- author = "Jan Zegklitz",
- title = "Enhancements of Genetic Programming-based Symbolic Regression Algorithms",
- school = "Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University in Prague",
- year = "2021",
- address = "Prague, Czech Republic",
- month = jul,
- keywords = "genetic algorithms, genetic programming, symbolic regression, linear regression, machine learning, feature construction, fitness predictors",
-
URL = "
http://hdl.handle.net/10467/96767",
-
URL = "https://dspace.cvut.cz/entities/publication/c313c932-5b70-4353-b425-bde45541d0d8",
-
URL = "https://dspace.cvut.cz/bitstreams/87c8bcbd-1b7e-4d61-93a8-d2773365f383/download",
- size = "148 pages",
-
abstract = "Symbolic regression (SR) is a kind of regression task
where the goal is to find a model in the form of a
mathematical expression that is as small as possible
and, preferably, human-readable. SR methods are mostly
based on genetic programming (GP).
we focus on improving the GP-based SR algorithms. We conduct a benchmarking study that compares a number of recent SR algorithms on a common set of benchmarks. The results indicate that the SR methods are not superior to the classic machine learning methods, but there is a room for improvement.
We propose a new type of node for GP-based SR algorithms which allows for encoding affine transformations of feature space. The results show that these nodes improve the performance of the algorithm.
We introduce the task of dynamic SR which is closely connected to reinforcement learning. We show the good applicability of the proposed new type of node for this kind of task, as the tunable transformations allow for fast target tracking.
Lastly, we focus on the topic of fitness prediction based on subsampling the training dataset. We propose modifications of previous methods which significantly simplify the algorithms and decrease the number of parameters that need to be set. The results indicate, that these simplifications are a viable alternative to the commonly used coevolutionary approach.",
-
notes = "In english
F3
Ph.D. programme: Electrical Engineering and Information Technology (P2612) Branch of study: Artificial Intelligence and Biocybernetics (3902V035)
Supervisor: Petr Posik",
