Machine learning is seeping into every fabric in various practical domains such as autonomous driving, wearable computing, and smart buildings. However, in the actual development and integration, especially when the learning-based components are frequently included as components of large complex systems where the physical instances can be included as interactable components, they often pose significant data challenges, including data noise, insufficient training data, or lacking annotations, which could significantly hinder the learning process. In this dissertation, we will introduce several approaches to tackle these challenges. In building-related cyber-physical systems, a transfer learning-based approach is proposed to tackle the challenge of long training time in model-free DRL for building HVAC control. Then we also consider accelerating online Deep Reinforcement Learning (DRL) for building HVAC control with the help of heterogeneous expert guidances. Besides, to handle the problem of HVAC control under corrupted sensor inputs, a learning-based framework is proposed for sensor fault-tolerant building HVAC control. In the vision task domain, we investigate the challenge of using the cross-domain unlabeled data and weak annotator to mitigate the data insufficiency in the target domain. Moreover, I also propose a novel approach for handling multi-label classification with unseen classes in the testing stage.

Creator

• Xu, Shichao

DOI

• https://doi.org/10.21985/n2-12hh-vr30

Subject

• Computer engineering
• Computer science

Language

• en

Alternate Identifier

• etdadmin_upload_983783
• http://dissertations.umi.com/northwestern:16519

Keyword

• Transfer Learning
• Reinforcement Learning
• Computer Vision
• Building HVAC Control
• Image Classification
• Machine Learning

Date created

• 2023-01-01

Resource type

• Dissertation

Rights statement

• In Copyright