Computational Intelligence Methods in Metalearning

author = "Jakub Smid",

title = "Computational Intelligence Methods in Metalearning",

school = "Faculty of Mathematics and Physics, Charles University in Prague",

year = "2016",

address = "Czech republic",

keywords = "genetic algorithms, genetic programming, Metalearning, Machine Learning, Metric, Attribute Assignment",

URL = "

http://hdl.handle.net/20.500.11956/82405",

URL = "

https://dspace.cuni.cz/handle/20.500.11956/82405",

URL = "

https://dspace.cuni.cz/bitstream/handle/20.500.11956/82405/IPTX_2011_2_11320_0_394065_0_123234.pdf",

size = "158 pages",

abstract = "This thesis focuses on the algorithm selection problem, in which the goal is to recommend machine learning algorithms to a new dataset. The idea behind solving this issue is that algorithm performs similarly on similar datasets. The usual approach is to base the similarity measure on the fixed vector of meta-features extracted out of each dataset. However, as the number of attributes among datasets varies, we may be loosing important information. Herein, we propose a family of algorithms able to handle even the non-propositional representations of datasets. Our methods use the idea of attribute assignment that builds the distance measure between datasets as a sum of distance given by the optimal assignment and an attribute distance measure. Furthermore, we prove that under certain conditions, we can guarantee the resulting dataset distance to be a metric. We carry out a series of meta-learning experiments on the data extracted from the OpenML repository. We build up attribute distance using Genetic Algorithms, Genetic Programming and several regularization techniques such as multi-objectivization, coevolution, and bootstrapping. The experiment indicates that the resulting dataset distance can be successfully applied on the algorithm selection problem. Although we use the proposed distance measures exclusively...",

abstract = "Tato prace je zamerena na problematiku vyberu algoritmu, ktera ma za cil doporucit algoritmus strojoveho uceni k nove uloze. Reseni problemu vychazi z myslenky, ze se algoritmy chovaji podobne na podobnych datech. Tato podobnost je casto zalozena na extrakci pevneho poctu metaatributů z kazde ulohy. Vzhledem k tomu, ze pocet atributů se u různych uloh typicky lisi, ztracime tak důlezite informace. V teto praci popiseme tridu algoritmů, ktera dokaze zpracovat take informace o jednotlivych atributech. Nase metody jsou zalozeny na prirazovani atributů. Vysledna vzdalenost mezi ulohami je dana jako soucet vzdalenosti mezi atributy urcenymi optimalnim prirazenim. Dale dokazeme, ze za urcitych podminek můzeme zarucit, ze vysledna vzdalenost mezi ulohami je metrika. Provedeme sadu experimentů na datech extrahovanych z OpenML repozitare. Vytvorime vzdalenost mezi atributy prostrednictvim genetickych algoritmů, genetickeho programovani a nekolika regularizacnich technik, jako je koevoluce a zavedeni vicekriteriality. Vysledky experimentů naznacuji, ze vysledna vzdalenost mezi ulohami můze byt uspesne pouzita na problematiku vyberu algoritmu. Ackoliv jsme nase metody pouzili vyhradne k metauceni, lze je aplikovat i v jinych oblastech. Navrzene algoritmy jsou aplikovatelne kdekoliv, kde mame definovanou vzdalenost...",

notes = "Supervisor: Roman Neruda",