Efficient MIMO Detection with Imperfect Channel

Knowledge - A Deep Learning Approach

Qian Chen† , Shunqing Zhang† , Shugong Xu† and Shan Cao†
† Shanghai Institute for Advanced Communication and Data Science,

Key laboratory of Specialty Fiber Optics and Optical Access Networks,

Shanghai University, Shanghai, 200444, China
Email:{chenqian, shunqing, shugong, cshan}@shu.edu.cn
arXiv:1903.07831v1 [eess.SP] 19 Mar 2019

Abstract—Multiple-input multiple-output (MIMO) system is authors in [9] proposed DetNet and show that the gradient
the key technology for long term evolution (LTE) and 5G. projection based detection algorithm can be well approximated
The information detection problem at the receiver side is in by a single layer neural network and the corresponding de-
general difficult due to the imbalance of decoding complexity and
decoding accuracy within conventional methods. Hence, a deep coding performance is computationally inexpensive with near
learning based efficient MIMO detection approach is proposed semidefinite relaxation (SDR) detection accuracy. Although
in this paper. In our work, we use a neural network to directly the above machine learning based approaches show promising
get a mapping function of received signals, channel matrix gain over the traditional signal detection methods, the follow-
and transmitted bit streams. Then, we compare the end-to-end ing issues in the conventional MIMO detectors have not been
approach using deep learning with the conventional methods in
possession of perfect channel knowledge and imperfect channel addressed based on our current investigation.
knowledge. Simulation results show that our method presents a • Robust Detection with Imperfect Channel Knowledge One
better trade-off in the performance for accuracy versus decoding
complexity. At the same time, better robustness can be achieved of the key issues in the MIMO detection design lies
in condition of imperfect channel knowledge compared with in the imperfectness of channel knowledge due to the
conventional algorithms. practical channel estimation method, the time-varying
Index Terms—MIMO detection, imperfect channel estimation, nature of communication devices and limited number of
deep learning reference signals [3]. Since the deep learning framework
has powerful generalization ability with respect to the
I. I NTRODUCTION input datasets, a straight forward question is whether we
MIMO systems [1], proposed by Telatar 20 years ago, have can rely on this generalization ability to handle the robust
been selected as one of the key features in the current LTE detection problem with imperfect channel knowledge.
systems and the massive deployment of MIMO systems [5] • Efficient Deep Learning Framework for Detection An-
are often regarded as a breakthrough for 5G systems. Although other issue that has not been solved is the efficient
the transmit side processing for MIMO transmission is often deep learning network architecture for the MIMO de-
straight forward, the information detection problem at the tection. Although the auto-encoder based solution has
receiver side is in general difficult [14]. been proposed in [10] to jointly combine the equalizer
and demodulator, the associated processing complexity
Traditional MIMO detection methods include linear and
is still significant if compared with MMSE or ZF based
non-linear approaches. For example, linear MIMO detector
detection. Therefore, an efficient learning framework with
includes zero forcing (ZF) [3] or minimum mean square error
only receiver side knowledge will be desirable.
(MMSE) [4] equalizers, while non-linear MIMO detector often
rely on minimum distance based maximum likelihood (ML) In this paper, we exploit the generalization capability of neu-
detection [2]. Although the ML-type decoding provides opti- ral networks to address the robust MIMO detection problem as
mal detection performance in theory, the associated decoding illustrated before. To be more specific, we test convolutional
complexity is in general unaffordable with current technology neural network (CNN) and deep neural network (DNN) based
[15]. In order to provide a better tradeoff, sphere decoding approach to model the non-linear transfer function between
[6] and successive interference cancellation [8] [7] are often MIMO transmitters and receivers, the result show DNN-based
applied in practice. approach has better BER performance and lower complexity
Recently, with the development of machine learning, more then CNN-based method. Through numerical examples, we
complicated problems in wireless communications can be show that the proposed DNN-based method outperforms the
formulated and efficiently solved by this framework. In the conventional ZF or MMSE based schemes by using the neural
physical layer process, for instance, [16] proposes an iterative network to directly get a mapping from end to end of the
brief propagation - convolutional neural network architecture bit stream. Meanwhile, DNN-based solution provides a better
for channel decoding, and variational auto-encoder has been tradeoff between the decoding performance and the decoding
applied for blind channel equalization [17]. Moreover, the complexity, which may pave the way for future efficient
 Channel Estimation

bit stream bit stream

Modulation Demodulation

bit stream bit stream

Modulation MIMO Demodulation
Channel
ZF/MMSE/ML

Detection
bit stream bit stream
Modulation Demodulation

bit stream bit stream

Modulation Demodulation

Figure 1. An overview of MIMO system. In this system, the bit streams are sent to the transmitting antenna after modulation. Then, they are transmitted
through MIMO channel which contains interference and are finally received by the receiving antenna.

