338 J. GUIDANCE VOL. 10, NO.

Direct Trajectory Optimization Using Nonlinear

Programming and Collocation
C.R. Margraves* and S.W. Paris*
Boeing Aerospace Company, Seattle, Washington

An algorithm for the direct numerical solution of an optimal control problem is given. The method employs
cubic polynomials to represent state variables, linearly interpolates control variables, and uses collocation to
satisfy the differential equations. This representation transforms the optimal control problem to a mathematical
programming problem which is solved by sequential quadratic programming. The method is easy to program for
a very general trajectory optimization problem and is shown to be very efficient for several sample problems.
Results are compared with solutions obtained with other methods.
Downloaded by TUFTS UNIVERSITY on June 30, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.20223

Introduction number of details are different (e.g., different order splines

DIRECT trajectory optimization method that represents are used, the collocation points are different, the independent
te and control variables by piecewise polynomials is variables were spline coefficients, etc.).
described in this paper. An implicit integration scheme based
on Hermite interpolation is used to convert the optimal con- Mathematical Method
trol problem to a nonlinear programming problem. The
Statement of the Problem
primary advantage of this method is that it is much easier to
extend to general problems involving path constraints, discon- Trajectory optimization problems that can be described by a
tinuous states, and control inequalities. sequence of vehicle/flight stages are considered. The control
Numerous approaches are possible for direct trajectory op- functions u(t), the time points delimiting the various stages
timization. The authors have described in earlier papers a (Eit called events), and the vehicle design parameters (co) are
method employing piecewise Chebyshev polynomials. Penalty sought which minimize a given performance index
functions were used to convert optimal control problems to
unconstrained parameter optimization problems which were J=<t>[x(E)tu(E),u,E] (1)
solved with a full second-order modified Newton
algorithm.1'3 The basic ideas used in this approach were The /th stage is assumed to be a dynamical system subject to
described earlier by Johnson4 and Hahn and Johnson.5 A the differential constraint
similar approach was described by Balakrishnan.6 Additional
direct approaches were described by Kelley,7 Rader and Hull,8 ' =/ (x,u,u,t) (2)
and Brauer et al.9
Birkbff and de Boor10 hae shown that piecewise on the interval te[EifEi+l]t where * is a vector of states
polynomials have many desirable properties for representing governed by first-order differential equations, u a vector of
smooth curves, de Boor11 and Russell and Shampine12 controls (e.g., pitch angle, and o> a vector of design param-
described implicit methods for the solution of boundary value eters (e.g., planform area, rocket nozzle diameter, etc.).
problems in ordinary differential equations using piecewise The prime denotes differentiation with respect to time. At
polynomials. Dickmanns and Wells13 and Dickmanns14'15 ap- each event Eit nonlinear boundary conditions may be imposed
plied the maximum principal to reduce the optimal control of the form
problem to a boundary value problem which they solved using
a collocation method; the method employed Hermite inter- (3)
polation and collocation.16
The solution of optimal control problems using nonlinear where £BE is the vector lower-limit on the boundary conditions
programming has been demonstrated for a number of distinc- and (7BE is the upper-limit. Equality constraints are possible
tive methods. Betts et al.17 described a procedure in which a by setting the appropriate elements of £BE and £/BE equal to
problem is divided into a series of subproblems which are each other. Path constraints of the form
solved by an indirect method. Boundary conditions on the
subproblems are solved by nonlinear programming. Kraft18 (4)
parameterized the control variables and integrated explicitly
the equations. A direct and an indirect method were utilized may also be imposed on the system. Denote the collection of
and compared. Evtushenko19 used nonlinear programming to events by ET= (El9E2,...9EN+l), where N is the number of
define a variety of optimal control algorithms. Renes20 pro- stages.
posed a procedure which is similar to that described in the next The functions/, a', and hk are assumed to be smooth (class
section. Renes did not give any numerical results and a C2 or better) within a given stage. However, no continuity
restrictions are imposed at events. Specifically, discontinuities
of the form

