April 13, 2010 15:13 00232

International Journal of Neural Systems, Vol. 20, No. 2 (2010) 149–158

c World Scientific Publishing Company
DOI: 10.1142/S0129065710002322

A DISCRETE-TIME RECURRENT NEUROFUZZY

NETWORK FOR BLACK-BOX MODELING OF INSULIN
DYNAMICS IN DIABETIC TYPE-1 PATIENTS

MARCOS A. GONZÁLEZ-OLVERA∗, ANA G. GALLARDO-HERNÁNDEZ†

and YU TANG‡
Faculty of Electrical Engineering
National Autonomous University of Mexico
Mexico City, 04510, Mexico
∗
mangel@verona.fi-p.unam.mx
†
anagabygh@gmail.com
‡
tang@unam.mx
MARIA CRISTINA REVILLA-MONSALVE§
and SERGIO ISLAS-ANDRADE¶
Unidad de Investigación de Enfermedades Metabólicas
Instituto Mexicano del Seguro Social Mexico City, Mexico
§
cristina revilla@hotmail.com
¶
sergioislas@prodigy.net.mx

In this work we present a data-driven modeling of the insulin dynamics in different in silico patients
using a recurrent neural network with output feedback. The inputs for the identification is the rate
of insulin (µU/dl/min) applied to the patient, and blood glucose concentration. The output is insulin
concentration (µU/ml) present in the blood stream. Once completed the off-line modeling, this model
could be used for on-line monitoring of the insulin concentration for a better treatment. The learning
law of the recurrent neural network is inspired by adaptive observer theory, and proven to be convergent
in the parameters and stable in the Lyapunov sense, even with only 13 samples available. Simulation
results are shown to validate the presented modeling.

Keywords: Identification; nonlinear systems; biological Systems; recurrent neurofuzzy networks;

Lyapunov stability; diabetes.

1. Introduction glucose can be measured in real-time, it is difficult to

In insulin dynamics, as in other control problems,
a good model of the system to be controlled is the
first step in the design of a control scheme. However,
some common problems arise when obtaining such a
model: parametric uncertainty and/or non-modeled
dynamics in the models of each subsystems involved
in this system. There exists a vast literature that
covers the modeling of insulin dynamics.1–5 There
are simple models like Bergman Minimal Model1 or
very detailed models as Sorensen model.3 Even when
apply deterministic methods to estimate the insulin
based on these models because of parameter uncer-
tainties: for instance, Sorensen model is a 21st-order
metabolic model, which involves a several parame-
ters that are not always possible to obtain for every
patient, due to complexity associated with the mea-
surements and the costs.
Neural networks and fuzzy systems have proven
useful for identifying nonlinear systems. In their first
applications, besides being considered as classifiers

‡
Corresponding author.

