Documente Academic
Documente Profesional
Documente Cultură
George Saikalis
Hitachi America, Ltd.
Research and Development Division
34500 Grand River Avenue
Farmington Hills, Michigan 48335
Feng Lin
Wayne State University
Department of Electrical and Computer Engineering
5050 Anthony Wayne Drive
Detroit, Michigan 48202
Abstract
Keywords:
1.
Introduction
Since their rebirth in 1980s, neural networks have found applications in many
engineering fields, including control. For example, neural networks have been
used for system identification [1] [2] [3] and adaptive control [4] [5] [6] [7]. Neural
network controllers can control not only linear systems but also nonlinear
systems [8] [9] [10] [11] [12] [13]. Neural network control designs are divided into
two major categories: (1) the direct design where the controller is a neural
network [14] [15] and (2) the indirect design where the controller is not itself a
neural network, but uses neural networks in its design and adaptation [16] [17].
Issues such as robustness [18] and stability [19] have also been discussed.
Many books on neural network control have been published, including [20] [21]
[22] [23].
There are two major factors that contribute to the popularity of neural networks.
The first factor is the ability of neural networks to approximate arbitrary nonlinear
functions [24] [25]. This is important because in many cases control objectives
can be more effectively achieved by using a nonlinear controller. The second
factor is the capability of neural networks to adapt [25] [26]. In fact, the way for
neural networks to adapt is very natural. It requires no model building or
parameter identification. Such a natural adaptivity is rather unique among manmade systems (but abundant in natural systems). It makes control design a much
easy job. For example, we all know how difficult it is to design a nonlinear
controller. However, if we can let a neural network controller to adapt itself, then
we can sit back and relax. (We know that this will make some people nervous, as
they will insist on the proof of stability.)
To adapt neural networks, many learning (or adaptation) algorithms have been
proposed, the two essential categories being the supervised learning and the
unsupervised learning [5] [25] [26]. Within each of these categories, there are
algorithms for feedback and feedforward neural networks. For the unsupervised
However, there is one main obstacle in the way to adapt neural network
controllers. That is, some most efficient adaptation algorithms such as backpropagation algorithm cannot be applied directly to neural network controllers. To
use back-propagation algorithm, the system must consist of pure neurons. This
is because the back-propagation algorithm relies on a dedicated feedback
network to propagate the error back. No such network can be constructed if the
original system does not consist of pure neurons. However, the neural network
control system is hybrid because the plant to be controlled is usually not a
neural network. Therefore is it not possible to apply back-propagation algorithm
to adapt the controller directly.
To bypass this obstacle, people have tried to approximate the plant with a neural
network. But this may not always work because of the error in approximation. So,
what can we do? Fortunately, there is one adaptation algorithm proposed by
Brandt and Lin [27] that can do the same job as the back-propagation algorithm
but requires no feedback networks.
Using Brandt-Lin algorithm, the errors required for adaptation is inferred from
local information in such a way that the error back-propagation is done implicitly
rather than explicitly. As a result, Brandt-Lin algorithm can be implemented in a
simple
and
straightforward
manner
without
using
feedback
network.
The rest of the paper will be organized as follows. In Section 2, we will introduce
the theory of adaptive interaction and review Brandt-Lin algorithm for adaptation
in neural networks. In Section 3, we will propose our adaptive neural network
controller and apply Brandt-Lin algorithm to derive the adaptation law for the
controller. Simulation results will be presented in Section 4.
2.
n n ,
where n and n are the input and output spaces respectively. Therefore, the
relation between input and output of the nth device is given by:
y n ( t ) = (Fn o x n )( t ) = Fn [ x n ( t )],
PreC1
C1
Device 1
Device 2
C3
C2
C4
Device 3
Device 4
Device 5
In this paper, the following notations are used to represent relations between
devices and connections:
prec is the device whose output is conveyed by connection c,
postc is the device whose input depends on the signal conveyed by c,
In = { c : prec = n } is the set of input interactions for the nth device, and
On = { c : postc = n } is the set of output interactions for the nth device.
