Page 51 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 6 – Wireless communication systems in beyond 5G era

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

1.E+00
1.E+00
NR LDPC
Deepcode, [1] 1.E-01
1.E-01
Deepcode (reproduced)
DEF code (ext.enc.)
1.E-02 Deep-LSTM code 1.E-02
DEF-LSTM code (ext enc/dec)
DEF-LSTM code (ext.enc) 1.E-03
1.E-03
BLER 1.E-04 BLER 1.E-04
1.E-05
1.E-05
1.E-06
1.E-06
NR LDPC
1.E-07 Pseudo-Deepcode
1.E-07
DEF-LSTM code, ext. input
1.E-08
1.E-08 3 4 5 6
-1 0 1 2 SNR [dB]
SNR [dB]
Fig. 4 – Performance comparison of Deepcode, DEF codes, LSTM‑based Fig. 5 – Performance comparison of Deepcode, pseudo‑Deepcode, and
Deepcode, and DEF‑LSTM codes. Spectral ef iciency is 0.67 bits/s/Hz DEF‑LSTM code with extended encoder input. Spectral ef iciency is 1.33
( = 2, = 2) . bits/s/Hz ( = 4, = 2).

– produces signi icant gains. The investigation of perfor‑ • The DEF‑LSTM code with extended encoder input
mance with larger feedback extensions is left for future and DEF‑LSTM code with extended encoder/decoder
work. Details of the evaluated architectures are reported input have similar performance except for high SNRs,
in Table 5. where the former performs slightly better.
Fig. 4 shows the Block Error Rate (BLER) vs. forward SNR • DEF‑LSTM codes (green and blue curve) outperform
of several codes with = 0.67 bits/s/Hz. The plot NR LDPC (dashed black curve) by at least three or‑
shows Deepcode [1] (pink curve), Deepcode obtained by ders of magnitude BLER for all SNRs.
the training method of Section 3 (solid black curve), DEF
code with extended encoder input (orange curve), Deep‑ • The training method of Section 3 (black curve) pro‑
code with LSTM‑based encoder and decoder NNs (pur‑ duces codes with better performance than the train‑
ple curve), DEF code with extended encoder input (green ing method of [1] (pink curve).
curve) and DEF code with extended encoder and decoder
Based on the irst observation above, it can be concluded
input (blue curve). All DNN‑based codes use second‑
that encoder input extension produces performance im‑
order modulation (i.e., = 2) and = 2 parity sym‑
provements. Subsequent observations highlight that the
bols per systematic symbol. Thus, the corresponding SE
encoder input extension provides performance improve‑
is 0.67 bits/s/Hz. The performance of the NR LDPC code
ments when combined with LSTM. However, based on the
as reported in [12] with the same SE (QPSK modulation, observation in the third bullet, we can conclude that de‑
code rate 1/3) is shown by a dashed black curve. coder input extension brings no bene its compared to en‑
Based on the data shown in Fig. 4, the following observa‑
coder input extension. Moreover, the above performance
tions are made:
evaluations show that usage of LSTM in the encoder and
• The DEF code with extended encoder input (orange decoder provides signi icant performance improvements
curve) has better performance than Deepcode (solid compared to RNN/GRU based codes.
black curve). Figure 5 shows the BLER performance of DNN‑based
codes with modulation order = 4, the correspond‑
• The DEF‑LSTM codes (green and blue curves) have
ing SE is 1.33 bits/s/Hz. As Deepcode [1] is not de ined
the best performance among all the evaluated codes. for SEs higher than 0.67 bits/s/Hz,

Table 4 – Evaluation parameters. Table 5 – Evaluated architectures.

DEF code parameter Selected values Code Encoder NN Decoder NN
[symbols] 50 (type, #layers) (type, #layers)
2 Deepcode RNN, 1 bidir. GRU, 2
0 50 DEF code RNN, 1 bidir. GRU, 2
# zero‑padding bits 1 Deep‑LSTM code LSTM, 1 bidir. LSTM, 2
Encoder input extensions ( , , ) = (1, 2, 2) DEF‑LSTM code LSTM, 1 bidir. LSTM, 2
2
1
0
Decoder input extensions ( , , ) = (1, 1, 1)
2
1
0
© International Telecommunication Union, 2021 39