
[0132] At 6.10, the UE 101 uses the encoder and the considered VQ to compress and quantize the CSI.
[0133] At 6.11, the UE 101 permutes the bits of the compressed VQ after quantization based on the permutation received at 6.8.
[0134] At 6.12, the UE 101 provides permuted bits of the compressed CSI to the gNB.
[0135] At 6.13, the gNB 111 uses inverse permutation on the received bits.
[0136] At 6.14, the gNB 111 uses the VQ codebook to de-quantize the bits.
[0137] At 6.15, the gNB 111 uses the decoder to reconstruct the CSI based on the de-quantized signals.
[0138] Fig 7 relates to an example including a one-sided joint training scheme (type 1), where the UE determines the permutation and indicates the determined permutation to the gNB.
[0139] At 7.1 , the UE 101 collects training data.
[0140] At 7.2, the UE 101 trains a full autoencoder and obtains a VQ codebook.
[0141] At 7.3, the gNB 111 provides a mapping function and a distance metric to the UE 101.
[0142] At 7.4, the UE 101 provides to the gNB 111 and indication of determining the permutation at the UE. The indication may define that the UE is able to provide permutation for the VQ codebook.
[0143] At 7.5, the UE 101 determines permutation for the VQ codebook.
[0144] At 7.6, the UE 101 provides to the gNB111 the decoder model and weights obtained at step 7.2.
[0145] At 7.7, the UE 101 provides to the gNB111 the VQ codebook obtained at step7.2.
[0146] At 7.8, the UE 101 provides the permutation to the gNB 111.
[0147] At 7.9, the UE 101 estimates the CSI.
[0148] At 7.10, the UE 101 uses the encoder and the considered VQ to compress and quantize the CSI.
[0149] At 7.11 , the UE 101 permutes the bits of the compressed VQ after quantization based on the permutation determined at 7.5.
[0150] Steps 6.12-6.15 correspond to respective steps in Fig. 6.
[0151] Fig. 8 relates to an example including a joint two-sided autoencoder training scheme (type 2). The UE determines the permutation and indicates the determined permutation to the gNB.
[0152] At 8.1 , the gNB 111 provides a mapping function and a distance metric to the UE 101.
[0153] At 8.2, the UE 101 provides to the gNB 111 and indication of determining the permutation at the UE. The indication may define that the UE is able to provide permutation for the VQ codebook.
[0154] At 8.3, the UE 101 and the gNB 111 jointly train an autoencoder and obtain a VQ codebook.
[0155] At 8.4, the UE 101 determines permutation for the VQ codebook.
[0156] At 8.5, the UE 101 provides to the gNB111 the VQ codebook obtained at step 8.3.
[0157] At 8.6, the UE 101 provides the permutation to the gNB 111.
[0158] At 8.9, the UE 101 estimates the CSI.
[0159] At 8.10, the UE 101 uses the encoder and the considered VQ to compress and quantize the CSI.
[0160] At 8.11 , the UE 101 permutes the bits of the compressed VQ after quantization based on the permutation determined at 8.4.
[0161] Steps 6.12-6.15 correspond to respective steps in Fig. 6.
[0162] Fig. 9 relates to an example including separate training of encoder at the UE and training of decoder at the gNB (type 3). The gNB determines the permutation and indicates the determined permutation to the UE. The UE may be unable to determine the permutation.
[0163] At 9.1, the UE 101 collects training data.
[0164] At 9.2, the UE 101 trains a full autoencoder and obtains a VQ codebook.
[0165] At 9.3, the gNB 111 provides a mapping function and a distance metric to the UE 101.
[0166] At 9.4, the UE 101 provides to the gNB 111 and indication of not determiningthe permutation at the UE. The indication may define that the UE is not able to provide permutation for the VQ codebook.
[0167] At 9.5, the UE 101 provides to the gNB111 the VQ codebook obtained at step 9.2.
[0168] At 9.6, the UE 101 provides to the gNB 111 the CSI and corresponding latent vectors according to the codebook.
[0169] At 9.7, the gNB 111 trains decoder based on the data received at 9.6 including original CSI and compressed CSI.
[0170] At 9.8, the gNB 111 determines permutation for the VQ codebook.
[0171] At 7.8, the gNB 11 provides the permutation to the UE 101.
[0172] Steps 6.9-6.15 correspond to respective steps in Fig. 6.
[0173] Figs. 10-11 show block diagrams of apparatuses according to example embodiments.
[0174] Fig. 10 shows a block diagram of an apparatus 1000 according to an example embodiment. The apparatus 1000 may operate as a network entity, such as a base station or gNB of Figs. 1-9. The apparatus 1000 generally comprises a memory 1040 including a computer program code 1050. The apparatus 1000 further comprises a processor 1020 for controlling the operation of the apparatus 1000 using the computer program code 1050, and a communication unit 1010 for communicating with other network entities and other devices.