149
April 13, 2010 15:13 00232
150 M. A. González-Olvera et al.
or controllers,6 they were used as static function
identifiers,7, 8 provide that the functions to be iden-
tified were smooth. The interpretation of a fuzzy
system as a neural network, called neurofuzzy net-
work, has been used to combine the best properties
of both: having the space partition via membership
functions, and take advantage of the training algo-
rithms designed for neural networks.9
When dealing with identification of dynamic
models, neurofuzzy networks incorporate dynamics
into the structure. This kind of neurofuzzy net-
works are then called recurrent neurofuzzy net-
works (RNFN). One of the very first approaches
given10 considered one-bit memory in each neu-
ron. In this context, several models have been
proposed, using different structures.9, 11–15, 37, 38 In
training of recurrent networks, the usual train-
ing schemes (as backpropagation or gradient tech-
niques16 ) are not longer useful and schemes that take
into account the dynamics of the network ought to
be used, such as Backpropagation-Through-Time,17
or Recurrent-Learning-Through-Time.18 However,
these techniques tend to be sluggish and unstable
in the presence of disturbances.
Regarding models to be used in control or esti-
mation problems, structures in state-space repre-
sentation are preferred in some applications.19, 20
However, they require the actual measurement of the
physical states of the system, which is not always
the case. Previously,21 a network with a stable train-
ing algorithm is proposed, where it generates its own
states in order to have the same input-ouput dynam-
ics as a real system, alleviating the need of measuring
those of the system.
In this paper we propose a data-driven modeling
in state-space representation of the insulin dynam-
ics in diabetic Type-1 patients, by analyzing the
response of the blood stream insulin concentration
when insulin is directly fed into the body and when
the glucose concentration is measured in real-time.
The training of the network is done using lab tests
that can measure both glucose and insulin concen-
trations in the bloodstream. Once the training is
completed, the objective is to have the network
to reproduce the insulin dynamics when only glu-
cose measurements and the injected insulin rate are
available.
This work is structured in the following man-
ner: In Sec. 2 the basic modeling problem of
the insulin dynamics is discussed. For this sys-
tem, a black-box structure based on Recurrent Neu-
ral Networks is presented in Sec. 3, jointly with
a training algorithm, where also the boundness
and convergence of the training algorithm are dis-
cussed. Examples for the validation of the perfor-
mance of the neural identifier are shown in Sec. 4.
Finally, conclusions and future work are drawn in
Sec. 5.
2. Problem Statement
Glucose is the principal energy substrate of the
human body, but at least 80% of body cells can not
absorb it by their own, so they need insulin to help
them to do it. Insulin is a hormone produced in pan-
creas by Beta cells (β cells), that also regulates the
glucose storage by liver, muscles, and adipose tissue,
in order to have glucose available for fasting periods.
When insulin is higher than its basal level these tis-
sues absorb glucose to store it, and the lack of insulin
activates the mechanisms to release it.22
In Diabetes Type-1 β cells are destroyed by
autoimmune mechanisms, so the pancreas can not
produce insulin anymore. In consequence, the glu-
cose concentration is always higher than its basal
level. Current medical treatment suggests three or
four daily glucose measurement, with the same num-
ber of subcutaneous insulin injections.23 Insulin can
also be delivered by an insulin pump, and there are
real time glucose monitors with implantable sensors.
This represents the sensor and actuator for a closed
loop control for glucose. There are several research
works about this topic.24–27
Insulin therapy can have some complications. In
the first place, the accumulation of insulin injections
can lead to a excessive concentration of insulin in
blood (also known as hyperinsulinemia) that pro-
duces hypoglycemia, a life-threatening condition.
If hyperinsulinemia persists, but within euglycemia
range, it can produce insulin resistance, and then
impairing the effect of the therapy. Otherwise, pro-
longed lack of insulin, can lead to excessive lipol-
ysis and unrestrain fatty acid oxidation producing
metabolic acidosis and dehydration.28 So, it is desir-
able to design a system that can estimate the con-
centration of insulin level in the bloodstream to help
that those limit conditions are not neither surpassed
nor fall below a certain level.
April 13, 2010 15:13 00232

A Discrete-Time Recurrent Neurofuzzy Network for Black-Box Modeling of Insulin Dynamics 151

and necessary conditions30 in order to transform (1)

through a local diffeomorphism zk = Φ(xk ) into the
system (2).
Now, the objective is to find a discrete-time
recurrent neural network that, using only input and
output measurements uk and yk , can generate a
dynamic mapping by a recurrent network uk → ŷk
such that makes yk , supt≥0 |ŷk − yk | is as small as
possible.
We propose the recurrent neurofuzzy network
(RNFN) to approximate f¯ in (2) via ϕ(y, u)θη by:

