Solving a platformer game through machine learning

Ovim radom opisane su korištene tehnologije i objašnjeni su elementi strojnog učenja koji su korišteni za treniranje agenta. Cilj je istrenirati agenta za rješavanje razina platformske igre koristeći ML-Agents alat. Izrađene su dvije okoline u Unityju za treniranje agenta, jedna jednostavna i jedna napredna. Jednostavna okolina je korištena za usporedbu PPO i SAC algoritama koji se koriste za podržano strojno učenje. Na temelju te usporedbe algoritama, efikasniji, PPO algoritam korišten je za treniranje agenta u naprednoj okolini. Uspješnost treniranja ispitana je na tri razine platformske igre. Rezultat ovog rada je uspješno istrenirani agent koji na temelju naučenog dolazi do cilja razine izbjegavajući zamke. Platformska igra se sastoji od tri razine na kojima se igrač može suprotstaviti agentu u natjecanju za pobjedu i jedne razine koja je korištena kao napredna okolina za treniranje, a sada na njoj igrač sam može trenirati.This work describes the technologies used and explains the machine learning elements used for agent training. The goal is to train an agent to solve a platform game level using the ML-Agents Toolkit. Two environments were created in Unity for training an agent, one simple and one advanced. A simple environment is used to compare PPO and SAC algorithms used for reinforcement learning. Based on this comparison of algorithms, the more efficient, PPO algorithm was used to train the agent in the advanced environment. The success of training was tested on three levels of platform games. The result of this work is a successfully trained agent that, based on what it has learned, reaches the level target while avoiding traps. The platform game consists of three levels where the player can face the agent in a competition to win and one level which is used as an advanced training environment and now the player can alone train on it