BI-population CMA-ES Algorithms with Surrogate Models and Line Searches

State-of-the-art
Contribution
BI-population CMA-ES Algorithms with
Surrogate Models and Line Searches
Ilya Loshchilov1
, Marc Schoenauer2
and Michèle Sebag 2
1
LIS, École Polytechnique Fédérale de Lausanne
2
TAO, INRIA − CNRS − Université Paris-Sud
July 6th, 2013
Ilya Loshchilov, Marc Schoenauer and Michèle Sebag BI-population CMA-ES with Surrogate Models and Line Searches 1/ 26

State-of-the-art
Contribution
Historical overview: BBOB’2012
Expensive Optimization
Self-adaptive surrogate-assisted CMA-ES (IPOP-saACM-ES and
BIPOP-saACM-ES) on noiseless1
and noisy testbeds2
.
BIPOP-saACM-ES demonstrates one of the best performance
among the algorithms of the BBOB-2009, 2010 and 2012.
Multimodal Optimization
Alternative restart strategies (NBIPOP-aCMA-ES and
NIPOP-aCMA-ES) on noiseless testbed3
.
NBIPOP-aCMA-ES is TOP-3 algorithm of the CEC’2013
(preliminary results).
1[Loshchilov, Schoenauer and Sebag; GECCO-BBOB 2012] "Black-box optimization
benchmarking of IPOP-saACM-ES and BIPOP-saACM-ES on the BBOB-2012 noiseless testbed"
2[Loshchilov, Schoenauer and Sebag; GECCO-BBOB 2012] "Black-box optimization
benchmarking of IPOP-saACM-ES on the BBOB-2012 noisy testbed"
3[Loshchilov, Schoenauer and Sebag; GECCO-BBOB 2012] "Black-box Optimization
Benchmarking of NIPOP-aCMA-ES and NBIPOP-aCMA-ES on the BBOB-2012 Noiseless Testbed"

State-of-the-art
Contribution
This talk: BBOB’2013
Expensive Optimization
saACM with intensive surrogate model exploitation
(BIPOP-saACM-ES-k) on noiseless testbed4
.
BIPOP-saACM-ES-k further improves BIPOP-saACM-ES.
Optimization of separable and non-separable functions
BIPOP-aCMA-STEP: a hybrid of BIPOP-aCMA and STEP
algorithm.
BIPOP-aCMA-STEP demonstrates a cheap way to identify and
exploit the separability.
Efﬁcient Optimization
HCMA: a hybrid of BIPOP-saACM-ES-k, STEP and NEWUOA
algorithms.
4[Loshchilov, Schoenauer and Sebag; GECCO 2013] "Intensive Surrogate Model Exploitation in
Self-adaptive Surrogate-assisted CMA-ES (saACM-ES)"

State-of-the-art
Contribution
Content
1 State-of-the-art
Covariance Matrix Adaptation Evolution Strategy (CMA-ES)
s∗
ACM-ES: Self-Adaptive Surrogate-Assisted CMA-ES
2 Contribution
Intensive surrogate model exploitation

State-of-the-art
Contribution
s∗ACM-ES: Self-Adaptive Surrogate-Assisted CMA-ES
(µ, λ)-Covariance Matrix Adaptation Evolution Strategy
Rank-µ Update 5 6
yi ∼ N (0, C) , C = I
xi = m + σ yi, σ = 1
sampling of λ
solutions
Cµ = 1
µ yi:λyT
i:λ
C ← (1 − 1) × C + 1 × Cµ
calculating C from
best µ out of λ
mnew ← m + 1
µ yi:λ
new distribution
The adaptation increases the probability of successful steps to appear again.
Other components of CMA-ES: step-size adaptation, evolution path.
5[Hansen et al., ECJ 2003] "Reducing the time complexity of the derandomized evolution strategy with covariance matrix adaptation
(CMA-ES)"
6 the slide adopted by courtesy of Nikolaus Hansen

