givp 1.0.0

GRASP-ILS-VND with Path Relinking metaheuristic for continuous black-box optimization
Documentation
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
1029
1030
1031
1032
1033
1034
1035
1036
1037
# givp — GRASP-ILS-VND with Path Relinking

**Python**  
[![PyPI: givp v1.0.0](https://img.shields.io/badge/PyPI-givp%20v1.0.0-3775A9?logo=pypi&logoColor=white)](https://pypi.org/project/givp/)
[![Python versions](https://img.shields.io/badge/Python-3.10%E2%80%933.15-blue?logo=python&logoColor=white)](https://pypi.org/project/givp/)
[![CI Python](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-python.yml/badge.svg)](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-python.yml)
[![Codecov (python)](https://codecov.io/gh/Arnime/grasp_ils_vnd_pr/graph/badge.svg?flag=python)](https://codecov.io/gh/Arnime/grasp_ils_vnd_pr/flags/python)
[![Ruff](https://img.shields.io/badge/Linter-ruff-red?logo=ruff&logoColor=white)](https://github.com/astral-sh/ruff)
[![Checked with mypy](https://img.shields.io/badge/Type--Checked-mypy-blue?logo=python&logoColor=white)](https://mypy-lang.org/)

**Julia**  
[![JuliaHub: GIVPOptimizer v1.0.0](https://img.shields.io/badge/JuliaHub-GIVPOptimizer%20v1.0.0-9558B2?logo=julia&logoColor=white)](https://juliahub.com/ui/Packages/General/GIVPOptimizer)
[![Julia](https://img.shields.io/badge/Julia-1.9%2B-9558B2?logo=julia&logoColor=white)](https://julialang.org/)
[![CI Julia](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-julia.yml/badge.svg)](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-julia.yml)
[![Codecov (julia)](https://codecov.io/gh/Arnime/grasp_ils_vnd_pr/graph/badge.svg?flag=julia)](https://codecov.io/gh/Arnime/grasp_ils_vnd_pr/flags/julia)
[![JuliaFormatter](https://img.shields.io/badge/Formatter-JuliaFormatter.jl-9558B2?logo=julia&logoColor=white)](https://github.com/domluna/JuliaFormatter.jl)
[![JET](https://img.shields.io/badge/Type--Checked-JET.jl-9558B2?logo=julia&logoColor=white)](https://github.com/aviatesk/JET.jl)
[![Aqua](https://img.shields.io/badge/Quality-Aqua.jl-9558B2?logo=julia&logoColor=white)](https://github.com/JuliaTesting/Aqua.jl)

**Rust**  
[![crates.io: givp v1.0.0](https://img.shields.io/badge/Crates-givp%20v1.0.0-E05D44?logo=rust&logoColor=white)](https://crates.io/crates/givp)
[![Rust](https://img.shields.io/badge/Rust-1.85%2B-000000?logo=rust&logoColor=white)](https://www.rust-lang.org/)
[![CI Rust](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-rust.yml/badge.svg)](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-rust.yml)
[![Codecov (rust)](https://codecov.io/gh/Arnime/grasp_ils_vnd_pr/graph/badge.svg?flag=rust)](https://codecov.io/gh/Arnime/grasp_ils_vnd_pr/flags/rust)
[![docs.rs](https://img.shields.io/docsrs/givp?cacheSeconds=300&logo=docs.rs&logoColor=white)](https://docs.rs/givp)
[![Clippy](https://img.shields.io/badge/Linter-Clippy-E05D44?logo=rust&logoColor=white)](https://doc.rust-lang.org/clippy/)
[![rustfmt](https://img.shields.io/badge/Formatter-rustfmt-E05D44?logo=rust&logoColor=white)](https://github.com/rust-lang/rustfmt)

**C++**  
[![header-only](https://img.shields.io/badge/header--only-yes-brightgreen?logo=cplusplus&logoColor=white)](cpp/include/givp/)
[![C++17](https://img.shields.io/badge/C%2B%2B-17-blue?logo=cplusplus&logoColor=white)](https://en.cppreference.com/w/cpp/17)
[![CI C++](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-cpp.yml/badge.svg)](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-cpp.yml)
[![Codecov (cpp)](https://codecov.io/gh/Arnime/grasp_ils_vnd_pr/graph/badge.svg?flag=cpp)](https://codecov.io/gh/Arnime/grasp_ils_vnd_pr/flags/cpp)
[![clang-format](https://img.shields.io/badge/Formatter-clang--format-blue?logo=llvm&logoColor=white)](https://clang.llvm.org/docs/ClangFormat.html)
[![clang-tidy](https://img.shields.io/badge/Linter-clang--tidy-blue?logo=llvm&logoColor=white)](https://clang.llvm.org/extra/clang-tidy/)

