Using Meta Learning to Initialize Bayesian Optimization · !Castuse experience intoanalgorithm. ......
Transcript of Using Meta Learning to Initialize Bayesian Optimization · !Castuse experience intoanalgorithm. ......
Using Meta Learning to InitializeBayesian Optimization
Albert-Ludwigs-Universität Freiburg
Matthias Feurer1 Jost Tobias Springenberg2 Frank Hutter11Research Group on Learning, Optimization, and Automated Algorithm Design2Machine Learning LabDepartment of Computer Science, University of Freiburg, GermanyECAI-2014 Workshop on Meta-learning & Algorithm Selection, 19 August 2014
Your task: Build an Iris classification system
Choose an algorithmbased on datasetcharacteristics, e.g. forthe Iris dataset this couldbe an SVMManual tuning-> fiddling withhyperparameters.Better: Use automatedmethods like PSO, GA orSMBOBest: AutoWeka
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 2 / 18
The iris pictures on this slide are from wikimedia commons and used under the following licenses: Top left: IrisVersicolor is public domain; Bottom left: Iris setosa is licensed by Radomil under CC BY-SA 3.0; Top right: IrisVirginica is licensed by C T Johansson under CC BY 3.0.
Your task: Build an Iris classification system
Choose an algorithmbased on datasetcharacteristics, e.g. forthe Iris dataset this couldbe an SVM
Manual tuning-> fiddling withhyperparameters.Better: Use automatedmethods like PSO, GA orSMBOBest: AutoWeka
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 2 / 18
The iris pictures on this slide are from wikimedia commons and used under the following licenses: Top left: IrisVersicolor is public domain; Bottom left: Iris setosa is licensed by Radomil under CC BY-SA 3.0; Top right: IrisVirginica is licensed by C T Johansson under CC BY 3.0.
Your task: Build an Iris classification system
Choose an algorithmbased on datasetcharacteristics, e.g. forthe Iris dataset this couldbe an SVMManual tuning-> fiddling withhyperparameters.
Better: Use automatedmethods like PSO, GA orSMBOBest: AutoWeka
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 2 / 18
The iris pictures on this slide are from wikimedia commons and used under the following licenses: Top left: IrisVersicolor is public domain; Bottom left: Iris setosa is licensed by Radomil under CC BY-SA 3.0; Top right: IrisVirginica is licensed by C T Johansson under CC BY 3.0.
Your task: Build an Iris classification system
Choose an algorithmbased on datasetcharacteristics, e.g. forthe Iris dataset this couldbe an SVMManual tuning-> fiddling withhyperparameters.Better: Use automatedmethods like PSO, GA orSMBO
Best: AutoWeka
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 2 / 18
The iris pictures on this slide are from wikimedia commons and used under the following licenses: Top left: IrisVersicolor is public domain; Bottom left: Iris setosa is licensed by Radomil under CC BY-SA 3.0; Top right: IrisVirginica is licensed by C T Johansson under CC BY 3.0.
Your task: Build an Iris classification system
Choose an algorithmbased on datasetcharacteristics, e.g. forthe Iris dataset this couldbe an SVMManual tuning-> fiddling withhyperparameters.Better: Use automatedmethods like PSO, GA orSMBOBest: AutoWeka
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 2 / 18
The iris pictures on this slide are from wikimedia commons and used under the following licenses: Top left: IrisVersicolor is public domain; Bottom left: Iris setosa is licensed by Radomil under CC BY-SA 3.0; Top right: IrisVirginica is licensed by C T Johansson under CC BY 3.0.
Adding the Iris Japonica to the dataset
Manual tuning:Use experience and startfrom the parametersfound on the Iris datasetAutomated methods-> start from scratch→ Cast use experienceinto an algorithm.
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 3 / 18
The iris pictures on this slide are from wikimedia commons and used under the following licenses: Top left: IrisVersicolor is public domain; Bottom left: Iris setosa is licensed by Radomil under CC BY-SA 3.0; Top right: IrisVirginica is licensed by C T Johansson under CC BY 3.0; Bottom right: Iris Japonica is licensed by KENPEI underCC BY-SA 3.0
Adding the Iris Japonica to the dataset
Manual tuning:Use experience and startfrom the parametersfound on the Iris dataset
Automated methods-> start from scratch→ Cast use experienceinto an algorithm.
