Completely Derandomized Self-Adaptation in

Evolution Strategies

Nikolaus Hansen hansen@[Link]

Technische Universität Berlin, Fachgebiet für Bionik, Sekr. ACK 1, Ackerstr. 71–76,
13355 Berlin, Germany
Andreas Ostermeier ostermeier@[Link]
Technische Universität Berlin, Fachgebiet für Bionik, Sekr. ACK 1, Ackerstr. 71–76,
13355 Berlin, Germany

Abstract
This paper puts forward two useful methods for self-adaptation of the mutation dis-
tribution – the concepts of derandomization and cumulation. Principle shortcomings of
the concept of mutative strategy parameter control and two levels of derandomization
are reviewed. Basic demands on the self-adaptation of arbitrary (normal) mutation dis-
tributions are developed. Applying arbitrary, normal mutation distributions is equiv-
alent to applying a general, linear problem encoding.
The underlying objective of mutative strategy parameter control is roughly to favor
previously selected mutation steps in the future. If this objective is pursued rigor-
ously, a completely derandomized self-adaptation scheme results, which adapts arbi-
trary normal mutation distributions. This scheme, called covariance matrix adaptation
(CMA), meets the previously stated demands. It can still be considerably improved by
cumulation – utilizing an evolution path rather than single search steps.
Simulations on various test functions reveal local and global search properties of
the evolution strategy with and without covariance matrix adaptation. Their per-
formances are comparable only on perfectly scaled functions. On badly scaled, non-
separable functions usually a speed up factor of several orders of magnitude is ob-
served. On moderately mis-scaled functions a speed up factor of three to ten can be
expected.
Keywords
Evolution strategy, self-adaptation, strategy parameter control, step size control, de-
randomization, derandomized self-adaptation, covariance matrix adaptation, evolu-
tion path, cumulation, cumulative path length control.

1 Introduction
The evolution strategy (ES) is a stochastic search algorithm that addresses the following
search problem: Minimize a non-linear objective function that is a mapping from search
space S ⊆ Rn to R. Search steps are taken by stochastic variation, so-called mutation, of
(recombinations of) points found so far. The best out of a number of new search points
are selected to continue. The mutation is usually carried out by adding a realization
of a normally distributed random vector. It is easy to imagine that the parameters
of the normal distribution play an essential role for the performance1 of the search
1 Performance, as used in this paper, always refers to the (expected) number of required objective function
evaluations to reach a certain function value.

c
°2001 by the Massachusetts Institute of Technology Evolutionary Computation 9(2): 159-195
N. Hansen and A. Ostermeier

Figure 1: One-σ lines of equal probability density of two normal distributions respec-
tively. Left: one free parameter (circles). Middle: n free parameters (axis-parallel ellip-
soids). Right: (n2 + n)/2 free parameters (arbitrarily oriented ellipsoids).

algorithm. This paper is specifically concerned with the adjustment of the parameters
of the normal mutation distribution.
Among others, the parameters that parameterize the mutation distribution are
called strategy parameters, in contrast to object parameters that define points in search
space. Usually, no particularly detailed knowledge about the suitable choice of strat-
egy parameters is available. With respect to the mutation distribution, there is typically
only a small width of strategy parameter settings where substantial search progress
can be observed (Rechenberg, 1973). Good parameter settings differ remarkably from
problem to problem. Even worse, they usually change during the search process (pos-
sibly by several orders of magnitude). For this reason, self-adaptation of the mutation
distribution that dynamically adapts strategy parameters during the search process is
an essential feature in ESs.
We briefly review three consecutive steps of adapting normal mutation distribu-
tions in ESs.

1. The normal distribution is chosen to be isotropic. Surfaces of equal probability

density are circles (Figure 1, left), or (hyper-)spheres if n ≥ 3. Overall variance of
the distribution –– in other words the (global) step size or the expected step length ––
is the only free strategy parameter.

2. The concept of global step size can be generalized.2 Each coordinate axis is as-
signed a different variance (Figure 1, middle) –– often referred to as individual step
sizes. There are n free strategy parameters. The disadvantage of this concept is the
dependency on the coordinate system. Invariance against rotation of the search
space is lost. Why invariance is an important feature of an ES is discussed in Sec-
tion 6.

3. A further generalization dynamically adapts the orthogonal coordinate system,

where each coordinate axis is assigned a different variance (Figure 1, right). Any
normal distribution (with zero mean) can be produced. This concept results in
2 There is more than one sensible generalization. For example, it is possible to provide one arbitrarily
oriented axis with a different variance. Then n + 1 parameters have to be adapted. Such an adaptation can
be formulated independent of the coordinate system.

160 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

(n2 + n)/2 free strategy parameters. If an adaptation mechanism is suitably formu-

lated, the invariance against rotations of search space is restored.

The adaptation of strategy parameters in ESs typically takes place in the concept of
mutative strategy parameter control (MSC). Strategy parameters are mutated, and a new
search point is generated by means of this mutated strategy parameter setting.
We exemplify the concept formulating an (1, λ)-ES with (purely) mutative control
of one global step size. x(g) ∈ Rn and σ (g) ∈ R+ are the object parameter vector and
step size3 of the parent at generation g. The mutation step from generation g to g + 1
reads for each offspring k = 1, . . . , λ

(g+1)
σk = σ (g) exp (ξk ) (1)
(g+1) (g) (g+1)
xk = x + σk zk , (2)

where:

ξk ∈ R, for k = 1, . . . , λ independent realizations of a random number with

zero
√ mean. Typically, ξk is normally distributed with standard deviation
1/ 2n (Bäck and Schwefel, 1993). We usually prefer to choose P(ξk = 0.3) =
P(ξk = −0.3) = 1/2 (Rechenberg, 1994).

z k ∼ N(0, I) ∈ Rn , for k = 1, . . . , λ independent realizations of a (0, I)-normally

distributed random vector, where I is the unity matrix. That is, components of
z k are independent and identically (0, 1)-normally distributed.
After λ mutation steps are carried out, the best offspring (with respect to the ob-
(g)
ject parameter vector xk ) is selected to start the next generation step. Equation (1)
facilitates the mutation on the strategy parameter level. The standard deviation of ξ
represents the mutation strength on the strategy parameter level.
This adaptation concept was introduced by Rechenberg (1973) for global and indi-
vidual step sizes. Schwefel (1981) expanded the mutative approach to the adaptation
of arbitrary normal mutation distributions, which is discussed in more detail in Section
3.1. Ostermeier et al. (1994b) introduced a first level of derandomization into strategy
parameter control that facilitates an individual step size adaptation in constantly small
populations. (If the standard mutative strategy parameter control is used to adapt in-
dividual step sizes, based on our experience, the population size has to scale linearely
with the problem dimension n.)
In this paper, a second level of derandomization is put forward: The original objec-
tive of mutative strategy parameter control –– that is to favor strategy parameter settings
that produce selected steps with high probability (again) –– is explicitely realized. Com-
plete derandomization, applied to the adaptation of arbitrary normal mutation distri-
butions, leads almost inevitably to the covariance matrix adaptation (CMA) described in
Section 3.2.
The paper is organized as follows. In Section 2, we discuss the basic shortcomings
of the concept of mutative strategy parameter control and review the derandomized
approach to strategy parameter control in detail. In Section 3, the different concepts are
investigated with respect to the adaptation of arbitrary normal mutation distributions.
3 In this paper, step size always refers to σ. Step size σ is the component-wise standard deviation
of the random vector√ σN(0, I ) ∈ R n . For this vector, nσ 2 can be interpreted as overall variance and
E[kσN(0, I )k] ≈ σ n is the expected step length.

Evolutionary Computation Volume 9, Number 2 161

N. Hansen and A. Ostermeier

General demands on such an adaptation mechanism are developed, and the CMA-ES
is introduced. In Section 4, the objective of strategy parameter control is expanded
using search paths (evolution paths) rather than single search steps as adaptation cri-
terion. This concept is implemented by means of so-called cumulation. In Section 5,
we formulate the CMA-ES algorithm, an evolution strategy that adapts arbitrary, nor-
mal mutation distributions within a completely derandomized adaptation scheme with
cumulation. Section 6 discusses the test functions used in Section 7, where various sim-
ulation results are presented. In Section 8, a conclusion is given. Appendix A provides
a MATLAB implementation of the CMA-ES.