Received June 9, 1986, revision received Dec. 10, 1986. Copyright (5)
© American Institute of Aeronautics and Astronautics, Inc., 1987.
All rights reserved. are allowed at the events. The 6,'s may be fixed quantities or
* Senior Specialist Engineers. Members AIAA. may be included in the design parameter set o>. Equations (5)
 JULY-AUGUST 1987 DIRECT TRAJECTORY OPTIMIZATION 339

account for physical events such as weight jettisons and im- The defects for each state evaluated at the center of each
pulsive velocity changes. segment constitute a set of nonlinear algebraic equations
which are a function of the states and controls at each node,
Method of Solution the events E9 and the design parameters co. Linear interpola-
The approach is to reduce the optimal control problem tion is used to obtain the controls at the segment centers. In-
stated in the previous section to a nonlinear programming itial experiments with cubic polynomial controls did not work
problem. To do this, the functions x(t) and u ( t ) are well. Also, quintic polynomials for the solution of second-
represented by piecewise polynomials and collocation is used order differential equations (without reduction to the first
to satisfy Eq. (2). order) were tried. This approach did not work as well as the
The length of each state is defined as Tsi=Ei+l -Ei9 as cubics. The boundary conditions [Eq. (3)] and the constraints
shown in Fig. 1. For each stage, the interval [EifEi+ {] is sub- [Eq. (4)] evaluated at both the nodes and centers of the
divided into Ns segments/Let the ratio of the length of theyth segments provide additional equations (constraints) to be
segment to T.^ be denoted as Tyi. Thus, the length of the yth satisfied. Equations (5), connecting the segments constitute
segment in the fth stage is r^T^. Using Hermite interpolation, additional linear equations to be satisfied. Now, collect all of
cubic polynomials are defined for each state on each segment the independent variables into a single vector P defined by
using values of the states at the nodes (the boundaries at the
segments) and the state time derivatives, as defined by the PT=[ZT,ET,uT] (12)
equations, at the nodes. The values of the states are then
selected (by nonlinear programming) to force the interpolated where
derivatives to agree with the differential equations at the
center of the segment. This procedure is illustrated in Fig. 2.
Downloaded by TUFTS UNIVERSITY on June 30, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.20223

The basic procedure can be derived as follows. Let the states

x be represented on each segment by cubics of the form Collecting all of the nonlinear equations into a single vector
2 3
equation gives
x = C0 + Q S + C2S + C3S (6)
(13)
where to simplify the argument, the segment length S is
transformed such that Se[0, 1]. Let x(Q) = xl9 x(l) = x2,
dx/ds(Q) = xl9 dx/ds(l) = x2. Differentiating Eq. (6) and where
evaluating at 0 and 1 gives
" 1 0 0 0 " " Co " A/:/ = defect for rth state at yth node
BN = collection of all nonlinear relationships from Eq. (3)
0.1 00 Q HN = collection of all nonlinear relationships from Eq. (4)
1111 C2 (7) It is now noted that our payoff function / is a function of P
only. The trajectory optimization problem stated above can be
0 1 2 3 _C3. expressed as
Inverting the 4 x 4 matrix gives

