Received 23 May 2023; revised 13 September 2023; accepted 4 November 2023.

Date of publication 9 November 2023; date of current version 18 December 2023.

The associate editor coordinating the review of this article and approving it for publication was B. Di.
Digital Object Identifier 10.1109/TMLCN.2023.3332021

RNN Based Channel Estimation in Doubly

Selective Environments
ABDUL KARIM GIZZINI 1 (Member, IEEE) AND MARWA CHAFII 2,3 (Senior Member, IEEE)
1 ETIS, UMR8051, ENSEA, CNRS, CY Cergy Paris Université, 95000 Cergy, France
2 Engineering Division, New York University (NYU) Abu Dhabi, Abu Dhabi, United Arab Emirates
3 NYU WIRELESS, NYU Tandon School of Engineering, Brooklyn, NY 11201 USA

CORRESPONDING AUTHOR: A. K. GIZZINI ([email protected])

ABSTRACT Doubly-selective channel estimation represents a key element in ensuring communication

reliability in wireless systems. Due to the impact of multi-path propagation and Doppler interference in
dynamic environments, doubly-selective channel estimation becomes challenging. Conventional symbol-by-
symbol (SBS) and frame-by-frame (FBF) channel estimation schemes encounter performance degradation
in high mobility scenarios due to the usage of limited training pilots. Recently, deep learning (DL) has been
utilized for doubly-selective channel estimation, where long short-term memory (LSTM) and convolutional
neural network (CNN) networks are employed in the SBS and FBF, respectively. However, their usage is
not optimal, since LSTM suffers from long-term memory problem, whereas, CNN-based estimators require
high complexity. For this purpose, we overcome these issues by proposing an optimized recurrent neural
network (RNN)-based channel estimation schemes, where gated recurrent unit (GRU) and Bi-GRU units are
used in SBS and FBF channel estimation, respectively. The proposed estimators are based on the average
correlation of the channel in different mobility scenarios, where several performance-complexity trade-offs
are provided. Moreover, the performance of several RNN networks is analyzed. The performance superiority
of the proposed estimators against the recently proposed DL-based SBS and FBF estimators is demonstrated
for different scenarios while recording a significant reduction in the computational complexity and the GPU
execution time.
INDEX TERMS Wireless communications, channel estimation, deep learning, RNN, LSTM, GRU, Bi-
GRU.

I. INTRODUCTION channel estimation is crucial, especially in high mobility

T HE recent advances in beyond 5G networks enable

high data rates and low latency mobile wireless
applications [1], [2]. Wireless communications offer mobility
scenarios [4].
In general, a few pilots are allocated within the transmitted
frame in order to maintain a good transmission data
to different nodes within the network, however, the mobility rate, where the state-of-the-art (SoA) channel estimation
feature has a severe negative impact on the communication schemes can be categorized into: (i) SBS estimators: the
reliability [3]. In such environment, the wireless channel channel is estimated for each received symbol separately [5],
is said to be doubly-selective, i.e. varies in both time and [6]. (ii) FBF estimators: where the previous, current, and
frequency. This is due to the propagation medium, where future pilots are employed in the channel estimation for
the transmitted signals propagate through multiple paths, each received symbol [7]. The higher channel estimation
each having a different power, delay, and Doppler shift accuracy can be achieved by using FBF estimators, since
effect resulting from the motion of network nodes. Knowing the channel estimation of each symbol takes advantage from
that the accuracy of the estimated channel influences the the knowledge of previous, current, and future allocated
system performance since it affects different operations at pilots within the frame. Unlike, SBS estimators, where only
the receiver like equalization, demodulation, and decoding. the previous and current pilots are exploited in the channel
Therefore, ensuring communication reliability using accurate estimation for each received symbol. However, the allocated

2023 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 2, 2024 1
pilots are insufficient for accurately tracking the doubly- estimation in doubly-selective environments for both SBS
selective channel. As a result, conventional SBS channel and FBF channel estimation, where an optimized RNN
estimation schemes use the demapped data subcarriers networks represented by a GRU and bi-directional (Bi)-
besides pilot subcarriers to accomplish the channel estimation GRU units are used in the proposed SBS and FBF channel
task. This procedure is known as data-pilot aided (DPA) estimators, respectively. Thus, having a low-complexity and
channel estimation, which is unreliable due to the demapping robust channel estimation in different mobility scenarios.
errors of the data subcarriers that are also enlarged from The proposed GRU-based SBS estimator uses only one
one symbol to another, leading to accumulated error in the GRU network instead of two as the case in the recently
channel estimation process. Moreover, the DPA-based chan- proposed LSTM-based estimator [22]. After that, DPA
nel estimation schemes such as spectral temporal averaging estimation is applied using the GRU estimated channel.
(STA) [5] and time-domain reliable test frequency domain Finally, unlike [22] where FNN network is used for noise
interpolation (TRFI) [6] are impractical solutions as they rely elimination, in the proposed GRU-based estimators, time
on many assumptions such as high correlation of the channel averaging (TA) processing is employed as a noise alleviation
within the received frame. In addition, they lack robustness technique where the noise alleviation ratio is calculated
in highly dynamic environments. On the other hand, several analytically. Moreover, motivated by the fact that Bi-RNN
2D interpolation methods, such as radial basis function is designed to perform 2D interpolation of unknown data
(RBF) [8] and average decision-directed with time truncation bounded between known data [25], the proposed Bi-GRU
(ADD-TT) [9] are employed in the FBF channel estimation. channel estimator is designed to overcome the limitations of
However, the performance of these interpolation methods the FBF CNN-based channel estimation schemes, where an
is limited when employed in high mobility scenarios, since end-to-end 2D interpolation is performed by the proposed
they use fixed interpolation parameters. Moreover, the well- Bi-GRU unit. In this context, the proposed Bi-GRU channel
known FBF estimator is the conventional 2D linear minimum estimator employs an adaptive frame design, where comb
mean square error (LMMSE) uses the channel and noise pilot allocation is replaced by full pilot allocated symbols
statistics in the estimation, thus, leading to comparable that are inserted periodically within the transmitted frame.
performance to the ideal case. However, it suffers from As a first step, the channel is estimated at the inserted
high complexity making it impractical in real-case scenarios. pilot symbols, after that, Bi-GRU acts as an end-to-end 2D
Therefore, investigating both SBS and FBF channel esti- interpolation unit to estimate the channel at the data symbols
mators with a good trade-off complexity vs. performance without the need to any initial estimation. By doing this
is a crucial need for improving the channel estimation interpolation, the proposed Bi-GRU based estimator is able to
accuracy as well as maintaining affordable computational further improve the estimation performance, unlike the CNN-
complexity. based estimators that work according to the noise mitigation
Recently, a great success of deep learning (DL) has principle [26], [27] rather than doing actual interpolation.
been witnessed in several wireless communications applica- Simulation results show the performance superiority of the
tions [10], [11], including localization [12], [13], [14], and proposed RNN-based channel estimation schemes against the
channel estimation [15], [16], [17], [18], particularly when SoA SBS and FBF channel estimators while recording an
integrated with conventional SBS and FBF estimators. This outstanding computational complexity reduction represented
success is due to the robustness, low-complexity, and good by the floating-point operations per second (FLOPS) and the
generalization ability of DL algorithms making their integra- graphics processing unit (GPU) execution time. To sum up,
tion into communication systems beneficial. Motivated by the contributions of this paper are listed below1 :
these advantages, DL algorithms have been integrated into • Proposing low-complexity and robust RNN-based chan-
doubly-selective channel estimators in two different manners: nel estimation schemes, where an optimized GRU, and
(i) feed-forward neural network (FNN) and LSTM networks Bi-GRU units are employed to accurately estimate the
with different architectures and configurations are employed doubly-selective channel in SBS and FBF fashions,
on top of SBS estimators [19], [20], [21], [22], [23]. (ii) CNNs respectively.
are integrated into the FBF estimators [8], [9], [24], where • Employing GRU unit as a pre-processing module to
the estimated channel for the whole frame is considered as a DPA and TA processing in SBS channel estimation.
2D low-resolution noisy image and CNN-based processing is Whereas an end-to-end 2D interpolation using Bi-GRU
applied as super-resolution and denoising techniques. These unit is proposed for FBF channel estimation.
SoA DL-based SBS and FBF still encounter a considerable • Analyzing the appropriate RNN architectures to be
performance degradation due to the poor accuracy of the employed according to the average channel correlation
employed initial channel estimation as in [20] and [21]. within the frame in different mobility scenarios, where
Moreover, they require high computational complexity due
to the employed DL architectures [8], [9], [24].
In order to achieve better performance-complexity trade- 1 We would like to mention that part of this work related to the Bi-RNN
off in different mobility scenarios according to the channel based FBF channel estimation has been accepted for publication in the IEEE
correlation, this paper sheds light on the RNN-based channel ICC 2023 conference [28].

