Classification of Human Learning Stages via Kernel Distribution Embeddings

Abstract

Adaptive automation, automation which is responsive to the human's performance via the alteration of control laws or level of assistance, is an important tool for training humans to attain new skills when operating dynamical systems. When coupled with cognitive feedback, adaptive automation has the potential to further facilitate human training, but requires precise assessments of human progression through various learning stages. This is challenging because of the underlying dynamics, as well as the stochasticity inherent to human action. We propose a data-driven approach to assess learning stages in a complex quadrotor landing task that is responsive to stochastic, human-in-the-loop quadrotor dynamics. We represent each learning stage as a distribution of canonical trajectories for that learning stage, then employ kernel distribution embeddings in combination with a rule-based heuristic, to determine which canonical distribution a sample landing trajectory is closest to. We demonstrate our approach on experimental human subject data, and use our approach to evaluate the efficacy of cognitively-based adaptive automation designed to calibrate self-confidence. Our approach is more accurate than standard classification methods, such as nearest centroid assignment, which rely on metrics that are not inherently suited to analysis of trajectories of stochastic dynamical systems.

Keywords

• Cyberphysical systems
• cognitive systems and control
• human-in-the-loop systems
• rule-based classification
• kernel methods