Return to search

Exploring the Effect of Different Numbers of Convolutional Filters and Training Loops on the Performance of AlphaZero

In this work, the algorithm used by AlphaZero is adapted for dots and boxes, a two-player game. This algorithm is explored using different numbers of convolutional filters and training loops, in order to better understand the effect these parameters have on the learning of the player. Different board sizes are also tested to compare these parameters in relation to game complexity. AlphaZero originated as a Go player using an algorithm which combines Monte Carlo tree search and convolutional neural networks. This novel approach, integrating a reinforcement learning method previously applied to Go (MCTS) with a supervised learning method (neural networks) led to a player which beat all its competitors.

Identiferoai:union.ndltd.org:WKU/oai:digitalcommons.wku.edu:theses-4090
Date01 October 2018
CreatorsPrince, Jared
PublisherTopSCHOLAR®
Source SetsWestern Kentucky University Theses
Detected LanguageEnglish
Typetext
Formatapplication/pdf
SourceMasters Theses & Specialist Projects

Page generated in 0.0018 seconds