The capability to adapt to dynamic environments and changes in data distribution is essential for the development of adaptive artificial intelligence (AI) systems that can operate effectively in the real world. Typically, adaptive learning requires an AI agent to autonomously gather the required information (i.e., training data) from the deployed environment to adapt to the local dynamic changes. This process falls within the broad area of semi-supervised learning, where partial training data is not manually labeled. This is especially critical for applications at the edge of cyber-physical systems, where available datasets are rather limited and resources for deep learning from big data are not available. While there exist effective semi-supervised learning methods based on feature smoothness assumptions, they become less robust in dynamic environments where there are shifts (e.g., concept drift) in data distribution. Recently, causality has been explored to address deployment domain shifts. This is based on the argument that causality, especially the causal direction, is consistent across different domains, making invariant features highly significant in adaptive learning.

This dissertation presents an interactive causality (IC) methodology, which utilizes directional and temporal causal events to facilitate the automated self-labeling of data. Interactive causality refers to the temporal state transitions of causes and effects during interactive activities, which are captured through additional observing channels. The methodology capitalizes on the domain-invariant causal relationships to capture information leakage during interactions from an additional observing channel, and is fortified by the rich domain knowledge in cyber-physical system contexts. Differing from big data ML approaches, the IC leverages the existing knowledge and experiences to create a knowledge graph with embedded temporal causality among interactive nodes and directly look at the interactivity among these nodes for adaptive learning. A theoretical foundation of the interactive causality driven self-labeling method is discussed and compared with other traditional semi-supervised learning algorithms. The proof is derived from the theory of dynamical system which represents the time-varying environments and the time-series data. A simulation using physics engine and a real-world example in semiconductor manufacturing are provided to demonstrate the effectiveness of the proposed method for adaptive learning and knowledge expansion. Overall, the experimental results successfully demonstrated the efficacy of our proposed adaptive learning in reducing manual data labeling efforts and robust domain adaptation.