Adaptive Modulation and Coding Using Neural Network Based SNR Estimation
Adaptive Modulation and Coding Using Neural Network Based SNR Estimation
Adaptive Modulation and Coding Using Neural Network Based SNR Estimation
net/publication/336275432
Adaptive Modulation and Coding Using Neural Network Based SNR Estimation
CITATIONS READS
8 150
3 authors:
Chang-Jun Ahn
Chiba University
208 PUBLICATIONS 666 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Kazuki Maruta on 06 April 2020.
Received August 15, 2019, accepted October 4, 2019, date of publication October 11, 2019, date of current version December 27, 2019.
Digital Object Identifier 10.1109/ACCESS.2019.2946973
ABSTRACT In this paper, we propose a novel Adaptive Modulation and Coding (AMC) scheme enabled by
Artificial Neural Network (ANN) aided Signal-to-Noise power Ratio (SNR) estimation. The Power Spectral
Density (PSD) values are trained for SNR classification and it is mapped to respective Modulation and
Coding Scheme (MCS) sets. Once trained, optimal MCS can be determined in low calculation complexity.
The proposed approach is robust especially in high mobility environment since the PSD appearance is hardly
influenced by the Doppler shift. Its effectiveness in terms of throughput is presented through computer
simulations compared to the existing Error Vector Magnitude (EVM) based link adaptation scheme.
INDEX TERMS SNR estimation, artificial neural network, adaptive modulation and coding.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
VOLUME 7, 2019 183545
S. Kojima et al.: Adaptive Modulation and Coding Using Neural Network Based SNR Estimation
computation resources and memory are limited or when the also be acceptable. Once the learning process is conducted,
strictly low processing latency is imposed. only the resultant optimized function should be installed to
The proposed method estimates SNR using ANN exploit- the transceiver. Therefore, the proposed AMC works in a
ing power spectral density (PSD) values of the received simplified manner and can be applied to small transceivers
signal. Since the estimator of the proposed method requires being driven with limited power. Furthermore, in the link
only PSD values, it can implement with a simple config- adaptation such as AMC, FBI for control is essential. The
uration. It is also a point that it is easy to acquire data computation complexity for SNR estimation and the amount
set of PSD value and teacher signal of SNR value. Thus, of FBI itself are factors of the feedback delay. The neural
this estimation method easily realizes a high accuracy and network aided proposed method can also alleviate it, since
low computational complexity by extracting features such these features, the proposed approach can contribute to the
as noise power from the PSD. In addition, the proposed massive connectivity and low latency applications such as
method is hardly affected by the Doppler shift since it focuses mission critical IoT and vehicle-to-vehicle communication.
on only the power-domain which excludes the phase offset
III. SYSTEM MODEL
impact. SNR estimation accuracy can be kept even in the high
A. CHANNEL MODEL
mobility environment. Its robustness and maximized through-
This section explains the time-varying multipath fading chan-
put performance are valuably disclosed through computer
nel that we assumed as the propagation model in this paper
simulations.
[28], [29]. In the case that the transmission bandwidth is
The reminder of this paper is organized as follows.
larger than the channel coherence bandwidth, the frequency
Section II reviews related works and validates our contribu-
components of the transmission signal exceeding the coherent
tion. The system model is described in Section III. Section IV
frequency bandwidth have different phase transitions and
introduces the proposed method that estimates SNR using
gains. This channel is called frequency selective channel [30].
