A Multiresolutional Estimated Gradient Architecture for Global Optimization

In this paper we present a novel optimization algorithm that estimates gradients over regions to search for optima of a non-convex function on both a local and global scale. The proposed architecture is based on three concepts: using the memory of previously evaluated points, multiresolutional search, and the estimation of gradients at these different resolutions to direct the search. This multiresolution estimated gradient architecture (MEGA) shows promise to perform competitively when compared to standard global searches. Comparisons on the Rosenbrock, Griewank, and sinusoidal test functions show that MEGA can converge faster than particle swarm optimization, particularly as dimensionality of a problem increases.