Hide keyboard shortcuts

Hot-keys on this page

r m x p   toggle line displays

j k   next/prev highlighted chunk

0   (zero) top of page

1   (one) first highlighted chunk

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66

67

68

69

70

71

72

73

74

75

76

77

78

79

80

81

82

83

84

85

86

87

88

89

90

91

92

93

94

95

96

97

98

99

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

360

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375

376

377

378

379

380

381

382

383

384

385

386

387

388

389

390

391

392

393

394

395

396

397

398

399

400

401

402

403

404

405

406

407

408

409

410

411

412

413

414

415

416

417

418

419

420

421

422

423

424

425

426

427

428

429

430

431

432

433

434

435

436

437

438

439

440

441

442

443

444

445

446

447

448

449

450

451

452

453

454

455

456

457

458

459

460

461

462

463

464

465

466

467

468

469

470

471

472

473

474

475

476

477

478

479

480

481

482

483

484

485

486

487

488

489

490

491

492

493

494

495

496

497

498

499

500

501

502

503

504

505

506

507

508

509

510

511

512

513

514

515

516

517

518

519

520

521

522

523

524

525

526

527

528

529

530

531

532

533

534

535

536

537

538

539

540

541

542

543

544

545

546

547

548

549

550

551

552

553

554

555

556

557

558

559

560

561

562

563

564

565

566

567

568

569

570

571

572

573

574

575

576

577

578

579

580

581

582

583

584

585

586

587

588

589

590

591

592

593

594

595

596

597

598

599

600

601

602

603

604

605

606

607

608

609

610

611

612

613

614

615

616

617

618

619

620

621

622

623

624

625

626

627

628

629

630

631

632

633

634

635

636

637

638

639

640

641

642

643

644

645

646

647

648

649

650

651

652

653

654

655

656

657

658

659

660

661

662

663

664

665

666

667

668

669

670

671

672

673

674

675

676

677

678

679

680

681

682

683

684

685

686

687

688

689

690

691

692

693

694

695

696

697

698

699

700

701

702

703

704

705

706

707

708

709

710

711

712

713

714

715

716

717

718

719

720

721

722

723

724

725

726

727

728

729

730

731

732

733

734

735

736

737

738

739

740

741

742

743

744

745

746

747

748

749

750

751

752

753

754

755

756

757

758

759

760

761

762

763

764

765

766

767

768

769

770

771

772

773

774

775

776

777

778

779

780

781

782

783

784

785

786

787

788

789

790

791

792

793

794

795

796

797

798

799

800

801

802

803

804

805

806

807

808

809

810

811

812

813

814

815

816

817

818

819

820

821

822

823

824

825

826

827

828

829

830

831

832

833

834

835

836

837

838

839

840

841

842

843

844

845

846

847

848

849

850

851

852

853

854

855

856

857

858

859

860

861

862

863

864

865

866

867

868

869

870

871

872

873

874

875

876

877

878

879

880

881

882

883

884

885

886

887

888

889

890

891

892

893

894

895

896

897

898

899

900

901

902

903

904

905

906

907

908

909

910

911

912

913

914

915

916

917

918

919

920

921

922

923

924

925

926

927

928

929

930

931

932

933

934

935

936

937

938

939

940

941

942

943

944

945

946

947

948

949

950

951

952

953

954

955

956

957

958

959

960

961

962

963

964

965

966

967

968

969

970

971

972

973

974

975

976

977

978

979

980

981

982

983

984

985

986

987

988

989

990

991

992

993

994

995

996

997

998

999

1000

1001

1002

1003

1004

1005

1006

1007

1008

1009

1010

1011

1012

1013

1014

1015

1016

1017

1018

1019

1020

1021

1022

1023

1024

1025

1026

1027

1028

""" 

differential_evolution: The differential evolution global optimization algorithm 

Added by Andrew Nelson 2014 

""" 

from __future__ import division, print_function, absolute_import 

import warnings 

 

import numpy as np 

from scipy.optimize import OptimizeResult, minimize 

from scipy.optimize.optimize import _status_message 

from scipy._lib._util import check_random_state, MapWrapper 

from scipy._lib.six import xrange, string_types 

from scipy.optimize._constraints import (Bounds, new_bounds_to_old) 

 

 

__all__ = ['differential_evolution'] 

 

_MACHEPS = np.finfo(np.float64).eps 

 

 

def differential_evolution(func, bounds, args=(), strategy='best1bin', 

maxiter=1000, popsize=15, tol=0.01, 

mutation=(0.5, 1), recombination=0.7, seed=None, 

callback=None, disp=False, polish=True, 

init='latinhypercube', atol=0, updating='immediate', 

workers=1): 

"""Finds the global minimum of a multivariate function. 

 

Differential Evolution is stochastic in nature (does not use gradient 

methods) to find the minimium, and can search large areas of candidate 

space, but often requires larger numbers of function evaluations than 

conventional gradient based techniques. 

 

The algorithm is due to Storn and Price [1]_. 