**R**  
[![R package: givp v1.0.0](https://img.shields.io/badge/R%20Package-givp%20v1.0.0-276DC3?logo=r&logoColor=white)](https://arnime.r-universe.dev/givp)
[![R](https://img.shields.io/badge/R-4.1%2B-276DC3?logo=r&logoColor=white)](https://www.r-project.org/)
[![CI R](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-r.yml/badge.svg)](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-r.yml)
[![Codecov (r)](https://codecov.io/gh/Arnime/grasp_ils_vnd_pr/graph/badge.svg?flag=r)](https://codecov.io/gh/Arnime/grasp_ils_vnd_pr/flags/r)
[![lintr](https://img.shields.io/badge/Linter-lintr-276DC3?logo=r&logoColor=white)](https://lintr.r-lib.org/)
[![testthat](https://img.shields.io/badge/Tests-testthat-276DC3?logo=r&logoColor=white)](https://testthat.r-lib.org/)

**Project**  
[![OpenSSF Scorecard](https://img.shields.io/ossf-scorecard/github.com/Arnime/grasp_ils_vnd_pr?cacheSeconds=300&logo=openssf&logoColor=white)](https://securityscorecards.dev/viewer/?uri=github.com/Arnime/grasp_ils_vnd_pr)
[![CI SonarQube](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-sonarqube.yml/badge.svg)](https://github.com/Arnime/grasp_ils_vnd_pr/actions/workflows/ci-sonarqube.yml)
[![OpenSSF Gold](https://img.shields.io/badge/OpenSSF-best%20practices-gold?logo=openssf&logoColor=white)](https://www.bestpractices.dev/projects/12627)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg?logo=open-source-initiative&logoColor=white)](https://opensource.org/licenses/MIT)
[![PRs welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg?logo=github&logoColor=white)](CONTRIBUTING.md)

A direction-agnostic metaheuristic optimizer for **continuous, integer or
mixed** black-box problems, available in five languages:

| Language | Distribution | Requires |
|----------|-------------|---------|
| **Python** (NumPy-native) | [PyPI `givp`]https://pypi.org/project/givp/ | Python 3.10+, NumPy |
| **Julia** | [JuliaHub `GIVPOptimizer`]https://juliahub.com/ui/Packages/General/GIVPOptimizer | Julia 1.9+ |
| **Rust** | [crates.io `givp`]https://crates.io/crates/givp | Rust 1.85+ |
| **C++17** | Header-only (vcpkg / FetchContent / copy) | C++17 compiler, CMake 3.21+ |
| **R** | [r-universe `givp`]https://arnime.r-universe.dev/givp / GitHub / local source (`r/`) | R 4.1+ |

The library bundles:

- **GRASP** — Greedy Randomized Adaptive Search Procedure
- **ILS** — Iterated Local Search
- **VND** — Variable Neighborhood Descent (with an adaptive variant)
- **Path Relinking** between elite solutions
- LRU evaluation cache, convergence monitor, optional thread-parallel candidate
  evaluation, and a wall-clock time budget

Code quality is enforced in CI with language-specific static analysis and
formatting checks — analogous to mypy + ruff in Python — across all ports:

| Language | Linter / static analysis | Formatter | Type checker |
|----------|--------------------------|-----------|-------------|
| **Python** | [ruff]https://github.com/astral-sh/ruff | ruff | [mypy]https://mypy-lang.org/ (strict) |
| **Julia** | [Aqua.jl]https://github.com/JuliaTesting/Aqua.jl | [JuliaFormatter]https://github.com/domluna/JuliaFormatter.jl | [JET.jl]https://github.com/aviatesk/JET.jl |
| **Rust** | [Clippy]https://doc.rust-lang.org/clippy/ (`-D warnings`, all targets) | [rustfmt]https://github.com/rust-lang/rustfmt | rustc (type system) |
| **C++** | [clang-tidy]https://clang.llvm.org/extra/clang-tidy/ (`--warnings-as-errors`) | [clang-format]https://clang.llvm.org/docs/ClangFormat.html (LLVM style) | compiler (`-Wall -Wextra -Wpedantic -Werror`) |
| **R** | [lintr]https://lintr.r-lib.org/ (`linters_with_defaults()`) |||

All checks run as mandatory CI gates. A monorepo SonarQube quality gate
workflow runs in addition (CI SonarQube).