ANN for AMC scheme. Section V presents the simulation
Its discrete expression can be written as follows:
results from various viewpoints. Section VI concludes this
K
paper. X k
h(t; τ ) = hk (t)δ(t − ) (1)
W
k=1
II. RELATED WORKS
K
At present, an application of machine learning to the wireless gk X
hk (t) = √ exp [j (2πfd t cosαk + φk )] (2)
communication field has attracted a great deal of attention K k=1
and research has been actively conducted [19]. Particularly,
where K is the number of discrete multipaths, and hk rep-
modulation recognition (MR) [20]–[25] and channel
resents the channel complex gain of the kth multipath com-
estimation by using a neural network [26], [27] have been
ponent. δ(t) denotes the Dirac’s delta function, W is the
extensively studied. Modulation recognition (MR) is a pio-
transmission signal bandwidth, gk indicates the k-th path
neering technology that fully exploits the potentiality of
amplitude. fd denotes the Doppler frequency, αk and φk indi-
ANN. MR aims at implementation of correct demodulation
cate the angle of arrival (AoA) of the kth incoming wave
in the receiver side and is an essential technique to improve
and its initial phase, respectively. We constructed 2atime-
the communication performance especially in the cognitive
varying channel based on Jakes’ model. K
P h = 1
k=1 E k
radio environment. Literature claims that MR is fundamental
where E [·] indicates the ensemble-average operation. Then,
for AMC, however, MR and AMC is substantially differ-
the frequency response H (t; f ) is obtained by the discrete
ent techniques. AMC requires information about channel
Fourier transform (DFT) of h(t; τ ) expressed as follows:
quality such as SNR at the receiver side and it is fed-back Z ∞
to the transmitter side. Our proposal introduces ANN to H (t; f ) = h(t; τ )exp (−j2πf τ ) dτ
accurately estimate SNR using PSD. MR is to improve the −∞
signal reception performance even at lower SNR situation, K
X k
but not to estimate SNR itself. On the other hand, the neu- = hk (t)exp −j2πf . (3)
W
ral network based channel estimation can also improve the k=1
estimation accuracy as well as communication performance. In regards to K , the following relationship holds,
The common objective of using neural networks among these K = bWTms c + 1, (4)
approaches including our proposal is to simplify the com-
where Tms represents the spread of multipath. From (4), since
plicated calculation processing for a high-speed and a high
K > 1, the frequency response of channel is not immutable
reliable communication. The proposed approach focused on
due to the variation of each channel gains. Such a fading envi-
frequency-domain real-valued signals which can be easily
ronment is called frequency selective. The Doppler power
obtained by Fourier transform of specified signal samples
spectra can be written as follows:
and can reduce the burden on the receiver. In terms of Z ∞
learning neural network, a data set of PSD can be acquired not 1 ∗
Dk (λ) = E hk (t)hk (t + τ ) · exp (−j2π λτ ) dτ,
only by signal processing but also commercial measurement −∞ 2
instruments such as a spectrum analyzer; captured images can for k = 1, 2, 3, . . . , K . (5)
As can be seen from the above equation, multipath can be C. RECEIVER STRUCTURE
modeled using Doppler spread and power spectra, and it is The receiver structure is shown in Fig.1(b). After remov-
greatly affected by Doppler frequency representing moving ing the GI and performing serial to parallel conversion,
speed of the transceiver. The influence of multipath phase the received signal that passed through the multipath fading
changes in OFDM signals that degrade channel tracking channel is expressed as follows:
performance depends on Doppler frequency. Thus, in the Z ∞
case of high Doppler frequency, channel estimation accuracy r(t) = h(t; τ )s(t − τ )dτ + z(t) (8)
−∞
deteriorates and detection performance also deteriorates.
