Strojno učenje obnašanja inteligentnih agentov v računalniških igrah

Penca, David (author), Bosnić, Zoran (mentor)

V diplomskem delu predstavljamo pristop programiranja igralcev v večigralskih spletnih igrah, ki temelji na metodah strojnega učenja. Pokazati želimo, da lahko posameznim likom določimo poteze, ki jih lahko izvajajo, jim podamo informacije o njihovem okolju in jih prepustimo, da si na podlagi bojev s človeškimi igralci ustvarijo igralno taktiko. Pristopi, ki temeljijo na sprotnem strojnem učenju taktik likov, lahko zmanjšajo čas, porabljen za programiranje, hkrati pa omogočajo prilagajanje nasprotnikov taktikam igralcev brez dodatnega dela programerjev. Tako dobimo igralce, ki se čez čas izboljšujejo in so robustni na izkoriščanje uveljavljenih taktik s strani človeškega igralca. Osredotočili smo se na spodbujevano učenje in na evolucijske algoritme, saj sta oba pristopa primerna za sisteme, ki se učijo na podlagi številnih interakcij s človeškimi nasprotniki. Naše rešitve smo implementirali v igralnem pogonu Unreal Engine 4.

• Language: Slovenian
• Keywords: strojno učenje, Q-učenje, genetski algoritmi, računalniške igre, inteligentni agenti
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2018
• PID: 20.500.12556/RUL-103413
• Publication date in RUL: 17.09.2018

PENCA, David, 2018, Strojno učenje obnašanja inteligentnih agentov v računalniških igrah [online]. Bachelor’s thesis. [Accessed 24 March 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=103413

Secondary language

• Language: English
• Title: Machine learning of character behavior in computer games

Abstract

In our thesis we present an approach for programming enemy characters in online multiplayer games that is based on machine learning algorithms. We wish to demonstrate, that it is possible to specify the available actions for specific characters, implement sensing of their environment and let them learn the tactics on their own, by fighting human players. Approaches based on machine learning have the potential to reduce the time needed for programming as well as enable the characters to adapt to current player tactics, without any additional programming. By using such programming methods we are able to create characters which get better over time and are not vulnerable to exploitation of established tactics by the players. We have focused mainly on reinforcement learning and evolutionary algorithms, because both approaches are suitable for use in systems that learn from numerous interactions with human players. We have implemented our prototype in the Unreal Engine 4 game engine.

• Keywords: machine learning, Q-learning, genetic algorithms, computer games, inteligent agents