In the last few years, deep reinforcement learning has found extensive application in computer games and simulators due to ability to generate a large number of interactive and visually-rich simulations. In this paper, a 3D environment is developed using Unity game engine, and the use of the ML-Agents Toolkit for training intelligent agents with a PyTorch implementation of the Proximal Policy Optimization (PPO) algorithm is described. The paper compares the performance of different configurations and discusses successful training outcomes. Through reinforcement learning, the agent has learned skills such as movement, environment exploration, enemy avoidance, and coin collection. Additionally, the paper covers the theory behind machine learning, reinforcement learning, deep neural networks, the PPO algorithm, and the use of ML-Agents Toolkit.

Zadnjih godina razvoj u dubokom podržanom učenju uvelike je imao primjenu u računalnim igricama i simulatorima zbog mogućnosti generiranja velikog broja interaktivnih i vizualno bogatih simulacija. U ovom radu, razvijeno je 3D okruženje koristeći pogonski sustav za računalne igre Unity te je opisana uporaba ML-Agents Toolkita za treniranje inteligentnih agenata pomoću gotove PyTorch implementacije algoritma Proximal Policy Optimization (PPO). U radu su uspoređene performanse različitih konfiguracija te opisano uspješno treniranje. Podržanim učenjem, agent je naučio vještine kretanja i istraživanja okoline, izbjegavanje neprijatelja te skupljanja novčića. Također, u radu je pokrivena teorija iza strojnog učenja, podržanog učenja, dubokih neuronskih mreža, algoritma PPO te korištenje programskog alata ML-Agents.