Bibliographic details and abstracts are available to all. Downloads of full-text dissertations are restricted to University of Portsmouth members who must login. MPhils may be accessed by all.

Abstract

Deep Reinforcement Learning has seen limited use and research in video games, despite its rapid growth and implementation in other fields. This study seeks to test the validity and possibilities of using deep reinforcement learning as an alternative to traditional AI techniques, by creating an artefact that tests the navigational abilities of an agent with advanced movement techniques in a 3D environment. Vision observations were used with agents to encourage exploration and to gain an understanding of the environment in an attempt to imitate human players.

The developed artefact was used as a comparator to human play, to draw reasonable conclusions about its potential effectiveness as an AI agent in a video game, and found promising results with convincing behaviour. There were notable drawbacks to this approach, however, with long training hours, difficult debugging, and a high-performance demand for training. Conversely, machine learning can improve the efficiency of developers, enabling the creation of other aspects of a video game whilst agents train, offsetting one of the most critical disadvantages of deep reinforcement learning.

Course: Computer Games Technology - BSc (Hons) - C1671

Date Deposited: 2022-11-04

URI/permalink: https://library.port.ac.uk/dissert/dis14054.html