2 Derandomization of Mutative Strategy Parameter Control

In the concept of mutative strategy parameter control (MSC), the selection probabil-
ity of a strategy parameter setting is the probability to produce (with these strategy
parameters) an object parameter setting that will be selected. Selection probability of
a strategy parameter setting can also be identified with its fitness. Consequently, we
assume the following aim behind the concept of MSC:4 Find the strategy parameters
with the highest selection probability or, in other words, raise the probability of mu-
tation steps that produced selected individuals.5 This implicitly assumes that strategy
parameters that produced selected individuals before are suitable parameters in the
(near) future. One idea of derandomization is to increase the probability of producing
previously selected mutation steps again in a more direct way than MSC.
Before the concept of derandomization is discussed in detail, some important
points concerning the concept of MSC are reviewed.

• Selection of the strategy parameter setting is indirect. The selection process oper-
ates on the object parameter adjustment. Comparing two different strategy para-
meter settings, the better one has (only) a higher probability to be selected –– due
to object parameter realization. Differences between these selection probabilities
can be quite small, that is, the selection process on the strategy parameter level is
highly disturbed. One idea of derandomization is to reduce or even eliminate this
disturbance.

• The mutation on the strategy parameter level, as with any mutation, produces dif-
ferent individuals that undergo selection. Mutation strength on the strategy para-
meter level must ensure a significant selection difference between the individuals.
In our view, this is the primary task of the mutation operator.

• Mutation strength on the strategy parameter level is usually kept constant

throughout the search process. Therefore, the mutation operator (on the strategy
parameter level) must facilitate an effective mutation strength that is virtually in-
dependent of the actual position in strategy parameter space. This can be compli-
cated, because the best distance measure may not be obvious, and the position-
independent formulation of a mutation operator can be difficult (see Section 3.1).
4 We assume there is one best strategy parameter setting at each time step. Alternatively, to apply different
parameter settings at the same time means to challenge the idea of a normal search distribution. We found
this alternative to be disadvantageous.
5 Maximizing the selection probability is, in general, not identical with maximizing a progress rate. This is
one reason for the often observed phenomenon that the global step size is adapted to be too small by MSC.
This problem is addressed in the path length control (Equations (16) and (17)) by the so-called cumulation
(Section 4).

162 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

• The possible (and the realized) change rate of the strategy parameters between
two generations is an important factor. It gives an upper limit for the adaptation
speed. When adapting n or even n2 strategy parameters, where n is the problem
dimension, adaptation speed becomes an important factor for the performance of
the ES. If only one global step size is adapted, the performance is limited by the
possible adaptation speed only on a linear objective function (see discussion of
parameter d from (3) below).
On the one hand, it seems desirable to realize change rates that are as large as
possible –– achieving a fast change and consequently a short adaptation time. On
the other hand, there is an upper bound to the realized change rates due to the
finite amount of selection information. Greater changes cannot rely on valid in-
formation and lead to stochastic behavior. (This holds for any adaptation mech-
anism.) As a simple consequence, the change rate of a single strategy parameter
must decrease with an increasing number of strategy parameters to be adapted,
assuming a certain constant selection scheme.
In MSC, the change rate of strategy parameters obviously depends (more or less
directly) on the mutation strength on the strategy parameter level. Based on this
observation, considerable theoretical efforts were made to calculate the optimal
mutation strength for the global step size (Schwefel, 1995; Beyer, 1996b). But, in
general, the conflict between an optimal change rate versus a significant selection
difference (see above) cannot be resolved by choosing an ambiguous compromise
for the mutation strength on the strategy parameter level (Ostermeier et al., 1994a).
The mutation strength that achieves an optimal change rate is usually smaller than
the mutation strength that achieves a suitable selection difference. The discrep-
ancy increases with increasing problem dimension and with increasing number of
strategy parameters to be adapted. One idea of derandomization is to explicitly
unlink the change rate from the mutation strength resolving this discrepancy.
Parent number µ and the recombination procedure have a great influence on the
possible change rate of the strategy parameters between (two) generations. As-
suming a certain mutation strength on the strategy parameter level, the possible
change rate can be tuned downwards by increasing µ. This is most obvious for
intermediate multi-recombination: The mean change of µ recombined individu-
√
als is approximately µ times smaller than the mean change of a single individ-
ual. Within the concept of MSC, choosing µ and an appropriate recombination
mechanism is the only way to tune the change rate independently of the mutation
strength (downwards). Therefore, it is not a surprising observation that a success-
ful strategy parameter adaptation in the concept of MSC strongly depends on a
suitable choice of µ: In our experience µ has to scale linearly with the number
of strategy parameters to be adapted. One objective of derandomization is to fa-
cilitate a reliable and fast adaptation of strategy parameters independent of the
population size, even in small populations.

We start our discussion of derandomization from the ES with mutative strategy

parameter control formulated in Equations (1) and (2). For the sake of simplicity, we
still consider adaptation of one global step size only. (The technique of derandomiza-
tion becomes especially important if a large number of strategy parameters has to be
adapted.) The first level of derandomization (Ostermeier et al., 1994a) facilitates inde-
pendent control of mutation strength and change rate of strategy parameters. This can

Evolutionary Computation Volume 9, Number 2 163

N. Hansen and A. Ostermeier

be achieved by slight changes in the formulation of the mutation step. For k = 1, . . . , λ,

µ ¶
(g+1) ξk
σk = σ (g) exp (3)
d
(g+1)
xk = x(g) + σ (g) exp (ξk ) z k , (4)
where d ≥ 1 is the damping parameter, and the symbols from Equations (1) and (2)
in Section 1 are used. This can still be regarded as a mutative approach to strategy
parameter control: If d = 1, Equations (3) and (4) are identical to (1) and (2), because
(g+1)
σ (g) exp (ξk ) in Equation (4) equals σk in Equation (2). Enlarging the damping para-
(g+1)
meter d reduces the change rate between σ (g) and σk leaving the selection relevant
mutation strength on the strategy parameter level (the standard deviation √of ξk in Equa-
tion (4)) unchanged. Instead of choosing the standard deviation to be 1/ 2n and d = 1,
as in the purely p mutative approach, it is more sensible to choose the standard deviation
1/4 and d = n/8, assuming n ≥ 8, facilitating the same rate of change with a larger
mutation strength. Large values for d scale down fluctuations that occur due to stochas-
tic effects, but decrease
p the possible change rate. For standard deviation 1/4 within the
interval 1 ≤ d ≈ < n/4, neither the fluctuations for small d nor the limited change rate
P
for large d decisively worsen the performance on the sphere objective function i x2i .
But if n or even n2 strategy parameters have to be adapted, the trade-off between large
fluctuations and a large change rate come the fore. Then the change rate must be tuned
carefully.
As Rechenberg (1994) and Beyer (1998) pointed out, this concept –– mutating big
and inheriting small –– can easily be applied to the object parameter mutation as well.
It will be useful in the case of distorted selection on the object parameter level and is
implicitly implemented by intermediate multi-recombination.
In general, the first level of derandomization yields the following effects:
• Mutation strength (in our example, standard deviation of ξk ) can be chosen com-
paratively large in order to reduce the disturbance of strategy parameter selection.
• The ratio between change rate and mutation strength can be tuned downwards (in
our example, by means of d) according to the problem dimension and the number
of strategy parameters to be adapted. This scales down stochastic fluctuations of
the strategy parameters.
• µ and λ can be chosen independent of the adaptation mechanism. They, in partic-
ular, become independent of the number of strategy parameters to be adapted.
The second level of derandomization completely removes the selection distur-
bance on the strategy parameter level. To achieve this, mutation on the strategy para-
meter level by means of sampling the random number ξk in Equation (3) is omitted.
Instead, the realized step length kzk is used to change σ. This leads to
µ ¶
(g+1) (g) kz k k − E[kN (0, I) k]
σk = σ exp (5)
d
(g+1)
xk = x(g) + σ (g) z k (6)
kz k k − E[kN (0, I) k] in Equation (5) replaces ξk in Equation (3). Note that E[kN (0, I) k]
is the expectation of kz k k under random selection.6 Equation (5) is not intended to
6 Random selection occurs if the objective function returns random numbers independent of x.

