Page 36 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 4 – AI and machine learning solutions in 5G and future networks

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 4

-17.5 -20 -24
-20.5
-21
NMSE [dB] -18 NMSE [dB] -21.5 NMSE [dB] -24.5

-18.5
-22
-22.5 -25
-19 -23 -25.5
0.7 0.75 0.8 0.85 0.9 0.95 1 0.7 0.75 0.8 0.85 0.9 0.95 1 0.7 0.75 0.8 0.85 0.9 0.95 1
Detection threshold δ Detection threshold δ Detection threshold δ

Fig. 6 – NMSE behavior over the decision threshold of the 3 training datasets.
Table 1 – NMSE table for training data
SNR (dB) Algorithm −15 −10 −5
SW‐OMP −1.45 dB −5.70 dB −9.68 dB
Pilot Frames: 20 MLGS‐SBL −4.29 dB −9.13 dB −12.34 dB
PCSBL‐DDT −8.16 dB −10.62 dB −11.07 dB
PC‐OMP −8.34 dB −12.36 dB −16.15 dB
SW‐OMP −3.95 dB −7.95 dB −11.87 dB
Pilot Frames: 40 MLGS‐SBL −7.55 dB −11.19 dB −14.15 dB
PCSBL‐DDT −10.56 dB −12.14 dB −12.62 dB
PC‐OMP −12.66 dB −16.33 dB −19.78 dB
SW‐OMP −7.33 dB −11.60 dB −15.63 dB
Pilot Frames: 80 MLGS‐SBL −13.02 dB −16.37 dB −18.94 dB
PCSBL‐DDT −11.90 dB −13.10 dB −13.63 dB
PC‐OMP −18.70 dB −21.49 dB −24.48 dB

Table 2 – NMSE table for test data
SNR (dB) Algorithm [−20, −11) [−11, −6) [−6, 0]
MLGS‐SBL −7.66 dB −10.97 dB −12.34 dB
Pilot Frames: 20 PCSBL‐DDT −8.94 dB −9.99 dB −10.31 dB
PC‐OMP −9.09 dB −12.45 dB −14.22 dB
MLGS‐SBL −11.87 dB −12.79 dB −14.20 dB
Pilot Frames: 40 PCSBL‐DDT −10.82 dB −11.33 dB −11.89 dB
PC‐OMP −13.79 dB −15.24 dB −16.79 dB
MLGS‐SBL −13.62 dB −16.23 dB −20.08 dB
Pilot Frames: 80 PCSBL‐DDT −11.74 dB −12.47 dB −12.98 dB
PC‐OMP −16.32 dB −19.07 dB −23.91 dB

6. NUMERICAL RESULTS We note that while the three new algorithms presented
in this paper have been ine‐tuned based on the training
dataset, the baseline algorithm, SW‐OMP, has been imple‐
w discuss performance of
mented as‐is from the literature. On the other hand, in our
proposed tr testing
implementation of SW‐OMP, we consider the case where
data generated Raymobtime, ra tr based
the true AoDs and AoAs are contained in the sparsifying
mmWave channel generation tool. We train the mmWave
dictionary. While the proposed algorithms do suffer from
channel estimation algorithms using 10, 000 independent
off‐grid effects, SW‐OMP insulated
r consisting of 100 be‐
from the performance degradation caused by them.
tw T Rx. Mor about channel
generation methodology can be found in [24 We used
20 40 and 80 pilot frames during both the training and
testing phases of the proposed For the train‐
ing phase, we used SNR values of {−15, −10, −5} dB. We
performance of proposed al‐
gorithms reference state‐of‐the‐art model‐based
greedy search algorithm called SW-OMP [4].

20 © International Telecommunication Union, 2021