Title: Christopher D
1Trajectory OptimizationFrom Euler to Lawden
to Today
- Christopher DSouza
- The Charles Stark Draper Laboratory
- Houston, TX
2Why Optimize?
- Engineers are always interested in finding the
best solution to the problem at hand - Fastest
- Fuel Efficient
- Optimization theory allows engineers to
accomplish this - Often the solution may not be easily obtained
- In the past, it has been surrounded by a certain
mystique - This seminar is aimed at demystifying trajectory
optimization - Practical trajectory optimization is now within
reach - State of the art computers
- State of the art algorithms
- In order to fully appreciate trajectory
optimization, however, one must understand
something about its history - We need to understand where weve been in order
to appreciate where we are
3The Greeks started it!
- Queen Dido of Carthage (7 century BC)
- Daughter of the king of Tyre
- Fled Tyre to Tunisia
- Agreed to buy as much land as she could enclose
with one bulls hide - Set out to choose the largest amount of land
possible, with one border along the sea - A semi-circle with side touching the ocean
- Founded Carthage
- Fell in love with Aeneas but committed suicide
when he left - Story immortalized in Homers Aeneid
4The Italians Countered
- Joseph Louis Lagrange (1736-1813)
- His work Mécanique Analytique (Analytical
Mechanics) (1788) was a mathematical masterpiece - Invented the method of variations which
impressed Euler and became calculus of
variations - Invented the method of multipliers (Lagrange
multipliers) - Sensitivities of the performance index to changes
in states/constraints - Became the father of Lagrangian Dynamics
- Euler-Lagrange Equations
- Obtained the equilibrium points of the Earth-Moon
and Earth-Sun system
5The Multi-Talented Mr. Euler
- Euler (1707-1783)
- Friend of Lagrange
- Published a treatise which became the de facto
standard of the calculus of variations - The Method of Finding Curves that Show Some
Property of Maximum or Minimum - He solved the brachistachrone (brachistos
shortest, chronos time) problem very easily - Minimum time path for a bead on a string
- Cycloid
6The Plot Thickens Hamilton
and Jacobi
- William Hamilton (1805-1865)
- Published work on least action in mechanical
systems that involved two partial differential
equations - Inventor of the quaternion
- Karl Gustav Jacob Jacobi (1804-1851)
- Discovered conjugate points in the fields of
extremals - Gave an insightful treatment to the second
variation - Jacobi criticized Hamiltons work
- Only one PDE was required
- Hamilton-Jacobi equation
- Became the basis of Bellmans work 100 years later
7The Chicago School
- At the beginning of the twentieth century
Gilbert Bliss and Oskar Bolza gathered a number
of mathematicians at the University of Chicago - Made major advances in calculus of variations
following on the work of Karl Wilhelm Theodor
Weierstrass - Applied this to the field of ballistics during WW
I - Artillery firing tables
- Second Variation Conditions (conjugate point
conditions) - Built on the work of Legendre, Jacobi, and
Clebsch - Graduated many of the premiere applied
mathematicians of the early/mid 20th century - M. R. Hestenes
- E. J. McShane
8Derek and the Primer
- During the 1950s, Derek Lawden applied the
calculus of variations to exo-atmospheric rocket
trajectories - Published Optimal Space Trajectories for
Navigation - Concerned with thrusting and coasting arcs
- Invented the primer vector
- Direction is along the thrust direction
- Directly related to the velocity Lagrange
multiplier - Provided a methodology for determining optimal
space trajectories
9The Russians are Coming
Pontryagin
- In the mid 1950s a group of Russian Air Force
officers went to the Steklov Mathematical
Institute outside of Moscow to find out whether
the mathematicians could determine a particular
set of optimal aircraft maneuvers - Pontryagin, the director of the Institute,
accepted the challenge and went on to invent a
new calculus of variations - The Maximum Principle
- Used the concept of control parameters,
upravlenie, or u - Solved the original problem and in the process
revolutionized optimal control and trajectory
optimization
10The American Response Bryson
- Arthur Bryson, then at Harvard, an
aerodynamicist, came across the paper by
Pontryagin and immediately recognized its value - He applied it to a problem of finding an minimum
time to climb trajectory and presented it to the
military - It was sent to Pax River and was demonstrated by
Lt. John Young (using an altitude vs Mach number
table at 1000 ft intervals) - 338 seconds vs the predicted 332 seconds
- Path
- Accelerate to M 0.84 at just about ground level
