Playing Frogger using Multi-step Q-learning

Maijers, Jip (2019) Playing Frogger using Multi-step Q-learning. Bachelor's Thesis, Artificial Intelligence.

Preview	Text bAI_2019_MaijersJWJ.pdf Download (613kB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (119kB)

Abstract

This thesis is an extension on the Bachelor's thesis by Tomas van der Velde about training an agent to play the game Frogger using Q-learning. In this thesis 2-step Q-learning and 4-step Q-learning with 2 different reward functions (an action-based reward function and a distance-based reward function) are used to see if it is possible for agents trained with these algorithms to outperform agents trained with regular Q-learning. The game Frogger can be described as a two-part game, where the aim of the first part is crossing a road and where the goal of the second part is crossing a river. In previous research it was found that crossing the river was the biggest obstacle for agents, which is why two separate neural networks were again used. The results obtained show that the action-based reward function outperforms a distance-based reward function and that 4-step Q-learning and 2-step Q-learning improved upon the performance of regular Q-learning, with 4-step Q-learning performing the best of the three in terms of road completion, and 2-step Q-learning in terms of points gained and overall win rate.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Wiering, M.A.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	08 May 2019
Last Modified:	10 May 2019 10:44
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/19442

Actions (login required)

View Item