Академический Документы
Профессиональный Документы
Культура Документы
C. H. TSENG A N D J. S. A R O R A
Optimal Design Laboratory, College o f Engineering, The University o f Iowa, Iowa City,
IA 52242-1593, U.S.A.
The present paper demonstrates the use of an interactive design optimization environment for the dynamic
response application, including control of systems. It is
shown that not only the efficiency of the optimal design
process can be improved but also that some difficult
minimum time optimal control problems can be solved
with an alternate formulation. During the interactive
session, the designer can refine the model for the
problem, based on the insights gained, to obtain a
more accurate solution. Therefore interaction in the
computer-aided design optimization process can be very
useful for practical applications.
In Section 2, a general mathematical model for
optimal design and control of dynamic systems is
defined. A prototype interactive environment for the
selected application is described in Section 3. Example
problems are presented in Section 4 and conclusions are
given in Section 5.
I"
(1)
to
to <~t <~tf
(2)
~ = g~(b,z(tf),tf)+
F~(b,u(t),z(t),t) dt
to
= 0; c = 1
r'
~<0; a = r ' + 1 ..... r
(3)
~z(b,u(t),z(t),t)
= 0;/3 = 1 , . . . , q '
~ < 0 ; / 3 = q ' + l ..... q
(4)
In the preceding definition, z(t)~ R n is the state variable that is assumed to be a continuous function of
time, and u(t)ER m is assumed to be bounded and
measurable on the interval [to, tf], where to and tf are
the initial and terminal times; to is kept fixed and tf is
167
Data Entry I
Design
UserSupplied
Problem
Definition
I/11
Interactive
Graphics
Output:
Final Design
IDESIGN
Fig. 1. Conceptual layout
optimization environment
of
interactive
design
IDESIGN
3. I N T E R A C T I V E E N V I R O N M E N T FOR
DYNAMIC RESPONSE OPTIMIZATION
To create an interactive design optimization environment for the dynamic systems modeled by equations (1)
to (4), proper hardware and software are needed. In
addition, several software components must be properly integrated. Here we describe a possible environment for the dynamic response applications. Using this
prototype, an interactive design optimization environment for other applications can be created.
The Apollo workstations, the general purpose design
optimization shell IDESIGN 1'2'6, and other software
components are used to create the present environment.
Figure 1 shows the overall arrangement of the system.
An application is installed into IDESIGN through usersupplied subroutines that define problem functions and
their gradients. Two levels of interaction are possible,
as shown in the conceptual layout of the system in
Fig. 1. The first level is the general purpose interaction
shown on the left side. This type of interaction is not
tied to any application and uses only general optimization terminology. It can be easily incorporated in the
general purpose optimizer. The domain-specific interaction is shown on the right side of the figure. This type
of interaction uses the domain-specific terminology and
commands, and must be developed for each application. This has been done for the present application, as
will be described later in the paper. The software con-
168
7-
;-
Fig. 2. Conceptual arrangement of user-supplied subroutines for dynamic response optimization application
taining dynamic response optimization capabilities is
called OCP (Optimal Control Problem).
The interactive capabilities of IDESIGN and their
usage are described in detail in Refs 1 and 2, so they will
not be repeated here. Additional capabilities for the
dynamic response application will be discussed. The
capability for optimum design of dynamic systems
modeled by equations (1) to (4) is installed into
IDESIGN by creating the following four subroutines:
USERMF for the cost function, USERCF for the
constraint functions, USERMG for the cost function
gradient, and USERCG for the constraint gradients.
Through these routines, application-dependent software, such as differential equation solvers, interpolation routines, gradient evaluation routines and integral
evaluation routines, are called to evaluate problem
functions and their gradients. Domain-specific interactive commands can be added in these subroutines.
Figure 2 shows the arrangement of the user-supplied
TBFIT:
CTRLPR:
CTRLFT:
ANALYS:
CTRLF:
INGSOL:
PREPRO:
ZOFN:
DIFSOL:
Control subroutine that calculates the performance index (cost function), partitions
the incore array, and keeps track of its
usage.
Subroutine to read additional input data
for the dynamic and control optimization
problem.
Subroutine to interpolate the control function with respect to the design variable
vector.
Subroutine to set up initial conditions for
the state equation and to calculate state
trajectories; it calls several other routines
to accomplish its function.
Control subroutine to calculate coefficients
of the interpolation function, function
values, or their first derivatives at a given
grid point. Subroutine calls TBFIT and
TGVAL.
Subroutine to control method for numerical integration.
Subroutine to calculate value of the
problem functions, defined in equations
(1), (3) and (4), at a specified time t.
Initial state z(t0) evaluation subroutine.
