World Model Agents with Change-based Intrinsic Motivation

Ferrao, Jeremias Lino (2024) World Model Agents with Change-based Intrinsic Motivation. Bachelor's Thesis, Artificial Intelligence.

Preview	Text bAI2024JeremiasFerrao.pdf Download (5MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (155kB)

Abstract

Sparse reward environments present significant challenges in reinforcement learning due to the infrequency of feedback, making it difficult for agents to learn effective policies. This thesis explores the performance of the DreamerV3 agent, enhanced with CBET, in comparison to the IMPALA agent within such environments. By leveraging a world model to internalize environment dynamics, DreamerV3 achieves superior sample efficiency and faster convergence. The CBET variant of DreamerV3 further improves performance by incorporating intrinsic motivation to guide exploration, proving especially beneficial in the challenging minigrid and crafter environments. Despite a brief anomaly in the minigrid transfer experiment where IMPALA outperforms DreamerV3, the overall results demonstrate the efficacy of DreamerV3 and its CBET variant in optimizing extrinsic returns and accelerating learning processes in sparse reward settings. Future research should address the limitations observed in policy transfer methods to enhance the robustness and generalizability of these findings. Additionally, exploring the interpretability of reinforcement learning algorithms is crucial to understand the decision-making processes of agents and the impact of intrinsic rewards. Reducing the computational resources required for DreamerV3 without compromising performance will also be a key focus, aiming to make these advanced techniques more accessible and practical for a wider range of applications.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Fernandes Cunha, R.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	07 Aug 2024 09:30
Last Modified:	07 Aug 2024 09:30
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/33851

Actions (login required)

View Item