IMRT Course 2016

• Optimization algorithms find a solution by minimizing a function.

• Virtually all algorithms use gradient information for the downhill search.

Optimization Algorithms

• The size of the optimization problem limits the number of available

algorithms to

quasi-Newton methods

and

Conjugate Gradient search

If implemented well, none should be superior to the others.

• These algorithms are deterministic, so that given infinite time, they find

ESTRO IMRT course 2016

IMRT optimization: algorithms&cost functions

London April 4th, 2016

-20-

a minimum, which is unique in the absence of delivery constraints.

• Some delivery constraints make the problem virtually unsolvable.

• If nothing else helps: stochastic algorithms (like Simulated Annealing)

which perform a sparse, random search of the entire solution

space (which is very, very big). Stochastic algorithms are the last

resort when the problem is otherwise intractable.