Page 46 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 6 – Wireless communication systems in beyond 5G era
P. 46

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 6




                                                                Transmitter        Receiver
                                                                     1 st
                   Modulator                                        phase
                                 Buffer
                                      (  )                 Power reall. (a,w)

                          Parity Symbol Generator                   2 nd
                                th
                               (k iteration)                       phase                  1 st
                                                                                         phase      
                                                                                                       ...
                                                                                                       ...  ...  ...
                                                                D
                                            0 (   −    0 )  ...     0 (   − 1)     0 (  )  phase         −1  ...
                                                                                                     ...
                                                                                                    ...
                                                                                          2
                                                                                           nd
                                ...                         -  +        1 st
                    ...          0 (   −    1 )     0 (   − 2)     0 (   − 1)  D  ...  D  Buffer     0  S  phase  Codeword
                      −1 (   −       )       −1 (   − 2)       −1 (   − 1)  D  ...  D     0  ...  De-  - S  +  phase
                                                                       nd
                                                                      2
                 D  ...  D                           −1  mux       −1         −1
          Fig. 1 – DEF encoder structure. Each “  ” block represents a unit‑time delay. Blue blocks and signals denote new functionalities compared to prior
          solutions.


          approach could be used in the decoder to extend its input   denotes the sub‑vector that contains the elements of b
          so as to comprise delayed versions of received signals in a   with indices in [   ∶   ]. Boldface upper case letters like
          given time window. However, it can be shown that such a   A denote matrices;      ,    represents the element of A in the
                                                                th
                                                                          th
          generalization of the decoder does not bring any bene it      row and    column. Notation   (v), where    is a func‑
          and therefore it will not be considered in the de inition of   tion taking a scalar input, indicates the vector obtained by
          DEF codes.  The extended‑feedback encoder architecture   applying    to each element of v. Hadamard (i.e., element‑
          is combined with different NN architectures of recurrent   wise) product is denoted by ∘.
          type, namely RNN, GRU and LSTM. The DEF code general‑
          izes Deepcode [1] along several directions. Its major ben‑   2.  DEFINITION OF DEEP EXTENDED FEED‑
          e its can be summarized as follows:
                                                                    BACK CODE
            • Improved error correction capability obtained    The Deep Extended Feedback (DEF) code is the set of
             by feedback extension. The DEF code generates     codewords produced by the DEF encoder shown in Fig. 1.
             parity symbols based on feedbacks in a longer time  Blue blocks and signals in Fig. 1 denote the new func‑
             window, thereby introducing long‑range dependen‑  tionalities of the DEF code compared to Deepcode [1], ex‑
             cies between parity symbols. As the above long‑   tended feedback is shown by the unit‑time delay blocks la‑
             range dependencies are a necessary ingredient of  beled “  ” and their corresponding input/ouput signals;
             all good error correcting codes, it is expected that  QAM/PAM symbols are produced by the block labeled
             feedback extension will bring performance improve‑  “Modulator”. DEF code and Deepcode operate according
             ments.
                                                               to the same encoding procedure as described later on.
            • Higher spectral ef iciency obtained by usage     The novel DEF code features will be treated in dedicated
             of QAM/PAM modulations. The DEF code uses         subsections.
             Quadrature Amplitude Modulation (QAM) with ar‑    The encoding procedure consists of two phases.  In
             bitrary order, thereby potentially achieving higher  the  irst phase, an   ‑bit information message m  =
             spectral ef iciency.                              (  (0), … ,   (  −1)) is mapped to a sequence of real sym‑
                                                               bols x = (  (0), … ,   (   − 1)), hereafter called systematic
          In this work, we initially focus on DEF codes’ performance  symbols.
          evaluation over channels with noiseless feedback, where  The modulation sequence x is transmitted on the forward
          the forward‑channel output observations are sent uncor‑  channel. The corresponding sequence ̄x observed by the
          rupted to the encoder.                               receiver is given by
          Notation: Lower case and upper case letters denote
          scalar (real or complex) values. For any pair of positive                ̄ x = x + n 0             (1)
          integers    and    with    <   , [   ∶   ] denotes the sequence of
          integers [  ,    + 1, … ,   ], sorted in increasing order. Bold‑  where n 0  represents Additive White Gaussian Noise
          face lower case letters (e.g., b) denote vectors; unless oth‑  (AWGN) and other possible forward‑channel impair‑
          erwise speci ied, all vectors are assumed to be column  ments. In the performance evaluations of Section 4, n is
                                                                                                             0
                                th
          vectors.   (  ) denotes the    element of b; b(   ∶   ),    <   ,  modeled as a sequence of white Gaussian noise samples.



          34                                 © International Telecommunication Union, 2021
   41   42   43   44   45   46   47   48   49   50   51