 

Parameters 

---------- 

func : callable 

The objective function to be minimized. Must be in the form 

``f(x, *args)``, where ``x`` is the argument in the form of a 1-D array 

and ``args`` is a tuple of any additional fixed parameters needed to 

completely specify the function. 

bounds : sequence or `Bounds`, optional 

Bounds for variables. There are two ways to specify the bounds: 

1. Instance of `Bounds` class. 

2. ``(min, max)`` pairs for each element in ``x``, defining the finite 

lower and upper bounds for the optimizing argument of `func`. It is 

required to have ``len(bounds) == len(x)``. ``len(bounds)`` is used 

to determine the number of parameters in ``x``. 

args : tuple, optional 

Any additional fixed parameters needed to 

completely specify the objective function. 

strategy : str, optional 

The differential evolution strategy to use. Should be one of: 

 

- 'best1bin' 

- 'best1exp' 

- 'rand1exp' 

- 'randtobest1exp' 

- 'currenttobest1exp' 

- 'best2exp' 

- 'rand2exp' 

- 'randtobest1bin' 

- 'currenttobest1bin' 

- 'best2bin' 

- 'rand2bin' 

- 'rand1bin' 

 

The default is 'best1bin'. 

maxiter : int, optional 

The maximum number of generations over which the entire population is 

evolved. The maximum number of function evaluations (with no polishing) 

is: ``(maxiter + 1) * popsize * len(x)`` 

popsize : int, optional 

A multiplier for setting the total population size. The population has 

``popsize * len(x)`` individuals (unless the initial population is 

supplied via the `init` keyword). 

tol : float, optional 

Relative tolerance for convergence, the solving stops when 

``np.std(pop) <= atol + tol * np.abs(np.mean(population_energies))``, 

where and `atol` and `tol` are the absolute and relative tolerance 

respectively. 

mutation : float or tuple(float, float), optional 

The mutation constant. In the literature this is also known as 

differential weight, being denoted by F. 

If specified as a float it should be in the range [0, 2]. 

If specified as a tuple ``(min, max)`` dithering is employed. Dithering 

randomly changes the mutation constant on a generation by generation 

basis. The mutation constant for that generation is taken from 

``U[min, max)``. Dithering can help speed convergence significantly. 

Increasing the mutation constant increases the search radius, but will 

slow down convergence. 

recombination : float, optional 

The recombination constant, should be in the range [0, 1]. In the 

literature this is also known as the crossover probability, being 

denoted by CR. Increasing this value allows a larger number of mutants 

to progress into the next generation, but at the risk of population 

stability. 

seed : int or `np.random.RandomState`, optional 

If `seed` is not specified the `np.RandomState` singleton is used. 

If `seed` is an int, a new `np.random.RandomState` instance is used, 

seeded with seed. 

If `seed` is already a `np.random.RandomState instance`, then that 

`np.random.RandomState` instance is used. 

Specify `seed` for repeatable minimizations. 

disp : bool, optional 

Display status messages 

callback : callable, `callback(xk, convergence=val)`, optional 

A function to follow the progress of the minimization. ``xk`` is 

the current value of ``x0``. ``val`` represents the fractional 

value of the population convergence. When ``val`` is greater than one 

the function halts. If callback returns `True`, then the minimization 

is halted (any polishing is still carried out). 

polish : bool, optional 

If True (default), then `scipy.optimize.minimize` with the `L-BFGS-B` 

method is used to polish the best population member at the end, which 

can improve the minimization slightly. 

init : str or array-like, optional 

Specify which type of population initialization is performed. Should be 

one of: 

 

- 'latinhypercube' 

- 'random' 

- array specifying the initial population. The array should have 

shape ``(M, len(x))``, where len(x) is the number of parameters. 

`init` is clipped to `bounds` before use. 

 

The default is 'latinhypercube'. Latin Hypercube sampling tries to 

maximize coverage of the available parameter space. 'random' 

initializes the population randomly - this has the drawback that 

clustering can occur, preventing the whole of parameter space being 

covered. Use of an array to specify a population subset could be used, 

for example, to create a tight bunch of initial guesses in an location 

where the solution is known to exist, thereby reducing time for 

convergence. 

atol : float, optional 

Absolute tolerance for convergence, the solving stops when 

``np.std(pop) <= atol + tol * np.abs(np.mean(population_energies))``, 

where and `atol` and `tol` are the absolute and relative tolerance 

respectively. 

updating : {'immediate', 'deferred'}, optional 

If ``'immediate'``, the best solution vector is continuously updated 

within a single generation [4]_. This can lead to faster convergence as 

trial vectors can take advantage of continuous improvements in the best 

solution. 

With ``'deferred'``, the best solution vector is updated once per 

generation. Only ``'deferred'`` is compatible with parallelization, and 

the `workers` keyword can over-ride this option. 

 

.. versionadded:: 1.2.0 

 

workers : int or map-like callable, optional 

If `workers` is an int the population is subdivided into `workers` 

sections and evaluated in parallel 

(uses `multiprocessing.Pool <multiprocessing>`). 

Supply -1 to use all available CPU cores. 

Alternatively supply a map-like callable, such as 

`multiprocessing.Pool.map` for evaluating the population in parallel. 

