This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts

for publication in the IEEE ICC 2010 proceedings

A Neural Network Based Spectrum Prediction Scheme

for Cognitive Radio
Vamsi Krishna Tumuluru, Ping Wang and Dusit Niyato
Center for Multimedia and Network Technology (CeMNeT)
School of Computer Engineering, Nanyang Technological University, Singapore.
AbstractThe Cognitive Radio (CR) technology enables the
unlicensed users to share the spectrum with the licensed users
on a non-interfering basis. Spectrum sensing is an important
function for the unlicensed users to determine availability of
a channel in the licensed users spectrum. However, spectrum
sensing consumes considerable energy which can be reduced by
employing predictive methods for discovering spectrum holes.
Using a reliable prediction scheme, the unlicensed users will
sense only those channels which are predicted to be idle. By
achieving a low probability of error in predicting the idle
channels, the spectrum utilization can also be improved. Since the
traffic characteristics of most licensed user systems encountered
in real life are not known a priori, we design the spectrum
predictor using the neural network model, multilayer perceptron
(MLP), which does not require a prior knowledge of the traffic
characteristics of the licensed user systems. The performance of
the spectrum predictor is analyzed through extensive simulations.

Keywords Cognitive radio, channel status prediction, neural

networks, spectrum utilization.
I. I NTRODUCTION
A spectrum survey conducted by the U.S. Federal Communications Commission (FCC) [1] showed that the present static
spectrum allocation policies have caused spectrum scarcity
among the unlicensed user systems and under utilization
of spectrum among the licensed user systems (e.g., UHF
band). Opportunistic spectrum access (OSA) techniques can be
adopted by the unlicensed user systems to meet the spectrum
requirements of their users. Recently, the FCC has approved
opportunistic spectrum access by the unlicensed users in TV
spectrum (UHF and VHF bands) [2]. This concept can be
extended to other licensed user systems and is popularly
called as the Cognitive Radio (CR) technology [3]. A wireless
network designed based on the CR technology is referred to
as a cognitive radio network (CRN). A CRN is composed of
two types of users, namely, the primary users (licensed users)
and the secondary users (unlicensed users). The primary users
have a higher priority than the secondary users in accessing the
channels in the licensed spectrum. In most cases, the secondary
users in a CRN logically divide the channels allocated to the
primary users into slots [4]. Within each slot the secondary
user has to sense the primary user activity for a short duration
and accordingly accesses the slot when it is sensed idle. The
idle slots are also called spectrum holes or white spaces. Thus
the secondary users can access the licensed spectrum without
causing any harmful interference to the primary users.
To minimize the interference to the primary users, the
secondary users need a reliable spectrum sensing mechanism.

Several spectrum sensing mechanisms were proposed in literature [5], [6], [7], in some of which the secondary users
are assumed to be able to sense the full spectrum. However,
the secondary users usually sense only a part of the spectrum
in a slot due to hardware constraints. On the other hand, the
secondary users may want to conserve their sensing energy by
avoiding the busy portions of the spectrum during sensing.
To let the secondary users efficiently manage the sensing
mechanism, channel1 status prediction becomes important. The
secondary users may predict the status of a channel based on
the sensing history and sense only if a channel is predicted to
be idle in the next time slot. Thereby, the secondary users can
save the sensing energy. Besides, using spectrum prediction,
the effective bandwidth in the next slot may be estimated which
allows the secondary users to adjust the data rates in advance.
In this paper, we design the channel status predictor using
the neural network model multilayer perceptron (MLP). In
cognitive radio networks, it is difficult to obtain the statistics
of channel usage by the primary users a priori. The advantage
of neural networks over statistical models is that it does not
require a priori knowledge of the underlying distributions of
the observed process. Therefore, the neural networks offer an
attractive choice for modeling the channel status predictor.
The rest of the paper is organized as follows. In Section II,
we present some related work. In Section III, we propose the
channel status predictor design using the MLP neural network.
In Section IV, we provide the simulations for the MLP
predictor and demonstrate the effect of spectrum prediction
in improving the spectrum utilization and saving spectrum
sensing energy. Finally, Section V concludes this paper.
II. R ELATED W ORK
In [9], the spectrum prediction was done using a linear filter
model followed by a sigmoid transform. The performance of
the predictor suffered due to the non-deterministic nature of
the binary series. In [10], a hidden Markov model (HMM)
based spectrum predictor was proposed. The primary user
traffic follows Poisson process with 50% traffic intensity (i.e.,
50% channel time is occupied by the primary users). However,
in [10] the accuracy of channel status prediction was not
provided. Another HMM based predictor was also proposed
in [11], but it only deals with deterministic traffic scenarios,
making it not applicable for the actual environment. There
are several drawbacks in using the HMM based spectrum
prediction schemes [10], [11] such as, determining an optimal
1 In

this paper, channel and spectrum are used interchangeably.

