You are on page 1of 10

PSO based Optimal Weight Selection of LQ Regulator for Linear Inverted

Pendulum System
Vinodh Kumar. E* Jovitha Jerome**

*Department of Instrumentation and Control Systems Engineering, PSG College of Technology, Coimbatore,
India- 641004 (vinothmepsg@gmail.com)

**Department of Instrumentation and Control Systems Engineering, PSG College of Technology, Coimbatore,

India- 641004 (jjovitha@yahoo.com)

Abstract: This paper presents a Particle Swarm Optimization (PSO) based method to determine the optimal weights
of LQR controller for a Self Erecting Single Inverted Pendulum (SESIP). Normally, the weights of LQR controller
are selected based on trial and error approach and there is no specific rule for selecting the weights, which makes it
tedious to tune the controller for optimum performance. To overcome this problem, an intelligent approach
employing PSO based optimization is proposed. Swarm intelligence is used to obtain the optimal weights, which
provides superior performance than the conventional trial and error approach. In order to assess the performance of
the proposed approach, weight selection using GA is compared with the PSO. The performance of the controller is
evaluated not only for stabilizing the pendulum in upright position, but also for tracking the given reference signal.

Keywords: Inverted Pendulum, LQR Controller, Riccatti Equation, Particle Swarm Optimization, Genetic
Algorithm, Weighting Matrices

1. INTRODUCTION is a powerful controller for many control systems,


which has become one of the standard techniques for
The linear Inverted Pendulum system is a fourth control systems design. LQR method is increasingly
order, nonlinear, unstable system, which can be used in aspects like control of induction motors,
treated as a typical control problem to study various vehicular drive-shaft control, etc. In an LQ regulation
modern control theories. It is a well established problem with the quadratic performance index J, the
benchmark problem that provides many challenging choice of the parameters in penalty matrices Q and R
problems to control design [1]. Because of their is crucial for the performance of feedback-controlled
nonlinear nature, pendulums have maintained their system under consideration. The problem of selecting
usefulness and they are now used to illustrate many weighting matrices has been investigated by various
of the ideas emerging in the field of nonlinear control methods. Kalman [9] firstly presented a method for
[2]. The challenges of control made the inverted weighting matrices selection based on the given
pendulum systems a classic tool in control poles. Wang [10] expanded the researches in Kalman
laboratories. According to control purposes of direction. The design methods only compute the
inverted pendulum, the control of inverted pendulum weighting matrices based on the given poles, and
can be divided into three aspects. The first aspect that dont guarantee the performance and constraints of
is widely researched is the swing-up control of the system. Recently, more researches have been
inverted pendulum [3-4]. The second aspect is the performed for LQR controller design using GA
stabilization of the inverted pendulum [56]. The method. This method has some disadvantages in
third aspect is tracking control of the inverted designing of LQR controller such as high
pendulum [7-8]. In practice, stabilization and computational time and low rate of successful
tracking control is more useful for plenty of real time optimization. In order to overcome these difficulties,
applications. The Linear Quadratic Regulator (LQR) a new approach is proposed which employs the
optimization method called PSO. The main Figure 2: Pendulum cart system block diagram
contribution of this paper is to compare the
performances of two different evolutionary Applying Newtons second law of motion to the cart
algorithms namely, PSO and GA for reference signal system, the equation of motion can be represented as
tracking and controlling the Inverted pendulum follows:
system. Our goal is to obtain optimal weighting M x =F Beq x F a
matrices for the design of a LQR controller which not
only stabilizes the inverted pendulum but also tracks (1)
the given reference signal.
The rest of this paper is organized as follows. Section Where F is cart driving force produced by the motor
II contains the description of the Inverted Pendulum and Fa is inertia force due to the armature motor
system. Controller specifications are given in section rotation. The driving force acting on the cart through
III. Section IV deals with LQR Controller and PV the motor pinion can be expressed as:
controller designs. Optimal weight selection using
PSO algorithm is detailed in section V. Simulation g K g T m
results are presented in section VI and the conclusion F=
is given in VII. rm
(2)