164 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

be used in practice, because the selection relevance of kz k k vanishes for increasing n.7
One can interpret Equation (5) as a special mutation on the strategy parameter level
or as an adaptation procedure for σ (which must only be done for selected individu-
als). Now, all stochastic variation of the object parameters –– namely originated by the
random vector realization z k in Equation (6) –– is used for strategy parameter adapta-
tion in Equation (5). In other words, the actually realized mutation step on the object
parameter level is used for strategy parameter adjustment. If kz k k is smaller than ex-
pected, σ (g) is decreased. If kz k k is larger than expected σ (g) is increased. Selecting a
small/large mutation step directly leads to reduction/enlargement of the step size.
The disadvantage of Equation (5) compared to Equation (1) is that it cannot be
expanded easily for the adaptation of other strategy parameters. The expansion to the
adaptation of all distribution parameters is the subject of this paper.
In general, the completely derandomized adaptation follows the following princi-
ples:

1. The mutation distribution is changed, so that the probability to produce the se-
lected mutation step (again) is increased.

2. There is an explicit control of the change rate of strategy parameters (in our exam-
ple, by means of d).

3. Under random selection, strategy parameters are stationary. In our example, ex-
pectation of log σ (g+1) equals log σ (g) .

In the next section, we discuss the adaptation of all variances and covariances of the
mutation distribution.

3 Adapting Arbitrary Normal Mutation Distributions

We give two motivations for the adaptation of arbitrary normal mutation distributions
with zero mean.

• The suitable encoding of a given problem is crucial for an efficient search with
an evolutionary algorithm.8 Desirable is an adaptive encoding mechanism. To
restrict such an adaptive encoding/decoding to linear transformations seems rea-
sonable: Even in the linear case, O(n2 ) parameters have to be adapted. It is easy
to show that in the case of additive mutation, linear transformation of the search
space (and search points accordingly) is equivalent to linear transformation of the
mutation distribution. Linear transformation of the (0, I)-normal mutation dis-
tribution always yields a normal distribution with zero mean, while any normal
distribution with zero mean can be produced by a suitable linear transformation.
This yields equivalence between an adaptive general linear encoding/decoding
and the adaptation of all distribution parameters in the covariance matrix.

• We assume the objective function to be non-linear and significantly non-

separable –– otherwise the search problem is comparatively easy, because it can
be solved by solving n one-dimensional problems. Adaptation to non-separable
objective functions is not possible if only individual step sizes are adapted. This
7 For an efficient, completely derandomized global step size adaptation, the evolution path p (g+1)
=
p p k
(1 − c) p(g) + c(2 − c) z k , where 2/(n + 2) ≤ c ≤ 2/(n + 1), must be used instead of z k in Equation (5).
This method is referred to as cumulative path length control and implemented with Equations (16) and (17).
8 In biology, this is equivalent to a suitable genotype-phenotype transformation.

Evolutionary Computation Volume 9, Number 2 165

N. Hansen and A. Ostermeier

demands a more general mutation distribution and suggests the adaptation of ar-
bitrary normal mutation distributions.

Choosing the mean value to be zero seems to be self-evident and is in accordance

with Beyer and Deb (2000). The current parents must be regarded as the best approxi-
mation to the optimum known so far. Using nonzero mean is equivalent to moving the
population to another place in parameter space. This can be interpreted as extrapola-
tion. We feel extrapolation is an anti-thesis to a basic paradigm of evolutionary com-
putation. Of course, compared to simple ESs, extrapolation should gain a significant
speed up on many test functions. But, compared to an ES with derandomized strategy
parameter control of the complete covariance matrix, the advantage from extrapolation
seems small. Correspondingly, we find algorithms with nonzero mean (Ostermeier,
1992; Ghozeil and Fogel, 1996; Hildebrand et al., 1999) unpromising.
In our opinion, any adaptation mechanism that adapts a general linear encod-
ing/decoding has to meet the following fundamental demands:

Adaptation: The adaptation must be successful in the following sense: After an adap-
tation phase,
P progress-rates comparable to those on the (hyper-)sphere objective
function i x2i must be realized on any convex-quadratic objective function.9 This
must hold especially for large problem conditions (greater than 104 ) and non-
systematically oriented principal axes of the objective function.

Performance: The performance (with respect to objective function evaluations) on the

(hyper-)sphere function should be comparable to the performance of the (1, 5)-ES
(g) (g+n)
with optimal step size, where ln(fbest /fbest )/λ ≈ 0.14.10 A loss by a factor up to
ten seems to be acceptable.

Invariance: The invariance properties of the (1, λ)-ES with isotropic mutation distribu-
tion with respect to transformation of object parameter space and function value
should be preserved. In particular, translation, rotation, and reflection of object
parameter space (and initial point accordingly) should have no influence on strat-
egy behavior as well as any strictly monotonically increasing, i.e., order-preserving
transformation of the objective function value.

Stationarity: To get a reasonable strategy behavior in cases of low selection pressure,

some kind of stationarity condition on the strategy parameters should be fulfilled
under purely random selection. This is analogous to choosing the mean value
to be zero for the object parameter mutation. At least non-stationarity must be
quantified and judged.

From a conceptual point of view, one primary aim of strategy parameter adapta-
tion is to become invariant (modulo initialization) against certain transformations of
the object parameter space. For any fitness function f : x 7→ f (c · A · x), the global
step size adaptation can yield invariance against c 6= 0. An adaptive general linear
R
9 That is any function f : x 7→ xT Hx + bT x + c, where the Hessian matrix H ∈ n×n is symmetric and
R R
positive definite, b ∈ n and c ∈ . Due to numerical requirements, the condition of H may be restricted to,
e.g., 1014 . We use the expectation of n
k
(g) (g+k)
N
ln((fbest − f ∗ )/(fbest − f ∗ )) with k ∈ as the progress measure
(where f ∗ is the fitness value at the optimum), because it yields comparable values independent of H , b,
c, n, and k. For H = I and large n, this measure yields values close to the common normalized progress
n
measure ϕ? := (g) (r(g) − E[r(g+1) ]) (Beyer, 1995), where r is the distance to the optimum.
r
10 Not taking into account weighted recombination, the theoretically optimal (1 + 1)-ES yields approxi-
mately 0.20.

166 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

encoding/decoding can additionally yield invariance against the full rank n × n ma-
trix A. This invariance is a good starting point for achieving the adaptation demand.
Also, from a practical point of view, invariance is an important feature for assessing a
search algorithm (compare Section 6). It can be interpreted as a feature of robustness.
Therefore, even to achieve an additional invariance seems to be very attractive.
Keeping in mind the stated demands, we discuss two approaches to the adaptation
of arbitrary normal mutation distributions in the next two sections.
3.1 A (Purely) Mutative Approach: The Rotation Angle Adaptation
Schwefel (1981) proposed a mutative approach to the adaptation of arbitrary normal
mutation distributions with zero mean, often referred to as correlated mutations. The
distribution is parameterized by n standard deviations σi and (n2 −n)/2 rotation angles
αj that undergo a selection-recombination-mutation scheme.
The algorithm of a (µ/ρI , λ)-CORR-ES as used in Section 7.2 is briefly described.
At generation g = 1, 2, . . . for each offspring ρ parents are selected. The (component
wise) arithmetic mean of step sizes, angles, and object parameter vectors of the ρ par-
rec rec rec
ents, denoted with σi=1,...,n , αj=1,...,(n2 −n)/2 , and x , are starting points for the muta-
tion. Step sizes and angles read component wise
µ µ ¶ µ ¶¶
(g+1) 1 1
σi = σirec · exp N 0, + Ni 0, √ (7)
2n 2 n
à à µ ¶2 ! !
(g+1) rec 5
αj = αj + Nj 0, π + π mod 2π − π (8)
180

The random number N(0, 1/(2n)) in Equation p (7), which denotes a normal distribution
with zero mean and standard deviation 1/(2n), is only drawn once for all i = 1, . . . , n.
(g+1)
The modulo operation ensures the angles to be in the interval −π ≤ αj < π, which
is, in our experience, of minor relevance. The chosen distribution variances reflect the
typical parameter setting (see Bäck and Schwefel (1993)). The object parameter muta-
tion reads
 
(g+1)
³ ´  σ1
(g+1) (g+1) .. 
x(g+1) = xrec + R α1 , . . . , α(n2 −n)/2 ·  .  · N (0, I) (9)

