Benchmarking Global Optimization Algorithms

Abstract

We benchmark seven global and three local algorithms by comparing their performance and speed in optimizing difficult objective functions. We apply the algorithms to optimize a small suite of multidimensional test functions that are commonly used to benchmark algorithms in computational mathematics. To understand optimizers’ performance in applications that are common to economics, we apply the same optimizers to maximize the objective function of a GMM estimation problem that targets 297 moments to estimate 7 parameters. Our results show that the reliability and speed of all algorithms vary substantially depending on the dimensionality and characteristics of the problem. Experimenting with different algorithms can therefore be very helpful. We find that StoGo and Tiki-Taka Algorithm (Tiktak) are most reliable and computationally efficient in the optimization of test functions. For the economic estimation, the most reliable and efficient algorithms are Multi-Level Single-Linkage and Tiktak

Publication
Date