2. SYSTEM MODEL Where T m is torque generated by the motor.


Using Kirchoffs voltage law, the following equation
The linear Self Erecting Single Inverted Pendulum is obtained (by disregarding the motor inductance
(SESIP) consists of a pendulum system which is since Lm Rm ):
attached to a cart equipped with a motor that drives it
along a horizontal track. The schematic diagram of
inverted pendulum system is shown in Figure 1.
V Eemf V K m m
i m= =
R R
(3)

TABLE 1
List of parameters

Symbol Description Value/Unit

R Motor armature resistance 2.6


L Motor armature inductance 0.18mH
Kt Motor torque constant 0.00767 Nm/A
m Motor efficiency 100%
Km Motor EMF constant 0.00767 Ns/rad
J Rotor moment of inertia 3.9x10-7 kgm2
Kg Gearbox ratio 3.71
Figure 1: Schematic of Cart-Inverted Pendulum
g Gearbox efficiency 100%
rm Motor pinion radius 6.35x10-3 m
The user can dictate the position and velocity of the
cart through the motor and the track restricts the cart rp Position pinion radius 1.48x10-2 m
movement in the horizontal direction [11]. Encoders Beq Equivalent viscous damping
are attached to the cart and the pivot in order to coefficient at motor 5.4 Nms/rad
measure the cart position and pendulum joint angle, Bp Viscous damping coefficient
respectively. The block diagram of the overall set up at pendulum pivot 5.4 Nms/rad
is given in Figure 2.
l Pendulum length from
pivot to centre of mass 0.3302 m
I Pendulum moment of inertia 7.88x10-3 kgm2
Mp Pendulum mass 0.23kg
M Cart mass 0.94kg
Vm Motor nominal input voltage 5V
2
K J x
Fa = g g 2
The torque generated by the DC motor is rm
proportional to the armature current as: (11)

T m=m K t i m Finally, substituting (11) and (7) into (1), applying


(4) Laplace transform and rearranging, yields the open
loop transfer function for the cart system:
Substituting (4) and (3) into (2) leads to:
X (s ) a1
g m K g K t ( V K m m x ) G ( s )= =
F= V ( s ) ( b 1 s+b 2 ) s
R rm (12)
(5)
Where:
The linear velocity of the cart and motor angular
velocity can be related by: a1=r m g m K g K t

K g x b1=RM r 2m + R g K 2g J
m= (6)
rm
2 2
b2=g K g m K t K m + Beq R r m
Therefore (5) can be rearranged to:
The nonlinear model of the inverted pendulum
g m K g K t ( r m V K g K m x ) motion can be derived using Lagranges equation.
F= For brevity, the nonlinear equation of motions can be
R r m2
obtained as follows:
(7)

As seen at the motor pinion, the armature inertia


( M + M p ) x + Beq x M p l cos + M p l 2 sin =F
force due to the motor rotation can be expressed as: (13)

K T B p M p gl sin =0
M p l x cos + ( I + M p l 2 ) +
Fa = g g a
rm (14)
(8)
The nonlinear model can be linearized which is valid
nd near the equilibrium point (upright pendulum) so that
Applying Newtons 2 law of motion to the motor
shaft:
sin = and cos =1 and also neglecting
higher order term. The linearized model is written in
state space in order to allow the design of state
J =T a feedback controller for upright pendulum
(9) stabilization.

Where m is motor shaft rotation angle. Moreover, the X =AX + BU


mechanical configuration of the carts pinion system (15)
gives the following equation:
Y =CX
Kgx
m = (10) Where
T
X =[ x x ] ,U =V and
rm
T
Y =[ x ]
Using (10) and (9) and substituting into (8), gives:
part involves the design of an optimal state feedback
controller for the linearized model that will stabilize
the pendulum around the upright position. The
second part involves the design of a controller that
swings the pendulum up to the unstable equilibrium.
When the pendulum approaches the linearized point,
the control will switch to the stabilizing controller
which will balance the pendulum around the upright
position. The control scheme of SESIP consists of
two main control loops and decision making logic to
switch between the two control schemes. One control
loop is a proportional velocity controller on the cart
position that follows a set point designed to swing up
the pendulum from the suspended to the inverted
posture. The other control loop is active when the
pendulum is around the upright position and consists
of a Linear Quadratic Regulator (LQR) maintaining
With q=( M + M p ) I + M M p l 2 the inverted pendulum in vertical position. The
control scheme of the self-erecting inverted
The state model of the system obtained by pendulum is shown in Figure 3.
substituting the parameters presented in Table 1 is
given as:

[][ ][ ] [ ]
xc 0 0 1 0 xc 0
= 0 0 0 1 + 0 u
xc 0 2.2643 15.8866 0.0073 xc 2.2772
0 27.8203 36.6044 0.0896 5.2470
(16)

Figure 3: Control scheme of inverted pendulum

[]
xc

[
y= 1 0 0 0
0 1 0 0 xc ] LQR design specification

The primary objective of the LQR scheme is to catch,


in a first time, swing-up pendulum and then to
(17) maintain it balanced in the inverted posture. The
linear cart should track a desired (square wave)
The eigen values of the system matrix are -16.2577, position set point and at the same time the controller
-4.5611, 0, 4.8426. It can be clearly seen that the should also minimize the control effort. The purpose
open loop system has one pole in the Right Half of optimal control is to allow for best tradeoff
Plane (RHP) i.e., positive pole. Therefore the system between performance and cost of control. The gain of
is unstable in open loop. As a consequence, in order the LQR scheme is tuned to control the inverted
to maintain the pendulum balanced in the inverted pendulum and linear cart system to satisfy the
position, a controller is to be designed such that all following design requirements.
the resulting closed loop poles lie in the Left Half 1. The pendulum angle should be regulated
Plane (LHP). around its upright position and never exceed
a 1 degree deflection.
3. FORMULATION OF CONTROLLER 2. Rise time 2s
DESIGN PROBLEM 3. Control effort Vm should be minimum and is
not allowed to reach the saturation level.
The controller design for the inverted pendulum
system is broken up into two components. The first
4. CONTROLLER DESIGN
s 2 +2 s+ 2=0 (21)
PV Controller
the PV controller gains are obtained as
This controller aims at swinging up the pendulum K p=274.62 and K v =5.53 .
from the rest (=1800=-1800) while keeping the cart
travels within the limited horizontal distance. Many Switching Algorithm
different control algorithms can be used to perform
the swing up control such as, trajectory tracking, A transition algorithm is designed to switch from
rectangular reference input swing up type and Pulse swing-up controller to stabilizing controller. This is
Width Modulation (PWM), in a controlled manner performed by smooth conditional switching that can
that the energy is gradually added to the system to be expressed as follows:
bring the pendulum to the inverted position. In this If ||>250, only swing-up control is active.
work, a Proportional-Velocity (PV) controller is used If ||<200, only stabilizing control is active.
because of its simple structure, effectiveness and easy If 200 ||250, swing-up and stabilizing control
tuning. The block diagram representation of swing up signal are averaged/weighted. The last condition is
controller is shown in Figure 4. The proportional made as transition region to avoid what known as
velocity position controller for servo plant introduces hard switching.
two corrective terms. One is proportional (Kp_c) to the
cart position error while the other is proportional
LQR Controller
(Kd_c) to the cart velocity. The resulting PV control
law is given as:
The stabilizing controller, the focus of this paper, is
designed based on linearized state space model in
V m ( t ) =K p ( xc ( t ) xc ( t ) ) + K v
c d c [ d
x (t )
dt c ] (18 (15). It is a state feedback control with integral action
for tracking system. It is assumed that the system
given in (15) is completely state controllable and all
)
state variables are available for feedback [12]. One
can use full state feedback control with integral gain
as shown in Figure 5.