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 3 / 18
The iris pictures on this slide are from wikimedia commons and used under the following licenses: Top left: IrisVersicolor is public domain; Bottom left: Iris setosa is licensed by Radomil under CC BY-SA 3.0; Top right: IrisVirginica is licensed by C T Johansson under CC BY 3.0; Bottom right: Iris Japonica is licensed by KENPEI underCC BY-SA 3.0
Adding the Iris Japonica to the dataset
Manual tuning:Use experience and startfrom the parametersfound on the Iris datasetAutomated methods-> start from scratch
→ Cast use experienceinto an algorithm.
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 3 / 18
The iris pictures on this slide are from wikimedia commons and used under the following licenses: Top left: IrisVersicolor is public domain; Bottom left: Iris setosa is licensed by Radomil under CC BY-SA 3.0; Top right: IrisVirginica is licensed by C T Johansson under CC BY 3.0; Bottom right: Iris Japonica is licensed by KENPEI underCC BY-SA 3.0
Adding the Iris Japonica to the dataset
Manual tuning:Use experience and startfrom the parametersfound on the Iris datasetAutomated methods-> start from scratch→ Cast use experienceinto an algorithm.
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 3 / 18
The iris pictures on this slide are from wikimedia commons and used under the following licenses: Top left: IrisVersicolor is public domain; Bottom left: Iris setosa is licensed by Radomil under CC BY-SA 3.0; Top right: IrisVirginica is licensed by C T Johansson under CC BY 3.0; Bottom right: Iris Japonica is licensed by KENPEI underCC BY-SA 3.0
Sequential Model-based Bayesian Optimization(SMBO)
ML Algorithm A
ConfigurationSpace Λ of A
Dataset D
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 4 / 18
Sequential Model-based Bayesian Optimization(SMBO)
Configuration Task
ML Algorithm A
ConfigurationSpace Λ of A
Dataset D
Configuration λ ∗
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 4 / 18
Sequential Model-based Bayesian Optimization(SMBO)
Configuration Task
ML Algorithm A
ConfigurationSpace Λ of A
Dataset D
Configuration λ ∗
Fit regressionmodel on
(λ ,Aλ (D)) pairs
Select promisingconfiguration
λ ∈ Λ
Evaluate Aλ (D)
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 4 / 18
Metalearning-Initialized SMBO (MI-SMBO)
Configuration Task
ML Algorithm A
ConfigurationSpace Λ of A
Dataset Dnew
Fit regressionmodel on pairs of
(λ ,Aλ (Dnew))
Select promisingconfiguration
λ ∈ Λ
EvaluateAλ (Dnew)
Configuration λ ∗
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 5 / 18
Metalearning-Initialized SMBO (MI-SMBO)
Configuration Task
ML Algorithm A
ConfigurationSpace Λ of A
Dataset Dnew
Fit regressionmodel on pairs of
(λ ,Aλ (Dnew))
Select promisingconfiguration
λ ∈ Λ
EvaluateAλ (Dnew)
Configuration λ ∗
Find Datasets Disimilar to Dnew
Initialize Searchwith λ ∗Di
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 5 / 18
Metafeatures
# training examples: 150# classes: 3# features: 4# numerical features: 4# categorical features: 0missing values? No
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 6 / 18
The iris pictures on this slide are from wikimedia commons and used under the following licenses: Top left: IrisVersicolor is public domain; Bottom left: Iris setosa is licensed by Radomil under CC BY-SA 3.0; Top right: IrisVirginica is licensed by C T Johansson under CC BY 3.0.
Metalearning-Initialized Bayesian Optimization
For a new dataset Dnew :Sort known datasets D1:N by distance to Dnew .For each of these datasets, extract the best knownhyperparameter configuration λ ∗Di
.Initialize SMBO with the first k hyperparameterconfigurations from the sorted list.