(g+1)
σn
The spherical N (0, I) distribution is multiplied by a diagonal step size matrix deter-
(g+1) (g+1)
mined by σ1 , . . . , σn . The resulting axis-parallel mutation ellipsoid is newly ori-
ented by successive rotations in all (i.e., (n2 −n)/2) two-dimensional subspaces spanned
by canonical unit vectors. This complete rotation matrix is denoted with R(.). This al-
gorithm allows one to generate any n-dimensional normal distribution with zero mean
(Rudolph, 1992).
It is generally recognized that the typical values for µ = 15 and λ = 100 are
not sufficient for this adaptation mechanism (Bäck et al., 1997). Due to the mutative
approach, the parent number has presumably to scale with n2 . Choosing µ ≈ n2 /2,
which is roughly the number of free strategy parameters, performance on the sphere
problem declines with increasing problem dimension and becomes unacceptably poor
for n ≈> 10. The performance demand cannot be met. This problem is intrinsic to the
(purely) mutative approach.
The parameterization by rotation angles causes the following effects:

Evolutionary Computation Volume 9, Number 2 167

N. Hansen and A. Ostermeier

Figure 2: Lines of equal probability density of two normal distributions (n = 2) with

5
axis ratios of 1 : 3 (left) and 1 : 300 (right) before and after rotation by 180 π. The right
figure is enlarged in the y-direction; the rectangle covers the identical area in the left
and in the right. Comparing the cross section areas of the ellipsoids, the distributions
with axis ratio of 1 : 3 are very similar, while the distributions with axis ratio 1 : 300
differ substantially.

• The effective mutation strength strongly depends on the position in strategy para-
meter space. Mutation ellipsoids with high axis ratios are mutated much more
strongly than those with low axis ratios: Figure 2 shows two distributions with
different axis ratios before and after rotation by the typical variance. The resulting
cross section area, a simple measure of their similarity, differs in a wide range. For
axis ratio 1 : 1 (spheres, not shown) it is 100%. For axis ratio 1 : 3 (Figure 2 left)
it is about 90%, while it decreases to roughly 5% for axis ratio 1 : 300 (Figure 2
right). The section area tends to become zero for increasing axis ratios. The mu-
tation strength (on the rotation angles) cannot be adjusted independently of the
actual position in strategy parameter space.

• In principle, invariance against a new orientation of the search space is lost, be-
cause rotation planes are chosen with respect to the given coordinate system. In
simulations, we found the algorithm to be dependent even on permutations of the
coordinate axes (Hansen et al., 1995; Hansen, 2000)! Furthermore, its performance
depends highly on the orientation of the given coordinate system (compare Section
7.2). The invariance demand is not met.

Taking into account these difficulties, it is not surprising to observe that the adaptation
demand is not met. Progress rates on convex-quadratic functions with high axis ratios
can be several orders of magnitude lower than progress rates achieved on the sphere
problem (Holzheuer, 1996; Hansen, 2000, and Section 7.2).
When the typical intermediate recombination is applied to the step sizes, they in-
crease unbounded under random selection. The systematic drift is slow, usually causes
no problems, and can even be an advantage in some situations. Therefore, we assume
the stationarity demand to be met.
Using another parameterization together with a suitable mutation operator can
solve the demands for adaptation and invariance without giving up the concept of
MSC (Ostermeier and Hansen, 1999). To satisfy the performance demand the concept
of MSC has to be modified.

3.2 A Completely Derandomized Approach: The Covariance Matrix Adaptation

(CMA)
The covariance matrix adaptation (Hansen and Ostermeier, 1996) is a second-level (i.e.,
completely) derandomized self-adaptation scheme. First, it directly implements the

168 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

aim of MSC to raise the probability of producing successful mutation steps: The covari-
ance matrix of the mutation distribution is changed in order to increase the probability
of producing the selected mutation step again. Second, the rate of change is adjusted
according to the number of strategy parameters to be adapted (by means of ccov in
Equation (15)). Third, under random selection, the expectation of the covariance ma-
trix C is stationary. Furthermore, the adaptation mechanism is inherently independent
of the given coordinate system. In short, the CMA implements a principle component
analysis of the previously selected mutation steps to determine the new mutation dis-
tribution. We give an illustrative description of the algorithm.
Consider a special method to produce realizations of a n-dimensional normal dis-
tribution with zero mean. If the vectors z 1 , . . . , z m ∈ Rn , m ≥ n, span Rn , and N (0, 1)
denotes independent (0, 1)-normally distributed random numbers, then

N (0, 1) z 1 + . . . + N (0, 1) z m (10)

is a normally distributed random vector with zero mean.11 Choosing z i , i = 1, . . . , m,

appropriate any normal distribution with zero mean can be realized.
The distribution in Equation (10) is generated by adding the “line distributions”
N (0, 1) z i ∼ N(0, z i z T
i ). If the vector z i is given, the normal (line) distribution with
zero mean, which produces the vector z i with the highest probability (density) of all
normal distributions with zero mean, is N(0, z i z T i ). (The proof is simple and omitted.)
The covariance matrix adaptation (CMA) constructs the mutation distribution out
of selected mutation steps. In place of the vectors z i in Equation (10), the selected
mutation steps are used with exponentially decreasing weights. An example is shown
in Figure 3, where n = 2. The isotropic initial distribution is constructed by means of
the unit vectors e1 and e2 . At every generation g = 1, 2, . . ., the selected mutation step
(g)
z sel is added to the vector tuple. All other vectors are multiplied by a factor q < 1.12
According to Equation (10), after generation g = 3, the distribution reads
(1) (2) (3)
N (0, 1) q 3 e1 + N (0, 1) q 3 e2 + N (0, 1) q 2 z sel + N (0, 1) qz sel + N (0, 1) z sel (11)

This mutation distribution tends to reproduce the mutation steps selected in the
past generations. In the end, it leads to an alignment of the distributions before and af-
ter selection, i.e., an alignment of the recent mutation distribution with the distribution
of the selected mutation steps. If both distributions become alike, as under random
selection, in expectation, no further change of the distributions takes place (Hansen,
1998).
This illustrative but also formally precise description of the CMA differs in three
points from the CMA-ES formalized in Section 5. These extensions are as follows:

• Apart from the adaptation of the distribution shape, the overall variance of the
mutation distribution is adapted separately. We found the additional adaptation
of the overall variance useful for at least two reasons. First, changes of the overall
variance and of the distribution shape should operate on different time scales. Due
to the number of parameters to be adapted, the adaptation of the covariance ma-
trix, i.e., the adaptation of the distribution shape, must operate on a time scale of
n2 . Adaptation of the overall variance should operate on a time scale n, because the
variance should be able to change as fast as required on simple objective functions
11 The covariance matrix of this normally distributed random vector reads z 1 z T
1 + . . . + zm zm .
T
12 q adjusts the change rate, and q 2 corresponds to 1 − ccov in Equation (15).

Evolutionary Computation Volume 9, Number 2 169

N. Hansen and A. Ostermeier

e2 qe2

e1 qe1

(1)
z sel

g=0 g=1

d11 b1
q 2 e2 3
q e2 (2)
(2)
z sel qz sel

q 2 e1 q 3 e1

(3)
z sel (1)
(1)
qz sel q 2 z sel d22 b2

g=2 g=3

Figure 3: Construction of the mutation distribution in the CMA, where n = 2. The

initial configuration (upper left) consists of the orthogonal unit vectors e1 and e2 . They
(1) (2) (3)
produce an isotropic mutation distribution (dashed). The vectors z sel , z sel , and z sel
are added successively at generations g = 1, 2, 3, while old vectors are multiplied by
q = 0.91. The covariance matrix of the distribution after generation g = 3 reads C (3) =
P3 ³ ´T
2(3−i) (i) (i)
q 6 e1 eT 6 T
1 + q e2 e2 + i=1 q z sel z sel = d211 b1 bT 2 T 2
1 + d22 b2 b2 . The numbers d11
and d222 are eigenvalues of C (3) ; b1 and b2 are the corresponding eigenvectors with unit
length. Realizations from the distribution can be drawn with Equation (11) or more
easily with N (0, 1) d11 b1 + N (0, 1) d22 b2 .

P
like the sphere i x2i . Second, if overall variance is not adapted faster than the
distribution shape, an (initially) small overall variance can jeopardize the search
process. The strategy “sees” a linear environment, and adaptation (erroneously)
enlarges the variance in one direction.

