This article was downloaded by: [University of Ottawa]

On: 13 March 2013, At: 11:08

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Control

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tcon20

Optimal control via collocation and non-linear

programming
a a a
T. H. TSANG , D. M. HIMMELBLAU & T. F. EDGAR
a
Department of Chemical Engineering, The University of Texas at Austin, Austin, Texas,
78712, |
Version of record first published: 12 Mar 2007.

To cite this article: T. H. TSANG , D. M. HIMMELBLAU & T. F. EDGAR (1975): Optimal control via collocation and non-linear
programming, International Journal of Control, 21:5, 763-768

To link to this article: http://dx.doi.org/10.1080/00207177508922030

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to
anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should
be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims,
proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in
connection with or arising out of the use of this material.
 INT.J. CONTROL, 1975, VOL. 21, NO.5, 763-768

Optimal control via collocation and non-linear programming

T. H. TSANGt, D. 1\1. HIMl\j]~LBLAUt and T. F. EDGARt

The collocation method meshed with non-linear programming techniques provides an

efficient strategy for the numerical solut.ion of optimal control problems. Good
accuracy can be obtained for the state and the control trajectories as well as for
the value of t.he objective function. In addition. the control strategy can he quite
flexible in form. However. it is necessary to select the appropriate number of collo-
cation points and number of parameters in t.he approximating functions with care.

1. Introduction
Downloaded by [University of Ottawa] at 11:08 13 March 2013

Recently Neuman and Sen (1974) demonstrated how the method of weighted
residuals (l\HVR) (Finlayson 1972) could be used to transform a dynamic
optimization problem into a static non-linear programming (NLP) problem.
The latter represented an approximation to the optimal control problem, but
the value of the cost function resulting from the suboptimal control was very
close to the optimal value. They demonstrated how both the Galerkin
procedure and the collocation method could be employed, but we devote our
attention to the latter method because it. does not require the evaluation of
integrals of the control modes. The main difference between the procedure
proposed here and that of Neuman and Sen is that we make use of constrained
NLP whereas they employed unconstrained NLP (effected by solving for certain
of the variables in the constraints and substituting for them in the objective
function). The use of a constrained NLP structure enables the direct applica-
tion of this method to complicated state equations and objective functions as
well as the incorporation of state and control inequality and equality con-
straints. We also explore the effect of different assumed control modes.

2. The collocation method

We consider the same problem as did Neuman and Sen and use the same
notation. The problem is
I
Minimize: J(u)=! J [x 2(t ) + U2(t )] dt (1)
o
Subject to: i(t) = -X(t)+1t(t) (2)
x(O) = ],0 (3)

The optimal control problem (1) to (3) is a linear-quadratic control problem

whose analytical non-iterative solution via the Riccati equation is well known.
However, generally the solution to an optimal control problem and associated
two-point boundary-value problem must be carried out by iterative methods
and requires the repeated integration of two sets of differential equations.

Received 14 October 1974.

t Department of Chemical Engineering, The University of Texas at Austin,
Austin, Texas 78712.
 764 T. H. Tsang et al,

In the collocation method both the state variable, x(t), and the control
variable, u(t), are approximated by functions of t. Here, as a simple example,
second-order polynomials are assumed.
x(t) = 1 + alt + a 2t2= 0,0 + alt + a 2t2 = aT <I>(t) (4)

u(t) = bo + bIt + b2t 2 = bT.y(t) (5)

Note that the approximate solution x(t) satisfies the initial condition. Intro-
duction of the approximating functions into (I) and subsequent integration
gives
J(a; b) = [oW + b0 2)t + !(al + bobl)t 2+ -li(2a 2 + a l2bob2+ b12)t3
+ !(a la2+ blb2)t4 + io-(a22+ b1)t 5 J:::: A (6)
Downloaded by [University of Ottawa] at 11:08 13 March 2013

The system eqn. (2) becomes, with x = + 2a2t,

0,1

(/,1 + 2a2t+ 1 + alt + a 2t2- bo - bIt - b2t2= 0 (7)

and the residual for each collocation point (t;) is

Rj=(1 +al-bo)+(al+202-bl)tj+(a2-b2)t/, i> 1,2, ... , N (8)
A different choice of functional form for x(t) or u(t) results in a different approxi-
mation, of course. The number of collocation points determines the number of
constraints for this exam pIe.
Thus, the solution to the approximation of the optimal control problem (1)
to (3) becomes a problem in minimizing (6) with respect to the vectors a arid
b subject to (8), a set of equality constraints. The equality constraints are
linear if the system is linear (as here) and non-linear if the system equation is
non-linear.

3. Numerical results
Four sets of collocation points were chosen for the example problem:
Case I 'I = J., t2=-~

Case 2 11 =!, t2='~

Case 3 : t. = J, t 2 = -~, t3= !
Case 4 : t 1 =!, t 2 =~, t 3 =1, i, = ltr
Two non-lineal' programming codes GREG (Guigan 1971) and COMET (Staha
I !)7:J) thu.t can accommodate both non-linear equality and inequality constraints
were used to solve the NLP problem specified by thc minimization of eqn. (6)
subject to eqns. (8)_
The table shows that there was essentially no difference between the results
obta.ined for the two codes. All of the constraints were satisfied to at least
within 10-'. Two sets of starting vectors were used: A (a; = b, = 1'0) and
13(0; = b, = 5'0); only slight differences in the answers resulted from the
different starting points. The value of the objective function, J, calculated
for cach case was well within 1·0% of the value of the objective function for the
optimal control (denoted by J*).
 Collocation and non-linear programming 765

Coefficients

Deviation
Calculated of
objective objective
Case a, &, function function
J (%)
100 (J -J*)
J

lA - 1·22027 0·46345 -0'13847 -0,78422 1·1l791 0·193202 0·151

IB - 1·22023 0'46345 -0·13843 -0,78410 l·ll 783 0·193202 0·151
2A - 1-48971 0·88098 -0·02678 -1,37040 1·64267 0·192293 0·320
3A - 1·25134 0·53476 -0,25134 -0,18182 0·53477 0·194296 0'718
Downloaded by [University of Ottawa] at 11:08 13 March 2013

3B - 1·25142 0·53483 -0·25142 -0·18176 0·53484 0·194296 0·718

4A -1'25ll6 0·53466 -0·25116 -0,18185 0'53465 0·194296 M18

J* =0· 19291.
Table I. Numerical results via NLP.

