Implementing Multi-Task Learning for Bayesian Optimization Structure Search

No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
School of Science | S harjoitus- ja seminaarityöt
Date
2020
Major/Subject
Monimutkaiset Systeemit
Complex Systems
Mcode
SCI3060
Degree programme
Language
en
Pages
30 + app. 34
Series
Abstract
Machine learning algorithms are highly dependent on the quality of the training data. Problems with the data, like accuracy and expenses in gathering it project directly to the cost and performance of the algorithm. Multi-task learning methods try to overcome this problem by combining information from multiple sources of data. In this study I focused on two existing methods for multi-output Gaussian processes, linear model of coregionalisation and intrinsic coregionalisation model. My aim in this study was to implement these algorithms to Bayesian Optimization Structure Search (BOSS) tool and test their potential for accelerating atomistic structure search space navigation with a simulation study. I found potential for reducing computational cost of optimization of expensive structure searches with multi-task learning.
Description
Supervisor
Rinke, Patrick, Associate Professor, Aalto University, Department of Applied Physics, Finland
Thesis advisor
Todorovic, Milica, Research Fellow, Aalto University, Department of Applied Physics, Finland
Remes, Ulpu, Postdoctoral Researcher, University of Helsinki, Department of Mathematics and Statistics, Finland
Keywords
Bayesian Optimization Structure Search, Gaussian Processes, Multi-task Learning, Coregionalization,Linear Model of Coregionalization, Electronic Structure Theory, Machine Learning for Materials Science, Transfer Learning, Aineen rakenne, Bayesilainen optimointi, Monilähteinen koneoppiminen, Gaussin prosessi, Siirtolähteinen koneoppiminen, Laskennallinen materiaalitutkimus
Other note
Citation