Explanation-Guided Adaptive Learning for Human-Centered Cyber-Physical Systems

Abstract

Human-centered Cyber-Physical Systems (H-CPS), such as smart rehabilitation monitors and intelligent tutoring systems, promise transformative, personalized assistance but often struggle under real-world deployment constraints due to their reliance on conventional end-to-end optimization. Specifically, deployed H-CPS face significant challenges in providing actionable, trustworthy explanations required in high-stakes domains and suffer from catastrophic forgetting when adapting to shifting environments under strict data privacy regulations. This dissertation addresses these critical gaps by proposing Explanation-Guided Adaptive Learning, a framework that reorients explainability from a passive, post-hoc diagnostic tool into an active computational mechanism. By defining "explanation" broadly as an interpretable, structured context that encompasses formal guarantees, geometric constraints, and symbolic knowledge, this work uses such a context to actively guide system behavior across three dimensions: assurance, adaptation, and interaction. First, we develop frameworks that transform explanation into an active mechanism, using attribution refinement and formal verification to produce actionable, targeted explanations with provable robustness guarantees. Second, we introduce frameworks that use structural priors to constrain how models adapt, leveraging geometric preservation and channel-aware mechanisms to enable privacy-safe continual learning and zero-shot transfer without replay buffers. Finally, we present a neuro-symbolic simulation framework that imposes hard architectural constraints from cognitive models onto large language models, producing synthetic human trajectories that are behaviorally faithful, where unconstrained baselines exhibit severe competency bias. Validated across physical rehabilitation, autonomous driving, and educational domains, this dissertation demonstrates that embedding interpretable context, geometric structure, and symbolic knowledge into the learning loop yields AI systems that are measurably more trustworthy, resilient, and human-faithful.