- author = "Julia Reuter",
- title = "Development of symbolic models using genetic programming and domain knowledge",
- school = "Otto von Guericke University",
- year = "2025",
- address = "Magdeburg, Germany",
- month = "18 " # jun,
- keywords = "genetic algorithms, genetic programming, Kuenstliche Intelligenz, Evolution, Biologie, ANN, Symbolic regression, Coevolution, IK-CCGP, MOGP, Island Model, IMGP, Unit Aware, Inverse Kinematics",
- URN = "urn:nbn:de:gbv:ma9:1-1981185920-1212625",
-
URL = "
https://portal.dnb.de/opac.htm?method=simpleSearch&cqlMode=true&query=idn%3D1043471448",
-
URL = "https://katalog.dnb.de/EN/resource.html?id=041975545&sortA=bez&sortD=-dat&v=plist",
-
URL = "https://www.ci.ovgu.de/Publications/PhD+Thesis.html",
-
URL = "https://opendata.uni-halle.de/handle/1981185920/121262",
-
URL = "
https://www.ci.ovgu.de/is_media/PhDThesis/Reuter_Julia_Dissertation_2025.pdf",
-
DOI = "
10.25673/119304",
- size = "213 pages",
-
abstract = "Problems from the science and engineering area come
with challenging characteristics: They are often
high-dimensional, highly non-linear, and have not yet
been solved by humans due to their enormous complexity.
While black-box machine learning methods such as deep
neural networks can achieve high accuracies on such
problems, the underlying relations remain opaque.
However, it is essential for experts in science and
engineering to analyze and understand the learned
models. Symbolic regression (SR) is the construction of
mathematical expressions from data to identify the
relation between input and target variables. While
traditional regression methods assume an underlying
model structure, SR learns both the model structure and
its associated parameters. Genetic programming (GP), a
method from the family of evolutionary algorithms, is a
widely used method for SR, because of its white box
nature and its capability to optimize multiple
objectives simultaneously. Often, domain knowledge is
available that can contribute to discovering novel
symbolic models. The goal of this thesis is to develop
symbolic regression algorithms that can tackle complex
science and engineering problems by making use of the
valuable knowledge provided by domain experts.
In the related literature, various approaches to integrating domain knowledge into algorithms have been proposed and applied to real-world applications. This thesis aims to close a gap in this field by outlining and classifying these methods. Moreover, various techniques have been proposed to improve components of the GP algorithm that are relevant in practice, some of which are further improved upon in this thesis.
This thesis proposes two benchmark problems from robotics and fluid mechanics, and establishes a comparative baseline to evaluate the efficacy of the newly developed methods. To reduce the number of features of the high-dimensional problems, an inductive bias fitting the nature of the problem is proposed. Given the high complexity of the approached problems and the non-deterministic nature of the GP algorithm, methods to improve the repetition stability of a GP algorithm are additionally investigated. Furthermore, this thesis proposes methods to fulfill important expert requirements, such as methods to handle physical unit constraints. In this context, multi-objective optimization plays a pivotal role, allowing for the exploration of a diverse set of solutions while effectively optimizing multiple criteria. Empirical evaluations and case studies considering various problems with both known and unknown equations from the engineering and science area validate the proposed approaches. The results demonstrate how domain knowledge can improve the accuracy of symbolic models, while tackling increased problem complexities and developing meaningful equations for domain experts.",
- language = "English",
-
notes = "also known as
\cite{Share_it_1981185920-121262}
idn=1043471448
Supervisor: Sanaz Mostaghim",
