Font Size:
Spectral gradient method for stochastic optimization
Last modified: 2014-04-28
Abstract
We consider constrained optimization problems with the objective function in a form of mathematical expectation and convex, compact feasible set. At every iteration, the proposed method uses a sample average in order to approximate the objective function. Sample size is updated at every iteration and it does not have to be monotonically increasing. Line search along the direction of a spectral projected gradient is employed. Feasibility is maintained throughout the whole optimization process. Almost sure convergence is analyzed.