Figure 4: Block diagram of PV controller


Figure 5: State feedback control with integral gain for
The closed loop transfer function of the cart servo tracking system
can be expressed as:
Controller gains k:=(k1,k2,k3, k4) can be determined
X (s ) K p G ( s) by common method such as pole placement or LQR
= method. However, the issue in pole placement
X r ( s ) 1+ K p G ( s )+s K v G ( s ) method becomes a matter of trial and error to meet
(19) the desired response, same as choosing the
appropriate element of Q and R matrices in LQR
This leads to a second order system as follows: method. The problem becomes more intractable when
X (s ) 1.6 K p the system has multiple outputs. Therefore, the
= 2 objective of this work is to obtain the optimal values
X r ( s ) s + s ( 12.23+1.6 K v ) +1.6 K p
of Q and R matrices of LQR method using PSO
(20) algorithm. The optimal weights are in turn used to
determine the state feedback controller gains which
The desired performance is =0.59 and =26 rad/s. can make the system not only to stabilize in the
Comparing characteristic equation in (19) with the upright position but also to track the given
standard second order form: reference signal. A brief theory of LQR
controller is presented here. Consider a linear time Commonly, a trial and error method has been used to
invariant (LTI) system represented by the following construct the matrices Q and R elements. This
state equation. method is very simple and very familiar in LQR
application. However, it takes long time to choose the
x ( t )= Ax ( t )+ Bu ( t ) , t 0, x ( 0 )=x 0 (22) best values for matrices Q and R. To overcome this
problem, a Particle Swarm Algorithm is proposed to
obtain the optimal weights of Linear Quadratic
Where x ( t ) is the state vector, u(t) is the input
Regulator.
vector, A and B indicate the constant system model
parameters. The pair (A, B) is assumed to be
stabilizable. The full state feedback control law is
defined as,
5. PARTICLE SWARM
OPTIMIZATION
u ( t ) = Kx ( t )
(23) Particle Swarm Optimization (PSO) is an
evolutionary computation technique, developed by
Where K is state feedback matrix. The linear Russell Eberhart James Kennedy [15-17] in 1995,
quadratic cost function is defined as, and was inspired by the social behavior of bird
flocking and fish schooling. PSO has its roots in
artificial life and social psychology as well as in
J ( K )= [ x T ( t ) Qx ( t )+u T ( t ) Ru ( t ) ] dt engineering and computer science. It utilizes a
0
population of particles that fly through the
(24) problem hyperspace with given velocities. At each
iteration, the velocities of the individual particles are
where Q is a nonnegative definite matrix that stochastically adjusted according to the historical best
penalizes the departure of system states from the position for the particle itself and the neighborhood
equilibrium and R is a positive definite matrix that best position. Both the particle best and the
penalizes the control input. The solution of the LQR neighborhood best are derived according to a user
problem can be obtained via a Lagrange multiplier defined fitness function [16]. The movement of each
based optimization technique and is given by particle naturally evolves to an optimal or near-
optimal solution. The word swarm comes from the
1 T irregular movements of the particles in the problem
K=R B P space, now more similar to a swarm of mosquitoes
(25) rather than a flock of birds or a school of fish [18].
PSO is a computational intelligence based technique
Where P R n m is a nonnegative definite matrix that is not largely affected by the size and
satisfying the matrix Riccati Equation, nonlinearity of the problem, and can converge to the
optimal solution in many problems where most
A T P+ PA+Q PB R1 B T P=0 analytical methods fail to converge. Moreover, PSO
(26) has some advantages over other similar optimization
techniques such as Genetic Algorithm (GA), namely:
The weighting matrices Q and R are important i. PSO is easier to implement and there are fewer
components of an LQR optimization process. The parameters to adjust.
compositions of Q and R elements have great ii. In PSO, every particle remembers its own
influences on system performance [13-14]. The previous best value as well as the
number of matrices Q and R elements are dependent neighborhood best; therefore, it has a more
on the number of state variable (n) and the number of effective memory capability than the GA.
input variable (m), respectively. The diagonal-off iii. PSO is more efficient in maintaining the
elements of these matrices are zero for simplicity. If diversity of the swarm [19] (more similar to
diagonal matrices are selected, the quadratic the ideal social interaction in a community),
performance index is simply a weighted integral of since all the particles use the information
the squared error of the states and inputs. The related to the most successful particle in order
designer is free to choose the matrices Q and R, but to improve themselves, whereas in GA, the
the selection of matrices Q and R is normally based worse solutions are discarded and only the
on an iterative procedure using experience and good ones are saved; therefore, in GA the
physical understanding of the problems involved.
population evolves around a subset of the best
individuals. 3. Compare evaluation with particles previous best
The position vector x and the velocity vector v of n value ( pbest i ):
particles in the D dimensional search space can be
if current value < pbesti then
represented respectively as
pbesti = current value
x=[ x d1 , x d2 , . x di . x dn ] , i {1,2, n } pbestdi = current position in D dimensions