detection design. distributed (i.i.d.) complex Gaussian distribution with zero

The rest of this paper is organized as follows. In Section II, mean and scaled identity covariance matrix [11], i.e., ∆H ∼
we briefly introduce the background of MIMO detection prob- CN(0, σe2 INr ), where σe 2 = NpN·E
t
p
with Np and Ep represent-
lem and formulate the corresponding mathematical problems ing the number and the power of pilot symbols respectively.
in Section III. The deep learning method based MIMO detec- With the imperfect CSI H b and the observed symbols y(t),
tion framework is discussed in Section IV and the numerical the detection process at the receiver side can be modeled
results are shown in Section V. Finally, concluding remarks through,
are given in Section VI.  
x̂(t) = arg min g y(t), Hb (3)
x(t)
II. S YSTEM M ODEL
where x̂ is selected from all the possible modulated symbols.
Consider a downlink MIMO transmission as shown in
By applying the inverse of modulation function, the detected
Fig. 1, where Nt transmit antennas are delivering messages to
bits for time slot t is given by,
Nr antennas using vertical Bell Labs layered space-time (V-  
BLAST) technique [13]. Given the time slot t and the modula- b̂(t) = f −1 x̂(t), M (4)
tion size 2M , a binary information stream b(t) ∈ B(M ×Nt )×1
are modulated through function f (·, M ) and the modulated The following assumption is made through the rest of this
symbols, x(t) ∈ CNt ×1 , are given by x(t) = f b(t), M .

paper. We consider block channel fading assumption, where
Denote y(t) ∈ CNr ×1 to be the received symbols and the the channel condition H remain static within each period T
mathematical model for y(t) are given by, and varies independently among different fading periods.
III. P ROBLEM F ORMULATION AND T RADITIONAL
y(t) = Hx(t) + n(t) (1)
A PPROACH
where H ∈ CNr ×Nt denotes the flat Rayleigh fading coeffi- In this section, we formulate the bit error rate (BER)
cients among transmit and receive antenna pairs and n(t) ∈ minimization through a general optimization framework. With
CNr ×1 is the additive white Gaussian noise with zero mean the above mathematical models of the decoding process, we
and unit variance, i.e. n(t) ∼ CN(0, INr ). can write the BER minimization problem as follows.
In the practical systems, due to the limited power and Problem 1 (BER Minimization):
resources for pilot symbols, the estimated channel state infor- N nT
mation (CSI) at the receiver side can not be perfect in general. 1 X X k b̂(t) − b(t) k2
minimize lim (5)
For illustration purpose, we denote Hb to be the imperfect CSI b̂(t) N →∞ N
n=1
T × M × Nt
t=(n−1)T +1
at the receiver side and the mathematical expression is given

subject to x(t) = f b(t), M , (6)
by [11],
y(t) = Hn x(t) + n(t),
H
b = H + ∆H (2) ∀t ∈ [(n − 1)T + 1, nT ] , (7)
n o
where ∆H denotes the estimation errors. If maximum like- Hb n = Hn + ∆H, H b = H bn , (8)
lihood (ML) based channel estimation mechanism is ap-  
plied, ∆H can be modeled through independent identical b̂(t) = h y(t), H
b . (9)
where h(·) denotes a combined procedure of symbol detection provide outstanding generalization capability, the BER perfor-
(3) and demodulation (4), and we add subscripts n to the mance under imperfect channel conditions can be improved
channel conditions H b and H to represent the nth fading with the deep learning framework. As problem 1 defines
period. an end-to-end BER evaluation framework under imperfect
To directly solve Problem 1 is in general challenging, since channel knowledge with binary decision of b̂(t), it belongs
the above formulation contains nonlinear functions, such as to the classification problem in the machine learning area [10]
f (·) and h(·). In addition, the searching space of b̂(t) is and the loss function that commonly adopted is the cross-
discrete and the brute force approach to obtain the optimal entropy between the estimation b̂(t) and the original bits b(t),
solution requires exponential complexity. To address this chal- i.e.,
lenge, conventional schemes reformulate Problem 1 as the N nT h  i
following symbol error rate (SER) minimization problem1 .
P P
CE b̂(t), b(t)
Problem 2 (SER Minimization): n=1 t=(n−1)T +1
L=− , (19)
N nT
M × Nt × N T
X X S (x(t) − x̂(t))
minimize lim (10) where CE(a, b) is defined to be a ln b + (1 − a) ln(1 − b).
x̂(t) N →∞
n=1 t=(n−1)T +1
N × T × Nt
A. Data Preprocessing
subject to y(t) = Hn x(t) + n(t),
We use the MIMO channel model we have mentioned in
∀t ∈ [(n − 1)T + 1, nT ] , (11)
n o Section II and get the complex time-domain vectors of both
Hb n = Hn + ∆H, H b = H bn , (12) y(t) and Ĥ, then convert them into real domain vectors. The
  details of the transition can be found in the following equation:
x̂(t) = arg min g y(t), Hb . (13)
x̂
   
