Blind Packet-Based Receiver Chain Optimization
Using Machine Learning
Mohammed Radi
∗†
, Emil Matus
∗†
, Gerhard Fettweis
∗†
∗
Vodafone Chair Mobile Communications Systems, Technische Universit
¨
at Dresden, Dresden, Germany
†
Center for Advancing Electronics Dresden cfAED, Technische Universit
¨
at Dresden, Germany
Abstract—The selection of the most appropriate equalization-
detection-decoding algorithms in wireless receivers is a challeng-
ing task due to the diversity of application requirements, algo-
rithm performance-complexity trade-offs, numerous transmission
modes, and channel properties. Typically, the fixed receiver-
chain is employed for specific application scenario that may
support iterative processing for better adaptation to variable
channel conditions. We propose a novel method for optimizing
receiver efficiency in the sense of maximizing packet transmission
reliability while minimizing receiver processing complexity. We
achieve this by packet-wise dynamic selection of the least complex
receiver that enables error-f ree packet reception out of set of
available receivers. The scheme employs convolutional neural
network (CNN) and supervised deep learning approach for
packet classification and subsequent prediction of the optimum
receiver using raw baseband signals. The proposed scheme aims
to approach a packet error rate close to the rate of the most
complex receiver architecture while using a combination of both
low and high complexity architectures. This is achieved by em-
ploying the neural network based classifier to dynamically select
packet-specific optimum architecture; i.e. instead of using the
most complex receiver for all packets, the approach dynamically
assigns the packet to the most appropriate receiver in terms of
equalization-detection-decoding capability and the least pos sible
complexity. We analyze the performance of the proposed scheme
considering various channel scenarios. The system demonstrates
excellent packet classification performance resulting in the signif-
icant performance increase and the reduction of the usage of the
functional blocks that can go up to 96% of the time in different
scenarios.
Index Terms—Receiver architecture, Neural networks, Deep
neural networks (DNN), CNN, deep learning, classification,
MIMO, OFDM, Iterative Receivers.
I. INTRODUCTION
One of the main challenges in designing and implementing
wireless receivers is finding an efficient technique to dispatch
and schedule different computational tasks. These techniques
should consider the dynamic changes when it comes to tasks
frequency, functionality, and number of resources required, or
else, it is designed for the worst-case; i.e. highest required
resources to achieve a target packet error rate (PER).
Considering the dynamic changes would affect the total
requirements of the design when it comes to the number of
processing elements (PE), latency, power, and area needed for
the wireless receiver.
For OFDM based systems, Joint equalization, detection, and
decoding by means of a maximum a posteriori probability
(MAP) leads to the best communication performance and rep-
resents a performance bound for other sub-optimal approaches.
One of the most powerful receiver techniques approaching
MAP is the iterative approach which incrementally improves
the PER performance by iterating over different receiver
sections i.e. equalization, detection, and decoding [1]. One
of the main challenges regarding iterative receivers is the
huge number of possible arrangements of iteration loops, and
also selecting the proper algorithms for detectors, equalizers,
and decoders should be used. There is an optimum s chedule
of receiver tasks for a specific combination of transmission
scenario(s) and codeword(s) under certain channel conditions
affected by inter-symbol interference (ISI), inter carrier inter-
ference (ICI), and inter antenna interference (IAI).
Typically, the fixed receiver-chain is employed for a specific
application scenario that may support iterative processing to
better adapt the communication performance to variable chan-
nel conditions. The usage of complex advanced receivers for
all traffic regardless of instantaneous channel properties results
in poor receiver efficiency as the fixed complex algorithms are
used even for packets that could be decoded correctly using
a simpler receiver. On the other side, adopting the iterative
techniques causes unpredictable processing time that poses a
challenge on task scheduling and resource allocation in multi-
processor modem architecture.The motivation of this work is
to develop a technique that allows a dynamic optimization
of the receiver chain to get over the performance-complexity
trade-off. The decision of selecting a receiver chain is based
on the instantaneous properties of the channel, receivers
architectures/algorithms, and transmitted data. In order to
cope with the aforementioned problem, we propose a method
that dynamically assigns the packets individually to the least
complex receiver-chain that is capable of correctly process
and decode the packet. We achieve this by introducing a new
scheme for predicting the most proper receiver architecture for
every IQ data packet, and so; predicting the schedule of the
required processes in the receiver.
Besides that, by knowing the schedule before processing the
incoming data, it is more efficient and faster to schedule all the
processing tasks in the receiver, giving a smoother and more
efficient performance of wireless receivers on multiprocessor
system on chip (MPSOC), as scheduling the tasks there is one
of the main challenges. We achieve this by introducing a CNN
based classifier (pre-trained offline) on top of the receiver, and
by proper training it can accurately decide the most proper
(i.e least complex) schedule of processes and tasks needed
to maintain the required performance of the system. We take
This document is a preprint of: M. Radi, E. Matus and G. Fettweis, “Blind Packet-Based Receiver Chain Optimization Using Machine Learning,” in
Proceedings of IEEE Wireless Communications and Networking Conference (WCNC 2020), Seoul, Korea, Apr 2020.
© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
f
f
Enc
Enc
Mod
Mod
M-DFT SCM
N-IDFT
h
++
N-DFT
N-DFT
SC
DM
SC
DM
M-IDFT
M-IDFT
MIMO
CHANNEL
TRANSMITTER
FDE
Det
Dec
Dec
RECEIVER
M-DFT SCM
N-IDFT
x
1
x
2
n
2
n
1
y
1
y
2
Y
1
Y
2
Mod
M-DFT SCM
Mod
M-DFT SCM
{
{
!
!
"
!
#
#
"
f
f
Enc
Enc
Mod
Mod
M-DFT SCM
N-IDFT
h
++
N-DFT
N-DFT
SC
DM
SC
DM
M-IDFT
M-IDFT
MIMO
CHANNEL
TRANSMITTER
FDE
Det
Dec
Dec
RECEIVER
M-DFT SCM
N-IDFT
x
1
x
2
n
2
n
1
y
1
y
2
Y
1
Y
2
Mod
M-DFT SCM
Mod
M-DFT SCM
{
{
!
!
"
!
#
#
"
Fig. 1: Transmitter receiver chain system model
SCM/SCDM: Subcarrier mapping/demapping, FDE: Frequency domain equalizer
advantage of using CNNs, which showed promising results in
communication system optimization and modeling. The nature
of CNN as an efficient feature extractor allows a fast and
accurate way to make a pre-processing decision based on a
previously learned combination of features and the related
proper decisions. Thanks to different enhanced techniques
based on stochastic gradient descent (SGD)[2], activation
layers or rectifiers[3], regularization techniques, it became
easier and faster to learn a higher number of features, and
even it becomes faster with using high-performance graphical
processing units (GPUs) and parallelization techniques.
The main contributions of this work include:
1) Introducing an offline pre-trained CNN based classifier
to predict the least complex optimum receiver architec-
ture ( including equalization, detection, and decoding).
2) The proposed solution gives the opportunity to dy-
namically use different receiver architectures/schedules
achieving communication performance equal or close to
the most complex one.
3) The nature of the proposed solution as a pre-processor
gives a smoother and more efficient way to schedule
tasks/processes on Multi-user wireless receivers; leading
to more efficient usage of time and power.
4) Analyzing how the prediction performance depends on
the channel properties, receiver set diversity, and how
prediction affects the receiver(s) complexity and effi-
ciency.
II. SYSTEM MODEL
A. Transmission System
In this paper, we consider an LTE/5G-NR uplink trans-
mission scenario combining single carrier frequency division
multiple access (SC-FDMA) and MIMO spatial multiplexing
with 2 transmit and 2 receive antennas as shown in Fig. 1.
Source packet of information bits i.e. transport block u is at-
tached with CRC to facilitate error detection and subsequently
segmented into a t uple of two code blocks u = (u
1
, u
2
).
The channel encoder converts u
1
and u
2
into corresponding
codewords followed by symbol modulation according to pre-
defined modulation and coding scheme (MCS). Blocks of M
consecutive modulated symbols are transformed to the fre-
quency domain by a M -point DFT and subsequently mapped
to a continuous block of subcarriers of N-point inverse-DFT
(IDFT) to create baseband SC-FDMA signals for each antenna
port. Finally, a cyclic prefix (CP) is added to each transmit
symbol resulting in a signal tuple x = (x
1
, x
2
).
The SC-FDMA symbols at two input ports of synchronised
receiver are represented by tuple y = (y
1
, y
2
) with
y
j
= [x
1
∗ h
j,1
+ x
2
∗ h
j,2
]
N
+ n
j
j = 1, 2 (1)
where ∗ and [.]
N
are convolution and CP removal opera-
tions, respectively. h
j,l
is channel impulse response from l-
th transmit to j-receive antenna and n
j
is complex valued
additive white gaussian noise vector (AWGN). We assume
perfectly synchronised receiver, hence, the receiver baseband
signal processing considers only demodulation, equalization
[4], detection [5], [6], and decoding components. In addition,
we employ turbo principle in the receiver which provides
means to approach the maximum a posteriori probability
(MAP) communications performance by exchanging soft in-
formation between the receiver components. For this purpose,
the receiver implements soft-in soft-out (SISO) fr equency-
domain equalization (FDE) which merges adaptive MMSE
filtering with parallel interference cancellation according to
[8], [9].