---

## Table of contents

- [givp — GRASP-ILS-VND with Path Relinking]#givp--grasp-ils-vnd-with-path-relinking
  - [Table of contents]#table-of-contents
  - [Install]#install
    - [Python installation]#python-installation
    - [Julia installation]#julia-installation
    - [Rust installation]#rust-installation
    - [C++ installation]#c-installation
    - [R installation]#r-installation
  - [Python]#python
    - [Quick start]#quick-start
    - [Choosing the optimization sense]#choosing-the-optimization-sense
      - [Boolean flag (recommended)]#boolean-flag-recommended
      - [String flag (SciPy/Optuna compatible)]#string-flag-scipyoptuna-compatible
    - [Bounds, integer variables and mixed problems]#bounds-integer-variables-and-mixed-problems
    - [Object-oriented API and multi-start]#object-oriented-api-and-multi-start
    - [Configuration cookbook]#configuration-cookbook
    - [Inspecting progress (callback and verbose)]#inspecting-progress-callback-and-verbose
    - [API reference]#api-reference
      - [`givp(...) -> OptimizeResult`]#givp---optimizeresult
      - [`class GIVPOptimizer`]#class-givpoptimizer
      - [`class GIVPConfig` (dataclass)]#class-givpconfig-dataclass
      - [`class OptimizeResult`]#class-optimizeresult
    - [Glossary of hyper-parameters]#glossary-of-hyper-parameters
    - [Adapting to a domain-specific model]#adapting-to-a-domain-specific-model
  - [Julia]#julia
    - [Julia quick start]#julia-quick-start
    - [Julia bounds and integer variables]#julia-bounds-and-integer-variables
    - [Julia configuration cookbook]#julia-configuration-cookbook
    - [Julia result fields]#julia-result-fields
    - [Julia progress monitoring]#julia-progress-monitoring
    - [Running Julia tests and benchmarks]#running-julia-tests-and-benchmarks
  - [Rust]#rust
    - [Rust quick start]#rust-quick-start
    - [Rust bounds and integer variables]#rust-bounds-and-integer-variables
    - [Rust configuration cookbook]#rust-configuration-cookbook
    - [Rust result fields]#rust-result-fields
    - [Running Rust tests and benchmarks]#running-rust-tests-and-benchmarks
  - [C++]#c
    - [C++ quick start]#c-quick-start
    - [C++ configuration]#c-configuration
    - [Running tests and benchmarks]#running-tests-and-benchmarks
  - [R]#r
    - [R quick start]#r-quick-start
    - [R bounds and integer variables]#r-bounds-and-integer-variables
    - [R configuration cookbook]#r-configuration-cookbook
    - [Running R tests and lint]#running-r-tests-and-lint
  - [Comparison with other optimizers]#comparison-with-other-optimizers
  - [Empirical results]#empirical-results
    - [Sphere — $f(\\mathbf{x}) = \\sum x\_i^2$, global minimum 0 at **0**]#sphere--fmathbfx--sum-x_i2-global-minimum-0-at-0
    - [Rosenbrock — global minimum 0 at **1**]#rosenbrock--global-minimum-0-at-1
    - [Rastrigin — multimodal, global minimum 0 at **0**]#rastrigin--multimodal-global-minimum-0-at-0
    - [Ackley — multimodal, global minimum 0 at **0**]#ackley--multimodal-global-minimum-0-at-0
      - [**References**]#references
  - [Troubleshooting]#troubleshooting
  - [License]#license

---

## Install

### Python installation

```bash
pip install givp
```

> **Recommended:** install with the optional `cache` extra to use
> [xxhash]https://github.com/ifduyue/python-xxhash for a ~10× faster
> evaluation cache:
>
> ```bash
> pip install givp[cache]
> ```
>
> Without `xxhash`, the cache falls back to Python's built-in `hashlib`
> (SHA-256), which works correctly but is slower.

From source (editable):

```bash
git clone https://github.com/Arnime/grasp_ils_vnd_pr.git
cd grasp_ils_vnd_pr
pip install -e .[dev]
```

Requires Python 3.10+ and NumPy.

### Julia installation

```bash
git clone https://github.com/Arnime/grasp_ils_vnd_pr.git
cd grasp_ils_vnd_pr/julia
julia --project=. -e 'using Pkg; Pkg.instantiate()'
```

Requires Julia 1.9+.

### Rust installation

Add to your `Cargo.toml`:

```toml
[dependencies]
givp = "1.0.0"
```

From source:

```bash
git clone https://github.com/Arnime/grasp_ils_vnd_pr.git
cd grasp_ils_vnd_pr/rust
cargo build --release
cargo test
```

Requires Rust 1.85+ (edition 2021).

### C++ installation

The C++ port is **header-only**.

Install with vcpkg:

```bash
vcpkg install givp
```

Then use in CMake:

```cmake
find_package(givp CONFIG REQUIRED)
target_link_libraries(my_app PRIVATE givp::givp)
```

Alternatively, use CMake `FetchContent`:

```cmake
include(FetchContent)
FetchContent_Declare(
    givp
    GIT_REPOSITORY https://github.com/Arnime/grasp_ils_vnd_pr.git
    GIT_TAG        v1.0.0
    SOURCE_SUBDIR  cpp
)
FetchContent_MakeAvailable(givp)

target_link_libraries(my_app PRIVATE givp::givp)
```

Or copy the `cpp/include/givp/` directory into your project and add it to your
include path. No external dependencies are required at runtime.

Requires a C++17 compiler (GCC 9+, Clang 10+, MSVC 2019+) and CMake 3.21+
(for the `FetchContent` approach).

Dependency-manager style alternatives:

- **vcpkg manifest mode** (recommended for application repos):

```json
{
  "dependencies": [
    "givp"
  ]
}
```

- **FetchContent/CPM-style git dependency** (no local copy required):
  use the `FetchContent` snippet above pinning `GIT_TAG`.

### R installation

Install from **r-universe** (binary/source, no local clone):

```r
install.packages(
  "givp",
  repos = c("https://arnime.r-universe.dev", "https://cloud.r-project.org")
)
```

Template files for the `Arnime.r-universe.dev` repository are available in
`docs/r-universe/`.

Install directly from GitHub:

```r
install.packages("remotes")
remotes::install_github("Arnime/grasp_ils_vnd_pr", subdir = "r")
```

From source:

```bash
git clone https://github.com/Arnime/grasp_ils_vnd_pr.git
cd grasp_ils_vnd_pr
R CMD INSTALL r
```

Requires R 4.1+.

