Quantitative Performance Measures for Dynamic Optimization Problems

Abstract

Measuring the performance of algorithms over dynamic optimization problems (DOPs) presents some important differences when compared to static ones. One of the main problems is the loss of solution quality as the optimization process advances in time. The objective in DOPs is in tracking the optima as the landscape changes; however, it is possible that the algorithm gets progressively further from the optimum after some changes happened. The main goal of this chapter is to present some difficulties that may appear while reporting the results on DOPs, and introduce two new performance tools to overcome these problems. We propose a measure based on linear regression to measure fitness performance degradation, and analyze our results on the moving peaks problem, using several measures existing in the literature as well as our performance performance degradation measure. We also propose a second measure based on the area below the curve defined by some population attribute at each generation (e.g., the best-of-generation fitness), which is analyzed in order to see how it can help in understanding the algorithmic search behavior.