B. Multi-stage receiver architecture
The advantage of iterative receiver approach is in the capa-
bility of packet-based performance adaptation to the i nstanta-
This document is a preprint of: M. Radi, E. Matus and G. Fettweis, “Blind Packet-Based Receiver Chain Optimization Using Machine Learning,” in
Proceedings of IEEE Wireless Communications and Networking Conference (WCNC 2020), Seoul, Korea, Apr 2020.
© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
neous transmission and channel properties. Let k = 0, 1, ...K
denotes the number of receiver iterations, with non-iterative
case for k = 0. The packet-based adaptation is achieved by
iterating the feedback loop (or loops if available) until the
CRC check error flag e
k
(i) at decoder output of k-th iteration
indicates that the i-th data packet u(i) is decoded correctly
i.e.
e
k
(i) =
(
0, if u(i) = ˆu(i),
1, otherwise.
(2)
This can be formally represented by multi-stage receiver ar-
chitecture using the set S of receivers R
k
⊂ S, k = 0, 1, ...K,
each with a predefined constant number of iterations k as in
Fig. 2. Consequently, receiver R
k
implements unique combi-
nation of equalization, demodulation, detection and decoding
algorithms that results in the differences in the packet error
rate (PER) and corresponding average packet error probability
1−p
k
(¯γ), where p
k
(¯γ) represents packet success rate (PSR) as
a function of average signal to noise ratio (SNR) ¯γ. Note that
the receiver PSR is improving with k for specific SNR value ¯γ
∗
such that p
k−1
(¯γ
∗
) < p
k
(¯γ
∗
) for k = 1, 2, ...K as illustrated
in Fig. 3. Consequently, an i ncremental success rate gain
between stages k − 1 and k is ∆p
k
(¯γ
∗
) = p
k
(¯γ
∗
) − p
k−1
(¯γ
∗
)
whilst ∆p
0
(¯γ
∗
) = p
0
(¯γ
∗
). Consequently, the total accumu-
lated success rate at stage k is (Fig. 3)
p
k
(¯γ
∗
) =
k
X
m=0
∆p
m
(¯γ
∗
). (3)
In order to reach the target performance level p
k
(¯γ
∗
), up to
k receiver stages have to be passed during packet processing
depending on the instantaneous packet quality. This, however,
results in the variable and unpredictable packet processing time
and costs which poses a huge challenge on scheduling and
resource allocation in practical receiver implementations.
y
ˆ
u
u
ˆ
u
ˆ
u
u
y
ˆ
u
u
ˆ
u
ˆ
u
u
Fig. 2: Conventional combined multi-stage receiver
C. Optimization Problem
Let define a set of receiver costs as C = {C
k
= Cost(R
k
) :
k = 0, 1, ...K} where C
k
corresponds to costs associated with
packet processing at receiver R
k
. In this work we consider
costs C
k
as general quantity that may be associated with
algorithm complexity, power consumption or processing time.
The goal is to minimise the packet processing cost while
maximizing communication performance. This optimization
problem can be formulated as maximization of the difference
'p
0
J

k
p
J
0
R
k
R
k
R
K
R

k
p
J
'

k
p
J
'

K
p
J
'
J
J
3(5
ϭ
Fig. 3: Packet error rate (PER) for different receiver architectures
∆
k
= C
K
−C
k
between the assumed maximum costs C
K
(i.e.
C
K
> C
k
, k = 0, 1...K) which must be allocated for packet
processing when no apriori knowledge is available on actual
packet quality and actual costs C
k
of receiver R
k
. This can
be formalised for any packet y(i) by solving
k
opt
= argmax
k
(∆
k
) (4)
subject to
e
k
(i) = 0 (5)
To solve this combinatorial optimization problem, we pro-
pose a predictor based on a deep neural network.
III. PROPOSED RECEIVER ARCHITECTURE
A. Multi-branch receiver
This work addresses the receiver efficiency and predictabil-
ity problem by introducing parallel multi-branch receiver
architecture with blind packet-based branch-prediction. Typi-
cally, the receiver adaptation algorithms rely on the knowledge
of t he instantaneous SNR and channel impulse response.
However, these parameters are based on the pilot symbol
measurements and, hence, reflect only the behavioral of lin-
ear channel section observed by pilots. These methods do
not consider end-to-end data packet path and, particularly,
non-linear effects associated with channel encoding/decoding,
bit/symbol-mapping and receiver algorithms as (iterative) de-
tection, decoding etc. To cope with this problem, we propose a
machine learning approach using CNN-based packet classifier
for branch-prediction in the multi-branch receiver.