<(y(t)) <(x(t))
ẏ(t) = , ẋ(t) = ,
where S(·) calculates the number of non-zero elements in the =(y(t)) =(x(t))
inner vector. 
<(Ĥ) −=(Ĥ)

By configuring different objective functions of g(·), tradi- Ḣ = . (20)
=(Ĥ) <(Ĥ)
tional detection algorithms minimize the SER results through
maximizing the likelihood function [18], minimizing the MSE where <(·) and =(·) are the real and imaginary part of
[4] or zero-forcing the noise values [11], where the mathemat- complex vector respectively.
ical expressions are2 ,
B. Network Architecture
x̂(t)M L = arg min k y(t) − Hx(t)
b k2 , (14) To find a network structure suitable for MIMO detection,
x(t)
bH −1 bH we have tested two neural networks, CNN and DNN. DNN is
x̂(t)ZF = (H H)
b H y(t), (15)
often used for joint channel equalization and decoding, while
x̂(t)M M SE = (H H + σ 2 IN )−1 H
b Hb b H y(t).
n t (16) CNN is widely applied to do feature extraction and process
With the imperfect CSI (ICSI) model, the received signal correlated noise [19]. The layout of these two networks we
vector y(t) can be written as, used are given in the following Table I. For DNN, there are
four hidden layers, besides the output layer, input layer and a
b − ∆H)x(t) + n(t) = Ĥx(t) + n̄(t).
y(t) = (H (17) Batch Normalization layer used to accelerate convergence and
prevent overfitting. And for CNN, two convolutional layers are
Define n̄(t) , n(t) − ∆Hx(t) to be the equivalent noise with
followed by two dense layers.
the channel estimation errors, and the equivalent noise variance
The architectures of DNN and CNN model are shown in
is given by,
Fig. 2 and Fig. 3 respectively. The first parameter below each
2 1 convolutional layer represents the number of filters in that
Tr E n̄(t)n̄(t)H = Nt σe2 + σn2 ,
  
σn̄(t) = (18)
Nr layer, while the second and third numbers show the size of
where Tr(·) denotes the matrix trace operation. Since the each filter. For the two dense layers, there are 128, 64 neurons.
equivalent noise variance scales with the estimation error σe2 ,
the post-processing Signal-to-Noise Ratio
(SNR) as well as • Input Layer: Joint matrix of ẏ(t) and Ḣ which have been
the BER scales with the order O (σe2 )−1 . converted to real domain from complex domain.
• Hidden Layers: Deploy  neural network to estimate the
IV. N EURAL N ETWORK F RAME
mapping function: h y(t), H b and the activation func-
Due to the non-convex approximation capability provided tion is rectified linear unit (ReLU).
by neural networks, we can directly solve problem 1 with • Output Layer: b̂(t), and the activation function is sig-
sufficient training data. In addition, since the neural networks moid, because we can take this problem as a classification
1 By assuming Gray coding, BER can be well approximated by a linear problem.
scaling of SER [2], and therefore, Problem 1 and Problem 2 are equivalent. • Optimizer: We choose to use adam as the optimizer which
2 (·)H denotes the matrix transpose and conjugation operation. is widely used in deep learning.
 Table I Table II
A N OVERVIEW OF N ETWORK C ONFIGURATIONS AND PARAMETERS . S IMULATION PARAMETERS .