170 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

• The CMA is formulated for parent number µ ≥ 1 and weighted multi-recombina-

(g)
tion, resulting in a more sophisticated computation of z sel .
• The technique of cumulation is applied to the vectors, that construct the distribu-
(g)
tion. Instead of the single mutation steps z sel , evolution paths p(g) are used. An
evolution path p(g) represents a sequence of selected mutation steps. The tech-
nique of cumulation is motivated in the next section.
The adaptation scheme is formalized by means of the covariance matrix of the
distribution, because storing as many as g vectors is not practicable. That is, in every
generation g, after selection of the best search points has taken place,
1. The covariance matrix of the new distribution C (g) is calculated due to C (g−1) and
(g)
the vector z sel (with cumulation p(g) ). In other words, the covariance matrix of the
distribution is adapted.
2. Overall variance is adapted by means of the cumulative path length control. Re-
ferring to Equation (5), z k is replaced by a “conjugate” evolution path (pσ in Equa-
tions (16) and (17)).
3. From the covariance matrix, the principal axes of the new distribution and their
variances are calculated. To produce realizations from the new distribution the n
line distributions are added, which correspond to the principal axes of the muta-
tion distribution ellipsoid (compare Figure 3).
Storage requirements are O(n2 ). We note, that “generally, for any moderate n, this is
an entirely trivial disadvantage” (Press et al., 1992). For computational and numerical
requirements, refer to Sections 5.2 and 7.1.

4 Utilizing the Evolution Path: Cumulation

The concept of MSC utilizes selection information of a single generation step. In con-
trast, it can be useful to utilize a whole path taken by the population over a number of
generations. We call such a path an evolution path of the ES. The idea of the evolution
path is similar to the idea of isolation. The performance of a strategy can be evaluated
significantly better after a couple of steps are taken than after a single step. It is worth
noting that both quantity and quality of the evaluation basis can improve.
Accordingly, to reproduce successful evolution paths seems more promising than
to reproduce successful single mutation steps. The former is more likely to maximize
a progress rate while the latter emphasizes the selection probability. Consequently,
we expand the idea of strategy parameter control: The adaptation should maximize
the probability to reproduce successful, i.e., actually taken, evolution paths rather than
single mutation steps (Hansen and Ostermeier, 1996).
An evolution path contains additional selection information compared to single
steps – correlations between single steps successively taken in the generation sequence
are utilized. If successively selected mutation steps are parallel correlated (scalar prod-
uct greater than zero), the evolution path will be comparatively long. If the steps are
anti-parallel correlated (scalar product less than zero), the evolution path will be com-
paratively short. Parallel correlation means that successive steps are going in the same
direction. Anti-parallel correlation means that the steps cancel each other out. We
assume both correlations to be inefficient. This is most obvious if the correlation/anti-
correlation between successive steps is perfect. These steps can exactly be replaced
with the enlarged/reduced first step.

Evolutionary Computation Volume 9, Number 2 171

N. Hansen and A. Ostermeier

w
v
w
v v
v
−w −w

Figure 4: Two idealized evolution paths (solid) in a ridge topography (dotted). The
distributions constructed by the single steps (dashed, reduced in size) are identical.

Consequently, to maximize mutation step efficiency, it is necessary to realize longer

steps in directions where the evolution path is long –– here the same distance can be cov-
ered by fewer but longer steps. On the other side, it is appropriate to realize shorter
steps in directions where the evolution path is short. Both can be achieved using evo-
lution paths rather than single mutation steps for the construction of the mutation dis-
tribution as described in Section 3.2.
We calculate an evolution path within an iterative process by (weighted) summa-
tion of successively selected mutation steps. The evolution path p ∈ Rn obeys
p (g+1)
p(g+1) = (1 − c) p(g) + c(2 − c) z sel , (12)

where 0 < c ≤ 1 and p(0) = 0. This procedure, introduced in Ostermeier et al. (1994b),
p (g+1)
is called cumulation. c(2 − c) is a normalization factor: Assuming z sel and p(g)
(g+1)
in Equation (12) to be independent and identically distributed yields p ∼ p(g)
(g+1)
independently of c ∈]0; 1]. Variances of p and p are identical because 12 =
(g)
p 2 (g+1)
(1 − c)2 + c(2 − c) . If c = 1, no cumulation takes place, and p(g+1) = z sel . The life
(g)
span of information accumulated in p is roughly 1/c (Hansen, 1998): After about 1/c
generations the original information in p(g) is reduced by the factor 1/e ≈ 0.37. That
means c−1 can roughly be interpreted as the number of summed steps.
The benefit from cumulation is shown with an idealized example. Consider the
two different sequences of four consecutive generation steps in Figure 4. Compare any
evolution path, that is, any sum of consecutive steps in the left and in the right of the
figure. They differ significantly with respect to direction and length. Notice that the
difference is only due to the sign of vectors two and four.
Construction of a distribution from the single mutation steps (single vectors in
the figure) according to Equation (10) leads to exactly the same result in both cases
(dashed). Signs of constructing vectors do not matter, because the vectors are multi-
plied by a (0, 1)-normally distributed random number that is symmetrical about zero.
The situation changes when cumulation is applied. We focus on the left of Figure 4.
Consider a continuous sequence of the shown vectors –– i.e., an alternating sequence of
(2i−1) (2i)
the two vectors v and w. Then z sel = v and z sel = w for i = 1, 2, . . ., and according
(g)
to Equation (12), the evolution path p alternately reads
p ³ ´
(1)
p(g)
v := c(2 − c) v + (1 − c)w + (1 − c)2 v + (1 − c)3 w + . . . + (1 − c)g−1 z sel

172 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

c−1 = 1 c−1 ≈ 1.7

c−1 ≈ 3.1 c−1 ≈ 5.6 c−1 = 10

Figure 5: Resulting cumulation vectors pv and pw and distribution ellipsoids for an

alternating selection of the two vectors v and w from Figure 4, left. For c = 1, no
cumulation takes place,
√ √ and√pv = √ v and pw = w (upper left). Shown is the result for
values of c−1 = 1; 3; 10; 30; 100, respectively.

at odd generation number g and

p ³ ´
(1)
p(g)
w := c(2 − c) w + (1 − c)v + (1 − c)2 w + (1 − c)3 v + . . . + (1 − c)g−1 z sel

at even generation number. For g → ∞ we get

1
p(g)
v → pv = p (v + (1 − c)w) and
c(2 − c)
1
p(g)
w → pw = p (w + (1 − c)v)
c(2 − c)

After 3/c generations, the deviation from these limits is under 5%.
To visualize the effect of cumulation, these vectors and the related distributions are
shown in Figure 5 for different values of c. With increasing life span c−1 , the vectors pv
and pw become more and more similar. The distribution scales down in the horizontal
direction, increasing in the vertical direction. Already for a life span of approximately
three generations, the effect appears to be very substantial. Notice again that this effect

Evolutionary Computation Volume 9, Number 2 173

N. Hansen and A. Ostermeier

is due to correlations between consecutive steps. This information cannot be detected

utilizing single mutation steps separately. Because additional information is evaluated,
cumulation can make adaptation of the distribution parameters faster and more reliable.
In our research, we repeatedly found cumulation to be a very useful concept. In
the following CMA-ES algorithm, we use cumulation in two places: primarily for the
cumulative path length control, where cumulation takes place in Equation (16), and the
step size σ is adapted from the generated evolution path pσ in Equation (17); second,
for the adaptation of the mutation distribution as exemplified in this section and taking
place in Equation (14).

