Evolving Artificial Neural Networks for Multi-objective Tasks

• Published in: Applications of Evolutionary Computation Vol. 10784; pp. 671 - 686
• Main Authors: Künzel, Steven, Meyer-Nieberg, Silja
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2018; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Evolutionary algorithms; Multi-objective; NEAT; Neuroevolution
• ISBN: 9783319775371; 3319775375
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-77538-8_45

Neuroevolution represents a growing research field in Artificial and Computational Intelligence. The adjustment of the network weights and the topology is usually based on a single performance criterion. Approaches that allow to consider several – potentially conflicting – criteria are only rarely taken into account. This paper develops a novel combination of the NeuroEvolution of Augmenting Topologies (NEAT) algorithm with modern indicator-based evolutionary multi-objective algorithms, which enables the evolution of artificial neural networks for multi-objective tasks including a large number of objectives. Several combinations of evolutionary multi-objective algorithms and NEAT are introduced and discussed. The focus lies on variants with modern indicator-based selection since these are considered as efficient methods for higher dimensional tasks. This paper presents the first combination of these algorithms and NEAT. The experimental analysis shows that the novel algorithms are very promising for multi-objective Neuroevolution.