opt.+opt.+de

class opt.de.NumpyRandomState

Bases: handle

NumpyRandomState Bit-exact port of the subset of numpy.random.RandomState consumed by scipy’s differential_evolution. Backed by opt.de.MT19937.

Implements random_sample, uniform, the default-dtype randint (32-bit masked rejection), and shuffle/permutation (Fisher-Yates driven by random_interval), each reproducing numpy’s exact draw sequence and word consumption, so the optimizer follows the identical trajectory to the native-Python line-opt for a given integer seed.

Constructor Summary

NumpyRandomState(seed)

Property Summary

gen: opt.de.MT19937 handle

Method Summary

fillMask(~, v)

permutation(n)

randint(low, high)

randomInterval(maxv)

randomSample()

randomSampleN(n)

shuffle(arr)

uniformN(low, high, n)

uniformScalar(low, high)

class opt.de.MT19937

Bases: handle

MT19937 Bit-exact reimplementation of numpy’s legacy MT19937 core, matching the generator underlying numpy.random.RandomState.

Reproduces numpy’s exact 32-bit output stream so that NumpyRandomState and the differential-evolution optimizer generate identical draws to the native-Python line-opt for a given seed. Seeding follows numpy’s _legacy_seeding: a scalar seed that fits in 32 bits uses init_genrand; otherwise the seed words are fed to init_by_array.

Constructor Summary

MT19937(seed)

Property Summary

mt: 1x624 uint32 state

mti: scalar position (numpy get_state pos)

Method Summary

getStateKey()

initByArray(initKey)

initGenrand(s)

nextUint32()

seed(s)

class opt.de.DifferentialEvolution

Bases: handle

DifferentialEvolution Self-contained port of scipy’s differential_evolution that reproduces its trajectory bit-for-bit for a given integer seed, using opt.de.NumpyRandomState for all draws.

Covers the configuration used by line-opt: binomial strategies (default best1bin), updating=’immediate’, dithered mutation, latinhypercube initialization, polish=false, and penalty-based constraints handled inside the objective. The exact numpy draw order is preserved: LHS init (uniform grid then per-column permutation), a per-generation dither uniform, and per candidate randint (fill point), shuffle (sample selection) and uniform (crossover), plus a data-dependent uniform in the bound-repair step.

Constructor Summary

DifferentialEvolution(objective, low, high, strategy, popsizeMult, maxiter, ditherLow, ditherHigh, recombination, tol, seed)

Property Summary

atol

bestPerGen: cell array of 1xd vectors

callback: function handle (bestX, nit) -> logical stop, or []

ditherHigh

ditherLow

energies: 1 x numMembers

high: 1xd upper bounds

low: 1xd lower bounds

maxiter

nfev

numMembers

objective

energy = f(x)

Type:: function handle

paramCount

popsizeMult

population: numMembers x d, scaled to [0,1]

randomIndex: persistent shuffle buffer (0-based values)

recombination

rng: opt.de.NumpyRandomState

scale

strategy

tol

Method Summary

best()

bprime(candidate, s): candidate and s are 0-based member indices; convert to 1-based rows

calculateInitialEnergies()

converged()

ensureConstraint(trial)

initPopulationLhs()

mutate(candidate): candidate is 0-based member index

next()

promoteLowestEnergy()

scaleArg1(j)

scaleArg2(j)

scaleParameters(trial)

selectSamples(candidate, numberSamples): candidate is 0-based member index

solve()