20.04.2014 Views

design space pruning heuristics and global optimization method for ...

design space pruning heuristics and global optimization method for ...

design space pruning heuristics and global optimization method for ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

more robust convergence. The solution is generally less sensitive to the initial guesses<br />

<strong>and</strong> those initial guesses are more physically intuitive, which make direct <strong>method</strong>s<br />

preferable <strong>for</strong> implementing within an automated <strong>global</strong> <strong>optimization</strong> scheme.<br />

Differential dynamic programming (DDP) also parameterizes the control<br />

variables, providing a large convergence domain <strong>and</strong> decreasing the sensitivity to poor<br />

initial guesses. As compared to direct <strong>method</strong>s, DDP is less sensitive to the high<br />

dimensionality of the low-thrust trajectory <strong>optimization</strong> problem as it trans<strong>for</strong>ms the large<br />

problem into a succession of low dimensional sub-problems. Quadratic programming is<br />

then used on each resulting quadratic sub-problem to solve <strong>for</strong> controls that improve the<br />

trajectory locally. The states <strong>and</strong> objective function are then calculated <strong>for</strong>ward in time<br />

using the updated controls, <strong>and</strong> the process is repeated until the problem has converged.<br />

One disadvantage of DDP is that it is most effective <strong>for</strong> smooth unconstrained problems.<br />

Low-thrust problems, however, tend to include numerous constraints <strong>and</strong> can be highly<br />

non-smooth. In recent work, Lantoine <strong>and</strong> Russell have modified the traditional DDP<br />

algorithm to create a hybrid differential dynamic programming algorithm that addresses<br />

some of the weaknesses of DDP. The hybrid approach uses first- <strong>and</strong> second-order state<br />

transition matrices to calculate the partial derivates required <strong>for</strong> <strong>optimization</strong>, <strong>and</strong><br />

combines DDP with NLP techniques to increase its robustness <strong>and</strong> efficiency. 222324<br />

Finally, there also exist hybrid <strong>method</strong>s which numerically integrate the Euler-<br />

Lagrange equations <strong>and</strong> control the <strong>space</strong>craft based on the primer vector.14 ,25<br />

As in the<br />

direct <strong>method</strong>, hybrid <strong>method</strong>s solve a nonlinear programming problem, but with the<br />

Lagrange multipliers making up part of the parameter vector while maximizing or<br />

minimizing some cost function. Hybrid <strong>method</strong>s search numerically <strong>for</strong> the set of<br />

parameters that extremize the cost function, while explicitly satisfying kinematic<br />

boundary constraints. According the work by Gao <strong>and</strong> Kleuver, the advantages of hybrid<br />

trajectory <strong>optimization</strong> <strong>method</strong>s include a significant reduction in the <strong>design</strong> <strong>space</strong> <strong>and</strong><br />

7

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!