Logistic regression for BLER prediction in 5G

Juan Carlos Ruiz-Sicilia, Mari Carmen Aguayo-Torres

[email protected], [email protected]
Dpto. Ingenierı́a de Comunicaciones. Universidad de Málaga. ETSI Telecomunicación, 29071, Málaga, Spain

Abstract—In this work, a block error rate (BLER) predictor Scheduler

for 5G based on logistic regression is presented. The regression
is fed with transmission parameters and channel statistics. Transport block Block error indication
With these features, the predictor can model the behaviour of
the transmission chain, including the low parity channel code TB CRC attachment TB CRC checking
(LDPC). In particular, for each modulation and coding scheme
Segmentation in CBs  CBs concatenation 
(MCS), the regression model uses as features the mean of the
SINR over the allocated resources and the squared distance to CRC attachment per CB CRC checking per CB
the mean. Moreover, a single model able to cope with a set of
modulation and coding schemes (MCSs) at the expense of certain Channel coding Channel decoding
accuracy loss is also proposed, and its performance evaluated.
Possible applications for the regression models such as end-to- Rate matching Zero padding
end modelling or as part of the adaptive modulation and coding
(AMC) function are explored. Results show that the model has CBs concatenation CBs segmentation
excellent accuracy in a wide set of scenarios.
Scrambling Descrambling

I. I NTRODUCTION Modulation Demodulation

In recent years, the main aspects of 5G have been defined Mapping to PRBs PRBs demapping
and deployment of the network has begun [1]. One of the most OFDM mod OFDM demod
important features of this standard is the distinction between
three types of service requirements depending on the use
Channel (, fD, s)
case: enhanced Mobile Broadband (eMBB), massive machine
type communications (MMTC) and ultra-reliable low-latency Fig. 1. Transmission chain for a transport block along one slot.
communications (URLLC). Due to the huge differences be-
tween their requirements, a fine optimization of the radio
access network (RAN) is required. Machine learning (ML) The rest of the paper is organized as follows. First, the
techniques are considered as the way to reach the optimization system model is presented. Then, two BLER prediction mod-
and one of the main keys for 6G development [2]. els based on logistic regression are described and the dataset
A wide set of works can be found in literature which used to train is examined. In Section IV, their performance is
proposes the use of ML in different areas of wireless networks evaluated. Finally, some concluding remarks are exposed.
[3]. ML techniques allow not only implementing functionali-
ties but also system modelling. Within the physical and link II. S YSTEM MODEL
layer field in 5G, block error rate (BLER) modelling is easily At 5G, transmission at both downlink and uplink is carried
identified as a candidate for ML application [4]. out by modulating QAM symbols via orthogonal frequency
A transport block in 5G [5] is a set of bits jointly coded by division multiplexing (OFDM) [6]. The total number of avail-
a low parity channel code (LDPC) after appending a cyclic able subcarriers as well as the subcarrier spacing, ∆f , can be
redundancy check (CRC). BLER modelling can be described adjusted to the specific scenario. Specifically, ∆f is given by
as the estimation of the probability for a transport block 15 · 2µ (kHz), being µ the system numerology. One slot is
to reach the receiver with errors, that is, BLER modelling formed by M OFDM symbols and its duration is also variable:
considers link-level transmission. At that stage, the influence as ∆f grows, the slot lasts shorter. A simplified transmission
of implementation details are huge, thus evaluating perfor- chain at each slot is shown in Fig. 1.
mance has to be done per equipment version. A good BLER The scheduler (out of the scope of this paper) allocates
model should be accurate while simple enough to allow a user L physical resource blocks (PRBs) to the user during one
equipment (UE) to carry out the predictor calculation. slot. The transport block is thus transmitted over N resource
This work presents a technique to predict the BLER by elements (REs) in the time-frequency grid of one slot, with
logistic regression based on some transmission parameters and N evaluated as M · L minus those REs devoted to signalling
channel statistics. This could be useful in several 5G functions or reference signals. Moreover, the scheduler assigns the
such as adaptive modulation and coding (AMC). An analysis modulation and coding scheme (MCS) to be used, which in
of this extremely important technology in 5G serves as test turns determines the size of the QAM modulation as well as
for the instantaneous BLER model. Another utility for BLER the coding rate to be employed.
prediction is serving to abstract the link-level at a system level The transport block (TB) in Fig. 1 is formed by transport
design. block size (TBS) bits. The TBS is determined by the number
of allocated REs N and the spectral efficiency of the employed 0.16
MCS1 . The TB is appended a CRC which is checked at
the receiver to detect errors. In 5G, each transport block is 0.14

