Theses and Dissertations

Date of Award

5-2024

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

First Advisor

Dongchul Kim

Second Advisor

Emmett Tomai

Third Advisor

Eric Enriquez

Abstract

Deep Reinforcement learning (DRL) has gained importance in optimizing control policies, while Graph Neural Networks (GNNs) offer a robust approach for modeling complex relationships within systems represented as graphs. This thesis investigates the integration of DRL and GNNs to optimize control policies for robotic tasks, with a focus on locomotion. It compares static and dynamic GNN architectures for control policy predictions, revealing their strengths and limitations in adapting to locomotion predictions. The study assesses the impact of model structure complexity on GNNs' predictive capabilities, showcasing how intricate model structure can maximize GNNs' potential in capturing spatial and relational dependencies when learning control policies. Through comprehensive experiments and analysis, this research contributes insights into the optimal integration of GNNs in DRL, particularly in continuous locomotion tasks, addressing a gap in the literature. It also introduces a new implementation for conducting RL-GNN experiments with a 3-D simulation physics engine by employing PyTorch, PyTorch Geometric, and PyTorch Geometric Temporal, enhancing the methodological repertoire and promoting reproducibility in research. This study advances the understanding and application of DRL-GNN frameworks in robotics, facilitating the development of more intelligent and adaptive robotic systems.

Comments

Copyright 2024 Gaukhar Nurbek.

https://go.openathens.net/redirector/utrgv.edu?url=https://www.proquest.com/pqdtglobal1/dissertations-theses/exploring-graph-neural-networks-reinforcement/docview/3085279182/sem-2?accountid=7119

Download

Share

COinS