978-1-4244-6404-3/10/$26.00 2010 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings

model (number of states in the HMM) is difficult and even

impossible in practice, a huge memory space is needed to store
a large number of past observations, and model estimation has
high computational complexity.
III. S PECTRUM P REDICTION USING N EURAL N ETWORK
The channel status prediction problem is treated as a binary
series prediction problem. We design the binary series predictor using neural networks. Neural networks are nonlinear
parametric models which create a mapping function between
the input and output data. The input and output data are
obtained from the binary series xT1 = {x1 , x2 , .., xt , .., xT } which
is generated for the channel by sensing the channel status at
every slot for a duration T . The channel status in a slot can
be either busy or idle depending on whether a primary user
activity is present or not. The channel status busy and idle are
denoted by binary symbols 1 and 1 respectively. Using the
binary series, the neural network predictor is trained to predict
the channel status in the next slot based on the slot status
history. In a multiple channel system, a predictor is assigned
to each channel.
Unlike the HMM based prediction schemes [10], [11] which
require continuous training, the neural network predictor is
trained only once in an offline fashion when the observed
process is stationary. Once the neural networks are trained,
the computational complexity is significantly reduced. The
neural network model, multilayer perceptron (MLP), has been
used in applications, e.g., system identification and time series
prediction [12], [13].
A. MLP Predictor Design
The MLP network is a multilayered structure consisting
of an input layer, an output layer, and a few hidden layers.
Excluding the input layer, every layer contains a certain
number of computing units (referred to as neurons) which
calculate a weighted sum of the inputs and perform a nonlinear
transform on the sum. The nonlinear transform is implemented
using a hyperbolic tangent function. Neurons belonging to
different layers are connected through adaptive weights. The
output of a neuron j in the lth layer, denoted by yjl , can be
represented as
yjl =


1 exp(vjl )
1 + exp(vjl )

(1)

l ) is the weighted sum of the inputs

where vjl (= i yil1 wji
coming from the output of the neurons in the (l 1)th layer
l connecting the
using the adaptive weights (or parameters) wji
th
neuron j in the l layer and neuron i in the (l 1)th layer.
Due to the nonlinear transform on vjl in (1), yjl lies in the
range of [1, +1]. If the inputs come from the input layer, vjl is
calculated using the corresponding inputs instead of yil1 . The
total number of inputs in the input layer is referred to as the
order of the MLP network and is denoted by .
The number of hidden layers and the number of neurons
in each layer depend on the application. For channel status
prediction problem, we found the MLP network with two
hidden layers to be sufficient. The first hidden layer has 15

Fig. 1.

MLP predictor training.

neurons, the second hidden layer has 20 neurons, and the

output layer has only one neuron. The order of the MLP
predictor is set to 4.
B. MLP Predictor Training
The MLP predictor training process is illustrated in Fig. 1.
The parameters of the MLP predictor are updated using the
batch backpropagation (BP) algorithm [8]. The training patterns are obtained by ordering the entire binary series xT1 into
input vectors xtt +1 = {xt , xt1 , ..., xt +2 , xt +1 } of length
and the corresponding desired value xt+1 . For each input
vector presented to the input layer of the MLP, the outputs
of the neurons in each layer are calculated proceeding from
the first hidden layer to the output layer using equation (1).
This computation is called forward pass. The output of the
neuron in the output layer y1o is referred to as the MLP output
and is denoted by xt+1 . xt+1 is treated as an estimate of the
corresponding desired value xt+1 . The difference between the
desired value and its estimate is called as the error et which
can be expressed as follows:
et = xt+1 x
t+1 = xt+1 y1o .

