@phdthesis{,
    author = {Künzel, Steven},
    title = {Evolving Artificial Neural Networks for Multi-Objective Tasks},
    editor = {},
    booktitle = {},
    series = {},
    journal = {},
    address = {},
    publisher = {},
    edition = {},
    year = {2021},
    isbn = {},
    volume = {},
    number =  {},
    pages = {},
    url = {},
    doi = {},
    keywords = {Neuroevolution, Multi-Objective Optimization, Evolutionary Algorithms},
    abstract = {Neuroevolution is an active research field in artificial intelligence. It aims at evolving artificial neural networks using evolutionary methods. Today, artificial intelligence is of continuously growing importance. Among other techniques, neural networks do also play a key role. This thesis develops nNEAT, a novel neuroevolutionary algorithm based on the quasi-standard NEAT, considering multiple objectives being optimized concurrently. Furthermore, the aspect of automatic parameter control is addressed to increase the usability of the new algorithm. Due to the relevance of sorting solutions qualitatively in evolutionary algorithms, a novel sorting framework for multi-objective environments is introduced. It allows combining various measures from different spaces ad-hoc, without requiring considerations about scaling and integration. A promising quality measure is the R2 indicator. To foster its application, a new algorithm has been developed to compute the exact R2 contribution of each solution within a set. It requires less computational effort by factor μ (set size) in comparison to the conventional approach. Further results of the thesis are a multi-objective variant of the well-known Double Pole Balancing problem, as well as a lightweight and fast version of the Fighting Game AI Competition simulator. Additionally, an abstract racing car simulation with a basic physics engine was developed, which can be applied for training neural networks for realistic racing car simulations like TORCS. It provides a speed-up of more than factor 110 in comparison to TORCS on our test system. The main contribution of this work is the design and implementation of a multi-objective neuroevolutionary algorithm that can be applied without expert knowledge and thus supports multi-objective neuroevolution becoming a more frequently used utility.},
    note = {},
    
    school = {Universität der Bundeswehr München},
    
}