where z(t) is the additive white Gaussian noise (AWGN)
B. TRANSMITTER STRUCTURE
which has power spectral density of Z0 . The received signal
after fast Fourier transform is given by
The transmitter structure is shown in Fig.1(a). The time
1 Ts +nT
Z
domain transmission signal is represented as follows: −j2πm(t − nT )
r(m, n) = r(t) exp dt
T Ts
∞
s ss nT
X 2P 2P X
Nc
1 Ts
Z
j2π(a − m)t
s(t) = · p(t − nT ) = d(a, n) · exp
n=−∞
Nc Nc Ts 0 Ts
"N a=1
c # Z ∞
X j2πm(t − nT )
· d(m, n) · exp (6) · h(τ, t + nT )p(t − τ )
Ts −∞
m=1
−j2πaτ
where P is the average transmission power, Nc denotes · exp dτ dt + z(m, n) (9)
Ts
the number of subcarriers and d(m, n) is the m-th subcar- where z(m, n) is AWGN having variance of 2Z0 /Ts with zero-
rier of the n-th modulated symbol. Here, d(m, n) satisfies mean. Suppose the maximum chromatic dispersion between
E [|d(m, n)|] = 1. T is the symbol length of OFDM signal, subcarriers is smaller than Tg , the integral with respect to τ
Ts indicates the effective symbol length of OFDM signal not can be expressed as
including guard interval (GI). It is assumed that GI length is
Nc Z ∞
Tg and Tg is a value that satisfies T = Ts + Tg . The frequency X −j2πaτ
h(τ, t + nT )p(t − τ ) exp dτ
interval between adjacent orthogonalized OFDM subcarriers −∞ Ts
a=1
is 1/Ts . p(t) is the transmission pulse, and expressed as Nc Z Ts
X −j2πaτ
= h(τ, t + nT ) exp dτ
(
1 −Tg ≤ t ≤ Ts 0 Ts
p(t) = (7) a=1
0 otherwise. = H (m, t + nT ) . (10)
where f is the activation function, and various differentiable FIGURE 4. The power spectrum of the received signal according to SNR
functions are used depending on the application. From (16), level.
it can be rewritten as follows:
yc = f c wc yc−1 + bc . (17) between the teacher signal and the output signal, so that the
The initial value of the input signal is the received signal given output signal approaches to the teacher signal.
by (12) and expressed as follows: The proposed method estimates SNR from the PSD values
of received signal as shown in Fig.4. From Fig.4(a), (c),
y0 = [r̃(1)T , r̃(2)T , . . . , r̃(n)T ]T , (18) the PSD under only AWGN channel has a large feature
according to SNR level. Thus, it is possible to estimate coarse
where r̃(n) indicates the n-th received symbol. Here, the num- SNR level from these spectrum image in human eyes. Mean-
ber of neurons in the first layer is equal to the number of the while, as shown in Fig.4(b), (d), the PSD under frequency
received signal. Assuming that the combination of the teacher selective fading channel shows intense fluctuation, and it
signal and the input signal is [t, y0 ], the output of the network seems difficult to extract the features in terms of SNR at a
is represented by glance. It is expected to be possible to determine the SNR
by extracting these complicated features using the neural
yC = f C wC yC−1 + bC . (19)
network. Furthermore, the feature quantities involved in SNR
Supposing that the cost function is a square error function, estimation that obtained by the PSD value are not affected by
the evaluation index of the network ξ is expressed as Doppler frequency, because the effect of Doppler frequency
appears as a rotation of the received signals’ phase com-
1 T
ξ = t − yC t − yC ponents; its effect can be excluded on the power spectrum.
2 Although the PSD value is the key point in the proposed
1
2
=
t − f C wC yC−1 + bC
. (20) method, acquisition of this value can be easily acquired via
2 simple signal processing or measurement instruments such as
In back propagation algorithm, the steepest descent method a spectrum analyzer. Captured images can also be acceptable
is generally used. The gradient of the weight and the bias are to learn the network. In this case, the burden on the receiving
expressed as follows, side can be significantly reduced.
∂ξ
1wcuv = −η (21) B. APPLICATION OF SNR ESTIMATION BY ANN TO AMC
∂wcu,v
∂ξ In the proposed method, the estimator learns in advance using
1bcu = −η c , (22) PSD values and corresponding SNR values. It enables sim-
∂bu plification of the signal processing for the operation period.
where η represents the learning rate. As described above, If channel environment was changed, newly trained neural
the weights and biases are sequentially updated from the error network can be simultaneously updated to a large number of
user terminals via software-based means. After that, substitut- of the proposed method is demonstrated with these basic
ing the PSD value of the received signal to the learned estima- parameters, other FEC codes, such as Turbo code, LDPC and
tor as an input signal, estimated SNR can be obtained. AMC Polar code, can also be effective as same.
