Skip to content

Latest commit

 

History

History
320 lines (266 loc) · 10.3 KB

README.md

File metadata and controls

320 lines (266 loc) · 10.3 KB

NLSolvers

Source Build Status
Source Build Status
Codecov branch Build Status

NLSolvers provides optimization, curve fitting, and equation solving functionalities for Julia. The goal is to provide a set of robust and flexible methods that run fast. Currently, the package does not try to implement any automatic generation of unspecified functions (gradients, Hessians, Hessian-vector products) using AD.

NLSolvers.jl uses different problem types for different problems

  • OptimizationProblem for optimization problems
  • NEqProblem for non-linear equations problems
  • FixedPointProblem for non-linear equations problems

Optimization Problems

Take the following scalar objective (with scalar input)

using NLSolvers
function objective(x)
    fx = x^4 + sin(x)
end
function gradient(∇f, x)
    ∇f = 4x^3 + cos(x)
    return ∇f
end
objective_gradient(∇f, x) = objective(x), gradient(∇f, x)
function hessian(∇²f, x)
    ∇²f = 12x^2 - sin(x)
    return ∇²f
end
function objective_gradient_hessian(∇f, ∇²f, x)
    f, ∇f = objective_gradient(∇f, x)
    ∇²f = hessian(∇²f, x)
    return f, ∇f, ∇²f
end
scalarobj = ScalarObjective(f=objective,
                            g=gradient,
                            fg=objective_gradient,
                            fgh=objective_gradient_hessian,
                            h=hessian)
optprob = OptimizationProblem(scalarobj; inplace=false) # scalar input, so not inplace

solve(optprob, 0.3, LineSearch(Newton()), OptimizationOptions())

With output

Results of minimization

* Algorithm:
  Newton's method with default linsolve with backtracking (no interp)

* Candidate solution:
  Final objective value:    -4.35e-01
  Final gradient norm:      3.07e-12

  Initial objective value:  3.04e-01
  Initial gradient norm:    1.06e+00

* Stopping criteria
  |x - x'|              = 6.39e-07 <= 0.00e+00 (false)
  |x - x'|/|x|          = 1.08e-06 <= 0.00e+00 (false)
  |f(x) - f(x')|        = 9.71e-13 <= 0.00e+00 (false)
  |f(x) - f(x')|/|f(x)| = 2.23e-12 <= 0.00e+00 (false)
  |g(x)|                = 3.07e-12 <= 1.00e-08 (true)
  |g(x)|/|g(x₀)|        = 2.88e-12 <= 0.00e+00 (false)

* Work counters
  Seconds run:   7.15e-06
  Iterations:    6

The problem types are especially useful when manifolds, bounds, and other constraints enter the picture.

Let's take the same problem as above but write it with arrays and mutating code style. The inplace keyword argument to the OptimizationProblem is used to apply the desired code paths internally. If set to true, cache arrays will be updated inplace and mutation is promised to be allowed for the input type(s). If set to false, no operations will mutate or happen in place.

using NLSolvers
function objective_ip(x)
    fx = x[1]^4 + sin(x[1])
end
function gradient_ip(∇f, x)
    ∇f[1] = 4x[1]^3 + cos(x[1])
    return ∇f
end
objective_gradient_ip(∇f, x) = objective_ip(x), gradient_ip(∇f, x)
function hessian_ip(∇²f, x)
    ∇²f[1] = 12x[1]^2 - sin(x[1])
    return ∇²f
end
function objective_gradient_hessian_ip(∇f, ∇²f, x)
    f, ∇f = objective_gradient_ip(∇f, x)
    ∇²f = hessian_ip(∇²f, x)
    return f, ∇f, ∇²f
end
scalarobj_ip = ScalarObjective(f=objective_ip,
                            g=gradient_ip,
                            fg=objective_gradient_ip,
                            fgh=objective_gradient_hessian_ip,
                            h=hessian_ip)
optprob_ip = OptimizationProblem(scalarobj_ip; inplace=true)

solve(optprob_ip, [0.3], LineSearch(Newton()), OptimizationOptions())

which gives

Results of minimization

* Algorithm:
  Newton's method with default linsolve with backtracking (no interp)

* Candidate solution:
  Final objective value:    -4.35e-01
  Final gradient norm:      3.07e-12

  Initial objective value:  3.04e-01
  Initial gradient norm:    1.06e+00

* Stopping criteria
  |x - x'|              = 6.39e-07 <= 0.00e+00 (false)
  |x - x'|/|x|          = 1.08e-06 <= 0.00e+00 (false)
  |f(x) - f(x')|        = 9.71e-13 <= 0.00e+00 (false)
  |f(x) - f(x')|/|f(x)| = 2.23e-12 <= 0.00e+00 (false)
  |g(x)|                = 3.07e-12 <= 1.00e-08 (true)
  |g(x)|/|g(x₀)|        = 2.88e-12 <= 0.00e+00 (false)

* Work counters
  Seconds run:   1.10e-05
  Iterations:    6

as above. Another set of examples could be SArray's and MArray's from the StaticArrays.jl package.

using StaticArrays
solve(optprob_ip, @MVector([0.3]), LineSearch(Newton()), OptimizationOptions())

which gives

Results of minimization

* Algorithm:
  Newton's method with default linsolve with backtracking (no interp)

* Candidate solution:
  Final objective value:    -4.35e-01
  Final gradient norm:      3.07e-12

  Initial objective value:  3.04e-01
  Initial gradient norm:    1.06e+00

* Stopping criteria
  |x - x'|              = 6.39e-07 <= 0.00e+00 (false)
  |x - x'|/|x|          = 1.08e-06 <= 0.00e+00 (false)
  |f(x) - f(x')|        = 9.71e-13 <= 0.00e+00 (false)
  |f(x) - f(x')|/|f(x)| = 2.23e-12 <= 0.00e+00 (false)
  |g(x)|                = 3.07e-12 <= 1.00e-08 (true)
  |g(x)|/|g(x₀)|        = 2.88e-12 <= 0.00e+00 (false)

* Work counters
  Seconds run:   5.68e-04
  Iterations:    6

So numbers, mutating array code and non-mutating array code is supported depending on the input to the problem type and initial x or state in general.

