FighterZero

pristop samo-igranja za učenje igranja pretepaške igre z globokim spodbujevalnim učenjem

Matej Vitek (Author), Peter Peer (Mentor)

Abstract

Področje globokega učenja je v zadnjem desetletju doživelo precejšen razcvet. Uporablja se za reševanje premnogih problemov, v zadnjih petih letih pa precej tudi za igranje iger. Dva pomembna dosežka sta bila globoke Q-mreže (DQN) in AlphaZero. DQN se je naučila igrati klasične igre za Atari 2600 (Pong, Space Invaders, itd.), AlphaZero pa se je s samo-igranjem naučil igrati šah, šogi in Go. Mi smo na temelju AlphaZero poskusili zgraditi agenta FighterZero, ki bi se prav tako s samo-igranjem naučil igrati pretepaške računalniške igre. Rezultati so bili manj uspešni, kot smo pričakovali, saj se je časovna zahtevnost izkazala za nepremagljivo oviro.

Keywords

umetna intiligenca;inteligentni agent;igre;samo-igranje;globoko učenje;spodbujevalno učenje;drevesno preiskovanje Monte Carlo;nevronske mreže;razvoj iger;

Data

Language:	Slovenian
Year of publishing:	2018
Typology:	2.09 - Master's Thesis
Organization:	UL FRI - Faculty of Computer and Information Science
Publisher:	[M. Vitek]
UDC:	004
COBISS:	18432089
Views:	1506
Downloads:	343
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	FighterZero: a self-playing deep reinforcement learning agent for fighting game AI
Secondary abstract:	Deep learning has been a field of great academic interest and substantial breakthroughs over the last decade. Its applications are many and over the last five years it has spread also to the field of game playing, owing largely to two chief accomplishments of Google's DeepMind team: Deep Q-Networks (DQN), which learned to play classic Atari 2600 games, and AlphaZero, which learned, strictly through self-play, to play the board games chess, shogi and Go. In this thesis we attempted to build on the success of AlphaZero by adapting its self-playing architecture to fighting games, a popular genre of video games. The results were, however, less successful than we had expected and hoped, as the time constraints proved to be an insurmountable obstacle.
Secondary keywords:	artificial intelligence;intelligent agent;games;self-playing;deep learning;reinforcement learning;Monte Carlo tree search;neural networks;game development;
Type (COBISS):	Master's thesis/paper
Study programme:	0
Thesis comment:	Univ. v Ljubljani, Fak. za matematiko in fiziko, Oddelek za matematiko, Računalništvo in matematika - 2. stopnja
Pages:	XV, 36 str.
ID:	10959377