5 The (µW , λ)-CMA-ES Algorithm

We formally describe an evolution strategy utilizing the methods illustrated in Sections
3.2 and 4, denoted by CMA-ES.13 Based on a suggestion by Ingo Rechenberg (1998),
weighted recombination from the µ best out of λ individuals is applied, denoted with
(µW , λ).14 Weighted recombination is a generalization of intermediate multi-recombi-
nation, denoted (µI , λ), where all recombination weights are identical. A MATLAB
implementation of the (µW , λ)-CMA-ES algorithm is given in Appendix A.
(g) (g)
The transition from generation g to g + 1 of the parameters x1 , . . . , xλ ∈ Rn ,
(g) (g)
pc ∈ Rn , C (g) ∈ Rn×n , pσ ∈ Rn , and σ (g) ∈ R+ , as written in Equations (13) to (17),
(0) (0)
completely defines the algorithm. Initialization is pc = pσ = 0 and C (0) = I (unity
(0)
matrix), while the initial object parameter vector hxiw and the initial step size σ (0) have
to be chosen problem dependent. All vectors are assumed to be column vectors.
(g+1)
The object parameter vector xk of individual k = 1, . . . , λ reads
(g+1) (g+1)
xk = hxi(g)
w +σ
(g)
B (g) D (g) z k , (13)
| {z }
∼ N(0,C (g) )

where the following apply:

(g+1)
xk ∈ Rn , object parameter vector of the k th individual in generation g + 1.
(g) Pµ (g)
hxiw := Pµ 1 wi i=1 wi xi:λ , wi ∈ R+ , weighted mean of the µ best individuals of
i=1
generation g. The index i : λ denotes the ith best individual.
σ (g) ∈ R+ , step size in generation g. σ (0) is the initial component wise standard
deviation of the mutation step.
(g+1)
zk ∈ Rn , for k = 1, . . . , λ and g = 0, 1, 2, . . . independent realizations of a (0, I)-
(g+1)
normally distributed random vector. Components of z k are independent
(0, 1)-normally distributed.
C (g) Symmetrical positive definite n × n-matrix. C (g) is the covariance matrix of
the normally distributed random vector B (g) D (g) N(0, I). C (g) determines
¡ ¢T ¡ ¢2 ¡ ¢T
B (g) and D (g) . C (g) = B (g) D (g) B (g) D (g) = B (g) D (g) B (g) , which
is a singular value decomposition of C (g) . Initialization C (0) = I.
13 The algorithm described here is identical to the algorithms in Hansen and Ostermeier (1996) setting
1
µ = 1, cc = cσ , and dσ = β χ b ; in Hansen and Ostermeier (1997) setting w1 = . . . = wµ , cc = cσ and
n
1
dσ = D ; and in Hansen (1998) setting w1 = . . . = wµ .
14 This notation follows Beyer (1995), simplifying the (µ/µ , λ)-notation for intermediate multi-recombi-
I
nation to (µI , λ), avoiding the misinterpretation of µ and µI being different numbers.

174 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

D (g) n × n-diagonal matrix (step size matrix). dij = 0 for i 6= j and diagonal
(g)
elements dii of D (g) are square roots of eigenvalues of the covariance matrix
(g)
C .
B (g) orthogonal n×n-matrix (rotation matrix) that determines the coordinate sys-
tem, where the scaling with D (g) takes place. Columns of B (g) are (defined
as) the normalized eigenvectors of the covariance matrix C (g) . The ith di-
(g) 2
agonal element of D (g) squared dii is the corresponding eigenvalue to the
th (g) (g) (g) (g) 2 (g)
i column of B , bi . That is, C (g) bi = dii bi for i = 1, . . . , n. B is
−1 T
orthogonal, i.e., B =B .
(g+1)
The surfaces of equal probability density of D (g) z k are axis-parallel (hyper-)
ellipsoids. B (g) reorients these distribution ellipsoids to become coordinate system
independent. The covariance matrix C (g) determines B (g) and D (g) apart from signs
of the columns in B (g) and permutations of columns in both matrices accordingly.
2
Conferring to Equation (13), notice that σ (g) B (g) D (g) N(0, I) is (0, σ (g) C (g) )-normally
distributed.
(g+1)
C (g) is adaptated by means of the evolution path pc . For the construction of
(g+1) (g+1)
pc the “weighted mean selected mutation step” B (g) D (g) hziw is used. Notice
(g) (g) (g)
that the step size σ is disregarded. The transition of pc and C reads

p(g+1)
c = (1 − cc ) · p(g) u
c + cc · cw B
(g) (g)
D hzi(g+1)
w (14)
| {z }
 
cw (g+1) (g)
= hxiw −hxiw
σ (g)
³ ´T
C (g+1) = (1 − ccov ) · C (g) + ccov · p(g+1)
c p(g+1)
c , (15)

where the following apply:

(g+1) (0)
pc ∈ Rn , sum of weighted differences of points hxiw . Initialization pc = 0.
³ ´T
(g+1) (g+1)
Note that pc pc is a symmetrical n × n-matrix with rank one.
cc ∈ ]0, 1] determines the cumulation time for pc , which is roughly 1/cc .
p
cuc := cc (2 − cc ) normalizes the variance of pc , because 12 = (1 − cc )2 + cuc 2
(see
Pµ Section 4).
wi (g+1) (g+1)
√
cw := Pi=1 µ 2
is chosen that under random selection cw hziw and z k
i=1 wi
have the same variance (and are identically distributed).
(g+1) Pµ (g+1) (g+1)
hziw := Pµ 1 wi i=1 wi z i:λ , with z k from Equation (13). The index i : λ
i=1
(g+1) (g+1)
denotes the index of the ith best individual from x1 , . . . , xλ . The
(g+1)
weights wi are identical with those for hxiw .
ccov ∈ [0, 1[, change rate of the covariance matrix C. For ccov = 0, no change
takes place.
Equation (15) realizes the distribution change as exemplified in Section 3.2 and Figure
3.
An additional adaptation of the global step size σ (g) is necessary, taking place on a
considerably shorter time scale. For the cumulative path length control in the CMA-ES,

Evolutionary Computation Volume 9, Number 2 175

N. Hansen and A. Ostermeier

(g+1)
a “conjugate” evolution path pσ is calculated, where scaling with D (g) is omitted.

p(g+1)
σ = (1 − cσ )· p(g) u
σ + cσ · cw B
(g)
hzi(g+1) (16)
| {z w }
−1 −1 cw
 
(g+1) (g)
= B (g) (D (g) ) (B (g) ) hxiw −hxiw
σ (g)
à !
(g+1)
(g+1) (g) 1 kpσ k−χ bn
σ = σ · exp · , (17)
dσ χ
bn

where the following apply:

(g+1) (0)
pσ ∈ Rn , evolution path not scaled by D (g) . Initialization pσ = 0.
(g)
cσ ∈ ]0,
p1[ determines the cumulation time for pσ , which is roughly 1/cσ .
cuσ := cσ (2 − cσ ) fulfills (1 − cσ )2 + cuσ 2 = 1 (see Section 4).
dσ ≥ 1, damping parameter, determines the possible change rate of σ (g) in the
generation sequence √ (compare
¡ ¢ Section
¡ n ¢ 2, in particular Equations (3) and (5)).
χ
bn = E[kN(0, I)k] = 2 · Γ n+1 2 /Γ 2 , expectation of the length of a (0, I)-
normally
√ ¡ distributed
¢ random vector. A good approximation is χ bn ≈
1 1
n 1 − 4n + 21n 2 (Ostermeier, 1997).

Apart from omitting the transformation with D (g) , Equations (16) and (14) are identi-
(g+1) (g+1)
cal: In (16) we use B (g) hziw for the cumulation, instead of B (g) D (g) hziw . Under
random selection, B hziw is (0, I)-normally distributed, independently of C (g) .
(g)

Thereby step lengths in different directions are comparable, and the expected length of
pσ , denoted χ bn , is well known.15 The cumulation in Equation (14) often speeds up the
adaptation of the covariance matrix C but is not an essential feature (cc can be set to 1).
For the path length
p control, the cumulation is essential (c−1
σ must not be considerably
smaller than n/2). With increasing n, the lengths of single steps become more and
more similar and therefore more and more selection irrelevant.
While in Equation (13) the product matrix B (g) D (g) only has to satisfy the equation
¡ ¢T
C (g) = B (g) D (g) B (g) D (g) , cumulative path length control in Equation (16) requires
its factorization into an orthogonal and a diagonal matrix.

