This item is under embargo and not available online per the author's request. For access information, please visit http://digital.wustl.edu/publish/etd-search.html.

Abstract

Sparse reward environments present a significant challenge in reinforcement learning (RL) due to the limited feedback from extrinsic rewards. This research proposes a target-driven exploration framework that leverages intrinsic rewards derived from self-supervised feature extraction to guide agents toward meaningful states. The framework employs three feature extraction methods—Raw Pixels, Random Fixed Features, and Variational Autoencoders (VAEs)—to represent state embeddings and compute intrinsic motivation. Evaluations on benchmark environments, such as SpaceInvaders, demonstrate improved exploration efficiency and learning capabilities. The results highlight the framework's ability to utilize intrinsic rewards effectively, enabling agents to navigate and learn in complex, sparse reward settings. Future work will refine feature extraction techniques and further optimize the model to address more complex environments.

Document Type

Article

Author's School

McKelvey School of Engineering

Author's Department

Electrical and Systems Engineering

Class Name

Electrical and Systems Engineering Undergraduate Research

Language

English (en)

Date of Submission

12-3-2024

Available for download on Tuesday, December 02, 2025

Share

COinS