Variable Importance and Interaction in Differential Evolution: Online Learning and a New Recombination Operator

• Renato Tinós USP
• Francisco Chicano University of Málaga
• Michal W. Przewozniczek Wroclaw University of Science and Technology

Resumo

Offline Linkage Learning (LL) techniques have been used for estimating variables interaction. An online LL technique for Differential Evolution (DE) is proposed. In addition, the proposed technique estimates variable importance. Variable importance and interaction can be relevant to understand properties of optimization instances. We estimate variable importance and interaction in problems of the IEEE CEC’17 Competition on Single Objective Real-Parameter Numerical Optimization using the proposed technique. The resulting importance and interaction graph provides valuable insights, revealing key properties of the problems. Variable importance and interaction can also be useful for designing efficient crossover. We propose the Real Variable Interaction Crossover (RVIntX), which explores the epistatic structure of the problem and the variables common to both parents to find good recombination masks. Experimental verification confirms that RVIntX can improve the quality of the results. Additionally, although some operators based on linkage information are not useful in solving problems with high epistasis, RVIntX does not suffer from this disadvantage.