ETD

• Explainable AI
• Human-in-the-Loop
• Hybrid Decision-Making Systems
• Learning to Defer
• Skeptical Learning

As Machine Learning models become increasingly pervasive into high-stakes decision making environments, human-AI interaction evolves from complete control towards more complex, interactive collaboration. Hybrid Decision Making Systems (HDMS) based on Artificial Intelligence allow for close alignment between human expertise and machine automation. In this context, this thesis develops and implements BRIDGET, a novel phase-shifting HDMS designed to ``bridge the gap`` between two canonical frameworks: Learning to Defer, which focuses on efficient task deferral, and Learning Together, which promotes co-evolutionary collaboration between agents.
BRIDGET dynamically transitions between Human-in-Command and Machine-in-Command decision making, by understanding which agent should be in charge at a certain point. By including Skeptical Learning, BRIDGET aids the user in labeling data efficiently, while budget-aware exit conditions and deferral policies address the problem of human fatigue.
The proposal is tested using established benchmark datasets, while the experimental setup features synthetic experts and state-of-the-art deferral policies.