Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

DOI 10.1007/s00521-016-2387-4

ORIGINAL ARTICLE

Generalized self-tuning regulator based on online support vector

regression
Kemal Uçak1 • Gülay Öke Günel1

Received: 25 August 2015 / Accepted: 21 May 2016 / Published online: 8 June 2016
 The Natural Computing Applications Forum 2016

Abstract This paper introduces a novel generalized self- support vector regression  PID-type STR  Self-tuning
tuning regulator based on online support vector regression regulator  Stability analysis
(OSVR) for nonlinear systems. The main idea is to
approximate the parameters of an adaptive controller by
optimizing the regression margin between reference input 1 Introduction
and system output. For this purpose, ‘‘closed-loop margin’’
which depends on tracking error is defined, and the Adaptation is a vital characteristic of living organisms that
parameters of the adaptive controller are optimized so as to helps to increase their resistance to adverse environmental
minimize the tracking error which leads simultaneously to conditions. Every adaptation leads to some loss, in the form
the optimization of the closed-loop margin. The overall of material, energy or information for the organism [1].
architecture consists of an online SVR which computes a The organism can be strengthened against adverse condi-
forward model of the system, an adaptive controller with tions as a result of repeated adaptation which is an accu-
tunable parameters and an adaptation mechanism realized mulation of experiences that it can evaluate to minimize
by separate online SVRs to estimate each tunable controller the losses involved in adaptation [1]. Thus, the organism
parameter. The proposed architecture is implemented with can manage to learn how to alter its own characteristics
adaptive proportional–integral–derivative (PID) and adap- against contingency.
tive fuzzy PID in the controller block. The performance of By imitating the adaptation features of biological sys-
the generalized self-tuning regulator mechanism has been tems, a variety of solutions can be developed for problems
examined via simulations performed on a bioreactor in many engineering fields. Adaptation, when interpreted in
benchmark system, and the results show that the general- terms of control theory, helps to analyse and control sys-
ized adaptive controller and OSVR model attain good tems with changing parameters, model uncertainties or
control and modeling performances. varying operating conditions. An adaptive control system
includes a feedback structure to measure the quantity of
Keywords Fuzzy PID-type STR  Generalized adaptive adaptation and a mechanism, preferably incorporating
controller  Generalized parameter estimator  Online some sort of intelligence, to use this quantitative infor-
mation to design a controller. By introducing adaptation to
a conventional controller, it is possible to use it to cope
with strong nonlinearities, time delays and time-varying
& Kemal Uçak dynamics of systems. The necessity for adaptive controllers
[email protected] arises since the environment is continuously changing for
Gülay Öke Günel many real-world systems [2].
[email protected] The parameter adaptive control can be examined under
1 three main headings in a common framework according to
Department of Control and Automation Engineering, Faculty
of Electrical-Electronics Engineering, Istanbul Technical Astrom [3]: gain scheduling, model reference control, and
University, Maslak, 34469 Istanbul, Turkey self-tuning regulators.

123
S776 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

Gain scheduling can be reckoned as a mapping from objective functions. Since support vector machines (SVM)
previously defined scenarios between system parameters or possess convex objective functions, they ensure global
system operating conditions to controller parameters [2]. In minimum and have better generalization capability with
gain scheduling, the wide range of operating conditions of very few training data [13–15] compared to ANN and
the system is firstly divided into small ranges via a priori ANFIS. Hence SVM-based identification and control
information, and robust/optimal controllers are designed techniques have recently been utilized in adaptive control
for each small range. Decision trees or lookup tables are techniques instead of ANN and ANFIS [16–19].
employed so as to model the relationship between system In technical literature, there are various controller
conditions and controller parameters. The proper controller structures related to STR design based on soft computing
parameters most convenient to the situation of the system methods. Akhyar and Omatu [6] derived a self-tuning PID
are selected. controller using an ANN parameter estimator to control
Model reference adaptive control (MRAC) is applied to linear and nonlinear systems. Wang et al proposed an ANN
compel the closed-loop system to exhibit the same parameter tuner to approximate the parameters of a con-
behavior as a reference model. The reference model assigns ventional PI controller depending on various operating
the transient and steady-state specifications of the closed- conditions since unmodeled system dynamics and distur-
loop system. The goal of MRAC is to ensure convergence bances hamper to determine suitable scheduling points in
of the static and dynamic characteristics of the gain scheduling [7]. Ponce et al. [20] designed a self-tuning
adjustable system, to the characteristics of the reference control system based on an ANN controller trained by
model [1]. The desired closed-loop behavior of the con- tracking error instead of net output error to control non-
trolled system is specified by a model, error is computed as linear systems. In this approach, it is only required to know
the difference between the model and closed-loop outputs, the sign of the system Jacobian to ensure the convergence
and controller parameters are obtained to minimize this of the weighting coefficients and the estimation of the sign
error [2]. of system Jacobian is uncomplicated compared to esti-
Self-tuning regulators (STR) are among the most con- mating system Jacobian [20]. Flynn et al. [21] proposed to
venient adaptive control methods for nonlinear systems. use radial basis function neural network (RBFNN) to
They are able to adapt controller parameters automatically approximate system Jacobian for a turbogenerator system
[4]. This class of regulators is generally comprised of three since the convergence for RBFNN is faster than multilay-
parts: a model estimator, a controller and a block in which ered feedforward neural networks (MLP), and RBF net-
the controller parameters are determined from the esti- works can be trained much more rapidly and conveniently
mated model parameters [5]. The regulator has two loops than MLPs. Abdullah et al. [22] utilized self-tuning pole-
called as inner and outer loops. The inner loop is composed zero placement controller for nonlinear unstable systems
of the system to be controlled and an ordinary feedback based on RBFNN. The nonlinear dynamics of the system is
controller [3]. The outer loop comprises a recursive model represented with a model including a simple linear time-
estimator and a design calculation to compute controller varying sub-model derived via recursive least squares
parameters [3]. For STR, controller design alternatives are algorithm and a nonlinear sub-model identified using
very rich since it is possible to utilize various types of RBFNN. Wahyudi et al. [23] deployed an RBFNN
controllers and parameter estimators, by combining the parameter estimator trained with extended minimal
powerful features of these components, more flexible and resource allocation algorithm (EMRAN) which is a
robust adaptive controllers can be successfully achieved. sequential learning technique and extended Kalman filter
For example, by combining the nonlinear function (EKF) to directly estimate the parameters of a PID con-
approximation ability of artificial neural networks (ANN) troller. Firstly, nominal values of PID parameters are
and robustness of PID controllers, PID-type ANN con- obtained with a standard controller design method, and
trollers can be designed to effectively control nonlinear then these parameters are tuned via RBF parameter esti-
systems [6, 7]. mator that is sequentially trained to compensate for system
Model structure selection and its parametrization are parameter variations. Guo and Yang [24] adapted genetic
significant issues for closed-loop control performance in algorithm (GA) to optimize the initial weights of an ANN
self-tuning regulators [2]. In model-based control (MBC) parameter estimator to forecast the parameters of a PID
methodologies, controller performance is influenced by controller for a hydro-turbine governor system. System
modeling inaccuracies. Although ANNs and adaptive Jacobian which is required to adjust the parameters of the
neuro-fuzzy inference systems (ANFIS) have been suc- ANN parameter estimator is approximated via a second
cessfully employed in the identification and control of ANN. Kang et al. [25] employed an ANN controller for
numerous nonlinear systems [8–12], their functionalities speed control of a servo motor without using system model.
are impressed by local minimas resulting from non-convex In order to tune controller parameters, a linear combination

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S777

of the tracking error and its derivative is deployed in place parameters of a fuzzy PID controller to control an electrical
of system Jacobian. Pham and Karaboga [26] utilized a DC drive. In fuzzy logic-based control methodology, fuzzy
recurrent neural network (RNN) system model trained with rule base generally depends on the system to be controlled
GA for fuzzy STR to control linear and nonlinear systems. and the type of the controller to be implemented, so fuzzy
Initially, a RNN model of the system is determined by rule base is established by intuition or practical experience
offline training with a set of input output data pairs col- [29]. In order to obtain a fuzzy system with suboptimal/
lected from the system. Then, the RNN model of the sys- optimal parameters, ANFIS-based STRs combining the
tem is gradually improved during online control [26]. learning ability of ANNs with reasoning feature of fuzzy
Fuzzy systems have frequently been employed to con- systems have been designed as in [35–37]. When back-
stitute adaptive mechanisms for controllers. Fuzzy logic- propogation algorithm is used to train ANFIS directly as a
based control methods can be categorized mainly into two parameter estimator, the model of the controlled system is
classes, methods in the first category employ fuzzy estima- needed. Li and Priemer [35] employed a modified random
tors to tune the parameters of conventional controllers as in optimization method to train a neural network-based fuzzy
[27–29]. Methods in the second category use fuzzy logic logic parameter estimator without requiring model of the
controllers with tunable parameters which are updated via system being controlled. Bishr et al. [36] developed an
self-tuning algorithms as in [30, 31]. He et al. [27] reduced online training algorithm for ANFIS to estimate the
the three parameters of the PID controller to a single parameters of a self-tuning PID controller. Lu et al. [37] used
unknown variable inspired by Ziegler–Nichols formula and a wavelet type-2 fuzzy neural network system model for
considered a fuzzy adaptation mechanism to estimate this self-tuning predictive PID controllers to control liquid-level
new single parameter. Gautam and Ha [28] proposed a fuzzy and heating processes. The parameters of the PIDs are
self-tuning estimator for PID parameters to control a updated using gradient descent method and the system
quadrotor. They used EKF algorithm to update the param- Jacobians are approximated via wavelet type-2 fuzzy neural
eters of the fuzzy estimator. Ahn et al. [29] approximated the network system model.
parameters of a PID controller via three separate fuzzy In this paper, a generalized self-tuning regulator based
estimators. The tunable parameters of the fuzzy estimator on SVR methodology is proposed to control nonlinear
are adjusted by back-propagation algorithm. Evolutionary dynamic systems. The main contribution of the paper is
algorithms have also been used for finding the optimal utilizing an online SVR to approximate the optimal
parameters of fuzzy inference mechanisms [32, 33]. parameter values of a self-tuning regulator. For this pur-
Bandyopadhyay et al. [32] deployed a fuzzy-genetic pose, the ‘‘closed-loop margin’’ notion proposed in [38] has
approach to tune the parameters of a self-tuning PID con- been expanded to design STRs and online SVR update
troller. The adjustable parameters of the PID controller are equations are derived. The proposed mechanism is used to
reduced to a single parameter using dead-beat control. The optimize the parameters of two different type of con-
controller parameter is predicted by a fuzzy inference trollers, namely PID and fuzzy PID controllers. Another
mechanism and the rule base of fuzzy model is optimized via contribution of the paper, unlike the existing research in
genetic algorithm. Sharkawy [33] applied three independent literature, is using online learning method for estimating
fuzzy parameter estimator mechanisms to tune the param- the system model. Stability of the closed-loop system has
eters of an incremental PID controller. Each PID parameter also been analyzed. The performance of the proposed
is tuned with a first order Takagi-Sugeno (TS) fuzzy infer- generalized STRs has been examined on a nonlinear
ence system, whose parameters are optimally determined bioreactor system, and the performance of the generalized
offline using a modified genetic algorithm (GA) [33]. Qiao STRs has been compared with SVM-based PID controller
and Mizumato [30] proposed a peak observer-based tuning proposed by Iplikci [19]. The results show that the pro-
mechanism to adjust the scaling coefficients of a fuzzy PID posed generalized STR structure and online SVR model
controller. The mechanism updates the scaling coefficients attain good modeling and control performances.
when the system output has a peak point. Since the coeffi- The organization of the paper is presented as follows:
cients are not adjusted up to another peak, the proposed Sect. 2 describes the basic principles of online SVR.
tuning mechanism has a limited field of use and is practical Construction of optimization problem so as to utilize SVR
only for step type reference signals. In order to overcome directly as an adaptive parameter estimator and the pro-
drawbacks of the peak observer-based tuning mechanism, posed STRs are explained in detail in Sect. 3. Additionally,
Woo et al. [31] proposed to tune scaling coefficients of a the stability analysis of the closed-loop system is per-
fuzzy controller using a function of tracking error during the formed. In Sect. 4, the performance of the proposed
course of control. The controller parameters are successfully mechanism is simulated and also, the performance of the
adapted even the system has no overshoot. Bouallègue et al. proposed method is compared with a SVM-based PID
[34] utilized particle swarm optimization to adjust the controller. The study is briefly concluded in Sect. 5.