parameters the determined according to SNR estimated value. The ANN structure used in the proposed method consists
The above simplified mechanism is expected to contribute to of C = 3 layers, an input layer, a hidden layer, and an
light-weight signal processing with lower latency. output layer, respectively. The hidden layer is composed of
55 neurons and the learning rate η is set to 0.01. 10000 sets of
PSD values and SNR values were used for each SNR as a data
set. Sigmoid function is employed as the activation function
given by
1
y = . (23)
1 + e−x
PSD data sets are generated using MATLAB. The train-
ing network is designed using MATLAB Deep Learning
Toolbox. Operating environment for training and testing the
network is NVIDIA Quadro GV100 GPU.
FIGURE 5. Frame structure.
B. SIMULATION RESULTS
V. COMPUTER SIMULATION Fig. 6 shows the accuracy of estimated SNR with various
A. SIMULATION PARAMETERS SNR intervals to be classified. When the interval is small,
The frame structure is based on wide spread OFDM as shown the accuracy of SNR estimation is low. For this reason, in the
in Fig.5. One frame consists of 2 pilot symbols and 30 data case of small interval, the spectra of adjacent SNR values
symbols where one OFDM symbol consists of 52 subcarriers are very similar. Therefore it is difficult to extract different
and 16 guard interval samples. These parameters are selected feature quantities by ANN; erroneous judgment increases.
based on the great majority of legacy wireless LAN standard. Meanwhile, increasing SNR interval to 3 dB, its accuracy
The duration of effective symbol is 3.2 µs and the guard inter- exceeds 80%. Thus, the proposed method can demonstrate
val duration is 0.8 µs. Table 2 summarizes detailed simulation sufficient AMC performance by setting the interval of SNR
parameters. The bandwidth of transmission signal is 20 MHz. threshold to 3 dB or more.
Rayleigh fading channel consists of 15 multipaths with the Fig. 7 compares estimated SNR and actual SNR for EVM.
interval of 50 ns. Max Doppler frequencies were set to 10 and The accuracy of the estimated value by EVM deteriorates as
100 Hz representing low and high mobilities, respectively. the Doppler frequency increases. Since the EVM estimates
As for the error correction code, the convolutional code with SNR from the difference between the received IQ symbols
rates of 1/2, 2/3, 3/4 and 5/6 was applied. If the effectiveness and the reference one, it is largely influenced by the Doppler
FIGURE 7. Estimated SNR by EVM versus actual SNR. FIGURE 9. Throughput performance at Doppler frequency 10 Hz.
ACKNOWLEDGMENT
This work was supported by the Grant of Scientific Research
No. 17K06415 from the Japan Society for the Promotion of
Science (JSPS).
REFERENCES
[1] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. Soong,
and J. C. Zhang, ‘‘What will 5G be?’’ IEEE J. Sel. Areas Commun., vol. 32,
no. 6, pp. 1065–1082, Jun. 2014. doi: 10.1109/JSAC.2014.2328098.
[2] F. Boccardi, R. W. Heath, Jr., A. Lozano, T. L. Marzetta, and P. Popovski,
‘‘Five disruptive technology directions for 5G,’’ IEEE Commun. Mag.,
vol. 52, no. 2, pp. 74–80, Feb. 2014. doi: 10.1109/MCOM.2014.6736746.
[3] J. Joung and A. H. Sayed, ‘‘Multiuser two-way amplify-and-forward relay
processing and power control methods for beamforming systems,’’ IEEE
Trans. Signal Process., vol. 58, no. 3, pp. 1833–1846, Mar. 2010. doi:
10.1109/TSP.2009.2038668.