State-of-the-art
Contribution
Invariance: Guarantee for Generalization
Invariance properties of CMA-ES
Invariance to order-preserving
transformations in function space
true for all comparison-based algorithms
Translation and rotation invariance
thanks to C
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
CMA-ES is almost parameterless (as a consequence of invariances)
Tuning on a small set of functions Hansen & Ostermeier 2001
Default values generalize to whole classes
Exception: population size for multi-modal functions a b
a[Auger & Hansen, CEC 2005] "A restart CMA evolution strategy with increasing population size"
b[Loshchilov et al., PPSN 2012] "Alternative Restart Strategies for CMA-ES"

State-of-the-art
Contribution
BIPOP-CMA-ES
BIPOP-CMA-ES: 7
(BIPOP-aCMA-ES 8
)
Regime-1 (large populations, IPOP part):
Each restart: λlarge = 2 ∗ λlarge , σ0
large = σ0
default
Regime-2 (small populations):
Each restart:
λsmall = λdefault
1
2
λlarge
λdefault
U[0,1]2
, σ0
small = σ0
default × 10−2U[0,1]
where U[0, 1] stands for the uniform distribution in [0, 1].
BIPOP-CMA-ES launches the ﬁrst run with default population size
and initial step-size. In each restart, it selects the restart regime
with less function evaluations used so far.
7Hansen (GECCO BBOB 2009). "Benchmarking a BI-population CMA-ES on the BBOB-2009
function testbed"
8Loshchilov, Schoenauer and Sebag (GECCO BBOB 2012). "Black-box Optimization
Benchmarking of NIPOP-aCMA-ES and NBIPOP-aCMA-ES on the BBOB-2012 Noiseless Testbed"

State-of-the-art
Contribution
Contents
1 State-of-the-art
s∗
2 Contribution

State-of-the-art
Contribution
s∗
Using Ranking SVM as the surrogate model
Build a global model using Ranking SVM 9
xi ≻ xj iff ˆF(xi) < ˆF(xj)
Comparison-based surrogate models → invariance to
rank-preserving transformations of F(x)
How to choose an appropriate Kernel?
Use covariance matrix C adapted by CMA-ES in Gaussian
kernel10
K(xi, xj) = e−
(xi−xj )T (xi−xj)
2σ2
; KC(xi, xj) = e−
(xi−xj )T C−1(xi−xj )
2σ2
Invariance to rotation of the search space thanks to C
9[Runarsson et al., PPSN 2006] "Ordinal Regression in Evolutionary Computation"
10[Loshchilov et al., PPSN 2010] "Comparison-based optimizers need comparison-based
surrogates"

State-of-the-art
Contribution
s∗
Surrogate-assisted
CMA-ES with online
adaptation of model
hyper-parameters a
a[Loshchilov et al., GECCO 2012]
"Self-Adaptive Surrogate-Assisted Covariance
Matrix Adaptation Evolution Strategy"

State-of-the-art
Contribution
Results on Black-Box Optimization Competition
BIPOP-s∗aACM and IPOP-s∗aACM (with restarts) on 24 noiseless 20 dimensional functions
0 1 2 3 4 5 6 7 8 9
log10 of (ERT / dimension)
0.0
0.5
1.0
Proportionoffunctions
RANDOMSEARCH
SPSA
BAYEDA
DIRECT
DE-PSO
GA
LSfminbnd
LSstep
RCGA
Rosenbrock
MCS
ABC
PSO
POEMS
EDA-PSO
NELDERDOERR
NELDER
oPOEMS
FULLNEWUOA
ALPS
GLOBAL
PSO_Bounds
BFGS
ONEFIFTH
Cauchy-EDA
NBC-CMA
CMA-ESPLUSSEL
NEWUOA
AVGNEWUOA
G3PCX
1komma4mirser
1plus1
CMAEGS
DEuniform
DE-F-AUC
MA-LS-CHAIN
VNS
iAMALGAM
IPOP-CMA-ES
AMALGAM
IPOP-ACTCMA-ES
IPOP-saACM-ES
MOS
BIPOP-CMA-ES
BIPOP-saACM-ES
best 2009
f1-24

State-of-the-art
Contribution
Contents
1 State-of-the-art
s∗
2 Contribution

State-of-the-art
Contribution
The only difference between BIPOP-saACM-k and BIPOP-saACM:
Intensive exploitation: when optimizing ˆF, λ = kλλdef , µ = µdef .
kλ = 1 for D<10 and kλ = 10, 100, 1000 for 10, 20, 40-D.
Divergence Prevention: kλ > 1 is used only of ˆn ≥ ˆnkλ
, ˆnkλ
= 4.