123
S778 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

2 Online support vector regression 1 XN

LPr ¼ kwk2 þ C ðni þ nH
i Þ
2 i¼1
This section briefly reviews online support vector regres-
sion. The basic principles of support vector regression and X
N
 bi ðe þ ni  yi þ hw; Uðxi Þi þ bÞ
online learning method are presented in Sects. 2.1 and 2.2, i¼1
respectively. ð5Þ
X
N
 bH
i ðe þ nH
i þ yi  hw; Uðxi Þi  bÞ
2.1 An overview of support vector regression i¼1
X
N
H
Consider the input–output training instances  ðgi ni þ gH
i ni Þ
i¼1
T ¼ fxi ; yi gNi¼1 x i 2 X  Rn ; y i 2 R ð1Þ
The optimality conditions for Lagrangian in (5) are
where N and n indicate the number of the training samples acquired as:
and the dimension of the input samples, respectively. SVR XN
oLPr
model (2) can be deployed in order to capture the con- ¼ 0 ! w  bi hw; Uðxi Þi ¼ 0 ð6Þ
nection among input–output instances in (1). ow i¼1

y^i ¼ hw; Uðxi Þi þ b; i ¼ 1; 2; . . .; N ð2Þ oLPr XN

¼ 0 ! ðbi  bH
i Þ ¼0 ð7Þ
ob i¼1
where ‘‘w’’ represents the weights of the SVR network
in feature space (F), ‘‘Uðxi Þ’’ is the projection of the oLPr
¼ 0 ! C  bi  gi ¼ 0; i ¼ 1; 2; . . .N ð8Þ
input samples to feature space, ‘‘b’’ typifies the bias of oni
regressor and h:; :i is inner product in F [39]. The
oLPr
essence of the optimization problem for support vector ¼ 0 ! C  bH H
i  gi ¼ 0; i ¼ 1; 2; . . .N ð9Þ
machines (SVM) is based on finding the optimal sepa- onH
i

rator or regressor. In classification, the separator that Thus, substituting (6)–(9) in (5), dual form of the opti-
maximizes the margin between two different classes is mization problem is acquired in (10)–(11) as a quadratic
searched. Similarly, in regression, the aim is to find the programming (QP) problem:
optimal regressor within a predefined margin via e-in-
sensitive loss function. The primal form for optimization 1X N X N
JD ¼ ðb  bH H
i Þðbj  bj ÞKij
problem is formulated using e-insensitive loss function 2 i¼1 j¼1 i
as: ð10Þ
X
N X
N

XN þe ðbi þ bH
i Þ  yi ðbi  bH
i Þ
1
min JPr ¼ kwk2 þ C ðni þ nH
i Þ ð3Þ i¼1 i¼1
w;b;n;nH 2 i¼1
with the following constraints
with the following constraints
0  bi  C
yi  hw; Uðxi Þi  b  e þ ni 0  bH
i C
ð11Þ
hw; Uðxi Þi þ b  yi  e þ nH
i ð4Þ X
N
ðbi  bH
i Þ ¼ 0; i ¼ 1; 2; . . .N
ni ; n H
i  0 ; i ¼ 1; 2; . . .N i¼1

where e is the upper value of tolerable error, n’s and nH ’s where Kij ¼ Uðxi ÞT Uðxi Þ. Thus, the optimization problem
denote the deviation from e tube and called as slack vari- is degraded to a problem with single type unknown
ables [13, 39]. The primal form has non-convex objective parameter (b). Various QP algorithms can be implemented
function and the solution may get stuck at local minima. to obtain Lagrange multipliers (b). The solution of the
The dual form for the optimization problem can be regression problem in (2) can be approximated as in (12):
obtained using Lagrangian multiplier method. Thus, the X
problem in (3)–(4) can be rendered to convex problem. For y^ðxÞ ¼ ki Kðx; xi Þ þ b ; ki ¼ bi  bH
i ð12Þ
i2SV
this purpose, a Lagrange function is derived by introducing
nonnegative Lagrange multipliers b, bH , g and gH as dual where ‘‘SV’’ emblematizes support vectors. This method
variables to compound objective function in (3) and con- computes the solution offline, however online learning
straints in (4) as follows: algorithms for support vector regression can also be

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S779

derived via the objective function and the constraints given Note that Fig. 1 shows how the sample in R immigrates to
in (10), (11). class S and newly learned sample enters into class E. In
online learning, the optimal regression surface transumes
2.2 Online e-support vector regression when a new data is transcluded to training samples or a
formerly trained instance is forgotten. In order to provide
Notion of ‘‘margin’’ is the key to fully comprehend the optimal representation of all existing instances by the
fundamental idea of online SVR. Let us predefine an error regressor, it is required to adjust the parameters of the
margin function as: previous model by ensuring the KKT conditions for
X
N training or forgetting phases. Derivation of online learning
hðxi Þ ¼ f ðxi Þ  yi ¼ kj Kij þ b  yi ð13Þ rules necessitates a Lagrange function which is a combi-
j¼1 nation of a dual objective function and corresponding
constraints. The Lagrange function for dual formulation
where hðxi Þ is the margin value, f ðxi Þ is the output of
can be expressed as follows via (10), (11).
regressor and yi is the actual output corresponding to input
xi from the data set. The training samples are separated into 1X N X N XN

error support (E), margin support (S) and remaining sup- LD ¼ ðbi  bH
i Þðb j  b H
j ÞKij þ e ðbi þ bH
i Þ
2 i¼1 j¼1 i¼1
port (R) subsets depending on their locations with respect
to the margin function and Lagrange multipliers as follows: X
N X
N
 yi ðbi  bH
i Þ ðdi bi þ dH H
i bi Þ
E ¼ fi j jki j ¼ C; jhðxi Þj  eg i¼1 i¼1

S ¼ fi j 0\jki j\C; jhðxi Þj ¼ eg ð14Þ XN X

N
H
 ui ðC  bi Þ þ uH
i ðC  bi Þ þ z ðbi  bH
i Þ
R ¼ fi j jki j ¼ 0; jhðxi Þj  eg i¼1 i¼1
ð16Þ
When a new sample xc is imbibed by the regressor, it is
required to include the new sample xc into one of these KKT optimality conditions for dual Lagrangian function
subsets (E, S, R) depending on the margin and Lagrange in (16) are derived in (17) via dual variables:
multiplier value of the new sample. During this learning
oLD X N
process, KKT conditions must be satisfied automatically ¼ ðbj  bH
j ÞKij þ e  yi  di þ ui þ z ¼ 0
for all training instances [18]. Assuming that the Lagrange obi j¼1
multiplier of the new added data is kc ¼ 0, from (13), its oLD XN
margin value is acquired as: ¼  ðbj  bH H H
j ÞKij þ e þ yi  di þ ui  z ¼ 0
obH
i j¼1
X
N ðHÞ ðHÞ ðHÞ ðHÞ ðHÞ ðHÞ
hðxc Þ ¼ f ðxc Þ  yc ¼ kj Kðxj ; xc Þ þ b  yc ð15Þ di  0; ui  0; di bi ¼ 0; ui ðC  bi Þ ¼ 0
j¼1 ð17Þ
The Lagrange value of current data (kc ) and Lagrange ðHÞ
where superscript di represents both di and dH i . KKT
values of previously added samples are gradually updated H
condition indicates that at most one of bi or bi should be
to provide all samples satisfy KKT conditions. As a result
nonzero and both are nonnegative [40]. The margin for the
of the adjustment process, current data move into one of
ith sample xi can be designated with (18):
the three sets (E, S, R) and the sets of the previously
learned samples may change since Lagrange multiplier (ki ) X
N

and margin values of the previously learned samples (hðxi ) hðxi Þ ¼ f ðxi Þ  yi ¼ kj Kij þ b  yi ð18Þ
j¼1
may alter because of admittance of the new data into the
regression problem. Thus, the new data is successfully The convergence and migration of data in learning or for-
ingested by the regressor. This situation is illustrated in getting phases occur according to the following conditions:
Fig. 1. Figure 1a, b depicts the margin before and after
training, respectively. As seen in Fig. 1a, the regressor hðxi Þ  e; ki ¼ C
cannot predict correctly for current data since it actually hðxi Þ ¼ e; C\ki \0
belongs to set E but the regressor result gives it as R for  e  hðxi Þ  e; ki ¼ 0 ð19Þ
initial value of kc . For this reason, the Lagrange multiplier hðxi Þ ¼ e; 0\ki \C
of the current data has to be adjusted. The incorporation of
hðxi Þ   e; ki ¼ C
the newly added data changes the structure of the regres-
sion problem. All Lagrange multipliers are adjusted to The variation on margin function values of previously
yield low prediction error and correct classification of data. learned samples (Dhðxi Þ) are derived as in (20) via (13)–

123
S780 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

Fig. 1 E, S and R subsets (a) Y (b) Y

before (a) and after (b) training
[13, 18, 38, 40]

f ( x) + ε
E E
f ( x) h( x )
f ( x) − ε f ( x) + ε
R S
S f ( x) h( x )
yˆ c = f ( xc ) output of svr S f ( x) − ε
h ( xc ) yˆ c = f ( xc )
yc new sample h ( xc )
yc

x x
xc xc

(19) depending on Lagrange multiplier of the current 2 3

Db
sample (Dkc ) and Db [40, 41]. 6 Dks 7
6 1 7

X
N Dk ¼ 6 7
6 .. 7 ¼ bDkc ð24Þ
Dhðxi Þ ¼ Kic Dkc þ Kij Dkj þ Db ð20Þ 4 . 5
j¼1 Dksk
The new added data has to satisfy the dual constraints in where
(11) at every parameter update step, so 2 3 2 3 2 31
b 1 0 1  1
X
N 6 7 6 Ks c 7 61
kc þ kj ¼ 0 ð21Þ 6 bs1 7 6 1 7 6 Ks1 s1  Ks1 sk 7
7
6 7
j¼1
7
b ¼ 6 . 7 ¼ H6 . 7H ¼ 6
6
6 .. .. .. .. 7
7
6 .. 7 4 .. 5 4. . . . 5
4 5
is extracted. As given in (14), Lagrange multiplier values of the bs k Ksk c 1 Ksk s1  Ksk sk
vectors belonging to subsets E or R are equal to ‘‘0’’ or ‘‘C’’.
The migration between subsets especially affects the Lagrange ð25Þ
values of vectors in S. If a sample which belongs to set S re- as given in [40]. Thus, the feasible increment directions
mains in set S again, there is no change on margin values of for the bias and the Lagrange multipliers of the samples
mentioned sample, that is Dhðxi Þ ¼ 0; i 2 S [41]. Thus, rela- in S can be obtained for a given Dkc using (23)–(25). The
tion between increments of current data (Dkc ) and parameters derivation and calculation of the Lagrange multiplier of
of the previously obtained model can be formulated: current sample (Dkc ) is detailed in [38]. The variation in
X
N margin values as a result of increment Dkc for non-sup-
Kij Dkj þ Db ¼ Kic Dkc port samples can be calculated as follows using (18), (20),
j¼1 ð22Þ (24):
X
Dkj ¼ Dkc 2 3 2 3 2 3
j2SV Dhðxz1 Þ Kz1 c 1 Kz1 s1  K z1 sl
6 Dhðxz Þ 7 6 Kz c 7 6 1 Kz s  K z2 sl 7
6 2 7 6 2 7 6 2 1 7
and in matrix form is given as 6 7 ¼ cDkc ; c ¼ 6 7 6 7
6 .. 7 6 .. 7 þ 6 .. .. .. .. 7b
2 32 3 2 3 4 . 5 4 . 5 4. . . . 5
0 1  1 Db 1
Dhðxzm Þ Kzm c 1 Kzm s1  Kzm sl
6 1 Ks s    Ks s 76 Dks 7 6 Ks c 7
6 1 1 1 k 76 1 7 6 1 7 ð26Þ
6 76 7 ¼ 6 7 ð23Þ
6 .. .. .. .. 76 .. 7 6 .. 7Dkc
4. . . . 54 . 5 4 . 5
where z1 ; z2 ; . . .; zm are the indices of non-support samples,
1 Ksk s1  Ksk sk Dksk Ksk c c are margin sensitivities and c ¼ 0 for samples in S. The
alternation of the matrix H for learning and forgetting
where the indices of the samples in support vector set are
stages and detailed information related to recursive algo-
defined as S ¼ fs1 ; s2 ; s3 ; . . .; sk g. As a consequence, Dk is
rithm can be attained via [18, 38, 40, 41].
attained as:

