Scalable Neuroevolution of Ensemble Learners

In recent years, machine learning has become increasingly important in daily life. One of the most popular machine learning models used in many applications is an Artificial Neural Network (ANN). While in applications such as automatic speech recognition there is sufficient knowledge about the expected behavior for each input to use supervised learning, other applications like robot control define only an overall target so that the expected output for a given input can be ambiguous, making supervised learning inapplicable. Therefore, Topology and Weight Evolving ANN (TWEANN) has been developed in the past to evolve ANN topologies and connection weights. However, challenges of TWEANN are the design of genetic recombination and the exploration of huge search spaces for suitable solutions induced in particular by large-scale problems which can lead to impractical runtimes. To address the aforementioned issues, this paper proposes a novel evolutionary framework to evolve ensemble learners as specialized ANNs in a divide-and-conquer manner. Specifically, genetic recombination is heavily orchestrated for ANN combinations to effectively find suitable solutions in the search space restricted by ensemble methods. Experiments on various benchmark problems for solving control tasks show that the proposed framework clearly outperforms existing TWEANN algorithms.