5.1 Parameter Setting

Besides population size λ and parent number µ, the strategy parameters (w1 , . . . , wµ ),
cc , ccov , cσ , and dσ , connected to Equations (13), (14), (15), (16), and (17), respectively,
have to be chosen.16
The default parameter settings are summarized in Table 1. In general, the selection
related parameters µ, λ, and w1 , . . . , wµ are comparatively uncritical and can be chosen
in a wide range without disturbing the adaptation procedure. We strongly recommend
always choosing µ ≤ λ/2 and the recombination weights according to w1 ≥ . . . ≥ wµ .
By definition, all weights are greater than zero. In real world applications, the default
settings from Table 1 are good first guesses. Only for n < 10 does the default value
yield λ > n. For a quick impression, Table 2 lists the default λ values for a few n,
Arnold (2000) suggested a remarkable simplification of Equation (17), replacing (kpσ
15 Dirk (g+1)
k −
bn with (kp
(g+1) 2
bn )/χ
χ σ k − n)/(2n). This formulation fulfills demands analogously to those from Hansen
(1998, 18f) on (17) and avoids the unattractive approximation of χ bn . In preliminary investigations, both
variants seem to perform equally well. If this result holds, we would prefer the latter variant.
16 On principle, the definition of parameter µ is superfluous, because µ could be implicitely defined by
setting w1 , . . . , wµ > 0 and wµ , . . . , wλ = 0.

176 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

Table 1: Default parameter setting for the (µW , λ)-CMA-ES.

λ µ wi=1,...,µ cc ccov cσ dσ
¡ λ+1 ¢ 4 2 4
4 + b3 ln(n)c bλ/2c ln 2 − ln(i) n+4
√ 2
n+4 c−1
σ +1
(n+ 2)

Table 2: Default λ values, small numbers indicate the truncated portion.

n 2 3 4 6 8 11 15 21 40 208
4 + b3 ln(n)c 6.08 7.30 8.16 9.38 10.24 11.19 12.12 13.13 15.07 20.01

where λ(n − 1) < λ(n). In the (µI , λ)-CMA-ES, where w1 = . . . = wµ ,17 we choose
µ = bλ/4c as the default. To make the strategy more robust or more explorative in case
of multimodality, λ can be enlarged, choosing µ accordingly. In particular, for unimodal
and non-noisy problems, λ = 9 is often most appropriate even for large n.
Partly for historical reasons, in this paper, we use the (µI , λ)-CMA-ES. Based on
our experience, for a given λ, an optimal choice of µ and w1 , . . . , wµ only achieves
speed-up factors less than two compared to the (µI , λ)-CMA-ES, where µ ≈ 0.27λ. The
optimal recombination weights depend on the function to be optimized, and it remains
an open question whether the (µI , λ) or the (µW , λ) scheme performs better overall
(using the default parameters accordingly).
If overall simulation time does not substantially exceed, say, 2n generations, dσ
should be chosen smaller than in Table 1, e.g., dσ = 0.5 c−1 −1
σ + 1. Increasing ccov , dσ or µ
and λ by a factor α > 1 makes the strategy more robust. If this increases the number of
needed function evaluations, as generally to be expected, it will be typically by a factor
less than α. The explorative behavior can be improved by increasing µ and λ, or µ up to
λ/2 in the (µI , λ)-CMA-ES, or dσ in accordance with a suffiently large initial step size.
The parameter settings are discussed in more detail in the following.

λ: (Population size) In general, the equations

λ≥5 and < λ < 2n + 10

2µ ≈ ≈

give a reasonable choice for λ. Large values linearly worsen progress rates, e.g., on
simple problems like the sphere function. Also on more sophisticated problems,
performance can linearly decrease with increasing λ, because the adaptation time
(in generations) is more or less independent of λ. Values considerably smaller than
ten may decline the robustness of the strategy.

µ: (Parent number) We recommend choosing 2 ≤ µ ≈ < n. In the (µI , λ)-CMA-ES, in

most cases, µ ≈ 0.27λ will suffice (Beyer, 1996a; Herdy, 1993). To provide a robust
strategy, large µ and if need be a larger ratio of µ/λ up to 0.5 are preferable (Hansen
and Ostermeier, 1997). In particular for n ≈ < 5, even µ = 1 can occasionally be the
best choice.
17 The algorithm is independent from multiplication of w = (w , . . . , w ) with a real number greater than
1 µ
zero.

Evolutionary Computation Volume 9, Number 2 177

N. Hansen and A. Ostermeier

cσ : (Cumulation for step size) Investigations based on Hansen (1998) give strong evi-
dence that cσ < 1 can and must be chosen with respect to the problem dimension
r
n < −1 <
c n ,
2 ≈ σ ≈
while for large n, most sensible values are between n/10 and n/2. Large values, i.e.,
c−1 −1
σ ≈ n, slow down the possible change rate of the step size because dσ = cσ + 1
but can still be a good choice for certain problem instances.
dσ : (Damping for step size) According to principle considerations, the damping para-
meter must be chosen dσ ≈ α c−1 σ , where α is near one, and dσ ≥ 1. Therefore we
define dσ = α c−1σ + 1. Choosing α smaller than one, e.g., 0.3, can yield (slightly)
larger progress rates on the sphere problem. Depending on n, cσ , µ, and λ, a factor
up to three can be gained. This is recommended if overall simulation time is con-
siderably shorter than 3nλ function evaluations. Consequently one may choose
α = 1 − min(0.7, n/gmax ). If α is chosen too small, oscillating behavior of the step-
size can occur, and strategy performance may decline drastically. Larger values for
α linearly slow down the possible change rate of the step size and (consequently)
make the strategy more robust.
cc : (Cumulation for distribution) Based on empirical observations, we suspect cσ ≤
cc ≤ 1 to be a sensible choice for cc . Especially when long axes have to be learned,
cc ≈ cσ should be most efficient without compromising the learning of short axes.
ccov : (Change rate of the covariance matrix) With decreasing change rate ccov , reliability
of the adaptation increases (e.g., with respect to noise) as well as adaptation time.
For ccov → 0, the CMA-ES becomes an ES with only global step size adaptation.
For strongly disturbed problems,
√ it can be appropriate to choose smaller change
rates like ccov = α/(n + α)2 with α ≈ < 1 instead of α = 2.

5.2 Limitations and Practical Hints

Besides the limitations of any general linear encoding/decoding scheme, the limita-
tions for the CMA, revealed so far, result from shortage of valid selection information
(Hansen and Ostermeier, 1997) or numeric precision problems. The former can be due
to selection irrelevant parameters or axes, weak or distorted selection, or also, numeric
precision problems. In any of these cases, a random walk on strategy parameters will
appear. In the CMA, the condition of the covariance matrix C increases unbounded
bringing search, with respect to the short principal axes of C, to a premature end.
The initial step size σ (0) should be chosen such that σ does not tend to increase
significantly within the initial 2/cσ generations. Otherwise the initially learned distri-
bution shape can be inefficient and may have to be unlearned consuming a consider-
able number of additional function evaluations. This effect of a too small σ (0) can be
avoided keeping the mutation distribution shape (i.e., the covariance matrix) initially
constant in case of an initially increasing σ. Furthermore, the prominent effect of the
initial step size on the global search performance (see Sections 7.4 and 6) should be
kept in mind.
In the practical application, a minimal variance for the mutation steps should be
ensured. Remember that dii is the ith diagonal element of the step size matrix D. To
ensure a minimal variance, the standard deviation of the shortest principal axis of the
mutation ellipsoid, σ · mini (dii ), must be restricted. If it falls below the given bound,

178 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

step size σ should be enlarged (compare Appendix A). Problem specific knowledge
must be used for setting the bound. Numerically, a lower limit is given by the demand
(g) (g)
hxiw 6= hxiw + 0.2 · σ B (g) D (g) ei for each unit vector ei . With respect to the selection
(g) (g)
procedure, even f (hxiw ) 6= f (hxiw + 0.2 · σ B (g) D (g) ei ) seems desirable.
To avoid numerical errors, a maximal condition number for C (g) , e.g., 1014 , should
be ensured. If the ratio between the largest and smallest eigenvalue of C (g) exceeds
(g) (g)
1014 , the operation C (g) := C (g) +(maxi (dii )2 /1014 −mini (dii )2 ) I limits the condition
number in a reasonable way to 1014 + 1.
The numerical stability of computing eigenvectors and eigenvalues of the symmet-
rical matrix C usually poses no problems. In MATLAB, the built-in function eig can be
used. To avoid complex solutions, the symmetry of C must be explicitly enforced (com-
pare Appendix A). In C/C++, we used the functions tred2 and tqli from Press et
al. (1992), substituting float with double and setting the maximal iteration number
in tqli to 30n. Another implementation was done in Turbo-PASCAL. For condition
numbers up to 1014 , we never observed numerical problems in any of these program-
ming languages.
If the user decides to manually change object parameters during the search pro-
cess – which is not recommended – the adaptation procedure (Equations (14) to (17))
must be omitted for these steps. Otherwise, the strategy parameter adaptation can be
severely disturbed.
Because change of C (g) is comparatively slow (time 2
√ scale n ), it is possible to up-
date B (g) and D (g) not after every generation but after n or even after, e.g., n/10 gen-
erations. This reduces the computational effort of the strategy from O(n3 ) to O(n2.5 ) or
O(n2 ), respectively. The latter corresponds to the computational effort for a fixed linear
encoding/decoding of the problem as well as for producing a realization of an (arbi-
trarily) normally distributed, random vector on the computer. In practical applications,
it is often most appropriate to update B (g) and D (g) every generation.
A simple method to handle constraints repeats the generation step, defined in
Equation (13), until λ or at least µ feasible points are generated before Equations (14)
to (17) are applied. If the initial mutation distribution generates sufficiently numer-
ous feasible solutions, this method can remain sufficient during the search process due
to the symmetry of the mutation distribution. If the minimum searched for is located
at the edge of the feasible region, the performance will usually be poor and a more
sophisticated method should be used.

