154

AIAA JOURNAL

VOL. 3, NO. 1

Note that En(q, r) [Fn(q, r)] is not the Laplace transform of In these formulas (41-43), ft is used to refer to a particular eigenvalue q = ftn, j, satisfying (39); the temperature (41) involves a sum over all these eigenvalues (j = 1, 2 . . . , In an earlier paper, Warren1 gives a solution n = 0, 1, . . .). to this problem. The present method has computational advantages over Warren's solution, particularly in the neighborhood of the inner boundary r = b and for small times. Indeed Warren's approach converges nonuniformly near the inner boundary r = b and takes the value zero at that boundary rather than the nonhomogeneous value of Eq. (32). References Warren, W. E., "A transient axisymmetric thermoelastic problem for the hollow sphere," AIAA J. 1, 2569 (November 1963). 2 Carslaw, H. S. and Jaeger, J. C., Conduction of Heat in Solids (Oxford University Press, 1959), p. 309. 3 Watson, G. N., Bessel Functions (Cambridge University Press, 1958), p. 80.
1

The technique mentioned in Ref. 2 is applicable to the class of problems for which small variations in the constraint functions and in the performance index are related to small changes in the control variables (either time dependent or invariant) in the following manner:

, du dt

A12 =

J to

(1)
(2)

du dt

In order to apply the technique of Ref. 2, the multiple-arc problem must be expressed in the form of Eqs. (1) and (2). For general multiple-arc problems, the governing set of differential equations for the state variables can be written as follows:
/(a)
=

(a) -

= Q

Downloaded by Beihang University (CNPIEC - XI'AN BRANCH) on May 12, 2013 | http://arc.aiaa.org | DOI: 10.2514/3.2809

A Steepest-Ascent Solution of Multiple-Arc Optimization Problems

jf() g h m r T t ta u v w x /3 (At) Az B X <p Q

= = = = = = = = = = = = = = = = = = =

n-dimensional vector of known functions of x, u, and t gravitational acceleration altitude mass range thrust magnitude independent variable, time corner times r-dimensional vector of steering functions horizontal velocity vertical velocity n-dimensional vector of state variable histories mass flow rate A-dimensional vector of corrections to the corner times s-dimensional vector of time invariant control variable corrections thrust angle n-dimensional vector of Lagrange multipliers performance index g-dimensional vector of constraint functions, q < n

for ta-i <t<ta,a = l,...A (3) A complete solution to the differential equations consists of a number of segments. Each segment is referred to as a subarc, and the juncture of two subarcs is referred to as a corner. The length and position of a subarc are defined by the corner times; in particular, the ath subarc begins at ta-i and ends at ta. It is assumed that either a priori knowledge or some logical procedure is available to determine the number of subarcs in the solution. The superscript (a) notation in Eq. (3) is used because the expressions governing the derivatives of the state variables x are not necessarily continuous from one subarc to the next. Discontinuities can result from a change in the algebraic form of the derivatives, from a discontinuity in u, or from a discontinuity in some parameter in the system. For the present, it is assumed that to and all initial conditions for the differential equations are given. The following. constraints must be satisfied at the terminal point IA' 12 = Q[x(tA), tA] = 0 (4) Subject to the differential constraints (3) and the terminal constraints (4) it is desired to optimize som performance index: performance index = (p[x(tA)} tA] Because of the existence of corners, the standard relationship between the adjoint system and the linearized form of Eq. (3) is not directly applicable. An analogous result can be derived, however, which does consider the existence of corners and the associated discontinuities in the derivatives. By predicting the effects of small changes in the corner times on the terminal conditions, this analogous result permits a logical correction procedure that will ultimately yield an optimal multiple-arc solution. Before proceeding with the derivation, one comment on notation must be made. The derivative expressions / (a) are defined for t < ta but are not rigorously defined for t = ta. It is assumed, however, that lim/<fl)(a;, u, t) for ta-i < t < ta
t-*ta

HE problem of trajectory optimization has received a great deal of attention in recent years. The problem, as usually stated, involves the determination of one or more steering functions, such as angle of attack or throttle setting, such that some performance index is optimized, subject to specified constraints. One of the most successful approaches to this problem has been the so-called steepest-ascent or gradient technique, developed by Bryson and Denham1' 2 and Kelley.3 The problems that have been treated with this approach, however, are actually special cases in that they involve only one subarc. This note extends the steepestascent technique for the simultaneous optimization of time dependent functions and time invariant quantities, originally mentioned by Denham,2 to multiple-arc problems with the associated unknown corner times. Received April 14, 1964; revision received August 11, 1964. * Member of the Technical Staff, Performance Analysis Department. Associate Member AIAA.

