Learning Long-Term Behavior of Unknown Dynamical Systems from Limited Data
Project Description
This project focuses on inferring long-term dynamical behavior from short-term, limited time-series observations. The goal is to develop models that remain stable and accurate over long forecasting horizons despite having only brief trajectories for training. The work includes building a reproducible pipeline for data preprocessing, training data-efficient forecasting models, and evaluating long-horizon rollouts with metrics that capture both short-term accuracy and long-term stability. Emphasis is placed on understanding failure modes such as error accumulation and drift, and on testing strategies that improve long-term consistency under sparse and noisy data.
Supervisor
CHEN, Junfeng
Quota
2
Course type
UROP1100
UROP2100
UROP3100
UROP4100
Applicant's Roles
The role is to carry out a guided research workflow to infer long-term dynamical behavior from short-term time-series data. Key responsibilities include:
- Reviewing core background on dynamical systems and time-series forecasting.
- Preparing datasets: cleaning, normalization, train/validation/test splits, and handling noise/sparsity.
- Implementing baseline models in Python (NumPy/PyTorch) and running controlled experiments.
- Designing evaluation for long-horizon rollouts (error accumulation, stability, drift) and reporting results with clear metrics and plots.
- Using Git/GitHub to maintain a reproducible codebase, track experiments, and document progress.
- Delivering a final package consisting of well-organized code, experiment logs, and a concise research-style report summarizing methods, findings, and limitations.
- Reviewing core background on dynamical systems and time-series forecasting.
- Preparing datasets: cleaning, normalization, train/validation/test splits, and handling noise/sparsity.
- Implementing baseline models in Python (NumPy/PyTorch) and running controlled experiments.
- Designing evaluation for long-horizon rollouts (error accumulation, stability, drift) and reporting results with clear metrics and plots.
- Using Git/GitHub to maintain a reproducible codebase, track experiments, and document progress.
- Delivering a final package consisting of well-organized code, experiment logs, and a concise research-style report summarizing methods, findings, and limitations.
Applicant's Learning Objectives
- Learn foundational knowledge in dynamical systems and time-series modeling.
- Gain proficiency in Python programming and Git/GitHub for version control and collaborative workflows.
- Acquire practical skills in data processing and model implementation using Python scientific/ML tools (e.g., NumPy, PyTorch).
- Become familiar with the core concepts, common practices, and research culture of scientific machine learning.
- Gain proficiency in Python programming and Git/GitHub for version control and collaborative workflows.
- Acquire practical skills in data processing and model implementation using Python scientific/ML tools (e.g., NumPy, PyTorch).
- Become familiar with the core concepts, common practices, and research culture of scientific machine learning.
Complexity of the project
Moderate