On the Surprising Effectiveness of Spectral Clipping in Learning Stable Linear Dynamics

• This is a repository containing the code for the paper "On the Surprising Effectiveness of Spectral Clipping in Learning Stable Linear Dynamics".

• Some of the code is forked from "Memory-Efficient Learning of Stable Linear Dynamical Systems for Prediction and Control"

• Code still under development.

Code Instructions

• Please download the dataset to your local directory.

  • Please download the datasets through this link.

  • Extract datasets.zip to \spec_clip. The directory tree should be like:

    • \spec_clip
      • ucsd.m
      • dtdb.m
      • kodex.m
      • \datasets
        • \ucsd
        • \dtdb
        • \kodex
• Run ucsd.m with MATLAB. It will learn the linear dynamics of UCSD dataset with LS, SC, SOC, WLS and CG respectively. The script will automatically create a new directory \results_ucsd to save system matrices, training time and reconstruction error.
• Run dtdb.m with MATLAB. It will learn the linear dynamics of DTDB dataset with LS, SC and SOC respectively. The script will automatically create a new directory \results_dtdb to save system matrices, training time and reconstruction error.
• Run kodex.m with MATLAB. It will learn the linear dynamics of KODex dataset with LS, SC and SOC respectively. The script will automatically create a new directory \results_kodex to save system matrices, and training time.