We assume linear interaction among devices and external signal un(t), that is,
x n ( t) = un (t) +
cIn
y pre c ( t ),
c In
y pre c ( t ) ],
nN
The goal of the adaptation algorithm is to adapt the connection weights c so the
performance index E(y1, .,yn, u1, .,un) as a function of the inputs and outputs
will be minimized. To present the algorithm, we must first introduce the Frchet
derivatives [30]. As described in [30], let T be a transformation defined on an
open domain D in a normed space X and having range in a normed space Y. If
for a fixed xD and each hX there exists T(x;h)Y which is linear and
continuous with respect to h such that
lim
|| T ( x + h) T ( x ) T ( x ; h) ||
=0
||h||0
|| h ||
The adaptation algorithm is given in the following theorem [27]. For the sake of
simplicity, the explicit reference to time is removed.
Theorem:
For the system with dynamics given by
y n = Fn [u n +
cIn
y pre c ],
nN
dE
s [ x post s ]
o F post
dy post s
E
c [ x post c ] o y pre c ,
) o F post
& c = ( s& s
dE
y
s O post c
post
c
s [ x post s ] o y post c
o F post
dy post s
cC
-------- (1)
where > 0 is the adaptation coefficient. If (1) has a unique solution for & c , cC
(that is, the Jacobian determinant must not be zero in the region of interest), then
the performance index E(y1, .,yn, u1, .,un) will decrease monotonically with
time and the following equation is always satisfied:
& c =
dE
,
d c
cC
E
. s & s Fpost
-------- (2)
c .[ x post c ].y pre c .
sO
y
post
c
post c
The above equations can be applied to a very general class of systems, including
neural networks, as shown below.
x1
r1
w1
p3
(.)
r3
w5
Log-Sig
w2
w3
x2
r2
w4
p4
(.)
r4
w6
Log-Sig
p5
(.)
r5
1
.
1 + e x
sDn
s pre s
rn = (p n )
If we denote
n =
1 d
2 dt
s A n
2
s
-------- (3)
&s
w
s A n
then by applying the adaptation law in (2), the weight adaptation becomes:
& s = rpre ( post ( p post ) + fpost )
w
s
s
s
s
-------- (4)
Equations (3) and (4) describe Brandt-Lin algorithm for adaptation in neural
networks. As shown in [27], it is equivalent to back-propagation algorithm but
requires no feedback network to back-propagate the error.
3.
Input Excitation
Signal
error
Neural Network
Controller
Proposed
Neural Network
Adaptation
Algorithm
Plant
Gp(s)
W1
W2
|
|
Wn
To be more specific, the neural network controller have two inputs e1 and e2. e1 is
the error between the set point and the plant output and e2 is a delayed signal
based on e1.
The reason for introducing e2 is as follows. Since the neural network controller is
itself a memory-less device, in order for control output to depend not only on the
current input (error in our case), but also on past inputs, some delayed signals
must be introduced. In this paper, we will consider only one simple delayed
signal. However, in principle, multiple delayed signals can be introduced (that is,
the neural network controller will have more than two inputs). Hence, the
configuration of the neural network controller is further described in Figure 4.
e1
Neural Network
Controller
Gp(s)
Delay
e2
If we use the simple neural network with two hidden neurons as in Figure 2, then
the neural network controller is shown in Figure 5. More sophisticated neural
network can be used to improve the performance.
e1
r1
w1
p3
r3
w5
A
Log-Sig
w2
p5
+/- Sig
w3
e2
r2
w4
w6
p4
Log-Sig
The reason behind the tangent sigmoid (tan-sig) is the ability to provide a dual
polarity signal to the output. Based on simulation results, the simple constant
gain output will also work and often provide a better result.