"Q- " 1 0 0 0 " subject to

Q 0 1 0 0
(8) (14)
C2 -3 -2 3 -1
C3 2
1 -2 1
where NP is the dimension of P, AP is composed of all the
Now using Eq. (8), evaluating Eq. (6) at S= 1/2, and trans- linear relationships from Eqs. (3-5), £and u are the upper and
forming to segment length T=Tjit Tsi, we see that the inter- lower bounds. For equality constraints £/ = «/, and inequality
polated value of x at the center of the segment is constraints are handled by £/= — oo or «, = oo. Equation (14)
constitutes a nonlinear programming problem.
(9) A large amount of literature in recent years has been
devoted to the solution of nonlinear programming problems.
where// is Eq. (2) evaluated at xt (the superscript denoting the Recent treatments of the subject are presented by Gill et al.,21
stage was dropped for simplicity). (Note x^ = Tf[xt, u(tt), tif Evtushenko,22 and Reklaitis.23 These books contain long lists
a?].) In the same way, the slope of references on earlier work in this field.
-x2)/2T- (10) Programming Considerations
The computer code implementing the aforementioned pro-
is obtained. Evaluating Eq. (2) at xc gives fc. Define the defect cedure is called Nonlinear Programming for Direct Optimiza-
at the center of the segment as tion of Trajectories (NPDOT). NPDOT uses a nonlinear pro-
gramming package called NPSOL which was developed by the
A =/,-*; Systems Optimization Laboratory at Stanford University.24
NPSOL uses sequential quadratic programming (SQP). SQP
(11) methods were popularized by Biggs,25 Han,26 and Powell.27
NPSOL requires as inputs the matrix A and an initial guess
x{ and x2 are varied to drive A = 0. for P. Subroutines must be provided which evaluate <t>(P)t
If the cubic polynomial is capable of representing the solu- C(P), and their first derivatives with respect to P. In the
tion on the given segment, then selecting x1 and x2 to drive A method described here, these derivatives are computed by
to zero will produce an accurate approximation to the solution finite difference. This task is greatly simplified by the fact that
ofEq. (2). the defects depend on states at adjacent nodes only and only
 340 C.R. MARGRAVES AND S.W. PARIS J. GUIDANCE

on the time length of the stage in which it occurs. The Jaco- The following problems have been solved using NPDOT: 1)
bian has the banded structure shown in Fig. 3. Van der Pohl (Ref. 31 and 32), 2) Brachistochrone (Ref. 33),
To obtain accurate solutions to realistic problems, some 3) minimum time interceptor (Ref. 34), 4) advanced booster.
refinements must be added to the procedure. The key The results of NPDOT agreed with the previously published
refinements involve problem scaling, partial derivative com- results for problems 1-3. A comparison of the solution times
putation, data smoothing, node selection, and node refine- between NPDOT and CTOP is given in Fig. 4. The times
ment. A discussion of equation scaling as applied to optimiza- shown in Fig. 4 are adjusted to be comparable for the Cray X-
tion problems may be found in Ref. 28. As a minimum, the MP/24. (For approximate comparison with other computers,
independent variables should be scaled so that they have the X-MP has a clock cycle time of 9 ns.) Discussions of prob-
similar mangitudes. The following scaling procedure worked
well. (No systematic comparison with other approaches was
attempted.) States and controls were scaled by

where XN is the nominal trajectory and Sx is an estimate of an

upper bound for X-XN. Times were scaled by
TS = T/ST
Downloaded by TUFTS UNIVERSITY on June 30, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.20223

where ST is the estimated upper bound for T.

For realistic trajectory problems, the computation of Nodes 1 2 3 4 5 10 11 12
analytical derivatives can be a formidable task. It is usually Events EI E2

preferable to compute derivatives directly by finite difference.

Partial derivatives must be accurate to obtain convergence of Fig. 1 Piecewise polynomial representation.
the nonlinear programming algorithm. An algorithm for selec-
ting the finite-difference perturbation size is given in Ref. 28
(pp. 341-345). The code described here uses this algorithm in-
itially and retains the perturbation step sizes for subsequent
iterations.
If tabular data are used, it is important that they be smooth.
Smoothing data without changing the physical meaning can be
a difficult problem. Good results were obtained with splines
[see Ref. 29)] for tables with one independent variable. A
satisfactory solution for multivariable tables has not yet been
found.
The smoothness of data is closely related to node spacing
since, in general, rapid changes in tabular functions will be ac- Fig. 2 Implicit integration illustration.
ceptable only if there are sufficient nodes in the region of
rapid change. This can be difficult to accomplish since the
nodes are distributed in thie whereas tabular data may be
functions of other variables such as Mach number of altitude.
Regions where large changes occur as a function of Mach
number for the nominal and optimal solution may be very dif-
ferent regions in time. An initial node selection can be made
based on the nominal. Unless a nominal that is close to
opitimal is known, an adaptive procedure is requried.
The node selection can then be updated after initial con-
vergence or when optimizer behavior indicates possible prob-
lems due to the polynomial distribution. Dickmanns14 has
described several methods for refining node selection. The
basic idea is to represent each segment by a higher-order
method (polynomial, numerical integration, etc.) and add B.C.
nodes when disagreement is found. After several refinement
cycles, the adequacy of the total representation can be treated
by explicit numerical integration of the trajectory using the Fig. 3 Jacobian structure.
polynomial control function and comparing the results with
the polynomial representation of the trajectory.

