The Optimization Ability of Evolved Strategies

Abstract

Hyper-Heuristics (HH) is a field of research that aims to automatically discover effective and robust algorithmic strategies by combining low-level components of existing methods and by defining the appropriate settings. Standard HH frameworks usually comprise two sequential stages: Learning is where promising strategies are discovered; and Validation is the subsequent phase that consists in the application of the best learned strategies to unseen optimization scenarios, thus assessing its generalization ability.

Evolutionary Algorithms are commonly employed by the HH learning step to evolve a set of candidate strategies. In this stage, the algorithm relies on simple fitness criteria to estimate the optimization ability of the evolved strategies. However, the adoption of such basic conditions might compromise the accuracy of the evaluation and it raises the question whether the HH framework is able to accurately identify the most promising strategies learned by the evolutionary algorithm. We present a detailed study to gain insight into the correlation between the optimization behavior exhibited in the learning phase and the corresponding performance in the validation step. In concrete, we investigate if the most promising strategies identified during learning keep the good performance when generalizing to unseen optimization scenarios. The analysis of the results reveals that simple fitness criteria are accurate predictors of the optimization ability of evolved strategies.