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S781

3 Contruction of optimization problem 3.2 Generalized STR structure based on SVR

for self-tuning regulator
The tuning mechanism of the proposed STR architecture
3.1 An overview of self-tuning regulators based on online SVR is depicted in Fig. 3. There are two
separate SVR structures in the proposed mechanism:
The main components of a self-tuning regulator (STR) are SVRestimator to calculate the controller parameters and
system model, parameter estimator and controller blocks as SVRmodel which predicts the future behavior of the con-
given in Fig. 2 where h and Xc indicate the controller trolled system. Since SVR has multi-input–single-output
parameters and input vector of the controller, respectively. (MISO) structure, a separate SVRestimator structure is
In order to minimize tracking error and estimate feasible deployed for each approximated parameter. Therefore, the
controller parameters, the future behavior of the system is number of the SVRestimator structures to be used in param-
required, so system model block is essential to approximate eter estimator block depends on the number of the
the dynamics of the system. Parameter estimator block adjustable parameters in the controller. For instance, three
computes new controller parameters by regarding the SVRestimator structures are employed for PID-type STR to
future behavior of the plant via the obtained system model, forecast Kp , Ki and Kd parameters.
and then adjusted controller parameters are implemented in The controller parameters are estimated via SVRestimator
the controller block to make system track reference signal as:
accurately. Any controller with adjustable parameters can hm ¼ festimatorm ðPmc Þ
be utilized in the generalized controller block given in X
Fig. 2. In this work, the proposed STR structure is imple- ¼ amk Kestimatorm ðPmc ; Pmk Þ þ bestimatorm ;
k2SVestimatorm
mented with PID and fuzzy PID controllers as explained in
detail below. Depending on the controlled systems and m 2 f1; 2; . . .pg ð27Þ
design techniques, numerous self-tuning architectures are
where Pmc indicates the current input of mth estimator,
possible in the parameter estimation block [5]. As for the
Kestimatorm ð:; :Þ is the kernel matrix, amk , Pmk and bestimatorm
system model part, various intelligent modeling techniques
are the parameters of the mth parameter estimator,
such as ANN [8, 9, 42], ANFIS [12, 43] etc have been
festimatorm ð:; :Þ is the regression function to be optimized in
applied to identify the dynamics of system. In our con-
training. The controller computes a control signal as:
troller structure, SVR is employed to model the dynamics
of the controlled system since it has high generalization un ¼ gcontroller ð½h1 ðP1c Þ; . . .; hm ðPmc Þ; Xc Þ ð28Þ
capacity and ensures global minimum in training. Subse-
where gcontroller indicates the control law computed as the
quently, another SVR is used as parameter estimator to
output of the controller, hm is the mth parameter of con-
approximate controller parameters.
troller and Xc is the current input vector of the controller.

Fig. 2 Self-tuning regulator Controller Parameters

Adjustment
Mechanism
θ̂

Parameter
Estimator

ym

System Model

r
u y
Generalized Controller
System
u = f c (θ , Χ c )

123
S782 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

Fig. 3 Generalized self-tuning regulator based on online SVR

SVRmodel is employed to forecast system behavior and it algorithm for the proposed adaptive control architecture is
calculates the model output as given in Sect. 3.5.
y^nþ1 ¼ fmodel ðMc Þ
X 3.3 PID-type STR based on SVR
¼ kj Kmodel ðMc ; Mj Þ þ bmodel ð29Þ
j2SVmodel PID controller still predominates in the process industries
where fmodel and Kmodel are the regression function and the due to its robustness, effectiveness for wide range of
kernel matrix of the system model, respectively, Mj ’s are operating conditions and its functional simplicity [6]. It has
support vectors, Mc is current input, and kj and bmodel are been widely used in industry due to its simplicity, good
the parameters of the system model to be adjusted. control performance and excellent robustness to uncer-
SVRestimator and SVRmodel are both used online to perform tainties [44]. The classical incremental PID controller
learning, prediction and control consecutively. Ideally, produces a control signal as follows [1, 17, 19, 44–46]:
during the course of online working, it is expected that un ¼ un1 þ Dun
y^nþ1 will eventually converge to ynþ1 . Therefore, when Dun ¼ Kpn ½en  en1  þ Kin ½en  ð30Þ
the parameters of SVRestimator are optimized, in order to
þ Kdn ½en  2en1 þ en2 
calculate and observe the impact of the computed control
signal (un ) on system behavior and train SVRestimator where Kpn , Kin and Kdn respectively indicate the parameters of
precisely, un is applied to SVRmodel at every step of proportional, integral, and derivative parts of the controller to
training phase of parameter estimator to predict system be tuned. For the proposed mechanism given in Sect. 3.2, the
behavior (ynþ1 ). The control signal applied to the real incremental PID control law can be extracted as:
system is obtained via the trained parameter estimator and
the actual output ynþ1 is determined after applying the un ¼ gcontroller ðh; Xc Þ ¼ un1 þ hT Xc
2 3
calculated control signal to the system. Thus, the current en  en1
6 7 ð31Þ
input of system model Mc and output ynþ1 can be com- ¼ un1 þ ½ Kpn Kin Kdn 4 en 5
puted for training phase of SVRmodel . The detailed en  2en1 þ en2

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S783

In an adaptive control scheme, the initially assigned values where the scaling factors Ken , Kden , Wn and bn are the
of the controller parameters will generally not be optimal controller parameters to be optimized. Wn and bn are the
[19], hence, it is required to adjust the parameters via parameters for the PD and PI parts of the fuzzy PID con-
optimization methods [19, 47]. The controller parameters troller, fFLCPD is the fuzzy controller, efn and e_fn are scaled
Kp , Ki and Kd are calculated via online SVR parameter tracking error and derivative of tracking error, respectively.
estimator (SVRestimator ). For this purpose, an online SVR The derivative and integral parts of the fuzzy controller can
has been utilized for each controller parameter since SVR be associated depending on the requirements of the con-
has multi-input–single-output structure, so parameter esti- trolled system via input–output scaling coefficients. In our
mator is composed of three separate SVR identifiers. The simulations, triangular membership functions with cores
PID controller parameters are estimated via SVRestimator as: f1; 04; 0; 0:4; 1g as in [30] are chosen for both efn and
2 3 2 3 e_fn as shown in Fig. 5 where Czn denotes the zth-fired fuzzy
K^pn festimatorp ðPpc Þ
6 7 6 7 rule. The formulation of fired fuzzy rule is given as:
h ¼ 4 K^in 5 ¼ 4 festimatori ðPic Þ 5
K^d festimatord ðPdc Þ Czn ¼ frules ðefn ; e_fn ; kz1 ; kz2 Þ ¼ Sz þ Pz  1 ; z 2 f1; 2; 3; 4g
2 Pn 3 ð32Þ
k2SVestimatorp apk Kestimatorp ðPpc ; Ppk Þ þ bestimatorp ð35Þ
6 P 7
¼64 P k2SVestimatori aik Kestimatori ðPic ; Pik Þ þ bestimatori 5
7
where
k2SVestimator adk Kestimatord ðPdc ; Pdk Þ þ bestimatord
d
1 1
Sz ¼ k
; Pz ¼
Since the parameters of the controller are estimated via 1þe z1 e f n 1 þ ekz2 e_fn
SVR, it is named as ‘‘PID-type STR based on SVR’’. This
and k11 ¼ k12 ¼ 4, k21 ¼ k22 ¼ 5, k31 ¼ k32 ¼ 6,
structure inherits both the robustness of PID controllers and
k41 ¼ k42 ¼ 7. The fuzzy rule surface depicted in Fig. 6 is
the nonlinear generalization performance of SVR method.
composed of two sigmoid functions as in (35). Center of
gravity method has been utilized for defuzzification. Thus,
3.4 Fuzzy PID-type STR based on SVR
the output of the fuzzy controller can be computed as [30]
P4
In Fig. 4, the structure of an incremental fuzzy PID con- X4 X4
z¼1 wzn Czn
troller [34] is depictured, where en denotes the difference fFLCPD ðefn ; e_fn Þ ¼ P 4
¼ w C
zn zn ; wzn ¼ 1
z¼1 wzn z¼1 z¼1
between reference signal and system output and uFLCPIDn is
ð36Þ
the output of fuzzy PID controller at time index n. A fuzzy
PID controller is also implemented as the generalized where
controller in Fig. 3 and its parameters are tuned using an
w1n ¼ Ai ðefn ÞBj ðe_fn Þ
online SVRestimator . Inputs of the fuzzy controller are:
w2n ¼ Aiþ1 ðefn ÞBj ðe_fn Þ
efn ¼ Ken en
ð33Þ w3n ¼ Ai ðefn ÞBjþ1 ðe_fn Þ
e_fn ¼ Kden ½en  en1 
w4n ¼ Aiþ1 ðefn ÞBjþ1 ðe_fn Þ
The output of the controller in Fig. 4 is computed as
[30, 31]: and
eiþ1  efn ef  ei
uFLCPDn ¼ Wn fFLCPD ðefn ; e_fn Þ Ai ðefn Þ ¼ ; Aiþ1 ðefn Þ ¼ n
eiþ1  ei eiþ1  ei
DuFLCPIn ¼ bn fFLCPD ðefn ; e_fn Þ e_jþ1  e_fn e_f  e_j
uFLCPIn ¼ uFLCPIn1 þ DuFLCPIn ð34Þ Bj ðe_fn Þ ¼ ; Bjþ1 ðe_fn Þ ¼ n
e_jþ1  e_j e_jþ1  e_j
uFLCPIDn ¼ fFLCPID ðefn ; e_fn ; Wn ; bn Þ
are membership functions values. C ¼ ½C1n C2n C3n C4n  ¼
¼ uFLCPIn þ uFLCPDn ½uij uiþ1j uijþ1 uiþ1jþ1  indicate the fired rules in rule