2 VOLUME 2, 2024
Gizzini and Chafii: RNN-Based Channel Estimation in Doubly Selective Environments

the advantages of using the proposed optimized GRU set of subcarriers denoted Kd and Kp , respectively. The
unit instead of regular LSTM unit are discussed. received frequency-domain OFDM subcarrier denoted as
• Showing that the proposed RNN-based channel estima- ỹi [k] is expressed as follows
tors record a significant superiority over the SoA SBS
and FBF channel estimators in terms of bit error rate
(BER) and throughput for different modulation orders, ỹi [k] = h̃i [k]x̃i [k] + ṽi [k], k ∈ Kon . (2)
mobility scenarios, and frame lengths.
• Illustrating the advantage of using the ensemble learning Here, h̃i ∈ CKon ×1 refers to the frequency response of the
(EL) algorithm [29] in the generalization of one DL doubly-selective channel at the i-th OFDM symbol. ṽi ∈
model that is robust against a range of Doppler CKon ×1 signifies the additive white Gaussian noise (AWGN)
frequencies. of variance σ 2 . As a matrix form, (2) can be expressed as
• Providing a detailed computational complexity analysis follows
for the studied channel estimators, where we show that
the proposed RNN-based channel estimators achieve Ỹ [k, i] = H̃[k, i]X̃[k, i] + Ṽ [k, i], k ∈ Kon , (3)
substantial reduction in the required number of FLOPS,
and the GPU execution time in comparison with the SoA where Ỹ ∈ CKon ×I and X̃ ∈ CKon ×I denotes the received
SBS and FBF channel estimators. and transmitted frames in frequency domain. Moreover, Ṽ ∈
The remainder of this paper is organized as follows: CKon ×I and H̃ ∈ CKon ×I denote the AWGN noise and the
Section II presents the system model. The SoA DL-based doubly-selective frequency response of the channel for all
channel estimation schemes are thoroughly investigated symbols within the transmitted OFDM frame, respectively.
and discussed in Section III. Section IV illustrates the
framework of the proposed RNN-based channel estimation
III. SoA DL-BASED CHANNEL ESTIMATION
schemes, besides providing a brief overview of the main
This section presents the recently proposed SoA DL-based
RNN networks integrated into the doubly-selective channel
SBS and FBF channel estimation schemes, where the
estimation. In Section V, different modulation orders are used
processing steps applied in each estimator are presented.
to present simulation results, wherein the performance of the
studied estimators is examined in terms of BER. Detailed
A. DL-BASED SBS CHANNEL ESTIMATION SCHEMES
computational complexity analysis is provided in Section VI.
Finally, Section VII concludes this study. In general, FNN and LSTM networks are employed in
Notations: Throughout the paper, vectors are defined with the SBS channel estimation, where optimized FNNs are
lowercase bold symbols x. Time and frequency domain integrated as a post-processing unit with conventional SBS
vectors are represented by x and x̃ respectively. Moreover, channel estimators as the case in the DPA-FNN [19],
the (i, k) element of x̃ is represented by x̃i [k], where i and STA-FNN [20], and TRFI-FNN [21]. On the other hand,
k denote the time and the subcarrier indices. We note that LSTM networks are utilized as a pre-processing unit in the
x̃i ∈ CK ×1 , x̃i,d ∈ CKd ×1 , and x̃i,p ∈ CKp ×1 refer to the full LSTM-FNN-DPA [22], and LSTM-DPA-TA [23] channel
OFDM symbol, data subcarriers, and the pilots, respectively, estimators. Both implementations are helpful in improving
where Kd , and Kp are the number of data and pilot subcarriers. the accuracy of the channel estimation. However, the LSTM-
Matrices are written as uppercase bold symbols X. Finally, based estimation illustrates a considerable superiority over
E [.] denotes the expectation operator and the conjugate the FNN-based estimation. In this context, and since we
matrix of X is signified by X ∗ . are focusing on RNN-based channel estimation, this section
presents the steps applied in the LSTM-based channel
II. SYSTEM MODEL estimators.
Consider a frame consisting of I orthogonal frequency We would like to mention that the conventional DPA
division multiplexing (OFDM) symbols. The i-th transmitted channel estimation aims at tracking the channel variation
frequency-domain OFDM subcarrier x̃i [k], is denoted by within the received OFDM frame by applying a three-step
operation. First of all, the current received OFDM symbol is
 x̃i,d [k], k ∈ Kd .

equalized by the previously estimated channel, resulting in
x̃i [k] = x̃i,p [k], k ∈ Kp . (1) an initial channel estimation. Next, a demapping operation is
0, k ∈ Kn .

applied to the initial estimated channel, such that
where k refers to the subcarrier index, where 0 ≤ k ≤ ỹi [k]
ˆ
K − 1. Moreover, d and p indices refer to the transmitted d̃ i [k] = D , h̃DPA0 [k] = h̃ˆ LS [k]. (4)
data and pilot subcarriers, respectively. The total number of h̃ˆ DPAi−1 [k]
subcarriers is divided into Kon = Kd + Kp subcarriers in
addition to Kn null guard band subcarriers, where x̃i,d ∈ D(.) refers to the demapping operation to the nearest con-
CKd ×1 and x̃i,p ∈ CKp ×1 represent the modulated data stellation point in accordance with the employed modulation
symbols and the predefined pilot symbols allocated at a order. h̃ˆ LS signifies the LS estimated channel at the received

VOLUME 2, 2024 3
 FIGURE 1. Block diagram of the SoA LSTM-based SBS estimators.

preambles, such that x̃i,p ∈ CKp ×1 refers to the predefined pilot subcarriers. The
P LS estimated channels in (7) are fed as an input to both LSTM
ỹ(p)
P
u [k] and FNN networks, where the LSTM-FNN estimated channel
h̃ˆ LS [k] =
u=1 is expressed as follows
, k ∈ Kon , (5)
P3̃[k]
ỹi,d [k]
where 3̃ denotes the frequency domain predefined preamble .


d̃ LSTM-FNNi,d [k] = D (8)
sequence. The final DPA channel estimates are updated in the h̃ˆ LSTM-FNNi−1,d [k]
following manner
Finally, the DPA channel estimation is applied using
ỹ [k]
h̃ˆ DPAi [k] = i . (6) d̃ LSTM-FNNi,d , such that
d̃ i [k]
It is worth mentioning that conventional DPA estimation ỹi,d [k]
h̃ˆ DLi,d [k] = . (9)
is based on the basic LS estimation that suffers from d̃ LSTM-FNNi,d [k]
noise enhancement. In addition to that, due to the doubly-
selective channel variation, a significant demapping error
We note that at the beginning of the frame (i = 1), h̃ˆ i−1,p
is recorded in (4). This demapping error is enlarged and
denotes the LS estimated channel at the received preamble
accumulated over the received frame resulting in a significant
symbols as shown in (5). While this estimator can outperform
degradation in performance. In order to address the limitation
the FNN-based estimators, it encounters a high complexity
of the conventional DPA channel estimation, the recently
cost arising from the employment of two DL networks.
proposed DL-based channel estimators employ LSTM prior
to the DPA estimation so that the demapping error can
2) LSTM-DPA-TA
be implicitly reduced, thus, improving the accuracy of the
channel estimation. In our previous work [23], an optimized LSTM network is
used instead of two as implemented in the LSTM-FNN-DPA
1) LSTM-FNN-DPA estimator. In addition, noise compensation is made possible
The work proposed in [22] shows that employing the LSTM by applying TA processing. This methodology only requires
processing prior to the DPA estimation could lead to a the previous pilots h̃ˆ i−1,p besides the LSTM estimated
significant improvement in the overall performance. In this channel as an input. Then, the LSTM estimated channel is
context, two cascaded LSTM and FNN networks for both employed in the DPA estimation as follows
channel estimation as well as noise compensation. The
ỹi [k]
ˆ
LSTM-FNN-DPA estimator employs the least squares (LS) d̃ LSTMi [k] = D , h̃LSTM0 [k] = h̃ˆ LS [k],
estimated channel at the current and previous received pilots h̃ˆ LSTMi−1 [k]
denoted by h̃ˆ i,p ∈ CKp ×1 and h̃ˆ i−1,p ∈ CKp ×1 , such that (10)
ỹi [k]
ỹi,p [k] ˆ ỹi−1,p [k] h̃ˆ LSTM-DPAi [k] = . (11)
h̃ˆ i,p [k] = , h̃i−1,p [k] = , k ∈ Kp . (7) d̃ LSTMi [k]
x̃i,p [k] x̃i,p [k]
4 VOLUME 2, 2024
Gizzini and Chafii: RNN-Based Channel Estimation in Doubly Selective Environments

FIGURE 2. Block diagram of the SoA CNN-based FBF estimators.

