Show simple item record

dc.contributor.authorFraser, Warren Duncan
dc.date.accessioned2017-12-21T23:27:50Z
dc.date.available2017-12-21T23:27:50Z
dc.date.issued2016-05
dc.identifier.urihttp://hdl.handle.net/11122/8044
dc.descriptionMaster's Project (M.S.) University of Alaska Fairbanks, 2016en_US
dc.description.abstractThis report covers scaling neural networks for training Go artificial intelligence. The Go board is broken up into subsections, allowing for each subsection to be calculated independently, and then factored into an overall board evaluation. This modular approach allows for subsection networks to be translated to larger board evaluations, retaining knowledge gained. The methodology covered shows promise for significant reduction in training times required for unsupervised training of Go AI. A brief history of artificial neural networks and an overview of Go and the specific rules that were used in this project are presented. Experiment design and results are presented, showing a promising proof of concept for reducing training time required for evolutionary Go AI. The codebase for the project is Apache 2.0 licensed and is available on GitHub.en_US
dc.language.isoen_USen_US
dc.subjectGo (Game)en_US
dc.subjectArtificial intelligenceen_US
dc.titleGo artificial intelligence: a scalable evolutionary approachen_US
dc.typeOtheren_US
dc.type.degreems
dc.identifier.departmentDepartment of Computer Science
dc.contributor.chairHay, Brian
dc.contributor.committeeLawlor, Orion
dc.contributor.committeeGenetti, Jon
refterms.dateFOA2020-03-05T14:55:18Z


Files in this item

Thumbnail
Name:
Fraser_W_2016.pdf
Size:
8.414Mb
Format:
PDF

This item appears in the following Collection(s)

Show simple item record