Geodesic and contour optimization using conformal mapping

摘要

We propose a novel optimization algorithm for differentiable functions utilizing geodesics and contours under conformal mapping. The algorithm can locate multiple optima by first following a geodesic curve to a local optimum then traveling to the next search area by following a contour curve. Alongside we implement a jumping mechanism which we call shadow casting to help geodesics jump to locations closer to the global optimum. To improve the efficiency, local search methods such as the Newton-Raphson algorithm are also employed. For functions with many optima or when the global optimum is very close to a local one, numerical analyses have shown that the resulting algorithm, SGEO-QN, can outperform recent derivative-free DIRECT variants in number of function/gradient evaluations. The results also indicate that under certain conditions, number of function/gradient evaluations for SGEO-QN scales nearly linearly with increasing dimensionality. Lastly, SGEO-QN appears to be less affected by rotational transforms of the objective functions than the variants of DIRECT compared.