ParOpt Overview
ParOpt is a parallel optimization library for large-scale optimization using distributed design variables. ParOpt is implemented in C++ with a Python wrapper generated using Cython. ParOpt contains both an interior-point optimization algorithm with a line search globalization strategy and an \(l_{\infty}\) trust region algorithm with an \(l_{1}\) penalty function.
ParOpt is designed to solve optimization problems that take the form:
where \(x\) is a vector of distributed design variables with lower and upper bounds \(l\) and \(u\). The constraints \(c(x)\), which we refer to as dense constraints, are a small number of global constraints. The constraints \(c_{w}(x)\) are sparse and separable, with the Jacobian \(A_{w}(x) = \nabla c_{w}(x)\). The separable constraints satisfy the property that the matrix \(A_{w} C A_{w}^{T}\) is block diagonal when \(C\) is a diagonal matrix.
Quick example
The following code is an example of ParOpt applied to the optimization problem:
import numpy as np
from mpi4py import MPI
from paropt import ParOpt
class Problem(ParOpt.Problem):
def __init__(self):
nvars = 2
ncon = 1
super(Problem, self).__init__(MPI.COMM_SELF, nvars, ncon)
def getVarsAndBounds(self, x, lb, ub):
"""Get the variable values and bounds"""
lb[:] = -3.0
ub[:] = 3.0
x[:] = -2.0
return
def evalObjCon(self, x):
"""Evaluate the objective and constraint values"""
fail = 0
fobj = -10*x[0]**2 + 10*x[1]**2 + 4*np.sin(x[0]*x[1]) - 2*x[0] + x[0]**4
cons = np.array([x[0] + x[1] - 0.5])
return fail, fobj, cons
def evalObjConGradient(self, x, g, A):
"""Evaluate the objective and constraint gradients"""
fail = 0
g[0] = -20*x[0] + 4*np.cos(x[0]*x[1])*x[1] - 2.0 + 4*x[0]**3
g[1] = 20*x[1] + 4*np.cos(x[0]*x[1])*x[0]
A[0][0] = 1.0
A[0][1] = 1.0
return fail
problem = Problem()
options = {}
opt = ParOpt.Optimizer(problem, options)
opt.optimize()
x, z, zw, zl, zw = opt.getOptimizedPoint()
Recommended settings
Often the default options work well. A complete description of the options is described here.
For the trust region algorithm, we recommend starting with the following set of parameters. These control the trust region algorithm and the interior point solution algorithm that solves the trust region subproblems. Different globalization strategies are implemented with the trust region method. These control how the \(\ell_{1}\) penalty parameters are updated. A filter globalization strategy is also implemented.
options = {
"algorithm": "tr",
"tr_output_file": "paropt.tr", # Trust region output file
"output_file": "paropt.out", # Interior point output file
"tr_max_iterations": 100, # Maximum number of trust region iterations
"tr_infeas_tol": 1e-6, # Feasibility tolerace
"tr_l1_tol": 1e-5, # l1 norm for the KKT conditions
"tr_linfty_tol": 1e-5, # l-infinity norm for the KKT conditions
"tr_init_size": 0.05, # Initial trust region radius
"tr_min_size": 1e-6, # Minimum trust region radius size
"tr_max_size": 10.0, # Max trust region radius size
"tr_eta": 0.25, # Trust region step acceptance ratio
"tr_adaptive_gamma_update": True, # Use an adaptive update strategy for the penalty
"max_major_iters": 100, # Maximum number of iterations for the IP subproblem solver
"qn_subspace_size": 10, # Subspace size for the quasi-Newton method
"qn_type": "bfgs", # Type of quasi-Newton Hessian approximation
"abs_res_tol": 1e-8, # Tolerance for the subproblem
"starting_point_strategy": "affine_step", # Starting point strategy for the IP
"barrier_strategy": "mehrotra", # Barrier strategy for the IP
"use_line_search": False, # Don't useline searches for the subproblem
}
For MMA, we recommend starting with the settings shown below. These settings set the primary MMA parameters as well as several key interior point parameters that control the manner in which the subproblems are solved. Options that begin with mma_ control the output file name, maximum iterations, feasibility and optimality tolerances for the outer iterations. The options max_major_iters, abs_res_tol, starting_point_strategy, barrier_strategy and use_line_search all are passed to the interior point method. These suggested settings are selected to solve each subproblem to a tight tolerance.
options = {
"algorithm": "mma",
"mma_output_file": "paropt.mma", # MMA output file name
"output_file": "paropt.out", # Interior point output file
"mma_max_iterations": 100, # Maximum number of iterations for MMA
"mma_infeas_tol": 1e-6, # Feasibility tolerance for MMA
"mma_l1_tol": 1e-5, # l1 tolerance on the on the KKT conditions for MMA
"mma_linfty_tol" : 1e-5, # l-infinity tolerance on the KKT conditions for MMA
"max_major_iters": 100, # Max iterations for each subproblem
"abs_res_tol": 1e-8, # Tolerance for each subproblem
"starting_point_strategy": "affine_step", # IP initialization strategy
"barrier_strategy": "mehrotra", # IP barrier strategy
"use_line_search": False, # Don't use line searches on the subproblem
}
Please cite us
The key contributions and development of ParOpt are described in the following paper: Ting Wei Chin, Mark K. Leader, Graeme J. Kennedy, A scalable framework for large-scale 3D multimaterial topology optimization with octree-based mesh adaptation, Advances in Engineering Software, Volume 135, 2019.
@article{Chin:2019,
title = {A scalable framework for large-scale 3D multimaterial topology optimization with octree-based mesh adaptation},
journal = {Advances in Engineering Software},
volume = {135},
year = {2019},
doi = {10.1016/j.advengsoft.2019.05.004},
author = {Ting Wei Chin and Mark K. Leader and Graeme J. Kennedy}}
Installation
The installation process for ParOpt consists of first compiling the C++ source and then installing the python interface. The C++ source requires a BLAS/LAPACK installation, an MPI compiler for C++ and [Metis](https://github.com/KarypisLab/METIS). Using the Python interface additional requires numpy, mpi4py and Cython.
Currently ParOpt does not use an build system so all options are set in the file Makefile.in.
Default settings for these options are contained within the file Makefile.in.info`.
The first step in compling ParOpt is to first copy the file Makefile.in.info to Makefile.in.
Key parameters that should be set within the Makefile.in file are:
PAROPT_DIRthe root ParOpt directoryCXXthe MPI-enabled C++ compilerLAPACK_LIBSthe link command for the BLAS/LAPACK librariesMETIS_INCLUDEandMETIS_LIBpoint to the Metis include directory and the static library for Metis.
By default, the shared and static library are complied to the directory PAROPT_DIR/lib.
Installation of the python interface is performed using pip install -e .\[all\].