ηk+1 = Aηk + ϕ(ŷk , uk )θη ,

(3)
Fig. 1. Identification objective. ŷk = Cηk ,
From the theory given by Refs. 31–33, this is a
This can be considered as a system with two
dual identification problem. Under persistent excita-
inputs and one output, given that the inputs are
tion condition for the regressor ϕ(y, u) embedded in
the glucose concentration in the bloodstream and the
the neurofuzzy network, an adaptive observer can be
insulin injected, and the output is the insulin concen-
designed such that jointly estimates η̂ and θη via a
tration in the bloodstream. So, the modeling problem
decoupled algorithm.
is that, given input-output measurements, to obtain
a discrete-time dynamic model based on RNFN that
can learn this dynamics and generate an estimate of 3. Proposed Recurrent Neurofuzzy
the output, as depicted in Fig. 1. Network and Training Algorithm
This is now an identification problem, repre- The neurofuzzy system (3) is described by the fol-
sented by an unknown MISO system lowing r fuzzy rules:

xk+1 = f (xk , uk ),  R1 : If (uk , yk ) is F1 then
(1) 



yk = h(xk ), 
 ηk+1 = Aηk + θu1,1 u1,k



 + θu2,1 u2,k + θy,1 ρ(yk ),
where uk = [u1,k u2,k ]T ∈ 2 and yk ∈  are the 



input and the output, respectively, and x ∈ n is the 

..

 .
state of the system, f : n × 2 → n , h : n → , 



 Ri : If (uk , yk ) is Fi then
are unknown smooth functions, dependent on the
ηk+1 = Aηk + θu1,i u1,k
dynamics of the absorption of the glucose and the 


 + θu2,i u2,k + θy,i ρ(yk ),
insulin present in the bloodstream. For this applica- 



tion, It is assumed that (1) is observable in the region 

..

 .
of interest.29 The number of states n varies depend- 


 Rr : If (uk , yk ) is Fr then


ing on the precision required from the model, from 3 


 ηk+1 = Aηk + θu1,r u1,k
states1 to 21.3 

Following well-known results,30 we also hypothe- + θu2,r u2,k + θy,r ρ(yk ),
size (and then assume) that this system pertains to where the fuzzy sets Fi are characterized by the
the class of nonlinear systems transformable to an radial basis functions and u1 (t) and u2 (t) are
output feedback form the injected insulin rate and glucose concentration
zk+1 = Azk + f¯(yk , uk ), respectively. The consequent parameters are
(2)
yk = Czk , θu1,i = [ θu1,i,1 θu1,i,2 . . . θu1,i,n ]T ,

with (A, C) observable, and f¯(y, u) smooth. In the θu2,i = [ θu2,i,1 θu2,i,2 . . . θu2,i,n ]T ,
same way, it is assumed that (1) fulfills the sufficient θy,i = [ θy,i,1 θy,i,2 . . . θy,i,n ]T .
April 13, 2010 15:13 00232

152 M. A. González-Olvera et al.

Remark 1. The main function that the fuzzy be obtained as a function of the input and output sig-
structure has in the neurofuzzy network is to pro- nals from (1). The consequent part in the fuzzy rules
vide with a space partition i, and then each section is designed in such a way to meet with the persistence
be modeled by a nonlinear system in the form of the of excitation condition (see below). However, other
fuzzy rule Ri . type of fuzzy rules may be considered, as long as they
satisfy the persistence of excitation condition.
If the defuzzification is defined as a weighted aver-
r r
age ηk+1 = i=1 Fi ηk+1i
/ i Fi , then the regressor Remark 4. Similarly to some previous work,34 the
ϕ can be defined by regressor in this proposed scheme can be proven to
comply with the persistence of excitation condition.
ϕ(yk , uk ) = [ϕu1 (yk , uk )ϕu2 (yk , uk )ϕy (yk , uk )]
∈ n×3nr , (4) Remark 5. By construction, it follows from (5)
that the regressor ϕ is always bounded if u is
and ϕu1 , ϕu2 , ϕy ∈ 4 can be expressed as bounded.
[ F1 In×n . . . Fr In×n ]u1,k
ϕu1 (yk , uk ) = r , 3.1. Observation-training algorithm
i Fi