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 7 / 18
Similarity of Datasets
wavef
orm
-500
0iri
s
pend
igits
yeas
t
satim
age
balanc
e-sc
ale
ecoli
cmc
page
-block
sta
egl
ass
nurs
ery
spam
base
vehi
cle
segm
ent
car
optd
igits
prim
ary-
tum
or
mfe
at-fo
urier
mfe
at-k
arhu
nen
mfe
at-zer
nike
abalon
e
vowel
mfe
at-fa
ctor
s
derm
atolog
y
lette
r
mfe
at-m
orph
olog
ical
mfe
at-p
ixel
euca
lypt
us
arrh
ythm
ia
zoo
audi
olog
y
auto
s
soyb
ean
cylin
der-b
ands
vote
iono
sphe
re
tic-ta
c-to
e
braz
iltou
rism
hepa
titisla
bor
hear
t-sta
tlog
brea
st-w
mus
hroo
m
hear
t-c
hear
t-h
post
oper
ative-
patie
nt-d
ata
lym
ph
brea
st-c
ance
r
diab
etesliv
er-d
isord
ers
kr-v
s-kp
cred
it-a
anne
al.O
RIG
cred
it-g
sona
r
habe
rman
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 8 / 18
Finding the nearest datasets (1)
waveform-5000
iris
pendigits
yeast
satimage
balance-scale
ecoli
cmc
page-blocks
tae
glass
nursery
spambase
vehicle
segment
car
optdigits
primary-tumor
mfeat-fourier
mfeat-karhunen
mfeat-zernike
abalone
vowel
mfeat-factors
dermatology
letter
mfeat-morphological
mfeat-pixel
eucalyptus
arrhythmia
zoo
audiology
autos
soybean
cylinder-bands
vote
ionosphere
tic-tac-toe
braziltourism
hepatitisla
bor
heart-statlog
breast-w
mushroom
heart-c
heart-h
postoperative-patient-data
lymph
breast-cancer
diabetesliv
er-disorders
kr-vs-kp
credit-a
anneal.ORIG
credit-g
sonar
haberman
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 8 / 18
Finding the nearest datasets (2)
waveform-5000
iris
pendigits
yeast
satimage
balance-scale
ecoli
cmc
page-blocks
tae
glass
nursery
spambase
vehicle
segment
car
optdigits
primary-tumor
mfeat-fourier
mfeat-karhunen
mfeat-zernike
abalone
vowel
mfeat-factors
dermatology
letter
mfeat-morphological
mfeat-pixel
eucalyptus
arrhythmia
zoo
audiology
autos
soybean
cylinder-bands
vote
ionosphere
tic-tac-toe
braziltourism
hepatitisla
bor
heart-statlog
breast-w
mushroom
heart-c
heart-h
postoperative-patient-data
lymph
breast-cancer
diabetesliv
er-disorders
kr-vs-kp
credit-a
anneal.ORIG
credit-g
sonar
haberman
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 8 / 18
Finding the nearest datasets (3)
waveform-5000
iris
pendigits
yeast
satimage
balance-scale
ecoli
cmc
page-blocks
tae
glass
nursery
spambase
vehicle
segment
car
optdigits
primary-tumor
mfeat-fourier
mfeat-karhunen
mfeat-zernike
abalone
vowel
mfeat-factors
dermatology
letter
mfeat-morphological
mfeat-pixel
eucalyptus
arrhythmia
zoo
audiology
autos
soybean
cylinder-bands
vote
ionosphere
tic-tac-toe
braziltourism
hepatitisla
bor
heart-statlog
breast-w
mushroom
heart-c
heart-h
postoperative-patient-data
lymph
breast-cancer
diabetesliv
er-disorders
kr-vs-kp
credit-a
anneal.ORIG
credit-g
sonar
haberman
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 8 / 18
Finding the nearest datasets (3)
waveform-5000
iris
pendigits
yeast
satimage
balance-scale
ecoli
cmc
page-blocks
tae
glass
nursery
spambase
vehicle
segment
car
optdigits
primary-tumor
mfeat-fourier
mfeat-karhunen
mfeat-zernike
abalone
vowel
mfeat-factors
dermatology
letter
mfeat-morphological
mfeat-pixel
eucalyptus
arrhythmia
zoo
audiology
autos
soybean
cylinder-bands
vote
ionosphere
tic-tac-toe
braziltourism
hepatitisla
bor
heart-statlog
breast-w
mushroom
heart-c
heart-h
postoperative-patient-data
lymph
breast-cancer
diabetesliv
er-disorders
kr-vs-kp
credit-a
anneal.ORIG
credit-g
sonar
haberman
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 8 / 18
Finding the nearest datasets (4)
waveform-5000
iris
pendigits
yeast
satimage
balance-scale
ecoli
cmc
page-blocks
tae
glass
nursery
spambase
vehicle
segment
car
optdigits
primary-tumor
mfeat-fourier
mfeat-karhunen
mfeat-zernike
abalone
vowel
mfeat-factors
dermatology
letter
mfeat-morphological
mfeat-pixel
eucalyptus
arrhythmia
zoo
audiology
autos
soybean
cylinder-bands
vote
ionosphere
tic-tac-toe
braziltourism
hepatitisla
bor
heart-statlog
breast-w
mushroom
heart-c
heart-h
postoperative-patient-data
lymph
breast-cancer
diabetesliv
er-disorders
kr-vs-kp
credit-a
anneal.ORIG
credit-g
sonar
haberman
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 8 / 18
Distance metric (1)
Commonly used in literature, the L1 norm:
d(Dnew,Dj) = ∑i|mnew
i −mji | (1)