segmentated in a set of code blocks (CBs), each with its own

0.12
CRC in order to allow retransmissions per code block or CB
group2 .
0.1
The resulting coded bits are used to modulate the QAM
symbols and create the OFDM waveform after mapped to the 0.08
allocated PRBs. At downlink, data from other users might
be multiplexed over the same OFDM resource grid, while at 0.06
uplink zeros are set over those unused PRBs. However, for
this work, it is supposed only one user. 0.04
In this work, the Tapped Delay Line Channel (TDL)
channel model as given in [7] has been considered with delay 0.02
spread given by τs . In this multipath scenario, the received
signal after OFDM demodulation can be written as 0
-10 -5 0 5 10 15 20 25 30

yn = hn xn + ωn (1)
¯ R = 20
Fig. 2. PDF of σ from a TDL-A channel with fD = 50 Hz and SN
where xn is the QAM complex symbol transmitted over dB
the nth RE and ωn is the Gaussian noise and interference.
hn is the complex channel response for each RE. At the
same OFDM symbol, correlation among channel responses at N N
different subcarriers is given by the Fourier transform of the 1 X 2 1 X
γ= γn ; σ = (γn − γ)2 . (4)
power delay profile (PDP) of the channel [6]. At the same N n=1 N n=1
subcarrier, correlation among channel responses at different The construction of the feature vector Ψ based on these
OFDM symbols depends on the Doppler frequency fD . statistics allows the full consideration of the channel response.
Note that, although OFDM is thought to avoid the inter- On one hand, γ is the traditional parameter for error rate
ference between symbol and subcarriers, adverse effects of determination given a channel. Note that γ mean is the average
wireless channels such as high Doppler shift or a high time channel SINR, notated by γ̄ in this work. As more REs are
dispersive channel response produce intersymbol interference taken, more similar its value will be to γ̄. On the other hand,
(ISI) and intercarrier interference (ICI) [6]. Thus, ωn is not σ describes how fast channel changes in the allocated REs.
only thermal noise but includes other RE interference. Its An example of the probability density function (PDF) of this
power can be measured from the demodulation reference second statistic can be observed in Fig. 2. In this figure, it is
symbols (DRS) as [8] shown the incidence of a variation of the delay spread in σ. A
NT −1 longer delay spread implies a minor coherence bandwidth and,
1 X
Neq = |yn − hn xn |2 , (2) hence, faster variations within the TB. The same behaviour
NT n=0 is obtained varying the Doppler shift: as it increases, the
with NT the number of RE over which the averaging has coherence time is reduced so σ also grows.
been carried out. By this estimation, and being S the received In the first model proposed, named as model A, instanta-
power of the signal, the signal-to-interference-plus-noise ratio neous BLER for the ith MCS is modeled using as feature
(SINR) for each RE can be calculated as vector Ψ = [1, γ, σ], that is,
1
S iBLERi (γ1 , γ2 , ..., γn ) ≈ Bi (γ, σ) = .
γn = |hn |2 . (3) 1 + exp(−αiT Ψ)
Neq (5)
(i) (i) (i)
III. BLER PREDICTION The values for αiT = (α0 α1 α2 ) are to be found
to approximate the BLER for a specific MCS used over N
In this work, it has been employed the logistic regression to
resource elements and using a certain subcarrier spacing. One
model the BLER. This ML technique is accurate in situations
set of coefficient has to be stored in memory per MCS to be
where the output of the estimation is a probability.
predicted.
Considering the aforementioned transmission model, an as- Another scheme for the logistic regression, referred here-
sociated vector of SINR [γ1 , γ2 , γ3 , ..., γN ] could be assigned inafter as model B, is presented in order to include the MCS
to each TB. In order to compress it into a couple of scalars, in the feature vector. By taking Ψ = [1, γ, σ, i], only one
it is possible to define the average SINR of the TB, γ, and set of coefficients αT = (α0 α1 α2 α3 ) has to be stored
the average squared distance to the mean, σ, which are given in memory. As counterpart, generalization in the logistic
by: regression implies certain accuracy loss.
1 We assume single antenna transmission and reception, but the proposed
A. Dataset generation and analysis
modeling method can be easily extended to multiple input multiple output
procedures. It is necessary to collect a huge amount of data in order to fit
2 Retransmissions are not taken into consideration in this work. the logistic regression storing γ, σ and an indicator regarding
 TABLE I
S IMULATION PARAMETERS FOR DATASET GENERATION