[ F1 In×n . . . Fr In×n ]u2,k In order to initialize the membership functions, a

ϕu2 (yk , uk ) = r , (5) simple clustering algorithm is to be used. In this
i Fi
case, we consider to use the Fuzzy C-Means35, 36 for
[ F1 In×n . . . Fr In×n ]ρ(yk ) this purpose, but other clustering methods may be
ϕy (y, u) = r ,
i Fi employed. We take the number of clusters is to be
equal to the number of rules in the fuzzy part of the
with ρ :  → , ρ = ρ(yk ) a sigmoid function,
network.
monotonically increasing and
We now let the network defined by (3) to work
lim ρ(y) = ±1, ρ(0) = 0. as a parallel model for (2), which in turn is equal
y→±∞
to the order of the system (3). Following the same
So, treatment as in Refs. 21 and 32, the system can
ηk+1 = Aηk + [ϕu1 (yk , uk )ϕu2 (yk , uk ) be trained by jointly estimating η and θη , with the
observable pair (A, C). Designing a vector K such
× ϕy (yk , uk )]θ, (6) that the eigenvalues of A − KC be in any desired
= Aηk + ϕ(yk , uk )θ. (7) position, the training equations are

Consequently, the parameter vector θ is Υ[k + 1] = (Ā − KC)Υ[k] + ϕ[k] (9)

