Engee documentation

SciPy.jl

The page is in the process of being translated.

SciPy is a mature Python library that offers a rich family of optimization, root—​finding and linear‐programming algorithms. OptimizationSciPy.jl gives access to these routines through the unified Optimization.jl interface just like any native Julia optimizer.

Note OptimizationSciPy.jl relies on PythonCall. A minimal Python distribution containing SciPy will be installed automatically on first use, so no manual Python set-up is required.

Installation: OptimizationSciPy.jl

import Pkg
Pkg.add("OptimizationSciPy")

Methods

Below is a catalogue of the solver families exposed by OptimizationSciPy.jl together with their convenience constructors. All of them accept the usual keyword arguments maxiters, maxtime, abstol, reltol, callback, progress in addition to any SciPy-specific options (passed verbatim via keyword arguments to solve).

Local Optimizer

Derivative-Free

  • ScipyNelderMead() — Simplex Nelder—​Mead algorithm

  • ScipyPowell() — Powell search along conjugate directions

  • ScipyCOBYLA() — Linear approximation of constraints (supports nonlinear constraints)

Gradient-Based

  • ScipyCG() — Non-linear conjugate gradient

  • ScipyBFGS() — Quasi-Newton BFGS

  • ScipyLBFGSB() — Limited-memory BFGS with simple bounds

  • ScipyNewtonCG() — Newton-conjugate gradient (requires Hessian-vector products)

  • ScipyTNC() — Truncated Newton with bounds

  • ScipySLSQP() — Sequential least-squares programming (supports constraints)

  • ScipyTrustConstr() — Trust-region method for non-linear constraints

Hessian—​Based / Trust-Region

  • ScipyDogleg(), ScipyTrustNCG(), ScipyTrustKrylov(), ScipyTrustExact() — Trust-region algorithms that optionally use or build Hessian information

Global Optimizer

  • ScipyDifferentialEvolution() — Differential evolution (requires bounds)

  • ScipyBasinhopping() — Basin-hopping with local search

  • ScipyDualAnnealing() — Dual annealing simulated annealing

  • ScipyShgo() — Simplicial homology global optimisation (supports constraints)

  • ScipyDirect() — Deterministic DIRECT algorithm (requires bounds)

  • ScipyBrute() — Brute-force grid search (requires bounds)

Linear & Mixed-Integer Programming

  • ScipyLinprog("highs") — LP solvers from the HiGHS project and legacy interior-point/simplex methods

  • ScipyMilp() — Mixed-integer linear programming via HiGHS branch-and-bound

Root Finding & Non-Linear Least Squares (experimental)

Support for ScipyRoot, ScipyRootScalar and ScipyLeastSquares is available behind the scenes and will be documented once the APIs stabilise.

Examples

Unconstrained minimisation

using Optimization, OptimizationSciPy

rosenbrock(x, p) = (p[1] - x[1])^2 + p[2] * (x[2] - x[1]^2)^2
x0 = zeros(2)
p = [1.0, 100.0]

f = OptimizationFunction(rosenbrock, Optimization.AutoZygote())
prob = OptimizationProblem(f, x0, p)

sol = solve(prob, ScipyBFGS())
@show sol.objective   # ≈ 0 at optimum
7.717288356613562e-13

Constrained optimisation with COBYLA

using Optimization, OptimizationSciPy

# Objective
obj(x, p) = (x[1] + x[2] - 1)^2

# Single non-linear constraint: x₁² + x₂² ≈ 1 (with small tolerance)
cons(res, x, p) = (res .= [x[1]^2 + x[2]^2 - 1.0])

x0 = [0.5, 0.5]
prob = OptimizationProblem(
    OptimizationFunction(obj; cons = cons),
    x0, nothing, lcons = [-1e-6], ucons = [1e-6])  # Small tolerance instead of exact equality

sol = solve(prob, ScipyCOBYLA())
@show sol.u, sol.objective
([0.9999995099640485, 1.1653740143742524e-5], 1.2462829129061485e-10)

Differential evolution (global) with custom options

using Optimization, OptimizationSciPy, Random, Statistics
Random.seed!(123)

ackley(x, p) = -20exp(-0.2*sqrt(mean(x .^ 2))) - exp(mean(cos.(2π .* x))) + 20 + ℯ
x0 = zeros(2)                    # initial guess is ignored by DE
prob = OptimizationProblem(ackley, x0; lb = [-5.0, -5.0], ub = [5.0, 5.0])

sol = solve(prob, ScipyDifferentialEvolution(); popsize = 20, mutation = (0.5, 1))
@show sol.objective
4.440892098500626e-16

Passing solver-specific options

Any keyword that Optimization.jl does not interpret is forwarded directly to SciPy. Refer to the SciPy optimisation API for the exhaustive list of options.

sol = solve(prob, ScipyTrustConstr(); verbose = 3, maxiter = 10_000)

Troubleshooting

The original Python result object is attached to the solution in the original field:

sol = solve(prob, ScipyBFGS())
println(sol.original)  # SciPy OptimizeResult

If SciPy raises an error it is re-thrown as a Julia ErrorException carrying the Python message, so look there first.

Contributing

Bug reports and feature requests are welcome in the Optimization.jl issue tracker. Pull requests that improve either the Julia wrapper or the documentation are highly appreciated.