forked from scikit-learn/scikit-learn.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathgaussian_process.html
804 lines (772 loc) · 70.7 KB
/
gaussian_process.html
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
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="generator" content="Docutils 0.19: https://docutils.sourceforge.io/" />
<meta property="og:title" content="1.7. Gaussian Processes" />
<meta property="og:type" content="website" />
<meta property="og:url" content="https://scikit-learn/stable/modules/gaussian_process.html" />
<meta property="og:site_name" content="scikit-learn" />
<meta property="og:description" content="Gaussian Processes (GP) are a generic supervised learning method designed to solve regression and probabilistic classification problems. The advantages of Gaussian processes are: The prediction int..." />
<meta property="og:image" content="https://scikit-learn/stable/_images/sphx_glr_plot_gpr_noisy_003.png" />
<meta property="og:image:alt" content="scikit-learn" />
<meta name="description" content="Gaussian Processes (GP) are a generic supervised learning method designed to solve regression and probabilistic classification problems. The advantages of Gaussian processes are: The prediction int..." />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>1.7. Gaussian Processes — scikit-learn 1.4.dev0 documentation</title>
<link rel="canonical" href="http://scikit-learn.org/stable/modules/gaussian_process.html" />
<link rel="shortcut icon" href="../_static/favicon.ico"/>
<link rel="stylesheet" href="../_static/css/vendor/bootstrap.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="../_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="../_static/plot_directive.css" type="text/css" />
<link rel="stylesheet" href="https://fonts.googleapis.com/css?family=Vibur" type="text/css" />
<link rel="stylesheet" href="../_static/jupyterlite_sphinx.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-binder.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-dataframe.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-rendered-html.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<script id="documentation_options" data-url_root="../" src="../_static/documentation_options.js"></script>
<script src="../_static/js/vendor/jquery-3.6.3.slim.min.js"></script>
</head>
<body>
<nav id="navbar" class="sk-docs-navbar navbar navbar-expand-md navbar-light bg-light py-0">
<div class="container-fluid sk-docs-container px-0">
<a class="navbar-brand py-0" href="../index.html">
<img
class="sk-brand-img"
src="../_static/scikit-learn-logo-small.png"
alt="logo"/>
</a>
<button
id="sk-navbar-toggler"
class="navbar-toggler"
type="button"
data-toggle="collapse"
data-target="#navbarSupportedContent"
aria-controls="navbarSupportedContent"
aria-expanded="false"
aria-label="Toggle navigation"
>
<span class="navbar-toggler-icon"></span>
</button>
<div class="sk-navbar-collapse collapse navbar-collapse" id="navbarSupportedContent">
<ul class="navbar-nav mr-auto">
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../install.html">Install</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../user_guide.html">User Guide</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="classes.html">API</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../auto_examples/index.html">Examples</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" target="_blank" rel="noopener noreferrer" href="https://blog.scikit-learn.org/">Community</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../getting_started.html" >Getting Started</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../tutorial/index.html" >Tutorial</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../whats_new/v1.4.html" >What's new</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../glossary.html" >Glossary</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../developers/index.html" >Development</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../faq.html" >FAQ</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../support.html" >Support</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../related_projects.html" >Related packages</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../roadmap.html" >Roadmap</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../governance.html" >Governance</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../about.html" >About us</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="https://github.com/scikit-learn/scikit-learn" >GitHub</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="https://scikit-learn.org/dev/versions.html" >Other Versions and Download</a>
</li>
<li class="nav-item dropdown nav-more-item-dropdown">
<a class="sk-nav-link nav-link dropdown-toggle" href="#" id="navbarDropdown" role="button" data-toggle="dropdown" aria-haspopup="true" aria-expanded="false">More</a>
<div class="dropdown-menu" aria-labelledby="navbarDropdown">
<a class="sk-nav-dropdown-item dropdown-item" href="../getting_started.html" >Getting Started</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../tutorial/index.html" >Tutorial</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../whats_new/v1.4.html" >What's new</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../glossary.html" >Glossary</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../developers/index.html" >Development</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../faq.html" >FAQ</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../support.html" >Support</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../related_projects.html" >Related packages</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../roadmap.html" >Roadmap</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../governance.html" >Governance</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../about.html" >About us</a>
<a class="sk-nav-dropdown-item dropdown-item" href="https://github.com/scikit-learn/scikit-learn" >GitHub</a>
<a class="sk-nav-dropdown-item dropdown-item" href="https://scikit-learn.org/dev/versions.html" >Other Versions and Download</a>
</div>
</li>
</ul>
<div id="searchbox" role="search">
<div class="searchformwrapper">
<form class="search" action="../search.html" method="get">
<input class="sk-search-text-input" type="text" name="q" aria-labelledby="searchlabel" />
<input class="sk-search-text-btn" type="submit" value="Go" />
</form>
</div>
</div>
</div>
</div>
</nav>
<div class="d-flex" id="sk-doc-wrapper">
<input type="checkbox" name="sk-toggle-checkbox" id="sk-toggle-checkbox">
<label id="sk-sidemenu-toggle" class="sk-btn-toggle-toc btn sk-btn-primary" for="sk-toggle-checkbox">Toggle Menu</label>
<div id="sk-sidebar-wrapper" class="border-right">
<div class="sk-sidebar-toc-wrapper">
<div class="btn-group w-100 mb-2" role="group" aria-label="rellinks">
<a href="neighbors.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="1.6. Nearest Neighbors">Prev</a><a href="../supervised_learning.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="1. Supervised learning">Up</a>
<a href="cross_decomposition.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="1.8. Cross decomposition">Next</a>
</div>
<div class="alert alert-danger p-1 mb-2" role="alert">
<p class="text-center mb-0">
<strong>scikit-learn 1.4.dev0</strong><br/>
<a href="http://scikit-learn.org/dev/versions.html">Other versions</a>
</p>
</div>
<div class="alert alert-warning p-1 mb-2" role="alert">
<p class="text-center mb-0">
Please <a class="font-weight-bold" href="../about.html#citing-scikit-learn"><string>cite us</string></a> if you use the software.
