Machine Learning Control Project

The Machine Learning Control Project explores how modern data-driven techniques can augment or replace traditional control design methodologies. While classical control theory depends heavily on first-principles modeling, many real-world systems are too complex, nonlinear, or uncertain for purely analytic models to capture effectively. This project investigates approaches in which machine learning—ranging from sparse regression to deep neural networks—plays an active role in modeling, identification, and controller synthesis.

Research topics include dynamic mode decomposition (DMD) for reduced-order modeling, sparse identification of nonlinear dynamics (SINDy) for interpretable model discovery, nonlinear system identification for control-oriented modeling, and model predictive control (MPC) informed by learned dynamics. The project also studies nonlinear optimal control, deep reinforcement learning for policy learning, and hybrid schemes that integrate learning-based models with robust or adaptive control architectures.

Applications span robotics, autonomous navigation, and other cyber-physical systems, with an emphasis on real-time feasibility and safety-critical operation. A recent example is the Master's thesis work by Ariful Islam (2024), which developed a semantic command–based obstacle avoidance controller for a mobile robot using deep learning to predict safe navigation trajectories. This study demonstrated how learned models can replace or complement conventional control pipelines while retaining interpretable structure for system verification.

The broader goal is to develop a principled framework for machine learning–enhanced control systems—one that bridges rigorous control theory with flexible, adaptive learning methods to enable robust, high-performance control in challenging environments.