---

## Python

### Quick start

```python
import numpy as np
from givp import givp

def sphere(x: np.ndarray) -> float:
    return float(np.sum(x ** 2))

result = givp(sphere, bounds=[(-5.0, 5.0)] * 10)
print(result.x)        # best vector found
print(result.fun)      # best objective value
print(result.nfev)     # number of evaluations performed
```

Default behavior:

- **Minimization** (`minimize=True` / `direction="minimize"`).
- All variables treated as continuous.
- Default hyper-parameters (`GIVPConfig()`).

---

### Choosing the optimization sense

The library is **agnostic** to whether you want the lowest or the highest
value of `func`. Two equivalent ways to declare it:

#### Boolean flag (recommended)

```python
from givp import givp

def gain(x):
    return float((x ** 2).sum())  # higher is better

result = givp(gain, [(-5, 5)] * 10, minimize=False)
assert result.direction == "maximize"
```

#### String flag (SciPy/Optuna compatible)

```python
result = givp(gain, [(-5, 5)] * 10, direction="maximize")
```

Both flags are accepted on `givp`, on `GIVPOptimizer` and on
`GIVPConfig`. Setting **both** simultaneously is allowed only when they
agree; conflicting values raise `ValueError`.

> **Internal note.** The core algorithm always minimizes. When you ask for
> maximization the public API wraps your objective with a sign flip and
> restores the sign on `result.fun`. This means `result.fun` is always
> reported in your original sign — no need to negate it back yourself.

---

### Bounds, integer variables and mixed problems

`bounds` is accepted in two equivalent forms:

```python
# SciPy style: list of (low, high) per variable
bounds = [(-5.0, 5.0), (0.0, 10.0), (-1.0, 1.0)]

# (lower, upper) tuple of two equally-sized sequences
bounds = ([-5.0, 0.0, -1.0], [5.0, 10.0, 1.0])
```

By default every variable is continuous. To declare a **mixed** problem (some
continuous variables followed by some integer variables in the decision
vector), use `integer_split` on the configuration:

```python
from givp import GIVPConfig, givp

n_cont, n_int = 12, 8
bounds = [(-5.0, 5.0)] * n_cont + [(0.0, 4.0)] * n_int

cfg = GIVPConfig(integer_split=n_cont)  # indices >= n_cont are integer

result = givp(my_objective, bounds, config=cfg)
```

Special cases:

| `integer_split` | Meaning                                |
|-----------------|----------------------------------------|
| `None` (public API default: `num_vars`) | All-continuous problem.    |
| `0`             | All-integer problem.                   |
| `n_vars`        | All-continuous problem (explicit).     |
| `k` (0 < k < n) | First `k` continuous, rest integer.    |

---

### Object-oriented API and multi-start

When you want to keep configuration around, run the optimizer multiple times
and track the best result automatically, use `GIVPOptimizer`:

```python
from givp import GIVPConfig, GIVPOptimizer

opt = GIVPOptimizer(
    func=sphere,
    bounds=[(-5.0, 5.0)] * 10,
    minimize=True,
    config=GIVPConfig(max_iterations=50, time_limit=30.0),
    verbose=True,
)
for _ in range(5):
    opt.run()
print("best across 5 restarts:", opt.best_fun)
print("history length:", len(opt.history))
```

`opt.best_x` and `opt.best_fun` always reflect the best result observed across
all `run()` calls, in the **user's original sign**.

---

### Configuration cookbook

```python
from givp import GIVPConfig

# 1) Fast triage (small budget, no warm-up)
cfg_fast = GIVPConfig(
    max_iterations=20,
    vnd_iterations=50,
    ils_iterations=5,
    use_elite_pool=False,
    use_convergence_monitor=False,
    use_cache=True,
)

# 2) Production-quality run with wall-clock budget
cfg_quality = GIVPConfig(
    max_iterations=200,
    vnd_iterations=300,
    ils_iterations=15,
    elite_size=10,
    path_relink_frequency=5,
    adaptive_alpha=True,
    alpha_min=0.05,
    alpha_max=0.20,
    time_limit=600.0,         # stop after 10 minutes
    n_workers=4,              # parallelize candidate evaluation
)

# 3) Expensive objective: maximize cache reuse, keep evaluations few
cfg_expensive = GIVPConfig(
    num_candidates_per_step=8,
    cache_size=50_000,
    use_cache=True,
    early_stop_threshold=40,  # stop earlier on stagnation
)

# 4) Maximization with hourly-shaped layout (3 plants × 24 hours)
cfg_hydro = GIVPConfig(
    minimize=False,
    integer_split=72,         # first 72 vars continuous, rest integer
    max_iterations=120,
    time_limit=300.0,
)
```

---

### Inspecting progress (callback and verbose)

Both `givp` and `GIVPOptimizer` accept:

- `verbose=True` — prints per-iteration cost and cache statistics.
- `iteration_callback=fn` — calls `fn(iteration_index, best_cost, best_solution)`
  once per outer GRASP iteration. The callback receives the cost in the
  **internal minimization sign** (i.e., already sign-flipped if you asked for
  maximization). Useful to plot convergence or persist intermediate results.

```python
costs = []

def log_iter(i, cost, sol):
    costs.append(cost)

result = givp(
    sphere,
    [(-5, 5)] * 10,
    iteration_callback=log_iter,
    verbose=True,
)
```

---

### API reference

#### `givp(...) -> OptimizeResult`

```python
givp(
    func: Callable[[np.ndarray], float],
    bounds: Sequence[tuple[float, float]] | tuple[Sequence[float], Sequence[float]],
    *,
    num_vars: int | None = None,
    minimize: bool | None = None,
    direction: str | None = None,         # 'minimize' or 'maximize'
    config: GIVPConfig | None = None,
    initial_guess: Sequence[float] | None = None,
    iteration_callback: Callable[[int, float, np.ndarray], None] | None = None,
    seed: int | None = None,
    verbose: bool = False,
) -> OptimizeResult
```

#### `class GIVPOptimizer`

Same constructor signature, exposes `.run() -> OptimizeResult` and tracks
`.best_x`, `.best_fun`, `.history`.

#### `class GIVPConfig` (dataclass)

All hyper-parameters listed in the [glossary](#glossary-of-hyper-parameters).

#### `class OptimizeResult`

| Field       | Type        | Meaning                                                    |
|-------------|-------------|------------------------------------------------------------|
| `x`         | `np.ndarray`| Best solution vector.                                      |
| `fun`       | `float`     | Objective value at `x`, in the **user's original sign**.   |
| `nit`       | `int`       | GRASP outer iterations executed.                           |
| `nfev`      | `int`       | Number of objective evaluations.                           |
| `success`   | `bool`      | True when at least one feasible solution was produced.     |
| `message`   | `str`       | Human-readable termination reason.                         |
| `direction` | `str`       | `'minimize'` or `'maximize'`.                              |
| `meta`      | `dict`      | Algorithm-specific extras (cache stats, etc.).             |

For backward compatibility the result is iterable: `x, fun = result` works.

---

### Glossary of hyper-parameters

| Field                       | Default | Meaning                                                            |
|-----------------------------|---------|--------------------------------------------------------------------|
| `max_iterations`            | 100     | GRASP outer iterations.                                            |
| `alpha`                     | 0.12    | Initial RCL randomization (0 = greedy, 1 = uniform).               |
| `vnd_iterations`            | 200     | Maximum VND inner iterations.                                      |
| `ils_iterations`            | 10      | Iterated Local Search loops per outer iteration.                   |
| `perturbation_strength`     | 4       | Magnitude of ILS perturbation (number of variables jolted).        |
| `use_elite_pool`            | True    | Maintain a diverse pool of elite solutions for path relinking.     |
| `elite_size`                | 7       | Maximum number of elite solutions kept.                            |
| `path_relink_frequency`     | 8       | Every N GRASP iterations, run path relinking on elite pairs.       |
| `adaptive_alpha`            | True    | If True, alpha varies in `[alpha_min, alpha_max]` over iterations. |
| `alpha_min` / `alpha_max`   | 0.08 / 0.18 | Bounds for adaptive alpha.                                     |
| `num_candidates_per_step`   | 20      | Candidates evaluated per construction step.                        |
| `use_cache`                 | True    | Memoize evaluations via LRU cache.                                 |
| `cache_size`                | 10000   | LRU cache capacity.                                                |
| `early_stop_threshold`      | 80      | Iterations without improvement before terminating.                 |
| `use_convergence_monitor`   | True    | Enable diversification/restart heuristics.                         |
| `n_workers`                 | 1       | Threads used to evaluate candidates concurrently in the GRASP construction phase. Set to `os.cpu_count()` or a fixed value (2–4) for CPU-bound objectives. Due to the Python GIL, gains are most visible when `func` releases the GIL (e.g. NumPy-heavy code). |
| `time_limit`                | 0.0     | Wall-clock budget in seconds (`0` = unlimited).                    |
| `minimize`                  | `None`  | Boolean direction flag. `True` = minimize, `False` = maximize.     |
| `direction`                 | `'minimize'` | String direction flag (alternative form).                     |
| `integer_split`             | `None`  | Index where integer variables begin in the decision vector.        |

---

### Adapting to a domain-specific model

The library knows nothing about your problem. Wrap your domain code so it
exposes a `func(x: np.ndarray) -> float` and a list of bounds. Penalty terms,
repair operators and constraint handling all live in your project.

Minimal pattern:

```python
def make_objective(model):
    def f(x):
        try:
            return float(model.evaluate(x))
        except (ValueError, RuntimeError):
            return float("inf")  # treat infeasibility as worst possible cost
    return f

result = givp(make_objective(my_model), bounds=my_bounds)
```

For an end-to-end example with a mixed continuous/integer hydropower model,
see the SOG2 adapter in the upstream project repository
(`givp.py`).

---

## Julia

The Julia port exposes the same algorithm with an idiomatic Julia API.
Install via the [Julia installation](#julia-installation) instructions above.

