Estimating Hypervolume using Population Features from Dynamic Compartmental Models

抄録

Characterizing an evolutionary algorithm's behavior and performance is a step towards having tools to automatically select and configure the algorithm that better solves the problem at hand. A promising way to characterize algorithms is to use models that capture their dynamics. Dynamic compartmental models are inspired by epidemiology models to study the dynamics of multi- and many-objective evolutionary algorithms. These models have been used as a tool for algorithm analysis, algorithm comparison, and algorithm configuration, assuming that the Pareto optimal set is known. In this work, we relax this assumption by considering the most recent non-dominated set and propose features that allow the use of dynamic compartmental models on large problems. We then introduce a model to estimate the hypervolume from the changes observed on non-dominated solutions in the population. We use several instances of MNK-landscapes with 3, 4, and 5 objectives, and we show that the models can effectively learn algorithm behavior and estimate the search performance of a multi-objective algorithm on those instances. We also show that the models produce good estimates on unseen instances of the same class of problems, and capture the variability of the algorithm when initialized with different populations.