Numerical Examples Constrained Brachistochrone

Several test problems that have been solved with the new
technique are described in this section. The problems range 1.8
from simple analytical problems to contemporary aerospace
performance-prediction problems. A stepped approach was
used to validate NPDOT against known results, to gain ex-
perience before attempting new unknown solutions, and to CTOP NPDOT CTOP NPDOT
demonstrate the method's inherent flexibility. The Chebyshev
Trajectory Optimization Program (CTOP), described in Ref. • All times are Cray X-MP/24 CPU seconds
• CTOP - Chebyshev Trajectory Optimization Program (Boeing,
• CTOP. NPDOT min time to climb =
. "J^ ^ ^ ^ = m ^
1, served as the performance reference and "truth" model for • POST - Program to Optimize Simulated Trajectories (Martin) * CTOp multiple runs
• NPDOT single runs
NPDOT. The academic problems were also compared to solu-
tions using the maximum principle.30 Fig. 4 CPU time comparison.
 JULY-AUGUST 1987 DIRECT TRAJECTORY OPTIMIZATION 341

„ 190,000
^150,000
1110,000
* 70,000
30,000,
2,500 5,000 7,500
8 segment
Chebyshev
distribution

2,500 5,000 7,500

-5,000
20 30 40 50 60 70
Range, X, nmi
Fig. 5 Supersonic interceptor minimum-time climb trajectories.
2,500 5,000 7,500
Great circle range, nmi
1,700
o 1,500 Fig. 7 Advanced booster trajectory.
J 1,300
Downloaded by TUFTS UNIVERSITY on June 30, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.20223