This evaluation is carried out as ``workers(func, iterable)``. 

This option will override the `updating` keyword to 

``updating='deferred'`` if ``workers != 1``. 

Requires that `func` be pickleable. 

 

.. versionadded:: 1.2.0 

 

Returns 

------- 

res : OptimizeResult 

The optimization result represented as a `OptimizeResult` object. 

Important attributes are: ``x`` the solution array, ``success`` a 

Boolean flag indicating if the optimizer exited successfully and 

``message`` which describes the cause of the termination. See 

`OptimizeResult` for a description of other attributes. If `polish` 

was employed, and a lower minimum was obtained by the polishing, then 

OptimizeResult also contains the ``jac`` attribute. 

 

Notes 

----- 

Differential evolution is a stochastic population based method that is 

useful for global optimization problems. At each pass through the population 

the algorithm mutates each candidate solution by mixing with other candidate 

solutions to create a trial candidate. There are several strategies [2]_ for 

creating trial candidates, which suit some problems more than others. The 

'best1bin' strategy is a good starting point for many systems. In this 

strategy two members of the population are randomly chosen. Their difference 

is used to mutate the best member (the `best` in `best1bin`), :math:`b_0`, 

so far: 

 

.. math:: 

 

b' = b_0 + mutation * (population[rand0] - population[rand1]) 

 

A trial vector is then constructed. Starting with a randomly chosen 'i'th 

parameter the trial is sequentially filled (in modulo) with parameters from 

``b'`` or the original candidate. The choice of whether to use ``b'`` or the 

original candidate is made with a binomial distribution (the 'bin' in 

'best1bin') - a random number in [0, 1) is generated. If this number is 

less than the `recombination` constant then the parameter is loaded from 

``b'``, otherwise it is loaded from the original candidate. The final 

parameter is always loaded from ``b'``. Once the trial candidate is built 

its fitness is assessed. If the trial is better than the original candidate 

then it takes its place. If it is also better than the best overall 

candidate it also replaces that. 

To improve your chances of finding a global minimum use higher `popsize` 

values, with higher `mutation` and (dithering), but lower `recombination` 

values. This has the effect of widening the search radius, but slowing 

convergence. 

By default the best solution vector is updated continuously within a single 

iteration (``updating='immediate'``). This is a modification [4]_ of the 

original differential evolution algorithm which can lead to faster 

convergence as trial vectors can immediately benefit from improved 

solutions. To use the original Storn and Price behaviour, updating the best 

solution once per iteration, set ``updating='deferred'``. 

 

.. versionadded:: 0.15.0 

 

Examples 

-------- 

Let us consider the problem of minimizing the Rosenbrock function. This 

function is implemented in `rosen` in `scipy.optimize`. 

 

>>> from scipy.optimize import rosen, differential_evolution 

>>> bounds = [(0,2), (0, 2), (0, 2), (0, 2), (0, 2)] 

>>> result = differential_evolution(rosen, bounds) 

>>> result.x, result.fun 

(array([1., 1., 1., 1., 1.]), 1.9216496320061384e-19) 

 

Now repeat, but with parallelization. 

 

>>> bounds = [(0,2), (0, 2), (0, 2), (0, 2), (0, 2)] 

>>> result = differential_evolution(rosen, bounds, updating='deferred', 

... workers=2) 

>>> result.x, result.fun 

(array([1., 1., 1., 1., 1.]), 1.9216496320061384e-19) 

 

Next find the minimum of the Ackley function 

(https://en.wikipedia.org/wiki/Test_functions_for_optimization). 

 

>>> from scipy.optimize import differential_evolution 

>>> import numpy as np 

>>> def ackley(x): 

... arg1 = -0.2 * np.sqrt(0.5 * (x[0] ** 2 + x[1] ** 2)) 

... arg2 = 0.5 * (np.cos(2. * np.pi * x[0]) + np.cos(2. * np.pi * x[1])) 

... return -20. * np.exp(arg1) - np.exp(arg2) + 20. + np.e 

>>> bounds = [(-5, 5), (-5, 5)] 

>>> result = differential_evolution(ackley, bounds) 

>>> result.x, result.fun 

(array([ 0., 0.]), 4.4408920985006262e-16) 

 

References 

---------- 

.. [1] Storn, R and Price, K, Differential Evolution - a Simple and 

Efficient Heuristic for Global Optimization over Continuous Spaces, 

Journal of Global Optimization, 1997, 11, 341 - 359. 