</p>
</div>
<div class="sk-sidebar-toc">
<ul>
<li><a class="reference internal" href="#">1.7. Gaussian Processes</a><ul>
<li><a class="reference internal" href="#gaussian-process-regression-gpr">1.7.1. Gaussian Process Regression (GPR)</a></li>
<li><a class="reference internal" href="#gpr-examples">1.7.2. GPR examples</a><ul>
<li><a class="reference internal" href="#gpr-with-noise-level-estimation">1.7.2.1. GPR with noise-level estimation</a></li>
<li><a class="reference internal" href="#comparison-of-gpr-and-kernel-ridge-regression">1.7.2.2. Comparison of GPR and Kernel Ridge Regression</a></li>
<li><a class="reference internal" href="#gpr-on-mauna-loa-co2-data">1.7.2.3. GPR on Mauna Loa CO2 data</a></li>
</ul>
</li>
<li><a class="reference internal" href="#gaussian-process-classification-gpc">1.7.3. Gaussian Process Classification (GPC)</a></li>
<li><a class="reference internal" href="#gpc-examples">1.7.4. GPC examples</a><ul>
<li><a class="reference internal" href="#probabilistic-predictions-with-gpc">1.7.4.1. Probabilistic predictions with GPC</a></li>
<li><a class="reference internal" href="#illustration-of-gpc-on-the-xor-dataset">1.7.4.2. Illustration of GPC on the XOR dataset</a></li>
<li><a class="reference internal" href="#gaussian-process-classification-gpc-on-iris-dataset">1.7.4.3. Gaussian process classification (GPC) on iris dataset</a></li>
</ul>
</li>
<li><a class="reference internal" href="#kernels-for-gaussian-processes">1.7.5. Kernels for Gaussian Processes</a><ul>
<li><a class="reference internal" href="#gaussian-process-kernel-api">1.7.5.1. Gaussian Process Kernel API</a></li>
<li><a class="reference internal" href="#basic-kernels">1.7.5.2. Basic kernels</a></li>
<li><a class="reference internal" href="#kernel-operators">1.7.5.3. Kernel operators</a></li>
<li><a class="reference internal" href="#radial-basis-function-rbf-kernel">1.7.5.4. Radial basis function (RBF) kernel</a></li>
<li><a class="reference internal" href="#matern-kernel">1.7.5.5. Matérn kernel</a></li>
<li><a class="reference internal" href="#rational-quadratic-kernel">1.7.5.6. Rational quadratic kernel</a></li>
<li><a class="reference internal" href="#exp-sine-squared-kernel">1.7.5.7. Exp-Sine-Squared kernel</a></li>
<li><a class="reference internal" href="#dot-product-kernel">1.7.5.8. Dot-Product kernel</a></li>
<li><a class="reference internal" href="#references">1.7.5.9. References</a></li>
</ul>
</li>
</ul>
</li>
</ul>
</div>
</div>
</div>
<div id="sk-page-content-wrapper">
<div class="sk-page-content container-fluid body px-md-3" role="main">
<section id="gaussian-processes">
<span id="gaussian-process"></span><h1><span class="section-number">1.7. </span>Gaussian Processes<a class="headerlink" href="#gaussian-processes" title="Permalink to this heading">¶</a></h1>
<p><strong>Gaussian Processes (GP)</strong> are a generic supervised learning method designed
to solve <em>regression</em> and <em>probabilistic classification</em> problems.</p>
<p>The advantages of Gaussian processes are:</p>
<blockquote>
<div><ul class="simple">
<li><p>The prediction interpolates the observations (at least for regular
kernels).</p></li>
<li><p>The prediction is probabilistic (Gaussian) so that one can compute
empirical confidence intervals and decide based on those if one should
refit (online fitting, adaptive fitting) the prediction in some
region of interest.</p></li>
<li><p>Versatile: different <a class="reference internal" href="#gp-kernels"><span class="std std-ref">kernels</span></a> can be specified. Common kernels are provided, but
it is also possible to specify custom kernels.</p></li>
</ul>
</div></blockquote>
<p>The disadvantages of Gaussian processes include:</p>
<blockquote>
<div><ul class="simple">
<li><p>They are not sparse, i.e., they use the whole samples/features information to
perform the prediction.</p></li>
<li><p>They lose efficiency in high dimensional spaces – namely when the number
of features exceeds a few dozens.</p></li>
</ul>
</div></blockquote>
<section id="gaussian-process-regression-gpr">
<span id="gpr"></span><h2><span class="section-number">1.7.1. </span>Gaussian Process Regression (GPR)<a class="headerlink" href="#gaussian-process-regression-gpr" title="Permalink to this heading">¶</a></h2>
<p>The <a class="reference internal" href="generated/sklearn.gaussian_process.GaussianProcessRegressor.html#sklearn.gaussian_process.GaussianProcessRegressor" title="sklearn.gaussian_process.GaussianProcessRegressor"><code class="xref py py-class docutils literal notranslate"><span class="pre">GaussianProcessRegressor</span></code></a> implements Gaussian processes (GP) for
regression purposes. For this, the prior of the GP needs to be specified. The
prior mean is assumed to be constant and zero (for <code class="docutils literal notranslate"><span class="pre">normalize_y=False</span></code>) or the
training data’s mean (for <code class="docutils literal notranslate"><span class="pre">normalize_y=True</span></code>). The prior’s
covariance is specified by passing a <a class="reference internal" href="#gp-kernels"><span class="std std-ref">kernel</span></a> object. The
hyperparameters of the kernel are optimized during fitting of
GaussianProcessRegressor by maximizing the log-marginal-likelihood (LML) based
on the passed <code class="docutils literal notranslate"><span class="pre">optimizer</span></code>. As the LML may have multiple local optima, the
optimizer can be started repeatedly by specifying <code class="docutils literal notranslate"><span class="pre">n_restarts_optimizer</span></code>. The
first run is always conducted starting from the initial hyperparameter values
of the kernel; subsequent runs are conducted from hyperparameter values
that have been chosen randomly from the range of allowed values.
If the initial hyperparameters should be kept fixed, <code class="docutils literal notranslate"><span class="pre">None</span></code> can be passed as
optimizer.</p>
<p>The noise level in the targets can be specified by passing it via the
parameter <code class="docutils literal notranslate"><span class="pre">alpha</span></code>, either globally as a scalar or per datapoint.
Note that a moderate noise level can also be helpful for dealing with numeric
issues during fitting as it is effectively implemented as Tikhonov
regularization, i.e., by adding it to the diagonal of the kernel matrix. An
alternative to specifying the noise level explicitly is to include a
WhiteKernel component into the kernel, which can estimate the global noise
level from the data (see example below).</p>
<p>The implementation is based on Algorithm 2.1 of <a class="reference internal" href="#rw2006" id="id1"><span>[RW2006]</span></a>. In addition to
the API of standard scikit-learn estimators, GaussianProcessRegressor:</p>
<ul class="simple">
<li><p>allows prediction without prior fitting (based on the GP prior)</p></li>
<li><p>provides an additional method <code class="docutils literal notranslate"><span class="pre">sample_y(X)</span></code>, which evaluates samples
drawn from the GPR (prior or posterior) at given inputs</p></li>
<li><p>exposes a method <code class="docutils literal notranslate"><span class="pre">log_marginal_likelihood(theta)</span></code>, which can be used
externally for other ways of selecting hyperparameters, e.g., via
Markov chain Monte Carlo.</p></li>
</ul>
</section>
<section id="gpr-examples">
<h2><span class="section-number">1.7.2. </span>GPR examples<a class="headerlink" href="#gpr-examples" title="Permalink to this heading">¶</a></h2>
<section id="gpr-with-noise-level-estimation">
<h3><span class="section-number">1.7.2.1. </span>GPR with noise-level estimation<a class="headerlink" href="#gpr-with-noise-level-estimation" title="Permalink to this heading">¶</a></h3>
<p>This example illustrates that GPR with a sum-kernel including a WhiteKernel can
estimate the noise level of data. An illustration of the
log-marginal-likelihood (LML) landscape shows that there exist two local
maxima of LML.</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpr_noisy.html"><img alt="../_images/sphx_glr_plot_gpr_noisy_003.png" src="../_images/sphx_glr_plot_gpr_noisy_003.png" /></a>
</figure>
<p>The first corresponds to a model with a high noise level and a
large length scale, which explains all variations in the data by noise.</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpr_noisy.html"><img alt="../_images/sphx_glr_plot_gpr_noisy_004.png" src="../_images/sphx_glr_plot_gpr_noisy_004.png" /></a>
</figure>
<p>The second one has a smaller noise level and shorter length scale, which explains
most of the variation by the noise-free functional relationship. The second
model has a higher likelihood; however, depending on the initial value for the
hyperparameters, the gradient-based optimization might also converge to the
high-noise solution. It is thus important to repeat the optimization several
times for different initializations.</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpr_noisy.html"><img alt="../_images/sphx_glr_plot_gpr_noisy_005.png" src="../_images/sphx_glr_plot_gpr_noisy_005.png" /></a>
</figure>
</section>
<section id="comparison-of-gpr-and-kernel-ridge-regression">
<h3><span class="section-number">1.7.2.2. </span>Comparison of GPR and Kernel Ridge Regression<a class="headerlink" href="#comparison-of-gpr-and-kernel-ridge-regression" title="Permalink to this heading">¶</a></h3>
<p>Both kernel ridge regression (KRR) and GPR learn
a target function by employing internally the “kernel trick”. KRR learns a
linear function in the space induced by the respective kernel which corresponds
to a non-linear function in the original space. The linear function in the
kernel space is chosen based on the mean-squared error loss with
ridge regularization. GPR uses the kernel to define the covariance of
a prior distribution over the target functions and uses the observed training
data to define a likelihood function. Based on Bayes theorem, a (Gaussian)
posterior distribution over target functions is defined, whose mean is used
for prediction.</p>
<p>A major difference is that GPR can choose the kernel’s hyperparameters based
on gradient-ascent on the marginal likelihood function while KRR needs to
perform a grid search on a cross-validated loss function (mean-squared error
loss). A further difference is that GPR learns a generative, probabilistic
model of the target function and can thus provide meaningful confidence
intervals and posterior samples along with the predictions while KRR only
provides predictions.</p>
<p>The following figure illustrates both methods on an artificial dataset, which
consists of a sinusoidal target function and strong noise. The figure compares
the learned model of KRR and GPR based on a ExpSineSquared kernel, which is
suited for learning periodic functions. The kernel’s hyperparameters control
the smoothness (length_scale) and periodicity of the kernel (periodicity).