θ̂[k + 1] = θ̂[k] + µ[k]Υ[k]T C T (y[k] − C η̂k) (10)
T
θ = [ θu1 T
θu2 θyT ]T ∈ 3nr , (8)
η̂[k + 1] = (Ā − KC)η̂[k] + ϕ[k]θ̂[k] + Ky[k]
with
+ Υ[k + 1](∆θ̂[k]) (11)
T
θu1 = [ θu1,1 T
. . . θu1,r ]T ∈ nr ,
where η̂ is the estimation of η, A − KC is a fil-
T
θu2 = [ θu2,1 T
. . . θu2,r ]T ∈ nr , ter matrix for ϕ with eigenvalues inside the uni-
T
θy = [ θy,1 T ]T ∈ nr .
. . . θy,r tary circle, Γ is a positive definite gain matrix. The
parameter and estimation error are exponentially
Remark 2. The objective of including a the sig- convergent provided that the regressor ϕ[k] is per-
moid function ρ is twofold: to provide with more non- sistently exciting, as stated in the next theorem:
linear identification capabilities to the network, and
to guarantee for this structure that ϕy (0, 0)θy = 0. Theorem 1. Consider the system (1) and the net-
work (3). Assume that the approximation error is
Remark 3. The structure of the network is such sufficiently small, then the training algorithm results
that the research in adaptive observers made by in an exponential convergence of the parameters
Refs. 30, 32 and 34 can be used, so the regressor can as well as the states to their true values (that is,
April 13, 2010 15:13 00232
A Discrete-Time Recurrent Neurofuzzy Network for Black-Box Modeling of Insulin Dynamics
η̃ = η̂ − η → 0 and θ̃ = θ̂ − θ → 0), provided that
the persistence of excitation condition
1 
k+L−1
ΥTi [i]C T CΥ[i] ≥ αI
L where k is the current epoch, Γf inal is the desired
i=k
is fulfilled.
Proof. Design Ao = Ā − KC to be Hurwitz. In the
same sense as Ref. 33, define the error signals
η̃[k] = η̂[k] − η[k], θ̃[k] = θ̂[k] − θ[k].
If the approximation error ε is sufficiently small,
then the error dynamics can be found as
η̃[k + 1] = Ao η̃[k] + ϕ[k]θ̃[k]
+ Υ[k + 1](θ̃[k + 1] − θ̃[k]).
Defining the error ζ[k] = η̃[k] − Υ[k]θ̃[k], its
dynamics is
ζ[k + 1] = A0 ζ[k] + (Ao Υ[k] + ϕ[k]
− Υ[k + 1])θ̃[k].
So using (11), one obtains the exponentially stable
equation
ζ[k + 1] = A0 ζ[k].
Now, taking the parametric error θ̃[k], its dynam-
ics is
θ̃[k + 1] = θ̃[k] − ΓΥT [k]C T C η̃[k]
= (I − ΓΥT [k]C T CΥ[k])θ̃[k]
−ΓΥ [k]C Cζ[k]
T T
It can be proven that the origin of the homogeneous
part of the previous equation is exponentially stable,
provided that (12) is fullfilled.
Remark 6. The persistent excitation condition for
the particular case of the identification of the insulin
dynamics implies that sufficient measurements must
be obtained in the laboratory from the patient under
several conditions, like before and after meals, and
under different levels of injected insulin.
Remark 7. This algorithm can be used to train a
network by epochs, and then the matrix gain Γ can
be changed in order to make the training finer as the
153
epoch number increases. In general, we considered
this to be a linear relation in the form
Γf inal − Γinitial
(k − 1)
Γk = (18)
(12) Epochs − 1
value of Γ in the last epoch, Γinitial the gain at k = 1
(the first epoch), and Epochs the total number of
epochs.
4. Modeling and Identification of
Insulin Dynamics in Different
Patients
For validation purposes, the network is tested with
simulated data in the identification of the responses
of three different in silico patients, where the input
u(t) is the insulin rate in µU/ml/min that is injected
(13) into the patient, calculated by the Quasi-Continuous
algorithm described in Ref. 24, and glucose con-
centration, G (mg/dl) the output is the measured
insulin concentration in blood I (µU/ml). It is true
that in a healthy human being, the glucose dynamics
has a strong correlation with the amount of insulin
(14) present in the blood stream: however, as in diabetic
patients this relation suffers from a decoupling and
relatively cheap and fast sensors exists, the glucose
is here used as another input to the system.
(15) The data of in silico patients are obtained
using the well known Bergman Minimal Model with
parameters considered in Ref. 25. This is done in
two phases: First, an RNFN is trained assuming that
both glucose and insuline data can be sampled 5
times per second, in order to show the approximation
(16) capabilities of the network. However, in real exper-
iments this can not be expected possible, as exper-
(17)
imental data would require of actual blood samples
and only a limited number of them can be done in a
certain period of time, so results using 13 data sam-
pled each 8 minutes are shown (also, as each mea-
surement requires blood, a limit of blood extracted
from a patient must be taken into account). Once
the network has been trained, it is tested comparing
to the actual amount of insulin present in the blood-
stream, assuming it can not be measured and only
estimated by the RNFN.
In all cases, the time unit is considered to be
one minute, with three rules (r = 3) and three
states (n = 3, the same as the Bergman model).
These number of rules was chosen as the nonlin-
ear dynamics is reasonably well represented with
April 13, 2010 15:13 00232

154 M. A. González-Olvera et al.

these zones, and also to avoid a parameter explo-

sion. The following design matrices and parameters
are used: λ{Ā} = [ −0.8 −0.6 −0.4 ], C = [ 1 0 0 ],
λ{Ā − KC} = [ −1.5 −1 −0.75 ], Epoch = 30,
Γinitial = 1, Γf inal = 0.1. So, a total of 27 parame-
ters are trained.

4.1. Simulation results under

unrestricted sampling conditions
As mentioned before, in this case ideal conditions are
considered, i.e. when an undetermined quantity of

Fig. 4. Training: Ideal case, Patient #2.

Fig. 2. Membership functions: Ideal case, Patient #1,

#2, #3. Fig. 5. Training: Ideal case, Patient #3.

