Use of reinforcement learning for training game agents in fighting games

V tej magistrski nalogi smo se osredotočili na razvoj igralnega agenta z uporabo okrepitvenega učenja in posebej obravnavali vprašanje, ali so ti agenti dosegli ali presegli raven profesionalnih človeških igralcev. Delo je razdeljeno na teoretični in praktični del. Teoretični del zajema osnovne koncepte okrepitvenega učenja in opisuje uporabljena orodja. V praktičnem delu smo izvedli tri eksperimente z različnimi velikostmi korakov in časi usposabljanja modela ter nato primerjali njihovo uspešnost. Naša analiza kaže, da se v našem konkretnem primeru s povečevanjem časa učenja pri uporabi tehnik okrepitvenega učenja postopoma približujemo optimalnim rezultatom. Naša raziskava potrjuje učinkovitost okrepitvenega učenja pri oblikovanju igralnih agentov, katerih primerjava je relevantna zgolj ob zadostnem obsegu usposabljanja.In this master\u27s thesis, we focused on developing a game agent using reinforcement learning, specifically addressing whether these agents have reached or surpassed the level of professional human players. The work is divided into theoretical and practical sections. The theoretical part covers fundamental reinforcement learning concepts and describes the tools used. In the practical section, we conducted three experiments with varying step sizes and training times for the model, subsequently comparing their performance. Our findings indicate that reinforcement learning techniques yield optimal results only when the agent is trained sufficiently long. In conclusion, our study supports the effectiveness of reinforcement learning in advancing game agents that can only be compared effectively after undergoing extensive training