Outperforming the Baseline: Transfer Learning in Atari via Parallelized Q-Networks

Matic, Andjela (2025) Outperforming the Baseline: Transfer Learning in Atari via Parallelized Q-Networks. Bachelor's Thesis, Artificial Intelligence.

Preview	Text bAI2025MaticA.pdf Download (1MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (243kB)

Abstract

Transfer Learning (TL) is an emerging technique present in reinforcement learning (RL), showing great potential and contributions to decreasing computational costs, improving efficiency, and maximizing generalization. This study looks at TL in the scope of the Parallelized Q-Network (PQN) algorithm, which has demonstrated vast improvements in cost-effectiveness and overall performance compared to peer algorithms, such as Rainbow, IQL, PPO, etc. Building on previously conducted research in the MinAtar environment, this paper focuses on applying and evaluating TL in the Atari environment. Due to its favourable dynamics and structure, the Atari environment addresses limitations of previous research and presents a suitable extension. The results from multiple configurations show significant improvements in TL simulations compared to the baseline. In the example of the Pong-Breakout pair, training is noticeably accelerated, with convergence being reached more than twice as fast compared to the baseline. These findings illustrate the potential of TL in complex environments, especially when the source and target environments share underlying visual and mechanical features.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Fernandes Cunha, R.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	12 Aug 2025 08:48
Last Modified:	12 Aug 2025 08:48
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/36746

Actions (login required)

View Item