Designing Efficient Evolutionary Algorithms for Cluster Optimization: A Study on Locality

Abstract

Cluster geometry optimization is an important problem from the Chemistry area. Hybrid approaches combining evolutionary algorithms and gradient-driven local search methods are one of the most efficient techniques to perform a meaningful exploration of the solution space to ensure the discovery of low energy geometries. Here we perform a comprehensive study on the locality properties of this approach to gain insight to the algorithm's strengths and weaknesses. The analysis is accomplished through the application of several static measures to randomly generated solutions in order to establish the main properties of an extended set of mutation and crossover operators. Locality analysis is complemented with additional results obtained from optimization runs. The combination of the outcomes allows us to propose a robust hybrid algorithm that is able to quickly discover the arrangement of the cluster's particles that correspond to optimal or near-optimal solutions.