Fig. 6. Identification results: Ideal case, Patient #1. I

is the insulin concentration in the bloodstream, Iˆ is the
Fig. 3. Training: Ideal case, Patient #1. estimation given by the network.
April 13, 2010 15:13 00232

A Discrete-Time Recurrent Neurofuzzy Network for Black-Box Modeling of Insulin Dynamics 155

Fig. 7. Identification results: Ideal case, Patient #2. I

is the insulin concentration in the bloodstream, Iˆ is the
estimation given by the network.

Fig. 8. Identification results: Ideal case, Patient #3. I

is the insulin concentration in the bloodstream, Iˆ is the
estimation given by the network.

blood can be obtained from the patient in any desired
time length. For this case, a total of 400 min and a
sampling time Ts = 0.2[min] are taken. Three dif-
ferent patient models are identified, using the input
signals (insulin and glucose) depicted in Figs. 4–6 for
each patient. As indicated before, the Fuzzy C-Means
algorithm is run in order to obtain the initial mem-
bership functions, which are depicted in Fig. 2. The
convergence of the training algorithms, as well as the
progression of the RMS error in normalized data are
Fig. 9. Training and identification: Real case — Patient
#1. yd is the response of the in silico patient, y is the
output identified by the network.
shown in Figs. 3–5 for each patient. It can be seen
that the parameters remain stable and convergent.
After training, the network is compared in parallel
with the output obtained from the in silico patient.
These results are depicted in Figs. 6–8.
April 13, 2010 15:13 00232
156 M. A. González-Olvera et al.
Fig. 10. Training and identification: Real case —
Patient #2. yd is the response of the in silico patient,
y is the output identified by the network.
4.2. Simulation results under restricted
conditions
Once the RNFN has been tested in unrestricted
conditions of sampling time and testing time, more
restrictive (and real) conditions must be considered.
Fig. 11. Training and identification: Real case —
Patient #3. yd is the response of the in silico patient,
y is the output identified by the network.
As in reality only a small amount of blood can be
taken from a certain patient, the tests are limited
in resources and time, and the patient can only be
confined for few hours (typically 2), in this case
the sampling time is increased to T s = 8[min],
and only 13 samples (less than two hours). In this
case, the same number of rules and states are used,
and the design matrices are modified to be λ{Ā} =
[ −0.4 −0.3 −0.2 ], C = [ 1 0 0 ], λ{Ā − KC} =
[ −0.9 0.9 −0.8 ]. As there are few data, more epochs
are used, so Epoch = 60, and Γinitial = 0.1, Γf inal =
0.02. Once again, for this limited resource of data,
the membership functions are obtained using a Fuzzy
April 13, 2010 15:13 00232
A Discrete-Time Recurrent Neurofuzzy Network for Black-Box Modeling of Insulin Dynamics
C-Means algorithm using the input and output sig-
nals depicted in Figs. 5–7 (corresponding to patient
#1, #2 and #3 respectively), and then the training
algorithm is run and the system and the network are
compared in parallel. In this example, the dynamics
are compared using a validation set where more data
is available. It can be seen that the parameters are
still convergent, as can be seen in Figs. 9–11, and that
the dynamics is captured by the network, although
some high-frequency information is lost because of
the increased sampling time.
5. Conclusions
In this work we presented a data-driven modeling of
the insulin dynamics in different in silico patients
using a discrete-time recurrent neural network with
output feedback, where the input signals are consid-
ered as the rate of insulin (in µU/dl/min) applied to