DNN CNN parameters

Input Layer 4*5/8*9 8*9*1 computer setup Intel core i5-6500 CPU @3.20GHz with a 12GB memory
Layer1 Dense 512-ReLU 72*4*5-ReLU training set 5.4 × 105 symbols
Layer2 BatchNormalization BatchNormalization test set 7.2 × 105 symbols
Layer3 Dense 256-ReLU 128*2*3-ReLU validation set 1.8 × 105 symbols
Layer4 Dense 128-ReLU Dense 96-ReLU
Layer5 Dense 64-ReLU Dense 32-ReLU
Output Layer 4-way sigmoid 4-way sigmoid
detect 7.2 × 105 symbols needed by these two neural network.
On the one hand, it took DNN 15 seconds to finish the
Total Parameters 215372 208012
detection. On the other hand, CNN took 58 seconds, which
is nearly 4 times that of DNN. In summary, DNN is more
suitable for MIMO detection in term of BER performance and
decoding rate.

-1
B E R v s S N R in 4 x 4 M IM O w ith B P S K m o d u la tio n
1 0
D N N
C N N
...
...

...

...

...
...

-2
1 0

Input Output
data data

B E R
Dense1 Dense2 Dense3 Dense4 1 0 -3

512 256 128 64

Figure 2. Architecture of the DNN model used in MIMO detection.

-4
1 0
8 9 1 0 1 1 1 2 1 3
E b /N o (d B )
• Loss Function: Our problem is a discrete Ont-Hot vector
of a classification problem. So we used cross entropy as Figure 4. The BER performance of CNN versus DNN with perfect CSI. In
this experiment, DNN-based detection method performs much better CNN-
our loss function. based method.
C. Detection Performance of DNN versus CNN
In order to test which network is more suitable for MIMO V. S IMULATION R ESULTS
detection, we apply these two networks with a close number of We have compared the DNN-based method with ZF (Base-
total parameters to fairly perform 4 × 4 MIMO detection with line 1), MMSE (Baseline 2), DetNet [9] (Baseline 3) and
BPSK modulation. The details of the experiment are given ML-based detection algorithm (Baseline 4) in term of BER
in Section V. The result is shown in Fig. 4. It proves that performance, robustness and decoding rate. The parameters of
DNN has better BER performance than CNN. We believe this experiments are shown in Table II. Besides, we consider the
is on account of the strong ability of DNN to process one- spatial correlation mentioned in document [12]. We assume
dimensional data. Besides, we have measured the run time to that the fading of the receiver is spatially uncorrelated but
transmit-correlated, which is typical in the downlink channel
of a mobile communication system.
full connections
feature maps
A. BER Performance with Perfect CSI
feature maps Output data
In this section, DNN-based approach is compared with
Input data DetNet and traditional methods, such as ZF and MMSE under
different SNRs with full knowledge of the CSI (Np · Ep →
+∞, i.e. Ĥ = H). In all the simulations, our networks are
Conv1 Dense2
32
trained on data sets generated at 8dB SNR. The performance
72 4×5
Conv2
Dense1
for 4 × 4 MIMO detection with BPSK modulation is evaluated
128 2×3
96 and the result is given in Fig. 5. As we can see from the
figure, the BERs of DNN-based method are lower than MMSE
Figure 3. Architecture of the CNN model used in MIMO detection. and ZF at various SNRs, and even has an advantage of
more than 5dB. So our network has a strong capability of -1
B E R v s S N R in 2 x 2 M IM O w ith Q P S K m o d u la tio n
1 0
generalization. Besides, for BER of 10−3 , the DNN-based B a s e
B a s e
lin
lin
e 1 -C S I
e 2 -C S I
B a s e
B a s e
lin
lin
e 1 -IC S I
e 2 -IC S I

method outperforms DetNet in our channel model for 4.5 dB. B a s e

D N N
lin
-C S I
e 4 -C S I B a s e
D N N
lin
-IC
e 4 -IC S I
S I

B E R v s S N R in 4 x 4 M IM O w ith B P S K m o d u la tio n -2
1 0 -1 1 0
B a s e lin e 1 -C S I
B a s e lin e 2 -C S I
B a s e lin e 3 [1 3 ]-C S I

B E R
D N N -C S I

-3
-2 1 0
1 0
B E R

-4
-3
1 0
1 0 8 9 1 0 1 1 1 2 1 3
E b /N o (d B )

(a) BER versus SNR in 2 MIMO with QPSK modulation.

