MAP-Elites for Genetic Programming-Based Ensemble Learning: An Interactive Approach [AI-eXplained]

Evolutionary ensemble learning is an emerging research area, and designing an appropriate quality-diversity optimization algorithm to obtain a set of effective and complementary base learners is important. However, how to maintain such a set of learners remains an open issue. This paper proposes using cosine similarity-based dimensionality reduction methods to maintain a set of effective and complementary base learners within the MAP-Elites framework for evolutionary ensemble learning. Additionally, this paper proposes a reference point synthesis strategy to address the issue of individuals being unevenly distributed in semantic space. The experimental results show that the ensemble model induced by the cosine similarity-based dimensionality reduction method outperforms the models induced by the other seven dimensionality reduction methods in both interactive examples and large-scale experiments. Moreover, reference points are shown to be helpful in improving the algorithm's effectiveness. The main contribution of this paper is to provide an interactive approach to explore the methods and results, which is detailed in the full paper presented in IEEE Xplore.