- author = "Henning Cui",
- title = "Improving Genetic Operators of Cartesian Genetic Programming",
- school = "Institut fuer Informatik, Fakultaet fuer Angewandte Informatik, Universitaet Augsburg",
- year = "2025",
- address = "Germany",
- month = "27 " # nov,
- keywords = "genetic algorithms, genetic programming, Cartesian Genetic Programming, Genetische Programmierung",
-
URL = "
https://opus.bibliothek.uni-augsburg.de/opus4/frontdoor/index/index/year/2026/docId/126712",
-
URL = "
https://opus.bibliothek.uni-augsburg.de/opus4/frontdoor/deliver/index/docId/126712/file/Cui_Diss.pdf",
-
URL = "https://opus.bibliothek.uni-augsburg.de/opus4/126712",
- size = "240 pages",
-
abstract = "Cartesian Genetic Programming (CGP) belongs to the
family of Evolutionary Algorithms and can be used for a
wide range of supervised machine learning applications.
It has been extensively used to automatically generate
digital circuits, shows great potential in the
automated design of deep neural networks, and has also
been successfully used for (interpretable medical)
image processing tasks, to name a few examples.
Additionally, CGP profits from generating relatively
small programs, further increasing its potential as
this characteristic improves the interpretability of
its solutions.
Since the invention of CGP, many advantages and shortcomings have been researched by various scientists. Moreover, new operators and extensions have been proposed to mitigate its disadvantages and/or improve its performance. Although there is still a lot of foundational research to be done, as some CGP specific mechanisms and behaviours are not fully understood. By further enhancing its evolutionary operators, CGP performance can also be improved upon. Hence, this thesis aims to broaden the general understanding of CGP and to improve its performance by introducing new operators and extensions.
All findings of this dissertation are of empirical nature, which is why a fast and reliable framework is of utmost importance. Thus, the first contribution of this thesis is a new CGP specific toolbox written in the Rust programming language. This framework is highly modular, which enables the simple implementation of different configurations and combinations of operators and/or extensions. In addition, its fast execution time allows for the extensive study of numerous training runs in a short amount of time.
The second contribution of this thesis is the further development of an existing extension called Reorder. CGP suffers from a bias called positional bias. It occurs due to its specific representation of solutions, resulting in potential performance degradations. The original extension should remove this bias; however, this work finds a flaw in Reorder. It is able to mitigate positional bias but is not able to completely eliminate it. By introducing several extensions adjacent to Reorder in this dissertation, this bias can be completely removed. There is also the option to purposefully introduce new biases, which might aid CGP evolutionary search process. Both options have the potential to improve CGP performance, resulting in faster convergence times and/or better fitness values.
Afterwards, two new mutation operators for CGP are introduced, improving CGP performance on real-world datasets. In addition, new insights can be gained by closely examining the outcomes of both operators.
The following chapter introduces weights to influence CGP mutation mechanism. This allows for small performance gains and mitigates the aforementioned positional bias. Further insights into CGP behaviour and workings are discussed. The most important finding is: Positional bias is lessened but this does not automatically lead to an increased performance. This goes against the current understanding of CGP, and more research should be done in this direction.
Crossover is an important evolutionary operator for Evolutionary Algorithms. CGP, however, generally does not use standard crossover operators as it generally leads to no improvements or even worse performances. The next chapter investigates if positional bias is at fault for this phenomenon. While this hypothesis could not be verified, the chapter still has two major findings. At first, CGP combined with tournament selection degrades the performance on Boolean benchmarks. Secondly, if correctly configured and tuned, CGP combined with crossover converges faster and/or improves its final fitness values. Especially the last point goes against the general research trend. Potential reasons regarding this opposite conclusion are also given.
The last contribution of this dissertation discusses CGP robustness against redundant attributes in datasets. Especially real-world datasets might contain attributes that are of redundant nature or even regarded as noise. Data preprocessing and data mining techniques are normally used to remove most of these unwanted attributes. However, some redundant attributes might not be found, or this process might lead to subpar results of the learning algorithm. By design, CGP is able to choose its own input attributes. Through evolutionary mechanisms, CGP should be able to ignore redundant attributes. The last chapter of this thesis confirms this hypothesis for different types and quantities of artificially inserted redundant attributes.",
-
notes = "In english
Also know as \cite{Cui2026}
https://nbn-resolving.org/urn:nbn:de:bvb:384-opus4-1267125
Dewey Decimal Classification: 0 Informatik, Informationswissenschaft, allgemeine Werke / 00 Informatik, Wissen, Systeme / 004 Datenverarbeitung; Informatik
Supervisor: Joerg Haehner",