.. [2] http://www1.icsi.berkeley.edu/~storn/code.html 

.. [3] http://en.wikipedia.org/wiki/Differential_evolution 

.. [4] Wormington, M., Panaccione, C., Matney, K. M., Bowen, D. K., - 

Characterization of structures from X-ray scattering data using 

genetic algorithms, Phil. Trans. R. Soc. Lond. A, 1999, 357, 

2827-2848 

""" 

 

# using a context manager means that any created Pool objects are 

# cleared up. 

with DifferentialEvolutionSolver(func, bounds, args=args, 

strategy=strategy, 

maxiter=maxiter, 

popsize=popsize, tol=tol, 

mutation=mutation, 

recombination=recombination, 

seed=seed, polish=polish, 

callback=callback, 

disp=disp, init=init, atol=atol, 

updating=updating, 

workers=workers) as solver: 

ret = solver.solve() 

 

return ret 

 

 

class DifferentialEvolutionSolver(object): 

 

"""This class implements the differential evolution solver 

 

Parameters 

---------- 

func : callable 

The objective function to be minimized. Must be in the form 

``f(x, *args)``, where ``x`` is the argument in the form of a 1-D array 

and ``args`` is a tuple of any additional fixed parameters needed to 

completely specify the function. 

bounds : sequence or `Bounds`, optional 

Bounds for variables. There are two ways to specify the bounds: 

1. Instance of `Bounds` class. 

2. ``(min, max)`` pairs for each element in ``x``, defining the finite 

lower and upper bounds for the optimizing argument of `func`. It is 

required to have ``len(bounds) == len(x)``. ``len(bounds)`` is used 

to determine the number of parameters in ``x``. 

args : tuple, optional 

Any additional fixed parameters needed to 

completely specify the objective function. 

strategy : str, optional 

The differential evolution strategy to use. Should be one of: 

 

- 'best1bin' 

- 'best1exp' 

- 'rand1exp' 

- 'randtobest1exp' 

- 'currenttobest1exp' 

- 'best2exp' 

- 'rand2exp' 

- 'randtobest1bin' 

- 'currenttobest1bin' 

- 'best2bin' 

- 'rand2bin' 

- 'rand1bin' 

 

The default is 'best1bin' 

 

maxiter : int, optional 

The maximum number of generations over which the entire population is 

evolved. The maximum number of function evaluations (with no polishing) 

is: ``(maxiter + 1) * popsize * len(x)`` 

popsize : int, optional 

A multiplier for setting the total population size. The population has 

``popsize * len(x)`` individuals (unless the initial population is 

supplied via the `init` keyword). 

tol : float, optional 

Relative tolerance for convergence, the solving stops when 

``np.std(pop) <= atol + tol * np.abs(np.mean(population_energies))``, 

where and `atol` and `tol` are the absolute and relative tolerance 

respectively. 

mutation : float or tuple(float, float), optional 

The mutation constant. In the literature this is also known as 

differential weight, being denoted by F. 

If specified as a float it should be in the range [0, 2]. 

If specified as a tuple ``(min, max)`` dithering is employed. Dithering 

randomly changes the mutation constant on a generation by generation 

basis. The mutation constant for that generation is taken from 

U[min, max). Dithering can help speed convergence significantly. 

Increasing the mutation constant increases the search radius, but will 

slow down convergence. 

recombination : float, optional 

The recombination constant, should be in the range [0, 1]. In the 

literature this is also known as the crossover probability, being 

denoted by CR. Increasing this value allows a larger number of mutants 

to progress into the next generation, but at the risk of population 

stability. 

seed : int or `np.random.RandomState`, optional 

If `seed` is not specified the `np.random.RandomState` singleton is 

used. 

If `seed` is an int, a new `np.random.RandomState` instance is used, 

seeded with `seed`. 

If `seed` is already a `np.random.RandomState` instance, then that 

`np.random.RandomState` instance is used. 

Specify `seed` for repeatable minimizations. 

disp : bool, optional 

Display status messages 

callback : callable, `callback(xk, convergence=val)`, optional 

A function to follow the progress of the minimization. ``xk`` is 

the current value of ``x0``. ``val`` represents the fractional 

value of the population convergence. When ``val`` is greater than one 

the function halts. If callback returns `True`, then the minimization 

is halted (any polishing is still carried out). 

polish : bool, optional 

If True, then `scipy.optimize.minimize` with the `L-BFGS-B` method 

is used to polish the best population member at the end. This requires 

a few more function evaluations. 

maxfun : int, optional 

Set the maximum number of function evaluations. However, it probably 

makes more sense to set `maxiter` instead. 

init : str or array-like, optional 

Specify which type of population initialization is performed. Should be 

one of: 

 

- 'latinhypercube' 

- 'random' 

- array specifying the initial population. The array should have 

shape ``(M, len(x))``, where len(x) is the number of parameters. 

`init` is clipped to `bounds` before use. 

 

The default is 'latinhypercube'. Latin Hypercube sampling tries to 

maximize coverage of the available parameter space. 'random' 

initializes the population randomly - this has the drawback that 

clustering can occur, preventing the whole of parameter space being 

covered. Use of an array to specify a population could be used, for 

example, to create a tight bunch of initial guesses in an location 

where the solution is known to exist, thereby reducing time for 

convergence. 

atol : float, optional 

Absolute tolerance for convergence, the solving stops when 

``np.std(pop) <= atol + tol * np.abs(np.mean(population_energies))``, 

where and `atol` and `tol` are the absolute and relative tolerance 

respectively. 

updating : {'immediate', 'deferred'}, optional 

If `immediate` the best solution vector is continuously updated within 

a single generation. This can lead to faster convergence as trial 

vectors can take advantage of continuous improvements in the best 

solution. 