Moreover, the noise level
of the data is learned explicitly by GPR by an additional WhiteKernel component
in the kernel and by the regularization parameter alpha of KRR.</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_compare_gpr_krr.html"><img alt="../_images/sphx_glr_plot_compare_gpr_krr_005.png" src="../_images/sphx_glr_plot_compare_gpr_krr_005.png" /></a>
</figure>
<p>The figure shows that both methods learn reasonable models of the target
function. GPR provides reasonable confidence bounds on the prediction which are not
available for KRR. A major difference between the two methods is the time
required for fitting and predicting: while fitting KRR is fast in principle,
the grid-search for hyperparameter optimization scales exponentially with the
number of hyperparameters (“curse of dimensionality”). The gradient-based
optimization of the parameters in GPR does not suffer from this exponential
scaling and is thus considerably faster on this example with 3-dimensional
hyperparameter space. The time for predicting is similar; however, generating
the variance of the predictive distribution of GPR takes considerably longer
than just predicting the mean.</p>
</section>
<section id="gpr-on-mauna-loa-co2-data">
<h3><span class="section-number">1.7.2.3. </span>GPR on Mauna Loa CO2 data<a class="headerlink" href="#gpr-on-mauna-loa-co2-data" title="Permalink to this heading">¶</a></h3>
<p>This example is based on Section 5.4.3 of <a class="reference internal" href="#rw2006" id="id2"><span>[RW2006]</span></a>.
It illustrates an example of complex kernel engineering and
hyperparameter optimization using gradient ascent on the
log-marginal-likelihood. The data consists of the monthly average atmospheric
CO2 concentrations (in parts per million by volume (ppmv)) collected at the
Mauna Loa Observatory in Hawaii, between 1958 and 1997. The objective is to
model the CO2 concentration as a function of the time t.</p>
<p>The kernel is composed of several terms that are responsible for explaining
different properties of the signal:</p>
<ul class="simple">
<li><p>a long term, smooth rising trend is to be explained by an RBF kernel. The
RBF kernel with a large length-scale enforces this component to be smooth;
it is not enforced that the trend is rising which leaves this choice to the
GP. The specific length-scale and the amplitude are free hyperparameters.</p></li>
<li><p>a seasonal component, which is to be explained by the periodic
ExpSineSquared kernel with a fixed periodicity of 1 year. The length-scale
of this periodic component, controlling its smoothness, is a free parameter.
In order to allow decaying away from exact periodicity, the product with an
RBF kernel is taken. The length-scale of this RBF component controls the
decay time and is a further free parameter.</p></li>
<li><p>smaller, medium term irregularities are to be explained by a
RationalQuadratic kernel component, whose length-scale and alpha parameter,
which determines the diffuseness of the length-scales, are to be determined.
According to <a class="reference internal" href="#rw2006" id="id3"><span>[RW2006]</span></a>, these irregularities can better be explained by
a RationalQuadratic than an RBF kernel component, probably because it can
accommodate several length-scales.</p></li>
<li><p>a “noise” term, consisting of an RBF kernel contribution, which shall
explain the correlated noise components such as local weather phenomena,
and a WhiteKernel contribution for the white noise. The relative amplitudes
and the RBF’s length scale are further free parameters.</p></li>
</ul>
<p>Maximizing the log-marginal-likelihood after subtracting the target’s mean
yields the following kernel with an LML of -83.214:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="mf">34.4</span><span class="o">**</span><span class="mi">2</span> <span class="o">*</span> <span class="n">RBF</span><span class="p">(</span><span class="n">length_scale</span><span class="o">=</span><span class="mf">41.8</span><span class="p">)</span>
<span class="o">+</span> <span class="mf">3.27</span><span class="o">**</span><span class="mi">2</span> <span class="o">*</span> <span class="n">RBF</span><span class="p">(</span><span class="n">length_scale</span><span class="o">=</span><span class="mi">180</span><span class="p">)</span> <span class="o">*</span> <span class="n">ExpSineSquared</span><span class="p">(</span><span class="n">length_scale</span><span class="o">=</span><span class="mf">1.44</span><span class="p">,</span>
<span class="n">periodicity</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="o">+</span> <span class="mf">0.446</span><span class="o">**</span><span class="mi">2</span> <span class="o">*</span> <span class="n">RationalQuadratic</span><span class="p">(</span><span class="n">alpha</span><span class="o">=</span><span class="mf">17.7</span><span class="p">,</span> <span class="n">length_scale</span><span class="o">=</span><span class="mf">0.957</span><span class="p">)</span>
<span class="o">+</span> <span class="mf">0.197</span><span class="o">**</span><span class="mi">2</span> <span class="o">*</span> <span class="n">RBF</span><span class="p">(</span><span class="n">length_scale</span><span class="o">=</span><span class="mf">0.138</span><span class="p">)</span> <span class="o">+</span> <span class="n">WhiteKernel</span><span class="p">(</span><span class="n">noise_level</span><span class="o">=</span><span class="mf">0.0336</span><span class="p">)</span>
</pre></div>
</div>
<p>Thus, most of the target signal (34.4ppm) is explained by a long-term rising
trend (length-scale 41.8 years). The periodic component has an amplitude of
3.27ppm, a decay time of 180 years and a length-scale of 1.44. The long decay
time indicates that we have a locally very close to periodic seasonal
component. The correlated noise has an amplitude of 0.197ppm with a length
scale of 0.138 years and a white-noise contribution of 0.197ppm. Thus, the
overall noise level is very small, indicating that the data can be very well
explained by the model. The figure shows also that the model makes very
confident predictions until around 2015</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpr_co2.html"><img alt="../_images/sphx_glr_plot_gpr_co2_003.png" src="../_images/sphx_glr_plot_gpr_co2_003.png" /></a>
</figure>
</section>
</section>
<section id="gaussian-process-classification-gpc">
<span id="gpc"></span><h2><span class="section-number">1.7.3. </span>Gaussian Process Classification (GPC)<a class="headerlink" href="#gaussian-process-classification-gpc" title="Permalink to this heading">¶</a></h2>
<p>The <a class="reference internal" href="generated/sklearn.gaussian_process.GaussianProcessClassifier.html#sklearn.gaussian_process.GaussianProcessClassifier" title="sklearn.gaussian_process.GaussianProcessClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GaussianProcessClassifier</span></code></a> implements Gaussian processes (GP) for
classification purposes, more specifically for probabilistic classification,
where test predictions take the form of class probabilities.
GaussianProcessClassifier places a GP prior on a latent function <span class="math notranslate nohighlight">\(f\)</span>,
which is then squashed through a link function to obtain the probabilistic
classification. The latent function <span class="math notranslate nohighlight">\(f\)</span> is a so-called nuisance function,
whose values are not observed and are not relevant by themselves.
Its purpose is to allow a convenient formulation of the model, and <span class="math notranslate nohighlight">\(f\)</span>
is removed (integrated out) during prediction. GaussianProcessClassifier
implements the logistic link function, for which the integral cannot be
computed analytically but is easily approximated in the binary case.</p>
<p>In contrast to the regression setting, the posterior of the latent function
<span class="math notranslate nohighlight">\(f\)</span> is not Gaussian even for a GP prior since a Gaussian likelihood is
inappropriate for discrete class labels. Rather, a non-Gaussian likelihood
corresponding to the logistic link function (logit) is used.