Subroutine to control the solution method
for differential equations.
169
zl~__
,2t _
C.G,@ )+
~J z 5
m2g
c,
z'L
Cs~b
k2~>ljC2
.
f2())
jC4
,,I,)l_r_L"1
v(y)
L .....
36o"
5"
:"
(a)
Pro, f l i t ,
N,~. I
v(y)
t
~.y
(b)
Profile
~n. 2
lmeraetive Change
I D F r W C = 2; N G P T = 101:
INTGPT = 32
RELERR = 3.E-5; ABSERR = I.E-5
ACS = 0.5; A C V = 0.01
5.07E-01
....
1.986E+02
Design change
No. Value
5 2.500E+01
6 2.500E+01
1.23E-01
---
1.986E+02
I
8
by linear exttapolatlon
Suggest
Choice
5.449E+01 5.449E+01
8.000E+01 8.000E+01
Numbea'ofuallsforfuuetiodaevalualion = 29
Total n u ~
of gradients evaluated
= 1568
CPU = 38528 seconds for APOLLO DN460
With worst case ~,~,at~em o f ' d y n a m i c ~ s t r a i n t s and batch environment [51
Cost function value = 9.894E+01
Numbeaofc~lls fc function evaluation = 60
Total number of gradients evaluated
= 159
CPU = 59464 seconds for APOLLO DN460
Q
maximum acceleration of the driver's seat over the specified time and the road conditions. The problem formulation can be transcribed into the form of equations
(1) to (4).
The problem is solved using a two phase process.
During the initial phase, design changes are made
according to the suggestions given by IDESIGN
without doing any optimization. The numerical
schemes are selected according to their performance
with respect to efficiency. During the second phase, the
numerical schemes are selected to obtain a more
accurate solution. The step-by-step procedure is illustrated in Table 1 and explained as follows:
Stage I: Inaccurate solution is obtained with the following parameter values: N G P T = 101, I N T G P T = 32,
RELERR=3.E-5
and A B S E R R = I . E - 5
(error
tolerance for integrating differential equations),
I D P T W C = 2 (the max treatment of dynamic constraints among the grid pointsa), ACV = 10070 (tolerance for constraint violation), and A C S = 0 . 5
(convergence parameter for IDESIGN).
Iteration 1 :
Maximum violation of constraints is large (50.770).
IDESIGN gives suggestions, without optimization,
for changing design variable numbers 5 and 6 to
improve constraint violations. After the suggested
design changes, the maximum violation is still large
(12.3070). Another suggestion is given for changing
the design variables 1, 4, and 5. The maximum violation is reduced to 1.2070 after the suggested changes
are made. At this stage, a command to obtain the
optimum solution is given.
Iteration 8:
Optimum solution is found with relaxed convergence
criteria. The minimum acceleration is about 12070
higher than the true optimum.
Stage 11." To obtain a more accurate solution, the
iterative process is restarted from iteration number 8
with the parameters and schemes as N G P T = 1001,
I N T G P T = 128,
R E L E R R = 3" E - 7,
ABSERR =
1.E-7,
IDPTWC=4
(the hybrid treatment of
dynamic constraint described in Refs 4 and 5), and
ACV = 0.1070 and ACS = 0.01 for IDESIGN.
Iteration 8:
The maximum violation is the same as in the first
stage. The command to obtain optimum solution is
given.
Iteration 22:
The optimum solution found is the same as the one
given in the literature 4'5'7.
Data for comparison between interactive and batch
solution procedures are given in Table 1. It can be seen
that the problem is solved more efficiently with the
interactive procedure. As seen in Table 1, the total
number of iterations and the CPU time are reduced
substantially compared with results given in Refs 4 and
5.
From the results for this example, we see that with
reasonable interactive changes, we can solve dynamic
response optimization problems more efficiently as well
as accurately.
171
7,0
Stage
Max. Vio.
Conv. Para.
P1
1
2
3
4
9
5.0000E+O0
1.2410E-05
2.9292E-06
2.1550E-06
1.8534E-05
1.0000E+00
1.1959E-02
1.1557E-02
9.6765E-03
7.0102E-04
0.0000E+O0
|.6382E-01
1.6363E-01
1.6277E-01
1.6001E-OI
Optimal
I1
9
19
1.2500E+00
1.1030E-05
9.9785E-01
3.3716E-04
1.6001E-01
7.6930E-0 l
2.5
Optimal
II1
19
20
2I
22
23
4.0000E-01
3,7878E-01
3.7862E-01
3.7860E-01
3.6666E-01
9.1396E+00
1.2292E+03
5.5097E+01
5.2948E+00
7.8681E+00
7.6930E-01
8.3701E-01
8.6526E-01
8.6533E-01
8.9445E-01
2.0
Restart/19
IV
19
23
9.5994E-02
2.9773E-02
9.7985E-01
3.1322E+00
7.6930E-0!