4. Compare evaluation with groups previous best


d {1,2, . D } ,
value ( pbest g ):
v =[ v 1 , v 2 , . v di . v dn ] ,i { 1,2, n }
d d
if current value < pbest g ): then
g = particles array index
d {1,2, . D } ,
5. Change the velocity and position of the particle
according to equation (27) and (28)
According to a user defined fitness function, the best
position pbest of each particle and the best of the
6. Change the inertia weight factor and
fittest particle gbest found so far can be represented
acceleration coefficients c1 and c2 .
respectively as

d 7. Loop to step 2 and repeat until the terminating


pbest=pbest i ,i {1,2, n } , d {1,2, . D } condition is met.

gbest= pbestdg , g { 1,2, n } , d { 1,2, . D } PSO algorithm stages for searching proper weighting
matrices are as follows:
Then, the new velocities and the positions of the First, the lower and upper bounds of the parameters
particles for the next fitness evaluation are calculated are specified and the population of the particles is
using the following two equations: initialized randomly. The controller gains are
calculated using LQR command in MATLAB
software. It is important to note that two conditions
v di ( t +1 )=w v di ( t )+ c1 rand 1 ( pbest di ( t ) xdi ( t ) ) + c2 det(R) d
rand 2 ( pbest
0 and ( ) x di (>t )0) must be satisfied. After
g tdet(Q)
(27) that, the control energy matrix (u) and state variable
x di ( t+1 )= xdi ( t ) +vid ( t+1 ) (28) (x) are calculated. Then, the linear quadratic cost
function (J) is evaluated for each particle. If the cost
Where is the inertia weight factor, c1 and c2 are for local best solution is less than cost of the current
constants known as acceleration coefficients, and are global best solution, the global solution is replaced
two separately generated uniformly distributed with the local solution. In each stage, the program
random numbers in the range [0, 1]. Motivated by the saves the cost value and minimum error value and
desire to better control the scope of the search, Shi according to equation (26). If the number of iterations
and Eberhart [20-21] have observed that the optimal reaches the maximum designated by the user, the
solution can be improved by varying the value of the latest global best solution is recorded and the
inertia weight from 0.9 at the beginning of the algorithm comes to the end. The pseudo code of the
search to 0.4 at the end of the search for most PSO based LQ regulator is given as follows:
problems. For each particle:
Initialize each particle
PSO Steps End
Do
The steps of particle swarm optimization algorithm For each particle:
are as follows [22]: Solve P from algebraic Riccati equation (26)
Calculate the feedback gain K by (25)
1. Initialize n number of particles with random Calculate the Integral Absolute Error (IAE),
positions x and velocities v on D dimensions in the x
search space. 2
( d (i ) x (i )) , xd is the desired state
2. For each particle, evaluate the desired optimization E IAE=v
fitness function in D variables. i
trajectory. If the present IAE value is less than the such as GA and trial and error. The settling time of
previous best IAE value, update the present value the system for PSO tuning is also lesser which makes
(pbest), set current value as the new pbest. the pendulum system to stabilize in minimum time.
End When the upright position has been reached, the
Choose the particle with the minimum IAE value system is switched to stabilizing control mode to
of all the particles as the gbest. stabilize the pendulum. By proper selection of the
For each particle: weight matrices in LQR design, the optimal control
Calculate particle velocity by (27). signal is generated which can satisfy the design
Update particle position according to (28). criteria mentioned in section1. The optimal values of
If one of the elements of the Q and R is Q and R obtained manual tuning are obtained as
negative, retain the current particle position