(2)

The objective of the training algorithm is to minimize this error

l such that the MLP output
by adapting the parameters wji
approximately represents the desired value. In other words,
the MLP predictor tries to create a mapping function between
the input vector and the desired value. According to the BP
algorithm [8], it is easier to minimize a mean square error

criterion E (= 12 e2t ) than to directly minimize the error et .
Based on the BP algorithm [8], the parameters are updated as
follows:

et

wt

wt

wt1 + wt
E

+ wt1 .
wt

(3)
(4)

l at
In equations (3) and (4), wt represents the parameter wji
l
time instant t (also denoted by wji t ) while and represent
the learning rate and the momentum term respectively. can
be chosen from the range (0, 1) while can be chosen from
the range [0.5, 0.9]. In the simulations, the values of and
are set to 0.2 and 0.9, respectively. The partial derivative
E/wt in equation (4) is calculated successively for each

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings

neuron by proceeding backwards from the output layer to the

input layer. This computation is called backward pass. The
partial derivative E/wt can be expressed in terms of the
l
variables et , yjl , vjl and wji
using chain rule as follows:
t
l
l
E
E
E et yj vj
=
=
.
l
l
l
l
wt
et yj vj wji
wji
t

(5)

l
in equation (5)
The partial derivatives yjl /vjl and vjl /wji
t
l as
are calculated based on the relation between yjl , vjl and wji
l
l
l
l
explained in Section III-A. Hence yj /vj = (1 yj )(1 + yj ) and
l
vjl /wji
= yil1 .
t
The partial derivative E/wt is calculated in two ways
depending on whether the neuron j has a desired output
value or not. For the neuron in output layer o, a desired
output value xt+1 exists. Therefore, E/wt for the parameter
o
wt = w1it
connecting the output layer neuron to a neuron i in
the preceding layer (o 1) is expressed as

E
E
E
E et y1o v1o
=
=
=
.
o
o
l
wt
w1it
et y1o v1o w1it
wji

(6)

l
o
By substituting vjl = v1o , yil1 = yio1 and wji
= w1i
and using
t
t
l as explained in Section
the relation between yjl , vjl and wji
o
in equation
III-A, the partial derivatives y1o /v1o and v1o /w1i
t
(6) are calculated as (1 y1o )(1 + y1o ) and yio1 respectively. In
equation (6), E/et = et whereas et /y1o = 1 using equation
(2). Hence equation (6) can be expressed as



E
E
=
= (et )(1) (1 y1o )(1 + y1o ) (yio1 ) .
o
wt
w1i
t

l
E
E yj
=
.
vjl
yjl vjl

(8)

In order to calculate the local gradient jl (t) at time instant t

for every neuron j in a hidden layer l, jl (t) is initialized at
the output layer neuron and calculated recursively for every
neuron j in layer l by proceeding backwards from the output
layer to the first hidden layer. The initialization is given by
1o (t) =



E
E y1o
E et y1o
=
=
= (et )(1) (1y1o )(1+y1o ) .
o
o
o
o
o
v1
y1 v1
et y1 v1

(9)

The recursive equation for jl (t) is given by

jl (t) = (1 yjl )(1 + yjl )

l+1
kl+1 (t)wkj
.
t

(10)

l
The partial derivative E/wt for wt = wji
connecting the
t
neuron j in a hidden layer l to a neuron i in the layer (l 1)
is expressed in terms of jl (t) as follows:
l
l
E
E
E yj vj
=
=
= jl (t)yil1 .
l
l
l
l
wt
wji
yj vj wji
t

(11)

After calculating wt at each instant t (t = 1, 2, . . . , T ),

the parameter w is updated by the average value of wt as
follows:
wnew

T

1
= wold +
wt
T t=1

if xt+1 0 then xt+1

+1

if xt+1 < 0 then xt+1

1 .

(12)

(13)

IV. S IMULATION AND A NALYSIS

In this section, we use simulation to evaluate the accuracy
of the MLP predictor. The HMM based spectrum prediction
schemes [10], [11] mentioned in Section II do not provide
details of the model like number of states and length of the
observation sequence, hence making it difficult for comparison
with the proposed MLP spectrum prediction scheme. For the
purpose of simulation, the primary user traffic on a channel is
assumed to follow Poisson process2 . The ON/OFF time of the
channel is drawn from geometric distributions. For different
traffic scenarios, we vary the traffic intensity and the mean
inter-arrival time tinter of the traffic bursts. The traffic intensity
is related to the mean inter-arrival time as follows.