[4] Z. Yun-Xiao, L. Yi-Fe, M. Shi-Bo, and J. Yue, ‘‘Power control algorithm
FIGURE 10. Throughput performance at Doppler frequency 100 Hz. based on SNR cost function in cognitive radio system,’’ in Proc. IEEE 18th
Int. Conf. Parallel Distrib. Syst., Singapore, Dec. 2012, pp. 913–917. doi:
10.1109/ICPADS.2012.150.
where Nsym is the number of symbols and Nsub is the number [5] M. A. Azza, M. El Yahyaoui, and A. El Moussati, ‘‘Throughput perfor-
mance of adaptive modulation and coding schemes for WPAN transceiver’’
of subcarriers. From (24), 0EVM is significantly affected by in Proc. Int. Symp. Adv. Elect. Commun. Technol. (ISAECT), Rabat,
the number of received symbols. On the other hand, the com- Morocco, Nov. 2018, pp. 1–4. doi: 10.1109/ISAECT.2018.8618853.
putational complexity by the proposed system is expressed as [6] C. Y. Wong, R. S. Cheng, K. B. Lataief, and R. D. Murch, ‘‘Multiuser
OFDM with adaptive subcarrier, bit, and power allocation,’’ IEEE J.
follows, Sel. Areas Commun., vol. 17, no. 10, pp. 1747–1758, Oct. 1999. doi:
C−1 10.1109/49.793310.
0prop = Nuc × Nvc + 1 , (25) [7] Q. Liu, S. Zhou, and G. B. Giannakis, ‘‘Cross-layer combining of adaptive
modulation and coding with truncated ARQ over wireless links,’’ IEEE
where C indicates the number of layers, Nuc and Nvc represent Trans. Wireless Commun., vol. 3, no. 5, pp. 1746–1755, Sep. 2004. doi:
10.1109/TWC.2004.833474.
the number of neuron in the c-th layer and the (c − 1)-th [8] X. Qiu and K. Chawla, ‘‘On the performance of adaptive modulation in
layer, respectively. From (25), 0prop is greatly affected by the cellular systems,’’ IEEE Trans. Commun., vol. 47, no. 6, pp. 884–895,
number of layers. In the proposed system, C = 3 is sufficient Jun. 1999. doi: 10.1109/26.771345.
for highly accurate SNR estimation, therefore the calculation [9] R. Zeng, T. Liu, X. Yu, and Z. Zhang, ‘‘Novel channel quality indicator
prediction scheme for adaptive modulation and coding in high mobil-
complexity is lower than the EVM method. Substituting the ity environments,’’ IEEE Access, vol. 7, pp. 11543–11553, 2019. doi:
parameters shown in Table2 results in 0EVM = 228, 556, 830 10.1109/ACCESS.2019.2892228.
and 0prop = 5, 808, 056. Thus the proposed method can [10] S. Kojima, K. Maruta, and C. J. Ahn, ‘‘Throughput maximization by
adaptive switching with modulation coding scheme and frequency sym-
reduce significantly the computational complexity by 97.5%. bol spreading,’’ J. Commun. Softw. Syst., vol. 14, no. 4, pp. 332–339,
The above assessment revealed overall advantage of the Nov. 2018. doi: 10.24138/jcomss.v14i4.616.
proposed AMC method based on ANN, which can achieve [11] R. A. Shafik, M. S. Rahman, and A. R. Islam, ‘‘On the extended relation-
ships among EVM, BER and SNR as performance metrics,’’ in Proc. Int.
improved throughput performance as well as reduced com- Conf. Elect. Comput. Eng., Dhaka, Bangladesh, Dec. 2006, pp. 408–411.
putation complexity. doi: 10.1109/ICECE.2006.355657.