### Julia quick start

```julia
using GIVPOptimizer

function sphere(x::Vector{Float64})::Float64
    return sum(x .^ 2)
end

result = givp(sphere, [(-5.0, 5.0) for _ in 1:10])
println(result.x)       # best vector found
println(result.fun)     # best objective value
println(result.nfev)    # number of evaluations
```

Maximization:

```julia
result = givp(my_score, bounds; direction=maximize)
```

### Julia bounds and integer variables

`bounds` is a vector of `(lower, upper)` tuples, one per variable.
Use `integer_split` to declare mixed continuous/integer problems:

```julia
n_cont, n_int = 8, 4
bounds = vcat(fill((-5.0, 5.0), n_cont), fill((0.0, 3.0), n_int))
cfg = GIVPConfig(; integer_split=n_cont)          # indices >= n_cont are integer
result = givp(my_obj, bounds; config=cfg)
```

### Julia configuration cookbook

```julia
# 1) Fast triage
cfg_fast = GIVPConfig(;
    max_iterations=20, vnd_iterations=50, ils_iterations=5,
    use_elite_pool=false, use_convergence_monitor=false,
)

# 2) Production-quality with wall-clock budget
cfg_quality = GIVPConfig(;
    max_iterations=200, vnd_iterations=300, ils_iterations=15,
    elite_size=10, path_relink_frequency=5,
    adaptive_alpha=true, alpha_min=0.05, alpha_max=0.20,
    time_limit=600.0,
)

# 3) Maximization
cfg_max = GIVPConfig(; direction=maximize, max_iterations=100)

result = givp(sphere, bounds; config=cfg_quality, seed=42, verbose=true)
```

### Julia result fields

| Field       | Type               | Meaning                                          |
|-------------|--------------------|--------------------------------------------------|
| `x`         | `Vector{Float64}`  | Best solution vector.                            |
| `fun`       | `Float64`          | Objective value at `x`, in the original sign.    |
| `nit`       | `Int`              | Outer GRASP iterations executed.                 |
| `nfev`      | `Int`              | Total objective-function evaluations.            |
| `success`   | `Bool`             | True when `fun` is finite.                       |
| `message`   | `String`           | Human-readable termination reason.               |
| `direction` | `Direction`        | `minimize` or `maximize` (enum).                 |
| `meta`      | `Dict{String,Any}` | Algorithm-specific extras (cache stats, etc.).   |

Tuple unpacking works: `x, fun = result`.

### Julia progress monitoring

```julia
costs = Float64[]

function log_iter(i, cost, sol)
    push!(costs, cost)
end

result = givp(sphere, bounds; iteration_callback=log_iter, verbose=true)
```

### Running Julia tests and benchmarks

```bash
cd julia
julia --project=. -e 'using Pkg; Pkg.test()'
julia --project=. benchmarks/benchmarks.jl
```

---

## Rust

The Rust port is a zero-dependency (no NumPy), native-performance implementation
suitable for embedding in systems code or for maximum throughput.
Install via the [Rust installation](#rust-installation) instructions above.

### Rust quick start

```rust
use givp::{givp, GivpConfig};

let sphere = |x: &[f64]| -> f64 { x.iter().map(|v| v * v).sum() };
let bounds: Vec<(f64, f64)> = vec![(-5.12, 5.12); 5];

let result = givp(sphere, &bounds, GivpConfig::default()).unwrap();
println!("best: {:.6}", result.fun);
println!("x:    {:?}", result.x);
println!("nfev: {}", result.nfev);
```

### Rust bounds and integer variables

Bounds are a `&[(f64, f64)]` slice. Use `integer_split` for mixed problems:

```rust
use givp::{givp, GivpConfig};

let n_cont = 8usize;
let mut bounds: Vec<(f64, f64)> = vec![(-5.0, 5.0); n_cont];
bounds.extend(vec![(0.0, 3.0); 4]);  // 4 integer variables

let config = GivpConfig {
    integer_split: Some(n_cont),  // indices >= n_cont are rounded to int
    ..Default::default()
};
let result = givp(my_obj, &bounds, config).unwrap();
```

### Rust configuration cookbook

```rust
use givp::{givp, GivpConfig, Direction};

// Maximization
let cfg_max = GivpConfig {
    direction: Direction::Maximize,
    seed: Some(42),
    ..Default::default()
};

// Production-quality run with wall-clock budget
let cfg_quality = GivpConfig {
    max_iterations: 200,
    vnd_iterations: 300,
    ils_iterations: 15,
    elite_size: 7,
    path_relink_frequency: 5,
    adaptive_alpha: true,
    alpha_min: 0.05,
    alpha_max: 0.20,
    time_limit: 600.0,
    seed: Some(0),
    ..Default::default()
};

let result = givp(sphere, &bounds, cfg_quality).unwrap();
```

### Rust result fields

| Field       | Type         | Meaning                                         |
|-------------|--------------|-------------------------------------------------|
| `x`         | `Vec<f64>`   | Best solution vector.                           |
| `fun`       | `f64`        | Objective value at `x`, in the original sign.   |
| `nit`       | `usize`      | Outer GRASP iterations executed.                |
| `nfev`      | `usize`      | Total objective-function evaluations.           |
| `success`   | `bool`       | True when `fun` is finite.                      |
| `message`   | `String`     | Human-readable termination reason.              |
| `termination` | `TerminationReason` | Typed termination enum.               |

The function returns `Result<OptimizeResult, GivpError>` — use `.unwrap()` for
quick scripts or pattern-match for production code.