Parameter Values
Channel Model TDL-A, TDL-B, TDL-E
Delay spread (τs ) 100, 200, 400 (ns)
Doppler shift (fD ) 50, 210, 370 (Hz)
Number of allocated RE 504
MCS index From 0 to 28
∆f 30 kHz

if the TB has been corrupted or not for each transport block. A

set of simulations have been carried out varying the Doppler
shift and the delay spread. Channel parameters are all the
possible combinations from those shown in Table I. Note (a) MCS = 10
that also values of τs longer than the cyclic prefix has been
simulated, producing ISI and ICI. intercarrier interference also
appears at high values of fD .
Once data is prepared, it is opportune to analyze its
characteristics before its use to train the logistic regression. As
illustrated in Fig. 3, there are two clearly differentiated regions
what support the accuracy of γ and σ as descriptors of the
channel response. It has been used the logarithm scale so as
to improve the fit by limiting the variability of the features,
specially γ. Furthermore, for low MCSs, the bound which
separates the error/non-error regions is approximately linear,
while higher MCSs have a more non-linear behaviour. This
causes a better fit for low MCSs.

IV. R ESULTS
A. Average BLER modelling
(b) MCS = 26
System level simulators are usually too complex to allow
Fig. 3. Scatter plot of data used to train regression for TB size of 6 PRBs.
detailed link level simulations. The average BLER is easily
obtained by time averaging:
the BLER below a predefined target which depends on the
B̄i (γ̄) = E[Bi (γ, σ)] (6)
service, that is,
Results will depend on the distribution of both regression
M CS = max{i, Bi (γ, σ) < BLERT } (7)
parameters, which in turn are function of the specific PDP
and Doppler frequency. where BLERT depends on the service, being 0.1 for eMBB
As shown in Fig 4, approximately the same average BLER services and 10−5 for URLLC services.
is obtained by both regression averaging and simulation. This As illustrated in Fig. 6, this algorithm can track the channel
figure illustrates the importance of σ as it is the distinguish- by selecting robust MCSs when there is a fading and high
ing element between both channels. Apart from this, it is efficient ones as the SINR improves. Observed delay of one
especially relevant the case of the channel with 400 ns of transport block between SINR and MCS is due to the causal
delay spread, which is longer than the cyclic prefix of OFDM. decision. AMC performance has a high dependence with the
Hence, there is a noise floor due to ISI which is detected by coherence time, being reduced as Doppler shift grows.
the logistic regression and the adequate noise measurement. Spectral efficiency is one of the most characteristic statistic
Fig. 5 shows a comparison between the proposed models. to measure the performance of AMC. In Fig. 7, it is shown the
As expected, the highest the MCS is, the worse performance reduction of performance as Doppler increases. In this case, it
both regression models have. Nevertheless, it is observed that is not appreciable the difference between model A and model
the regression fits better using model A than model B, as we B.
use different models per MCS.
V. C ONCLUSIONS AND FUTURE WORK
B. MCS selection In this work, it has been introduced a technique for
Probably, the most useful application for the instantaneous BLER prediction based on logistic regression. Furthermore,
BLER prediction model is its use in AMC. In short, AMC two possible models have been contrasted. Although for the
algorithm decides the most accurate MCS in order to keep average BLER the differences between them are appreciable,
 100 30 5.5
0 7 14 20
2 8 15 21 5
25
3 10 16 22
4.5
4 11 17 23
20 5 12 18
10-1 6 13 19
4