AWGN noise alleviation can be achieved by further and three pilot symbols in low, high, and very high mobility
applying TA processing such that scenarios, respectively. Following the selection of the frame
structure, the WI-CNN estimators proceed as follows
1 1
h̄ˆ DL-TAi,d = (1 − )h̄ˆ DL-TAi−1,d + h̄ˆ LSTM-DPAi,d . (12) • Pilot symbols channel estimation: In order to estimate
α α the channel at the inserted pilot symbols, the basic LS
Here, α denotes the utilized weighting coefficient. In [23], the denoted as simple LS (SLS) estimation is applied using
authors use a fixed α = 2 for simplicity. Therefore, the TA the received preambles as shown in (5), and using each
applied in (12) reduces the AWGN noise power σ 2 iteratively received pilot symbol such that
within the received OFDM frame according to the ratio
ỹ(p)
q [k]
 (q−1) X
1
q  (q−j+1)
1 4q−1 + 2 h̃ˆ SLSq [k] = = h̃q [k] + ṽq [k], k ∈ Kon . (14)
RDL-TAq = + = . p̃[k]
4 4 3 × 4q−1
j=2 where ṽq [k] represents the noise at the q-th received
(13) pilot symbol, 1 ≤ q ≤ Q denotes the inserted pilot
(p) (p)
symbol index, and Ỹ Q = [ỹ1 , . . . , ỹ(p)
q , . . . , ỹQ ] ∈
This corresponds to the AWGN noise power ratio of
CKon ×Q . Moreover, accurate LS (ALS) can be obtained
the estimated channel at the q-th estimated channel, where
by applying the discrete Fourier transform (DFT)
1 < q < I + 1 and RDL-TA1 = 1 denotes the AWGN noise
interpolation of ĥq,L such that
power ratio at h̃ˆ LS . From the derivation of RDL-TAq , it can
be seen that the noise power decreases over the received h̃ˆ ALSq = Fon ĥq,L , k ∈ Kon , (15)
OFDM frame, i.e. the SNR increases, resulting in an overall
improved performance. The full derivation of (13) is found with ĥq,L ∈ CL×1 denotes the estimated channel
in [23]. Even though the LSTM-DPA-TA improves the impulse response at the q-th received pilot symbol.
performance compared to the LSTM-FNN-DPA estimator, We note that the ALS and SLS are used for full pilot
it still suffers from high computational complexity. Moreover, (FP) allocation. However, if the number of channel taps
in Section IV, we show that employing LSTM unit in the L is known, the channel estimation requires only L pilots
channel estimation would affect the estimation accuracy in each pilot symbol, where DFT interpolation can be
negatively due to the long memory of the LSTM unit, applied to the estimated channel impulse response ĥq,L
where uncorrelated previous estimated channels are taken such that
into consideration to estimate the current channel. However,
the proposed GRU-based channel estimation provides a better h̃ˆ DFTq = Fon ĥq,L , k ∈ Kon , (16)
performance-complexity trade-off.
where Fon ∈ CKon ×L denotes the truncated DFT
B. CNN-BASED FBF CHANNEL ESTIMATION SCHEMES matrices obtained by selecting Kon rows, and L columns
In [24], a CNN aided weighted interpolation (WI) channel from the K -DFT matrix.
estimation schemes have been proposed. The WI-CNN • Data symbols channel estimation: After estimating the
estimators use adaptive frame structure according to the channel at the inserted Q pilot symbols, the WI-CNN
mobility scenario. The idea is to avoid using comb pilot estimator divides the received frame into several sub-
allocation and insert Q pilot OFDM symbols with different frames that are grouped as follows
configurations within the transmitted OFDM frame instead. ˆ = [h̃ˆ ˆ
In this context, the WI-CNN estimators employ one, two, H̃ q q−1 , h̃q ], q = 1, · · · Q, (17)

VOLUME 2, 2024 5
 h̃ˆ q refers to the implemented LS estimation. Then, the between them is in how the input data is processed by each
estimated channel for the i-th received data OFDM RNN.
symbol within each sub-frame is calculated as a The SRNN is useful when we need to look at recent
weighted summation of the estimated channels at the information only to perform a present task. Therefore, when
pilot symbols, such that an SRNN is exposed to long sequences, it tends to lose the
ˆ ˆ information because it cannot store long sequences since it
H̃ WIf = H̃ f C f , (18) focuses only on the latest output. This problem is commonly
ˆ ∈ CKon ×2 denotes LS estimated channels at referred to as vanishing gradients [33] that occurs during
where H̃ f the training phase, where useful gradients cannot propagate
the pilot symbols within the f -th sub-frame. C f ∈ R2×If from the output of the model back to the layers near the
denotes the interpolation weights of the If OFDM data input of the model. As a result, the RNN does not learn
symbols within the f -th sub-frame. The interpolation the effect of earlier inputs and it is too difficult for RNN to
weights of C f are calculated by minimizing the mean preserve information over many time steps, hence, causing
squared error (MSE) between the ideal channel H̃ f , and the short-term memory problem. To overcome this problem,
the LS estimated channel at the OFDM pilot symbols specialized versions of RNN like LSTM and GRU are
ˆ as derived in [30]. In the final step, optimized super
H̃ f created.
resolution CNN (SR-CNN) is employed on top of the The LSTM is a special kind of RNN capable of learning
WI estimators in a low mobility scenario, whereas an long-term sequences. Unlike SRNN, where the input and
optimized denoising CNN (DN-CNN) is considered in the previous output are processed by a simple activation
high mobility one. layer, in LSTM the process is slightly complex, where the
The WI-CNN estimators suffer from high computational LSTM unit takes at each time input from three different
complexity. Moreover, using noise alleviation CNNs is not states represented by the current input state, the short-
sufficient to accurately estimate the doubly-selective channel. term memory state, and the long-term memory state from
Therefore, we propose a Bi-GRU channel estimator that the previous LSTM unit. These inputs are controlled and
performs 2D interpolation, unlike the SR-CNN and DN-CNN filtered by three main gates to regulate the information to
networks, which are based on noise alleviation techniques. be kept or discarded before passing the updated information
As a result, performance superiority of the proposed Bi- to the next LSTM unit. First of all, we have the input gate
GRU channel estimator can be achieved while recording that decides what information is relevant to add from the
a significant decrease of the computational complexity in current input state. After that, the forget gate decides what
comparison to the WI-CNN estimators as illustrated in is relevant to keep from prior steps represented by the long-
Section V and Section VI. term memory state. Finally, the output gate determines what
the next output should be by processing the current input,
IV. PROPOSED RNN-BASED CHANNEL ESTIMATION the previous short-term memory, and the newly computed
SCHEMES long-term memory to produce new short-term memory which
In this section, RNN main concepts and extensions are will be passed on to the next time step. LSTM is capable
first briefly introduced. Then, a detailed explanation of the of learning long-term sequences, and predicting the current
proposed RNN and Bi-RNN based schemes for SBS and FBF output is influenced by the long sequence of previous outputs.
channel estimation are presented, respectively. However, LSTM is not useful in all scenarios, especially,
when the successive inputs become uncorrelated over time.
A. RECURRENT NEURAL NETWORKS: REVIEW Since predicting the current output will be affected by uncor-
RNN is a type of artificial neural network (ANN) designed related previous outputs, the prediction accuracy is negatively
to work with sequential data. This sequential data can be affected.
in the form of time series, text, audio, video etc. RNN To overcome the problems encountered by SRNN and
uses the previous information in the sequence to produce LSTM, GRU has been proposed to provide a trade-off by
the current output, where it is incorporated with memory to using a shorter previous memory in comparison to LSTM.
take information from prior inputs to influence the current GRU is based on the same concept as the LSTM, but with
output. This mechanism is the key essence to RNN success optimized architecture, where it has only two gates, a reset
in sequential problems, where the core concept of RNNs gate, and an update gate. The reset gate is used to decide how
is to keep/discard input data in a recurring manner. Hence, much of the past information needs to be neglected, while
the RNN can learn the important and unimportant data the update gate acts similarly to the forget and input gate of
contributing to predicting the current output. We note that an LSTM. It decides what information to throw away and
training and back-propagation operations in RNN are similar what new information to add from the current input as well as
to other forms of ANN, where RNN needs to be trained in determining the previous information that needs to pass along
order to produce accurate and desired outputs. In general, to the next time step. We note that fewer training parameters
there exist three main types of RNNs [31], [32]: (i) Simple and faster execution can be achieved by using GRU instead
RNN (SRNN), (ii) LSTM, and (iii) GRU. The main difference of LSTM.

