Dinamično spreminjanje stopnje v računalniški igri s pomočjo generativne nasprotniške nevronske mreže

Ćehić, Edin (author), Klemenc, Bojan (mentor)

V zadnjih letih smo priča izjemno hitremu razvoju umetne inteligence v splošni rabi. Specifično nas zanima uporaba umetne inteligence za ustvarjanje. Čeprav mnogi kritizirajo to smer, verjamemo, da je lahko zelo uporabna v prihodnosti, še posebej pri razvoju video iger. Umetna inteligenca, zlasti generativne nevronske mreže, se že uporabljajo za generiranje tekstur in grafik. V naši raziskavi pa smo se osredotočili na uporabo generativnih nasprotniških nevronskih mreže za ustvarjanje nivojev ali svetov v računalniških igrah, ki se dinamično prilagajajo uporabnikovim dejanjem. Uporabili smo odprtokodni klon igre Hill Climb Racing, zasnovan za učenje navidezno inteligentnih agentov. Stopnje smo generirali z našo nevronsko mrežo, jih vstavili v igro in nato prilagajali z uporabo nevronske mreže glede na uspešnost agentov pri reševanju nivojev.

• Language: Slovenian
• Keywords: računalniške igre, generativne nevronske mreže, spodbujevano učenje, generativne nasprotniške mreže
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-161478
• Publication date in RUL: 11.09.2024

Secondary language

• Language: English
• Title: Dynamic level modification in a computer game using a generative adversarial neural network

Abstract

In recent years, we have witnessed an extremely rapid development of artificial intelligence in general use. Specifically, we are interested in the application of artificial intelligence for generation of art and assets. Although many criticize this direction, we believe it can be very useful in the future, especially in video game development. Artificial intelligence, particularly generative neural networks, is already used for generating textures and graphics. In our research, however, we focused on using generative adversarial neural networks to create levels or worlds in computer games that dynamically adapt to the user's actions. We used an open-source clone of the game Hill Climb Racing, designed for training virtually intelligent agents. We generated levels with our generative neural network, integrated them into the game, and then modified the levels with the neural network based on the agents' performance in solving them.

• Keywords: computer games, generative neural networks, reinforcement learning, generative adversarial networks