Two strategies of adaptive cluster covering with descent and their comparison to other algorithms (Q1284411)

The methodology of global optimization is enriched by two strategies of randomized search based on clustering approaches: the adaptive cluster covering (ACCO) and the Adaptive Cluster Covering with descent (ACD). These methods apply to the set of multi-extremum minimization problems and can be compared to techniques as Controlled Random Search (CRS), as Genetic Algorithms (GA) and as multistart methods such as Multis and M-simplex. The basic ideas of adaptive clustering covering is to subdivide the domain of search for the global optimum into smaller subdomains by way of clustering; then the objective functions are assessed in the subdomains and using these results the subdomains are shrinked. The adaptive aspect of the strategy refers to a possible shifting in subregion of search together with a shrinking of it and a change in density of covering depending on the previous assessments of the global minimizer. As by these different steps the true localization of the global optimum is not yet assured, an action of periodic re-randomization of the initial population is added. The adaptive clustering covering with descent uses in addition the simplex descent features to look for better points within each subdomain obtained by clustering. An extensive comparative work has been performed to look for the best algorithm with respect to criteria such as effectiveness, efficiency and reliability. The problems considered varied from 2 to 10 dimensions and presented typically 1 to 10 local optima. Globally the ACCO method appeared to perform best, being comparable to CRS but faster. The ACD algorithm proved to be efficient and effective for functions with higher dimensions but more accurate tuning of its parameters is needed to improve its reliability.