Transfer Learning in Reinforcement Learning: When Task-Specific Adaptation Outperforms Generalization

Zaharia, Catalin (2025) Transfer Learning in Reinforcement Learning: When Task-Specific Adaptation Outperforms Generalization. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Transfer-Learning-in-Reinforcement-Learning-When-Task-Specific-Adaptation-Outperforms-Generalization.pdf Download (2MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (235kB)

Abstract

Transfer Learning has emerged as a transformative technique in Reinforcement Learning (RL), enabling algorithms to leverage knowledge from prior tasks to improve efficiency and generalization. This study investigates whether transfer learning can further enhance the already exceptional learning capabilities of the Parallelized Q-Network (PQN). Using the MinAtar environment as a controlled testbed, I integrate transfer learning techniques to assess their impact on learning time. By comparing the algorithm with and without transfer learning across multiple tasks, the results reveal that transfer learning consistently underperforms when compared to task-specific training. Contrary to expectations, transfer configurations fail to accelerate learning, with environment-specific training emerging as the superior approach. These findings underscore the limitations of transfer learning in RL and reaffirm the critical role of environment-specific adaptation in achieving efficient and robust learning outcomes.

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:	14 Mar 2025 10:17
Last Modified:	14 Mar 2025 10:17
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/34861

Actions (login required)

View Item