- author = "Dirk Schweim",
- title = "On Sampling Error and Bias in Evolutionary Algorithms",
- school = "Johannes Gutenberg-Universitaet Mainz",
- year = "2021",
- address = "Germany",
- month = "17 " # mar,
- keywords = "genetic algorithms, genetic programming, grammatical evolution, EDA, ANN, informatik",
- timestamp = "Fri, 16 Jul 2021 22:40:55 +0200",
-
biburl = "
https://dblp.org/rec/phd/dnb/Schweim21.bib",
- bibsource = "dblp computer science bibliography, https://dblp.org",
-
URL = "https://d-nb.info/1235777952",
-
URL = "https://portal.dnb.de/opac.htm?method=simpleSearch&cqlMode=true&query=idn%3D1235777952",
-
URL = "https://www.deutsche-digitale-bibliothek.de/item/YYC7M4ODKVH2F2MENMII3PVY6I3ON3SQ",
-
URL = "https://dblp.org/rec/phd/dnb/Schweim21.html?view=bibtex",
-
URL = "http://www.worldcat.org/title/1246629572",
- size = "176 pages",
- abstract = "When using a sample, there are usually differences between the properties of the statistical population and the information obtained from the sample. These differences are called errors. There are two main types of errors, sampling and non-sampling errors. Sampling errors are non-systematic errors, describing random variations caused by observing only a subset of the statistical population. The expected amount of sampling error can be reduced by using larger samples. On the other hand, non-sampling errors or biases are non-random, systematic errors. In optimization with evolutionary algorithms (EAs), both types of errors can be observed. This dissertation contributes to a better understanding of EAs by studying sampling error and bias in three EAs, genetic programming (GP), grammatical evolution (GE), and estimation of distribution programming (EDP). The thesis consists of five research articles. The first article studies how the initial GP population size affects the sampling error of subtree frequencies. A model of the expected frequencies of subtrees for GP populations initialized with full, grow, and ramped half-and-half is introduced. Based on the frequency model, another model is introduced that estimates the GP population size, given a threshold of sampling error and a confidence level. Then, both models are empirically validated and the models implications are discussed. The second and third paper investigate the impact of grammar design in GE during initialization and crossover. The analyses focus on the expected number of non-terminals chosen when mapping genotypes to phenotypes. The results show that differences in the expected number of non-terminals impact both, the redundancy of solutions as well as the number of invalid solutions that are created. The fourth paper studies bias of EDPs variation operator. Results of random walks indicate that the complexity of the EDP model is high since the model is overfitting the parent solutions. To prevent overfitting and ensure a minimum amount of variation in a population, EDP adds random noise to the learned probabilities. The results show that adding only a low amount of noise leads to a strong bias towards small trees. The bias gets stronger with an increased amount of noise. The biases of variation operators in GP and EDP are compared. The results show that the bias in GP is lower. Last, a new EDA-GP variant—called DAE-GP, is proposed that uses denoising autoencoder long short-term memory networks (DAE-LSTMs) as probabilistic model. DAE-LSTMs are artificial neural networks that first learn the properties of a parent population by mapping promising candidate solutions to a latent space and reconstructing the candidate solutions from the latent space. The trained model is then used to sample new offspring solutions. The model does not impose any assumptions about the relationships between learned variables and therefore, the expected bias is reduced compared to other EDA-GP models. As a consequence, preliminary results show that DAE-GP outperforms GP on a problem with high locality.",
-
notes = "in english
List of Papers
Chapter 2 Schweim, Dirk1; Wittenberg, David1; Rothlauf, Franz1 (2020). On Sampling Error in Genetic Programming. In: Natural Computing Journal. Springer. \cite{Schweim:NatComput}
Chapter 3 Schweim, Dirk1; Thorhauer, Ann1; Rothlauf, Franz1 (2018). On the Non-Uniform Redundancy of Representations for Grammatical Evolution: The Influence of Grammars. In: Handbook of Grammatical Evolution. Springer. 2018. \cite{Schweim:2018:hbge}
Chapter 4 Schweim, Dirk1 (2020). On the Influence of Grammars on Crossover in Grammatical Evolution. 24th European Conference on Genetic Programming. \cite{Schweim:2021:EuroGP}
Chapter 5 Schweim, Dirk1; Rothlauf, Franz1 (2018). An Analysis of the Bias of Variation Operators of Estimation of Distribution Programming. In: Proceedings of the 2018 Genetic and Evolutionary Computation Conference (GECCO 2018), Kyoto, Japan, July, 15-19, 2018. ACM 2018. \cite{Schweim:2018:GECCO}
Chapter 6 Wittenberg, David1; Rothlauf, Franz1; Schweim, Dirk1 (2020). DAE-GP: Denoising Autoencoder LSTM Networks as Probabilistic Models in Estimation of Distribution Genetic Programming. In: Proceedings of the 2020 Genetic and Evolutionary Computation Conference (GECCO 2020), Cancun, Mexico, July, 8-12, 2020. ACM 2020. \cite{Wittenberg:2020:GECCO}",