10
r1 = e1
and
r2 = e2
p3 = w1r1 + w2r2
and
p4 = w3r1 + w4r2
r3 = (p3)
and
r4 = (p3)
p5 = w5r3 + w6r4
Let
E = e 12 = ( r y )2 = r2 2.y.r + y2
Then
E
= 2 .r + 2 .y = 2 .( r y ) = 2 .e 1 .
y
The adaptation law for w5 and w6 is more complicated as it is linked to the plant
to be controlled. By Equation (2), since Opostc is empty, we have
w&
c .[ u ]. r 3 .( 2 .e 1 )
= .F post
Similarly,
w& 6 = .r 4 .e 1
11
4.
Simulation results
4.1.
Matlab/Simulink model
12
4.2.
This section covers the investigation of the effect of the initial weights on the
convergence of the algorithm. The following elements are set during the
simulation.
88 . 76
s ( s + 21 . 526 )( s + 2 .474 )
Plant:
G( s ) =
Input Signal:
Amplitude: 10
Type: Sinewave
requency: 0.01 Hz
Output Stage:
Tangent Sigmoid
Learning Rate:
=10
The results of four simulations with different initial weights are shown in Figures
7-10 and summarized in Table 1. It is observed that te initial weights must have
opposite signs in the hidden units of the neuron connection link.
Table 1: Effects of the initial weights on adaptation
Figure Number
Initial Weights
Results (500s)
Figure 7
Adapted
Adapted
Not Adapting
13
Not Adapting
Figure 7
Figure 8
14
Figure 9
Figure 10
15
4.3.
This section covers the effects of the learning rates on the adaptation. The
following elements are set during simulation.
88 .76
s(s + 21 .526 )(s + 2 .474 )
Plant:
G(s ) =
Input Signal:
Type: Sinewave
Amplitude: 5
Offset: 5
Frequency: 0.01 Hz
Output Stage:
Tangent Sigmoid
Initial Weights:
The results of three simulations with different learning rates are shown in Figures
11-13 and summarized in Table 2. It is observed that the larger the learning rate,
the faster the algorithm will adapt. However, if the learning rate is too large, the
output may not be robust and may lead to the system breaking-up. Also, the
weights converge to local minima depending on the different learning rates.
Table 2: Effects of the learning rate on the adaptation algorithm
Figure Number
Learning Rate
Results (500s)
Figure 11
=100
Adapted
Figure 12
=10
Adapted
Figure 13
=1
Adapted
16
Figure 11
Figure 12
17
Figure 13
4.4.
This section covers the effect of changing the output stage from a tan-sigmoid to
a constant gain. The following elements are set during simulation.
88.76
s(s + 21.526 )(s + 2.474 )
Plant:
G(s) =
Input Signal:
Type: Sinewave
Amplitude: 10
Learning Rate:
=10
Initial Weights:
The results of four simulations with two different frequencies are shown in
Figures 14-17 and summarized in Table 3. It is observed that it is easier for the
controller to adapt if the input frequency is low. Also, a constant gain output
provide better adaptation at higher input frequencies.
18
Input Frequency
Output Stage
Results (500s)
Figure 14
0.01 Hz
Tan-sigmoid
Adapted
Figure 15
0.01 Hz
Gain=0.001
Adapted
Figure 16
0.1 Hz
Tan-sigmoid
Not Adapted
Figure 17
0.1 Hz
Gain=0.001
Adapted
Figure 14
Figure 15
19
Figure 16
Figure 17
20
4.5.
To further validate the adaptation algorithm, the neural network based adaptive
controller is applied to different plants.
Plants:
Input Signal:
G 2( s ) =
1000
( s + 10 )( s + 5)
G 3( s ) =
5000
s ( s + 5)( s + 100 )
G 4( s ) =
5000
( s + 1)( s + 5 )( s + 100 )
Type: Sinewave
Amplitude: 10
Output Gain:
0.001
Initial Weights:
We change the learning rate and input frequency for these plants and see how
high the frequency can be increased. The results of five simulations are shown in
Figures 18-22 and summarized in Table 4. It is observed that the input frequency
can be increased to 10 Hz for G2(s). With third order plants G3(s) and G4(s), a
maximum of 1 Hz input signal is possible. Note that G3(s) is open loop unstable.
