Accelerating Model Based Reinforcement Learning Using GPU Through Parallelization of Dyna-Q Architecture

Stoian, Rares Stefan (2025) Accelerating Model Based Reinforcement Learning Using GPU Through Parallelization of Dyna-Q Architecture. Bachelor's Thesis, Artificial Intelligence.

Preview	Text BCsThesiss4775309.pdf Download (23MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (242kB)

Abstract

In this paper, we propose Parallelized Dyna Q-Network (PDQN), a fully online, GPU-accelerated reinforcement learning algorithm that integrates model-based planning with the recently introduced Parallelized Q-Network (PQN). By employing learned world models for short-horizon planning, PDQN seeks to further accelerate convergence while preserving the simplicity of fully online temporal-difference learning. Empirical results on multiple MinAtar environments show that PDQN achieves performance on par with PQN. Our analysis also reveals that selecting appropriate values for planning delay and planning steps is crucial for the performance to be at least on par with the baseline.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Fernandes Cunha, R. and Sabatelli, M.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	28 May 2025 06:39
Last Modified:	28 May 2025 06:39
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/35220

Actions (login required)

View Item