(7)

For neuron j in a hidden layer l, a desired output value does

not exist. In this case, E/wt is calculated in terms of an error
term called local gradient, denoted by jl (t). The local gradient
jl (t) for a neuron j in the hidden layer l can be expressed as
jl (t) =

where T is the length of the entire binary series xT1 .

The parameter updating is repeated until the minimum of
the mean square error or the maximum number of iterations
is reached. Once training is complete, we test the MLP
predictor by randomly observing successive slots xtt +1 and
computing the MLP output xt+1 . By using a decision threshold
at the MLP output, the predicted value can be expressed as
binary symbol

tserv
mean ON time
=
mean ON+OFF time tinter

(14)

where tserv is the mean time that the primary user is active
on a channel for each traffic burst. A minimum of 50% traffic
intensity is maintained for a channel. The training and testing
data are generated by observing the channel occupancy. The
length of the training and testing data are chosen as 1000 and
30,000 slots respectively.
The accuracy of the MLP spectrum predictor is evaluated
in terms of the wrong prediction probability, denoted by
Ppe (Overall). Of particular interest is the wrong prediction
probability given the real channel status is busy, denoted by
Ppe (Busy). Ppe (Busy) is an important measure from the primary
users standpoint because it indicates the level of interference
to the primary user. Ppe (Overall) is an important measure from a
secondary users perspective because the goal of the secondary
user is to minimize the interference to the primary users while
maximizing its own transmission opportunities.
We also investigate the probability of wrongly predicting the
idle channel status (i.e., the channel is predicted to be busy
when it is actually idle or so-called false-alarm probability),
denoted by Ppe (Idle). It is observed that for a given mean interarrival time tinter and different values of traffic intensity ,
there are only small changes in the probability Ppe (Idle). Hence
Ppe (Idle) is shown for different mean inter-arrival times tinter
but only for traffic intensity of = 0.5.
2 In the simulation, we use Poisson process as an example to generate
primary user traffic. However, the proposed MLP spectrum predictor is not
restricted to Poisson process and is applicable to any traffic distribution.

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings

A. Performance of the Predictor under Stationary Traffic Conditions

First, we evaluate the performance of the MLP predictor under stationary3 traffic conditions. Fig. 2 shows the performance
of the MLP predictor in predicting the busy channel status
for various traffic scenarios. It can be seen that for a given
mean inter-arrival time tinter , the MLP predictor performance
improves when the traffic intensity increases. The worst case
scenario occurs when the mean inter-arrival time is 10 slots
and the traffic intensity is 0.5 (mean ON time=mean OFF
time=5 slots). Fig. 3 shows the overall performance of the
0.1
Mean interarrival time
0.09

22
20
18
16
10

0.08

Pe(Busy)

0.07
0.06

TABLE I
P ERFORMANCE OF THE MLP PREDICTOR IN PREDICTING THE IDLE
CHANNEL STATUS

Mean ON time (slots)

5
8
9
10
11

Mean OFF time (slots)

5
8
9
10
11

Ppe (Idle)
0.101477
0.063528
0.058236
0.052713
0.047875

predictor is analyzed by calculating the percentage of wrong

predictions in the duration [t0 , t6 ]. For one time training of
the MLP predictor during the interval [t0 , t1 ], the percentage of
wrong predictions is 6.31%. The performance can be improved
by retraining the MLP predictor for a short duration in each
interval. When the MLP is retrained using 1000 observations
at the beginning of each interval, the percentage of wrong
predictions reduces to 4.07%.
TABLE II
E XAMPLE OF A LICENSED CHANNEL WITH NON - STATIONARY TRAFFIC

0.05
0.04

DISTRIBUTION
0.03

Time interval
[t0 , t1 ]
[t1 , t2 ]
[t2 , t3 ]
[t3 , t4 ]
[t4 , t5 ]
[t5 , t6 ]

0.02
0.01
0.5

0.55

0.6

0.65
Traffic intensity

0.7

0.75

0.8

Fig. 2. Performance of the MLP predictor in predicting the busy channel

status for various traffic scenarios.

