A direct search method for smooth and nonsmooth unconstrained optimization
A derivative free frame based method for minimizing C¹ and non-smooth functions is described. A 'black-box' function is assumed with gradients being unavailable. The use of frames allows gradient estimates to be formed. At each iteration a ray search is performed either along a direct search quasi-Newton direction, or along the ray through the best frame point. The use of randomly oriented frames and random perturbations is examined, the latter yielding a convergence proof on non-smooth problems. Numerical results on non-smooth problems show that the method is effective, and that, in practice, the random perturbations are more important than randomly orienting the frames. The method is applicable to nonlinear ℓ₁ and ℓ∞ data fitting problems, and other nonsmooth problems.