Contrastive Learning: An Alternative Surrogate for Offline Data-Driven Evolutionary Computation

Abstract

Offline data-driven evolutionary algorithms (DDEAs), which learn problem models from historical data and then perform optimization, have attracted significant attention in the data-driven age. Most existing studies build surrogate models based on regression methods to predict the fitness of each solution, which depends heavily on the quality and quantity of offline data. Considering the evolution trait of evolutionary algorithms (EAs), the absolute fitness of each individual is not essential, instead, the relative strengths of individuals are adequate. This article explores an alternative way to realize DDEAs by establishing a contrastive learning model that performs binary classification to determine the pros and cons between individuals. The task of binary classification is relatively simpler than regression, and meanwhile the training data is inherently augmented to the square of the origin. The proposed contrastive learning model is implemented based on a siamese neural network to measure the differences between solutions. Further, we use the predicted pairwise relationship between individuals to construct a directed graph and propose a topological sort algorithm on the graph to obtain the ranking of the population. During the topological sort, a regression model based on local principle is used to resolve some conflicting issues. Integrating the above components, an offline DDEA named contrastive learning-based DDEA (CL-DDEA) is put forward. Experiments and comparisons with state-of-the-arts validate the powerfulness of CL-DDEA, especially on high-dimensional problems.