BigFloat eigen

To use the NWI algorithm in a TrustRegion algorithm, it is necessary to first using GenericLinearAlgebra.

Systems of Nonlinear Equations NEqProblem

To solve a system of non-linear equations you should use the NEqProblem type. First, we have to define a VectorObjective. We can try to solve for the roots in the problem defined by setting the gradient of the Rosenbrock test problem equal to zero.

function F_rosenbrock!(Fx, x)
    Fx[1] = 1 - x[1]
    Fx[2] = 10(x[2]-x[1]^2)
    return Fx
end
function J_rosenbrock!(Jx, x)
    Jx[1,1] = -1
    Jx[1,2] = 0
    Jx[2,1] = -20*x[1]
    Jx[2,2] = 10
    return Jx
end
function FJ_rosenbrock!(Fx, Jx, x)
    F_rosenbrock!(Fx, x)
    J_rosenbrock!(Jx, x)
    Fx, Jx
end
function Jvop_rosenbrock!(x)
    function JacV(Fv, v)
        Fv[1] = -1*v[1]
        Fv[2,] = -20*x[1]*v[1] + 10*v[2]
    end
    LinearMap(JacV, length(x))
end
vectorobj = NLSolvers.VectorObjective(F_rosenbrock!, J_rosenbrock!, FJ_rosenbrock!, Jvop_rosenbrock!)

and define a probem type that lets solve dispatch to the correct code

vectorprob = NEqProblem(vectorobj)

and we can solve using two variants of Newton's method. One that globalizes the solve using a trust-region based method and one that uses a line search


julia> solve(vectorprob, [5.0, 0.0], TrustRegion(Newton(), Dogleg()), NEqOptions())
Results of solving non-linear equations

* Algorithm:
  Newton's method with default linsolve with Dogleg{Nothing}

* Candidate solution:
  Final residual 2-norm:      5.24e-14
  Final residual Inf-norm:    5.24e-14

  Initial residual 2-norm:    6.25e+04
  Initial residual Inf-norm:  2.50e+02

* Stopping criteria
  |F(x')|               = 5.24e-14 <= 0.00e+00 (false)

* Work counters
  Seconds run:   1.91e-05
  Iterations:    2


julia> solve(vectorprob, [5.0, 0.0], LineSearch(Newton()), NEqOptions())
Results of solving non-linear equations

* Algorithm:
  Newton's method with default linsolve with backtracking (no interp)

* Candidate solution:
  Final residual 2-norm:      0.00e+00
  Final residual Inf-norm:    0.00e+00

  Initial residual 2-norm:    2.50e+02
  Initial residual Inf-norm:  2.50e+02

* Stopping criteria
  |F(x')|               = 0.00e+00 <= 0.00e+00 (true)

* Work counters
  Seconds run:   1.00e-05
  Iterations:    2

Univariate optimization

The only method that is exclusively univariate is Brent's method for function minimization BrentMin.

Multivariate optimization

Sampling based

  • SimulatedAnnealing
  • PureRandomSearch
  • ParticleSwarm

Direct search

  • NelderMead

Quasi-Newton Line search

  • DBFGS
  • BFGS
  • SR1
  • DFP
  • GradientDescent
  • LBFGS

Conjugate Gradient Descent

  • ConjugateGradient

Newton Line Search

  • Newton

Gradient based (no line search)

  • Adam
  • AdaMax

Acceleration methods

  • Anderson

Krylov

  • InexactNewton

BB style

  • BB
  • DFSANE

Simple bounds (box)

  • ParticleSwarm
  • ActiveBox

Custom solve

Newton methods generally accept a linsolve argument.

Preconditioning

Several methods accept nonlinear (left-)preconditioners. A preconditioner is provided as a function that has two methods: p(x) and p(x, P) where the first prepares and returns the preconditioner and the second is the signature for updating the preconditioner. If the preconditioner is constant, both method will simply return this preconditioner. A preconditioner is used in two contexts: in ldiv!(pgr, factorize(P), gr) that accepts a cache array for the preconditioned gradient pgr, the preconditioner P, and the gradient to be preconditioned gr, and in mul!(x, P, y). For the out-of-place methods (minimize as opposed to minimize!) it is sufficient to have \(P, gr) and *(P, y) defined.

Beware, chaotic gradient methods!

Some methods that might be labeled as acceleration, momentum, or spectral methods can exhibit chaotic behavior. Please keep this in mind if comparing things like DFSANE with similar implemenations in other software. It can give very different results given different compiler optimizations, CPU architectures, etc. See for example https://link.springer.com/article/10.1007/s10915-011-9521-3 .

TODO

Documented in each type's docstring including LineSearch, BFGS, ....

Abstract arrays!!! :| manifolds Use user norms Banded Jacobian AD nan hessian

line search should have a short curcuit for very small steps

Next steps

Mixed complementatiry SAMIN, BOXES Univariate!! IP Newotn Krylov Hessian