Skip to content

Alirezalm/SCOT

Repository files navigation

Sparse Convex Optimization Toolkit (SCOT)

SCOT is an open-source distributed optimization solver designed to tackle Sparse Convex Optimization (SCO) problems within computational networks.

Overview

SCOT can be employed both as a standalone executable application and as a Python API known as SCOTPY, offering flexibility in integrating it into your projects.

Key Features

  • Solve Sparse Convex Optimization problems with ease.
  • Utilize SCOT via its Python API, SCOTPY.
  • Distribute optimization tasks efficiently across computational networks.
  • Compatible with various optimization algorithms for diverse problem domains.
  • Easily customizable and extensible for specific use cases.

Installation

See INSTALL file.

Usage

SCOT MPI Command-Line Interface (SMCLI) can be employed to utilize SCOT as a solver, provided appropriate input files in .dist.json format are available.

Example: Distributed Sparse Logistic Regression (DSLogR)

Here's how to solve a DSLogR problem using the SMCLI interface:

  1. Prepare problem data for each node in JSON format. Assuming two nodes, name the JSON files following this convention:
node_{mpi_rank}_{problem_name}.dist.json

For two nodes and the problem name logistic_regression, the files should be named as:

1. node_{0}_{logistic_regression}.dist.json
2. node_{1}_{logistic_regression}.dist.json

The exact JSON format for these files can be found in the data folder.

  1. Execute SCOT with default settings:
mpirun -n 2 ./bin/scot --dir=/path/to/.dist.json/files --input=logistic_regression --nz=2 
  1. After successful execution, two output files will be created in the scot_framework folder:
1. rank_0_output.json
2. rank_1_output.json

These files contain essential information about the solution to the given problem.

Numerical Benchmarks

Small scale instances

Alt Text

Medium scale instances

Alt Text

Large scale instances

Alt Text

Publications

  1. A. Olama, A Distributed Framework for Sparse Convex Optimization: Algorithms and Tools. PhD thesis, Federal University of Santa Catarina (UFSC), Brazil, 2023.
  2. A. Olama, E. Camponogara, and J. Kronqvist, Sparse convex optimization toolkit: a mixed-integer framework, Optimization Methods and Software, pp. 1–27, 2023.
  3. Olama, E. Camponogara, and P. R. Mendes, Distributed primal outer approximation algorithm for sparse convex programming with separable structures, Journal of Global Optimization, vol. 86, no. 3, pp. 637–670, 2023.
  4. A. Olama, N. Bastianello, P. R. Mendes, and E. Camponogara, Relaxed hybrid consensus ADMM for distributed convex optimisation with coupling constraints, IET Control Theory & Applications, vol. 13, no. 17, pp. 2828–2837, 2019.

Acknowledgement

This work is part of the project “Distributed Optimization for Cooperative Machine Learning in Complex Networks” (No PGR10067) which has received funding from Fundacao de Amparo a Pesquisa e Inovacao do Estado de Santa Catarina (FAPESC), in Brazil, under grant 2021TR2265 and the Ministero degli Affari Esteri e della Cooperazione Internazionale (MAECI), in Italy. Technical support from Digital Futures at KTH and C3.ai Digital Transformation Institute for the project "AI Techniques for Power Systems Under Cyberattacks" is also gratefully acknowledged.