where drag rise begins - Climb at constant Mach number to 30,000 ft
- Shallow dive to 24,000 ft followed by a slow
climb to 30000 ft, - increasing energy until the energy equals the
final energy - Climb very rapidly to desired altitude (20 km)
- Applied this new optimal control theory to
various aerospace engineering problems,
particularly those of interest to the US military
11The Inescapable Kalman
- Rudolf Kalman first came on the scene in the late
50s leading the way to the state space paradigm
of control theory along with the concepts of
controllability and observability - He then introduced an integral performance index
that had quadratic penalties on the state error
and control magnitude - Demonstrated that the optimal controls were
linear feedbacks of the state variables - Led to time varying linear systems and MIMO
systems - He later collaborated with Bucy to give us the
Kalman-Bucy filter
As some may know, these concepts were integral
to the success of the guidance and navigation
systems on the Apollo program
12Other Trajectory Optimization Legends
- Richard Bellman
- Introduced a new view and an extension of
Hamilton-Jacobi theory called Dynamic Programming
and the Hamilton-Jacobi-Bellman equation - Led to a family of extremal paths
- Provides optimal nonlinear feedback
- Curse of dimensionality
- John Breakwell
- Among the first to apply the calculus of
variations to optimal spacecraft and missile
trajectories - Prof. Angelo Miele
- Among the first to develop numerical procedures
for solving trajectory optimization problems
(SGRA) - Dr. Henry (Hank) Kelly
- Developed conditions for singular optimal control
problems (called the Kelley Conditions in Russia)
13So What?
- The brief reconnaissance into the history of
trajectory optimization is intended to
demonstrate the rich heritage which we possess - It was also intended to prepare us for a
discussion of where we are and where we are going - We began this seminar asking the question Why
optimize? - Because we are engineers and we want to find the
best solution - So, how do we go about optimizing?
14What to Optimize?
- Engineers intuitively know what they are
interested in optimizing - Straightforward problems
- Fuel
- Time
- Power
- Effort
- More complex
- Maximum margin
- Minimum risk
- The mathematical quantity we optimize is called a
cost function or performance index
15The Trajectory Optimization Nomenclature
- Dynamical constraints
- Examples equations of motion (Newtons Laws)
- Controls (u)
- Exogenous (independent) variables which operate
on the system - Examples Thrust, flight control surfaces
- States (x)
- Dependent variables which define the state of
the system - Examples position, velocity, mass
- Terminal constraints
- Conditions that the initial and final states must
satisfy - Example circular orbit with a particular energy
and inclination - Path constraints
- Conditions which must be satisfied at all points
of the trajectory - Example Thrust bounds
- Point constraints
- Conditions at particular points along the
trajectory - Examples way points, maximum heating
- Trajectory optimization seeks to obtain both the
states and the controls which optimize the chosen
performance index while satisfying the constraints
16The Optimal Control Problem
- The general trajectory optimization problem can
be posed as find the states and controls
which - subject to the dynamics
- which takes the system from to the
terminal constraints
17The Optimality Conditions and Pontryagins
Minimum Principle
These are also called the Euler-Lagrange equations
18The Optimality Conditions and Pontryagins
Minimum Principle
The boundary conditions are
There is one additional condition (sometimes
called the Weierstrass Condition) which
must satisfy
for any (the set of controls that
meet the constraints)
All of these conditions are collectively called
the Pontryagin Minimum Principle (PMP)
19Comments on the Pontryagin Minimum Conditions
- The Pontryagin conditions are very powerful tools
to help find optimal trajectories - Infinite Dimensional Conditions
- It is a two-point boundary value problem
- States are specified at the initial time
- Costates (Lagrange multipliers) are specified at
the final time - Some states (or combinations of states) are
specified at the final time - Equivalent to solving a PDE
- Most problems cannot be solved in closed form
- Closed form solutions lend themselves to analysis
- Need to use numerical methods to obtain solutions
for real-world problems - No guarantee of a solution
- Convergence issues
- Stability issues
- In the process we convert an infinite dimensional
problem into a finite dimensional problem - Implicit in numerical integration
20How to Optimize?