MLP predictor for various traffic scenarios. It can be seen that

Ppe (Overall) is slightly higher than Ppe (Busy) under the same
traffic scenario. This is because of the inclusion of the wrong
predictions when channel status is idle.
Table I shows the performance of the MLP predictor in
predicting the idle channel status. It can be observed that
Ppe (Idle) ranges from 4.8% to 10.1% when the mean ON/OFF
time varies from 11 slots to 5 slots, respectively.
0.1
Mean interarrival time
0.09

22
20
18
16
10

0.08

0.06

Pe(Overall)

0.07

0.05

0.04

0.03

0.02
0.5

0.55

0.6

0.65
Traffic intensity

0.7

0.75

0.8

Fig. 3. Overall performance of the MLP predictor in predicting the channel

status for various scenarios.

B. Performance of the Predictor under Non-Stationary Traffic

Conditions
For the non-stationary traffic condition, consider a licensed
channel with time varying statistics as shown in Table II.
Each interval has 6000 slots. The performance of the MLP
3 Stationary means that for each simulation scenario, the primary user traffic
statistics (i.e., the traffic intensity and the mean inter-arrival time tinter )
remain unchanged over time.

tinter
20
10
18
16
22
18

0.7
0.5
0.667
0.625
0.5
0.5

C. Performance of Cognitive Radio with Spectrum Prediction

We demonstrate the advantage of using spectrum prediction
based sensing using two performance measures, i.e.,
Percentage improvement in spectrum utilization, denoted
by SUimp (%),
Percentage reduction in sensing energy, denoted by
SEred (%).
1) Improvement in Spectrum Utilization: We consider a
primary user system contains Nch channels with different
primary user traffic distributions. Each secondary user is able
to sense only one channel during a slot due to the hardware
constraint. We also assume that every secondary user stores
a short history of the sensing results for every channel. This
information can be collected from neighbors over a common
control channel.
We consider two types of secondary users, CRsense device
and CRpredict device. A CRsense device randomly selects a
channel at every slot and senses the status of that channel,
while a CRpredict device individually predicts the status of all
channels based on their respective slot history, before sensing.
The channel to be sensed by the CRpredict device is randomly
selected among those channels with predicted idle status.
Spectrum utilization (SU ) can be defined as the ratio of the
number of idle slots discovered by the secondary user to the
total number of idle slots available in the system over a finite
period of time (e.g., 20000 slots).

SU =

Number of idle slots sensed

.
Total number of idle slots in Nch channels

(15)

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2010 proceedings

The percentage improvement in spectrum utilization due to

spectrum prediction can be expressed as
SUimp (%) = (SUpredict SUsense )/SUsense

(16)

where SUsense and SUpredict represent the spectrum utilization

for the CRsense and CRpredict devices, respectively. Substituting (15) in (16), SUimp (%) can be given by
SUimp (%) = (Ipredict Isense )/Isense

(17)

where Isense and Ipredict represent the number of idle slots

sensed by the CRsense and the CRpredict devices respectively.
In the simulation, we consider the system with different
number of channels Nch (the channels are added sequentially
according to Table III). Table IV shows that a CRpredict device
TABLE III
D IFFERENT CHANNEL MODELS FOR THE PRIMARY USER SYSTEM
Channel index
1
2
3
4
5
6
7
8

0.5625
0.6667
0.5
0.6818
0.5
0.7
0.5
0.5

TABLE IV
P ERCENTAGE IMPROVEMENT IN THE SPECTRUM UTILIZATION DUE TO
SPECTRUM PREDICTION

Isense
8039
7467
7952
7549
7911
8124

Ipredict
12884
14172
15659
15423
15733
15947

SUimp (%)
60.27
89.80
96.92
104.3
98.87
96.29

2) Reduction in Sensing Energy: Consider a single channel

scenario, a CRsense device senses all the slots whereas a
CRpredict device only senses when the channel status of the
slot is predicted to be idle. In other words, when the slot
status is predicted to be busy, the sensing operation is not
performed, thereby sensing energy is saved. If we assume one
unit of sensing energy is required to sense one slot, then the
total sensing energy required for a CRsense device in a finite
duration of time (e.g., 20000 slots) can be given by
SEsense

(Total

number of slots in the duration)

(unit sensing energy)
(18)

while the total sensing energy required by the CRpredict device

can be given by

SEpredict = SEsense (Bpredict ) (unit

SEsense SEpredict
=
SEsense

predict
. (20)
Total no. of slots
Table V shows the percentage reduction in the sensing energy
for traffic intensity of 50% and different mean inter-arrival
times when CRpredict device with MLP predictors is used.

