Extreme reactive portfolio (XRP): Tuning an algorithm population for global optimization

Given the current glut of heuristic algorithms for the optimization of continuous functions, in some case characterized by complex schemes with parameters to be hand-tuned, it is an interesting research issue to assess whether competitive performance can be obtained by relying less on expert developers (whose intelligence can be a critical component of the success) and more on automated self-tuning schemes. After a preliminary investigation about the applicability of record statistics, this paper proposes a fast reactive algorithm portfolio based on simple performance indicators: record value and iterations elapsed from the last record. The two indicators are used for a combined ranking and a stochastic replacement of the worst-performing members with a new searcher with random parameters or a perturbed version of a well-performing member. The results on benchmark functions demonstrate a performance equivalent or better than that obtained by offline tuning schemes, which require a greater amount of CPU time and cannot take care of individual structural variations between different problem instances.