Page 37 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 4 – AI and machine learning solutions in 5G and future networks
P. 37
ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 4
MLGS‐SBL we can predict when a change in the solution will happen
based on the selected threshold. Knowing this, we create
anal tr w maxi‐
a modi ied version of our approach that saves all chan‐
mum number of paths obtained from MLGS to = 10,
nel iterations together with the computed threshold re‐
and the number of levels in MLGS to = 5 We set the
number of columns in the initial AoD/AoA steering ma‐ quired for them to pass up to a minimum threshold value,
in our case 0.7. Once outside the function, we can evalu‐
trices to 256 We use the estimated noise variance after
ate these channel estimations and compute the error as
SBL to threshold the number of dominant delay taps of the
channel denoiser As this approach is primarily a model‐ a step‐wise function of the threshold. Then, we apply the
average operation to the error step‐wise functions for dif‐
based and uses few statistics from the training
ferent scenarios to get a better and smoother result of
data, it is suitable for general mmWave channel estima‐
the error behavior over different threshold values. Fig. 6
tion problems Further, the thresholds are set keep‐ shows the smoothed step‐wise error function for the dif‐
ing in mind the computational complexity of the MLGS‐
ferent datasets. The selected threshold value is ≈ 0.98.
B incr of out‐
The algorithm with the custom detection method is ap‐
put by MLGS, we can potentially improve the performance
plied to the test data and the obtained results are shown
of the algorithm, but at the cost of higher computational
in Table 2. The PC‐OMP algorithm outperforms the other
complexity. We include the NMSE values obtained for the
training and testing datasets in Table 1 and Table 2 re‐ two algorithms for the different data sets, as can be ob‐
served from the table. Specially, in lower SNRs where
spectively The inal performance score achieved, which
the channel estimation performance is lower due to noise,
is a weighted combination of the NMSE performance in
the PC‐OMP algorithm achieves gains of up to 3 dB com‐
Table 2 when the number of pilot frames is 20, using our
pared with the other two algorithms. At higher SNRs, e.g.,
proposed algorithm in the mmWave channel estimation
[−6, 0] dB, PC‐OMP outperforms the other two algorithms
challenge is −9.16 dB.
by up to 4 dB.
PCSBL‐DDT The inal performance score on the test dataset in the
channel estimation challenge of the PC‐OMP algorithm is
appr w adopted EM‐based sparse
−10.64 dB, outperforming the MLGS‐SBL and the PCSBL‐
Bayesian learning method to exploit the shared sparsity
betw differ dela taps pat‐ DDT methods by 1.48 dB and 1.15 dB, respectively. Also,
from Table 1, we can see that the PCSBL‐DDT and the
tern couplings between consecutive AoAs and AoDs. We
PC‐OMP algorithms are tuned better than the MLGS‐SBL
applied the algorithm to the time‐domain received signals
by only retaining the dominant delay taps to increase ef‐ method for the training dataset that result in their bet‐
fectiv signal used t form esti‐ ter NMSE performances at SNR −15 dB and pilot frames
{20, 40}. But the performance gap between MLGS‐SBL
mate. First, we used the pattern‐coupled Sparse Bayesian
and PCSBL‐DDT reduces for the testing data for SNR
t ground‐truth the
[−20, −11) dB and 20 pilot frames. Moreover, MLGS‐SBL
training dataset by adding a small noise to regularize the
performs better than PCSBL‐DDT at SNR [−20, −11) dB
data. In this way, we obtained the sparse representations
and 40 pilot frames. This can be attributed to the fact
for all the channels in the provided dataset Then, using
that extracting more features from a training dataset may
the respective sparse vectors and exploiting the density
result in an excellent performance during training but
map of joint AoA/AoD grids, we selected a non‐uniform
grid r ined patt betw hyper‐ slightly inferior performance while testing. This shows
that a good model‐based signal processing solution has to
parameters. The algorithm is applied to the test dataset
be combined with appropriate training, while taking into
to obtain the channel estimat The inal performance
account the training and testing performance trade‐off.
score in the channel estimation challenge is −9.49 dB.
The NMSE values for the speci ic scenarios are shown 7. CONCLUSION
in Table 2 for the testing dataset.
W hav presented thr nov signal processing ap‐
PC‐OMP
proaches t estimat mmWav hybrid
Befor evaluating performance of PC‐OMP w opti‐ analog‐digital MIMO W hav adapted model‐
mize value of det threshold value de‐ driv procedures t AoD A channel
scribed 5.2 order t improv results. gain information from a training dataset, and ine‐tuned
W creat ic method for struc‐ the algorithms to reduce the NMSE in the testing dataset.
tured problem that arises in our approach. This optimiza‐ We empirically showed that our algorithms unanimously
performed better purel model‐based approach
tion method is focused on reducing the optimization time
b lar mar giv tr data set Hence,
while being able to perform a high‐resolution grid search
approaches potentiall used
for the parameter values. We base our training algorithm
model‐driven approaches to
on the fact that our approach is a greedy algorithm with
ine‐tune them and thereby obtain better performance in
carefull chosen st condition. This that
physical layer wireless communication problems in real-
istic channel environments.
© International Telecommunication Union, 2021 21