[0175] The communication unit 1010 comprises, for example, one or more of: a local area network (LAN) port; a wireless local area network (WLAN) unit; Bluetooth unit; cellular data communication unit; or satellite data communication unit. The communication interface 1010 may support one or more different communication technologies. The communication interface 1010 may support Ethernet communications and / or IP based communications. The apparatus 1000 may also or alternatively comprise more than one of the communication interfaces 1010. The processor 1020 comprises, for example, any one or more of: a master control unit (MCU); a microprocessor; a digital signal processor (DSP); an application specific integrated circuit (ASIC); a field programmable gate array; and a microcontroller. Various parts may be implemented using more than one corresponding or different elements, such as memories and storages may be multiplied for capacity and / or redundancy purposes. Similarly, processing and / or communications may be implemented with multiple parallel or elements for capacity and / or redundancy purposes.
[0176] The computer program code 1050 may control the apparatus 1000 to provide one or more example embodiments of present disclosure, such as suitable parts of Figs. 1-9.
[0177] Fig. 11 shows a block diagram of an apparatus 1100 according to an example embodiment. The apparatus 1100 may operate as user equipment, such as the UE 101 ofFigs. 1-9. The apparatus 1100 generally comprises a memory 1140 including a computer program code 1150. The apparatus 1100 further comprises a processor 1120 for controlling the operation of the apparatus 1100 using the computer program code 1150, and a communication unit 1110 for communicating with other network entities and other devices. Further, the apparatus 1100 may comprise a user interface unit 1130.
[0178] The communication unit 1110 comprises, for example, one or more of: a wireless local area network (WLAN) unit; Bluetooth unit; cellular data communication unit; or satellite data communication unit. The communication interface 1110 may support one or more different communication technologies. The communication interface 1110 may support IP based communications. The apparatus 1100 may also or alternatively comprise more than one of the communication interfaces 1110. The processor 1120 comprises, for example, any one or more of: a master control unit (MCU); a microprocessor; a digital signal processor (DSP); an application specific integrated circuit (ASIC); a field programmable gate array; and a microcontroller. The user interface unit 1130 may comprise a circuitry for receiving input from a user of the apparatus 1100, e.g., via a keyboard; graphical user interface of a display; speech recognition circuitry; or an accessory device; such as a headset; and for providing output to the user via, e.g., a graphical user interface or a loudspeaker. Various parts may be implemented using more than one corresponding or different elements, such as memories and storages may be multiplied for capacity and / or redundancy purposes. Similarly, processing and / or communications may be implemented with multiple parallel or elements for capacity and / or redundancy purposes.
[0179] The computer program code 1150 may control the apparatus 1100 to provide one or more example embodiments of present disclosure, such as suitable parts of Figs. 1-9.
[0180] As used in this application, the term “circuitry” may refer to one or more or all of the following:(a) hardware-only circuit implementations (such as implementations in only analog and / or digital circuitry) and;(b) combinations of hardware circuits and software, such as (as applicable):(i) a combination of analog and / or digital hardware circuit(s) with software / firmware; and(ii) any portions of hardware processor(s) with software (including digital signal processor(s)), software, and memory(ies) that work together to cause an apparatus, such as a mobile phone or server, to perform various functions); and(c) hardware circuit(s) and or processor(s), such as a microprocessor(s) or a portion of a microprocessor(s), that requires software (e.g., firmware) for operation, but the software may not be present when it is not needed for operation.
[0181] This definition of circuitry applies to all uses of this term in this application, including in any claims. As a further example, as used in this application, the term circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and / or firmware. The term circuitry also covers, for example and if applicable to the particular claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or a similar integrated circuit in server, a cellular network device, or other computing or network device.
[0182] Without in any way limiting the scope, interpretation, or application of the claims appearing below, a technical effect of one or more of the example embodiments disclosed herein is a possibility to improve auto-encoder based CSI compression by making it more robust to reception errors. Various different quantization types may be supported by the embodiments of present disclosure.
[0183] Embodiments of the present disclosure may be implemented in software, hardware, application logic or a combination of software, hardware, and application logic. The software, application logic and / or hardware may reside e.g. on the first apparatus and / or the second apparatus of the foregoing disclosure. In an example embodiment, the application logic, software, or an instruction set is maintained on any one of various conventional computer-readable media. In the context of this document, a “computer-readable medium” may be any non-transitory media or means that can contain, store, communicate, propagate, or transport the instructions for use by or in connection with an instruction execution system, apparatus, or device, such as a computer, with one example of such device described and depicted in Figs. 10 and 11. A computer-readable medium may comprise a computer-readable storage medium that may be any media or means that can contain or store the instructions for use by or in connection with an instruction execution system, apparatus, or device, such as a computer.
