On Search Space Constraining Methods for Automatic Composition and Optimisation of Machine Learning Pipelines

Created by W.Langdon from gp-bibliography.bib Revision:1.8564

@PhdThesis{Tien-Dung_Nguyen:thesis,

• author = "Tien Dung Nguyen",
• title = "On Search Space Constraining Methods for Automatic Composition and Optimisation of Machine Learning Pipelines",
• school = "Faculty of Engineering and Information Technology, University of Technology Sydney",
• year = "2023",
• address = "Australia",
• month = mar,
• keywords = "genetic algorithms, genetic programming, TPOT",
• URL = "https://opus.lib.uts.edu.au/handle/10453/171452",
• URL = "http://hdl.handle.net/10453/171452",
• URL = "https://opus.lib.uts.edu.au/bitstream/10453/171452/1/thesis.pdf",
• size = "326 pages",
• abstract = "Automated machine learning (AutoML) has been developed and studied to automate the process of collecting, preprocessing and integrating data, composing and optimising ML pipelines, and deploying and maintaining predictive models. A machine learning (ML) pipeline is a work o w consisting of many components to perform data transformation, data preprocessing, feature engineering and classification/regression tasks. ML pipelines have been used to build predictive models for a variety of ML problems. One of the most important research topics of AutoML is ML pipeline composition and optimisation (PCO). PCO processes search for valid and well-performing ML pipelines in a given search space. A search space consists of ML components, i.e. preprocessing and predictor/meta-predictor components, components hyperparameters and pipeline structures that link ML components. Due to the large size of search spaces, it is challenging for PCO processes to find valid and well-performing pipelines. The aim of the dissertation is to study methods/strategies to constrain a search space to enable PCO processes to more easily and efficiently find valid and well-performing pipelines. This dissertation has three main contributions. The first contribution is a novel method, AVATAR, to constrain search space to consider only valid ML pipelines. The AVATAR eliminates invalid pipelines by assessing ML pipeline validity using a surrogate model, a Petri-net approach based on consideration of simplified pipelines not requiring the use of training data or its processing. The second contribution is a critical evaluation of the so called opportunistic meta-knowledge based on previous experience in the form of PCO runs and predictor evaluations with default hyperparameters. Although such meta-knowledge is inherently noisy and with high statistical variability, we found that it is still very useful for constraining search spaces with promising, well-performing ML components. The third contribution is an exploration of different search space reduction strategies employing that meta-knowledge. The results show that the reduction of search space should not be extreme. In addition, reducing the search space based on the knowledge of the problem itself enables PCO processes to deliver the best-performing ML pipelines, followed by the knowledge of strong general performers. Finally, the results obtained in this study can form the basis of and be extended in the future to dynamically control search spaces to find better ML pipelines by combining the knowledge of the ML problems solved in the past and the knowledge of the problem itself acquired from the run of the PCO processes.",
• notes = "Supervisors: Bogdan Gabrys and Katarzyna Musial",

}

Genetic Programming entries for Tien-Dung Nguyen

Citations