Safe Deep Reinforcement Learning for Super Mario Bros using Heuristics

Abdul Aziz, Amir (2024) Safe Deep Reinforcement Learning for Super Mario Bros using Heuristics. Bachelor's Thesis, Artificial Intelligence.

Preview	Text bAI2024AbdulAzizA.pdf Download (759kB) | Preview
	Text Abdul Aziz - ja.pdf Restricted to Registered users only Download (178kB)

Abstract

This thesis explores the integration of heuristic safety mechanisms within Deep Reinforcement Learning (Deep RL) frameworks to enhance agent survival in the complex, dynamic environment of Super Mario Bros. We examine the influence of heuristic-augmented Actor-Critic models on agent behavior, particularly in terms of safety without compromising learning efficiency. The study compares various configurations, including separate and shared neural network architectures and different combinations of frame stacks for the input layer. It quantitatively measures survival rates, highlighting a significant improvement in avoiding fatal encounters with the most effective configuration around the 10000th episode mark. However, these safety mechanisms may also constrain exploration and long-term reward optimization. Despite achieving a stable survival pattern and a consistent minimum mean loss, a tension between safety and reward maximization is evident. This work contributes to the domain of Safe RL, emphasizing the nuanced trade-offs in stochastic gaming environments.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Cardenas Cartagena, J. D.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	03 May 2024 09:19
Last Modified:	03 May 2024 11:06
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/32359

Actions (login required)

View Item