GaussianProcessClassifier approximates the non-Gaussian posterior with a
Gaussian based on the Laplace approximation. More details can be found in
Chapter 3 of <a class="reference internal" href="#rw2006" id="id4"><span>[RW2006]</span></a>.</p>
<p>The GP prior mean is assumed to be zero. The prior’s
covariance is specified by passing a <a class="reference internal" href="#gp-kernels"><span class="std std-ref">kernel</span></a> object. The
hyperparameters of the kernel are optimized during fitting of
GaussianProcessRegressor by maximizing the log-marginal-likelihood (LML) based
on the passed <code class="docutils literal notranslate"><span class="pre">optimizer</span></code>. As the LML may have multiple local optima, the
optimizer can be started repeatedly by specifying <code class="docutils literal notranslate"><span class="pre">n_restarts_optimizer</span></code>. The
first run is always conducted starting from the initial hyperparameter values
of the kernel; subsequent runs are conducted from hyperparameter values
that have been chosen randomly from the range of allowed values.
If the initial hyperparameters should be kept fixed, <code class="docutils literal notranslate"><span class="pre">None</span></code> can be passed as
optimizer.</p>
<p><a class="reference internal" href="generated/sklearn.gaussian_process.GaussianProcessClassifier.html#sklearn.gaussian_process.GaussianProcessClassifier" title="sklearn.gaussian_process.GaussianProcessClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GaussianProcessClassifier</span></code></a> supports multi-class classification
by performing either one-versus-rest or one-versus-one based training and
prediction. In one-versus-rest, one binary Gaussian process classifier is
fitted for each class, which is trained to separate this class from the rest.
In “one_vs_one”, one binary Gaussian process classifier is fitted for each pair
of classes, which is trained to separate these two classes. The predictions of
these binary predictors are combined into multi-class predictions. See the
section on <a class="reference internal" href="multiclass.html#multiclass"><span class="std std-ref">multi-class classification</span></a> for more details.</p>
<p>In the case of Gaussian process classification, “one_vs_one” might be
computationally cheaper since it has to solve many problems involving only a
subset of the whole training set rather than fewer problems on the whole
dataset. Since Gaussian process classification scales cubically with the size
of the dataset, this might be considerably faster. However, note that
“one_vs_one” does not support predicting probability estimates but only plain
predictions. Moreover, note that <a class="reference internal" href="generated/sklearn.gaussian_process.GaussianProcessClassifier.html#sklearn.gaussian_process.GaussianProcessClassifier" title="sklearn.gaussian_process.GaussianProcessClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">GaussianProcessClassifier</span></code></a> does not
(yet) implement a true multi-class Laplace approximation internally, but
as discussed above is based on solving several binary classification tasks
internally, which are combined using one-versus-rest or one-versus-one.</p>
</section>
<section id="gpc-examples">
<h2><span class="section-number">1.7.4. </span>GPC examples<a class="headerlink" href="#gpc-examples" title="Permalink to this heading">¶</a></h2>
<section id="probabilistic-predictions-with-gpc">
<h3><span class="section-number">1.7.4.1. </span>Probabilistic predictions with GPC<a class="headerlink" href="#probabilistic-predictions-with-gpc" title="Permalink to this heading">¶</a></h3>
<p>This example illustrates the predicted probability of GPC for an RBF kernel
with different choices of the hyperparameters. The first figure shows the
predicted probability of GPC with arbitrarily chosen hyperparameters and with
the hyperparameters corresponding to the maximum log-marginal-likelihood (LML).</p>
<p>While the hyperparameters chosen by optimizing LML have a considerably larger
LML, they perform slightly worse according to the log-loss on test data. The
figure shows that this is because they exhibit a steep change of the class
probabilities at the class boundaries (which is good) but have predicted
probabilities close to 0.5 far away from the class boundaries (which is bad)
This undesirable effect is caused by the Laplace approximation used
internally by GPC.</p>
<p>The second figure shows the log-marginal-likelihood for different choices of
the kernel’s hyperparameters, highlighting the two choices of the
hyperparameters used in the first figure by black dots.</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpc.html"><img alt="../_images/sphx_glr_plot_gpc_001.png" src="../_images/sphx_glr_plot_gpc_001.png" /></a>
</figure>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpc.html"><img alt="../_images/sphx_glr_plot_gpc_002.png" src="../_images/sphx_glr_plot_gpc_002.png" /></a>
</figure>
</section>
<section id="illustration-of-gpc-on-the-xor-dataset">
<h3><span class="section-number">1.7.4.2. </span>Illustration of GPC on the XOR dataset<a class="headerlink" href="#illustration-of-gpc-on-the-xor-dataset" title="Permalink to this heading">¶</a></h3>
<p>This example illustrates GPC on XOR data. Compared are a stationary, isotropic
kernel (<a class="reference internal" href="generated/sklearn.gaussian_process.kernels.RBF.html#sklearn.gaussian_process.kernels.RBF" title="sklearn.gaussian_process.kernels.RBF"><code class="xref py py-class docutils literal notranslate"><span class="pre">RBF</span></code></a>) and a non-stationary kernel (<a class="reference internal" href="generated/sklearn.gaussian_process.kernels.DotProduct.html#sklearn.gaussian_process.kernels.DotProduct" title="sklearn.gaussian_process.kernels.DotProduct"><code class="xref py py-class docutils literal notranslate"><span class="pre">DotProduct</span></code></a>). On
this particular dataset, the <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.DotProduct.html#sklearn.gaussian_process.kernels.DotProduct" title="sklearn.gaussian_process.kernels.DotProduct"><code class="xref py py-class docutils literal notranslate"><span class="pre">DotProduct</span></code></a> kernel obtains considerably
better results because the class-boundaries are linear and coincide with the
coordinate axes. In practice, however, stationary kernels such as <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.RBF.html#sklearn.gaussian_process.kernels.RBF" title="sklearn.gaussian_process.kernels.RBF"><code class="xref py py-class docutils literal notranslate"><span class="pre">RBF</span></code></a>
often obtain better results.</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpc_xor.html"><img alt="../_images/sphx_glr_plot_gpc_xor_001.png" src="../_images/sphx_glr_plot_gpc_xor_001.png" /></a>
</figure>
</section>
<section id="gaussian-process-classification-gpc-on-iris-dataset">
<h3><span class="section-number">1.7.4.3. </span>Gaussian process classification (GPC) on iris dataset<a class="headerlink" href="#gaussian-process-classification-gpc-on-iris-dataset" title="Permalink to this heading">¶</a></h3>
<p>This example illustrates the predicted probability of GPC for an isotropic
and anisotropic RBF kernel on a two-dimensional version for the iris-dataset.
This illustrates the applicability of GPC to non-binary classification.
The anisotropic RBF kernel obtains slightly higher log-marginal-likelihood by
assigning different length-scales to the two feature dimensions.</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpc_iris.html"><img alt="../_images/sphx_glr_plot_gpc_iris_001.png" src="../_images/sphx_glr_plot_gpc_iris_001.png" /></a>
</figure>
</section>
</section>
<section id="kernels-for-gaussian-processes">
<span id="gp-kernels"></span><h2><span class="section-number">1.7.5. </span>Kernels for Gaussian Processes<a class="headerlink" href="#kernels-for-gaussian-processes" title="Permalink to this heading">¶</a></h2>
<p>Kernels (also called “covariance functions” in the context of GPs) are a crucial
ingredient of GPs which determine the shape of prior and posterior of the GP.