> 1,100 tion and time deriatives of the polynomials were integrated to
~ 900
§ 700
> 500
acquire an error estimate. This was done over the eight-
segment trajectory. Five nodes were added to the center of the
segments which were producing the majority of the error. A
300, solution was generated for this 13-segment distribution and
100 200 300
another error estimate was produced. Two additional nodes
50 were added to yield a 15-unequal-segment distribution. This
40 distribution yielded a time to climb of 317.3 s and produced a
I 30 trajectory, shown as a solid line in Fig. 5, which matched our
5 20 previously obtained CTOP result and published energy state
| 10 results.35 The control function and velocity profile are shown
I" o in Fig. 6. The airplane accelerates to a high subsonic speed,
then climbs at a constant Mach number followed by a dive
-20. through Mach one and beyond, finishing up by a zoom climb
100 200 300
Time, T, sec to the final altitude. In the initial attempts for a solution, sim-
ple piecewise linear and piecewise cubic interpolation for the
Fig. 6 Supersonic interceptor control function and velocity profile. aero/propulsion data were used. This did not work well. The
program could not reduce the defects or obtain a small gra-
dient. This is consistent with previous trajectory optimization
lems 3 and 4 follow to provide some insight into the method as experience. These problems were resolved by using the taut
applied to typical aerospace flight mechanics problems. spline and tensor spline packages as defined and coded by
de Boor.36
Supesonic Interceptor Minimum-Time Climb
This problem involves finding the pitch function 0(0 to Advanced Booster
take a supersonic interceptor from sea level, Mach = 0.38, to This problem involves maximizing the final weight that can
an altitude of 20 km, Mach= 1.0, in minimum time. The in- be injected into low Earth orbit by an advanced booster which
itial weight was 42,000 Ibs. This problem was proposed and utilizes airbreathing propulsion. The propulsion system has
solved by Bryson.34 The equations of motion were those for several discrete phases which depend on the vehicle's Mach
planar flight above a flat Earth. The aerodynamic coefficients number. The first phase spans from Mach 0-2. The thrust and
were functions of Mach number. The thrust was a function of specific impulse were modeled as tabular functions of Mach
Mach number and altitude. The specific fuel consumption was number. The second propulsion phase goes from Mach 2-20.
constant. The values for the defining aerodynamic and pro- The thrust was modeled as a thrust coefficient multiplied by
pulsion data were taken from Ref. 35. The atmospheric dens- dynamic pressure. The thrust coefficient was a function of
ity and speed of sound were taken from the N ACA-1962 at- Mach number and angle of attack. The specific impulse is a
mosphere model. The initial-state values were obtained by function of Mach number. The final propulsion phase, Mach
linear interpolation between the initial and final values; the numbers greater than 20, was a rocket in which both the thrust
control (pitch) was set to 0.18 rad. The initial estimate of the and specific impulse were constant.
free variables was as follows: final range equal to 360,000 ft The equations of motion were for three-dimensional flight
and the final mass set to 1204 slugs. above a spherical nonrotating Earth. Limits were imposed on
A sequence of solutions is shown on Fig. 5. A preliminary dynamic pressure, altitude, and angle of attack. The second
solution, shown by short dashes, was obtained with 16 equal propulsion phase being proportional to dynamic pressure
segments. One can observe that this trajectory exploits the tended to drive the trejectory solutions to very high dynamic
discretization to meet the constraints at the nodes and centers pressure values. The aerodynamics for this vehicle were done
while violating such things as the altitude constraint in be- in body axis form, CA and CN = CNa -a. CA and CNa were
tween. To counter this activity, an eight-segment solution with functions of Mach number.
a Chebyshev node distribution was generated. This provided Initial attempts to generate trajectories using a conventional
the trajectory, shown by long dashes, which had a time to explicit integrating program for this class of vehicle required
climb of 325.2 s. This was not as good as the previous CTOP1 intensive "man-in-the-loop" interactions. The strategy
results, so a solution was obtained with 15 segments. To ob- employed was to accelerate the vehicle to a dynamic pressure
tain the node distribution, the absolute value of the difference <7, limit and fly that limit until a predetermined velocity was
between the time derivatives produced by the equations of mo- reached. At that time, a rocket burn is performed which
 342 C.R. HARGRAVES AND S.W. PARIS J. GUIDANCE

