RANDOMIZED BLOCK COORDINATE DESCENDANT STRONG FOR LARGE-SCALE STOCHASTIC OPTIMIZATION

STRONG is a response surface methodology based algorithm that iteratively constructs linear or quadratic fitness model to guide the searching direction within the trust region. Despite its elegance and convergence, one bottleneck of the original STRONG in high-dimensional problems is the high cost per iteration. This paper proposes a new algorithm, RBC-STRONG, that extends the STRONG algorithm with the Random Coordinate Descent optimization framework. We proposed a RBC-STRONG algorithm and proved its convergence property. Our numerical experiments also show that RBC-STRONG achieves better computational performance than existing methods.