- Two general types of methods exist for solving
optimal control problems - Direct Methods
- Discretize the states and controls at points in
time - Nodes
- Convert the problem into a parameter optimization
problem - States and controls at the nodes become the
optimizing parameters - Use an NLP (Non-Linear Program) to solve the
parameter optimization problem - Advantages Fast Solution
- Disadvantages Difficult to determine/prove
optimality - Indirect Methods
- Operate on the Pontryagin Necessary Conditions
- This is a two-point boundary value problem
- Use Shooting methods
- Advantages Easy to determine optimality
- Disadvantages (Very) difficult to converge
21Direct Methods
- Collocation
- A method in which you choose states and controls
at points in time along the trajectory - These points are called nodes
- States and control values at the nodes become the
optimizing variables - Convert the infinite dimensional problem into a
finite dimensional, parameter optimization
problem - Enforce the constraints at the nodes
- Dynamic
- Path
- Solved using a NonLinear Program (NLP)
- Types of Spacing
- Uniform spacing
- Nonuniform spacing
22Numerical Optimization Solvers
- The general form of the nonlinear programming
problem (NLP) is - My favorite is SNOPT developed by Philip Gill
- Sparse sequential quadratic programming (SQP)
- Can be used for problems with thousands of
constraints and variables - State of the art
23Trajectory Optimization Packages
- POST (Program to Optimize Simulated Trajectories)
- Direct/Multiple shooting FORTRAN program
originally developed in 1970 for Space Shuttle
Trajectory Optimization by NASA Langley - Generalized point mass, discrete parameter
targeting and optimization program. - Provides the capability to target and optimize
point mass trajectories for a powered or
unpowered vehicle near an arbitrary rotating,
oblate planet - SORT (Simulation and Optimization Rocket
Trajectories) - FORTRAN program originally developed for ascent
vehicle trajectories - Used to generate Space Shuttle guidance targets
and maintained by Lockheed-Martin - Can be used with a optimization package to
optimize the trajectory - Variable Metric Methods
- NPSOL
- OTIS (Optimal Trajectories through Implicit
Simulation) - FORTRAN program for simulating and optimizing
point mass trajectories of a wide variety of
aerospace vehicles from NASA Glenn supported by
Boeing (Steve Paris) in Seattle - Originally developed by Hargraves and Paris
- Designed to simulate and optimize trajectories of
launch vehicles, aircraft, missiles, satellites,
and interplanetary vehicles - Can be used to analyze a limited set of
multi-vehicle problems, such as a multi-stage
launch system with a fly back booster - Hermite-Simpson collocation method which uses
NZOPT as NLP
24State of the Art Optimizers for Optimal Control
- SOCS (Sparse Optimization for Control Systems)
- General-purpose FORTRAN software for solving
optimal control problems from Boeing (Seattle) - Trajectory optimization
- Chemical process control
- Machine tool path definition
- Uses Trapezoid, Hermite-Simpson or Runge-Kutta
integration - NLP is SPRNLP written by Betts and Huffman
- Uniform node spacing, but can have multiple
intervals - Provides mesh refinement for complex problems
- DIDO (Direct and InDirect Optimization)
- Also named after Queen Dido of Carthage
- General-purpose user-friendly MATLAB software for
solving optimal control problems from NPS - Non-uniform node spacing with multiple intervals
- Legendre-Gauss-Lobatto points
- Uses a sparse numerical optimization solver
(SNOPT) - Can determine if the necessary conditions are
satisfied - Has been used to solve a wide variety of missile
and spacecraft problems - Very fast even for complex problems
25The Wave of the Future Pseudospectral Methods
- Pseudospectral methods choose the collocation
points in such a way as to minimize integration
error - Number of nodes dependent on accuracy desired
- The nodes are non-uniformly spaced in time
- Quadratic spacing at the ends
- Number determines the spacing
- They use (global basis) functions which
(optimally) approximate the states and controls
and enforce the (dynamic and path) constraints at
the nodes over the interval -1, 1 - Chebyshev-Gauss
- Legendre-Gauss
- Chebyshev-Gauss-Lobatto
- Legendre-Gauss-Lobatto
- Pseudospectral methods yield spectral accuracy
- Optimal interpolation
- Particularly well suited for trajectory
optimization problems where much of the activity
occurs at the ends of the intervals
Includes the end points
26Pseudospectral Point Distribution (N 10)
Quadratic clustering at ends
27Launch Vehicle Example Three Stage to Orbit
- Suppose we wish to find the optimal trajectory
for a three stage vehicle to get the maximum
