# Oracle-Structured Bundle Distributed Optimization (OSBDO)

This repository accompanies the manuscript Implementation of an Oracle-Structured Bundle Method for Distributed Optimization.

We consider the optimization problem math \begin{array}{ll} \mbox{minimize} & h(x) = f(x) + g(x), \end{array} where $x \in {\mbox{\bf R}}^n$ is the variable, and $f, g:{\mbox{\bf R}}^n \to {\mbox{\bf R}}\cup {\infty}$ are the oracle and structured convex objective functions, respectively.

The oracle objective function $f$ is block separable, i.e., is of the form math f(x) = \sum_{i=1}^M f_i(x_i), where $xi \in {\mbox{\bf R}}^{ni}$ and $x=(x1, \ldots, xM)$. We refer to $xi$ as the public variable and $fi$ as the objective function of the agent $i$. Our access to $fi$ is only via an oracle that evaluates the function value and a subgradient at a given point $xi$, i.e., $fi(xi)$ and $qi \in \partial fi(x_i)$.

The function $g$ is structured in the sense that we are given its complete description in disciplined convex programming (DCP) format. Presumably the function $g$ couples the block variables $x1, \ldots, xM$.

In this repository we provide an implementation of the bundle-type method proposed in our manuscript.

Installation

OSBDO is available on the Python Package Index, use pip install osbdo Requirements * python >= 3.8 * CVXPY >= 1.2.0 * numpy >= 1.22.2 * matplotlib >= 1.16.0 * scipy >= 1.8.0

Getting started

To start using osbdo solver, follow the procedure below.

1. Describe each objective function $f_i$ in a class that inherits from osbdo.Agent

   • Agent_i(osbdo.Agent)
   • create dictionary with parameters params relevant for function $f_i$
     • dictionary params contains items with the dimension of public variable $x_i$ and its lower and upper bound
       • params = {"dimension"= ..., "lower_bound"=..., "upper_bound":..., }
   • implement methods
     • Agent_i.query(v): returns output of the subgradient oracle at point $v$ as a Point(v, q, f_i(v))
     • Agent_i._construct_params(): construct necessary parameters for $f_i$ from params
     • Agent_i.get_init_minorant(): returns initial minorant of agent's objective $\hat f^0_i$ as a cvxpy.Expression

2. Define a structured function $g$ as an osbdo.Coupling(agents, function, domain) by specifying

   • agents: list of $M$ agents of type Agent_i
   • function: function $g$ on its domain, given as a cvxpy.Expression
   • domain: domain of $g$ given as a list of cvxpy constraints

3. Define a distributed optimization problem as an osbdo.Problem(agents, g) by specifying

   • agents: list of $M$ agents of type Agent_i
   • g: a structured function $g$ of type Coupling

4. Solve a distributed optimization problem

   • osbdo.Problem.solve()
     • memory is an optional parameter that limits the memory (set to infinity by default); see the manuscript
   • osbdo.Problem.upper_bnd, osbdo.Problem.lower_bnd: upper and lower bounds on optimal problem value in each iteration, populated after calling the Problem.solve()

Hello world

We provide a guideline on how to use our method using the hello world example Jupyter notebook.

Example notebooks

We have example notebooks that show how to use our method on a number of different problems.

• supply chain problem
  
• resource allocation problem
• multi-commodity flow problem
• federated learning problem

Please consult our manuscript for the details of mentioned problems and their oracle-structured form.

Extra example

We use our method for finding the intersection of the convex sets. * intersection of convex sets

We also check the performance of OSBDO when applied to the action directed Walrasian equilibrium (over the primal variables) and price directed Walrasian equilibrium (over the dual variable). * Walrasian equilibrium

Citing

If you use OSBDO please cite the associated paper. @article{parshakova2022oraclestructured, title={An Oracle-Structured Bundle Method for Distributed Optimization}, author={Tetiana Parshakova and Fangzhao Zhang and Stephen Boyd}, year={2022}, eprint={2211.01418}, archivePrefix={arXiv}, primaryClass={math.OC} }