Implementacija igralca Backgammona z nevronsko mrežo

• Main Author: Gajski, Primož
• Other Authors: Šter, Branko
• Format: Bachelor Thesis
• Language: Slovenian
• Published: P. Gajski 2015
• Subjects: Backgammon; nevronska mreža; perceptron; spodbujevano učenje; algoritem vzvratnega učenja; računalništvo; univerzitetni študij; diplomske naloge; neural network; reinforcement learning; backpropagation algorithm; computer science; diploma; info:eu-repo/classification/udc/004.032.26(043.2); sami
• Online Access: https://repozitorij.uni-lj.si/IzpisGradiva.php?id=69596; https://repozitorij.uni-lj.si/Dokument.php?id=69607&dn=; http://www.cobiss.si/scripts/cobiss?command=DISPLAY&base=cobib&rid=7671124&fmt=11

Diplomsko delo Implementacija igralca Backgammona z nevronsko mrežo opisuje način in vsebuje implementacijo, kako se je računalnik zmožen naučiti igre backgammon z igranjem sam s sabo, torej brez človeškega faktorja in brez kakršnega koli vnaprejšnjega znanja o sami igri. Ocenjevanje pozicij za izbor najugodnejše poteze je narejeno z evaluacijsko funkcijo, ki je realizirana z nevronsko mrežo v obliki 2-nivojskega perceptrona. Učenje nevronske mreže poteka s spodbujevanim učenjem, ki je doseženo z algoritmom vzvratnega učenja. Mreža tako odigra veliko število iger (več milijonov), da osvoji zahtevnejši nivo igre backgammon. Testiranje je opravljeno tudi brez človeškega faktorja, in sicer je potekalo kot igranje proti t.i. igralcu pubeval, t.j. še ena evaluacijska funkcija, ki tudi ocenjuje pozicije žetonov. Po zaključenem učenju je nevronska mreža uspešno osvojila znanje ocenjevanja backgammon pozicije in obvlada tako osnovne kot tudi zahtevnejše elemente igre. Za konec je bilo potrebno realizirati uporabniški grafični vmesnik, ki uporabniku omogoča igranje proti naučeni nevronski mreži. Diploma thesis Implementation of Backgammon player with neural network describes implementation of how is a computer capable to learn the game of backgammon. It is achieved with computer playing against itself, without any human help and without any prior knowledge about the game. Evaluation function of game positions, according to the thrown dice and checker positions, was implemented with neural network. Neural network is presented as a 2-layer perceptron. The learning of neural network was achieved with reinforcement learning and backpropagation algorithm. Within backpropagation algorithm neural network plays several milions of backgammon games to achieve the advanced level of the game. Testing was also performed without any human interaction. Neural network played backgammon game against so called pubeval player, who also uses for its logic evaluation function to play on strong intermediate level. The testing showed that the neural network successfully accomplished all basic and also many of advanced features of game playing elements. Diploma work also involves realization of graphical user interface, which allows user to play backgammon against neural network.