Personalized Federated Learning for Inverse LQR

Written as part of a PhD research rotation in Winter 2021 as a part of the Autonomous Systems Lab under Professor Marco Pavone

Paper Link: https://drive.google.com/file/d/1GvfAKxxFkInUe_8MVJKWhXGdy0RAgIo9/view?usp=sharing

Repo Structure

1. fedadmm_*.py: Code to run a personalized federated learning
2. ADMM_*.py: Helper repos that run various versions of ADMM
3. utils.py: Helper scripts (mostly latexify() for nice figure generation)
4. tests\ repo
   • lqr_*.py: Test code to see how an LQR controller behaves given different configurations. Replaced by the function simulate() in fedadmm_*.py
   • test_*.py: Various test scripts I wrote over the course of the quarter to test different Python modules or configurations
   • _*.m: More test scripts, except in MATLAB
5. generate_figures repo: The scripts used to generate the paper's figures
6. data repo: Where all data from fedadmm_*.py scripts are stored
7. figures repo: Where all figures from fedadmm_*.py scripts are stored
8. figures_final repo: Where the paper's figures are stored

Note that the data and figures folders are empty right now. To access the actual data and figures I amassed over my quarter, an interested reader can revert back to the tagged commit final_writeup to download all the files.

Installation Instructions

Unfortunately, installation instructions for this repo are kinda a pain. Sorry! The main reason is that CVXPY's SCS solver implementation doesn't work for matrices larger than 2x2 (as of 3/22/2021) unless you have it installed with blas+lapack, but you can only install that if you have numpy with the mkl extension installed.

I used Python 3.7 in an Anaconda virtual environment, if that helps.

Install the following packages (in roughly the following order) from this link (or any alternative link that kindly provides binaries of these libraries for your given Python version: https://www.lfd.uci.edu/~gohlke/pythonlibs/)

• numpy+mkl
• scipy
• scs
• cvxopt
• ecos
• osqp
• cvxpy
• matplotlib
• slycot (you don't need it for this library, but it's super common for other controls packages and this is the only way I've gotten it to work on Windows. Life hacks!)

Instructions to Run Code

The default script I used for the paper is fedadmm.py, and the default ADMM script that I used was ADMM_slack.py

The installation was the hard part. Now you should be able to just run python fedadmm.py with your Conda environment active.