the patient, and the blood glucose concentration, and
the output is insulin concentration (µU/ml) present
in the blood stream. With a membership initializa-
tion based on the Fuzzy C-Means algorithm, and a
learning law inspired from adaptive observer theory,
it was proven to be convergent in the parameters
and stable in the Lyapunov sense even under model
mismatch and low sampling rates, and when just
few data is available. Simulation and experimental
results were shown in order to validate the presented
modeling, using only input-output data and prac-
tically no knowledge of the models of subsystems
involved in the system. Future work includes con-
trol schemes based in the identified model, as well as
analysis of the obtained model.
Acknowledgments
This work is partly supported by UNAM-PAPIIT
IN120009. The work of Marcos A. González-Olvera
and Ana G. Gallardo-Hernández is supported by
Conacyt.
References
1. R. N. Bergman, Y. Z. Ider, C. R. Bowden,
and C. Cobelli, Quantitative estimation of insulin
sensitivity, Am. J. Physiology, Endocrinology and
Metabolism 235 (1979) E667–E677.
2. V. Bolie, Coefficients of normal blood glucose regu-
lation, Journal of Applied Physiology 16 (1961) 783–
788.
157
3. J. T. Sorensen, A physiologic model of glucose
metabolism in man and its use to design and assess
improved insulin therapies for diabetes, Ph.D. disser-
tation, Massachusetts Institute of Technology. Dept.
of Chemical Engineering (1985).
4. M. Wilinska, L. J. Chassin, H. Schaller, L. Schaupp,
T. Pieber, and R. Hovorka, Insulin kinetics in type-
1 diabetes: continuous and bolus deliveryinsulin
kinetics in type-1 diabetes: continuous and bolus
delivery, Biomedical Engineering, IEEE Transac-
tions 52-1 (2005) 3–12.
5. C. D. Man, R. Rizza and C. Cobelli, Meal simulation
model of the glucose-insulin system, IEEE Trans-
action on Biomedial Engineering 54 (2007) 1740–
1749.
6. K. Narendra and K. Parthasarathy, Identification
and control of dynamical systems using neural net-
works, IEEE Tras. Neural Networks 1 (1990) 4–27.
7. G. Cybenko, Approximation by superpositions of a
sigmoidal function, Math. Con. Signal Syst. 2 (1989)
303–314.
8. J.-S. R. Jang, Anfis: Adaptive-network-based fuzzy
inference system, IEEE Transactions on Systems,
Man and Cybernetics 23 (1993) 665–685.
9. C.-F. Juang, A tsk-type recurrent fuzzy network for
dynamic systems processing by neural network and
genetic algorithms, IEEE Transactions of Fuzzy Sys-
tems 10(2) (2002) 155–170.
10. J. Hopfield, Neural networks and physical systems
with emergent collective computational abilities,
Proceedings of the National Academy of Sciences
79(8) (1982) 2554–2558.
11. V. Gorrini and H. Bersini, Recurrent fuzzy sys-
tems, Proc. IEEE Int. Conference on Fuzzy Systems
(1994), 193–198.
12. A. D. Back and A. C. Tsoi, A time series modeling
methodology using fir and iir synapses, Proc. Work-
shop on Neural Networks for Statistical and Eco-
nomic Data (1990), 187–194.
13. P. A. Mastorocostas and J. B. Theocharis,
A recurrent fuzzy-neural model for dynamic system
identification, IEEE Trans. onf Systems, Man and
Cybernetics — Part B: Cybernetics 32(2) (2002)
176–190.
14. A. Poznyak, E. Sanchez, and W. Yu, Differen-
tial Neural Networks for Robust Nonlinear Con-
trol: Identification, State Estimation and Trajectory
Tracking (World Scientific, 2001).
15. M. A. Gonzalez-Olvera and Y. Tang, Identification
and control of nonlinear dynamic systems via a con-
strained input-output neurofuzzy network, Advances
in Control Systems Theory and Applications (eds.)
G. Tao and J. Sun (2008).
16. O. Nelles, Nonlinear System Identification (Springer,
Berlin, 2001).
17. P. Werbos, Backpropagation through time: what it
does and how to do it, Proceedings of the IEEE
