Estimating stop conditions of swarm based stochastic metaheuristic algorithms
Peter Frank Perroni, Daniel Weingaertner, Myriam Delgado · Proceedings of the Genetic and Evolutionary Computation Conference · 2017
When dealing with metaheuristics, one important question is how many evaluations are worth spending in the search for better results. This work proposes a method to estimate the best moment to stop swarm iterations based on the analysis of the convergence behavior presented during optimization, aiming to provide an effective balance between saving fitness evaluations and keeping the optimization quality. An automated Convergence Stabilization Modeling operating in Online mode (CSMOn) is proposed based on a sequence of linear regressions using exponential and log-like curves. The method was tested on the CEC13 benchmark with CCPS02-IP E algorithm and on 30 random Max-Set functions with the swarm algorithms PSO, ABC and CCPS02-IP E. CEC13 results show that up to 90% less fitness evaluations are performed for functions where CCPS02-IP E has a steady convergence, and up to 49% for functions where convergence is erratic, while penalties for fitness are kept small. Max-Set results demonstrates the robustness of CSMOn for the search algorithms tested. We conclude that CSMOn is capable of adapting to an optimization in progress, producing a good trade-off between result quality and evaluation savings.