Simulation-Based Game Testing for Estimating Player Curiosity

Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Master's thesis

Date

2023-03-20

Department

Major/Subject

Autonomous Systems

Mcode

ELEC3055

Degree programme

Master's Programme in ICT Innovation

Language

en

Pages

30+15

Series

Abstract

Over the past few years, game testing based on simulating human-like agents instead of humans has grown in popularity in order to reduce the costs associated with manual testing and improve quality assurance. Although these human-like agents can successfully perform the game-testing task using deep reinforcement learning algorithms, they still lack some human characteristics, such as intrinsic drive, and cannot completely replace playtesters. This study focuses on granting intrinsic curiosity rewards to Ml-agents in order to provide them with intrinsic motivation. The thesis sets out to test the effectiveness of two different types of intrinsic curiosity rewards on training an agent capable of playing Super Mario Bros (SMB) and to understand how intrinsic reward is affected by the inclusion of different level design features. We utilize a factorial design and train a Unity ML-agent with two types of intrinsic reward (ICM, RND) on 20 modifications of SMB level 2-1. We find that both intrinsic rewards allow the agent to complete the levels like a human player, and Random Network Distillation has better sensitivity to different level design features.

Description

Supervisor

Guckelsberger, Christian

Thesis advisor

Takatalo, Jari
Perttu, Hämäläinen Perttu

Keywords

Super Mario Bros., deep reinforcement learning, aI playtesting, human curiosity

Other note

Citation