Comparing Sample Efficiency Between Model-Based and Model-Free Reinforcement Learning Methods

Wu, Dennis Chong Yi (2024) Comparing Sample Efficiency Between Model-Based and Model-Free Reinforcement Learning Methods. Bachelor's Thesis, Artificial Intelligence.

Preview	Text thesiss4752244.pdf Download (4MB) | Preview
	Text Toestemming Wu.pdf Restricted to Registered users only Download (130kB)

Abstract

Model-free Reinforcement Learning (RL) has been successfully applied to complex tasks, such as playing Atari games from image observations, but it typically requires a large amount of sample data. Model-based Reinforcement Learning attempts to address this issue by introducing a transition model, thereby reducing the number of samples needed to train an agent. This paper compares the sample efficiency of a Model-Free Double Deep Q-Network (DDQN) algorithm with a Model-Based Dyna-Q algorithm. Both agents were trained across various environments to determine which algorithm achieves a predefined reward threshold faster. Results indicate that while both agents can achieve the threshold, the model-based Dyna-Q algorithm consistently reaches it with fewer samples. However, this sample efficiency advantage comes at the cost of significantly higher computational resources.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Cardenas Cartagena, J. D. and Sabatelli, M.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	09 Aug 2024 12:07
Last Modified:	09 Aug 2024 12:07
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/33922

Actions (login required)

View Item