[12] S. Chang and B. Kelly, ‘‘A time-domain SNR estimator based on a periodic
preamble for wireless OFDM systems,’’ IEICE Electron. Express, vol. 8,
VI. CONCLUSION no. 24, pp. 2073–2079, Dec. 2011.
This paper proposed the novel SNR estimation method by [13] T. J. O’Shea, T. Erpek, and T. C. Clancy, ‘‘Physical layer deep learning
using ANN in AMC to improve the throughput performance of encodings for the MIMO fading channel,’’ in Proc. 55th Annu. Allerton
Conf. Commun., Control, Comput. (Allerton), Oct. 2017, pp. 76–80. doi:
in high mobility environment. The SNR estimation function 10.1109/ALLERTON.2017.8262721.
was learned by using power spectral density values in a [14] J. Kassab and S. Nagaraj, ‘‘Adaptive modulation in an OFDM com-
supervised manner. Once estimation function was learned, munications system with artificial neural networks,’’ in Proc. Int. Joint
Conf. Neural Netw., Atlanta, GA, USA, Jun. 2009, pp. 1547–1551. doi:
it can be installed to the communication terminals; required 10.1109/IJCNN.2009.5178657.
operation is only SNR classification based on the obtained [15] L. O. Chua and L. Yang, ‘‘Cellular neural networks: Theory,’’ IEEE
PSD values. It can alleviate the computation burden on the Trans. Circuits Syst., vol. CAS-35, no. 10, pp. 1257–1272, Oct. 1988. doi:
receiver side. Simulation results revealed that the proposed 10.1109/31.7600.
[16] L. O. Chua and L. Yang, ‘‘Cellular neural networks: Applications,’’ IEEE
SNR estimation method has high estimation accuracy being Trans. Circuits Syst., vol. 35, no. 10, pp. 1273–1290, Oct. 1988. doi:
hardly influenced by the Doppler shift. Furthermore, we 10.1109/31.7601.
[17] M. Riedmiller and H. Braun, ‘‘A direct adaptive method for faster back- [37] M. T. Hagan and M. B. Menhaj, ‘‘Training feedforward networks with
propagation learning: The RPROP algorithm,’’ in Proc. IEEE Int. Conf. the marquardt algorithm,’’ IEEE Trans. Neural Netw., vol. 5, no. 6,
Neural Netw., San Francisco, CA, USA, Apr. 1993, pp. 586–591. doi: pp. 989–993, Nov. 1994. doi: 10.1109/72.329697.
10.1109/ICNN.1993.298623. [38] P. J. Werbos, ‘‘Backpropagation through time: What it does and how
[18] Y.-J. Liu and S. Tong, ‘‘Adaptive NN tracking control of uncer- to do it,’’ Proc. IEEE, vol. 78, no. 10, pp. 1550–1560, Oct. 1990. doi:
tain nonlinear discrete-time systems with nonaffine dead-zone input,’’ 10.1109/5.58337.
IEEE Trans. Cybern., vol. 45, no. 3, pp. 497–505, Mar. 2015. doi: [39] F. R. Kschischang, B. J. Frey, and H.-A. Loeliger, ‘‘Factor graphs and
10.1109/TCYB.2014.2329495. the sum-product algorithm,’’ IEEE Trans. Inf. Theory, vol. 47, no. 2,
[19] Z. Qin, H. Ye, G. Li, and B. Juang, ‘‘Deep learning in physical layer pp. 498–519, Feb. 2001. doi: 10.1109/18.910572.
communications,’’ IEEE Wireless Commun., vol. 26, no. 2, pp. 93–99,
Apr. 2019. doi: 10.1109/MWC.2019.1800601.
[20] L. Guesmi and M. Menif, ‘‘Modulation formats recognition technique
using artificial neural networks for radio over fiber systems,’’ in Proc. SHUN KOJIMA received the B.E. and M.E.