Table 4: Effects of input frequency and learning rate on G2(s)
Figure Number
Plant
Input Frequency
Learning Rate
Results (500s)
Figure 18
G2(s)
0.01 Hz
=10
Adapted
Figure 19
G2(s)
0.01 Hz
=100
Adapted
Figure 20
G2(s)
10 Hz
=100
Adapted
Figure 21
G3(s)
1 Hz
=10
Adapted
Figure 22
G4(s)
1 Hz
=10
Adapted
21
Figure 18
Figure 19
22
Figure 20
Figure 21
23
Figure 22
4.6.
A non-minimum phase system has either a pole or a zero in the right-half of the
s-plane. Since it is well known that it is difficult to apply adaptive control to nonminimum phase systems, we decide to test the following non-minimum phase
system.
500
( s 1)( s + 5 )
Plants:
G 6(s) =
Input Signal:
Type: Sinewave
Amplitude: 10
Learning Rate:
=10
The results of four simulations with different frequencies, output stage and initial
weights are shown in Figures 23-26 and summarized in Table 5. It is observed
that the weight adaptation does occur. The adaptation convergence depends on
two factors: (1) Frequency of the input signal and (2) the magnitude of the initial
weights. The constant gain (= 0.001) is required when dealing with large initial
24
weights and higher frequency. It was found that the tangent sigmoid is suited
when the initial weights are small and the input frequency is low.
Table 6: Effects of the input frequency, output stage and initial weights on G6(s).
Figure
Input
Number
Frequency
Stage
Figure 23
0.01
Gain=
0.001
Gain=
0.001
tan-sig
Figure 24
Figure 25
0.01
Results (500s)
Adapted
Adapted
Adapted
Gain=
0.001
Figure 23
25
Adapted
Figure 24
Figure 25
26
Figure 26
5.
Conclusion
Learning works well with a variety of second and third order plants.
For higher input frequencies and large initial weights the output stage
with a constant gain works better.
This new approach does not require the transformation of the continuous time
domain plant into its neural network equivalent. Another benefit for applying the
27
6.
References
[1]
[2]
[3]
[4]
[5]
[6]
T. Yamada and T. Yabuta, Neural Network Controller Using Autotuning Method for
Nonlinear Functions, IEEE Transactions on Neural Networks, Vol. 3, pp 595-601, 1992.
[7]
[8]
[9]
[10]
[11]
I. Rivals and L. Personnaz, Non-linear Internal Model Control Using Neural Networks,
Application to Processes with Delay and Design Issues, IEEE Transactions on Neural
Networks, Vol. 11, pp 80-90, 2000.
[12]
28
[13]
[14]
[15]
J. Noriega and H. Wang, A Direct Adaptive Neural Network Control for Unknown
Nonlinear Systems and Its Application, IEEE Transactions on Neural Networks, Vol. 9,
pp 27-33, 1998.
[16]
[17]
[18]
[19]
[20]
J.J.E. Slotine and L. Weiping, Applied Nonlinear Control, Prentice Hall, 1989.
[21]
[22]
C. J. Harris, C. G. Moore and M. Brown, Intelligent Control: Aspects of Fuzzy Logic and
Neural Nets, World Scientific, Chap. 1.7 and 8, 1993.
[23]
H. Demuth, M. Beale, Neural Network Toolbox for MatLab, The Mathworks, Version 3,
1998.
[24]
[25]
D. S. Chen and R. C. Jain, A Robust Back Propagation Learning Algorithm for Function
Approximation, IEEE Transactions on Neural Networks, Vol. 5, pp 467-479, 1994.
[26]
[27]
[28]
[29]
29
[30]
30