With `deferred` the best solution vector is updated once per 

generation. Only `deferred` is compatible with parallelization, and the 

`workers` keyword can over-ride this option. 

workers : int or map-like callable, optional 

If `workers` is an int the population is subdivided into `workers` 

sections and evaluated in parallel 

(uses `multiprocessing.Pool <multiprocessing>`). 

Supply `-1` to use all cores available to the Process. 

Alternatively supply a map-like callable, such as 

`multiprocessing.Pool.map` for evaluating the population in parallel. 

This evaluation is carried out as ``workers(func, iterable)``. 

This option will override the `updating` keyword to 

`updating='deferred'` if `workers != 1`. 

Requires that `func` be pickleable. 

 

""" 

 

# Dispatch of mutation strategy method (binomial or exponential). 

_binomial = {'best1bin': '_best1', 

'randtobest1bin': '_randtobest1', 

'currenttobest1bin': '_currenttobest1', 

'best2bin': '_best2', 

'rand2bin': '_rand2', 

'rand1bin': '_rand1'} 

_exponential = {'best1exp': '_best1', 

'rand1exp': '_rand1', 

'randtobest1exp': '_randtobest1', 

'currenttobest1exp': '_currenttobest1', 

'best2exp': '_best2', 

'rand2exp': '_rand2'} 

 

__init_error_msg = ("The population initialization method must be one of " 

"'latinhypercube' or 'random', or an array of shape " 

"(M, N) where N is the number of parameters and M>5") 

 

def __init__(self, func, bounds, args=(), 

strategy='best1bin', maxiter=1000, popsize=15, 

tol=0.01, mutation=(0.5, 1), recombination=0.7, seed=None, 

maxfun=np.inf, callback=None, disp=False, polish=True, 

init='latinhypercube', atol=0, updating='immediate', 

workers=1): 

 

if strategy in self._binomial: 

self.mutation_func = getattr(self, self._binomial[strategy]) 

elif strategy in self._exponential: 

self.mutation_func = getattr(self, self._exponential[strategy]) 

else: 

raise ValueError("Please select a valid mutation strategy") 

self.strategy = strategy 

 

self.callback = callback 

self.polish = polish 

 

# set the updating / parallelisation options 

if updating in ['immediate', 'deferred']: 

self._updating = updating 

 

# want to use parallelisation, but updating is immediate 

if workers != 1 and updating == 'immediate': 

warnings.warn("differential_evolution: the 'workers' keyword has" 

" overridden updating='immediate' to" 

" updating='deferred'", UserWarning) 

self._updating = 'deferred' 

 

# an object with a map method. 

self._mapwrapper = MapWrapper(workers) 

 

# relative and absolute tolerances for convergence 

self.tol, self.atol = tol, atol 

 

# Mutation constant should be in [0, 2). If specified as a sequence 

# then dithering is performed. 

self.scale = mutation 

if (not np.all(np.isfinite(mutation)) or 

np.any(np.array(mutation) >= 2) or 

np.any(np.array(mutation) < 0)): 

raise ValueError('The mutation constant must be a float in ' 

'U[0, 2), or specified as a tuple(min, max)' 

' where min < max and min, max are in U[0, 2).') 

 

self.dither = None 

if hasattr(mutation, '__iter__') and len(mutation) > 1: 

self.dither = [mutation[0], mutation[1]] 

self.dither.sort() 

 

self.cross_over_probability = recombination 

 

# we create a wrapped function to allow the use of map (and Pool.map 

# in the future) 

self.func = _FunctionWrapper(func, args) 

self.args = args 

 

# convert tuple of lower and upper bounds to limits 

# [(low_0, high_0), ..., (low_n, high_n] 

# -> [[low_0, ..., low_n], [high_0, ..., high_n]] 

if isinstance(bounds, Bounds): 

self.limits = np.array(new_bounds_to_old(bounds.lb, 

bounds.ub, 

len(bounds.lb)), 

dtype=float).T 

else: 

self.limits = np.array(bounds, dtype='float').T 

 

if (np.size(self.limits, 0) != 2 or not 

np.all(np.isfinite(self.limits))): 

raise ValueError('bounds should be a sequence containing ' 

'real valued (min, max) pairs for each value' 

' in x') 

 

if maxiter is None: # the default used to be None 

maxiter = 1000 

self.maxiter = maxiter 

if maxfun is None: # the default used to be None 

maxfun = np.inf 

self.maxfun = maxfun 

 

# population is scaled to between [0, 1]. 

# We have to scale between parameter <-> population 

# save these arguments for _scale_parameter and 

# _unscale_parameter. This is an optimization 

self.__scale_arg1 = 0.5 * (self.limits[0] + self.limits[1]) 

self.__scale_arg2 = np.fabs(self.limits[0] - self.limits[1]) 

 

self.parameter_count = np.size(self.limits, 1) 

 

self.random_number_generator = check_random_state(seed) 

 

# default population initialization is a latin hypercube design, but 

# there are other population initializations possible. 

# the minimum is 5 because 'best2bin' requires a population that's at 

# least 5 long 

self.num_population_members = max(5, popsize * self.parameter_count) 

 

self.population_shape = (self.num_population_members, 

self.parameter_count) 

 

self._nfev = 0 

if isinstance(init, string_types): 

if init == 'latinhypercube': 

self.init_population_lhs() 

elif init == 'random': 

self.init_population_random() 

else: 

raise ValueError(self.__init_error_msg) 

else: 

self.init_population_array(init) 

 

self.disp = disp 

 

def init_population_lhs(self): 

""" 

Initializes the population with Latin Hypercube Sampling. 