12
allows the vehicle to coast in a transfer orbit with the ap- Russell, R.D. and Shampine, L.F., "A Collocation Method for
propriate apogee. A single apogee burn is then used to cir- Boundary Value Problems," Numerical Mathematics, Vol. 19, Jan.
cularize the orbit. All attempts to optimize this trajectory with 1972, pp. 1-28.
13
an explicit integrating parameter-optimization program failed. Dickmanns, F.D. and Wells, K.H., "Approximate Solution of
Optimal Control Problems Using Third Order Hermite Polynomial
The best results were achieved by using standard trade studies. Functions," Proceedings of the 6th Technical Conference on Op-
Attempts to use CTOP on this problem resulted in very slowly timization Techniques, Springer-Verlag, New York, IFIP-TC7, 1975.
converging trajectories. When NPDOT was applied to this 14
Dickmanns, E.D., "Efficient Convergence and Mesh Refinement
problem, only the initial airbreathing portions and rocket Strategies for Solving General Ordinary Two-Point Boundary Value
burn were optimized. This results in the problem of maximiz- Problems by Collocated Hermite Approximation," 2nd IFAC
ing the weight injected into a predefined transfer orbit. The Workshop on Optimization, Oberpfaff-Enhofen, FRG, Sept. 1980.
transfer orbit is defined by a fixed velocity and flight path 15
Dickmanns, E.D., "Collocated Hermite Approximation of Third
angle at a given altitude. This was done to benchmark and Fifth Order with Automatic Mesh Refinement for the Solution of
NPDOT with the existing solution. A coverged NPDOT tra- Optimal Control Boundary Value Problems," Hochschule der
jectory was obtained using 22 segments and resulted in a 10% Bundeswehr Munchen, FRG, LRT/WE, Oct. 1979.
16
gain in weight over the baseline trajectory. The trajectory, Dahlquist, G., Bjdrck, A., and Anderson, N., Numerical
pitch profile, and dynamic pressure as a function of time are Methods, Prentice-Hall Englewood Cliffs, NJ, 1974, p. 285.
17
shown in Fig. 7. It is interesting to note that the optimum tra- Betts, J., Bauer, T., Hoffman, W., and Zondervan, K., "Solving
jectory does not stay on the dynamic pressure constraint but the Optimal Control Problem Using a Nonlinear Programming
rather initially rides it and then gets off. To maximize ac- Technique, Part 1: General Formulation," AIAA Paper 84-2037,
Downloaded by TUFTS UNIVERSITY on June 30, 2018 | http://arc.aiaa.org | DOI: 10.2514/3.20223

celeration, the vehicle wants to fly at maximum q\ however, Aug. 1984.