The multi-branch receiver approach employs a parallel
branches with receivers R
k
, k = 0, 1, ...K, as illustrated in
Fig. 4. The branch-predictor solves the problem (4) and (5)
i.e. it analyses the instantaneous packet quality at time i
using raw baseband signals y(i) and, base on this, estimates
the least complexity branch k capable of decoding packet
without errors. In other words, each packet is routed only
to a selected branch k depending on the predicted packet
quality and receiver capability. The branch-prediction allows to
determine the exact packet-specific processing costs before the
actual receiver algorithms are processed. Consequently, system
This document is a preprint of: M. Radi, E. Matus and G. Fettweis, “Blind Packet-Based Receiver Chain Optimization Using Machine Learning,” in
Proceedings of IEEE Wireless Communications and Networking Conference (WCNC 2020), Seoul, Korea, Apr 2020.
© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
performance and utilization can be improved by packet-based
resource management.
We define the predictor conditional probability p(k|k
opt
)
expressing the probability of selecting branch k for packet pro-
cessing instead of using optimum branch k
opt
. The elements
p(k|k
opt
) forms a confusion matrix for k, k
opt
= 0, 1, 2, ...K.
The diagonal probabilities for k = k
opt
are associated with
correctly routed packets while misprediction probabilities are
represented by the off-diagonal matrix elements for k 6= k
opt
.
Clearly
P
K
m=1
p(k|k
opt
) = 1 for any k = 1, 2, ...K. The
confusion matrix gives an indication on the predictor’s packet
classification accuracy. An ideal predictor exhibits p(k|k
opt
) =
0 for all k 6= k
opt
which is impossible to achieve due to many
influencing factors in transmitter-receiver chain.
Predictor
...
...
y
ˆ
u
y
k
Predictor
...
...
y
ˆ
u
y
k
Fig. 4: Proposed combined multi-branch receiver with CNN-based packet
quality predictor
B. Two-Branch Predictor
One of the motivations of this work is to understand, if
the blind packet classification and optimum receiver selection
is possible and how the classification accuracy depends on
the channel properties and the difference of receivers PER
performance. In order to reflect these issues, we consider
the two-branch receiver architecture comprising basic receiver
(referred as class 1) and advanced receiver (class 2). We
consider two options for class 1 from the receiver set S i.e.
R
0
referred as MMSE without iteration (k = 0) and R
1
based
on k = 1 iteration. Receiver class 2 is the most advanced and
complex receiver R
8
with k = 8 iterations. The associated
predictor selects between 2 different receiver branches. The
confusion matrix of such predictor is shown in Table I. Here
the type I error with probability p(1|2) increases the error
rate of the system as the predictor would send packets that
need the more complex receiver architecture to be decoded to
the simple one. Type II error with probability p(2|1) would
not contribute in the error as it means that the predictor would
send packets that can be decoded using the simple architecture
receiver to the more complex one, but it will increase the
number of processes or iterations needed for the decoding.
C. Dataset generation
We construct the data sets by generating random data-bit
streams then passing it by the transmitter chain (explained in
next section) and different channel models with different SNRs
(8dB-30dB). Every 128 IQ samples are recorded as one packet,
then passed by the selected simple receiver architecture, if
packet is decoded with no errors, it is labeled as class 1
packet, or else it labeled as class 2 packet. First class (class 1)
contains the IQ data packets that can be processed (error free)
with simple receiver architecture (here as MMSE based or 1
full iteration receiver), while class 2 is the class of the packets
that need more complex receiver (8 iterations in this work) to
be processed.
One of the important notes regarding balancing the size
of both datasets is that with low SNRs; the number of the
resulting packets of both classes are balanced, but by increas-
ing the SNR level; the number of class 2 packets decreases
dramatically as the simple receiver architecture can decode
more packets easily. Training with these unbalanced dataset
sizes leads to huge bias towards class 1, especially in high
SNRs. By balancing them (50% each) the performance is
better, but still the network actually gives a little high type
II error. We found that a percentage of 2:1 of dataset size
gives better performance in general, as it keeps the fairness
of having more expected class 1 packets, without biasing the
network towards it. In the next sections, we consider p
s
(1)
and p
s
(2) are the probabilities of having packets of cla s s 1
and cl as s 2 respectively when running the system for the
corresponding SNR level and the receiver architecture. Fig. 5
shows the percentage of class 1 data size (with 0 or 1 iteration)
for different SNRs.
0 1 2 3 4 5 6 7 8
No. iterations
30
40
50
60
70
80
90
100
Frame success rate%
SNR=8dB
SNR=10dB
SNR=12dB
SNR=14dB
SNR=16dB
SNR=18dB
SNR=20dB
SNR=22dB
SNR=24dB
SNR=26dB
SNR=28dB
SNR=30dB
Fig. 5: Packet success rate for different SNR levels and different number of
iterations (EVA channel)
The generated datasets are used for training, validation, and
testing. In this work, every training dataset is 600000 labeled
IQ packets. It is divided (80% : 20%) For training:validation.