Fig. 4 Fuzzy PID controller

123
S784 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

Fig. 5 The membership

functions for inputs and rule −1 ei −2 Ai −2
base

−0.4 ei −1 Ai −1

Γ1n = uij Γ2n = uij +1

0 ei Ai
e fn
Γ3n = ui +1 j Γ 4n = ui +1 j +1
0.4 ei +1 Ai +1

1 ei +2 Ai + 2
B j −2 B j −1 Bj B j +1 B j+2

e j −2 e j −1 ej e j +1 e j +2
−1 −0.4 0 0.4 1

e fn

2 ^ 3 2 3
Fuzzy Control Surface K en festimatore ðPec Þ
6 K^ 7 6 festimator ðPdec Þ 7
6 den 7 6 de 7
0.8 h¼6 7¼6 7
4W ^ n 5 4 festimatorW ðPWc Þ 5
1 0.6
b^n festimatorb ðPbc Þ
0.4 2 P 3
k2SVestimatore aek Kestimatore ðPec ; Pek Þ þ bestimatore
0.5
0.2 6P 7
u 6 k2SVestimator adek Kestimatorde ðPdec ; Pdek Þ þ bestimatorde 7
0 6P de 7
0 ¼6 7
6 k2SVestimator aWk KestimatorW ðPWc ; PWk Þ þ bestimatorW 7
-0.5 -0.2
4 P W 5
k2SVestimator abk Kestimatorb ðPbc ; Pbk Þ þ bestimatorb
b
-0.4
-1
1 ð37Þ
-0.6
0.5 1
0 0.5 -0.8
Thus, the control signal in (33), (34) is obtained as:
e fn -0.5
-0.5
0
e fn uFLCPIDn ¼ gcontroller ðh; Xc Þ ¼ fFLCPID ðh; Xc Þ
-1 -1
¼ uFLCPIn1 þ ðWn þ bn ÞfFLCPD ðKen x1 ; Kden x2 Þ
Fig. 6 Fuzzy control surface
   
x1 en
Xc ¼ ¼
x2 en  en1
surface since four fuzzy rule are fired at a time depending ð38Þ
on the defined input membership functions. The parameter
estimator block tunes the controller vector h that consists of where h indicates the controller parameters and Xc is the
the Ken , Kden , Wn and bn . The controller parameters are input of the controller. By tuning fuzzy PID parameters
estimated as: using online SVRestimator , the strong characteristics of fuzzy

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S785

control technique and SVR methodology are merged to else

build a powerful controller for nonlinear systems. Continue with parameter estimator obtained at previous
step
3.5 Adaptive control algorithm for the generalized end
STR based on SVR Step 5: Prediction step for trained parameter esti-
mator (hþ m ) and computation of control input by trained
Input feature vector of parameter estimator (P) should estimator (uþ n)
contain convenient feature variables that can well represent –Calculate the controller parameters by trained
the closed-loop system’s operating conditions. In the pro- SVRestimator via (27).
posed STR, mainly reference signal and system output are –Calculate the control signal uþ n produced by the con-
utilized as input features. However, in order to enhance troller using the parameters obtained by trained SVRestimator
STR performance, the variables that are functions of ref- via (27)–(28).
erence and system output such as tracking error, integral of Step 6: Application of the control signal produced by
tracking error, derivative of tracking error, and control adapted controller
signal etc. have also been employed as described in Sect. 4. –Apply uþ n to system to calculate ynþ1 .
The control procedure for STR with ‘‘p’’ adjustable con- Step 7: Prediction and training step for system model
troller parameters can be summarized as follows (in the yþ
(^nþ1 )
algorithm given below u n indicates the control signal –Apply uþ n to SVRmodel and calculate y^þ nþ1 via (29).
predicted with controller parameters obtained at the pre-
–Calculate emodelnþ1 ¼ ynþ1  y^nþ1
vious step and uþ n stands for the control signal estimated If jemodelnþ1 j [ emodel
with trained controller parameters at the current step):
Train system model where emodelnþ1 ¼ ynþ1  y^nþ1
Step 1: Initialization of SVR estimator and SVR model
else
parameters.
Continue with system model parameters obtained at
–SVRestimator (estimator) parameters: amk ¼ bestimatorm ¼
previous step
0 ; m 2 f1; 2; . . .pg
end
–SVRmodel (system model) parameters: kj ¼ bmodel ¼ 0
Step 8: Incrementation of time step
Step 2: Prediction step for parameter estimator (h m) –Increment n ¼ n þ 1 and back to step 2.
–Set time step n.
–Constitute feature vector for parameter estimator
3.6 Generalized closed-loop system margin
(Pmc ).
Some examples for parameter estimator feature vector
In the proposed generalized adaptive STR architecture, two
are given as follows:
blocks contain separate online SVRs: In the system model
Pc ¼ ½rn . . .rnnr ; yn . . .ynny 
block SVRmodel computes an estimate of the system model
Pc ¼ ½Pn ; In ; Dn  where Pn ¼ en  en1 ,In ¼ en , Dn ¼ and in the parameter estimator block each tunable param-
en  2en1 þ en2 and en ¼ rn  yn . Combination of the eter of the controller is estimated by a different
reference signal, system output, and controller output can SVRestimator . The training dataset for SVRmodel consists of
also be utilized in the feature vector. pairs (Mc ,ynþ1 ) which are available during online opera-
Pc ¼ ½Pn ; In ; Dn ; rn . . .rnnr ; yn . . .ynny ; un1 . . .unnu  tion, therefore the training process is straightforward as
where nr , ny and nu represent the number of the past explained in Sect. 2.2. However, the training of SVRestimator
instances of features included in the feature vector. is not clear, since the input data (Pmc ) is known, but the
–Calculate the approximated controller parameters h m desired output of the estimator, namely the controller
by SVRestimator trained at previous step (n  1) via (27). parameters (h) to be implemented to produce control signal
Step 3: Computation of control signal (u n) and are not available to the designer in advance. This situation
prediction step for system model (^ y
nþ1 )) causes a significant dilemma to train SVR structures
–Calculate the control signal u n via (27)–(28). without the explicit information of desired output data.
–Constitute feature vector for system model (Mc ). However, a similar problem which hampers to utilize SVR
Mc ¼ ½u n . . .unnu ; yn . . .ynny  directly as a controller in control block was managed to
–Apply u n to SVRmodel and calculate y ^
nþ1 by (29). overcome in [38]. For this purpose, Uçak and Öke Günel
Step 4: Training step for parameter estimator proposed ‘‘closed-loop system margin’’ notion to solve this
–Calculate e^trnþ1 ¼ rnþ1  y^ nþ1 situation. In this section, since training SVRestimator without
If j^
etrnþ1 j [ eclosedloop a desired output dataset can be formulated in a similar way
Train parameter estimator via e^trnþ1 ¼ rnþ1  y^ nþ1
to the problem solved in [38], we will configure the

123
S786 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

(a) (b) (c)

Fig. 7 Margins of SVRestimator (a), adaptive controller (b) and SVRmodel (c)

‘‘closed-loop margin’’ notion so as to train SVRestimator estimators, but parameter estimator margins can only be
without the need to know the approximated controller affected indirectly through the combined actions of
parameters. Let us start by giving two main definitions. SVRmodel and the controlled closed-loop structure. We can
As defined in [38], the regression margin related to infer that optimization of closed-loop and SVRmodel mar-
SVRmodel is named as ‘‘open-loop margin.’’ This margin is gins will spontaneously lead to the optimization of
optimized by minimizing the feedforward modeling error SVRestimator s margins so SVRestimator s parameters can be
emodelnþ1 ¼ ynþ1  y^nþ1 , the error between the actual system obtained indirectly while we try to minimize tracking error
output and the output of the ‘‘learned model.’’ High mod- (or equivalently we optimize closed-loop margin). The
eling precision is needed to successfully identify the sys- margins of SVRestimator and SVRmodel are illustrated in
tem dynamics and compute the tracking error in the next Fig. 7. Parameter estimator SVRestimator margin is depicted
step. This information is used to tune the parameters of in Fig. 7a, adaptive controller is given in Fig. 7b, and
SVRestimator s which consequently compute the controller system model SVRmodel margin is shown in Fig. 7c where
parameters. Note that, we employ a separate SVRestimator festimatorm and fmodel denote the regression functions of mth
for each tunable controller parameter. Hence, the perfor- parameter estimator and system model, and gcontroller indi-
mances of SVRmodel and SVRestimator s are effective in the cates the control law. The input of SVRmodel is Mc and its
closed-loop success. The closed-loop performance of the output is ynþ1 while the input to SVRestimator is Pmc and the
overall system is affected by both the modeling error and output is hm , so in Fig. 7, the input–output axes of
the tracking error, so we define ‘‘closed-loop margin,’’ SVRmodel are denominated as M and Ysys while the axes for
which is a function of the system model margin and SVRestimator are named as Pm and hm . Considering the
parameter estimator margins, whereas it is defined as a margins of SVRmodel and the separate SVRestimator s for each
function of controller and system model margins in [38]. In tunable controller parameter, we can combine the inde-
tracking control, it is aimed to pursue reference signal as pendent subgraphs to yield a multidimensional graph
close as possible by minimizing the error which depicts the closed-loop system margin. In applica-
etrnþ1 ¼ rnþ1  ynþ1 , between reference input and closed- tions, there will generally be several tunable controller
loop output, so the closed-loop margin should be mini- parameters and input vectors will commonly be multi-di-
mized. The optimization of the SVRmodel margin is also mensional, resulting in hypersurfaces when margins are
important, since for good closed-loop control performance fused. As a representative graph in Fig. 8, we illustrate the
we need a system model with minimum error. emodel and case where all inputs and outputs are assumed to be one-
eclosedloop , the upper values of tolerable error for SVRmodel dimensional vectors, and there is only a single tunable
and the overall closed-loop system (SVRmodel and controller parameter, resulting in a hypercube. In Fig. 8,
SVRestimator s combined), respectively are set independently SVRestimator margin is drawn on the horizontal plane with
by the designer, so the margin of SVRmodel is optimized axes P and M while SVRmodel margin is given on the
independently from the closed-loop system margin. How- vertical plane, where axes M and Ysys represent its input
ever, the designer cannot set eestimator , the upper value of and output. Since vector M includes un and hm , the hori-
tolerable error for SVRestimator , and therefore does not have zontal regression surface representing SVRestimator can be
a direct influence on the margins of the parameter depicted with axes P and M instead of P and hm . During

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S787

Fig. 8 Closed-loop system

margin in three dimensions

(a) (b)

Fig. 9 Closed-loop margin before (a) and after (b) training

online operation of the whole control architecture, the rnþ1  ynþ1 is minimized. Based on this, we have used data
margins of the controller parameter estimators and model pairs (Pc ; rnþ1 ) in training. Hence, the input and output
are fused, and closed-loop margin between closed-loop axes for closed-loop system regression surface are termed
input and output is intuitively thought as a single margin, as P and R in Fig. 10 and axis R is utilized in place of Ysys
as depicted in Fig. 9. Here, controller parameter estimator for closed-loop system in Figs. 10 and 11.
and system model margins are combined resulting in the
‘‘closed-loop margin,’’ and this is projected onto P  Ysys
3.7 Online support vector regression for parameter
axes. Figure 9 illustrates this projection for a single con-
estimator
troller parameter estimator, before and after online training.
Closed-loop margin is represented with, hclosedloop ðPÞ,
Let the training data set used for the closed-loop system be:
and it is a function of model margin (hmodel ðMÞ) and
parameter estimator margin (hestimator ðPÞ). When training T ¼ fPmi ; riþ1 gNi¼1 Pmi 2 P  Rn ; riþ1 2 R ð39Þ
the closed-loop system, we require that the closed-loop
where N and n, respectively, indicate the number of the
output tracks the reference input, rnþ1 , and the error
training samples and the dimension of the input samples,
between closed-loop output and reference input etrnþ1 ¼
Pmi is input feature vector of mth parameter estimator and

123
S788 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

(a) (b)

Fig. 10 Closed-loop margin before (a) and after (b) training

Fig. 11 Closed-loop system

margin in three dimensions

riþ1 is the reference signal that system is required to track, Since parameters of SVRmodel are fixed and known in
the closed-loop margin function of the system for the ith training phase of SVRestimator , and the sole unknown vari-
sample Pmi can be defined as ables are the parameters of the SVRestimator , the closed-loop
margin can be rewritten as
hclosedloop ð½ P1i . . .Pmi Þ ¼ y^iþ1  riþ1
ð40Þ
¼ fmodel ðMi Þ  riþ1 hclosedloop ð½ P1i . . .Pmi Þ ¼ y^iþ1  riþ1
ð41Þ
¼ fclosedloop ð½ P1i . . .Pmi Þ  riþ1
where
X with respect to an input–output data pair of closed-loop
y^iþ1 ¼ fmodel ðMi Þ ¼ kj Kmodel ðMj ; Mi Þ þ bmodel
system (Pmi ; riþ1 ). Thus, using (Pmi ; riþ1 ) data pair and
j2SVmodel
closed-loop margin in (40), (41), the incremental learning
kj ¼ bj  bH
j algorithm for SVRestimator can be obtained. When new
Mi ¼ ½ui    uinu ; yi    yiny  sample Pmc is introduced, the coefficient amc corre-
ui ¼ gcontroller ð½ h1 ðP1i Þ    hm ðPmi Þ ; Xc Þ sponding this the new sample should be changed in a
finite number of discrete steps until it meets the KKT
hm ðPmi Þ ¼ festimatorm ðPmi Þ
X conditions while ensuring the existing samples in T
¼ amk Kestimatorm ðPmk ; Pmi Þ þ bestimatorm continue to satify the KKT conditions at each step [40].
k2SVestimatorm
The KKT conditions [38–40] that are fundamental in
amk ¼ gmk  gH mk ; m 2 f1; 2; . . .; pg convergence and migration of the closed-loop data are
Pmi ¼ ½ri . . .rinr ; yi . . .yiny ; ui1 . . .uinu  given as:

