Frank Wolfe algorithms for non-convex optimization

76 Views Asked by At

I have a non-convex function as follow

image

where: $\alpha < 1$ and $\theta \in \bigtriangleup_{K}$ means:

$\theta_1 + \theta_2 + ... + \theta_K = 1$ and $\theta_i \ge \epsilon > 0, \quad i=1,...,K$ ($\epsilon$ is a small constant)

I design a FW algorithm as follows:

click here to see image

How can I evaluate the convergence rate of the algorithm above? Can I change step size to achieve the convergence rate is O(1/t)? (t is number of iterations)