Separately generated datasets (with same scenario, different
random data bit streams and channel realizations of same
PDPs) of size 120000 packets are used for testing the t rained
networks.
This document is a preprint of: M. Radi, E. Matus and G. Fettweis, “Blind Packet-Based Receiver Chain Optimization Using Machine Learning,” in
Proceedings of IEEE Wireless Communications and Networking Conference (WCNC 2020), Seoul, Korea, Apr 2020.
© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
Table II shows the different datasets generated for the
simulations i n this work. Every dataset is generated for SNRs
from 8-30 dB , and the CNN is trained for all SNRs together.
The first 4 datasets consider 2 different nonlinear receiver
structures as classes, the simple less complex one is a receiver
with one full iteration and the complex one i s 8 full iterations,
and they are constructed using different channel models. The
datasets from 5 to 8 consider the less complex one as a
linear MMSE equalizer.The last dataset is to show that the
concept works also for mix of channel models (with different
performance of course), so it is constructed for packets come
from both EVA and EPA channel models.
TABLE I: confusion matrix
Actual class
Class 1 Class 2
Predicted class
Class 1
p(1|1)
True
positive
p(1|2)
False positive
Type I error
Class 2
p(2|1)
False negative
Type II error
p(2|2)
True
Negative
TABLE II: Datasets
Dataset Channel model(s) Rx classes
1 AWGN 1 iteration vs 8 iterations
2 EPA 1 iteration vs 8 iterations
3 EVA 1 iteration vs 8 iterations
4 ETU 1 iteration vs 8 iterations
5 AWGN MMSE vs 8 iterations
6 EPA MMSE vs 8 iterations
7 EVA MMSE vs 8 iterations
8 ETU MMSE vs 8 iterations
9 EVA and EPA 1 iteration vs 8 iterations
10 EVA and EPA MMSE vs 8 iterations
Fig. 1 shows the used transmitter-receiver chain model we
use for generating the datasets for training and testing. As
mentioned in dataset generation section, we consider 2 cases.
First we consider the basic receiver architecture Rx(1) as
the non iterative li near MMSE based receiver (wi ll be called
MMSE), and the advanced Rx(2) is 8 full iterations receiver
(will be called I8). The second we consider the basic receiver
architecture Rx(1) as one iteration receiver (will be called
I1), and the advanced Rx(1) is still I8. Table III shows the
parameters we consider during this work.
D. CNN architecture
In this paper, we propose 7-layer neural network architecture
as in Table IV. The input of the network during the training
or the test is simply the raw IQ samples arriving to the
receiver with no prepossessing at all. The network gets 128
complex samples (so input layer dimension is 2x128 for both
TABLE III: System parameters
Parameter values
Transmission mode Uplink SC-FDMA
MIMO 2x2
Bandwidth 1.4 MHz
Modulation QPSK
FFT/IFFT size 128
Number of occupied sub-carriers 72
Packet length 512 bits
code rate 3/4
Sampling frequency 1.92 MHz
Channel Quality Indicator (CQI) 6
Cyclic Prefix (CP) Normal 4.7µs
TABLE IV: CNN Parameters
Layer Activations learnables
Input 2 ×128 × 1 -
Convolution Layer 2 ×128 × 3 12 Weights, 3 Bias
RELU 2 ×128 × 3 -
Maxpool 2 ×64 × 3 -
Fully connected 1 ×1 × 2 384 Weights, 2 Bias
Softmax 1 ×1 × 2 -
classification output - -
real and imaginary parts of the samples). The table shows
the dimension of every layer with the number of learnable
parameters.The classification output here is deciding between
2 classes (class 1 or class) as mentioned earlier.
E. Simulation environment
The simulation was done using Matlab 2019a to take ad-
vantage of the easy and the efficient multi GPU parallelization
provided by the deep learning toolbox. Table VI shows the
learning process parameters for the incoming results.
In this paper, we consider the extended pedestrian A (EPA),
the extended vehicular A (EVA) model and extended typical
urban (ETU) model (see Table V) as possible power delay pro-
files (PDP) of the propagation channel. The signal bandwidth
is set to Bs = N ·15kHz, N = 128, where N is the FFT/IFFT
size. We take into consideration the effects of ISI, ICI, IAI
plus Doppler effect in all of these models. For the sake of
comparison and getting to know what features the performance
of the CNN predictor depends on, we also present the results
from adaptive white Gaussian noise (AWGN) channel.
