- author = "David Wittenberg and Franz Rothlauf",
- title = "Using denoising autoencoder LSTM networks to balance exploration and exploitation in estimation of distribution genetic programming",
- booktitle = "31st European Conference On Operational Research",
- year = "2021",
- address = "Athens",
- month = "11-14 " # jul,
- keywords = "genetic algorithms, genetic programming, ANN, Machine Learning, Combinatorial Optimization, Metaheuristics",
-
URL = "
https://www.euro-online.org/conf/euro31/treat_abstract?paperid=3310",
-
URL = "
https://liser.elsevierpure.com/ws/portalfiles/portal/32968251/abstract_book_euro31.pdf",
- abstract = "Estimation of distribution genetic programming (EDA-GP) algorithms are metaheuristics for variable-length combinatorial optimization problems that replace the standard recombination and mutation operators of genetic programming (GP) by sampling from a learned probabilistic model. An example of an EDA-GP is DAE-GP that uses denoising autoencoder long short-term memory networks as probabilistic model. DAE-GP is the first and only EDA-GP that uses neural networks as a model and outperforms standard GP. The key advantage of DAE-GP is that we can flexibly identify relevant relationships between problem variables and that we can apply denoising on input candidate solutions to control the generalization behavior of the model. However, current work only uses subtree mutation with fixed corruption strength. In this work, we therefore study alternative denoising strategies. We show on standard GP benchmark problems that denoising strongly influences the exploration and exploitation behavior in search. Adjusting the denoising strategy can therefore help to either exploit promising areas of the parent population or to explore new search spaces.",
-
notes = "https://euro2021athens.com/
Lehrstuhl fuer Wirtschaftsinformatik und BWL, Universitaet Mainz",
