TRAFS: A Nonsmooth Convex Optimization Algorithm with $\mathcal{O}\left(\frac{1}{\epsilon}\right)$ Iteration Complexity
This repository is the implementation of the TRAFS algorithm, a nonsmooth convex optimization algorithm that both provides better guarantees and converges significantly faster in numerical experiments compared to previous methods. Read our paper for more details.
The program has only been tested on Linux.
Make sure there is a recent C++ compiler, and gfortran is available. Install
julia to the system. Setup a python virtual env. In the virtual env, run pip install scipy numpy attrs clarabel mosek cython julia. Then use
./setup-deps.sh to compile and install other dependencies.
Use ./run_bench.sh to reproduce all the numerical experiments reported in the
paper. After the script finishes, use ./make_bench_table.py to generate the
LaTex tables summarizing the results.
Use run_*.py to run individual experiment.
All the optimization algorithms are in the opt module. The benchmark problems are in the prob module. The code should be self-explanatory.