Configuring differential evolution adaptively via path search in a directed acyclic graph for data clustering

Abstract

Abstract As an efficient data mining technique, data clustering has been widely-used for data analysis and extracting valuable hidden information. Leveraging the simplicity and effectiveness, the evolutionary optimization-driven clustering algorithms have exhibited promising performance and attracted tremendous attention. Up to the present, how to enable these algorithms to escape from local optima and accelerate convergence rates is an ongoing challenge. In this paper, we propose a novel adaptive Differential Evolution (DE) variant to deal with the above challenge when clustering data. In the improved DE algorithm, the four interdependent components, including mutation strategy, crossover strategy, scaling factor value, and crossover rate, are adaptively configured in an integrated manner via ant colony optimization (ACO) during the problem-solving process. To be specific, the relationships of four components in the DE algorithm are modeled as a directed acyclic graph, and a path in the graph exactly corresponds to a configuration for DE. During the optimization process, ant colony optimization is employed to search for a reasonable path for each individual of DE in terms of pheromones on arcs. In this manner, the configuration of the four interdependent components of DE will be generated dynamically, which is then used to guide the successive search behaviors of individuals in DE. Each individual has a path, representing a configuration for each component. After each iteration, individuals that generate promising solutions are allowed to deposit pheromone on the paths, resulting in more pheromones on the arcs appearing in better algorithm configurations (paths) more frequently. Through this manner, the search strategies and parameters of DE are comprehensively adapted by ACO. The proposed algorithm is named ACODE for short. To verify its effectiveness, the proposed ACODE is compared with four representative data clustering algorithms on eight widely-used benchmark datasets. The experimental results demonstrate the advantages of ACODE over half of the datasets.