The Effect of Temporal Dependence on the Performance of Stochastic Approximation Algorithms

,

Professor: Zhipeng Lou

Description: Stochastic approximation algorithms are widely used in online learning and statistical estimation, where data arrive sequentially and model parameters are updated as new observations become available. Many commonly used learning-rate schedules are developed and analyzed under the assumption that observations are independent. However, real-world sequential data often exhibit temporal dependence. For example, financial time series and temperature records can exhibit temporal dependence, in which observations are correlated with past observations.

This project will investigate how temporal dependence in sequential data influences the performance of online stochastic approximation algorithms. In particular, we will examine whether commonly used learning-rate schedules can achieve satisfactory finite-sample performance when data exhibit temporal dependence, focusing on both estimation accuracy and the validity of statistical inference. Through theoretical analysis and computational experiments, the student will explore how different forms and strengths of temporal dependence affect algorithm performance.

If existing learning-rate schedules perform poorly under temporal dependence, the project will explore modifications to these schedules or develop new ones tailored to dependent data. Where feasible, we will establish theoretical guarantees for the proposed methods, such as convergence results and bounds on estimation error. This project offers an opportunity to gain hands-on experience in statistical learning, online algorithms, and the theoretical analysis of stochastic methods.

Preferred Qualifications: Basic knowledge of probability and mathematical statistics, and programming experience in R or Python

Contact Us