- author = "Elena Bautu",
- title = "Intelligent Techniques for Data Modeling Problems",
- school = "Al. I. Cuza University",
- year = "2010",
- address = "Iasi, Romania",
- month = jun,
- note = "Romanian subtitle is Programare genetica pentru probleme de optimizare in Inteligenta artificiala",
- keywords = "genetic algorithms, genetic programming, gene expression programming, inverse problems, financial forecasting, data analysis, hypernetwork, hybridization",
-
URL = "
https://sites.google.com/site/ebautu/home/publications/thesis/thesis_elena_bautu.pdf",
- broken = "https://sites.google.com/site/ebautu/home/publications/thesis",
- size = "220 pages",
-
abstract = "Supervised learning deals with the problem of
discovering models from data as relationships between
input and output attributes. Two types of models are
distinguished: regression models for continuous output
and classification models (classiffiers) for discrete
output. This thesis addresses both regression and
classiffication problems, with an emphasis on new
applications and on proposing new evolutionary
techniques.
First, we address the regression domain. Symbolic regression by means of evolutionary techniques is recommended when there is little or no a priori information on the modelled process. It relies on a set of input-output observations to infer mathematical models, posing no constraints on the structure, the coefficients or the size of the model. We introduce inverse problems modeled by Fredholm integral equations of the first kind and the inverse problem of log synthesis to be modelled by symbolic regression by means of gene expression programming. A new genetic programming scheme is formulated for the problem of automatically designing quantum circuits. An adaptive version of the gene expression programming algorithm is presented, which automatically tunes the complexity of the model by a gene (de)activation mechanism. For modelling time series produced by dynamic processes, we propose an evolutionary approach that uses a novel representation (and suitable genetic operators) to partition the time series based on change points. Empirical results prove the approach to be promising.
Research on building classifiers for a given problem is also extensive, since there exists no best classifier at all tasks. The problem of predicting the direction of change of stock price on the market can be formulated as the search for a classifier that links past evolution to an increase or decrease. We explore new techniques for classification, in the context of predicting the direction of change of stock price, formulated as a binary classification",
- notes = "See also \cite{Bautu:book}",