17th Int. Conf. Transp. Opt. Netw. (ICTON), Budapest, Hungary, Jul. 2015, degrees in electrical and electronics engineering
pp. 1–4. doi: 10.1109/ICTON.2015.7193508. from Chiba University, Japan, in 2017 and 2018,
[21] A. Hazza, M. Shoaib, S. A. Alshebeili, and A. Fahad, ‘‘An overview of respectively, where he is currently pursuing the
feature-based methods for digital modulation classification,’’ in Proc. 1st Ph.D. degree with the Graduate School of Engi-
Int. Conf. Commun., Signal Process., Appl. (ICCSPA), Sharjah, United neering. His research interests include MIMO,
Arab Emirates, Feb. 2013, pp. 1–6. doi: 10.1109/ICCSPA.2013.6487244. cooperative communications, adaptive modulation
[22] B. Tang, Y. Tu, Z. Zhang, and Y. Lin, ‘‘Digital signal modulation clas- and coding, neural networks, and channel esti-
sification with data augmentation using generative adversarial nets in mation. He received the Best Paper Award from
cognitive radio networks,’’ IEEE Access, vol. 6, pp. 15713–15722, 2018. the 26th International Conference on Software,
doi: 10.1109/ACCESS.2018.2815741. Telecommunications and Computer Networks (SoftCOM2018).
[23] Y. Wang, M. Liu, J. Yang, and G. Gui, ‘‘Data-driven deep learning for auto-
matic modulation recognition in cognitive radios,’’ IEEE Trans. Veh. Tech.,
vol. 68, no. 4, pp. 4074–4077, Apr. 2019. doi: 10.1109/TVT.2019.2900460.
[24] S. Zheng, P. Qi, S. Chen, and X. Yang, ‘‘Fusion methods for KAZUKI MARUTA received the B.E., M.E., and
CNN-based automatic modulation classification,’’ IEEE Access, vol. 7, Ph.D. degrees in engineering from Kyushu Univer-
pp. 66496–66504, 2019. doi: 10.1109/ACCESS.2019.2918136. sity, Japan, in 2006, 2008, and 2016, respectively.
[25] S. Peng, H. Jiang, H. Wang, H. Alwageed, Y. Zhou, M. Sebdani, and From 2008 to 2017, he was with NTT Access
Y. Yao, ‘‘Modulation classification based on signal constellation diagrams Network Service Systems Laboratories and was
and deep learning,’’ IEEE Trans. Neural Netw. Learn. Syst., vol. 30, no. 3, involved in the research and development of inter-
pp. 718–727, Mar. 2019. doi: 10.1109/TNNLS.2018.2850703. ference compensation techniques for future wire-
[26] N. Taşpmar and M. N. Seyman, ‘‘Back propagation neural network less communication systems. He is currently an
approach for channel estimation in OFDM system,’’ in Proc. IEEE Int. Assistant Professor with the Graduate School of
Conf. Wireless Commun., Netw. Inf. Secur., Jun. 2010, pp. 265–268. doi: Engineering, Chiba University. His research inter-
10.1109/WCINS.2010.5541934. ests include OFDM, MIMO, adaptive array signal processing, channel esti-
[27] K. Hiray and K. V. Babu, ‘‘A neural network based channel estimation mation, and medium access control protocols. He is a member of the Institute
scheme for OFDM system,’’ in Proc. Int. Conf. Commun. Signal Pro- of Electronics, Information and Communication Engineers (IEICE) and
cess. (ICCSP), Melmaruvathur, India, Apr. 2016, pp. 0438–0441. doi: the Institute of Electrical Engineers of Japan (IEEJ). He received IEICE
10.1109/ICCSP.2016.7754174. Young Researcher’s Award, in 2012, IEICE Radio Communication Systems
[28] E. Biglieri, J. Proakis, and S. Shannai, ‘‘Fading channels: Information- (RCS) Active Researcher Award, in 2014, Asia-Pacific Microwave Confer-
theoretic and communications aspects,’’ IEEE Trans. Inf. Theory, vol. 44, ence (APMC) 2014 Prize, and IEICE RCS Outstanding Researcher Award,
no. 6, pp. 2619–2692, Oct. 1998. doi: 10.1109/18.720551.
in 2018. He was a co-recipient of the IEICE Best Paper Award in 2018.