1 0 -4 B E R v s S N R in 4 x 4 M IM O w ith B P S K m o d u la tio n
7 8 9 1 0 1 1 1 2 1 3 1 4 1 5 1 6 1 7 B a s e lin e 1 -C S I B a s e lin e 1 -IC S I
B a s e lin e 2 -C S I B a s e lin e 2 -IC S I
E b /N o (d B ) 1 0 -1
B a s
B a s
e lin
e lin
e 3 [1
e 4 -C
3 ]-C S I
S I
B a s
B a s
e lin
e lin
e 3 [1 3
e 4 -IC
]-IC S I
S I
D N N -C S I D N N -IC S I

Figure 5. BER versus SNR with perfect CSI. In this experiment, DNN-based -2
1 0
detection method outperforms ZF, MMSE and DetNet.
-3
1 0

B. Robustness with Imperfect CSI

B E R
-4
1 0
The proposed method is compared with traditional methods,
including ML with imperfect channel information in this 1 0 -5

Nt
section. We set Np · Ep = 400, i.e., ∆H ∼ CN(0, 400 INr ). -6
1 0
The result for 2 × 2 MIMO detection with QPSK modulation
and 4 × 4 MIMO detection with BPSK modulation are shown 8 9 1 0 1 1 1 2 1 3
in Fig. 6. We noticed that with imperfect channel knowledge, E b /N o (d B )
for BER of 10−2 , the proposed DNN-based method still
(b) BER versus SNR in 4 MIMO with BPSK modulation.
outperforms the MMSE and ZF methods for about 4.5 dB and
more than 5dB when performing 2 × 2 MIMO detection with Figure 6. BER versus SNR with perfect and imperfect channel information.
QPSK. Particularly, for 4 × 4 MIMO detection with BPSK, In this experiment, DNN-based detection method still outperforms MMSE,
ZF and DetNet.
DNN-based approach outperforms the DetNet for about 3.5
dB for BER of 2 × 10−3 , ZF and MMSE for more than 4
dB respectively. This indicates that the proposed method still detection algorithms in Python 3.5.2 using an Intel core i5-
has a good robustness with imperfect CSI. Moreover, we can 6500 CPU @3.20GHz with a 12GB memory. We run the test
notice that the BER of the ML method in the case of imperfect to detect 7.2 × 105 symbols and record the time consumption
CSI are significantly higher compared with that of perfect CSI. for 3 times, then calculate the average throughput. In Table
But the fluctuation of DNN caused by the perfection of CSI III, we compare the detection efficiency of different schemes,
is similar to other methods. It is worth mentioning that the including ZF, MMSE, ML and the proposed DNN-based
detecting result of DNN-based method using imperfect CSI detection method for 4 × 4 MIMO detection with BPSK
is even better than that of DetNet MIMO detection method modulation. From Table III, we can see the throughput of ZF is
with perfect channel information. Although the BER of the highest, and the proposed DNN-based method has very similar
proposed method is higher than that of the ML method, using performance with ZF algorithm. MMSE followed and ML
the model for MIMO detection has much lower complexity detection method is lowest. We can conclude that DNN-based
than ML algorithm, which is enough to make up for the method has a near-ZF throughput performance and achieve
deficiency in BER. more accurate decoding performance at the same time, which
means that the time complexity of DNN-based method is also
C. Throughput Comparison among Different Method low.
The higher time complexity the algorithm has, the slower
the decoding rate is. To compare the time complexity of differ- VI. C ONCLUSION
ent detection algorithms, we calculate the average throughput In this paper, we compared two MIMO detection methods
of them, which equals to the number of detected bits divided based on deep learning. The result shows that DNN has
by the time consumption. We program and run all MIMO better BER performance and higher decoding rate than CNN.
 Table III
T HROUGHPUT C OMPARISON AMONG D IFFERENT S CHEMES

ZF MMSE ML DNN-based
Throughput (Kbps) 4.8196 × 104 4.6933 × 104 8.8377 × 103 4.8 × 104