They encode the assumptions on the function being learned by defining the “similarity”
of two datapoints combined with the assumption that similar datapoints should
have similar target values. Two categories of kernels can be distinguished:
stationary kernels depend only on the distance of two datapoints and not on their
absolute values <span class="math notranslate nohighlight">\(k(x_i, x_j)= k(d(x_i, x_j))\)</span> and are thus invariant to
translations in the input space, while non-stationary kernels
depend also on the specific values of the datapoints. Stationary kernels can further
be subdivided into isotropic and anisotropic kernels, where isotropic kernels are
also invariant to rotations in the input space. For more details, we refer to
Chapter 4 of <a class="reference internal" href="#rw2006" id="id5"><span>[RW2006]</span></a>. For guidance on how to best combine different kernels,
we refer to <a class="reference internal" href="#duv2014" id="id6"><span>[Duv2014]</span></a>.</p>
<section id="gaussian-process-kernel-api">
<h3><span class="section-number">1.7.5.1. </span>Gaussian Process Kernel API<a class="headerlink" href="#gaussian-process-kernel-api" title="Permalink to this heading">¶</a></h3>
<p>The main usage of a <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.Kernel.html#sklearn.gaussian_process.kernels.Kernel" title="sklearn.gaussian_process.kernels.Kernel"><code class="xref py py-class docutils literal notranslate"><span class="pre">Kernel</span></code></a> is to compute the GP’s covariance between
datapoints. For this, the method <code class="docutils literal notranslate"><span class="pre">__call__</span></code> of the kernel can be called. This
method can either be used to compute the “auto-covariance” of all pairs of
datapoints in a 2d array X, or the “cross-covariance” of all combinations
of datapoints of a 2d array X with datapoints in a 2d array Y. The following
identity holds true for all kernels k (except for the <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.WhiteKernel.html#sklearn.gaussian_process.kernels.WhiteKernel" title="sklearn.gaussian_process.kernels.WhiteKernel"><code class="xref py py-class docutils literal notranslate"><span class="pre">WhiteKernel</span></code></a>):
<code class="docutils literal notranslate"><span class="pre">k(X)</span> <span class="pre">==</span> <span class="pre">K(X,</span> <span class="pre">Y=X)</span></code></p>
<p>If only the diagonal of the auto-covariance is being used, the method <code class="docutils literal notranslate"><span class="pre">diag()</span></code>
of a kernel can be called, which is more computationally efficient than the
equivalent call to <code class="docutils literal notranslate"><span class="pre">__call__</span></code>: <code class="docutils literal notranslate"><span class="pre">np.diag(k(X,</span> <span class="pre">X))</span> <span class="pre">==</span> <span class="pre">k.diag(X)</span></code></p>
<p>Kernels are parameterized by a vector <span class="math notranslate nohighlight">\(\theta\)</span> of hyperparameters. These
hyperparameters can for instance control length-scales or periodicity of a
kernel (see below). All kernels support computing analytic gradients
of the kernel’s auto-covariance with respect to <span class="math notranslate nohighlight">\(log(\theta)\)</span> via setting
<code class="docutils literal notranslate"><span class="pre">eval_gradient=True</span></code> in the <code class="docutils literal notranslate"><span class="pre">__call__</span></code> method.
That is, a <code class="docutils literal notranslate"><span class="pre">(len(X),</span> <span class="pre">len(X),</span> <span class="pre">len(theta))</span></code> array is returned where the entry
<code class="docutils literal notranslate"><span class="pre">[i,</span> <span class="pre">j,</span> <span class="pre">l]</span></code> contains <span class="math notranslate nohighlight">\(\frac{\partial k_\theta(x_i, x_j)}{\partial log(\theta_l)}\)</span>.
This gradient is used by the Gaussian process (both regressor and classifier)
in computing the gradient of the log-marginal-likelihood, which in turn is used
to determine the value of <span class="math notranslate nohighlight">\(\theta\)</span>, which maximizes the log-marginal-likelihood,
via gradient ascent. For each hyperparameter, the initial value and the
bounds need to be specified when creating an instance of the kernel. The
current value of <span class="math notranslate nohighlight">\(\theta\)</span> can be get and set via the property
<code class="docutils literal notranslate"><span class="pre">theta</span></code> of the kernel object. Moreover, the bounds of the hyperparameters can be
accessed by the property <code class="docutils literal notranslate"><span class="pre">bounds</span></code> of the kernel. Note that both properties
(theta and bounds) return log-transformed values of the internally used values
since those are typically more amenable to gradient-based optimization.
The specification of each hyperparameter is stored in the form of an instance of
<a class="reference internal" href="generated/sklearn.gaussian_process.kernels.Hyperparameter.html#sklearn.gaussian_process.kernels.Hyperparameter" title="sklearn.gaussian_process.kernels.Hyperparameter"><code class="xref py py-class docutils literal notranslate"><span class="pre">Hyperparameter</span></code></a> in the respective kernel. Note that a kernel using a
hyperparameter with name “x” must have the attributes self.x and self.x_bounds.</p>
<p>The abstract base class for all kernels is <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.Kernel.html#sklearn.gaussian_process.kernels.Kernel" title="sklearn.gaussian_process.kernels.Kernel"><code class="xref py py-class docutils literal notranslate"><span class="pre">Kernel</span></code></a>. Kernel implements a
similar interface as <a class="reference internal" href="generated/sklearn.base.BaseEstimator.html#sklearn.base.BaseEstimator" title="sklearn.base.BaseEstimator"><code class="xref py py-class docutils literal notranslate"><span class="pre">BaseEstimator</span></code></a>, providing the
methods <code class="docutils literal notranslate"><span class="pre">get_params()</span></code>, <code class="docutils literal notranslate"><span class="pre">set_params()</span></code>, and <code class="docutils literal notranslate"><span class="pre">clone()</span></code>. This allows
setting kernel values also via meta-estimators such as
<a class="reference internal" href="generated/sklearn.pipeline.Pipeline.html#sklearn.pipeline.Pipeline" title="sklearn.pipeline.Pipeline"><code class="xref py py-class docutils literal notranslate"><span class="pre">Pipeline</span></code></a> or
<a class="reference internal" href="generated/sklearn.model_selection.GridSearchCV.html#sklearn.model_selection.GridSearchCV" title="sklearn.model_selection.GridSearchCV"><code class="xref py py-class docutils literal notranslate"><span class="pre">GridSearchCV</span></code></a>. Note that due to the nested
structure of kernels (by applying kernel operators, see below), the names of
kernel parameters might become relatively complicated. In general, for a binary
kernel operator, parameters of the left operand are prefixed with <code class="docutils literal notranslate"><span class="pre">k1__</span></code> and
parameters of the right operand with <code class="docutils literal notranslate"><span class="pre">k2__</span></code>. An additional convenience method
is <code class="docutils literal notranslate"><span class="pre">clone_with_theta(theta)</span></code>, which returns a cloned version of the kernel
but with the hyperparameters set to <code class="docutils literal notranslate"><span class="pre">theta</span></code>. An illustrative example:</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="kn">from</span> <span class="nn">sklearn.gaussian_process.kernels</span> <span class="kn">import</span> <span class="n">ConstantKernel</span><span class="p">,</span> <span class="n">RBF</span>
<span class="gp">>>> </span><span class="n">kernel</span> <span class="o">=</span> <span class="n">ConstantKernel</span><span class="p">(</span><span class="n">constant_value</span><span class="o">=</span><span class="mf">1.0</span><span class="p">,</span> <span class="n">constant_value_bounds</span><span class="o">=</span><span class="p">(</span><span class="mf">0.0</span><span class="p">,</span> <span class="mf">10.0</span><span class="p">))</span> <span class="o">*</span> <span class="n">RBF</span><span class="p">(</span><span class="n">length_scale</span><span class="o">=</span><span class="mf">0.5</span><span class="p">,</span> <span class="n">length_scale_bounds</span><span class="o">=</span><span class="p">(</span><span class="mf">0.0</span><span class="p">,</span> <span class="mf">10.0</span><span class="p">))</span> <span class="o">+</span> <span class="n">RBF</span><span class="p">(</span><span class="n">length_scale</span><span class="o">=</span><span class="mf">2.0</span><span class="p">,</span> <span class="n">length_scale_bounds</span><span class="o">=</span><span class="p">(</span><span class="mf">0.0</span><span class="p">,</span> <span class="mf">10.0</span><span class="p">))</span>