6 VOLUME 2, 2024
Gizzini and Chafii: RNN-Based Channel Estimation in Doubly Selective Environments

Here, 9i is calculated for three mobility scenarios: (i) Low

mobility: fd = 250 Hz, (ii) High mobility: fd = 500 Hz,
and (iii) Very high mobility: fd = 1000 Hz. The detailed
properties of these channel models are provided in Section V.
As shown in Fig. 3, when the mobility increases, the
average correlation 9 starts to decrease exponentially.
However, as we can notice, 9i at the end of the received
frame reaches around 65% for low mobility scenario, while
it is around 40% in high and very high mobility scenarios,
with a drastic decrease in the overall 9i curve in very
high mobility scenarios. According to the 9i values in
different mobility scenarios, we can expect that the impact
of the estimated channels at earlier symbols would affect
negatively the accuracy of the estimated channel at advanced
symbols within the received OFDM frame. As a result, we
FIGURE 3. Correlation of the channel at the first and the last can conclude that, as the mobility increases, shorter RNN
OFDM symbol within the transmitted frame.
memory is required in the channel estimation in order to
guarantee the best possible performance. This is due to the
fact that, when long RNN memory is employed in a very high
B. PROPOSED SBS RNN-BASED CHANNEL ESTIMATOR mobility scenario, the older estimated channels negatively
It is worth mentioning that FNN network treats the initial impact the channel estimation at the current OFDM symbol
estimated channels separately, where it produces the output because the estimated channels become uncorrelated, i.e.
for each input. By doing this single input-output mapping, the value of 9 is low. In this context, the proposed RNN-
the FNN network is able to learn the frequency correlation based channel estimation scheme employs an optimized
of the doubly-selective channel, besides correcting the initial GRU unit instead of LSTM unit in the channel estimation
estimation error. On the contrary, RNN network treats the process due to its shorter memory. This results in improving
initial estimated channel as a correlated sequence, where the the accuracy of the channel estimation while recording a
current estimated channel is computed using the previous significant decrease in computational complexity. Moreover,
RNN estimated channels and the current initial estimated we study the performance of the SRNN unit in order to have
channel. This process allows the RNN network to learn a complete analysis of different RNN units.
both frequency and time correlation of the doubly-selective As illustrated in Fig. 4, the RNN unit is first employed to
channel, and thus, RNN outperforms FNN in the channel estimate the channel at the current data subcarriers, where
estimation task. We note that the SoA RNN-based estimators it takes as an input the previous LS estimated channels at
focus on the employment of different LSTM architectures pilot subcarriers denoted by h̄ˆ i−1,p ∈ R2Kp ×1 , concatenated
within the channel estimation process. However, the use with the previously RNN-based estimated channel at the data
of LSTM is not beneficial in all scenarios as discussed in subcarriers h̄ˆ 8-TAi−1,d ∈ R2Kd ×1 . Thus, the input and output
this section due to its high computational complexity and sizes of the RNN unit are 2Kon and 2Kd , respectively. After
long previous memory. In this context, the proposed RNN- that, the RNN output is fed as an input to the DPA estimation
based estimation scheme sheds light on the ability of GRU followed by TA processing in order to further mitigate
in estimating doubly-selective channels with high accuracy, the impact of noise. We note that our proposed estimators
tackling the limitations of the LSTM unit. consider both time and frequency selectivity, where the
As discussed in Section IV-A, SRNN takes advantage of RNN-based pre-processing deals with the time selectivity
the previously estimated channel only while estimating the and the DPA estimation with the frequency selectivity.
current one. Whereas, the LSTM has long-term memory, Moreover, the RNN-based estimated channel is fed as an
which means that estimating the channel at the current OFDM input to the DPA estimation block, which further improves the
symbol is affected by the older estimated channels. On the DPA estimation accuracy. The proposed RNN-based channel
other hand, GRU provides a trade-off between short-term estimation scheme proceeds as follows
memory and complexity. Therefore, in order to decide which
ȳi,d [k]

RNN performs better in doubly-selective channel estimation, d̄ 8i,d [k] = D , k ∈ Kd , (20)
we study the average correlation between the channel at the ˆh̄ [k]
8i−1,d
first symbol and all successive symbols within the transmitted ȳi,d [k]
OFDM frame, considering the frequency-time response, such h̄ˆ 8-DPAi,d [k] = , (21)
d̄ 8i,d [k]
that
where 8 ∈ {SRNN, GRU} refers to the used RNN unit, and
∗ h̄ˆ 80,d = h̄ˆ LS ∀ k ∈ Kd . Finally, to alleviate the impact of
h i
9i = E h̃1 h̃i , 2 ≤ i ≤ I. (19)
the AWGN noise, TA processing is applied to the estimated

VOLUME 2, 2024 7
 FIGURE 4. Proposed RNN-based channel estimation schemes.

TABLE 1. Parameters of the proposed RNN-based channel in most real case applications. For example, in vehicular
estimation scheme. communications, the vehicle velocity is a known parameter
that can be exchanged between all vehicular network nodes
and it must be regulated according to the road conditions.
In urban environments (inside cities) the car velocity must
not exceed 40 Kmphr, and thus, the model trained on low
mobility can be employed. Consequently, the RNN training
is performed for each mobility scenario separately using
the same architecture and training parameters summarized
in Table 1. However, when velocity information is not
available, EL algorithm can be used to combine the
weights of several trained models so that one generalized
model can be employed in all mobility scenarios as
discussed in V.

C. PROPOSED FBF Bi-RNN-BASED CHANNEL

ESTIMATOR
channel h̄ˆ 8-DPAi similarly as performed in (12), such that Bi-RNN networks are designed to predict unknown data
that are bounded within known data [25]. They are based
1 1
h̄ˆ 8-TAi,d = (1 − )h̄ˆ 8-TAi−1,d + h̄ˆ 8-DPAi,d , (22) on making the data flow through any RNN unit in both
α α directions forward (past to future), and backwards (future
where α = 2 for simplicity. We note that in doubly- to past). In regular RNN, the input flows in one direction,
selective channel, each two successive symbols are correlated whereas, in Bi-RNN the input flows in both directions to get
regardless of the mobility scenario, therefore, using α = the advantage of both past and future information. By doing
2 gives equal weights for the previous and current estimated so, the Bi-RNN network will be able to predict the unknown
channel. However, However, alpha can be fine-tuned by information in the middle based on its correlation with the
studying the average channel correlation between each two known past and future information.
successive OFDM symbols, and then assigning more accurate In this context, the proposed Bi-RNN channel estimator
weights to the previous and current estimated channels aims to utilize the interpolation ability of Bi-RNN networks
in (22). in the FBF channel estimation instead of employing high-
In the proposed scheme, RNN training is performed using a complexity CNN networks as it is the case in the SoA
high value of signal-to-noise ratio (SNR) = 40 dB to achieve channel estimation schemes. The proposed Bi-RNN channel
the best performance as observed in [34]. The reason is that estimation scheme uses Bi-GRU unit and it inherits the
when the training is performed for low noise impact, the RNN adaptive frame design from the WI-CNN estimators as shown
is able to better learn the channel correlation. In addition, due in Fig. 5. Recall that WI-CNN channel estimation performs
to its good generalization ability, it can still perform well in WI interpolation at the data symbols, where CNN processing
low SNR regions, where the noise is dominant. Moreover, is applied to alleviate the impact of noise. However, Bi-
intensive experiments are performed using the grid search RNNs perform 2D interpolation at the data symbols using
algorithm [35] in order to select the best suitable RNN hyper- the estimated channel at the pilot symbols without the need
parameters in terms of both performance and complexity. for any initial channel estimation at the data symbols. Thus,
Note that the mobility conditions can be assumed known the proposed Bi-RNN channel estimator can be adapted

8 VOLUME 2, 2024
Gizzini and Chafii: RNN-Based Channel Estimation in Doubly Selective Environments

FIGURE 5. Proposed Bi-RNN based channel estimator block diagram.