Latin Hypercube Sampling ensures that each parameter is uniformly 

sampled over its range. 

""" 

rng = self.random_number_generator 

 

# Each parameter range needs to be sampled uniformly. The scaled 

# parameter range ([0, 1)) needs to be split into 

# `self.num_population_members` segments, each of which has the following 

# size: 

segsize = 1.0 / self.num_population_members 

 

# Within each segment we sample from a uniform random distribution. 

# We need to do this sampling for each parameter. 

samples = (segsize * rng.random_sample(self.population_shape) 

 

# Offset each segment to cover the entire parameter range [0, 1) 

+ np.linspace(0., 1., self.num_population_members, 

endpoint=False)[:, np.newaxis]) 

 

# Create an array for population of candidate solutions. 

self.population = np.zeros_like(samples) 

 

# Initialize population of candidate solutions by permutation of the 

# random samples. 

for j in range(self.parameter_count): 

order = rng.permutation(range(self.num_population_members)) 

self.population[:, j] = samples[order, j] 

 

# reset population energies 

self.population_energies = np.full(self.num_population_members, 

np.inf) 

 

# reset number of function evaluations counter 

self._nfev = 0 

 

def init_population_random(self): 

""" 

Initialises the population at random. This type of initialization 

can possess clustering, Latin Hypercube sampling is generally better. 

""" 

rng = self.random_number_generator 

self.population = rng.random_sample(self.population_shape) 

 

# reset population energies 

self.population_energies = np.full(self.num_population_members, 

np.inf) 

 

# reset number of function evaluations counter 

self._nfev = 0 

 

def init_population_array(self, init): 

""" 

Initialises the population with a user specified population. 

 

Parameters 

---------- 

init : np.ndarray 

Array specifying subset of the initial population. The array should 

have shape (M, len(x)), where len(x) is the number of parameters. 

The population is clipped to the lower and upper bounds. 

""" 

# make sure you're using a float array 

popn = np.asfarray(init) 

 

if (np.size(popn, 0) < 5 or 

popn.shape[1] != self.parameter_count or 

len(popn.shape) != 2): 

raise ValueError("The population supplied needs to have shape" 

" (M, len(x)), where M > 4.") 

 

# scale values and clip to bounds, assigning to population 

self.population = np.clip(self._unscale_parameters(popn), 0, 1) 

 

self.num_population_members = np.size(self.population, 0) 

 

self.population_shape = (self.num_population_members, 

self.parameter_count) 

 

# reset population energies 

self.population_energies = (np.ones(self.num_population_members) * 

np.inf) 

 

# reset number of function evaluations counter 

self._nfev = 0 

 

@property 

def x(self): 

""" 

The best solution from the solver 

""" 

return self._scale_parameters(self.population[0]) 

 

@property 

def convergence(self): 

""" 

The standard deviation of the population energies divided by their 

mean. 

""" 

if np.any(np.isinf(self.population_energies)): 

return np.inf 

return (np.std(self.population_energies) / 

np.abs(np.mean(self.population_energies) + _MACHEPS)) 

 

def converged(self): 

""" 

Return True if the solver has converged. 

""" 

return (np.std(self.population_energies) <= 

self.atol + 

self.tol * np.abs(np.mean(self.population_energies))) 

 

def solve(self): 

""" 

Runs the DifferentialEvolutionSolver. 

 

Returns 

------- 

res : OptimizeResult 

The optimization result represented as a ``OptimizeResult`` object. 

Important attributes are: ``x`` the solution array, ``success`` a 

Boolean flag indicating if the optimizer exited successfully and 

``message`` which describes the cause of the termination. See 

`OptimizeResult` for a description of other attributes. If `polish` 

was employed, and a lower minimum was obtained by the polishing, 

then OptimizeResult also contains the ``jac`` attribute. 