is well-behaved. Whenever it is indicated that /(a) is evaluated at ta, it should be understood to mean the limit as t approaches ta. The derivation of the desired relationship begins with the formation of the following integral: A rt

= aE I ~a =l
J ta

XTJ(a) dt

(5>

Small variations in x, u, and ta are now introduced, and only first-order terms are retained:

=Tr
A-l a=l

a=l Jta~l

= 0

(6)

JANUARY 1965

TECHNICAL NOTES

155

By expanding J(a) in a Taylor series, integrating by parts, .and using the relationship between variations at fixed time points and total variations, the following result is obtained:

160
INITIAL GUESS TENTH ITERATION FINAL SOLUTION

AF =
,0=1

- E
0=1

140

-E l, o = l '"

rt
a

5w dt

120
Q>

% no | 80

In order to -eliminate the last term, the multiplier functions A are specified as solutions to the adjoint equations:
A

ta-1 < t<ta

a = 1, . . . A

40

100

200

300

4OO

5OO

6OO

700

8OO

90O

Downloaded by Beihang University (CNPIEC - XI'AN BRANCH) on May 12, 2013 | http://arc.aiaa.org | DOI: 10.2514/3.2809

(8) The influence functions are defined as:

TIME (SEC)

Fig. 1 Thrust angle vs time.

a = 1, . . . A

(9) Substituting 0) and (9) into (7), and assuming that x is continuous f across corners provides the desired relationship:
,4-1

moon is considered. The terminal constraints specify a final weight and a soft landing at a given altitude. The equations of motion are as follows: xl = v = (T/m) cos0 = /!
x2 = w = (T/m) sin0 - g = /2
^3 = f = V = /3

(17)
(18)
(19)

,-A*.-

\u 5u dt (10)
Equation (10) is an extension to the case with multiple subarcs of the standard relationship between linear differential equations and the associated adjoint set. By the appropriate choice of boundary conditions for X at i = tA, the form of Eq. (10) can be made identical to that of Eqs. (1) and (2). For those cases without corners, the boundary conditions on X are usually chosen such that the AiA term of Eq. (10) is cancelled. Since corrections to the corner times are the essence of this report, however, tA is used as the stopping condition on each trajectory; its value is corrected after each iteration using the generalized steepestascent technique. The following sets of boundary conditions on X are defined for the current problem:

x4 = h = w = /4

(20) (21)

x, = m = -/3 =/ 5

During the thrusting subarcs, T and ft are nonzero constants; during the coasting subarc, they are zero. The boundary conditions that must be satisfied at the terminal time are as follows:
w*
= 0

(22) t

The performance index to be maximized is

(ID
(12)

Combining Eqs. (10) and (12) and assuming fixed initial conditions, one obtains equations identical to (1) and (2) where the following definitions apply:

= [aoi 002 -

_aqi
- /<+)}

a = 1, . . . A-l

(13)
(14)

The pertinent boundary conditions and constants are as follows: v(t0) = 0; to(o) = 0; r(t0) = 0; h(tQ) = 0; m(tQ) = 10,000 Ibm; T - 2,000 Ib; ft = 8 Ibm/sec; g = 5.3 ft/sec2. The optimum steering program, as determined by the steepest-ascent procedure, is presented in Fig. 1. The procedure required approximately 20 iterations on an IBM 7090. In conclusion, the author feels that the technique described herein opens up some interesting avenues of investigation. Most of these avenues are as yet unexplored, but a procedure that is at least theoretically workable does exist. The procedure has been shown to be workable in practice for some cases, and it is thought that similar results can be obtained for many other cases.
References

k = 1,
0=1,

3 .A-l

(15)
(16)

k = l,...q

Using the generalized steepest-ascent technique, one can now proceed in an orderly fashion to an approximate optimum multiple-arc solution. As an application of the technique, the problem of maximizing the range of a rocket operating over an airless, flat
f The general case involving discontinuous state variables is developed in Ref. 4.

Bryson, A. E., Jr. and Denham, W. F., "A steepest-ascent method for solving optimum programming problems," J. Appl. Mech. 29,247-257 (1962). 2 Denham, W. F., "Steepest-ascent solution of optimal programming problems," Raytheon Co. Kept. BR-2393 (April 1963). 3 Kelley, H. J., "Gradient theory of optimal flight paths," ARS J. 30,947-953 (1960). 4 Graham, R. G., "The effects of state variable discontinuities in the solution of variational problems," Aerospace Corp. Rept. TDR-269 (4550-20 )-4 (July 1964).
v*

= 0: w*

0: h* = 0: m* = 5000 blm.