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 9 / 18
Experimental Setup
57 datasets from the OpenML repository
46 metafeatures from the literature:Split into five different subsets, including landmarking[Pfahringer et al. 2000]
Two case studiesSupport Vector Machine with MI-Spearmint [Snoek et al. 2012]AutoSklearn with MI-SMAC [Hutter et al. 2011]
Tried 5, 10, 20 and 25 initial configurationsran each instantiation 10 times on each dataset→ 26220 optimization runstherefore, precomputed a dense grid for every dataset
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 10 / 18
Experimental Setup
57 datasets from the OpenML repository46 metafeatures from the literature:
Split into five different subsets, including landmarking[Pfahringer et al. 2000]
Two case studiesSupport Vector Machine with MI-Spearmint [Snoek et al. 2012]AutoSklearn with MI-SMAC [Hutter et al. 2011]
Tried 5, 10, 20 and 25 initial configurationsran each instantiation 10 times on each dataset→ 26220 optimization runstherefore, precomputed a dense grid for every dataset
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 10 / 18
Experimental Setup
57 datasets from the OpenML repository46 metafeatures from the literature:
Split into five different subsets, including landmarking[Pfahringer et al. 2000]
Two case studiesSupport Vector Machine with MI-Spearmint [Snoek et al. 2012]AutoSklearn with MI-SMAC [Hutter et al. 2011]
Tried 5, 10, 20 and 25 initial configurationsran each instantiation 10 times on each dataset→ 26220 optimization runstherefore, precomputed a dense grid for every dataset
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 10 / 18
Experimental Setup
57 datasets from the OpenML repository46 metafeatures from the literature:
Split into five different subsets, including landmarking[Pfahringer et al. 2000]
Two case studiesSupport Vector Machine with MI-Spearmint [Snoek et al. 2012]AutoSklearn with MI-SMAC [Hutter et al. 2011]
Tried 5, 10, 20 and 25 initial configurations
ran each instantiation 10 times on each dataset→ 26220 optimization runstherefore, precomputed a dense grid for every dataset
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 10 / 18
Experimental Setup
57 datasets from the OpenML repository46 metafeatures from the literature:
Split into five different subsets, including landmarking[Pfahringer et al. 2000]
Two case studiesSupport Vector Machine with MI-Spearmint [Snoek et al. 2012]AutoSklearn with MI-SMAC [Hutter et al. 2011]
Tried 5, 10, 20 and 25 initial configurationsran each instantiation 10 times on each dataset→ 26220 optimization runs
therefore, precomputed a dense grid for every dataset
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 10 / 18
Experimental Setup
57 datasets from the OpenML repository46 metafeatures from the literature:
Split into five different subsets, including landmarking[Pfahringer et al. 2000]
Two case studiesSupport Vector Machine with MI-Spearmint [Snoek et al. 2012]AutoSklearn with MI-SMAC [Hutter et al. 2011]
Tried 5, 10, 20 and 25 initial configurationsran each instantiation 10 times on each dataset→ 26220 optimization runstherefore, precomputed a dense grid for every dataset
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 10 / 18
Combined Algorithm Selection andHyperparameter Optimization problem (CASH)
Max Features
SVM
gamma C(SVM)
Classifier
Random Forest LinearSVM
Criterion Min Samples Split loss C(LinearSVM)
[Auto-WEKA, Thornton et al. 2013]
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 11 / 18
Combined Algorithm Selection andHyperparameter Optimization problem (CASH)
Max Features
SVM
gamma C(SVM)
Classifier
Random Forest LinearSVM
Criterion Min Samples Split loss C(LinearSVM)
[Auto-WEKA, Thornton et al. 2013]
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 11 / 18
AutoSklearn: Hyperparameters
Component Hyperparameter # Values
Main λclassifier 3Main preprocessing 2SVM log2(C) 21SVM log2(γ) 19LinearSVM log2(C) 21LinearSVM penalty 2RF min splits 5RF max features 10RF criterion 2PCA variance to keep 2
1623 hyperparameter configurations
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 12 / 18
AutoSklearn: Hyperparameters
Component Hyperparameter # Values