State-of-the-art
Contribution
* smaller budget for surrogate-assisted search: 104
D for
BIPOP-saACM-k versus 106
D for BIPOP-saACM.

State-of-the-art
Contribution
Contents
1 State-of-the-art
s∗
2 Contribution

State-of-the-art
Contribution
Select the easiest point (STEP) 11 12
Simple line search method based on iterative interval division.
Great optimizer of one-dimensional multimodal functions.
An extension to multi-dimensional (sequential) search
+ simple idea: sequentially optimize one dimension after another.
- some stopping criteria should be set a priori, e.g., number of
evaluations or target precision.
- no hint whether the problem is separable or not is available.
11[Swarzberg et al., CEC 1994] "The easiest way to optimize a function"
12[Posík et al., ECJ 2012] "Restarted local search algorithms for continuous black box
optimization"

State-of-the-art
Contribution
Parallel multi-dimensional STEP
1. Check one new STEP point per each dimension.
2. Current estimate of the optimum x∗
= a solution composed of
best x∗
i -values from all variables.
3. If the current estimate is worse than the previous one, then the
problem is not separable.

State-of-the-art
Contribution
BIPOP-aCMA-STEP
1. BIPOP-aCMA-STEP and STEP are running in parallel, a fraction
ρST EP = 0.5 of function evaluations is allocated to STEP.
2. At each iteration after nMinIterST EP = 10 iterations the STEP
can be stopped if its best solution is worse than the one of
BIPOP-aCMA-ES.

State-of-the-art
Contribution
0 1 2 3 4 5 6 7 8
log10 of (# f-evals / dimension)
0.0
0.2
0.4
0.6
0.8
1.0
Proportionoffunction+targetpairs
BIPOP-aCMA
best 2009
BIPOP-aCMA-STEPf1-5,20-D

State-of-the-art
Contribution
0 1 2 3 4 5 6 7 8
0.0
0.2
0.4
0.6
0.8
1.0
BIPOP-aCMA
BIPOP-aCMA-STEP
best 2009f1-24,20-D

State-of-the-art
Contribution
HCMA = BIPOP-saACM-ES-k + STEP + NEWUOA13
1. NEWUOA with m = 2n + 1 for 10n functions evaluations.
2. BIPOP-saACM-ES-k and STEP with nMinIterST EP = 10 (e.g.,
10n evaluations).
13[Powell, 2006] "The NEWUOA software for unconstrained optimization without derivatives"

State-of-the-art
Contribution

State-of-the-art
Contribution
0 1 2 3 4 5 6 7 8
0.0
0.2
0.4
0.6
0.8
1.0
fmincon
lmm-CMA-ES
IPOP-texp
MOS
BIPOP-saACM-k
HCMA
best 2009f1-24,20-D

State-of-the-art
Contribution
Conclusion
Intensive surrogate model exploitatiom improves the
performance on unimodal functions.
STEP algorithm is a cheap tool to deal with separable problems.
HCMA demonstrates the best overall performance.
Perspective
Implement NEWUOA-like search within saACM-ES.
Use alternative restart strategies (NBIPOP and NIPOP) in
HCMA.

State-of-the-art
Contribution
Thank you for your attention!
Questions?

BI-population CMA-ES Algorithms with Surrogate Models and Line Searches

More Related Content

Viewers also liked (13)

Similar to BI-population CMA-ES Algorithms with Surrogate Models and Line Searches (20)

Recently uploaded (20)

BI-population CMA-ES Algorithms with Surrogate Models and Line Searches