Learning to Describe Collective Search Behavior of Evolutionary Algorithms in Solution Space

Evolutionary algorithms (EAs) are a kind of population-based meta-heuristic optimization methods, which have proven to have superiorities in solving NP-complete and NP-hard optimization problems. But until now, there is lacking in the researches of effective representation method to describe the collective search behavior of the Evolutionary Algorithm, while it is useful for researchers and engineers to understand and compare different EAs better. In the past, most of the theoretical researches cannot directly guide for practical applications. To bridge the gap between theoretical research and practice, we present a generic and reusable framework for learning features to describe collective behavior of EAs in this paper. Firstly, we represent the collective behavior of EAs with a parent-child difference of population distribution encoded by self-organizing map (SOM). Then, we train a Convolutional Neural Network (CNN) to learn problem-invariant features from the samples of EAs’ collective behavior. Lastly, experiment results demonstrate that our framework can effectively learn discriminative features representing collective behavior of EAs. In the behavioral feature space stretched by the obtained features, the collective behavior samples of various EAs on various testing problems exhibit obvious aggregations that highly correlated with EAs but very weakly related to testing problems. We believe that the learned features are meaningful in analyzing EAs, i.e. it can be used to measure the similarity of EAs according to their inner behavior in solution space, and further guide in selecting an appropriate combination of sub-algorithm of a hybrid algorithm according to the diversity of candidate sub-algorithm instead of blind.