78(10) (1990) 1550–1560.
April 13, 2010 15:13 00232
158 M. A. González-Olvera et al.
18. R. J. Williams and D. Zipser, A learning algorithm
for continually running fully recurrent neural net-
works, Neural Computation 2(1) (1989) 270–280.
19. A. Poznyak, T. Poznyak and I. Chairez, Dynamic
neural observers and their application for identifica-
tion and purification of water by ozone, Automation
and Remote Control 67(6) (2006) 887–899.
20. W. Yu and X. Li, Fuzzy identification using fuzzy
neural network, IEEE Trans. on Fuzzy Systems
12(3) (2004) 411–420.
21. M. A. Gonzalez-Olvera and Y. Tang, Identification
of a class of nonlinear systems by a continuous-
time recurrent neurofuzzy network, American Con-
trol Conference, to be published.
22. A. Guyton and J. Hall, Textbook of Medical Physiol-
ogy (W. B. Saunders, 2000).
23. A. D. Asociation, Standards of Medical Care in Dia-
betes (ADA, 2009).
24. G. Gallardo-Hernandez, L. Fridman, S. Islas-
Andrade and Y. Shtessel, Quasi-continuous high
order sliding modes controllers applied to glucose-
insulin regulatory system models, in Proceedings of
Decision and Control Conference, 9–11 December
2008.
25. P. Kaveh and Y. Shtessel, Blood glucose regula-
tion using higher-order sliding mode control, Int.
J. of Robust and Nonlinear Control 18 (2007) 557–
569.
26. R. Hovorka, V. Canonico, L. Chassin, U. Haueter,
M. Massi-Benedetti, M. Federici, T. Pieber,
H. Schaller, L. Schaupp, T. Vering and M. Wilinska,
Nonlinear model predictive control of glucose con-
centration in subjects with type 1 diabetes, Physio-
logical Measurements 25 (2004) 905–920.
27. R. Parker, F. Doyle III, J. Ward and N. Pep-
pas, Robust h∞ glucose control in diabetes using a
physiological model, American Institute of Chemical
Engineers (AIChE) 46(12) (2004) 2537–2549.
28. Larsen, Kronenberg, Melmed and Polonsky,
Williams Textbook of Endocrinology (Sanuders,
USA, 2003).
29. M. Spong, Robot Dynamics and Control (John Wiley
& Sons, Inc. New York, 1989).
30. R. Marino, Adaptive observers for single output non-
linear systems, Automatic Control, IEEE Transac-
tions on 35(9) (1990) 1054–1058.
31. G. Zhang, B. E. Patwo and M. Y. Hu, Forecast-
ing with artificial neural networks: The state of the
art, International Journal of Forecasting 14 (1998)
35–62.
32. Q. Zhang, Adaptive observer for multiple-input-
multiple-output (MIMO) linear time-varying sys-
tems, IEEE Transactions on Automatic Control
47(3) (2002) 525–529.
33. A. Guyader and Q. Zhang, Adaptive observer for
discrete time linear time varying systems, in Proc. of
3th IFAC/IFORS Symposium on Identification and
System Parameter Estimation, SYSID 2003.
34. G. Bastin and M. R. Gevers, Stable adaptive
observers for nonlinear time-varying systems, IEEE
Transactions on, Automatic Control 33(7) (1988)
650–658.
35. J. Dunn, A fuzzy relative of the ISODATA process
and its use in detecting compact well-separated clus-
ters, Journal of Cybernetics 3(3) (1973) 32–57.
36. J. Bezdek, Pattern Recognition with Fuzzy Objective
Function Algorithms (Kluwer Academic Publishers
Norwell, USA, 1981).
37. H. Adeli and X. Jiang, Dynamic fuzzy wavelet neu-
ral network model for structural system identifi-
cation, Journal of Structural Engineering 132(1)
(2006) 102–111.
38. G. Puscasu and B. Codres, Nonlinear system iden-
tification based on internal recurrent neural net-
works, International Journal of Neural Systems
19(2) (2009) 115–125.
Copyright of International Journal of Neural Systems is the property of World Scientific Publishing Company
and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright
holder's express written permission. However, users may print, download, or email articles for individual use.