payload to orbit - Performance index
- Differential constraints (equations of motion)
- Terminal constraints
- Throttle capability (minimum, maximum specified)
- Coast of at least 5 seconds between second and
third stage - Maximum of 115 seconds
28Problem Specific Issues
- Coordinate Systems
- Dynamics
- Inertial
- Spherical
- Equinoctial
- Controls
- Angles
- Thrust components
- Direction cosines
- Scaling
- For good convergence properties, we need all the
variables to be of order 1 - So we scale the states, the controls and the time
to achieve this - The art of trajectory optimization
- Tuning knobs
29Three Stage to Orbit Thrust Profile
Maximum Thrust
Minimum Thrust
Coast
30Three Stage to Orbit Thrust Direction Profile
Second Stage Separation
First Stage Separation
31Three Stage to Orbit Mass Profile
First Stage Separation
Second Stage Separation
Coast
32Orbit Transfer
- Optimal transfers between two orbits have been
the subject of directed research for the past 40
years - Much analytical and computational effort has been
devoted to this task - Primer vector theory has been applied
- Numerical solutions are sometimes difficult to
obtain - The Legendre PseudoSpectral (LPS) method has been
used to extensively analyze this problem - Impulsive burn approximations
- Finite burn effects
- Types of coordinate systems
- Cartesian
- Equinoctial
- Nonsingular orbital elements
33Impulsive Orbit Transfer
Elliptical-Elliptical Hohmann Transfer Analytic
Solution ?v1 2076.72 m/s ?v2 87.46 m/s LPS
Solution ?v1 2076.71 m/s ?v2 87.49 m/s
Elliptical-Elliptical Transfer with Inclination
Change Analytic Solution ?v1 2106.13 m/s ?v2
239.69 m/s LPS Solution ?v1 2106.17 m/s ?v2
239.65 m/s
34Finite Burn Orbit Transfer LEO (ISS) to LEO (Sun
Synchronous)
Orbital Elements Initial Final Orbit
a 6772 km 7062 km
e 7.08E-4 1.115E-3
i 51.6o 98.2o
W 58.6o 120.1o
w 238.3o 282.0o
- Finite Burn Accumulated DV
- DV 8027.5 m/s
- Impulsive Burn Accumulated DV
- DV 6548.6 m/s
35Further Applications of LPS
- ISS Momentum Desaturation
- Constellation Design
- Libration point formation designs
- Entry Trajectory Design
- Planetary Mission Design
36What is Next? -- MAHC
- Multi-Agent Hybrid Control (MAHC)
- 21st Century extension of 20th Century optimal
control - A general optimization framework for multiple
vehicles - Multiple constraints on each vehicle
- Allow for discrete decision variables
- Example
- Two stage vehicle
- Return vehicle must land at a particular point
- Latitude -28.25N 1 km
- Longitude -70.1 E 1 km
- Ascent vehicle continues to a desired orbit while
maximizing mass to orbit - The discrete state space is as follows
37Multi-Agent Hybrid Trajectory Optimization
Example Position Profile
38Multi-Agent Hybrid Trajectory Optimization
Example
39Hybrid Trajectory Optimization Example Control
History
40What is Next? -- Real-time Trajectory
Optimization
- Real-time trajectory optimization
- Computational capability is increasing with
Moores law - Time is approaching when these (direct) methods
can be implemented on board vehicles and
optimized in real-time - 1 Hz
- Guidance cycles (outer loop) slower than control
cycles (inner loop) - Application to orbit (transfer) problem
- Issues
- Convergence
- Stability of solutions
41What is Next? - NOG
- Neighboring Optimal Guidance (NOG)
- A real-time guidance scheme which determines a
new optimal path which is close to the nominal
(a priori) optimal path - Neighboring optimal
- Operates on deviations from the optimal
trajectory - Very robust
- Based upon the second variation sufficient
conditions
42Conclusion
- Trajectory optimization has advanced greatly over
the past 40 years - We are at the threshold of a new era for solving
exciting complex optimization problems - New methods exist for solving (general) optimal
control problems - Trajectory optimization problems are a subset of
this class - These methods give (reasonably) fast solutions
even given poor guesses - Fast computers
- Good algorithms
- Dont need to know the details of the methods or
devote your career to optimization - Just your problem
- Solution of complex trajectory optimization
problems is within reach of the practicing
engineer
43Selected References
- Lietmann, G., Optimization Techniques, Academic
Press, 1962. - Lawden, D.F., Optimal Trajectories for Space
Navigation, Butterworths, 1963. - Bryson, A.E. and Ho, Y-C., Applied Optimal
Control, Hemisphere Publishing Company, 1975. - Gill, P.E., Murray, W., and Wright, M.H.,
Practical Optimization, Academic Press, 1981. - Fletcher, R., Practical Methods of Optimization,
Wiley Press, 1987. - Betts, J.T., Practical Methods for Optimal
Control Using Nonlinear Programming, SIAM
Advances in Control and Design Series, 2001.
44Questions?