Smart Dog for Minecraft
Authors
Thomsen, Kim Arnold ; Dalhoff-Jensen, Rasmus Ditlev Cornelius
Term
4. term
Education
Publication year
2014
Submitted on
2014-06-10
Pages
109
Abstract
This thesis explores how tabular reinforcement learning can be combined with feature‑based reinforcement learning to give an agent both situation‑specific behavior and generalization in complex environments. The work proposes three ways to link the two learning paradigms (including different Q‑value update strategies) and implements them in a Minecraft mod as a “Smart Dog.” The system is modular, with sensing, decision making, and actuation running in a virtual real‑time loop, and employs Q‑learning techniques (e.g., ϵ‑greedy action selection, tabular state attributes, and feature weights). Through a set of controlled in‑game tests—including scenarios with items such as potions and food and experiments on knowledge transfer between situations—the three combination methods are evaluated. The results indicate that two of the three approaches provide a measurable benefit in this setting, supporting the claim that a combined strategy can deliver precise situational behavior while handling large state spaces in game‑based environments. Limitations, detailed metrics, and full results are beyond this excerpt, but the report covers theory, implementation, testing, and evaluation.
Denne afhandling undersøger, hvordan tabulær forstærkningslæring kan kombineres med feature‑baseret forstærkningslæring for at give en agent både situationsspecifik adfærd og generalisering i komplekse miljøer. Projektet formulerer tre tilgange til at koble de to læringsformer (herunder forskellige opdateringsstrategier for Q‑værdier) og implementerer dem i et Minecraft‑mod som en “Smart Dog”. Systemet er modulært med sansning, beslutningstagning og aktivering, og det arbejder i et virtuelt realtidsforløb med Q‑læring (bl.a. ϵ‑greedy valg, tabulære tilstande og feature‑vægte). Gennem en række kontrollerede tests i spillet, herunder scenarier med genstande som potions og mad samt forsøg med vidensoverførsel mellem situationer, vurderes de tre kombinationsmetoder. Resultaterne peger på, at to af de tre tilgange giver en målbar fordel i dette scenarie, hvilket understøtter, at en kombineret strategi kan forene præcis situationsadfærd med håndtering af store tilstandsrum i spilbaserede miljøer. Begrænsninger, detaljerede målinger og fulde resultater ligger uden for uddraget, men rapportens struktur dækker teori, implementering, test og evaluering.
[This apstract has been generated with the help of AI directly from the project full text]
Keywords