### Running Rust tests and benchmarks

```bash
cd rust
cargo test
cargo bench
```

---

## C++

The C++ port is a **header-only, zero-runtime-dependency** implementation.
It uses the same algorithm and identical default hyper-parameters as the
Python, Julia, and Rust ports.

### C++ quick start

```cpp
#include <givp/givp.hpp>
#include <iostream>
#include <vector>

int main() {
    auto sphere = [](const std::vector<double>& x) {
        double s = 0.0;
        for (auto v : x) s += v * v;
        return s;
    };

    std::vector<std::pair<double, double>> bounds(10, {-5.0, 5.0});
    givp::OptimizeResult r = givp::givp(sphere, bounds);

    std::cout << "best: " << r.fun  << "\n";  // best objective value
    std::cout << "nfev: " << r.nfev << "\n";  // evaluations used
    std::cout << "nit:  " << r.nit  << "\n";  // outer iterations
}
```

Maximization:

```cpp
givp::GivpConfig cfg;
cfg.direction = givp::Direction::Maximize;
auto result = givp::givp(my_func, bounds, cfg);
```

### C++ configuration

```cpp
givp::GivpConfig cfg;
cfg.max_iterations          = 50;
cfg.vnd_iterations          = 100;
cfg.time_limit              = 30.0;          // wall-clock seconds
cfg.seed                    = 42;
cfg.integer_split           = 8;             // first 8 vars continuous
cfg.verbose                 = true;

givp::OptimizeResult r = givp::givp(sphere, bounds, cfg);
```

`OptimizeResult` fields mirror all other ports:

| Field     | Type                    | Meaning                                          |
|-----------|-------------------------|--------------------------------------------------|
| `x`       | `std::vector<double>`   | Best solution vector.                            |
| `fun`     | `double`                | Objective value at `x`, in the original sign.    |
| `nit`     | `std::size_t`           | Outer GRASP iterations executed.                 |
| `nfev`    | `std::size_t`           | Total objective-function evaluations.            |
| `success` | `bool`                  | True when `fun` is finite.                       |
| `message` | `std::string`           | Human-readable termination reason.               |

### Running tests and benchmarks

```bash
# Configure and build
cmake -S cpp -B build -DCMAKE_BUILD_TYPE=Release -DGIVP_BUILD_TESTS=ON
cmake --build build

# Run the 41 Catch2 test cases
ctest --test-dir build --output-on-failure

# Build and run nanobench benchmarks
cmake -S cpp -B build_bench -DCMAKE_BUILD_TYPE=Release -DGIVP_BUILD_BENCHMARKS=ON
cmake --build build_bench
./build_bench/benchmarks/givp_benchmarks
```

---

## R

The R port mirrors the same algorithm and public API shape as the other
languages, with an idiomatic R interface.

### R quick start

```r
library(givp)

sphere <- function(x) sum(x^2)
bounds <- replicate(10, c(-5, 5), simplify = FALSE)

result <- givp(sphere, bounds)
print(result$fun)
print(result$nfev)
print(result$nit)
```

Maximization:

```r
result <- givp(my_score, bounds, direction = "maximize")
```

### R bounds and integer variables

You can pass bounds either as:

- list of pairs, e.g. `list(c(-5, 5), c(0, 10), c(-1, 1))`
- two-column numeric matrix

For mixed continuous/integer problems, set `integer_split` in config:

```r
cfg <- givp_config(integer_split = 8)
result <- givp(my_obj, bounds, config = cfg)
```

### R configuration cookbook

```r
# Fast triage
cfg_fast <- givp_config(
  max_iterations = 20,
  vnd_iterations = 50,
  ils_iterations = 5,
  use_elite_pool = FALSE,
  use_convergence_monitor = FALSE
)

# Quality run with wall-clock budget
cfg_quality <- givp_config(
  max_iterations = 200,
  vnd_iterations = 300,
  ils_iterations = 15,
  elite_size = 10,
  path_relink_frequency = 5,
  adaptive_alpha = TRUE,
  alpha_min = 0.05,
  alpha_max = 0.20,
  time_limit = 600
)
```

### Running R tests and lint

```bash
GIVP_R_TEST_PROFILE=quick Rscript -e "library(testthat); testthat::test_dir('r/tests/testthat')"
GIVP_R_TEST_PROFILE=full Rscript -e "library(testthat); testthat::test_dir('r/tests/testthat')"
Rscript -e "lintr::lint_package('r')"
```