""" 

nit, warning_flag = 0, False 

status_message = _status_message['success'] 

 

# The population may have just been initialized (all entries are 

# np.inf). If it has you have to calculate the initial energies. 

# Although this is also done in the evolve generator it's possible 

# that someone can set maxiter=0, at which point we still want the 

# initial energies to be calculated (the following loop isn't run). 

if np.all(np.isinf(self.population_energies)): 

self.population_energies[:] = self._calculate_population_energies( 

self.population) 

self._promote_lowest_energy() 

 

# do the optimisation. 

for nit in xrange(1, self.maxiter + 1): 

# evolve the population by a generation 

try: 

next(self) 

except StopIteration: 

warning_flag = True 

if self._nfev > self.maxfun: 

status_message = _status_message['maxfev'] 

elif self._nfev == self.maxfun: 

status_message = ('Maximum number of function evaluations' 

' has been reached.') 

break 

 

if self.disp: 

print("differential_evolution step %d: f(x)= %g" 

% (nit, 

self.population_energies[0])) 

 

# should the solver terminate? 

convergence = self.convergence 

 

if (self.callback and 

self.callback(self._scale_parameters(self.population[0]), 

convergence=self.tol / convergence) is True): 

 

warning_flag = True 

status_message = ('callback function requested stop early ' 

'by returning True') 

break 

 

if np.any(np.isinf(self.population_energies)): 

intol = False 

else: 

intol = (np.std(self.population_energies) <= 

self.atol + 

self.tol * np.abs(np.mean(self.population_energies))) 

if warning_flag or intol: 

break 

 

else: 

status_message = _status_message['maxiter'] 

warning_flag = True 

 

DE_result = OptimizeResult( 

x=self.x, 

fun=self.population_energies[0], 

nfev=self._nfev, 

nit=nit, 

message=status_message, 

success=(warning_flag is not True)) 

 

if self.polish: 

result = minimize(self.func, 

np.copy(DE_result.x), 

method='L-BFGS-B', 

bounds=self.limits.T) 

 

self._nfev += result.nfev 

DE_result.nfev = self._nfev 

 

if result.fun < DE_result.fun: 

DE_result.fun = result.fun 

DE_result.x = result.x 

DE_result.jac = result.jac 

# to keep internal state consistent 

self.population_energies[0] = result.fun 

self.population[0] = self._unscale_parameters(result.x) 

 

return DE_result 

 

def _calculate_population_energies(self, population): 

""" 

Calculate the energies of all the population members at the same time. 

 

Parameters 

---------- 

population : ndarray 

An array of parameter vectors normalised to [0, 1] using lower 

and upper limits. Has shape ``(np.size(population, 0), len(x))``. 

 

Returns 

------- 

energies : ndarray 

An array of energies corresponding to each population member. If 

maxfun will be exceeded during this call, then the number of 

function evaluations will be reduced and energies will be 

right-padded with np.inf. Has shape ``(np.size(population, 0),)`` 

""" 

num_members = np.size(population, 0) 

nfevs = min(num_members, 

self.maxfun - num_members) 

 

energies = np.full(num_members, np.inf) 

 

parameters_pop = self._scale_parameters(population) 

try: 

calc_energies = list(self._mapwrapper(self.func, 

parameters_pop[0:nfevs])) 

energies[0:nfevs] = calc_energies 

except (TypeError, ValueError): 

# wrong number of arguments for _mapwrapper 

# or wrong length returned from the mapper 

raise RuntimeError("The map-like callable must be of the" 

" form f(func, iterable), returning a sequence" 

" of numbers the same length as 'iterable'") 

 

self._nfev += nfevs 

 

return energies 

 

def _promote_lowest_energy(self): 

# promotes the lowest energy to the first entry in the population 

l = np.argmin(self.population_energies) 

 

# put the lowest energy into the best solution position. 

self.population_energies[[0, l]] = self.population_energies[[l, 0]] 

self.population[[0, l], :] = self.population[[l, 0], :] 

 

def __iter__(self): 

return self 

 

def __enter__(self): 

return self 

 

def __exit__(self, *args): 

# to make sure resources are closed down 

self._mapwrapper.close() 

self._mapwrapper.terminate() 

 

def __del__(self): 

# to make sure resources are closed down 

self._mapwrapper.close() 

self._mapwrapper.terminate() 

 

def __next__(self): 

""" 

Evolve the population by a single generation 

 

Returns 

------- 

x : ndarray 

The best solution from the solver. 

fun : float 

Value of objective function obtained from the best solution. 

""" 

# the population may have just been initialized (all entries are 

# np.inf). If it has you have to calculate the initial energies 

if np.all(np.isinf(self.population_energies)): 

self.population_energies[:] = self._calculate_population_energies( 

self.population) 

self._promote_lowest_energy() 

 

if self.dither is not None: 

self.scale = (self.random_number_generator.rand() 

* (self.dither[1] - self.dither[0]) + self.dither[0]) 

 

if self._updating == 'immediate': 

# update best solution immediately 

for candidate in range(self.num_population_members): 

if self._nfev > self.maxfun: 

raise StopIteration 

 

# create a trial solution 

trial = self._mutate(candidate) 

 

# ensuring that it's in the range [0, 1) 

self._ensure_constraint(trial) 

 

# scale from [0, 1) to the actual parameter value 

parameters = self._scale_parameters(trial) 

 

# determine the energy of the objective function 

energy = self.func(parameters) 

self._nfev += 1 

 

# if the energy of the trial candidate is lower than the 

# original population member then replace it 

if energy < self.population_energies[candidate]: 

self.population[candidate] = trial 

self.population_energies[candidate] = energy 

 

# if the trial candidate also has a lower energy than the 

# best solution then promote it to the best solution. 

if energy < self.population_energies[0]: 

self._promote_lowest_energy() 