<span class="gp">>>> </span><span class="k">for</span> <span class="n">hyperparameter</span> <span class="ow">in</span> <span class="n">kernel</span><span class="o">.</span><span class="n">hyperparameters</span><span class="p">:</span> <span class="nb">print</span><span class="p">(</span><span class="n">hyperparameter</span><span class="p">)</span>
<span class="go">Hyperparameter(name='k1__k1__constant_value', value_type='numeric', bounds=array([[ 0., 10.]]), n_elements=1, fixed=False)</span>
<span class="go">Hyperparameter(name='k1__k2__length_scale', value_type='numeric', bounds=array([[ 0., 10.]]), n_elements=1, fixed=False)</span>
<span class="go">Hyperparameter(name='k2__length_scale', value_type='numeric', bounds=array([[ 0., 10.]]), n_elements=1, fixed=False)</span>
<span class="gp">>>> </span><span class="n">params</span> <span class="o">=</span> <span class="n">kernel</span><span class="o">.</span><span class="n">get_params</span><span class="p">()</span>
<span class="gp">>>> </span><span class="k">for</span> <span class="n">key</span> <span class="ow">in</span> <span class="nb">sorted</span><span class="p">(</span><span class="n">params</span><span class="p">):</span> <span class="nb">print</span><span class="p">(</span><span class="s2">"</span><span class="si">%s</span><span class="s2"> : </span><span class="si">%s</span><span class="s2">"</span> <span class="o">%</span> <span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="n">params</span><span class="p">[</span><span class="n">key</span><span class="p">]))</span>
<span class="go">k1 : 1**2 * RBF(length_scale=0.5)</span>
<span class="go">k1__k1 : 1**2</span>
<span class="go">k1__k1__constant_value : 1.0</span>
<span class="go">k1__k1__constant_value_bounds : (0.0, 10.0)</span>
<span class="go">k1__k2 : RBF(length_scale=0.5)</span>
<span class="go">k1__k2__length_scale : 0.5</span>
<span class="go">k1__k2__length_scale_bounds : (0.0, 10.0)</span>
<span class="go">k2 : RBF(length_scale=2)</span>
<span class="go">k2__length_scale : 2.0</span>
<span class="go">k2__length_scale_bounds : (0.0, 10.0)</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="n">kernel</span><span class="o">.</span><span class="n">theta</span><span class="p">)</span> <span class="c1"># Note: log-transformed</span>
<span class="go">[ 0. -0.69314718 0.69314718]</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="n">kernel</span><span class="o">.</span><span class="n">bounds</span><span class="p">)</span> <span class="c1"># Note: log-transformed</span>
<span class="go">[[ -inf 2.30258509]</span>
<span class="go"> [ -inf 2.30258509]</span>
<span class="go"> [ -inf 2.30258509]]</span>
</pre></div>
</div>
<p>All Gaussian process kernels are interoperable with <a class="reference internal" href="classes.html#module-sklearn.metrics.pairwise" title="sklearn.metrics.pairwise"><code class="xref py py-mod docutils literal notranslate"><span class="pre">sklearn.metrics.pairwise</span></code></a>
and vice versa: instances of subclasses of <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.Kernel.html#sklearn.gaussian_process.kernels.Kernel" title="sklearn.gaussian_process.kernels.Kernel"><code class="xref py py-class docutils literal notranslate"><span class="pre">Kernel</span></code></a> can be passed as
<code class="docutils literal notranslate"><span class="pre">metric</span></code> to <code class="docutils literal notranslate"><span class="pre">pairwise_kernels</span></code> from <a class="reference internal" href="classes.html#module-sklearn.metrics.pairwise" title="sklearn.metrics.pairwise"><code class="xref py py-mod docutils literal notranslate"><span class="pre">sklearn.metrics.pairwise</span></code></a>. Moreover,
kernel functions from pairwise can be used as GP kernels by using the wrapper
class <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.PairwiseKernel.html#sklearn.gaussian_process.kernels.PairwiseKernel" title="sklearn.gaussian_process.kernels.PairwiseKernel"><code class="xref py py-class docutils literal notranslate"><span class="pre">PairwiseKernel</span></code></a>. The only caveat is that the gradient of
the hyperparameters is not analytic but numeric and all those kernels support
only isotropic distances. The parameter <code class="docutils literal notranslate"><span class="pre">gamma</span></code> is considered to be a
hyperparameter and may be optimized. The other kernel parameters are set
directly at initialization and are kept fixed.</p>
</section>
<section id="basic-kernels">
<h3><span class="section-number">1.7.5.2. </span>Basic kernels<a class="headerlink" href="#basic-kernels" title="Permalink to this heading">¶</a></h3>
<p>The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.ConstantKernel.html#sklearn.gaussian_process.kernels.ConstantKernel" title="sklearn.gaussian_process.kernels.ConstantKernel"><code class="xref py py-class docutils literal notranslate"><span class="pre">ConstantKernel</span></code></a> kernel can be used as part of a <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.Product.html#sklearn.gaussian_process.kernels.Product" title="sklearn.gaussian_process.kernels.Product"><code class="xref py py-class docutils literal notranslate"><span class="pre">Product</span></code></a>
kernel where it scales the magnitude of the other factor (kernel) or as part
of a <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.Sum.html#sklearn.gaussian_process.kernels.Sum" title="sklearn.gaussian_process.kernels.Sum"><code class="xref py py-class docutils literal notranslate"><span class="pre">Sum</span></code></a> kernel, where it modifies the mean of the Gaussian process.
It depends on a parameter <span class="math notranslate nohighlight">\(constant\_value\)</span>. It is defined as:</p>
<div class="math notranslate nohighlight">
\[k(x_i, x_j) = constant\_value \;\forall\; x_1, x_2\]</div>
<p>The main use-case of the <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.WhiteKernel.html#sklearn.gaussian_process.kernels.WhiteKernel" title="sklearn.gaussian_process.kernels.WhiteKernel"><code class="xref py py-class docutils literal notranslate"><span class="pre">WhiteKernel</span></code></a> kernel is as part of a
sum-kernel where it explains the noise-component of the signal. Tuning its
parameter <span class="math notranslate nohighlight">\(noise\_level\)</span> corresponds to estimating the noise-level.
It is defined as:</p>
<div class="math notranslate nohighlight">
\[k(x_i, x_j) = noise\_level \text{ if } x_i == x_j \text{ else } 0\]</div>
</section>
<section id="kernel-operators">
<h3><span class="section-number">1.7.5.3. </span>Kernel operators<a class="headerlink" href="#kernel-operators" title="Permalink to this heading">¶</a></h3>
<p>Kernel operators take one or two base kernels and combine them into a new
kernel. The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.Sum.html#sklearn.gaussian_process.kernels.Sum" title="sklearn.gaussian_process.kernels.Sum"><code class="xref py py-class docutils literal notranslate"><span class="pre">Sum</span></code></a> kernel takes two kernels <span class="math notranslate nohighlight">\(k_1\)</span> and <span class="math notranslate nohighlight">\(k_2\)</span>
and combines them via <span class="math notranslate nohighlight">\(k_{sum}(X, Y) = k_1(X, Y) + k_2(X, Y)\)</span>.
The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.Product.html#sklearn.gaussian_process.kernels.Product" title="sklearn.gaussian_process.kernels.Product"><code class="xref py py-class docutils literal notranslate"><span class="pre">Product</span></code></a> kernel takes two kernels <span class="math notranslate nohighlight">\(k_1\)</span> and <span class="math notranslate nohighlight">\(k_2\)</span>
and combines them via <span class="math notranslate nohighlight">\(k_{product}(X, Y) = k_1(X, Y) * k_2(X, Y)\)</span>.
The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.Exponentiation.html#sklearn.gaussian_process.kernels.Exponentiation" title="sklearn.gaussian_process.kernels.Exponentiation"><code class="xref py py-class docutils literal notranslate"><span class="pre">Exponentiation</span></code></a> kernel takes one base kernel and a scalar parameter
<span class="math notranslate nohighlight">\(p\)</span> and combines them via
<span class="math notranslate nohighlight">\(k_{exp}(X, Y) = k(X, Y)^p\)</span>.
