Physics Bucharest 2017

6/3/2017

Optimisation methods

Gradient descent Only accepts parameter changes which reduce cost function

Simulated annealing Will allow increases in cost function with reducing probability

ESTRO physics course 2017

33

The inverse problem Gradient descent can end in a local minima, whereas simulated annealing will allow, some “uphill” – like running down a ski slope.

ESTRO physics course 2017

34

C. Ma et al.