We performed simulation experiments in the case of perfect [11] Cheng Li, Fan Jiang, and Zijun Gong. Accurate analytical ber perfor-
and imperfect channel information respectively, and compared mance for zf receivers under imperfect channel in low snr region for
large receiving antennas. IEEE Signal Processing Letters, 2018.
the performance of the proposed method with the traditional [12] HA Abou Saleh and Walaa Hamouda. Performance of zero-forcing
method and DetNet under different SNR conditions. In the detectors over mimo flat-correlated ricean fading channels. IET Com-
case of all SNRs, DNN-based method has a certain distance munications, 3(1):10–16, January 2009.
[13] Tan Wei, Chen Xiao-min, Zhu Yi-min, Zhu Qiu-ming, and Qian Ke-
from the ML method, but better than the other detection cang. Power allocation algorithm for v-blast system with channel
schemes. In addition, it has similar decoding rate with ZF correlation and estimation error. In 2014 3rd Asia-Pacific Conference
method, which far exceeds ML method. Besides, better BER on Antennas and Propagation (APCAP), pages 741–744, July 2014.
[14] Shaoshi Yang and Lajos Hanzo. Fifty years of mimo detection: The
performance and robustness can be achieved in the case of road to large-scale mimos. IEEE Communications Surveys & Tutorials,
imperfect channel information, compared with the ZF, MMSE 17(4):1941–1988, 2015.
methods and DetNet. Furthermore, the proposed network is [15] J-S Kim, S-H Moon, and Inkyu Lee. A new reduced complexity ml
detection scheme for mimo systems. IEEE Transactions on Communi-
easy to be implemented on hardware owing to its simpler cations, 58(4):1302 – 1310, April 2010.
network structure. In general, DNN-based method is a proper [16] Fei Liang, Cong Shen, and Feng Wu. An iterative bp-cnn architecture for
approach for MIMO detection. channel decoding. IEEE Journal of Selected Topics in Signal Processing,
12(1):144–159, 2018.
[17] Avi Caciularu and David Burshtein. Blind channel equalization using
ACKNOWLEDGEMENT variational autoencoders. arXiv preprint arXiv:1803.01526, 2018.
[18] Uğur Yeşilyurt and Özgür Ertuğ. Hybrid ml-mmse adaptive multiuser
This work was supported by the National Natural Science detection based on joint channel estimation in sdma-ofdm systems. In
Foundation of China (NSFC) Grants under No. 61701293, 2017 25th International Conference on, Software, Telecommunications
and Computer Networks (SoftCOM), pages 1–5, September 2017.
the National Science and Technology Major Project Grants [19] Yaohan Wang, Zhizhao Zhang, and Xu Shugong Zhang Shunqing,
under No. 2018ZX03001009, the Huawei Innovation Research Cao Shan. A unified deep learning based polar-ldpc decoder for 5g
Program (HIRP), and research funds from Shanghai Institute communication systems. In 2018 The Tenth International Conference
on Wireless Communications and Signal Processing (WCSP), pages 1–6,
for Advanced Communication and Data Science (SICS). Hangzhou, China, 2018. IEEE.

R EFERENCES
[1] Emre Telatar. Capacity of multi-antenna gaussian channels. European
Transactions on Telecommunications, 10(6):585–595, 1999.
[2] Xu Zhu and Ross D Murch. Performance analysis of maximum
likelihood detection in a mimo antenna system. IEEE Transactions on
Communications, 50(2):187–191, 2002.
[3] Cheng Wang, Edward KS Au, Ross D Murch, Wai Ho Mow, Roger S
Cheng, and Vincent Lau. On the performance of the mimo zero-forcing
receiver in the presence of channel estimation error. IEEE Transactions
on Wireless Communications, 6(3):805–810, 2007.
[4] Eren Eraslan, Babak Daneshrad, and Chung-Yu Lou. Performance
indicator for mimo mmse receivers in the presence of channel estimation
error. IEEE Wireless Communications Letters, 2(2):211–214, 2013.
[5] Thomas L Marzetta. Noncooperative cellular wireless with unlimited
numbers of base station antennas. IEEE Transactions on Wireless
Communications, 9(11):3590–3600, 2010.
[6] Zhan Guo and Peter Nilsson. Algorithm and implementation of the k-
best sphere decoding for mimo detection. IEEE Journal on selected
areas in communications, 24(3):491–503, 2006.
[7] Jin Xu, Xiaofeng Tao, and Ping Zhang. Analytical ser performance
bound of m-qam mimo system with zf-sic receiver. In IEEE Interna-
tional Conference on Communications (ICC), pages 5103–5107. IEEE,
May 2008.
[8] Sabuj Sarkar. An advanced detection technique in mimo-psk wireless
communication systems using mmse-sic detection over a rayleigh fading
channel. CSI Transactions on ICT, 5(1):9–15, 2017.
[9] Tzvi Diskin Neev Samuel and Ami Wiesel. Deep mimo detection. In
2017 IEEE 18th International Workshop on Signal Processing Advances
in Wireless Communications (SPAWC), pages 1–5, July 2017.
[10] Xin Yan, Fei Long, Jingshuai Wang, Na Fu, Weihua Ou, and Bin
Liu. Signal detection of mimo-ofdm system based on auto encoder
and extreme learning machine. In 2017 International Joint Conference
on Neural Networks (IJCNN), pages 1602–1606, May 2017.