Learning from Mixed-Quality Data via Exploration-Guided Cognitive Manifolds

Yuntao Zou, Wulin Xie, xu zl, Qianqi Zhang

Keywords: Energy management systems, satellite battery arrays, cognition-driven adaptive learning, limited data learning, multi-objective optimization, hard constraint guarantees, aerospace engineering, dynamic policy regulation

Abstract: Data-driven learning paradigms typically require massive amounts of high-quality data to learn reliable decision models. In safety-critical domains, the cost of obtaining high-quality labeled data often exceeds practical budgets, while abundant low-quality data from routine operations remains unutilized. This creates a paradox: systems are data-rich but knowledge-poor, as traditional methods require extensive preprocessing that often removes valuable information. We introduce a learning framework that constructs cognitive manifolds to represent the system's evolving understanding, enabling principled utilization of mixed-quality data. Rather than filtering out imperfect data or treating all data equally, our approach dynamically evaluates each data point's potential contribution based on the learner's current cognitive state. We formalize this through exploration measures that capture the model's understanding uncertainty, enabling it to identify when low-quality data can fill knowledge gaps versus when it would introduce harmful noise. The framework transforms the traditional data quality problem into a cognitive matching problem: aligning data characteristics with the learner's current knowledge needs. Through this cognitive manifold representation, even highly noisy data becomes valuable during early learning stages when any information reduces uncertainty, while the same data is appropriately ignored once the model develops robust understanding. Experiments on multi-objective optimization tasks with mixed-quality data demonstrate that our approach extracts 3.6% more actionable knowledge from the same datasets compared to quality-filtering baselines, while maintaining the strict constraint satisfaction required in safety-critical applications.

Primary Area: reinforcement learning

Submission Number: 6194