TABLE V: Channel models used
Parameter EPA EVA ETU
No. of taps 7 9 9
Max Doppler Freq 5 70 70
IV. CLASSIFICATION PERFORMANCE EVALUATION AND
ANALYSIS
A. Overall classification performance
In this part we investigate the effect of the expected channel
models on the classification performance and the total system
performance (PER or PDE). And to investigate the effect of
This document is a preprint of: M. Radi, E. Matus and G. Fettweis, “Blind Packet-Based Receiver Chain Optimization Using Machine Learning,” in
Proceedings of IEEE Wireless Communications and Networking Conference (WCNC 2020), Seoul, Korea, Apr 2020.
© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
TABLE VI: Learning Parameters
Parameter Value
Optimizer SGD with momentum (SGDM)
Trainig datasets size 600000 packets
Training : Validation 80%:20%
Testing datasets size 120000 packets
Training datasets classes ratio 2:1
Initial learning rate 0.01
Number of epochs 6 - 500
Shuffling Every epoch
Validation frequency 1000 iterations
Hardware 2 x Tesla P100-PCIE-16GB
8 10 12 14 16 18 20 22 24 26 28 30
93
94
95
96
97
98
99
Percentage (%)
1 : Combined I1+I8(AWGN)
2 : Combined I1+I8 (EPA)
3 : Combined I1+I8 (EVA)
4 : Combined I1+I8 (ETU)
5 : Combined MMSE+I8 (AWGN)
6 : Combined MMSE+I8 (EPA)
7 : Combined MMSE+I8 (EVA)
8 : Combined MMSE+I8 (ETU)
9 : Combined I1+I8 (EVA&EPA mix)
10: Combined MMSE+I8 (EVA&EPA mix)
Fig. 6: Classification accuracy of different data sets
the different receiver architectures as well. We show the results
of 7 different experiments, each has separate training and
testing datasets as mentioned earlier. In experiments from 1-4,
we are testing the approach with 4 different PDPs (AWGN,
EPA,EVA, and ETU) using I1 and I8 receivers. Then we show
the difference between the performance when using a little
simpler receiver architecture linear MMSE instead of I1 with
I8 in experiments 5-8.Then in experiments 9 and 10, we try a
relatively harder datasets with a mix of EVA and EPA channel
models.
In Fig. 6, the different classification accuracy values with
different datasets are shown. In general the proposed CNN
is able to learn the features to decide if the IQ packet can
be decoded by the simple receiver or not, with accuracy that
reaches 99.86% for high SNRs. Looking at curves from 1 to
4, we notice that the classification accuracy increases with
the SNR level, as it is easier for the network to learn the
features of the IQ samples in less noisy medium.This happens
in every channel model except with EPA, we can notice that
the performance increase with SNR level to some certain level
(22 dB here), then it starts to drop relatively after that level.
But this drop in accuracy is due to Type II error only, while
type I drops to zero as we will show in Fig. 7; i.e. the PER
does not increase but actually decreases. The approach gives
10 15 20 25 30
0.5
1
1.5
2
2.5
3
3.5
Percentage (%)
1 : Combined I1+I8 (AWGN)
2 : Combined I1+I8 (EPA)
3 : Combined I1+I8 (EVA)
4 : Combined I1+I8 (ETU)
5 : Combined MMSE+I8 (AWGN)
6 : Combined MMSE+I8 (EPA)
7 : Combined MMSE+I8 (EVA)
8 : Combined MMSE+I8 (ETU)
9 : Combined I1+I8 (EVA&EPA mix)
10: Combined MMSE+I8 (EVA&EPA mix)
Fig. 7: Type I error of different data sets p(1|2)
a relatively better performance with multipath channels over
AWGN channel (0.5% - 3 % higher), showing that it is easier
to learn the features related to fading over the features related
to noise level.
With a dataset that mixed IQ samples from 2 different
channel models (EVA and EPA) , we can see the performance
is similar to EPA with a little higher accuracy than EVA, and
a little lower than EPA. Then for the curves from 5 to 8 with a
simpler receiver architecture; i.e. MMSE instead of I1, we can
see the performance is relatively better (up to 1.5%), showing
that it is easier to differentiate between a linear and non-linear
receiver systems over a system with non-linear receivers only.
This difference in performance also shows that the CNN based
predictor is learning features of the receivers and the data, not
the data only.
As the accuracy here is affected by both types of errors I
and II , and as the system PER is affected only by Type I
error, it is important to investigate separately the type I error
as shown in Fig. 7. We show that the error in all cases drops
by increasing the SNR level as well, reaching 0 for EPA for
example.
B. PER/PDE analysis
Now by introducing the predictor to the system as we
showed in Fig. 4, we can show the advantage of the system
by measuring the PER or the PDE of the system. The s ystem
gives a performance close to I8 receiver performance while
using a combination of both I1/I8 or MMSE/I8 receivers.
The performance is only 3.33% less than the I8 in the worst
case (combined I1+I8 in AWGN channel at 8dB SNR level).