to any existing protocols regardless of the pilot allocation low mobility (v = 45 Kmph, fd = 250 Hz) (ii) High mobility
scheme. However, the employed Bi-RNN architectures (v = 100 Kmph, fd = 500 Hz) (iii) Very high mobility (v =
should be fine-tuned accordingly to meet the required per- 200 Kmph, fd = 1000 Hz). The power-delay profiles of the
formance. The proposed Bi-RNN channel estimator proceeds employed channel models are provided in Table 2. Moreover,
as follows the performance of the studied channel estimators is further
• ALS estimation at the inserted pilot symbols as per- evaluated using the QuaDRiGa channel model [36]. We note
formed in (15), followed by zero insertion at all the that the QuaDRiGa channel model is a three-dimensional
data symbols. Thereafter, the initial estimated channels geometry-based stochastic channel model that offers several
ˆ ∈ CKon ×Id are converted to the real-valued domain
H̃ features including the small and large-scale fading as well
in
by stacking the real and imaginary values in one vector, as the continuous-time evolution, therefore, it presents the
ˆ ∈ R2Kon ×Id . practical real-world characteristics of the wireless channels.
where H̄ in
ˆ It is worth mentioning that in order to guarantee fairness
• Bi-RNN end-to-end interpolation, where H̄ in is fed as
in the conducted simulations, the studied channel estimators
an input to the optimized Bi-GRU unit. Accordingly,
are trained using the same parameters shown in Table 1 that
the Bi-GRU unit learns the weights of the estimated
are optimized using the grid search algorithm. Moreover,
channels at the OFDM data symbols. Employing the 2D
simulation parameters are based on the IEEE 802.11p
interpolation using the proposed Bi-GRU unit leads to
standard [37], where for the SBS channel estimation,
a considerable performance superiority in comparison
comb-pilot allocation is employed such that Kp I pilots are
with the WI-CNN estimators while recording a signifi-
used within the transmitted frame following the comb-
cant decrease in the required computational complexity,
pilot allocation. Concerning the FBF channel estimation,
as shown in Section V. Also here the proposed Bi-
the ChannelNet and TS-ChannelNet estimators use Kp I
GRU architecture is optimized using the grid search
pilots per frame, whereas the WI-CNN and the proposed
algorithm [35] and trained using the parameters listed
Bi-RNN channel estimators employ only Kon Q pilots per
in Table 1. Moreover, similarly as performed in
frame following the block-pilot allocation, where Kon =
Section IV-B, the performance of Bi-LSTM and Bi-
52 denotes the number of employed subcarriers within the
SRNN are investigated in Section V.
transmitted OFDM symbol, and Q is the number of inserted
pilot symbols within the transmitted frame ((i) Low mobility:
V. SIMULATION RESULTS Q = 1, (ii) High mobility: Q = 2, (iii) Very high mobility:
This section illustrates the performance evaluation of the Q = 3). Therefore, the proposed Bi-RNN based channel
SoA and the proposed RNN and Bi-RNN based channel estimator is able to outperform the recently proposed SoA
estimation schemes in terms of BER and throughput. FBF channel estimators employing fewer pilots with lower
Vehicular communications are considered as a simulation computational complexity, resulting in higher transmission
case study, where three mobility scenarios are defined as: (i) data rates as discussed in Section V-B. We also note that these

VOLUME 2, 2024 9
 FIGURE 6. BER for I = 100, mobility from left to right: low, high, very high.

simulations are implemented using QPSK, 16QAM, and A. SBS CHANNEL ESTIMATION
64QAM modulation orders, the SNR range is [0, 5, . . . , 40] 1) MODULATION ORDER
dB. In addition, the performance evaluation is performed For QPSK modulation order, we can notice from Fig. 6a
according to the employed modulation orders, the mobility that FNN-based channel estimators can implicitly learn
scenarios, and variable frame length. the channel frequency correlation apart from preventing a

10 VOLUME 2, 2024
Gizzini and Chafii: RNN-Based Channel Estimation in Doubly Selective Environments

TABLE 2. Characteristics of the employed channel models following jake’s doppler spectrum.

high demapping error arising from conventional DPA-based of SNR for BER = 10−4 and BER = 10−3 , respectively,
estimation, where STA-FNN and TRFI-FNN outperform in very high mobility scenarios, as illustrated in Fig. 6b
conventional STA and TRFI estimators by at least 15 dB and Fig. 6c. However, it can be noticed that FNN-based
gain in terms of SNR for BER = 10−3 . However, STA-FNN channel estimators suffer from severe performance degrada-
suffers from an error floor beginning from SNR = 20 dB, tion when 64QAM modulation is employed. This is because
particularly in very high mobility scenarios. This is due to of the remarkable accumulated DPA demapping error that
the STA frequency and time averaging operations that can cannot be eliminated by simple FNN architectures. A nice
alleviate the impact of noise and demapping error in low observation can be noticed from Fig. 6 where employing
SNR regions. On the other hand, the averaging operations SRNN in the channel estimation performs similarly to the
are not useful in high SNR regions since the impact of LSTM-FNN-DPA estimator in all mobility scenarios. This
noise is low, and the STA averaging coefficients are fixed. reveals that using SRNN combined with TA processing
Therefore, TRFI-FNN is used to improve the performance records similar performance as LSTM combined with FNN.
at high SNRs to compensate for the STA-FNN performance In other words, the performance degradation caused by
degradation in the high SNR regions. We can clearly observe the LSTM long-term memory is compensated by the FNN
that employing RNNs as a pre-processing unit rather than a network in the LSTM-FNN-DPA estimator. However, SRNN
simple FNN in the channel estimation brings a significant unit can be used instead to eliminate the LSTM long-
improvement in the overall performance. This is because term memory problem as well as mitigating the noise by
RNNs are capable of efficiently learning the time correlations simple TA processing as the case in the SRNN-DPA-TA
of the channel by taking the advantage of the previous estimator.
output apart from the current input in order to estimate the Fig. 7 illustrates the throughput of the studied SBS channel
current output. Even though the recently proposed LSTM- estimators employing QPSK modulation. It can be seen
based estimators are able to outperform the FNN-based that the proposed RNN-based channel estimators perform
estimator, but using LSTM in the channel estimation is not higher throughput than conventional and FNN-based channel
the best option, due to LSTM long-term memory problem. estimators, especially in low SNR regions. This is due to the
In contrast, we can notice that the proposed GRU-DPA-TA accurate channel prediction.
estimator is able to outperform the LSTM-DPA-TA estimator
by around 6 dB gain in terms of SNR for BER = 10−5 , 2) MOBILITY
especially, in very high mobility scenario. This is due to the The impact of mobility can be observed in Fig. 6. The
fact that LSTM employs long-term memory, thus, the current performance behavior is influenced by the following factors:
estimated channel is affected by older estimated ones. This (i) channel estimation error, (ii) time diversity due to
process harms the performance as the mobility increases, and increased Doppler spread, since the Doppler spread and the
the channel at successive received OFDM symbols becomes time diversity gain are proportional, i.e. more time diversity
uncorrelated. Whereas, the GRU uses shorter memory than gain can be obtained in very high mobility scenarios, and (iii)
LSTM, Thus, leading to the superiority of the proposed GRU- frame length. As shown in Fig. 6, where the frame length is
DPA-TA estimator in comparison with the LSTM-DPA-TA fixed (I = 100), the performance of all the studied channel
estimator. However, we can notice that in low mobility estimation schemes degrades with the increase of mobility.
scenario, both LSTM-DPA-TA and GRU-DPA-TA estimators This is because the channel estimation error increases with
achieve almost similar performance. This is because of the the increase of Doppler frequency. Moreover, we can notice
negligible impact of Doppler interference in low mobility that the conventional STA and TRFI channel estimators
scenario, thus, the channels at successive symbols within the suffer from severe performance degradation in the very high
received OFDM frame are highly correlated. So, considering mobility scenario, since the impact of the AWGN noise
long or short memory while estimating the current channel and DPA demapping error is much more dominant than
will not lead to considerable performance degradation. the time diversity gain. On the contrary, the time diversity
Concerning 16QAM and 64QAM modulation orders, the gain is dominant in DL-based channel estimators, since DL
proposed GRU-DPA-TA estimator outperforms the LSTM- networks are capable of reducing the channel estimation error
DPA-TA estimator by more than 5 dB and 7 dB gains in terms resulting from the AWGN noise and the DPA demapping

VOLUME 2, 2024 11
 FIGURE 7. Throughput employing QPSK, mobility from left to right: low, high, very high.

FIGURE 8. BER employing very high mobility and QPSK, frame length from left to right: I = 10, I = 50, I = 100.

error, leading to a performance improvement in very high GRU-DPA-TA channel estimator when the outdated esti-
mobility scenarios. Note that the net time diversity gain is mated channel is used. In this context, the received OFDM
also related to the employed frame length, since increasing symbols are equalized by the DL-based estimated channel at
the frame length increases the time diversity gain. This is the beginning of the frame. As shown in Fig. 8, equalizing
clearly illustrated in Fig. 8, where QPSK modulation order by the outdated estimated channel significantly degrades
with very high mobility is utilized. As we can notice, the the performance even when shorter frames are employed.
performance of the proposed RNN-based channel estimators Therefore, this shows the importance of applying channel
improves when a longer frame length is employed. It is worth tracking to guarantee good performance in different mobility
mentioning that, the proposed GRU-DPA-TA and LSTM- scenarios.
DPA-TA estimators perform similarly when I = 10, since Fig. 9 illustrates the robustness of the proposed GRU-
in shorter frames the impact of long and short-term memory DPA-TA channel estimator in addition to the LSTM-DPA-TA
cannot be clearly illustrated. On the contrary, when using and SRNN-DPA-TA channel estimators against the change
longer frames, i.e, I = 50 and I = 100, we can notice the in Doppler frequency, where QPSK modulation is employed.
superiority of using the GRU-based estimator instead of the RNN training is performed on one Doppler frequency and
LSTM-based one. tested on the entire range of Doppler frequencies. In this
In order to further illustrate the importance of channel context, the entire range of Doppler frequencies is divided
tracking, Fig. 8 shows the performance of the proposed into 3 ranges: (i) Low mobility (0 Hz - 300 Hz), (ii)

12 VOLUME 2, 2024
Gizzini and Chafii: RNN-Based Channel Estimation in Doubly Selective Environments

FIGURE 9. Robustness against Doppler frequencies, SNR = 40 dB. RNN from left to right: SRNN, GRU, LSTM.