Note that magic methods <code class="docutils literal notranslate"><span class="pre">__add__</span></code>, <code class="docutils literal notranslate"><span class="pre">__mul___</span></code> and <code class="docutils literal notranslate"><span class="pre">__pow__</span></code> are
overridden on the Kernel objects, so one can use e.g. <code class="docutils literal notranslate"><span class="pre">RBF()</span> <span class="pre">+</span> <span class="pre">RBF()</span></code> as
a shortcut for <code class="docutils literal notranslate"><span class="pre">Sum(RBF(),</span> <span class="pre">RBF())</span></code>.</p>
</section>
<section id="radial-basis-function-rbf-kernel">
<h3><span class="section-number">1.7.5.4. </span>Radial basis function (RBF) kernel<a class="headerlink" href="#radial-basis-function-rbf-kernel" title="Permalink to this heading">¶</a></h3>
<p>The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.RBF.html#sklearn.gaussian_process.kernels.RBF" title="sklearn.gaussian_process.kernels.RBF"><code class="xref py py-class docutils literal notranslate"><span class="pre">RBF</span></code></a> kernel is a stationary kernel. It is also known as the “squared
exponential” kernel. It is parameterized by a length-scale parameter <span class="math notranslate nohighlight">\(l>0\)</span>, which
can either be a scalar (isotropic variant of the kernel) or a vector with the same
number of dimensions as the inputs <span class="math notranslate nohighlight">\(x\)</span> (anisotropic variant of the kernel).
The kernel is given by:</p>
<div class="math notranslate nohighlight">
\[k(x_i, x_j) = \text{exp}\left(- \frac{d(x_i, x_j)^2}{2l^2} \right)\]</div>
<p>where <span class="math notranslate nohighlight">\(d(\cdot, \cdot)\)</span> is the Euclidean distance.
This kernel is infinitely differentiable, which implies that GPs with this
kernel as covariance function have mean square derivatives of all orders, and are thus
very smooth. The prior and posterior of a GP resulting from an RBF kernel are shown in
the following figure:</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpr_prior_posterior.html"><img alt="../_images/sphx_glr_plot_gpr_prior_posterior_001.png" src="../_images/sphx_glr_plot_gpr_prior_posterior_001.png" /></a>
</figure>
</section>
<section id="matern-kernel">
<h3><span class="section-number">1.7.5.5. </span>Matérn kernel<a class="headerlink" href="#matern-kernel" title="Permalink to this heading">¶</a></h3>
<p>The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.Matern.html#sklearn.gaussian_process.kernels.Matern" title="sklearn.gaussian_process.kernels.Matern"><code class="xref py py-class docutils literal notranslate"><span class="pre">Matern</span></code></a> kernel is a stationary kernel and a generalization of the
<a class="reference internal" href="generated/sklearn.gaussian_process.kernels.RBF.html#sklearn.gaussian_process.kernels.RBF" title="sklearn.gaussian_process.kernels.RBF"><code class="xref py py-class docutils literal notranslate"><span class="pre">RBF</span></code></a> kernel. It has an additional parameter <span class="math notranslate nohighlight">\(\nu\)</span> which controls
the smoothness of the resulting function. It is parameterized by a length-scale parameter <span class="math notranslate nohighlight">\(l>0\)</span>, which can either be a scalar (isotropic variant of the kernel) or a vector with the same number of dimensions as the inputs <span class="math notranslate nohighlight">\(x\)</span> (anisotropic variant of the kernel). The kernel is given by:</p>
<div class="math notranslate nohighlight">
\[k(x_i, x_j) = \frac{1}{\Gamma(\nu)2^{\nu-1}}\Bigg(\frac{\sqrt{2\nu}}{l} d(x_i , x_j )\Bigg)^\nu K_\nu\Bigg(\frac{\sqrt{2\nu}}{l} d(x_i , x_j )\Bigg),\]</div>
<p>where <span class="math notranslate nohighlight">\(d(\cdot,\cdot)\)</span> is the Euclidean distance, <span class="math notranslate nohighlight">\(K_\nu(\cdot)\)</span> is a modified Bessel function and <span class="math notranslate nohighlight">\(\Gamma(\cdot)\)</span> is the gamma function.
As <span class="math notranslate nohighlight">\(\nu\rightarrow\infty\)</span>, the Matérn kernel converges to the RBF kernel.
When <span class="math notranslate nohighlight">\(\nu = 1/2\)</span>, the Matérn kernel becomes identical to the absolute
exponential kernel, i.e.,</p>
<div class="math notranslate nohighlight">
\[k(x_i, x_j) = \exp \Bigg(- \frac{1}{l} d(x_i , x_j ) \Bigg) \quad \quad \nu= \tfrac{1}{2}\]</div>
<p>In particular, <span class="math notranslate nohighlight">\(\nu = 3/2\)</span>:</p>
<div class="math notranslate nohighlight">
\[k(x_i, x_j) = \Bigg(1 + \frac{\sqrt{3}}{l} d(x_i , x_j )\Bigg) \exp \Bigg(-\frac{\sqrt{3}}{l} d(x_i , x_j ) \Bigg) \quad \quad \nu= \tfrac{3}{2}\]</div>
<p>and <span class="math notranslate nohighlight">\(\nu = 5/2\)</span>:</p>
<div class="math notranslate nohighlight">
\[k(x_i, x_j) = \Bigg(1 + \frac{\sqrt{5}}{l} d(x_i , x_j ) +\frac{5}{3l} d(x_i , x_j )^2 \Bigg) \exp \Bigg(-\frac{\sqrt{5}}{l} d(x_i , x_j ) \Bigg) \quad \quad \nu= \tfrac{5}{2}\]</div>
<p>are popular choices for learning functions that are not infinitely
differentiable (as assumed by the RBF kernel) but at least once (<span class="math notranslate nohighlight">\(\nu =
3/2\)</span>) or twice differentiable (<span class="math notranslate nohighlight">\(\nu = 5/2\)</span>).</p>
<p>The flexibility of controlling the smoothness of the learned function via <span class="math notranslate nohighlight">\(\nu\)</span>
allows adapting to the properties of the true underlying functional relation.
The prior and posterior of a GP resulting from a Matérn kernel are shown in
the following figure:</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpr_prior_posterior.html"><img alt="../_images/sphx_glr_plot_gpr_prior_posterior_005.png" src="../_images/sphx_glr_plot_gpr_prior_posterior_005.png" /></a>
</figure>
<p>See <a class="reference internal" href="#rw2006" id="id7"><span>[RW2006]</span></a>, pp84 for further details regarding the
different variants of the Matérn kernel.</p>
</section>
<section id="rational-quadratic-kernel">
<h3><span class="section-number">1.7.5.6. </span>Rational quadratic kernel<a class="headerlink" href="#rational-quadratic-kernel" title="Permalink to this heading">¶</a></h3>
<p>The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.RationalQuadratic.html#sklearn.gaussian_process.kernels.RationalQuadratic" title="sklearn.gaussian_process.kernels.RationalQuadratic"><code class="xref py py-class docutils literal notranslate"><span class="pre">RationalQuadratic</span></code></a> kernel can be seen as a scale mixture (an infinite sum)
of <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.RBF.html#sklearn.gaussian_process.kernels.RBF" title="sklearn.gaussian_process.kernels.RBF"><code class="xref py py-class docutils literal notranslate"><span class="pre">RBF</span></code></a> kernels with different characteristic length-scales. It is parameterized
by a length-scale parameter <span class="math notranslate nohighlight">\(l>0\)</span> and a scale mixture parameter <span class="math notranslate nohighlight">\(\alpha>0\)</span>
Only the isotropic variant where <span class="math notranslate nohighlight">\(l\)</span> is a scalar is supported at the moment.
