Reinforcement learning example using a variant of the Dodgeball game

Rad opisuje osnove podržanog učenja i Unityev alat ML-Agents koji omogućava provođenje algoritama strojnog učenja u igrama razvijenim na platformi Unity. Analizirana je igra izbjegavanja lopti (engl. Dodgeball) i komponente koje se koriste za njezinu realizaciju. Proučen je način na koji se svaka scena u Unityju može pretvoriti u okolinu za učenje i način na koji se naučena ponašanja mogu koristiti za upravljanje likova u sceni. S Unity web-sjedišta je preuzeta igra Dodgeball i na temelju nje, izmjenom uvjeta treniranja i dodavanjem novih komponenti, napravljena je vlastita inačica igre. Demonstrirani su rezultati i napravljena je analiza uspješnosti istreniranog ponašanja agenata.The B. Sc. thesis describes the fundamentals of reinforcement learning and Unity's ML-Agents toolkit, which provides an implementation of reinforcement learning algorithms in games developed in Unity. The „Dodgeball“ game and the components that are used in its realization have been analyzed. The ways of converting a scene in Unity into a learning environment and using trained behaviors to control the game characters have been studied. The „Dodgeball“ game has been downloaded from Unity's official web page and modified by changing the training conditions and adding new components in order to create a new version of the game. The results from a modified version of the game have been demonstrated and analyses the behavior of trained agents