CASE
IA AND IB

- - - 2A

3A,38 AND 4A

- - - NEUMAN AND SEN (STEP)

ANALYTICAL SOLUTION

;::0r----1---+---I----+---+---t-----1f----+---,4-4-=__-l
::::>

0.1 02 0.3 0.4 0.5 0.6 0.7 O.B 0.9 1.0

Figure 1. Comparison of control trajectories (cases without constraints).

However, Fig. 1, which illustrates the control trajectories for the several
cases (including the optimal solution and the results of Neuman and Sen),
indicates that the controls calculated by the collocation technique differed
radically from the optimal control. Figure 2, moreover, shows that the state
variable trajectories resulting from the various approximations were not very
 766 T. H. Tsang et al.
CASE
IA AND IB
- - - - 2A
3A,3B AND 4A
--- NEUMAN AND SEN
ANAL YTICAL SOLUTION

0.6

- 0. .5
X
0.4

0.3

0.2t
Downloaded by [University of Ottawa] at 11:08 13 March 2013

0.1

I
o
o _-"--=,=,~'---:-:-----'-----,:':----'-----=':---'-~
0.2 0.4 0.6 0.8 1.0

Figure 2. Comparison of state trajectors (cases without constraints).

accurnte, indicating that the value of J is not too sensitive to whatever x(t) and
'u(t) may be. In comparing Case 3 (three collocation points) with Cases I and 2
(two collocation points each), it is seen that fewer collocation points yielded a
slightly smaller J. A decrease in J can be demonstrated analytically for this
one simple example, but the detailed derivations am too tedious and extensive
to be detailed here.
J -J* is also very sensitive to the relative weights applied to x(t) versus
1/.(t). Consider the objective funetion

1
J =I J [x 2(t )+ 1'l/2(t )] dt (9)
u

where r is an arbitrary scalar weighting factor. 1n the table the results are
based on r = I. When r was increased, the minimum of J deviated substantially
from J* for the second-order polynomials. For example,

% deviation
r from J*
5 30·3
10 42·9

1<'01' the above results, three collocation points were used, which left only two
degrees of freedom for optimization. 1"01' the two-segment piecewise constant
(i.c. step) function used by Neuman and Sen, we obtained much smaller devia-
tions of J from J*, as r increased but still noted unsatisfactory approximations
of the state variable x(t). Hence J -J*jJ was not very sensitive to increases
in r for the step approximation function.
The use of higher-order polynomials for both the state and control variables
leads to distinctly improved, much less sensitive approximations for x(t) and
 Collocation and non-linear programming 767

'u(t).If :I:(t) and u(t) are both approximated by fourth-order polynomials with
seven collocation points, we found

% deviation
r from J*
.5·0 x 10-5
.5 ,5·0 x 10- 5

The results for x(t) and ll,(t) are not illustrated because it is not possible to
distinguish in Figs. I or 2 between the analytical and the approximate trajec-
tories of x(t) and u(t). The computing time for each run on a CDC 6600
computer was 0'2 sec, in contrast to the second-order polynomials for which
the computing time was 0·03 sec.
Downloaded by [University of Ottawa] at 11:08 13 March 2013

CASE
IA

--- 2A

3A

4A
NEUMAN AND SEN. ANO
ANALYTICAL SOLUTION

.::ol---+--+---t---+---+--+----1'---;~7'''+-__::=_i
:::J

0.5
t
Figure 3. Comparison of control trajectories (cases with constraints).

4. Introduction of inequality constraints

It is quite feasible to add inequality constraints either on x(t) or u(t) if the
general NLP codes cited above are used. Suppose for the example problem
the control must be constrained so that

u(t) ~0 }
(10)
u(t);;, 0

The NLP problem now includes 2n additional inequality constraints, where n

is the number of collocation points. Figure 3 presents the results obtained
 768 Collocation and non-linear programming

minimizing (6) subject to (8) plus (~O) for the second order approximating
polynomials. The deviations from J* were
Case % deviation
lA 0·41
2A 0·33
3A 0·718
4A 1·04

S. Conclusions
The collocation method is quite useful in solving optimal control problems
in which both the approximating state and control vectors must have small
deviations from the exact solutions. NLP permits the introduction into the
Downloaded by [University of Ottawa] at 11:08 13 March 2013

control problem of extraneous equality and inequality constraints on either u(t)

or x(t). In principle the differential equation and constraints can be non-
lincar. For complex problems, the starting point, number of collocation
points, and form and number of parameters in the approximating function need
to bc carefully selected. 'For linear differential equations and constraints,
NLP together with collocation seems to be an effective tool for the approximate
solution of optimal control problems.

REFERENCES
FINLAYSON, B. A .• 1972, The Method o/Weighted Residuals and Variational Principles
(New York: Academic Press).
GUIGAN, G., 1971, 'Presentation et utilization du code Greg', Note HI 582/2,
Electricitc de France.
NEUMAN, C. P., and SEN, A., 1974, I.E.E.E. Trans. autom. Control, 19, 67.
STAHA, R. L., 1973, Ph.D. Dissertation, University of Texas at Austin.