[29] L. Cimini, ‘‘Analysis and simulation of a digital mobile channel using
orthogonal frequency division multiplexing,’’ IEEE Trans. Commun.,
vol. 33, no. 7, pp. 665–675, Jul. 1985. doi: 10.1109/TCOM.1985.1096357.
[30] C.-J. Ahn, S. Takahashi, H. Fujisaka, T. Kamio, and K. Haeiwa, ‘‘Power CHANG-JUN AHN received the Ph.D. degree
consumption for coherent optical orthogonal frequency division multiplex- from the Department of Information and Computer
ing with punctured LDPC codes and variable amplitude block codes,’’ Science from Keio University, Japan, in 2003.
J. Lightw. Technol., vol. 26, no. 14, pp. 2227–2234, Jul. 2008. doi: From 2001 to 2003, he was a Research Associate
10.1109/JLT.2008.920138. with the Department of Information and Com-
[31] S. Vishwanath and A. Goldsmith, ‘‘Adaptive turbo-coded modulation for puter Science, Keio University. From 2003 to
flat-fading channels,’’ IEEE Trans. Commun., vol. 51, no. 6, pp. 964–971, 2006, he was with the Communication Research
Jun. 2003. doi: 10.1109/TCOMM.2003.813180. Laboratory, Independent Administrative Institu-
[32] A. Svensson, ‘‘An introduction to adaptive qam modulation schemes tion (National Institute of Information and Com-
for known and predicted channels,’’ Proc. IEEE, vol. 95, no. 12, munications Technology). In 2006, he was on
pp. 2322–2336, Dec. 2007. doi: 10.1109/JPROC.2007.904442. assignment with ATR Wave Engineering Laboratories. In 2007, he was a
[33] R. Schmogrow, B. Nebendahl, M. Winter, A. Josten, D. Hillerkuss, Lecturer with the Faculty of Information Sciences, Hiroshima City Univer-
S. Koenig, J. Meyer, M. Dreschmann, M. Huebner, C. Koos, J. Becker,
sity. He is currently a Professor with the Graduate School of Engineering,
W. Freude, and J. Leuthold, ‘‘Error vector magnitude as a performance
Chiba University. His research interests include OFDM, MIMO, digital com-
measure for advanced modulation formats,’’ IEEE Photon. Technol. Lett.,
vol. 24, no. 1, pp. 61–63, Jan. 1, 2012. doi: 10.1109/LPT.2011.2172405. munication, channel coding, signal processing for telecommunications, and
[34] A. Georgiadis, ‘‘Gain, phase imbalance, and phase noise effects on error wireless power transfer. Dr. Ahn is a Senior Member of IEICE. From 2005 to
vector magnitude,’’ IEEE Trans. Veh. Technol., vol. 53, no. 2, pp. 443–449, 2006, he was an expert committee member for emergence communication
Mar. 2004. doi: 10.1109/TVT.2004.823477. committee, Shikoku Bureau of Telecommunications, Ministry of Internal
[35] H. A. Mahmoud and H. Arslan, ‘‘Error vector magnitude to SNR conver- Affairs and Communications (MIC), Japan. He received the ICF Research
sion for nondata-aided receivers,’’ IEEE Trans. Wireless Commun., vol. 8, Grant Award for Young Engineer, in 2002, the Funai Information Science
no. 5, pp. 2694–2704, May 2009. doi: 10.1109/TWC.2009.080862. Award for Young Scientist, in 2003, and the Distinguished Service Award
[36] C. Desset and L. van der Perre, ‘‘Validation of low-accuracy quantization from Hiroshima City, in 2010. He once served as an Associate Editor for
in massive MIMO and constellation EVM analysis,’’ in Proc. Eur. Conf. Special Section on Multi-dimensional Mobile Information Network for the
Netw. Commun. (EuCNC), Paris, France, Jun./Jul. 2015, pp. 21–25. doi: IEICE Transaction on Fundamentals.
10.1109/EuCNC.2015.7194033.