15 3.5

3
10
10-2 2.5

5 2

1.5
0
-3
10 1
-5
0.5

-10 0
0 50 100 150 200 250
-4
10
0 5 10 15 20 25 30 35 40

Fig. 6. The MCS decision for BLERT = 0.1 based on model A for a
¯ R = 7.5 dB.
TDL-A with fD = 50 Hz, τs = 150 ns and SN
Fig. 4. Comparison between simulations and predictions using a TDL-A
with fD = 50 Hz and MCS = 17.
3

100
2.5

2
10-1

1.5

10-2
1

0.5
10-3

0
0 5 10 15 20 25 30 35 40
-4
10
0 5 10 15 20 25 30 35 40
Fig. 7. Comparison of spectral efficiency for BLERT = 0.1 between model
A and model B and different fD using a TDL-A channel with τs = 100 ns.
Fig. 5. Comparison between simulations and predictions using a TDL-A
with fD = 50 Hz and τs = 150 ns.
R EFERENCES
[1] M. Shafi, A. F. Molisch, P. J. Smith, T. Haustein, P. Zhu, P. De Silva,
error prediction by model B is low enough not to make F. Tufvesson, A. Benjebbour, and G. Wunder, “5G: A tutorial overview
any significant difference in AMC application. Therefore, the of standards, trials, challenges, deployment, and practice,” IEEE Journal
on Selected Areas in Communications, vol. 35, no. 6, pp. 1201–1221,
advantages for this second model in terms of memory make 2017.
the difference for this application. [2] K. B. Letaief, W. Chen, Y. Shi, J. Zhang, and Y. A. Zhang, “The roadmap
However, the method can be improved to reduce the effect to 6G: AI empowered wireless networks,” IEEE Communications Mag-
azine, vol. 57, no. 8, pp. 84–90, 2019.
of causality in the MCS decision. Techniques such as time [3] C. Jiang, H. Zhang, Y. Ren, Z. Han, K. Chen, and L. Hanzo, “Machine
series prediction could allow a better decision and, therefore, learning paradigms for next-generation wireless networks,” IEEE Wire-
a higher spectral efficiency. In addition, the regression model less Communications, vol. 24, no. 2, pp. 98–105, 2017.
[4] Focus group on machine learning for future networks including 5G, FG-
could be generalized to include the TB size as feature. ML5G-ARC5G Unified architecture for machine learning in 5G and fu-
ture networks, Telecommunication standardization sector of International
Telecommunication Union (ITU-T) Std., January 2019.
ACKNOWLEDGEMENTS [5] 3GPP, NR; Physical layer; General description, 3rd Generation Partner-
ship Project (3GPP) TS 38.201, Jan 2020.
This work has been supported by the Spanish Government [6] A. Goldsmith, Wireless Communications. New York: Cambridge
University Press, 2005.
and FEDER under Grant TEC2016-80090-C2-1-R (Ministerio [7] 3GPP, Study on channel model for frequencies from 0.5 to 100 GHz
de Economı́a y Competitividad) and Beca de Colaboración (Release 16), 3rd Generation Partnership Project (3GPP) TS 38.901, Jan
19CO1/008714 (Ministerio de Educación y Formación Profe- 2020.
[8] S. M. Kay, Fundamentals of statistical signal processing. Vol. I, Estima-
sional), by the Junta de Andalucia under Grants P18-RT-3175 tion theory. Prentice Hall signal processing series, 1993.
and P18-TP-3587, and by the Universidad de Málaga.