[0184] If desired, the different functions discussed herein may be performed in a different order and / or concurrently with each other. Furthermore, if desired, one or more of the before-described functions may be optional or may be combined.
[0185] Although various aspects of present disclosure are set out in the independent claims, other aspects may comprise other combinations of features from the described embodiments and / or the dependent claims with the features of the independent claims, and not solely the combinations explicitly set out in the claims.
[0186] It is also noted herein that while the foregoing describes example embodiments, these descriptions should not be viewed in a limiting sense. Rather, there are several variations and modifications which may be made without departing from the scope ofthe present disclosure as defined in the appended claims.LIST OF ABBREVIATIONSBS base stationUE user equipment3D 3-dimensionalCSI Channel state informationMIMO multiple-input multiple-outputCR compression rateCS compressive sensingAE autoencoderDL deep learningCDF cumulative distribution function
Claims
CLAIMS1. A first apparatus comprising at least one processor; and at least one memory including executable instructions that, when executed by the at least one processor, cause the first apparatus to determine a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook; and to communicate the permutation to a second apparatus.
2. The first apparatus of claim 1 , wherein the first apparatus is caused to determine the permutation based on one of the following: exhaustive search, random search, an iterative algorithm.
3. The first apparatus of claim 1 or 2, wherein the first apparatus is caused to determine the permutation by choosing a desired permutation based on distortion calculated for a plurality of possible permutations.
4. The first apparatus of any preceding claim, wherein the first apparatus is caused to determine the permutation by choosing a permutation that results in minimum weighted average distortion among plurality of possible permutations.
5. The first apparatus of claim 1 , 2 or 3, wherein the first apparatus is caused to obtain a mapping function and a distance metric; and use the mapping function and the distance metric in the determination of the permutation, wherein determining the permutation comprises calculating distortion for a plurality of possible permutations based on the mapping function and the distance metric and choosing desired permutation based on the calculated distortion, wherein the mapping function is a function that transforms an input from one space into an output belonging to another space, wherein the distance metric is defined for being applied in the another space.
6. The first apparatus of claim 5, wherein the first apparatus is caused to obtain at least one of the mapping function and the distance metric by receiving the at least one of the mapping function and the distance metric from the second apparatus or bydetermining the at least one of the mapping function and the distance metric at the first apparatus.
7. The first apparatus of any preceding claim, wherein the first apparatus is caused to determine the permutation responsive to receiving, from the second apparatus, an indication of the permutation not being determined by the second apparatus.
8. The first apparatus of any preceding claim, wherein the first apparatus is caused to use the permutation in communications between the first apparatus and the second apparatus.
9. The first apparatus of any preceding claim, wherein the first apparatus is caused to use the permutation in at least one of the following: communicating channel state information, CSI; beamforming; radio resource management; positioning.
10. The first apparatus of any preceding claim, wherein the first apparatus is a user equipment and the second apparatus is a network entity, or vice versa.
11. A second apparatus comprising at least one processor; and at least one memory including executable instructions that, when executed by the at least one processor, cause the apparatus to receive, from a first apparatus, a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook.
12. The second apparatus of claim 11 , wherein the second apparatus is caused to determine at least one of a mapping function and a distance metric for calculating distortion for a plurality of possible permutations; and to communicate the at least one of the mapping function and the distance metric to the first apparatus, wherein the mapping function is a function that transforms an input from one space into an output belonging to another space, wherein the distance metric is defined for being applied in the another space.
13. The second apparatus of any one of claims 11-12, wherein the second apparatus is caused, responsive to receiving the permutation, to determine a new permutation and communicate the new permutation to the first apparatus.
14. The second apparatus of any one of claims 11-13, wherein the second apparatus is caused to use the permutation in communications between the first apparatus and the second apparatus.
15. The second apparatus of any one of claims 11-14, wherein the second apparatus is caused to use the permutation in at least one of the following: communicating channel state information, CSI; beamforming; radio resource management; positioning.
16. The second apparatus of any one of claims 11-15, wherein the first apparatus is a user equipment and the second apparatus is a network entity, or vice versa.
17. A method, comprising: determining, at a first apparatus, a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook; and communicating the permutation to a second apparatus.
18. A non-transitory computer readable medium comprising program instructions stored thereon for performing at least the following at a first apparatus: determining a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook; and communicating the permutation to a second apparatus.
19. A method, comprising: receiving, at a second apparatus from a first apparatus, a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communications network, wherein the permutation defines a sequence for arranging codewords of a quantization codebook.
20. A non-transitory computer readable medium comprising program instructions stored thereon for performing at least the following at a second apparatus: receiving, from a first apparatus, a permutation for quantization or de-quantization tasks or both in communications of channel-dependent information in a communicationsnetwork, wherein the permutation defines a sequence for arranging codewords of a quantization codebook.
Citation Information
Patent Citations
Efficient vector quantizer for FD-MIMO systems
US20160119045A1