TABLE 3. QuaDRiGa car-to-car channel model parameters.

degradation which is expected since the model is trained in

the absence of Doppler interference.
On the other hand, the proposed GRU-DPA-TA channel
estimator provides better robustness in comparison to the
FIGURE 10. Robustness analysis of the proposed GRU-DPA-TA other RNN-based channel estimators due to its ability to
channel estimator employing QPSK modulation. learn the time correlation efficiently using its short memory.
It is worth mentioning that, further model generalization
can be achieved by using the concept of ensemble learning
High mobility (300 - 600 Hz), and (iii) Very high mobility (EL) in case the velocity range is not known, where the
(600 Hz - 1000 Hz). As we can notice, training on one weights of several trained models can be averaged in order to
Doppler frequency and testing on the same one gives produce one generalized model as shown in Fig. 10, where the
the best performance. However, training on the highest model is trained on Doppler frequency and tested on different
Doppler frequency within each range shows a satisfactory Doppler frequencies considering the entire SNR range.
performance when tested on different Doppler frequencies We note that in Fig. 9, the EL results are obtained by aver-
within the considered range. This can be explained by the aging the weights of the trained models on 700 Hz, 800 Hz,
fact that when the model is trained on the worst conditions, and 900 Hz. Therefore, the performance of the trained models
i.e., high Doppler frequency, it can perform well when tested can be generalized according to the requirements of real-time
on better conditions, i.e., low Doppler frequency. However, applications.
the opposite does not work, especially, in the case of
SRNN-DPA-TA channel estimator which suffers from severe 3) PERFORMANCE EVALUATION USING QuaDRiGa
performance degradation when the trained model is tested CHANNEL MODEL
on higher Doppler frequencies due to its simple architecture. The QuaDRiGa channel model is used in order to further
Similarly, for the LSTM-DPA-TA channel estimator but with investigate the robustness of the proposed GRU-DPA-TA
smoother degradation. Therefore, training on lower Doppler channel estimation scheme. We consider a Car-to-Car chan-
frequencies, (for example, fd = 250 Hz) and testing on nel model with the dual mobility feature [38] (Section 4.14).
higher Doppler frequencies lead to a severe performance In this scenario, we have two vehicles moving toward each

VOLUME 2, 2024 13
 ALS-WI-CNN estimator there are no fixed parameters, and
the time correlation between the previous and the future
pilot symbols is considered in the WI interpolation operation.
These aspects lead to the performance superiority of the ALS-
WI-CNN compared to the ChannelNet and TS-ChannelNet
estimators. Although CNN processing is applied in the
ChannelNet, TS-ChannelNet, and ALS-WI-CNN estimators,
they suffer from a considerable performance degradation
that is dominant in very high mobility scenario. This show
that the CNN processing is not able to effectively alleviate
the impact of Doppler interference, especially in very high
mobility scenarios, where the proposed Bi-RNN based
channel estimation scheme outperforms the WI-ALS-CNN
estimator by at least 5 dB and 12 dB gain in terms of
SNR for a BER = 10−5 employing QPSK and 16QAM
FIGURE 11. BER performance employing QuaDRiGa channel
modulations, respectively. We note that the robustness of
model. I = 100 employing QPSK modulation. the proposed Bi-RNN based channel estimator against high
mobility is mainly due to the accuracy of the end-to-end
2D interpolation implemented by the utilized Bi-GRU unit.
Moreover, we can see that employing Bi-LSTM performs
other on linear tracks considering the setup parameters as
similarly to the ALS-Bi-GRU estimator, this is due to the used
shown in Table 3. QPSK modulation with I = 100 OFDM
frame structure, where the variation of the doubly-selective
symbols per frame.
channel within each sub-frame is low. However, it can be
Figure 11 shows the BER performance of the RNN-based
noticed that employing CNN performs better than the Bi-
channel estimators. Performance degradation is noticed in
SRCNN unit in low and high mobility scenarios, while using
comparison to the VTV-SDWW channel model, Fig. 6. This
Bi-SRCNN unit leads to around 2 dB gain in terms of SNR
is due to simulating the large-scale fading effects, i.e., the
for a BER = 10−4 in comparison with the ALS-WI-CNN
shadowing and the path loss in the QuaDRiGa channel model.
estimator in very high mobility scenario as shown in Fig. 12.
However, the proposed GRU-DPA-TA channel estimator
As a result, we can conclude that employing Bi-GRU unit
outperforms the other RNN-based channel estimators by at
instead of CNN network leads to more accurate channel
least 5 dB gain in terms of SNR for a BER = 10−3 . Again,
estimation with lower complexity. Finally, we note that the
the long memory of the LSTM shows a negative impact
performance of the 2D-LMMSE estimator is comparable to
on the BER performance, unlike the proposed GRU-based
the performance of the ideal channel but it requires huge
estimator. We note that employing LSTM unit with a larger
complexity as we discuss in the next section, which is
hidden size leads to a marginal performance improvement in
impractical in a real scenario. Moreover, the proposed Bi-
comparison to the optimized LSTM architecture. Therefore,
RNN based estimator records almost close performance as
the proposed GRU-DPA-TA channel estimator provides a
the 2D-LMMSE estimator. Therefore, the proposed Bi-RNN
good performance-complexity trade-off.
based channel estimator is an alternative to the 2D-LMMSE
estimator where it provides a good performance-complexity
B. FBF CHANNEL ESTIMATION trade-off.
In this section, performance evaluations of the CNN-based
estimators, conventional 2D LMMSE estimator as well as 2) MOBILITY
the proposed Bi-RNN based channel estimator are discussed The impact of mobility can be clearly observed in
using the same criteria as Section V-A. We note that we only Fig. 12b, where the performance of the ChannelNet and
consider the ALS-WI-CNN among the WI-CNN estimators TS-ChannelNet channel estimation schemes degrades as the
since it has the best performance. mobility increases, and the impact of the time diversity gain
is not dominant due to the high estimation error of the 2D
1) MODULATION ORDER RBF and ADD-TT interpolation techniques employed in
Fig. 12a and Fig. 12b depict the BER performance employing the ChannelNet and TS-ChannelNet estimators, respectively.
QPSK and 16QAM modulation orders, respectively. The In contrast, the time diversity gain is dominant in the
performance of channel network (ChannelNet) and tempo- ALS-WI-CNN and the proposed Bi-RNN based channel
ral spectral ChannelNet (TS-ChannelNet) accounts of the estimator, since the initial ALS and WI estimations are
predefined fixed parameters in the applied interpolation accurate, thus, the SR-CNN and DN-CNN networks are
scheme, where the RBF interpolation function and the ADD- capable of overcoming the Doppler interference. However,
TT frequency and time averaging parameters need to be using the ALS estimation at the pilot symbols followed
updated in a real-time manner. On the contrary, in the by Bi-GRU unit for 2D interpolation at the data symbols

14 VOLUME 2, 2024
Gizzini and Chafii: RNN-Based Channel Estimation in Doubly Selective Environments

FIGURE 12. BER for I = 100, mobility from left to right: low, high, very high.

reveal considerable robustness against mobility. This is due execution time, where the simulations are performed on
to the ability of the optimized Bi-GRU unit to significantly NVIDIA TITAN Xp GPU card (12 GB GDDR5X).
alleviating the impact of Doppler interference, where it can
be noticed that the proposed Bi-RNN estimator is able to out- A. SBS CHANNEL ESTIMATION
perform the ALS-WI-CNN estimators in different mobility Table 4 shows the FLOPS and the GPU execution time of the
scenarios. As a result, the proposed Bi-RNN based channel studied SBS channel estimation schemes. We can notice that
estimator provides a good performance-complexity trade- the FNN-based channel estimation schemes, especially, STA-
off between the CNN-based estimators and 2D-LMMSE FNN and TRFI-FNN record the lowest number of FLOPS
estimator. and execution time due to their optimized FNN architectures
in comparison to other schemes. Employing both LSTM
VI. COMPUTATIONAL COMPLEXITY ANALYSIS and FNN networks requires 255.84 kFLOPS as the case in
This section provides a detailed computational complexity the LSTM-FNN-DPA channel estimation scheme. Moreover,
analysis of the studied channel estimation schemes. The 86.01 kFLOPS are required by the LSTM-DPA-TA channel
computational complexity analysis is performed according estimator. This significant decrease in the number of FLOPS
to two criteria: (i) The number of FLOPS required by the is due to optimized LSTM architecture as well as using
DL architecture used in each channel estimator.2 (ii) The only one DL network in the channel estimation. However,
the proposed optimized GRU-DPA-TA estimator is able to
decrease the required number of FLOPS by around 82%
2 We note that the number of FLOPS are calculated using the pytorch- and 50% compared to the LSTM-FNN-DPA and LSTM-
OpCounter package [39]. DPA-TA channel estimators, respectively. In other words, the

VOLUME 2, 2024 15
 FIGURE 13. FLOPS of the studied DL-based SBS channel estimators.

