A PYTHONlibrary for optimizing CUTEST Unconstrained Problem and/or Empirical Risk Minimization(includingDeep Neural Networks).

[TOC]

---

Author	Xi He
Email	[email protected]

Introduction

The library contains a collection of first/second-order optimization algorithms for standard unconstrained optimization problem set CUTEST and/or Empirical Risk Minimization.

Dependencies

• numpy
• pytorch
• cutest (see cutest installation)

Tour of the UCLibrary

Here we exemplify a instance as to solve a problem in Cutest problem set.

Single Run

import ...
config.read('config.ini')
problem = CutestProblem('ROSENBR')
demo = Demo(problem)

problem.setInitialPoint()
optim = optimizers.Cubic(problem, lr=False, mode='exact', adaptive=True)

for _ in range(config.max_iters):
    optim.step()
    if optim.terminationCondition(mode='first_order', tol=config.tol):
        break

demo.addContourTrace(optim)
demo.drawPerformancePlot(optim)
demo.showPlot()
------------------
Default initial point by Cutest...
Cubic-exact-adaptive
('cholesky_decomp_counter', 132)
('cholesky_linear_solver_counter', 292)
('hess_counter', 30)
('step_counter', 30)
('grad_counter', 30)

Multiple Run

import ...

def run(problem, optim, demo):
    for _ in range(config.max_iters):
        optim.step()
        if optim.terminationCondition(mode='first_order', tol=config.tol):
            break
    demo.addContourTrace(optim)
    demo.drawPerformancePlot(optim)

config.read('config.ini')
problem = CutestProblem('ROSENBR')
problem.setInitialPoint()

demo = Demo(problem)

optim = optimizers.Cubic(problem, lr=False, mode='exact', adaptive=True)
run(problem, optim, demo)

optim = optimizers.TrustRegion(problem, mode='cauchy')
run(problem, optim, demo)

demo.showPlot()

List of Main Modules

Gradient Based Solver, optimizer.GD

• Vanilla GD
• Nesterov's acceleration
• Heavy Ball acceleration
• Dynamic momentum
• Optimal momentum
• Static momentum
• Restart scheme

Practical Curvature based, optimizer.Newton

• Constant damping
• Levenberg–Marquardt damping
• Truncated hessian

Trust Region, optimizer.TrustRegion

• Vanilla TR

Cubic Regularization, optimizer.Cubic

• Vanilla CR
• Adaptive CR
• CRm (CR with momentum)

Line Search, LineSearch

• Backtracking (Amijo linesearch)
• Strong-wolfe

Subsolvers, SubRoutine

• Exact solver for positive definite matrix
• Cauchy point
• Dog-leg
• CG for positive positive definite matrix
• Steinghaug-Toint CG
• Generalized Lanczos trust region
• exact tridiagonal matrix subsolver
• exact regularized subproblem solver
• AdaNT