Physical Sciences and Mathematics Commons^™

Time-Dependent Performance Comparison Of Stochastic Optimization Algorithms, David Shilane, Jarno Martikainen, Seppo Ovaska

U.C. Berkeley Division of Biostatistics Working Paper Series

This paper proposes a statistical methodology for comparing the performance of stochastic optimization algorithms that iteratively generate candidate optima. The fundamental data structure of the results of these algorithms is a time series. Algorithmic differences may be assessed through a procedure of statistical sampling and multiple hypothesis testing of time series data. Shilane et al. propose a general framework for performance comparison of stochastic optimization algorithms that result in a single candidate optimum. This project seeks to extend this framework to assess performance in time series data structures. The proposed methodology analyzes empirical data to determine the generation intervals in …