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S789

hclosedloop ð½ P1i . . .Pmi Þ  eclosedloop ; ai ¼ Cclosedloop where z1 ; z2 ; . . .; zr are the indices of non-support samples,
hclosedloop ð½ P1i . . .Pmi Þ ¼ eclosedloop ; Cclosedloop \ai \0 cm are margin sensitivities.
 eclosedloop  hclosedloop ð½ P1i . . .Pmi Þ  eclosedloop ; ai ¼ 0
3.8 Stability analysis of the closed-loop system
hclosedloop ð½ P1i . . .Pmi Þ ¼ eclosedloop ; 0\ai \Cclosedloop
hclosedloop ð½ P1i . . .Pmi Þ   eclosedloop ; ai ¼ Cclosedloop 3.8.1 Stability analysis for the generalized STR based
ð42Þ on SVR

The incremental algorithm for SVRestimator can be derived In this subsection, the Lyapunov stability analysis of the
by recasting Eqs. (18)–(26) using am , bestimatorm , hclosedloop , generalized STR based on SVR has been conducted. In
eclosedloop , Pmi and Kestimatorm in place of k, b, h, e, xi and order to clearly explain the stability analysis of the closed-
K. Thus, optimal parameters of SVRestimator , amk , bestimatorm loop system, firstly, the regression functions of estimator
are sought within eclosedloop tube by minimizing the and the system model are expressed in matrix or vector
tracking error via online learning algorithm given in form. Thus, the regression function of estimator can be
Sect. 2.2. Thus, the update direction vector for Lagrange expressed as
multipliers of support set samples in mth parameter esti- 2 3 2 3
bestimatorm T 1
mator Dam is attained as:
6 a 7 6 Km P 7
2 3 6 m1 7 6 1 mc 7
Dbestimatorm 6
hm ðPmc Þ ¼ festimatorm ðXm ; Pmc Þ ¼ 6 .. 7 6 7
7 6 .. 7
6 Das 7 4 . 5 4 . 5
6 1 7
Dam ¼ 6 6 ..
7 ¼ bm Damc
7 ð43Þ amk Kmk Pmc
4 . 5
¼ XTm Kestimatorm ðPmc Þ ; m 2 f1; 2. . .pg
Dask
ð46Þ
where
2 3 2 3 Using (28) and (46), the output of the controller is defined
b 1 as a function of X, P and X as follows:
6 7 6K 7
6 bs1 7 6 estimatorms1 c 7 un ¼ gcontroller ðX; P; XÞ ð47Þ
6 7 6 7
bm ¼ 6 . 7 ¼ Hm 6 .. 7;
6 .. 7 6 . 7
4 5 4 5 The regression function of the system model (29) which is
bsk Kestimatormsk c required to approximate system Jacobian in stability anal-
2 31 ð44Þ ysis is obtained as
0 1  1
61 K    Kestimatorms1 s 7
6 estimatorms1 s1 7 y^nþ1 ¼ fmodel ðk; Mc Þ
6 k
7 2 3 2 3
Hm ¼ 6 . . . . 7
6 .. .. .. .. 7 bmodel T 1
4 5 6 k 7 6 K1M 7
1 Kestimatormsk s1  Kestimatormsk sk 6 1 7 6 c 7 ð48Þ
¼6 7 6 7 T
6 .. 7 6 .. 7 ¼ k Kmodel ðMc Þ
4 . 5 4 . 5
The margin values of the non-support samples for Damc can
be calculated as follows: kk KkMc
2 3 In order to derive stability conditions, the following Lya-
Dhclosedloop ð½ P1z1  Pmz1 Þ
6 Dhclosedloop ð½ P1z  Pmz Þ 7 punov function is deployed
6 2 2 7
6 7 ¼ cm Dkc ; cm
6 .. 7 eTtrnþ1 P etrnþ1
4 . 5 Vðetrnþ1 Þ ¼ ð49Þ
2
Dhclosedloop ð½ P1zr  Pmzr Þ
2 3 2 3 where P ¼ I (identity matrix). Both the stability of the
Kestimatormz1 c 1 Kestimatormz1 s1  Kestimatormz1 sl
6K 7 61 K system and the convergence of the controller are guaran-
6 estimatormz2 c 7 6 estimatormz2 s1  Kestimatormz2 sl 7
7
6 7 6 7 teed when oV ot  0 [48]. Thus, the derivative of Lyapunov
¼6 7 þ 6 7 bm
6 ..
. 7 6 ... ..
.
..
.
..
. 7 oV
function ( ot ) is acquired as
4 5 4 5
Kestimatormzr c 1 Kestimatormzr s1  Kestimatormzr sl oVðetrnþ1 Þ
¼ eTtrnþ1 P e_trnþ1 ð50Þ
ð45Þ ot

123
S790 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

oetr oetr
where e_trnþ1 ¼ otnþ1 ¼ ounþ1 oun
ot . Considering a small devi-
Damc ¼ q minðjLm m Sm Em Rm
c1 j; jLc2 j; jL j; jL j; jL jÞ
n
ation from the equilibrium point, which corresponds to ¼ signðetrnþ1 Þ Wm ðetrnþ1 ; ami ; CÞ
local stability analysis using Eq. (51) [38, 48], the incre- etrnþ1 Wm ðetrnþ1 ; ami ; CÞ
mental change in the control signal (47) is obtained as ¼ Wm ðetrnþ1 ; ami ; CÞ ¼ etrnþ1
jetrnþ1 j jetrnþ1 j
p 
X 
ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ ¼ lm ðetrnþ1 ; ami ; CÞ etrnþ1
Dun ¼ DXm þ DPm
oXm oPm
m¼1 ð54Þ
i 
X 
ogcontroller ðX; P; XÞ
þ Dxk where lm ðetrnþ1 ; ami ; CÞ  0, Wm ðetrnþ1 ; ami ; CÞ  0. Thus,
k¼1
oxk
using (43) and (3.8.1) the adjustment rules for all Lagrange
ð51Þ multipliers in set Sm can be given as
where ‘‘p’’ indicates the number of the controller parame- DXm ¼ bm Damc ¼ bm lm ðetrnþ1 ; ami ; CÞ etrnþ1 ð55Þ
ters, ‘‘i’’ denotes the number of the inputs for the controller
oetrnþ1
and xk is the kth input of the controller. Substituting (51) in By substituting (55) in (52) and expanding as
oun
(50), the Eq. (50) can be rearranged as oetrnþ1 oynþ1
oynþ1 oun ¼  oyounþ1
n
, Eq. (52) can be finalized as follows:

2  3
Pp ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ
DXm þ DPm
oVðetrnþ1 Þ oetrnþ1 6
6
m¼1
oXm oPm 7
7
¼ eTtrnþ1 P 6   7 ð52Þ
ot oun 4 P ogcontroller ðX; P; XÞ 5
þ ik¼1 Dxk
oxk

with the assumption DPm ffi etrnþ1 and Dxk ffi etrnþ1 . Thus, as where oyounþ1
n
is the system Jacobian approximated via system
can be seen in (51)–(52), stability depends on the incre- model (fmodel ) and
ments in Lagrange multipliers of the estimator (DXm ). The

2  3
Pp ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ
bm lm ðetrnþ1 ; ami ; CÞ þ
oVðetrnþ1 Þ oetrnþ1 6
6
m¼1
oXm oPm 7
7
¼ eTtrnþ1 P 6   7etrnþ1
ot oun 4 Pi ogcontroller ðX; P; XÞ 5
þ k¼1 ð56Þ
oxk
oVðetrnþ1 Þ
¼ eTtrnþ1 ðQ þ W þ ZÞetrnþ1
ot

adjustment rule for DXm is derived in (43)–(44). The p  

oynþ1 X ogcontroller ðX; P; XÞ
Lagrange value for current data of the mth estimator is Q¼P bm lm ðetrnþ1 ; ami ; CÞ
oun m¼1 oXm
computed as
p  
oynþ1 X ogcontroller ðX; P; XÞ
Damc ¼ qm minðjLm m Sm Em Rm
ð53Þ W¼P
c1 j; jLc2 j; jL j; jL j; jL jÞ oun m¼1 oPm
 
where qm ¼ q ¼ signðhclosedloop ð½ P1i . . .Pmi Þ ¼ sign oynþ1 Xi
ogcontroller ðX; P; XÞ
Z¼P ð57Þ
ðetrnþ1 Þ and Lm m
c1 , Lc2 are variations of the current sample and oun k¼1 oxk
LSm ¼ ½LSi m ; i 2 Sm , LEm ¼ ½LEi m ; i 2 Em , LRm ¼ ½LR i ;i 2
m
In a nutshell, the stability conditions for closed-loop system
Rm  are the variations of the Pmi data in sets Sm , Em , Rm in the sense of Lyapunov can be attained as follows:
Sm Em Rm
respectively. The term minðjLm m
c1 j; jLc2 j; jL j; jL j; jL jÞ
in (53) is a positive function of etrnþ1 , ami and C. Therefore, • Condition 1: If VðtÞ  0 and oVðtÞ
ot  0, the stability of
Damc can be expressed as: the closed-loop system is ensured

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S791

• Condition 2: If Q  0, W  0 and Z  0, the stability of 3.8.2 Derivation of the sensitivity functions for PID-type
the closed-loop system is ensured STR based on SVR
• Condition 3: If Q  0, W  0 and Z  0 and
k Q þ W k  k Z k, the stability of the closed-loop In this subsection, the sensitivity of the control signal
system is ensured with respect to Lagrange multipliers of the parameters
• Condition 4: If Q  0, W  0 and Z  0 and estimator(ogcontroller ðX;P;XÞ
), the sensitivity of the control
oXm
k Q þ Z k  k W k, the stability of the closed-loop signal with respect to inputs of the parameters estima-
system is ensured ðX;P;XÞ
tor(ogcontroller ) and the sensitivity of the control signal
• Condition 5: If Q  0, W  0 and Z  0 and oPm

k Q k  k W þ Z k, the stability of the closed-loop with respect to inputs of the controller (ogcontrolleroxðX;P;XÞ
k
)
system is ensured have been derived for PID-type STR based on SVR. The
• Condition 6: If Q  0, W  0 and Z  0 and computation of ogcontroller ðX;P;XÞ
is as in (58).
oXm
k W þ Z k  k Q k, the stability of the closed-loop
system is ensured
• Condition 7: If Q  0, W  0 and Z  0 and
k W k  k Q þ Z k, the stability of the closed-loop
system is ensured

ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ ohm ofestimatorm ðXm ; Pmc Þ

¼
oXm ohm ofestimatorm ðXm ; Pmc Þ oXm
 
ogcontroller ðX; P; XÞ ofestimatorm ðXm ; Pmc Þ ofestimatorm ðXm ; Pmc Þ ofestimatorm ðXm ; Pmc Þ
¼ xm  ð58Þ
oXm obestimatorm oam1 oamk
ogcontroller ðX; P; XÞ
¼ xm Kestimatorm ðPmc Þ
oXm

ogcontroller ðX;P;XÞ
• Condition 8: If Q  0, W  0 and Z  0 and The sensitivity oPm is expressed as follows:
k Z k  k Q þ W k, the stability of the closed-loop
system is ensured

ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ ohm ofestimatorm ðXm ; Pmc Þ oKestimatorm ðPmc Þ
¼
oPm ohm ofestimatorm ðXm ; Pmc Þ oKestimatorm ðPmc Þ oPm
ð59Þ
ogcontroller ðX; P; XÞ oKestimatorm ðPmc Þ
¼ xm Xm
oPm oPm

In order to determine, Q, Z and W, it is required to com- ogcontroller ðX;P;XÞ

The sensitivity oxk is given as follows:
ðX;P;XÞ ogcontroller ðX;P;XÞ
pute ogcontrolleroXm , oPm and ogcontrolleroxðX;P;XÞ . The
k ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ ox1
derivations of ogcontroller ðX;P;XÞ
, ogcontroller ðX;P;XÞ
and ¼
oXm oPm oxk ox1 oxk
ogcontroller ðX;P;XÞ ogcontroller ðX; P; XÞ ox2
change depending on the controller to be ð60Þ
oxk þ
utilized. The derivations are given in Sects. 3.8.2 and 3.8.3 ox2 oxk
for PID-type STR and fuzzy PID-type STR respectively. ogcontroller ðX; P; XÞ ox3
þ
ox3 oxk

123
S792 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

Thus, the sensitivities for inputs can be obtained as ogcontroller ðX;P;XÞ ofFLCPID ðX;P;XÞ
The computation of ohm ¼ ohm is given
x1 ¼ en  en1 ¼ x2  en1 as
x2 ¼ en ¼ x2 ofFLCPID ðX; P; XÞ ofFLCPID ðX; P; XÞ ofFLCPD ðefn ; e_fn Þ oefn
¼
oKen ofFLCPD ðefn ; e_fn Þ oefn oKen
x3 ¼ en  2en1 þ en2 ¼ x1  2en1 þ en2 ¼ x2  en1 þ en2
ofFLCPD ðefn ; e_fn Þ
ogcontroller ðX;P;XÞ ¼ ðbn þ Wn Þ x1
oefn
¼ h1 þ h3 ¼ Kpn þ Kdn
ox1 ofFLCPID ðX; P; XÞ ofFLCPID ðX; P; XÞ ofFLCPD ðefn ; e_fn Þ oe_fn
¼ ð63Þ
ogcontroller ðX;P;XÞ oKden ofFLCPD ðefn ; e_fn Þ oe_fn oKden
¼ h1 þ h2 þ h3 ¼ Kpn þ Kin þ Kdn ofFLCPD ðefn ; e_fn Þ
ox2 ¼ ðbn þ Wn Þ x2
oe_fn
ogcontroller ðX;P;XÞ
¼ h3 ¼ Kdn ð61Þ ofFLCPID ðX; P; XÞ ofFLCPID ðX; P; XÞ
¼ ¼ fFLCPD ðefn ; e_fn Þ
ox3 obn oWn

ogcontroller ðX;P;XÞ
The sensitivity oPm is expressed as follows:
3.8.3 Derivation of the sensitivity functions for fuzzy PID-
type STR based on SVR ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ
¼
oPm ohm
The sensitivity of the control signal with respect to ohm ofestimatorm ðXm ; Pmc Þ oKestimatorm ðPmc Þ

Lagrange multipliers of the parameters estimator ofestimatorm ðXm ; Pmc Þ oKestimatorm ðPmc Þ oPm
(ogcontroller ðX;P;XÞ
), the sensitivity of the control signal with ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ oKestimatorm ðPmc Þ
oXm ¼ Xm
oPm ohm oPm
respect to inputs of the parameters estimator
ðX;P;XÞ ð64Þ
(ogcontroller
oPm ) and the sensitivity of the control signal with
ogcontroller ðX;P;XÞ ofFLCPID ðX;P;XÞ
respect to inputs of the controller (ogcontrolleroxðX;P;XÞ
k
) are The sensitivity oxk ¼ oxk is given as
derived for fuzzy PID-type STR based on SVR as follows. follows:
ogcontroller ðX;P;XÞ
The computation of oXm is as follows:

ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ ohm ofestimatorm ðXm ; Pmc Þ

¼
oXm ohm ofestimatorm ðXm ; Pmc Þ oXm
 
ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ ofestimatorm ðXm ; Pmc Þ ofestimatorm ðXm ; Pmc Þ ofestimatorm ðXm ; Pmc Þ
¼  ð62Þ
oXm ohm obestimatorm oam1 oamk
ogcontroller ðX; P; XÞ ogcontroller ðX; P; XÞ
¼ Kestimatorm ðPmc Þ
oXm ohm

(a) PID Type STR (a) Fuzzy PID Type STR

0.15 0.15
r(t) r(t)
y(t) y(t)
0.1 0.1

0.05 0.05

0 0
0 100 200 300 400 500 0 100 200 300 400 500

(b) (b)
0.6 0.6
u(t) u(t)

0.4 0.4

0.2 0.2

0 0
0 100 200 300 400 500 0 100 200 300 400 500
Time[sec] Time[sec]

Fig. 12 System output (a), control signal (b) for the case with no noise and parametric uncertainty

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S793

ofFLCPID ðX;P;XÞ ofFLCPID ðX;P;XÞ ofFLCPD ðefn ; e_fn Þ oefn ofFLCPD ðefn ; e_fn Þ oe_fn
¼ ð þ Þ
4 Simulation results
oxk ofFLCPD ðefn ; e_fn Þ oefn oxk oe_fn oxk
ofFLCPID ðX;P;XÞ ofFLCPD ðefn ; e_fn Þ oefn ofFLCPD ðefn ; e_fn Þ oe_fn
oxk
¼ ðbn þ Wn Þð
oefn oxk
þ
oe_fn oxk
Þ The performance of the generalized STR with adaptive PID
efn e_f ef e_f and adaptive fuzzy PID controllers are evaluated on the
¼ Ken ; n ¼ Kden ; n ¼ 0 ; n ¼ Kden
ox1 ox1 ox2 ox2 bioreactor benckmark problem. Nevertheless, the SVR-
ofFLCPID ðX;P;XÞ ofFLCPD ðefn ; e_fn Þ ofFLCPD ðefn ; e_fn Þ based self-tuning regulator scheme proposed in this paper
¼ ðbn þ Wn Þð Ke n þ Kden Þ
ox1 oefn oe_fn
can be implemented to control a diverse range of systems
ofFLCPID ðX;P;XÞ ofFLCPD ðefn ; e_fn Þ
ox2
¼ ðbn þ Wn Þð
oe_fn
Kden Þ and to successfully solve fundamental control problems
that frequently appear in practice such as nonlinearity,
ð65Þ
instability.
ofFLCPD ðefn ;e_fn Þ ofFLCPD ðefn ;e_fn Þ
The term oefn and oe_fn is computed as:

ofFLCPD ðefn ; e_fn Þ ðBj ðe_fn ÞC2n þ Bjþ1 ðe_fn ÞC4n Þ2  ðBj ðe_fn ÞC1n þ Bjþ1 ðe_fn ÞC3n Þ2
¼
oefn eiþ1  ei
k11 ek11 efn k21 ek21 efn
þ ðAi ðefn ÞBj ðe_fn ÞÞ2 2
þ ðAiþ1 ðefn ÞBj ðe_fn ÞÞ2
ð1 þ ek11 efn Þ ð1 þ ek21 efn Þ2
2 k31 ek31 efn 2 k41 ek41 efn
þ ðAi ðefn ÞBjþ1 ðe_fn ÞÞ þ ðA iþ1 ðe f ÞB jþ1 ð _
e f ÞÞ
ð1 þ ek31 efn Þ2 ð1 þ ek41 efn Þ2
n n

ð66Þ
ofFLCPD ðefn ; e_fn Þ ðAi ðefn ÞC3n þ Aiþ1 ðefn ÞC4n Þ2  ðAi ðefn ÞC1n þ Aiþ1 ðefn ÞC2n Þ2
¼
oe_fn e_jþ1  e_j
k12 ek12 e_fn k22 ek22 e_fn
þ ðAi ðefn ÞBj ðe_fn ÞÞ2 2
þ ðAiþ1 ðefn ÞBj ðe_fn ÞÞ2
ð1 þ ek12 e_fn Þ ð1 þ ek22 e_fn Þ2
2 k32 ek32 e_fn k42 ek42 e_fn
þ ðAi ðefn ÞBjþ1 ðe_fn ÞÞ 2
þ ðAiþ1 ðefn ÞBjþ1 ðe_fn ÞÞ2
ð1 þ ek32 e_fn Þ ð1 þ ek42 e_fn Þ2

1 4.1 Bioreactor system

0.8
Kp(t) 0.6 The bioreactor system is frequently used in technical lit-
0.4 erature as a benchmark nonlinear system so as to appraise
0.2 and compare the performances of proposed control
0
0 100 200 300 400 500
methodologies [19, 39, 49, 50]. A biorector is a vessel in
which water, cells (e.g., yeast or bacteria) and nutrients
0.8
(substrate) to be consumed by cells are mixed. As a result
of this consuming, product (both desired and undesired)
0.6
Ki(t) and more cells emerge [49]. This system is difficult to
0.4
control since it has highly nonlinear dynamics and exhibits
0.2 limit cycles [49]. The system dynamics can be represented
0 via the following differential equations:
0 100 200 300 400 500
x2 ðtÞ

1
x_1 ðtÞ ¼ x1 ðtÞuðtÞ þ x1 ðtÞð1  x2 ðtÞÞe cðtÞ
0.8 x2 ðtÞ 1 þ bðtÞ
Kd(t) x_2 ðtÞ ¼ x2 ðtÞuðtÞ þ x1 ðtÞð1  x2 ðtÞÞe cðtÞ
0.6 1 þ bðtÞ  x2 ðtÞ
0.4
ð67Þ
0.2
0 where x1 ðtÞ symbolizes the cell concentration , x2 ðtÞ indi-
0 100 200 300 400 500
Time[sec]
cates the amount of nutrients in tank, u(t) is the flow rate by
which the bioreactor is controlled, cðtÞ is nutrient inhibition
Fig. 13 PID-type STR parameters parameter, bðtÞ is grow rate parameter [19, 39, 49–51]. In

123
S794 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

0.2
0.6
0.15
Ke(t) 0.4 ψ(t)
0.1

0.2 0.05

0 0
0 100 200 300 400 500 0 100 200 300 400 500

0.3
0.6

Kde(t) 0.2 β(t) 0.4

0.1 0.2

0 0
0 100 200 300 400 500 0 100 200 300 400 500
Time[sec] Time[sec]

Fig. 14 Fuzzy PID-type STR parameters

(a) PID Type STR (a) Fuzzy PID Type STR

15 15

#svke(t)
10 10
#svkp(t) #svkde(t)

5 #svki (t) 5 #sv (t)

ψ
#svkd(t) #sv (t)
β
0 0
0 100 200 300 400 500 0 100 200 300 400 500

(b) (b)
15 15

10 10

5 5
#svm(t) #svm(t)
0 0
0 100 200 300 400 500 0 100 200 300 400 500
Time[sec] Time[sec]