The kernel is given by:</p>
<div class="math notranslate nohighlight">
\[k(x_i, x_j) = \left(1 + \frac{d(x_i, x_j)^2}{2\alpha l^2}\right)^{-\alpha}\]</div>
<p>The prior and posterior of a GP resulting from a <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.RationalQuadratic.html#sklearn.gaussian_process.kernels.RationalQuadratic" title="sklearn.gaussian_process.kernels.RationalQuadratic"><code class="xref py py-class docutils literal notranslate"><span class="pre">RationalQuadratic</span></code></a> kernel are shown in
the following figure:</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpr_prior_posterior.html"><img alt="../_images/sphx_glr_plot_gpr_prior_posterior_002.png" src="../_images/sphx_glr_plot_gpr_prior_posterior_002.png" /></a>
</figure>
</section>
<section id="exp-sine-squared-kernel">
<h3><span class="section-number">1.7.5.7. </span>Exp-Sine-Squared kernel<a class="headerlink" href="#exp-sine-squared-kernel" title="Permalink to this heading">¶</a></h3>
<p>The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.ExpSineSquared.html#sklearn.gaussian_process.kernels.ExpSineSquared" title="sklearn.gaussian_process.kernels.ExpSineSquared"><code class="xref py py-class docutils literal notranslate"><span class="pre">ExpSineSquared</span></code></a> kernel allows modeling periodic functions.
It is parameterized by a length-scale parameter <span class="math notranslate nohighlight">\(l>0\)</span> and a periodicity parameter
<span class="math notranslate nohighlight">\(p>0\)</span>. Only the isotropic variant where <span class="math notranslate nohighlight">\(l\)</span> is a scalar is supported at the moment.
The kernel is given by:</p>
<div class="math notranslate nohighlight">
\[k(x_i, x_j) = \text{exp}\left(- \frac{ 2\sin^2(\pi d(x_i, x_j) / p) }{ l^ 2} \right)\]</div>
<p>The prior and posterior of a GP resulting from an ExpSineSquared kernel are shown in
the following figure:</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpr_prior_posterior.html"><img alt="../_images/sphx_glr_plot_gpr_prior_posterior_003.png" src="../_images/sphx_glr_plot_gpr_prior_posterior_003.png" /></a>
</figure>
</section>
<section id="dot-product-kernel">
<h3><span class="section-number">1.7.5.8. </span>Dot-Product kernel<a class="headerlink" href="#dot-product-kernel" title="Permalink to this heading">¶</a></h3>
<p>The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.DotProduct.html#sklearn.gaussian_process.kernels.DotProduct" title="sklearn.gaussian_process.kernels.DotProduct"><code class="xref py py-class docutils literal notranslate"><span class="pre">DotProduct</span></code></a> kernel is non-stationary and can be obtained from linear regression
by putting <span class="math notranslate nohighlight">\(N(0, 1)\)</span> priors on the coefficients of <span class="math notranslate nohighlight">\(x_d (d = 1, . . . , D)\)</span> and
a prior of <span class="math notranslate nohighlight">\(N(0, \sigma_0^2)\)</span> on the bias. The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.DotProduct.html#sklearn.gaussian_process.kernels.DotProduct" title="sklearn.gaussian_process.kernels.DotProduct"><code class="xref py py-class docutils literal notranslate"><span class="pre">DotProduct</span></code></a> kernel is invariant to a rotation
of the coordinates about the origin, but not translations.
It is parameterized by a parameter <span class="math notranslate nohighlight">\(\sigma_0^2\)</span>. For <span class="math notranslate nohighlight">\(\sigma_0^2 = 0\)</span>, the kernel
is called the homogeneous linear kernel, otherwise it is inhomogeneous. The kernel is given by</p>
<div class="math notranslate nohighlight">
\[k(x_i, x_j) = \sigma_0 ^ 2 + x_i \cdot x_j\]</div>
<p>The <a class="reference internal" href="generated/sklearn.gaussian_process.kernels.DotProduct.html#sklearn.gaussian_process.kernels.DotProduct" title="sklearn.gaussian_process.kernels.DotProduct"><code class="xref py py-class docutils literal notranslate"><span class="pre">DotProduct</span></code></a> kernel is commonly combined with exponentiation. An example with exponent 2 is
shown in the following figure:</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/gaussian_process/plot_gpr_prior_posterior.html"><img alt="../_images/sphx_glr_plot_gpr_prior_posterior_004.png" src="../_images/sphx_glr_plot_gpr_prior_posterior_004.png" /></a>
</figure>
</section>
<section id="references">
<h3><span class="section-number">1.7.5.9. </span>References<a class="headerlink" href="#references" title="Permalink to this heading">¶</a></h3>
<div role="list" class="citation-list">
<div class="citation" id="rw2006" role="doc-biblioentry">
<span class="label"><span class="fn-bracket">[</span>RW2006<span class="fn-bracket">]</span></span>
<span class="backrefs">(<a role="doc-backlink" href="#id1">1</a>,<a role="doc-backlink" href="#id2">2</a>,<a role="doc-backlink" href="#id3">3</a>,<a role="doc-backlink" href="#id4">4</a>,<a role="doc-backlink" href="#id5">5</a>,<a role="doc-backlink" href="#id7">6</a>)</span>
<p><a class="reference external" href="https://www.gaussianprocess.org/gpml/chapters/RW.pdf">Carl E. Rasmussen and Christopher K.I. Williams,
“Gaussian Processes for Machine Learning”,
MIT Press 2006</a></p>
</div>
<div class="citation" id="duv2014" role="doc-biblioentry">
<span class="label"><span class="fn-bracket">[</span><a role="doc-backlink" href="#id6">Duv2014</a><span class="fn-bracket">]</span></span>
<p><a class="reference external" href="https://www.cs.toronto.edu/~duvenaud/cookbook/">David Duvenaud, “The Kernel Cookbook: Advice on Covariance functions”, 2014</a></p>
</div>
</div>
</section>
</section>
</section>
</div>
<div class="container">
<footer class="sk-content-footer">
© 2007 - 2023, scikit-learn developers (BSD License).
<a href="../_sources/modules/gaussian_process.rst.txt" rel="nofollow">Show this page source</a>
</footer>
</div>
</div>
</div>
<script src="../_static/js/vendor/bootstrap.min.js"></script>
<script>
window.ga=window.ga||function(){(ga.q=ga.q||[]).push(arguments)};ga.l=+new Date;
ga('create', 'UA-22606712-2', 'auto');
ga('set', 'anonymizeIp', true);
ga('send', 'pageview');
</script>
<script async src='https://www.google-analytics.com/analytics.js'></script>
<script defer data-domain="scikit-learn.org" src="https://views.scientific-python.org/js/script.js">
</script>
<script src="../_static/clipboard.min.js"></script>
<script src="../_static/copybutton.js"></script>
<script>
$(document).ready(function() {
/* Add a [>>>] button on the top-right corner of code samples to hide
* the >>> and ... prompts and the output and thus make the code
* copyable. */
var div = $('.highlight-python .highlight,' +
'.highlight-python3 .highlight,' +
'.highlight-pycon .highlight,' +
'.highlight-default .highlight')
var pre = div.find('pre');
// get the styles from the current theme
pre.parent().parent().css('position', 'relative');
// create and add the button to all the code blocks that contain >>>
div.each(function(index) {
var jthis = $(this);
// tracebacks (.gt) contain bare text elements that need to be
// wrapped in a span to work with .nextUntil() (see later)
jthis.find('pre:has(.gt)').contents().filter(function() {
return ((this.nodeType == 3) && (this.data.trim().length > 0));
}).wrap('<span>');
});
/*** Add permalink buttons next to glossary terms ***/
$('dl.glossary > dt[id]').append(function() {
return ('<a class="headerlink" href="#' +
this.getAttribute('id') +
'" title="Permalink to this term">¶</a>');
});
});
</script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-chtml.js"></script>
<script src="https://scikit-learn.org/versionwarning.js"></script>
</body>
</html>