18
Kraft, D., "On the Choice of Minimization Algorithms in
the vehicle's final conditions are orbital, and drag losses are Parametric Optimal Control, Springer-Verlag, New York, 1981.
largest at maximum q so there is a balance between the two. 19
Evtushenko, Y.C., Numerical Optimization Techniques, Op-
timization Software, Inc., Pub. Div., New York, 1985, pp. 388-398.
Conclusions 20
Renes, J.J., "On the Use of Splines and Collocation in a Trajec-
A trajectory optimization method which uses an embedded tory Optimization Algorithm Based on Mathematical Programming,"
collocation scheme in conjunction with mathematical pro- National Aerospace Lab., Amsterdam, the Netherlands, NLR-
gramming has been described. The method has been used to TR-78016 U, Sept. 1978.
21
solve a wide variety of test cases and found to be superior to Gill, P.E., Murray, W., and Wright, M.H., Practical Optimiza-
other procedures in terms of cost and robustness. tion, Academic Press, New York, 1981.
22
Evtushenko, Y.G., Numerical Optimization Techniques, Op-
timization Software, Inc., Pub. Div., New York, 1985, pp. 196-364.
Acknowledgment 23
Reklaitis, G.V., Ravindran, A., and Ragsdell, K.M., Engineering
This research was supported in part by the U.S. Air Force Optimization, Wiley, New York, 1983.
24
under Contract F33615-85-C-3009. Gill, P.E., Murray, W., Saunders, M.A., and Wright, M.H.,
"User's Guide for NPSOL (Version 2.1): A FORTRAN Package for
Nonlinear Programming," Systems Optimization Lab., Dept. of
References Operations Research, Stanford Univ., Stanford, CA, Technical Rept.
SOL 84-7, Sept. 1984.
Margraves, C.R., Johnson, F.T., Paris, S.W., and Rettie, I., 25
Biggs, M.C., "Constrained Minimization Using Recursive
"Numerical Computation of Optimal Atmospheric Trajectories," Equalty Quadratic Programming," Numerical Methods for Nonlinear
AIAA Journal of Guidance and Control, Vol. 4, July-Aug. 1981, pp. Optimization, edited by F.A. Lootsma, Academic Press, London,
406-414. 1972, pp. 411-428.
2
Paris, S.W., Joosten, B.K., and Fink, L.E., "Application of an 26
Han, S.P., "Superlinearly Convergent Variable Metric
Advanced Trajectory Optimization Method to Ramjet Propelled Algorithms for General Nonlinear Programming Problems,"
Missiles," Optimal Control Applications and Methods, Vol. 1, Mathematical Programming, Vol. 11, 1976, pp. 263-282.
Oct.-Dec.
3
1980, pp. 319-334. 27
Powell, M.J.D., "A Fast Algorithm for Nonlinearly Constrained
Hargraves, C.R., "Numerical Computation of Optimal At- Optimization Calculations," Univ. of Cambridge, England, Rept.
mospheric Trajectories Involving Staged Vehicles," AIAA Paper DAMTP 77/NA 2, 1977.
82-0360, AIAA 20th Aerospace Sciences Meeting, Orlando, FL, Jan. 28
Gill, P.E., Murray, W., and Wright, M.H., Practical Optimiza-
1982. tion, Academic Press, New York, 1981, pp. 341-345; pp. 346-356.
4
Johnson, F.T., * 'Approximate Finite-Thrust Trajectory Optimiza- 29
de Boor, C.R., A Practical Guide to Splines, Springer-Verlag,
tion," AIAA Journal, Vol. 7, June 1969, pp. 993-997. New York, 1978, pp. 299-314.
5
Hahn, D.W., and Johnson, F.T., "Final Report for Chebyshev 30
Pontryagin, L.S., Boltyanskii, V.G., Gamkrelidze, R.V., and
Trajectory Optimization Program (CHEBYTOP II)," The Boeing Mishchenko, E.F., The Mathematical Theory of Optimal Processes,
Co., Seattle, WS, Rept. D180-12916-1, prepared for NASA under Interscience, New York, 1962.
Contract NAS2-5994, June 1971. 31
Bullock, T.E., and Franklin, G.F., "A Second-Order Feedback
6
Balakrishnan, A.V., "On a New Computing Method in Optimal Method for Optimal Control Computations," IEEE Transactions on
Control," SIAM Journal on Control, Vol. 6, May 1968, pp. 149-173. Automatic Control, Vol. AC-12, No. 6, Dec. 1967.
7
Kelley, H.J., "Method of Gradients," Optimization Techniques, 32
Willoughby, J.K., "Adaptations of the Conjugate Gradient
edited by G. Leitmann, Academic Press, New York, 1962, pp. Method to Optimal Problems with Terminal State Constraints,"
205-254. Engineering Research Institute, Iowa State Univ. Technical Rept.
8
Rader J.E., and Hull, D.G., "Computation of Optimal Aircraft ERI-62500, Feb. 1970.
Trajectories Using Parameter Optimization Methods," Journal of 33
Bryson, A.E. and Ho, Y., Applied Optimal Control, Blaisdell
Aircraft, Vol. 12, Nov. 1975, pp. 864-866. Publishing, Waltham, MA, 1969, p. 81; 119-120.
9
Brauer, G.L., Cornick, D.E., and Stevenson, R., "Capabilities 34
Bryson, A.E., and Denham, W.F., "A Steepest-Ascent Method
and Applications of the Program to Optimize Simulated Trajectories for Solving Optimum Programming Problems," Journal of Applied
(POST)," NASA CR-2770, Feb. 1977. Mechanics, Vol. 29, June 1962.
10 35
Birkhoff, G. and de Boor, C.R., "Piecewise Polynomial Inter- Bryson, A.E., Desai, M.N., and Hoffman, W.C., "Energy-State
polation and Approximation," Approximation of Functions, edited Approximation in Performance Optimization of Supersonic Air-
by H.L. Garabedian, Elsevier, New York, 1965. craft," Journal of Aircraft, Vol. 6, Nov.-Dec. 1969, pp. 481-488.
n 36
de Boor, C.R., A Practical Guide to Splines, Springer-Verlag, de Boor, C.R., A Practical Guide to Splines, Springer-Verlag,
New York, 1978. New York, 1981.