Main λclassifier 3Main preprocessing 2SVM log2(C) 21SVM log2(γ) 19LinearSVM log2(C) 21LinearSVM penalty 2RF min splits 5RF max features 10RF criterion 2PCA variance to keep 2
1623 hyperparameter configurations
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 12 / 18
AutoSklearn: Results (1)
0 10 20 30 40 50#Function evaluations
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.07
Diffe
rence
to m
in f
unct
ion v
alu
e
SMAC
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 13 / 18
AutoSklearn: Results (1)
0 10 20 30 40 50#Function evaluations
0.00
0.02
0.04
0.06
0.08
0.10
Diffe
rence
to m
in f
unct
ion v
alu
e
SMACrandomTPEMI-SMAC(10,L1 ,landmarking)
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 14 / 18
AutoSklearn: Results (2)
0 10 20 30 40 501.8
2.0
2.2
2.4
2.6
2.8
3.0
SMACrandom
TPE MI-SMAC(10,L1 ,landmarking)
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 15 / 18
AutoSklearn: Results (3)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 10 20 30 40 50#Function evaluations
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
MI-SMAC(10,L1 ,landmarking) vs SMAC
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 15 / 18
AutoSklearn: Results (3)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 10 20 30 40 50#Function evaluations
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
MI-SMAC(10,L1 ,landmarking) vs SMAC
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 15 / 18
AutoSklearn: Results (3)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 10 20 30 40 50#Function evaluations
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
MI-SMAC(10,L1 ,landmarking) vs SMAC MI-SMAC(10,L1 ,landmarking) vs TPE
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 15 / 18
AutoSklearn: Results (3)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 10 20 30 40 50#Function evaluations
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
MI-SMAC(10,L1 ,landmarking) vs SMAC
MI-SMAC(10,L1 ,landmarking) vs TPE
MI-SMAC(10,L1 ,landmarking) vs random
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 15 / 18
Open questions
Does MI-SMBO scale to larger configuration spaces?What if gridsearch is too expensive?Can the metalearning component be added directly intothe SMBO procedure?
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 16 / 18
Take home messages
SMBO can be substantially improved by providing goodinitial configurations.Metalearning provides a sound framework to find theseconfigurations.MI-SMAC improves on state-of-the-art methods on a largeconfiguration space, namely AutoSklearn.
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 17 / 18
The end
Thank you for your attention.
Further questions: [email protected]
This presentation was partially supported by an ECCAI TravelAward and the ECCAI sponsors.
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 18 / 18
AutoSklearn: Results (5)
0 10 20 30 40 50#Function evaluations
0.00
0.02
0.04
0.06
0.08
0.10
Min
funct
ion v
alu
e
SMACrandomTPEMI-SMAC(10,L1 ,landmarking)
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 18 / 18
AutoSklearn: Results (7)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0 10 20 30 40 50
Function evaluations
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0.0
MI-SMAC(10,L1,all) vs MI-SMAC(10,L1,landmarking)MI-SMAC(10,L1,all) vs SMAC
MI-SMAC(10,L1,all) vs TPEMI-SMAC(10,L1,all) vs random
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 18 / 18
AutoSklearn: Results (8)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0 10 20 30 40 50#Function evaluations
0.6
0.5
0.4
0.3
0.2
0.1
0.0
SMAC vs MI-SMAC(10,L1 ,landmarking)
SMAC vs TPE
SMAC vs random
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 18 / 18
AutoSklearn: Results (9)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 10 20 30 40 50
Function evaluations
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0.0
MI-SMAC(25,L1,landmarking) vs MI-SMAC(10,L1,landmarking)MI-SMAC(25,L1,landmarking) vs MI-SMAC(20,L1,landmarking)MI-SMAC(25,L1,landmarking) vs MI-SMAC(5,L1,landmarking)
MI-SMAC(25,L1,landmarking) vs SMACMI-SMAC(25,L1,landmarking) vs TPEMI-SMAC(25,L1,landmarking) vs random
MetaSel ’14 Feurer, Springenberg and Hutter – MI-SMBO 18 / 18