---

## Comparison with other optimizers

| Library                                  | Sense convention                  | Discrete vars?  | Built-in cache | Built-in time budget | Language          |
|------------------------------------------|-----------------------------------|-----------------|----------------|----------------------|-------------------|
| `scipy.optimize.minimize`                | Always minimize                   | No              | No             | No                   | Python            |
| `scipy.optimize.differential_evolution`  | Always minimize                   | Continuous only | No             | Via callback         | Python            |
| `scipy.optimize.dual_annealing`          | Always minimize                   | No              | No             | `maxiter` only       | Python            |
| `optuna`                                 | Explicit (`direction`)            | Yes             | Per-trial only | Yes (`timeout`)      | Python            |
| `pygad`                                  | Always maximize                   | Yes             | No             | No                   | Python            |
| **`givp`** (Python / Julia / Rust)       | Explicit (`minimize`/`direction`) | Yes (mixed)     | LRU cache      | Yes (`time_limit`)   | Python+Julia+Rust |
| **`givp`** (C++)                         | Explicit (`Direction` enum)       | Yes (mixed)     | LRU cache      | Yes (`time_limit`)   | C++17             |
| **`givp`** (R)                           | Explicit (`direction`)            | Yes (mixed)     | LRU cache      | Yes (`time_limit`)   | R                 |

---

## Empirical results

Results of **30 independent runs** (seeds 0–29) on four standard continuous benchmarks
(`n = 10` dimensions) comparing the full GIVP pipeline against a GRASP-only baseline
(construction phase only, no ILS/VND/path relinking — equivalent to plain GRASP
as described by Feo & Resende, 1995).

Each run uses an explicit seed for full reproducibility.  To reproduce, run the
[`Notebooks/benchmark_literature_comparison.ipynb`](Notebooks/benchmark_literature_comparison.ipynb)
notebook.

> **Note**: Values below are representative results from the notebook execution.
> Run the notebook on your machine to obtain hardware-specific timings.

### Sphere — $f(\mathbf{x}) = \sum x_i^2$, global minimum 0 at **0**

| Algorithm   | Mean ± Std          | Best         | Median       | NFev (mean) |
|-------------|---------------------|--------------|--------------|-------------|
| GIVP-full   | 0.0002 ± 0.0001     | 6.2935e-05   | 1.9862e-04   | 605 626     |
| GRASP-only  | 2.1568 ± 0.4827     | 1.0564e+00   | 2.1947e+00   | 4 828       |

### Rosenbrock — global minimum 0 at **1**

| Algorithm   | Mean ± Std          | Best         | Median       | NFev (mean) |
|-------------|---------------------|--------------|--------------|-------------|
| GIVP-full   | 0.513 ± 0.325       | 1.0413e-02   | 5.1000e-01   | 571 765     |
| GRASP-only  | 5441.154 ± 2232.312 | 1.6380e+03   | 5.3735e+03   | 4 834       |

### Rastrigin — multimodal, global minimum 0 at **0**

| Algorithm   | Mean ± Std          | Best         | Median       | NFev (mean) |
|-------------|---------------------|--------------|--------------|-------------|
| GIVP-full   | 0.8492 ± 0.6247     | 9.1427e-03   | 1.0144e+00   | 524 825     |
| GRASP-only  | 28.0927 ± 4.9224    | 1.8621e+01   | 2.9218e+01   | 4 898       |

### Ackley — multimodal, global minimum 0 at **0**

| Algorithm   | Mean ± Std          | Best         | Median       | NFev (mean) |
|-------------|---------------------|--------------|--------------|-------------|
| GIVP-full   | 0.1525 ± 0.0277     | 1.0691e-01   | 1.5015e-01   | 477 840     |
| GRASP-only  | 8.9242 ± 1.0037     | 7.2171e+00   | 9.2427e+00   | 4 852       |

#### **References**

- Feo, T.A. & Resende, M.G.C. (1995). *Greedy randomized adaptive search procedures.*
  Journal of Global Optimization, 6, 109–133.
- Festa, P. & Resende, M.G.C. (2011). *GRASP: An annotated bibliography.*
  Essays and Surveys in Metaheuristics. Springer.

---

## Troubleshooting

**`ValueError: each element of upper must be strictly greater than lower`**
A bounds entry has `low >= high`. Even fixed values must use a strictly
positive interval (`(v - 1e-9, v + 1e-9)`) or be removed from the search.

**`ValueError: bounds length (...) does not match num_vars (...)`**
You passed `num_vars` explicitly but the bounds disagree. Drop `num_vars` to
let the library infer it from `bounds`, or fix the mismatch.

**`ValueError: 'minimize' and 'direction' disagree: ...`**
You passed both flags with conflicting values. Use one or the other (or pass
both with matching values).

**Optimization converges to `inf`.**
Your objective is raising or returning `nan`. The wrapper coerces non-finite
values to `+inf` so they are always comparable, but if *every* candidate is
infeasible the algorithm has nothing to improve. Lower `perturbation_strength`,
revisit your bounds, or relax the feasibility logic in `func`.

**Run is too slow.**
Try `use_cache=True` (with `givp[cache]` for maximum speed), increase
`cache_size`, raise `n_workers` (2–4 is a practical sweet spot under the
Python GIL; set `n_workers=os.cpu_count()` for NumPy-heavy objectives),
lower `num_candidates_per_step`, or set a `time_limit`. For very expensive
objectives, also reduce `vnd_iterations` and `ils_iterations`.

**Final solution looks too "rough" / integer values look noisy.**
Make sure `integer_split` is set correctly. With the default (`None` /
`num_vars`) all variables are treated as continuous and the integer-aware
neighborhoods are skipped.

---

## License

MIT