The performance is getting closer to I8 with higher SNRs till
reaching the I8 performance in some cases (combined I1+I8
in EPA channel after 22dB SNR level). Fig. 8, 9, and 10 show
the comparison of PER or PDE for EPA channel (combined
I1+I8), EVA channel (combined I1+I8), and EVA channel
This document is a preprint of: M. Radi, E. Matus and G. Fettweis, “Blind Packet-Based Receiver Chain Optimization Using Machine Learning,” in
Proceedings of IEEE Wireless Communications and Networking Conference (WCNC 2020), Seoul, Korea, Apr 2020.
© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
10 15 20 25 30
10
-3
10
-2
10
-1
10
0
PER
Single I1
Single I8
Combined I1+I8
14.5 15 15.5 16
0.06
0.08
0.1
25 26
4
5
6
10
-3
Fig. 8: Performance of combined Rx vs Rx with 1 and 8 iterations (EPA)
10 15 20 25 30
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
PER
Single I1
Single I8
Combined I1+I8
Extra PER over 0 due to missclassification
14 14.5 15 15.5 16
0.06
0.07
0.08
0.09
0.1
0.11
23
2
4
6
10
-3
Fig. 9: Performance of combined Rx vs Rx with 1 and 8 iterations (EVA)
(combined MMSE+I8) respectively. As in EVA channel s ce-
nario the I8 receiver can reach zero error in high SNRs, the
residual error comes from the Type I error of mis-classification
packets of class 2 to class 1, which is shown in green in Fig.
9, and 10.
Fig. 11 shows the advantage of using the combined I1 and
I8 receivers based on our approach, for example at 22 dB
EVA (combined I1+I8), more than 97% of the packets can be
decoded successfully using the I1 receiver while giving the
same performance of using the I8 receiver all the time.
C. Comparison with iSNR classifier
To be sure that our approach does not just depend on simple
factor as the instantaneous SNR (iSNR) we will compare
our approach with a theoretical classifier that depends on the
10 15 20 25 30
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
PER
14 15 16
0.06
0.08
0.1
22 23 24
2
4
6
10
-3
Single MMSE
Single I8
Combined MMSE+I8
Extra PER over 0 due to missclassification
Fig. 10: Performance of combined Rx vs Rx with MMSE and 8 iterations
(EVA)
8 10 12 14 16 18 20 22 24 26 28 30
45
50
55
60
65
70
75
80
85
90
95
100
Percentage (%)
1 : Combined I1+I8(AWGN)
2 : Combined I1+I8(EPA)
3 : Combined I1+I8(EVA)
4 : Combined I1+I8(ETU)
5 : Combined MMSE+I8(AWGN)
6 : Combined MMSE+I8(EPA)
7 : Combined MMSE+I8(EVA)
8 : Combined MMSE+I8(ETU)
9 : Combined I1+I8(EVA&EPA mix)
10: Combined MMSE+I8(EVA&EPA mix)
10 11 12
74
76
78
80
82
27 28 29 30
97.5
98
98.5
99
99.5
100
Fig. 11: The saving of the unnecessary usage of I8 receiver when using
combined I1+I8 based on the proposed predictor
full knowledge of the iSNR value for every packet (which
is not possible practically). This will be done by saving
the corresponding iSNR value for every packet i , which is
calculated by:
γ(i) =
P
s
(i)
P
n
(i)
|h(i)|
2
, (6)
where P
s
is the instantaneous transmitted signal power, P
n
is the instantaneous added noise power, and h is the corre-
sponding instantaneous channel r esponse. For every group of
packets with average SNR , there is a PER introduced by
the basic receiver, and hence there is a real knowledge of the
percentage of successful packets. We analyze the distribution
of the iSNR values for every average SNR and apply the same
percentage of packet success rate to decide the threshold iSNR
This document is a preprint of: M. Radi, E. Matus and G. Fettweis, “Blind Packet-Based Receiver Chain Optimization Using Machine Learning,” in
Proceedings of IEEE Wireless Communications and Networking Conference (WCNC 2020), Seoul, Korea, Apr 2020.
© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
to decide if the packet will be successfully processed or not
as shown in Fig. 12.
!
!
"#
$%&'' ( )&*+&,$-* ./0
$%&'' 1 )2&'3$ ./0
456
!
7
8
!
1 9 7
8
!
Fig. 12: Setting the iSNR threshold value using the percentage of successful
packets
Taking the case of EPA channel model, the performance
of the iSNR classifier compared to the propose approach is
shown in Fig. 13. With low average SNR values, the CNN
performance is only slightly better than iSNR classifier, means
that the noise is the dominant factor here, and the CNN
basically works in similar way as an iSNR detector. By looking
at higher average SNR values, we can see the big gap in
performance showing that the proposed approach is working
so much better; means that the approach is depending more on
learning features of fading channels, and learns the receiver
capabilities to come over the corresponding features of channel
or noise.
