Connectionist Machine Learning Techniques in a 2D Arcade Game

Langhorst, J. and Maathuis, H. (2016) Connectionist Machine Learning Techniques in a 2D Arcade Game. Bachelor's Thesis, Artificial Intelligence.

Preview	Text AI_BA_2016_JEROENLANGHORST.pdf - Published Version Download (347kB) | Preview
	Text Toestemming.pdf - Other Restricted to Backend only Download (532kB)

Abstract

Implementing intelligent behaviour in an agent for a computer game is often established by the usage of simple rules. Our research investigates the performance of Learning from Demonstration in combination with several techniques (Q-learning, SARSA and Supervised Learning) when learning policies for an agent in a 2D arcade game. For our experiment we constructed a 2D platforming arcade game using Java, in which the agent is controlled by Neural Networks. Experience Replay is explored to train the agent in an efficient manner. The goal was to see if the algorithms were capable of learning the desired policies. We also included several activation functions of the neural networks in our research: the traditional sigmoid, ReLU and a linear approach. We could not find any significant difference in success rate between implementations in the continuous environment. Also, we found that in a discretized environment Q-learning and SARSA in combination with the sigmoid should be avoided as it performs significantly worse than all other implementations.

Item Type:	Thesis (Bachelor's Thesis)
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	15 Feb 2018 08:25
Last Modified:	15 Feb 2018 08:25
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/14524

Actions (login required)

View Item