SEred (%) =

TABLE V
P ERCENTAGE REDUCTION IN SENSING ENERGY FOR TRAFFIC INTENSITY
= 50% AND DIFFERENT MEAN INTER - ARRIVAL TIMES tinter
tinter
10
16
18
20
22

Bpredict
10209
10605
10272
10251
10224

SEred (%)
51.04
53.03
51.36
51.25
51.12

V. C ONCLUSION
Channel status prediction in cognitive radio networks can
not only save the sensing energy greatly, but also improve
spectrum utilization. As the statistics of channel usage by
primary users in cognitive radio networks is not known a priori,
we have designed the channel status predictor using the neural
network model, multilayer perceptron (MLP). Analysis of the
MLP predictor performance has been presented using various
simulations.

tinter
16
18
10
22
20
10
18
22

using MLP predictors can discover more idle slots than a

CRsense device. The percentage of improvement in spectrum
utilization is more than 60%.

Nch
3
4
5
6
7
8

percentage reduction in the sensing energy can be given by

sensing energy) (19)

where Bpredict is the total number of busy slots predicted

by the CRpredict device. Therefore, using (18) and (19), the

R EFERENCES
[1] Federal Communications Commission, Spectrum policy task force, in
Rep. OET Docket No. 02-135, Nov.2002.
[2] Federal Communications Commission, Evaluation of performance of
prototype TV-band white space devices Phase II, in Rep. OET Docket
No. 08-TR-1005, Oct. 2008.
[3] S. Haykin, Cognitive radio: Brain-empowered wireless communications, IEEE Journal on Selected Areas in Communications, vol. 23,
no. 2, pp. 210230, Feb. 2005.
[4] T. V. Krishna and A. Das, A survey on MAC protocols in OSA
networks, Computer Networks, vol. 53, no. 9, pp. 13771394, 2009.
[5] N. Shah, T. Kamakaris, U. Tureli and M. Buddhikot, Wideband spectrum sensing probe for distributed measurements in cellular band, in
Proceedings of ACM International Workshop on Technology and Policy
for Accessing Spectrum, vol. 222, Aug. 2006.
[6] Y. Hur, J. Park, et al., A wide band analog multi-resolution spectrum
sensing (MRSS) technique for cognitive radio (CR) systems, in Proceedings of IEEE International Symposium on Circuits and Systems,
ISCAS 2006, May 2006.
[7] Z. Tian and G. B. Giannakis, A wavelet approach to wideband spectrum
sensing for cognitive radios, in Proceedings of International Conference
on Cognitive Radio Oriented Wireless Networks and Communications
(CROWNCOM), 2006, pp. 15, June 2006.
[8] S. Haykin, Neural Networks: A comprehensive foundation, 2nd ed.,
Prentice Hall, pp. 161175, 1999.
[9] S. Yarkan and H. Arslan H, Binary time series approach to spectrum
prediction for cognitive radios, in Proceedings of IEEE Conference on
Vehicular Technology (VTC), pp. 15631567, Sept. 2007.
[10] I. A. Akbar and W. H. Tranter, Dynamic spectrum allocation in
cognitive radio using hidden markov models: Poisson distributed case,
in Proceedings of IEEE SoutheastCon, pp. 196201, March 2007.
[11] C. H. Park, S. W. Kim, S. M. Lim and M. S. Song, HMM based
channel status predictor for cognitive radio, in Proceedings of AsiaPacific Microwave Conference (APMC), pp. 14, Dec. 2007.
[12] K. S. Narendra and K. Parthasarathy, Identification and control of
dynamical systems using neural networks, IEEE Transactions on Neural
Networks, vol. 1, no. 1, pp. 427, March 1990.
[13] P. J. Werbos, Backpropagation through time: What it does and how to
do it, in Proceedings of IEEE, vol. 78, no. 10, pp. 15501560, Oct.
1990.