9.0926E-01
2.4
Restart/19
19
23
4.9000E-02
2.6534E-07
3.5677E-01
3.1743E-04
7.6930E-01
8.4935E-01
2.45
Optim',d
VI
23
24
2.4745E-02
1.3464E-04
3.961 IE-01
5.3055E-04
8.4935E-01
9.1781E-01
2.425
Opdmal
VII
24
28
4.9858E-03
3.5331E-03
1.0743E-01
2.1537E-01
9.1781E-01
9.2441E-01
2.42
Restarl/24
VIII
24
28
1.9944E-03
8.8149E-04
4.2557E-02
1.0820E-01
9.1781E-01
9.2263E-01
2.423
RestarT24
IX
24
30
9.9563E-04
6.4601E-05
2.1182E-02
6.3473E-05
9A781E-01
9.2198E-01
2.424
Optimal
ts
z2 = u(t)
lu(t)l~<l.0
for
her.
tf
5.0
Zz(tf)=O
172
Iteration 9:
Optimum is found with the performance index (control effort) as 0.16.
Stage 11: Terminal time is reduced to 2.5 seconds and
the process restarted from Iteration 9.
Iteration 9:
Maximum violation is 125o70 and performance index
is 0.16.
Iteration 19:
Optimum is found with the performance index as
0.769.
Stage IlL" The terminal time is reduced to 2 seconds
and the process restarted from iteration 19.
Iteration 19:
Maximum violation is 4070, and the convergence parameter has large value, 9.14.
Iteration 23:
After 5 more trials, the constraint correction is very
slow and convergence parameter is still large. It is
concluded that the minimum control effort problem
with the terminal time as 2 seconds is infeasible.
Stage IV: Since the problem with the terminal time as
2 seconds in Stage III is not feasible, the terminal time
needs to be increased. It is increased to 2.4 seconds and
the process is restarted from iteration 19. After 5 more
trials, the behavior is almost the same as in Stage III.
Therefore, the problem is still unfeasible. The optimum
terminal time for the original minimum time control
problem is between 2.4 and 2.5 seconds. If the designer
is satisfied with this solution, the interactive process can
be terminated here.
5. CONCLUDING REMARKS
Interactive use of a prototype system for o p t i m u m
design and control of dynamic systems is demonstrated.
Advantages of designer interaction are demonstrated
by solving two simple examples. During problem
execution, the system allows the designer to control the
optimization process, and the numerical procedures
for analysis and design sensitivity analysis. The five
degrees of freedom vehicle suspension problem is
solved more efficiently and accurately by the use of
interactive facilities. The number of iterations and the
C P U time are reduced considerably compared to those
with the automated environment. The minimum time
control problem is solved using the interactive facilities
and the minimum control effort formulation.
The present study and other recent results 1'2 show
that designer interaction during the optimization
process is quite beneficial. For example, the designer
can change the problem formulation without quitting
the iterative process. For the present examples, the
number of time grid points, the method o f treatment of
time dependent constraints, and the accuracy controls
for generating the state trajectories and the final design,
are changed interactively. This interactive environment
can be easily expanded to include more facilities as well
as other applications. For the finite element analysis
applications, one m a y start with a coarser discretization
and interactively refine the model at a later stage. This
way, practical solutions with desired accuracy will be
obtained more efficiently. In addition to the above, the
designer can monitor the process for a few iterations
and if errors occur or are suspected in the formulation
or implementation, the p r o g r a m can be terminated at
an early stage, saving time and money.
For large and complex design problems, the interactive process can be slow because each iteration requires
time-consuming analysis and design sensitivity analysis
REFERENCES
1 Arora, J. S. and Tseng, C. H. Interactive design optimization,
Engineering Optimization, 1988, 13, 173-188
2 Arora, J. S. Introduction to Optimum Design, McGraw-Hill,
New York, 1989
3 Park, G. J. and Arora, J. S. Role of database management in
design optimization systems, J. of Aircraft, 1987, 24, (11)
745-750
4 Tseng, C. H. and Arora, J. S. Optimum design of systems for
dynamics and controls using sequential quadratic programming,
1989 AIAA J., 27 (12), 1793-1800
5 Tseng,C. H. and Arora, J. S. Optimal Design for Dynamics and
Control by using a Sequential Quadrature Programming Algorithm, Technical Report No. ODL-87.10, Optimal Design Lab-
7
8
9
173