[ ]
End
Until the stopping criterion is met. 2.5 0 0 0
0
Q= 0
In order to investigate the performance of PSO 8 0 R= [ 0.0001 ]
algorithm for determination of proper weighting
0 0 0 0
matrices in LQR controller design, the simulation is
carried out and the results are compared with the 0 0 0 0
results obtained by GA.
The corresponding feedback gain vector for the
weighing matrices is
6. SIMULATION RESULTS
The PSO based optimization and simulation work is
K=[ 158.11 390.40 116.01 38.67 ]
facilitated by Matlab 2006. For simulation purposes,
the pendulum is swung-up from the rest (downward GA
pendulum) by giving sinusoidal input. The travel The parameters used in GA are population size=50
distance of the cart is limited by the amplitude of the chromosomes, cross over rate=0.96, mutation
reference input to 0.3 m . Once the pendulum rate=0.1, search interval=[0,20] and generation
achieves the upward position, the control is number=60. The Q and R matrices determined by
transferred to stabilizing controller which maintains Genetic Algorithm are
the pendulum in upright position. The pendulum

[ ]
angle response of the system for the weights obtained 1 0 0 0
by three methods such as PSO, GA and trial and error 0 0
is shown in Figure 6. The swing up controller is able Q= 0 4.5 R= [ 0.0003 ]
to swing up the pendulum to upright position in less 0 0 0 0
than 5 sec. 0 0 0 0
1
GA
0.8 Trial and Error The corresponding feedback gain vector is
PSO
0.6

0.4 K=[ 57.74 182.27 50.52 19.51 ]


PSO
Angle (deg)

0.2

0 The parameters used in the PSO algorithm are


-0.2 population size=50 particles, search interval=[1,20],
-0.4
generation number=60 and acceleration constants:
c1=c2=1.5. The optimal weighting matrices of LQR
-0.6
by PSO algorithm is as follows:

[ ]
-0.8

-1
0 5 10 15 20 25 30
0.4 0 0 0
0 0
Q= 0 4.5
Time (s)
R= [ 0.0002 ]
Figure 6: Pendulum Angle 0 0 0 0
It is observed that the optimal weights obtained using
0 0 0 0
PSO algorithm has a maximum deviation of 0.2
degrees which is lesser than other two approaches The corresponding feedback gain vector is
K=[ 91.29 207.22 65.95 23.50 ]
In this paper, a PSO based optimal weight selection
25
PSO

of LQR Controller to stabilize the pendulum system
20 GA in upright position and to achieve the trajectory
TrialandError
15 ReferenceSignal
tracking has been presented. This approach has been
motivated especially by the fact that the designers
Cart position (cm)

