Conditional Gradient Method
Take some convex optimization problem:
minimize
subject to
Here, is some convex set. If
is hard to characterize, this optimization problem might be hard to handle in a direct fashion. For example,
might require enforcing some exponential number of constraints. In some problems, it nevertheless turns out to be possible to solve the following problem:
minimize
subject to
That is, for linear objective functions, the problem is solvable. (This might be, for example, by some graph algorithm). Anyway, the Conditional-Gradient method solves the original problem by repeatedly relinearizing it, and solving the linear optimization. One can then take a step to some point between the current position, and the solution to the linearized problem.
Initialize anywhere.
Repeat:
1)
2)
3) , for some
.
In step 3, controls how far
moves towards
. Notice that because
is convex, for any
,
The information I can find doesn’t make any argument for convergence. Clearly, never gets worse, but I want a proof of the following:
If is not at the global minimum, the line search will find some
.