10 15 20 25 30
10
-3
10
-2
10
-1
10
0
PER
Single I1
Single I8
Combined I1+I8
iSNR classifier
8 8.5 9 9.5 10
0.2
0.25
0.3
0.35
Fig. 13: The performance of the proposed solution compared to iSNR based
classifier
V. CONCLUSION AND FUTURE WORK
The proposed approach shows a promising technique of
using CNNs for receiver architecture prediction and selection
based on packet quality prediction. The approach gives the
advantage of using combination of low complexity receiver
architectures and higher complexity ones while giving the
same performance of the high complexity receivers. The
nature of the system as a pre-processor or predictor gives
the advantage of better scheduling for receiver processes and
results in lower latency and power consumption. The approach
is robust against changing the channel types, using mixes of
them, or different types of receiver architectures with slight
differences.
There is a wide range of more inspections and experiments
to see the effects of different datasets with more complex ones
(more receiver architectures, or more mixed data sets), and
changing the CNN architecture of fit these new complexities.
Also the approach will be testing more suggested techniques
and their effects in future. These techniques will include
training the CNNs with high SNRs samples of the datasets
first to increase accuracy before using the lower SNR samples
as suggested in [12] and [13].
Other techniques that will be inspected in future work will
include using complex valued neural networks, and forward
thinking CNNs to train layer by layer for certain f eature(s).
REFERENCES
[1] S. Ahmed, T. Ratnarajah, M. Sellathurai, and C. Cowan, “Iterative
Receivers for MIMO-OFDM and Their Convergence Behavior,” IEEE
Trans. Veh. Technol., vol. 58, no. 1, pp. 461–468, Jan. 2009.
[2] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
ArXiv preprint arXiv:1412.6980, 2014.
[3] X. Glorot, A. Bordes, and Y. Bengio, Deep Sparse Rectifier Neural
Networks. Proceedings of the Fourteenth International Conference on
Artificial Intelligence and Statistics, in PMLR 15:315-323, 2011.
[4] T. Seifert and G. Fettweis, “Soft-PIC Frequency-Domain Equalization
in Iterative MIMO Receivers for the LTE-A Uplink,” in Proc. IEEE 79th
Veh. Technol. Conf., May 2014, pp. 1–5.
[5] E. P. Adeva, T. Seifert, and G. Fettweis, “VLSI Architecture for MIMO
Soft-Input Soft-Output Sphere Detection,” J. Signal Process. Syst., vol.
70, no. 2, pp. 125–143, 2013.
[6] J. Ketonen, J. Karjalainen, M. Juntti, and T. Hanninen, “MIMO De-
tection in Single Carrier Systems,” in Proc. Eur. Signal Process. Conf.,
2011, pp. 654–658.
[7] M. Tuchler and A. Singer, “Turbo Equalization: An Overview,” IEEE
Trans. Inf. Theory, vol. 57, no. 2, pp. 920–952, Feb. 2011.
[8] N. Benvenuto, R. Dinis, D. Falconer, and S. Tomasin, “Single Carrier
Modulation With Nonlinear Frequency Domain Equalization: An Idea
Whose Time Has Come Again,” Proc. IEEE, vol. 98, no. 1, pp . 69–96,
2010.
[9] T. Seifert and G. Fettweis, “Soft-PIC Frequency-Domain Equalization
in Iterative MIMO Receivers for the LTE-A Uplink,” in Proc. IEEE 79th
Veh. Technol. Conf., May 2014, pp. 1–5.
[10] H. Robbins and S. Monro, “A stochastic app roximation method,” Annals
of Mathematical Statistics, vol. 22, pp. 400–407, 1951.
[11] Duchi, J., Hazan, E., Singer, Y. (2011). Adaptive Subgradient Methods
for Online Learning and Stochastic Optimization. Journal of Machine
Learning Research, 12, 2121–2159.
[12] D. George and E. A. Huerta, “Deep neural networks to enable real-time
multimessenger astrophysics,” arXiv preprint arXiv:1701.00008, 2016
[13] T. J. O’Shea, T. Roy, T. C. Clancy, ”Over-the-Air Deep Learning Based
Radio Signal Classification”, IEEE Journal of Selected Topics in Signal
Processing, vol. 12, no. 1, pp. 168-179, Feb. 2018.
This document is a preprint of: M. Radi, E. Matus and G. Fettweis, “Blind Packet-Based Receiver Chain Optimization Using Machine Learning,” in
Proceedings of IEEE Wireless Communications and Networking Conference (WCNC 2020), Seoul, Korea, Apr 2020.
© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.