FIGURE 14. FLOPS of the studied DL-based FBF channel estimators.

proposed GRU-DPA-TA estimator is 2x less complex than TABLE 4. FLOPS and GPU execution time of the studied SBS
the LSTM-DPA-TA estimator, at the same time, it achieves channel estimation schemes.
a significant BER performance gain as well as recording a
practical execution time.
We would like to mention that the number of FLOPS
can be further reduced by using the SRNN-DPA-TA channel
estimator since it is 2x less complex than the GPU-DPA-
TA channel estimator. However, the RNN-DPA-TA estimator
performs similarly to the LSTM-FNN-DPA estimator as
shown in Section V. Finally, we note that a trade-off
between the desired performance and the accepted complex-
ity should be taken into account, in order to optimize the
use of the RNN-based channel estimators. Fig. 13 shows
the required FLOPS by various examined SBS channel
estimators.

B. FBF CHANNEL ESTIMATION the CNN-based channel estimation schemes do not provide
Table 5 shows the FLOPS and the GPU execution time of the a good complexity vs. performance trade-off since a high
studied FBF channel estimation schemes. We can notice that number of FLOPS is required. In contrast, the complexity

16 VOLUME 2, 2024
Gizzini and Chafii: RNN-Based Channel Estimation in Doubly Selective Environments

TABLE 5. Flops and GPU execution time of the studied FBF REFERENCES
channel estimation schemes.
[1] H. Chang et al., ‘‘A novel nonstationary 6G UAV-to-ground wireless
channel model with 3-D arbitrary trajectory changes,’’ IEEE Internet
Things J., vol. 8, no. 12, pp. 9865–9877, Jun. 2021.
[2] M. Chafii, L. Bariah, S. Muhaidat, and M. Debbah, ‘‘Twelve scientific
challenges for 6G: Rethinking the foundations of communications
theory,’’ IEEE Commun. Surveys Tuts., vol. 25, no. 2, pp. 868–904,
2nd Quart., 2023.
[3] T. Ma, X. Jiang, Y. Wang, and F. Li, ‘‘A novel inter-carrier interference
cancellation scheme in highly mobile environments,’’ China Commun.,
vol. 17, no. 12, pp. 194–205, Dec. 2020.
[4] R. Bomfin, M. Chafii, A. Nimr, and G. Fettweis, ‘‘A robust baseband
transceiver design for doubly-dispersive channels,’’ IEEE Trans. Wireless
Commun., vol. 20, no. 8, pp. 4781–4796, Aug. 2021.
[5] J. A. Fernandez, K. Borries, L. Cheng, B. K. V. Kumar, D. D. Stancil,
and F. Bai, ‘‘Performance of the 802.11p physical layer in vehicle-to-
vehicle environments,’’ IEEE Trans. Veh. Technol., vol. 61, no. 1, pp. 3–14,
Jan. 2012.
[6] Y.-K. Kim, J.-M. Oh, Y.-H. Shin, and C. Mun, ‘‘Time and frequency
domain channel estimation scheme for IEEE 802.11p,’’ in Proc. 17th Int.
IEEE Conf. Intell. Transp. Syst. (ITSC), Oct. 2014, pp. 1085–1090.
[7] S. Ehsanfar, M. Chafii, and G. P. Fettweis, ‘‘On UW-based transmission
is significantly decreased by the proposed ALS-BiGRU for MIMO multi-carriers with spatial multiplexing,’’ IEEE Trans. Wireless
channel estimator which is 10x and 115x less complex than Commun., vol. 19, no. 9, pp. 5875–5890, Sep. 2020.
the ALS-WI-SRCNN and the ALS-WI-DNCNN estimators, [8] M. Soltani, V. Pourahmadi, A. Mirzaei, and H. Sheikhzadeh, ‘‘Deep
learning-based channel estimation,’’ IEEE Commun. Lett., vol. 23, no. 4,
respectively. Therefore, we can conclude that employing pp. 652–655, Apr. 2019.
the proposed optimized Bi-GRU network instead of CNN [9] X. Zhu, Z. Sheng, Y. Fang, and D. Guo, ‘‘A deep learning-aided
networks in the channel estimation is more efficient, and temporal spectral ChannelNet for IEEE 802.11p-based channel estimation
at the same time, it offers better performance and feasible in vehicular communications,’’ EURASIP J. Wireless Commun. Netw.,
vol. 2020, pp. 1–15, Dec. 2020.
execution time. Fig. 14 illustrates the FLOPS of the studied [10] T. O’Shea and J. Hoydis, ‘‘An introduction to deep learning for the physical
CNN-based FBF channel estimators. layer,’’ IEEE Trans. Cognit. Commun. Netw., vol. 3, no. 4, pp. 563–575,
Dec. 2017.
[11] M. Chafii, F. Bader, and J. Palicot, ‘‘Enhancing coverage in narrow band-
VII. CONCLUSION AND FUTURE PERSPECTIVES IoT using machine learning,’’ in Proc. IEEE Wireless Commun. Netw. Conf.
In this paper, RNN-based channel estimation in doubly- (WCNC), Apr. 2018, pp. 1–6.
[12] W. Njima, M. Chafii, A. Chorti, R. M. Shubair, and H. V. Poor, ‘‘Indoor
selective environments has been investigated. The recently localization using data augmentation via selective generative adversarial
proposed DL-based SBS and FBF channel estimators have networks,’’ IEEE Access, vol. 9, pp. 98337–98347, 2021.
been presented and their limitations have been discussed. [13] W. Njima, A. Bazzi, and M. Chafii, ‘‘DNN-based indoor localization under
In order to overcome these limitations, we have proposed limited dataset using GANs and semi-supervised learning,’’ IEEE Access,
vol. 10, pp. 69896–69909, 2022.
optimized RNN-based and Bi-RNN estimators for SBS [14] W. Njima, M. Chafii, A. Nimr, and G. Fettweis, ‘‘Deep learning based data
and FBF channel estimation respectively. Moreover, the recovery for localization,’’ IEEE Access, vol. 8, pp. 175741–175752, 2020.
performance of several RNNs architectures including SRNN, [15] Y. Yang, F. Gao, X. Ma, and S. Zhang, ‘‘Deep learning-based channel
LSTM, and GRU has been thoroughly analyzed based estimation for doubly selective fading channels,’’ IEEE Access, vol. 7,
pp. 36579–36589, 2019.
on the channel correlation within the received frame.
[16] J. Yuan, H. Q. Ngo, and M. Matthaiou, ‘‘Machine learning-based channel
Moreover, we show that the proposed GRU and Bi-GRU prediction in massive MIMO with channel aging,’’ IEEE Trans. Wireless
units result in a better performance-complexity trade-off Commun., vol. 19, no. 5, pp. 2960–2973, May 2020.
in different mobility scenarios. Simulation results have [17] H. Kim, S. Kim, H. Lee, C. Jang, Y. Choi, and J. Choi, ‘‘Massive MIMO
channel prediction: Kalman filtering vs. machine learning,’’ IEEE Trans.
shown the performance superiority of the proposed channel Commun., vol. 69, no. 1, pp. 518–528, Jan. 2021.
estimators over the recently proposed DL-based SBS and [18] C. Wu, X. Yi, Y. Zhu, W. Wang, L. You, and X. Gao, ‘‘Channel prediction
FBF estimators while recording a significant reduction in high-mobility massive MIMO: From spatio-temporal autoregression
in computational complexity and execution time. As a to deep learning,’’ IEEE J. Sel. Areas Commun., vol. 39, no. 7,
pp. 1915–1930, Jul. 2021.
future perspective, advanced DL algorithms such as transfer [19] S. Han, Y. Oh, and C. Song, ‘‘A deep learning based channel estimation
and meta-learning can be investigated in order to further scheme for IEEE 802.11p systems,’’ in Proc. IEEE Int. Conf. Commun.
improve the online performance of the proposed channel (ICC), May 2019, pp. 1–6.
estimators and their generalization capabilities. Moreover, [20] A. K. Gizzini, M. Chafii, A. Nimr, and G. Fettweis, ‘‘Deep learning based
channel estimation schemes for IEEE 802.11p standard,’’ IEEE Access,
working on interpretable and explainable theoretical DL vol. 8, pp. 113751–113765, 2020.
models is a crucial future step that would ensure the [21] A. K. Gizzini, M. Chafii, A. Nimr, and G. Fettweis, ‘‘Joint TRFI and
reliability and transparency of employing DL networks in the deep learning for vehicular channel estimation,’’ in Proc. IEEE Globecom
Workshops (GC Wkshps, Dec. 2020, pp. 1–6.
domain of wireless communications [40], especially, channel
[22] J. Pan, H. Shan, R. Li, Y. Wu, W. Wu, and T. Q. S. Quek, ‘‘Channel
estimation, where the intuitions behind our proposed work estimation based on deep learning in vehicle-to-everything environments,’’
can be further validated. IEEE Commun. Lett., vol. 25, no. 6, pp. 1891–1895, Jun. 2021.

