A Stability Principle for Learning Under Nonstationarity
提出一个稳定性原则，通过平衡历史数据带来的偏差与统计噪声，指导在变化环境中如何选择合适的数据量，使决策既稳健又适应变化。
Adapting to a Changing Environment with Steady Decisions Most real-world decision making unfolds in changing environments. When yesterday’s data may no longer depict today, how much past data should guide future decision making? Use too little and you chase noise; use too much and you overlook new trends. In “A Stability Principle for Learning under Nonstationarity,” Huang and Wang propose a simple stability principle: keep adding past data until the bias it incurs exceeds the natural statistical noise. Their theory shows that the proposed approach can optimally adapt to unforeseeable changes in the environment for a broad spectrum of statistical learning and decision-making problems. At its core are a novel similarity measure for different decision-making objectives and a segmentation technique that breaks down the nonstationary data stream into quasistationary pieces. Simulations and real-data experiments spanning several nonstationarity patterns confirm the approach’s ability to make steady yet responsive decisions.