 

elif self._updating == 'deferred': 

# update best solution once per generation 

if self._nfev >= self.maxfun: 

raise StopIteration 

 

# 'deferred' approach, vectorised form. 

# create trial solutions 

trial_pop = np.array( 

[self._mutate(i) for i in range(self.num_population_members)]) 

 

# enforce bounds 

self._ensure_constraint(trial_pop) 

 

# determine the energies of the objective function 

trial_energies = self._calculate_population_energies(trial_pop) 

 

# which solutions are improved? 

loc = trial_energies < self.population_energies 

self.population = np.where(loc[:, np.newaxis], 

trial_pop, 

self.population) 

self.population_energies = np.where(loc, 

trial_energies, 

self.population_energies) 

 

# make sure the best solution is updated if updating='deferred'. 

# put the lowest energy into the best solution position. 

self._promote_lowest_energy() 

 

return self.x, self.population_energies[0] 

 

next = __next__ 

 

def _scale_parameters(self, trial): 

"""Scale from a number between 0 and 1 to parameters.""" 

return self.__scale_arg1 + (trial - 0.5) * self.__scale_arg2 

 

def _unscale_parameters(self, parameters): 

"""Scale from parameters to a number between 0 and 1.""" 

return (parameters - self.__scale_arg1) / self.__scale_arg2 + 0.5 

 

def _ensure_constraint(self, trial): 

"""Make sure the parameters lie between the limits.""" 

mask = np.where((trial > 1) | (trial < 0)) 

trial[mask] = self.random_number_generator.rand(mask[0].size) 

 

def _mutate(self, candidate): 

"""Create a trial vector based on a mutation strategy.""" 

trial = np.copy(self.population[candidate]) 

 

rng = self.random_number_generator 

 

fill_point = rng.randint(0, self.parameter_count) 

 

if self.strategy in ['currenttobest1exp', 'currenttobest1bin']: 

bprime = self.mutation_func(candidate, 

self._select_samples(candidate, 5)) 

else: 

bprime = self.mutation_func(self._select_samples(candidate, 5)) 

 

if self.strategy in self._binomial: 

crossovers = rng.rand(self.parameter_count) 

crossovers = crossovers < self.cross_over_probability 

# the last one is always from the bprime vector for binomial 

# If you fill in modulo with a loop you have to set the last one to 

# true. If you don't use a loop then you can have any random entry 

# be True. 

crossovers[fill_point] = True 

trial = np.where(crossovers, bprime, trial) 

return trial 

 

elif self.strategy in self._exponential: 

i = 0 

while (i < self.parameter_count and 

rng.rand() < self.cross_over_probability): 

 

trial[fill_point] = bprime[fill_point] 

fill_point = (fill_point + 1) % self.parameter_count 

i += 1 

 

return trial 

 

def _best1(self, samples): 

"""best1bin, best1exp""" 

r0, r1 = samples[:2] 

return (self.population[0] + self.scale * 

(self.population[r0] - self.population[r1])) 

 

def _rand1(self, samples): 

"""rand1bin, rand1exp""" 

r0, r1, r2 = samples[:3] 

return (self.population[r0] + self.scale * 

(self.population[r1] - self.population[r2])) 

 

def _randtobest1(self, samples): 

"""randtobest1bin, randtobest1exp""" 

r0, r1, r2 = samples[:3] 

bprime = np.copy(self.population[r0]) 

bprime += self.scale * (self.population[0] - bprime) 

bprime += self.scale * (self.population[r1] - 

self.population[r2]) 

return bprime 

 

def _currenttobest1(self, candidate, samples): 

"""currenttobest1bin, currenttobest1exp""" 

r0, r1 = samples[:2] 

bprime = (self.population[candidate] + self.scale * 

(self.population[0] - self.population[candidate] + 

self.population[r0] - self.population[r1])) 

return bprime 

 

def _best2(self, samples): 

"""best2bin, best2exp""" 

r0, r1, r2, r3 = samples[:4] 

bprime = (self.population[0] + self.scale * 

(self.population[r0] + self.population[r1] - 

self.population[r2] - self.population[r3])) 

 

return bprime 

 

def _rand2(self, samples): 

"""rand2bin, rand2exp""" 

r0, r1, r2, r3, r4 = samples 

bprime = (self.population[r0] + self.scale * 

(self.population[r1] + self.population[r2] - 

self.population[r3] - self.population[r4])) 

 

return bprime 

 

def _select_samples(self, candidate, number_samples): 

""" 

obtain random integers from range(self.num_population_members), 

without replacement. You can't have the original candidate either. 

""" 

idxs = list(range(self.num_population_members)) 

idxs.remove(candidate) 

self.random_number_generator.shuffle(idxs) 

idxs = idxs[:number_samples] 

return idxs 

 

 

class _FunctionWrapper(object): 

""" 

Object to wrap user cost function, allowing picklability 

""" 

def __init__(self, f, args): 

self.f = f 

self.args = [] if args is None else args 

 

def __call__(self, x): 

return self.f(x, *self.args)