VOLUME 2, 2024 17
[23] A. K. Gizzini, M. Chafii, S. Ehsanfar, and R. M. Shubair, ‘‘Temporal ABDUL KARIM GIZZINI (Member, IEEE)
averaging LSTM-based channel estimation scheme for IEEE 802.11p stan- received the bachelor’s and master’s degrees in
dard,’’ in Proc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2021, computer and communication engineering from
pp. 1–7. IUL University, Lebanon, in 2015 and 2017,
[24] A. Karim Gizzini, M. Chafii, A. Nimr, R. M. Shubair, and G. Fettweis, respectively, and the Ph.D. degree in telecom-
‘‘CNN aided weighted interpolation for channel estimation in vehic- munications engineering from Cergy Paris CY
ular communications,’’ IEEE Trans. Veh. Technol., vol. 70, no. 12, University, France, in 2021. His master’s thesis
pp. 12796–12811, Dec. 2021.
was conducted in collaboration with the Lebanese
[25] B. Lindemann, T. Müller, H. Vietz, N. Jazdi, and M. Weyrich, ‘‘A survey on
National Council for Scientific Research (CNRS-
long short-term memory networks for time series prediction,’’ Proc. CIRP,
L). During 2020 he has been a visiting Researcher
vol. 99, pp. 650–655, Jan. 2021.
[26] C. Dong, C. C. Loy, K. He, and X. Tang, ‘‘Image super-resolution using at Vodafone Chair Mobile Communications Systems, Technical University
deep convolutional networks,’’ IEEE Trans. Pattern Anal. Mach. Intell., of Dresden - Germany. Moreover, he has an active research collaboration
vol. 38, no. 2, pp. 295–307, Feb. 2016. with the Technical University of Chemnitz - Germany, the National Institute
[27] K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, ‘‘Beyond a Gaussian of Telecommunications (Inatel) - Brazil, New York University (NYU) - Abu
denoiser: Residual learning of deep CNN for image denoising,’’ IEEE Dhabi, CNRS-L. In 2022, he was a Post-Doctoral Researcher and a Temporal
Trans. Image Process., vol. 26, no. 7, pp. 3142–3155, Jul. 2017. Lecturer at ENSEA (ETIS UMR8051). Since January 2023 he is a Post-
[28] A. K. Gizzini and M. Chafii, ‘‘Deep learning based channel estimation in Doctoral Researcher at the Center of digital systems (CERI-SN), IMT Nord
high mobility communications using bi-RNN networks,’’ in Proc. IEEE Europe, Institut Mines-Télécom, Lille, France. His research interests include
Int. Conf. Commun., May 2023, pp. 1–6. wireless communications, vehicular communications, channel estimation,
[29] Z. Hong, Z. Yang, H. Wang, D. Li, W. Nai, and Y. Xing, ‘‘The weighted and explainable artificial intelligence for wireless communications and
average ensemble learning based on polar bear algorithm with image processing.
T-distribution parameters,’’ in Proc. IEEE 9th Joint Int. Inf. Technol.
Artif. Intell. Conf. (ITAIC), vol. 9, Dec. 2020, pp. 1902–1905.
[30] Y. R. Zheng and C. Xiao, ‘‘Channel estimation for frequency-domain
equalization of single-carrier broadband wireless communications,’’ IEEE
Trans. Veh. Technol., vol. 58, no. 2, pp. 815–823, Feb. 2009.
[31] A. Sherstinsky, ‘‘Fundamentals of recurrent neural network (RNN)
and long short-term memory (LSTM) network,’’ Phys. D, Nonlinear
Phenomena, vol. 404, Mar. 2020, Art. no. 132306. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0167278919305974
[32] R. Dey and F. M. Salem, ‘‘Gate-variants of gated recurrent unit (GRU) MARWA CHAFII (Senior Member, IEEE)
neural networks,’’ in Proc. IEEE 60th Int. Midwest Symp. Circuits Syst.
received the master’s degree in the field of
(MWSCAS), Aug. 2017, pp. 1597–1600.
advanced wireless communication systems (SAR)
[33] A. Rehmer and A. Kroll, ‘‘On the vanishing and exploding gradient
problem in gated recurrent units,’’ IFAC-PapersOnLine, vol. 53, no. 2,
from CentraleSupélec, France, in 2013, and
pp. 1243–1248, 2020. the Ph.D. degree in electrical engineering from
[34] A. K. Gizzini, M. Chafii, A. Nimr, and G. Fettweis, ‘‘Enhancing least CentraleSupélec, in 2016. Between 2014 and
square channel estimation using deep learning,’’ in Proc. IEEE 91st Veh. 2016, she has been a visiting Researcher at
Technol. Conf. (VTC-Spring), May 2020, pp. 1–5. the Poznan University of Technology (Poland),
[35] F. J. Pontes, G. F. Amorim, P. P. Balestrassi, A. P. Paiva, and University of York (UK), Yokohama National
J. R. Ferreira, ‘‘Design of experiments and focused grid search for neural University (Japan), and University of Oxford
network parameter optimization,’’ Neurocomputing, vol. 186, pp. 22–34, (U.K.). She joined the Technical University of Dresden, Germany, in
Apr. 2016. 2018, as a Research Group Leader, and ENSEA, France, in 2019, as an
[36] S. Jaeckel, L. Raschkowski, K. Börner, and L. Thiele, ‘‘QuaDRiGa: A 3-D Associate Professor, where she held a Chair of Excellence on Artificial
multi-cell channel model with time evolution for enabling virtual field Intelligence from CY Initiative. Since September 2021, she has been an
trials,’’ IEEE Trans. Antennas Propag., vol. 62, no. 6, pp. 3242–3256, Associate Professor at New York University (NYU), Abu Dhabi, and NYU
Jun. 2014. WIRELESS, NYU Tandon School of Engineering. Her research interests
[37] A. Abdelgader and L. Wu, ‘‘The physical layer of the IEEE 802.11p include advanced waveform design, integrated sensing and communication,
WAVE communication standard: The specifications and challenges,’’ in and machine learning for wireless communications. She received the IEEE
Proc. World Congr. Eng. Comput. Sci., in Lecture Notes in Engineering
ComSoc Best Young Researcher Award for Europe Middle East and Africa
and Computer Science, vol. 2, Oct. 2014, pp. 22–24.
(EMEA) region, the prize of the best Ph.D. in France in the fields of Signal,
[38] S. Jaeckel, L. Raschkowski, K. Börner, L. Thiele, F. Burkhardt, and
E. Eberlein, ‘‘QuaDRiGa-quasi deterministic radio channel generator, user Image and Vision, and she has been nominated in the top 10 Rising Stars
manual and documentation,’’ Fraunhofer Heinrich Hertz Inst., Berlin, in Computer Networking and Communications by N2Women in 2020. She
Germany, Tech. Rep., Version v1.2.3-307, 2017. served as an Associate Editor at IEEE Communications Letters 2019 to
[39] V. Sovrasov. (2018–2023). Ptflops: A Flops Counting Tool for Neural 2021, where she received the Best Editor Award in 2020. Between 2018
Networks in PyTorch Framework. [Online]. Available: https://github.com/ and 2021, she was the Research Lead of the Women in AI organization. She
sovrasov/flops-counter.pytorch is currently an Associate Editor at IEEE Transactions on Communications,
[40] A. K. Gizzini, Y. Medjahdi, A. J. Ghandour, and L. Clavier, ‘‘Towards serving as Vice-Chair of the IEEE ComSoc ETI on Machine Learning for
explainable AI for channel estimation in wireless communications,’’ 2023, Communications and leading the Education working group of the IEEE
arXiv:2307.00952. ComSoc ETI on Integrated Sensing and Communications.

18 VOLUME 2, 2024