diplomsko delo
Deni Cerovac (Author), Aleksander Sadikov (Mentor)

Abstract

Diplomsko delo Spodbujevalno učenje na problemu igre Pacman prikazuje mojo motivacijo, ki me je pritegnila k izbiri takšnega projekta. V diplomskem delu smo opravili teoretičen pregled delovanja algoritmov na principu spodbujevalnega učenja, kjer smo tudi pregledali teoretično ozadje algoritmov Q-učenje in globoko Q-učenje, katera smo tudi implementirali in uporabili na igri Pacman. Naš pristop je bil poseben zaradi primerjave uspešnosti algoritmov v okolju, kjer imata algoritma zelo omejeno gibanje in vpliv na delovanje samega akterja. Glede na pridobljeno znanje smo v celoti projekta implementirali oba algoritma, ki sta se učila na dani igri in nam vrnila zanimive rezultate. Med implementacijo smo doživeli veliko izzivov, nekatere zabavne, nekatere pa ne, katere smo uspešno premagali. Na podlagi pridobljenih rezultatov smo ugotovili, da sta se kljub omejenem gibanju in vplivanju na karakterjevo obnašanje algoritma odrezala podobno in v nekaterih primerih bistveno boljše kot amaterski igralci igre Pacman.

Keywords

nevronske mreže;Q-učenje;globoko Q-učenje;strojno učenje;spodbujevalno učenje;računalništvo in informatika;univerzitetni študij;diplomske naloge;

Data

Language: Slovenian
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [D. Cerovac]
UDC: 004.8(043.2)
COBISS: 30969091 Link will open in a new window
Views: 860
Downloads: 155
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Pacman implementation using reinforcement learning
Secondary abstract: Bachelors project Pacman implementation using reinforcement learning, shows the reason and motivation that made me choose this project. In bachelors project we went over theoretical principals of Reinforcement learning algorithms, where we explained theoretical background of Q-learning and Deep Q-learning which we implemented and used on a game Pacman. Our approach was special because of our comparison of success between these two algorithms which were implemented on a game with restricted ability to impact on the movement and decision of our agent. Based on the accumulated knowledge in the course of our project we implemented both algorithms, that when finished returned some interesting results. Throughout our implementation we experienced a lot of challenges, some more fun than others, but in the end we successfully resolved all of them. Based on gathered results we found out that despite restricted movement of our agent, the algorithms were in average approximately as good or in some cases drastically better than average amateur Pacman players.
Secondary keywords: neural networks;Q-learning;deep Q-learning;machine learning;reinforcement learning;computer and information science;diploma thesis;
Type (COBISS): Bachelor thesis/paper
Study programme: 1000468
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 35 str.
ID: 12033197