10
often have to face trial and error approach to
5
determine the optimal weighting matrices in case of
0 LQ regulator design. The linearized model of
5 inverted pendulum is obtained to implement the state
10 feedback controller. The effectiveness of the
15
proposed control scheme is compared with the GA
based tuning of LQR and trial and error based
20
approach of LQR tuning. Simulation results
25
0 5 10 15 20 25 30 demonstrate that the proposed method is more
Time (s) efficient and robust compared with other heuristic
Figure 7: Cart Position method i.e., Genetic Algorithm (GA) method. It is
also observed that the speed of computation in PSO is
The reference signal tracking ability of the pendulum very less compare to GA and trial and error method.
system is shown in Figure 7. The overshoot and
settling time in case of PSO tuning is found to be less REFERENCES
compared to other two methods.
Table 2 1. Jia-Jun Wang, Simulation studies of inverted
Comparison of Performance Index of Tuning pendulum based on PID controllers, Simulation
Methods Modelling Practice and Theory, pp. 440449,
Performance Trial and 2011.
PSO GA 2. K.J. strm, K. Furuta, Swinging up a
Index Error
IAE 0.9897 1.022 1.045 pendulum by energy control, Automatica 36 (2)
(2000) 287295.
In order to assess the performance of the controller, 3. P. Mason, M. Broucke, B. Piccoli, Time optimal
the Integral Absolute Error is calculated for all the swing-up of the planar pendulum, IEEE
three methods and it is given in Table 2. The value of Transactions on Automatic Control 53 (8) (2008)
IAE in case of PSO is the least among the three 18761886.
methods and it corroborates for the better 4. C.W. Tao, J.S. Taur, T.W. Hsieh, C.L. Tsai,
performance of the PSO tuning. Design of a fuzzy controller with fuzzy swing-
up and parallel distributed pole assignment
schemes for an inverted pendulum and cart
7. CONCLUSION
system, IEEE Transactions on Control Systems system, IET Control Theory and Application 1
Technology 16 (6) (2008) 12771288. (4) (2007) 979986.
5. R. Shahnazi, T.M.R. Akbarzadeh, PI adaptive 9. Kalman, R. E., When is a linear control system
fuzzy control with large and fast disturbance optimal?, J. Basci Eng. Trans., ASME-86D:51-
rejection for a class of uncertain nonlinear 56,1964.
systems, IEEE Transactions on Fuzzy Systems 10. Wang Yaoqing, The determination of weighting
16 (1) (2008) 187197. matrices in LQ optimal control system, ACTA
6. N.A. Chaturvedi, N.H. McClamroch, D.S. Automatic Sinica, 1992.
Bernstein, Stabilization of a 3D axially 11.Johnny Lam,Control of an inverted pendulum,
symmetric pendulum, Automatica 44 (9) (2008) http://www.ece.ucsb.edu/~roy/student_projects/
22582265. Johnny_Lam_report_238.pdf
7. R.J. Wai, L.J. Chang, Adaptive stabilizing and 12. Mahmud Iwan Solihin, Rini
tracking control for a nonlinear inverted- Akmeliawati, Particle Swam
pendulum system via sliding-mode technique, Optimization for Stabilizing Controller
IEEE Transactions on Industrial Electronics 53 of a Self-erecting Linear Inverted
(2) (2006) 674692. Pendulum, International Journal of
8. L.H. Chang, A.C. Lee, Design of nonlinear Electrical and Electronic Systems
controller for bi-axial inverted pendulum Research, Vol.3, June 2010.
13. A.N.K. Nasir, 1M.A. Ahmad and M.F. Rahmat, 19. A.P. Engelbrecht, Particle swarm optimization:
Performance Comparison between LQR and where does it belong? in Proc. of the IEEE
PID controller for an Inverted Pendulum Swarm Intelligence Symposium 2006 (SIS'06),
system International Conference on Power May 2006, pp. 48-54.
Control and Optimization, Chiang May, 20. J. Kennedy, and R. Eberhart, "Particle swarm
Thailand, 18-20, July 2008. optimization," Proceedings of the IEEE
14. S. Mobaieen, B. Mohamady, H. Ghorbani and A. International Conference on Neural Networks
Rabii, Optimal Control Design Using (ICNN), Vol. 4, Nov. 1995, pp. 1942-1948.
Evolutionary Algorithms with Application to an 21. R. Eberhart, and J. Kennedy, "A new optimizer
Aircraft Landing System, Journal of Basic and using particle swarm theory," in Proc. 6th Int.
Applied Scientific Research, 2012. Symp. Micro Machine and Human Science
15. G. K. Venayagamoorthy, and R. G. Harley (MHS), Oct. 1995, pp. 39-43.
Swarm Intelligence for Transmission System 22. Che-Cheng Chang , Jichiang Tsai and Shi-Jia Pei
Control, Proceedings of the IEEE Conference on ,A Quantum PSO Algorithm for Feedback
Power Engineering, 2007. Control of Semi-Autonomous Driver Assistance
16. Y. Shi and R. C. Eberhart, Empirical study of Systems, 12th International Conference on ITS
particle swarm optimization, in Proc. IEEE Int. Telecommunications, 2012.
Congr. Evolutionary Computation, vol. 3, pp. 23.S. Yang, M. Wang, and L. Jiao, "A quantum
101106, (1999). particle swarm optimization," in Proc. of the
17. J.Hamidi, Control System Design Using IEEE Congress on Evolutionary Computation
Particle Swarm Optimization, International (CEC), Vol. 1, pp. 320-324, June 2004.
Journal of Soft Computing and Engineering
(IJSCE, ISSN: 2231-2307, Volume-1, Issue-6,
January 2012.
18. Ker-Wei, Yu, Zhi-Liang, Huang, LQ Regulator
Design Based on Particle Swarm Optimization,
Proceedings of the IEEE Conference on Systems,
Man, and Cybernetics, October, 2006.

You might also like