Vis enkel innførsel

dc.contributor.authorMagelssen, Kristoffer Lomeland
dc.contributor.authorStorebø, Trond
dc.date.accessioned2020-10-19T08:01:16Z
dc.date.available2020-10-19T08:01:16Z
dc.date.issued2020
dc.identifier.citationMagelssen, K. L. & Storebø, T. (2020) TsetlinGo : Solving the game of Go with Tsetlin Machineen_US
dc.identifier.urihttps://hdl.handle.net/11250/2683514
dc.descriptionMaster's thesis in Information- and communication technology (IKT590)en_US
dc.description.abstractThe Tsetlin Machine have already shown great promise on pattern recognition and text categorization. The board game GO is a highly complex game, and the Tsetlin Machine have not yet been tested extensively on strategic games like this. This thesis introduces TsetlinGO and aims to Solve the game of Go with Tsetlin Machine. For predicting the next moves a combination of Tsetlin Machine and Tree Search was used. In the thesis a 9x9 board size was used for the game of Go, to prevent the problem from becoming to complex. This thesis goes through hyper-parameter testing for classification of the Go board game. A solution with Tree Search and Tsetlin Machine combined is used to perform self-play and matches between Tsetlin Machines with different hyper-parameters. Based on the empirical results, our conclusion is that the Tsetlin Machine is more than capable for classification of the game of Go at various stages of play. Results from the experiments could be seen to achieve around 90%, while further climbing up to around 95% upon re-training. From examining the clauses, strong patterns was found that gave insight into how the machine works. The Tsetlin Machine was able to play complete games of Go, making connections on the board through use of patterns from the clauses. It was found that the size of the clauses had great impact as clauses with large patterns had trouble getting triggered in early play. The high accuracy from classification was found to not correlate with how strong the Tsetlin Machine would perform during self-play. This may indicate that producing training data directly from self-play may be required to fine tune the assessment of board positions faced during actual play. We can conclude that this thesis provide a benchmark for further research within the field of Tsetlin Machine and the game of Go.en_US
dc.language.isoengen_US
dc.publisherUniversity of Agderen_US
dc.rightsAttribution-NonCommercial-NoDerivatives 4.0 Internasjonal*
dc.rights.urihttp://creativecommons.org/licenses/by-nc-nd/4.0/deed.no*
dc.subjectIKT590en_US
dc.titleTsetlinGo : Solving the game of Go with Tsetlin Machineen_US
dc.typeMaster thesisen_US
dc.rights.holder© 2020 Kristoffer Lomeland Magelssen, Trond Storebøen_US
dc.subject.nsiVDP::Matematikk og Naturvitenskap: 400::Informasjons- og kommunikasjonsvitenskap: 420::Kunnskapsbaserte systemer: 425en_US
dc.source.pagenumber79en_US


Tilhørende fil(er)

Thumbnail

Denne innførselen finnes i følgende samling(er)

Vis enkel innførsel

Attribution-NonCommercial-NoDerivatives 4.0 Internasjonal
Med mindre annet er angitt, så er denne innførselen lisensiert som Attribution-NonCommercial-NoDerivatives 4.0 Internasjonal