Fig. 15 Number of the support vectors for controllers (a) and system models (b)

the closed-loop system, the aim is to control the cell con- added to the controlled output of the system. 3) Parametric
centration (yðtÞ ¼ x1 ðtÞ) by manipulating the flow rate uncertainty is introduced to the system. For nominal and
(u(t)). The limitations for the magnitude of the control measurement noise cases the input feature vectors are
signal are umin ¼ 0 and umax ¼ 2. The continuation period selected as Pp ¼ ½rn Pn un1 T ,
of control signal is kept constant at smin ¼ smax ¼ 0:5s. The Pi ¼ ½rn In In1 In2 yn un1 T and Pd ¼ ½rn Dn un1 T
bioreactor system has been simulated using the proposed
for SVRestimator of PID-type STR, and Pke ¼ ½rn Pn un1 T ,
STR architecture with both PID and fuzzy PID controllers
in the generalized controller block. Since we have assumed Pkde ¼ ½rn In yn T , PW ¼ ½rn Dn T , Pb ¼ ½rn In yn un1 T
that the dynamics of the system is not known, online SVR are employed as input feature vector for SVRestimator
has been utilized to identify the unknown dynamics using of fuzzy PID-type STR where Pn ¼ en  en1 , In ¼ en ,
the input–output data pairs. The input feature vector for Dn ¼ en  2en1 þ en2 and en ¼ rn  yn . For the case
with parametric uncertainty, the input feature vectors
SVRmodel is obtained as Mc ¼ ½un . . .unnu ; yn . . .ynny T
for SVRestimator of PID-type STR are utilized as
where nu ¼ ny ¼ 2. Simulations have been performed for
Pp ¼ ½rn Pn yn un1 T , Pi ¼ ½rn In In1 In2 yn un1 T
three separate cases. 1) Nominal case with no measurement
noise and parametric uncertainty. 2) Measurement noise is and Pd ¼ ½rn Dn yn un1 T , and Pke ¼ ½rn In In1

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S795

(a) PID Type STR (a) Fuzzy PID Type STR

0.15 0.15
r(t) r(t)
y(t) y(t)
0.1 0.1

0.05 0.05

0 0
0 100 200 300 400 500 0 100 200 300 400 500

(b) (b)
0.8 0.8
u(t) u(t)
0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 100 200 300 400 500 0 100 200 300 400 500
Time[sec] Time[sec]

Fig. 16 System output (a), control signal (b) for sinusoidal input

(a) PID Type STR (a) Fuzzy PID Type STR

0.15 0.15
r(t) r(t)
y(t) y(t)
0.1 0.1

0.05 0.05

0 0
0 100 200 300 400 500 0 100 200 300 400 500

(b) (b)
0.6
0.6 u(t) u(t)

0.4
0.4

0.2 0.2

0 0
0 100 200 300 400 500 0 100 200 300 400 500
Time[sec] Time[sec]

Fig. 17 System output (a), control signal (b) for the case with measurement noise (30 dB)

Dn yn un1 T , Pkdes ¼ PW ¼ Pb ¼ Pke are deployed as Number of the support vector for both controller and sys-
input feature vector for SVRestimator of fuzzy PID STR. tem models are illustrated in Fig. 15.
As can be seen from Fig. 16, the controllers track the
sinusoidal reference input accurately.
4.2 Nominal case with no noise and parametric
uncertainty 4.3 Measurement noise

The tracking performance of both controllers for the case The performance of the controllers under the influence of
when no noise and parametric uncertainty is applied to the measurement noise is evaluated by adding a 30-dB mea-
system and all parameters are fully known is given in surement noise to the system output. The tracking perfor-
Fig. 12. The system accurately tracks the reference input. mance and control input for controller are demonstrated in
The controller parameters are depicted in Figs. 13 and 14. Fig. 17.

123
S796 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

(a) PID Type STR (a) Fuzzy PID Type STR

0.15 0.15
r(t) r(t)
y(t) y(t)
0.1 0.1

0.05 0.05

0 0
0 100 200 300 400 500 0 100 200 300 400 500

(b) (b)
0.8
u(t) u(t)
0.6
0.6

0.4 0.4

0.2 0.2

0 0
0 100 200 300 400 500 0 100 200 300 400 500

(c) (c)
0.54 0.54
γ(t) γ(t)
0.52 0.52
0.5 0.5
0.48 0.48
0.46 0.46
0.44 0.44
0.42 0.42
0 50 100 150 200 250 0 50 100 150 200 250
Time[sec] Time[sec]

Fig. 18 System output (a), control signal (b) for the case with parametric uncertainty (c)

x 10
-4 PID Type STR x 10
-4 Fuzzy PID Type STR
(a) 6
(a)
5

4 4
V (t ) V (t ) 3
2 2
1
0 0
0 100 200 300 400 500 0 100 200 300 400 500
-3
x 10
(b) (b) 0
0
-2 -0.05
∂V (t )
-4 ∂V (t )
∂t
-6 ∂ t -0.1
-8 -0.15
-10 -0.2
0 100 200 300 400 500 0 100 200 300 400 500
Time[sec] Time[sec]

oVðtÞ
Fig. 19 V(t) (a) and ot (b) for the case with no noise and parametric uncertainty

4.4 Uncertainty in system parameters cðtÞ ¼ 0:48 þ 0:06 sinð0:016ptÞ. Figure 18 illustrates the
tracking performance of the controllers and control signal
The robustness of the controllers are examined with respect applied to the system in this case. By comparing Fig. 18
to parametric uncertainty, cðtÞ is presumed as the uncertain with the nominal case given in Fig. 12, it can be perceived
system parameter, where its nominal value is cnom ðtÞ ¼ that uncertainty caused by the time-varying parameter is
0:48 and alters around its nominal value as tolerated.

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S797

x 10
-4 PID Type STR x 10
-4 Fuzzy PID Type STR
(a) (a)
6 5
4
V (t ) 4 V (t ) 3

2 2
1
0 0
0 100 200 300 400 500 0 100 200 300 400 500
-4
x 10
(b) (b) 0
0
-2
-0.01
∂V (t ) -4 ∂V (t )
∂ t -6 ∂ t -0.02
-8
-10 -0.03

0 100 200 300 400 500 0 100 200 300 400 500
Time[sec] Time[sec]

oVðtÞ
Fig. 20 V(t) (a) and ot (b) for the case with measurement noise

-4
x 10
-4
PID Type STR x 10 Fuzzy PID Type STR
(a) (a)
6 6

4 4
V (t)
V (t)
2 2

0 0
0 100 200 300 400 500 0 100 200 300 400 500

-4 -4
x 10 x 10
(b) (b) 0
0
-0.5
∂V (t)
∂V (t) -5 -1
∂t
∂t
-1.5
-10 -2
-2.5
0 100 200 300 400 500 0 100 200 300 400 500
Time[sec] Time[sec]

oVðtÞ
Fig. 21 V(t) (a) and ot (b) for the case with parametric uncertainty

4.5 Closed-loop Lyapunov stability analysis 4.6 Comparison of the results with SVM-based PID
controller
The stability analysis of the proposed control methodology
derived in Sect. 3.8 is carried out, and the numerical ver- In order to evaluate the performance comparisons of the
ifications for both PID-type STR and fuzzy PID-type STR controllers, SVM-based PID controller proposed by Iplikci
are depicted in Figs. 19, 20 and 21 for noiseless case, and [19] is deployed to control bioreactor system for cases with
when measurement noise and parametric uncertainty are no noise and parametric uncertainty, with measurement
added respectively. For stability in the sense of Lyapunov, noise added to the system and with parametric uncertainty.
oVðetr Þ In a nutshell, the adjustment mechanism proposed for
both Vðetrnþ1 Þ  0 and ot
nþ1
 0 must be ensured simul-
taneously. As illustrated in Figs. 19, 20 and 21, it has been SVM-based PID controller optimizes the parameters of a
oVðetr Þ PID controller via Levenberg–Marquardt algorithm by
observed that Vðetrnþ1 Þ  0 and ot
nþ1
 0 for both con- utilizing K-step ahead future Jacobian of the system
troller during the course of control. In a nutshell, it can be behavior. The proposed mechanism includes five compo-
inferred that the closed-loop systems are stable for all nents: classical PID controller, controller parameter tuner,
cases. SVR NARX model of the system, control signal correction

123
S798 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

(a) PID Type STR (a) PID Type STR

0.15 0.15
r(t) r(t)
y(t) y(t)
0.1 0.1

0.05 0.05

0 0
0 100 200 300 400 500 0 100 200 300 400 500

(b) Fuzzy PID Type STR (b) Fuzzy PID Type STR
0.15 0.15
r(t) r(t)
y(t) y(t)
0.1 0.1

0.05 0.05

0 0
0 100 200 300 400 500 0 100 200 300 400 500

(c) SVM-based PID

(c) SVM-based PID
0.15 0.15
r(t) r(t)
y(t) y(t)
0.1 0.1

0.05 0.05

0 0
0 100 200 300 400 500 0 100 200 300 400 500
Time[sec] Time[sec]

Fig. 22 Comparison of controllers for the case with no noise and parametric uncertainty (left) and for the case with measurement noise (right)

block and line search block. The PID controller has three the prediction horizon (K) of SVM-based PID controller is
tunable parameters (Kp , Ki , Kd ) adjusted via Levenberg- set as ‘‘K ¼ 1.’’ The closed-loop performances of the
Marquardt algorithm in controller parameter tuner block. controllers for the case with no noise and parametric
SVR NARX model approximates K-step ahead Jacobians uncertainty and the case with measurement noise are
of the system in order to constitute Jacobian matrix which illustrated in Fig. 22 (left and right) respectively. The
is required in Levenberg-Marquardt optimization algo- robustness of the controllers with respect to parametric
rithm. Since the updated controller parameters may not uncertainty is evaluated in Fig. 23. In order to compare
good enough to force the system output to follow the controller performances numerically, the following per-
desired trajectory in transient state because of modeling formance index function is employed
inaccuracies and external disturbances, a control signal X1
correction term which is derived via Taylor expansion of Jcomp ¼ ½rnþ1  ynþ1 2 ð68Þ
control signal is utilized in control signal correction block n¼0

[19, 38]. Line search block calculates the optimal learning and the performance comparisons are depicted in Fig. 24.
rate for control signal correction term via golden section As can be seen from Fig. 24, PID-type STR has the best
method. The closed-loop tracking performances of the performance for all cases. The fuzz PID-type STR has
controllers proposed in this paper are compared with the better performance than SVM-based PID controller for all
closed-loop tracking competency of SVM-based PID con- cases. Comparison of the PID-type STR and SVM-based
troller given in [19]. When the prediction horizon of the PID performances is more meaningful since the adaptation
SVM-based PID controller is increased, it is possible to performance of the STR and Levenberg Marquardt can be
obtain better or similar results than PID-type STR and comparable. Therefore, SVR-based PID-type STR has
fuzzy PID-type STR. In order to compare the controller better tracking performances than SVM-based PID con-
under the same conditions and obtain meaningful results, troller for all cases. In order to exhaustively compare the

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S799

(a) PID Type STR transient state and steady-state behavior of the controllers,
0.15 the accuracies of the controllers are compared in terms of
r(t)
y(t) maximum and average values of steady-state errors, set-
0.1
tling times(according to 2 % error criterion) and over-
shoots separately for cases with no noise, with noise and
0.05
with parametric uncertainty. The results are tabulated in
0
Tables 1 and 2 respectively. As can be seen from Fig. 24
0 100 200 300 400 500 and Tables 1 and 2, the controller with the best transient
(b) Fuzzy PID Type STR and steady-state behavior is PID-type STR. The results
0.15
r(t)
given in Tables 1 and 2 verify the results given in Fig. 24.
y(t)
0.1

0.05
5 Conclusion

0 Support vector machines have successfully been employed

0 100 200 300 400 500
to cope with various classification and regression problems
(c) SVM-based PID for nearly two decades. Their performance is justified to be
0.15
r(t) superior than gradient-based intelligent systems like
y(t) ANNs, ANFISs due to their convex objective function and
0.1
better generalization property. However, they have not
0.05 been used as controllers since information about control
input to be applied to the system, which is required for
0 SVR training is not available beforehand. In this paper, a
0 100 200 300 400 500
Time[sec] novel architecture where an online SVR is used to tune the
parameters of a generalized STR which optimizes the
Fig. 23 Comparison of controllers for the case with parametric margin between reference input and system output has
uncertainty
been proposed. There are two online SVR structures
employed in the control system, SVRmodel calculates the
model of the controlled system and predicts its future
PID Type STR
behavior and SVRestimator estimates the controller parame-
0.06
Fuzzy PID Type STR ters of the STR. Two different controllers have been used
SVM-based PID
0.0484
in the controller block: adaptive PID and adaptive fuzzy
0.05
PID. A separate SVRestimator is designed for each tunable
0.0421

0.04 0.0385
parameter in the controllers.
0.0362 0.0368
Jcomp 0.0345 The performance of the proposed generalized adaptive
0.03 0.0281
0.0303
0.0293 control architecture and parameter estimator is evaluated for
both PID and fuzzy PID controllers on a bioreactor benchmark
0.02 system. A comprehensive stability analysis of the generalized
STR is performed. Furthermore, the closed-loop tracking
0.01 performances of the STRs are compared with SVM-based PID
controller proposed by Iplikci. The robustnesses of the con-
0 trollers have also been assessed for the noiseless case and
with no1noise 2
with 3
with parametric
when measurement noise and parametric uncertainty are
and measurement uncertainty
parametric noise added. Simulation results indicate that the proposed adapta-
uncertainty
tion mechanism for generalized STR accomplishes successful
Fig. 24 Performance comparison of controllers with respect to the tracking performance as well as good noise rejection and high
defined performance index (68) toleration to parametric uncertainties.

123
S800 Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801

Table 1 Maximum values of steady-state errors (ess ð%Þ), settling times (ts ð2 %Þ (s)) and overshoots (P.O. (%))
Controllers ess ð%Þ ts ð2 %Þ P.O.(%)
Noiseless Noisy Uncert. Noiseless Noisy Uncert. Noiseless Noisy Uncert.

PID-type STR 0.002 2 2 19.5 27.5 17 1.2605 18.0158 12.9661

Fuzzy PID-type STR 0.0019 2 2 27 34 19.5 0.0016 5.1059 14.5094
SVM-based PID 0.0559 1.185 2 31 36 72.5 0 3.5891 24.5359

Table 2 Average values of steady-state errors (ess ð%Þ), settling times (ts ð2 %Þ (s)) and overshoots (P.O. (%))
Controllers ess ð%Þ ts ð2 %Þ P.O. (%)
Noiseless Noisy Uncert. Noiseless Noisy Uncert. Noiseless Noisy Uncert.

PID-type STR 8:4611 104 1.4166 1.6691 10.6 14.5 7.9 0.2530 6.7918 4.5779
Fuzzy PID-type STR 8:0872 104 0.9027 1.7872 15.6 17.2 12.7 5:3 104 2.9685 6.1315
SVM-based PID 0.0363 0.5211 1.7716 25.3 27.2 27.8 0 1.9421 9.0258

In future works, new SVR type adjustment mechanisms 11. Spooner JT, Passino KM (1996) Stable adaptive control using
can be developed by employing closed-loop margin fuzzy systems and neural networks. IEEE Trans Fuzzy Syst
4(3):339–359. doi:10.1109/91.531775
notion. 12. Denai MA, Palis F, Zeghbib A (2004) ANFIS based modelling
and control of non-linear systems: a tutorial. In: IEEE interna-
Compliance with ethical standards tional conference on systems, man and cybernetics
13. Smola AJ, Schlkopf B (2004) A tutorial on support vector
Conflict of interest The authors declare that there is no conflict of regression. Stat Comput 14(3):199–222. doi:10.1023/B:STCO.
interests regarding the publication of this paper. 0000035301.49549.88
14. Cristianini N, Shawe-Taylor J (2000) An introduction to support
vector machines and other kernel-based learning methods.
References Cambridge University Press, Cambridge
15. Scholkopf B, Smola AJ (2002) Learning with kernels: support
vector machines, regularization, optimization, and beyond. MIT
1. Bobal V, Bohm J, Fessl J, Machacek J (2005) Digital self-tuning
Press, London
controllers. Advanced textbooks in control and signal processing.
16. Suykens JAK (2001) Nonlinear modelling and support vector
Springer, London
machines. In: IEEE instrumentation and measurement technology
2. Aström KJ, Wittenmark B (2008) Adaptive control. Dover
conference (IMTC/2001), Budapest
Publications, Mineola
17. Wanfeng S, Shengdun Z, Yajing S (2008) Adaptive PID con-
3. Aström KJ (1983) Theory and applications of adaptive control—a
troller based on online LSSVM identification. In: IEEE/ASME
survey. Automatica 19(5):471–486
international conference on advanced intelligent mechatronics
4. Wellstead PE, Liptak BG, Renganathan S (2006) Self-tuning
(AIM 2008), Xian
controllers. In: Liptak Bela G (ed) Instrument engineers hand-
18. Wang X, Du Z, Chen J, Pan F (2009) Dynamic modeling of
book: process control and optimization, vol 2, 4th edn. CRC
biotechnical process based on online support vector machine.
Press, Boca Raton, pp 345–350
J Comput 4(3):251–258. doi:10.4304/jcp.4.3.251-258
5. Aström KJ, Borisson U, Ljung L, Wittenmark B (1977) Theory
19. Iplikci S (2010) A comparative study on a novel model-based
and applications of self-tuning regulators. Automatica 13(5):
PID tuning and control mechanism for nonlinear systems. Int J
457–476
Robust Nonlinear Control 20(13):1483–1501. doi:10.1002/rnc.
6. Akhyar S, Omatu S (1993) Self-tuning PID control by neural
1524
networks. In: International joint conference on neural network
20. Ponce AN, Behar AA, Hernandez AO, Sitar VR (2004) Neural
(IJCNN’93), Nagoya
networks for self-tuning control systems. Acta Polytech
7. Wang GJ, Fong CT, Chang KJ (2001) Neural-network-based self-
44(1):49–52
tuning PI controller for precise motion control of PMAC motors.
21. Flynn D, McLoone S, Irwin GW, Brown MD, Swidenbank E,
IEEE Trans Ind Electron 48(2):408–415
Hogg BW (1997) Neural control of turbogenerator systems.
8. Efe MO, Kaynak O (2000) A comparative study of soft-com-
Automatica 33(11):1961–1973. doi:10.1016/S0005-
puting methodologies in identification of robotic manipulators.
1098(97)00142-8
Robot Auton Syst 30(3):221–230
22. Abdullah R, Hussain AT, Zayed A (1997) A new RBF neural
9. Hagan MT, Demuth HB, De Jesus O (2002) An introduction to
network based non-linear self-tuning pole-zero placement con-
the use of neural networks in control systems. Int J Robust
troller. In: 15th international conference on artificial neural net-
Nonlinear Control 12(11):959–985. doi:10.1002/rnc.727
works (ICANN 2005), Warsaw
10. Shu HL, Pi YG (2000) PID neural networks for time-delay sys-
23. Wahyudi S, Ahmad W, Htut MM (2009) Neural-tuned PID
tems. Comput Chem Eng 24(2–7):859–862. doi:10.1016/S0098-
controller for point-to-point (PTP) positioning system: model
1354(00)00340-9

123
Neural Comput & Applic (2017) 28 (Suppl 1):S775–S801 S801

reference approach. In: 5th international colloquium on signal In: International conference on fuzzy theory and its applications
processing and its applications (CSPA), Kuala (IFUZZY2012), Taichung
24. Guo AW, Yang J (2007) Self-tuning PID control of hydro-turbine 38. Uçak K, Günel GÖ (2015) An adaptive support vector regressor
governor based on genetic neural networks. In: International controller for nonlinear systems. Soft Comput. doi:10.1007/
symposium on intelligence computation and applications (ISICA s00500-015-1654-0
2007), Wuhan 39. Iplikci S (2006) Online trained support vector machines-based
25. Kang Y, Chu MH, Chang CW, Chen YW, Chen MC (1997) The generalized predictive control of non-linear systems. Int J Adapt
self-tuning neural speed regulator applied to DC servo motor. In: Control Signal Process 20(10):599–621. doi:10.1002/acs.919
International conference on natural computation (ICNC 2007), 40. Ma J, Theiler J, Perkins S (2003) Accurate online support vector
Haikou regression. Neural Comput 15(11):2683–2703. doi:10.1162/
26. Pham DT, Karaboga D (1999) Self-tuning fuzzy controller design 089976603322385117
using genetic optimisation and neural network modelling. Artif 41. Mario M (2002) On-line support vector machine regression. In:
Intell Eng 13(2):119–130. doi:10.1016/S0954-1810(98)00017-X 13th European conference on machine learning (ECML 2002),
27. He SZ, Tan SH, Xu FL (1993) PID self-tuning control using a Helsinki
fuzzy adaptive mechanism. In: International conference on fuzzy 42. Efe MO, Kaynak O (1999) A comparative study of neural net-
systems, San Francisco work structures in identification of nonlinear systems. Mecha-
28. Gautam D, Ha C (2013) Control of a quadrotor using a smart self- tronics 9(3):287–300. doi:10.1016/S0957-4158(98)00047-6
tuning fuzzy PID controller. Int J Adv Robot Syst 10:1–9. doi:10. 43. Jang JSR (1993) ANFIS: adaptive-network-based fuzzy inference
5772/56911 system. IEEE Trans Syst Man Cybern 23(3):665–685. doi:10.
29. Ahn KK, Truong DQ, Thanh TQ, Lee BR (2008) Online self- 1109/21.256541
tuning fuzzy proportional integral derivative control for hydraulic 44. Sung SW, Lee J, Lee IB (2009) Process identification and PID
load simulator. J Syst Control Eng 222(2):81–95. doi:10.1243/ control. IEEE Press, Wiley, Singapore
09596518JSCE484 45. Aström KJ, Hagglund T (1995) PID controllers: theory, design
30. Qiao WZ, Mizumoto M (1996) PID type fuzzy controller and and tuning. Instrument Society of America, Research Triangle
parameters adaptive method. Fuzzy Sets Syst 78(1):23–35 Park
31. Woo ZW, Chung HY, Lin JJ (2000) A PID type fuzzy controller 46. Visioli A (2006) Practical PID control. Springer, London
with self-tuning scaling factors. Fuzzy Sets Syst 115(2):321–326. 47. Luenberger DG, Ye Y (2008) Linear and nonlinear programming.
doi:10.1016/S0165-0114(98)00159-6 Springer, New York
32. Bandyopadhyay R, Chakraborty UK, Patranabis D (2001) Auto- 48. Saadia N, Amirat Y, Pontnau J, M’Sirdi NK (2001) Neural hybrid
tuning a PID controller: a fuzzy-genetic approach. J Syst Archit control of manipulators, stability analysis. Robotica 19:41–51.
47(7):663–673. doi:10.1016/S1383-7621(01)00022-4 doi:10.1017/S0263574700002885
33. Sharkawy AB (2010) Genetic fuzzy self-tuning PID controllers 49. Ungar LH (1990) Neural networks for control. In: Miller WT III,
for antilock braking systems. Eng Appl Artif Intell Sutton RS, Werbos PJ (eds) A bioreactor benchmark for adaptive
23(7):1041–1052. doi:10.1016/j.engappai.2010.06.011 network based process control. MIT Press, Cambridge,
34. Bouallègue S, Haggege J, Ayadi M, Benrejeb M (2012) PID- pp 387–402
type fuzzy logic controller tuning based on particle swarm opti- 50. Efe MO (2007) Discrete time fuzzy sliding mode control of a
mization. Eng Appl Artif Intell 25(3):484–493. doi:10.1016/j. biochemical process. In: 9th WSEAS international conference on
engappai.2011.09.018 automatic control, modeling and simulation (ACMOS’07),
35. Li C, Priemer R (1996) Self-learning general purpose PID con- Istanbul
troller. J Frankl Inst 334(2):167–189. doi:10.1016/S0016- 51. Efe MO, Abadoglu E, Kaynak O (1999) A novel analysis and
0032(97)81151-9 design of a neural network assisted nonlinear controller for a
36. Bishr M, Yang YG, Lee G (2000) Self-tuning PID control using bioreactor. Int J Robust Nonlinear Control 9(11):799–815. doi:10.
an adaptive network based fuzzy inference system. Intell Autom 1002/(SICI)1099-1239(199909)9:11\799:AID-RNC441[3.0.
Soft Comput 6(4):271–280 CO;2-U
37. Lu CH, Cheng CC, Liu CM, Guo JY (2012) Self-tuning predic-
tive PID controller using wavelet type-2 fuzzy neural networks.

123