6 Test Functions
According to Whitley et al. (1996), we prefer test functions that are non-linear, non-
separable, scalable with dimension n, and resistant to simple hill-climbing. In addition,
we find it reasonable to use test functions with a comprehensible topology even if n ≥ 3.
This makes the interpretation of observed results possible and can lead to remarkable
scientific conclusions. Table 3 gives the test functions used in this paper. The test suite
mainly strives to test (local) convergence speed. Only functions 10–12 are multimodal.18
Apart from functions 6–8, the global optimum is located at 0.
To yield non-separability, we set y := [o1 , . . . , on ]T x, where x is the object parame-
ter vector according to Equation (13). [o1 , . . . , on ]T implements an angle-preserving lin-
ear transformation of x, i.e., rotation and reflection of the search space. o1 , . . . , on ∈ Rn
18 For higher dimension, even function 8 fRosen has a local minimum near y = (−1, 1, . . . , 1)T . With the
given initialization of hxiw and σ (0) , the ES usually converges into the global minimum at y = 1.
(0)

Evolutionary Computation Volume 9, Number 2 179

N. Hansen and A. Ostermeier

Table 3: Test functions (to be minimized), where y = [o1 , . . . , on ]T x.

(0)
Function σ (0) hxiw f stop
Pn Pn
1. fsphere (x) = i=1 xi 2 1 i=1 oi 10−10
Pn ³Pi ´2 Pn
2. fSchwefel (y) = i=1 j=1 yi 1 oi 10−10
Pn Pi=1
n
3. fcigar (y) = y1 2 + i=2 (1000
P yi )2 1 Pi=1 oi 10−10
n n
4. ftablet (y) = (1000 y1 )2 + i=2 yi 2 1 i=1 oi 10−10
Pn ³ i−1
´2 Pn
5. felli (y) = i=1 1000 n−1 yi 1 i=1 oi 10−10
Pn
6. fparabR (y) = −y1 + 100p i=2 yi 2 1 0 −1000
Pn 2
7. fsharpR (y) = −y1 + 100 i=2 yi 1 0 −1000
Pn−1 ³ ¡ 2 ¢2
8. fRosen (y) = i=1 100 yi − yi+1 0.1 0 10−10
´
+(yi − 1)2
Pn i−1 Pn
9. fdiff pow (y) = i=1 |yi |2+10 n−1 0.1 i=1 oi 10−15
Pn ¡ Pn
10. fRast (y) = 10n + i=1 (ai yi )2 ¢ 100
i=1 ui oi
−10 cos (2π · ai yi ) , ui ∈ [−5.12; 5.12]
ai = 1
11. fRast10 (y) = fRast (y), 100 Pn
i−1 i=1 ui oi
where ai = 10 n−1 ui ∈ [−5.12; 5.12]
12. fRast1000 (y) = fRast (y), 100 Pn
i−1 i=1 ui oi
where ai = 1000 n−1 ui ∈ [−5.12; 5.12]

is a randomly oriented, orthonormal basis, fixed for each simulation run. Each oi is
realized equally distributed on the unit (hyper-)sphere surface, dependently drawn so
that hoi , oj i = 0 if i 6= j. An algorithm to generate the basis is given in Figure 6. For
(0) Pn (0)
the initial hxiw = i=1 oi , we yield y (0) = [o1 , . . . , on ]T hxiwP= (1, . . . , 1)T . For the
canonical basis, where oi = ei , it holds y = x and, obviously, i oi = (1, . . . , 1)T . For
a non-canonical basis, only fsphere remains separable.
Functions 1–5 are convex-quadratic and can be linearly transformed into the sphere
problem.19 They serve to test the adaptation demand stated in Section 3. The axis
scaling between longest and shortest principal axis of functions 3–5 is 1000, i.e., the
problem condition is 106 . Taking into account real-world problems, a search strategy
should be able to handle axis ratios up to this number. ftablet can be interpreted as a
sphere model with a smooth equality constraint in o1 direction.
fparabR and fsharpR facilitate straight ridge topologies. The ridge points into o1
direction that has to be enlarged unlimited. The sharp ridge function fsharpR is topo-
logically invariant from the distance to the ridge peak. In particular, the local gradient
is constant, independent of the distance, which is a hard feature for a local search strat-
egy. Results can be influenced by numerical precision problems (even dependent on the
implementation of the algorithm). Therefore, we establish for this problem a minimal
step size of 10−10 and a maximal condition number for C (g) of 1014 (compare Section
19 They can be written in the form f (x) = f
sphere (Ax) = (Ax) Ax = x A Ax = x Hx, where A is
T T T T

a full rank n × n-matrix, and the Hessian matrix H = AT A is symmetric and positive definite.

180 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

FOR i = 1 TO n

1. Draw components of oi independently (0, 1)-normally distributed

Pi−1
2. oi := oi − j=1 hoi , oj i oj (h., .i denotes the canonical scalar product)
3. oi := oi /koi k
ROF

Figure 6: Algorithm to generate a random orthonormal basis o1 , . . . , on ∈ Rn (Hansen

et al., 1995).

5.2 and Appendix A).

The generalized Rosenbrock function fRosen , sometimes called the “banana func-
tion,”Pfacilitates a bent ridge, where the global optimum is at y = 1, which means
n
x = i=1 oi . Like fRosen , the sum of different powers fdiff pow cannot be linearly trans-
formed into the sphere problem. Here the mis-scaling continually increases while ap-
proaching the optimum. A local search strategy presumably follows a gradually nar-
rowing ridge.
Functions 10–12 are multimodal. fRast is the generalized form of the well-known
Rastrigin function, while fRast10 and fRast1000 are mis-scaled versions of fRast . The
mis-scaling between longest and shortest axis is 10 for fRast10 and 1000 for fRast1000 .
For n = 20, the factor between “adjacent” axes is 1.13 and 1.44, respectively. We feel
the moderately mis-scaled Rastrigin function fRast10 is a much more realistic scenario
than the perfectly scaled one. On the multimodal test functions, the initial step size is
chosen comparatively large. With small initial step sizes, the local optimum found by
the ES is almost completely determined by the initial point that is, in each coordinate,
equally distributed in [−5.12; 5.12].
Invariance (compare also Section 3) is an important feature of a search strategy, be-
cause it facilitates the possibility of generalizing simulation results. Imagine, for exam-
ple, translation invariance is not given. Then it is not sufficient to test f : x 7→ f (x − a)
for, e.g., a = 0. One has to test various different values for a and may end up with
different results open to obscure interpretations. Typically, ESs are translation invari-
ant. Simple ESs are, in addition, invariant against rotation and reflection of the search
space, i.e., they are independent of the choice of the orthonormal coordinate system
o1 , . . . , on . More complex ES algorithms may be lacking this invariance (compare Sec-
tions 3.1 and 7.2).20 Any evaluation of search strategies (for example, by test functions)
has to take into account this important point.

7 Simulation Results
Four different evolution strategies are experimentally investigated.
• (µI , λ)-CMA-ES, where the default parameter setting from Table 1 is used apart
from λ and µ as given below, and wi = 1 for all i = 1, . . . , µ. To reduce the compu-
update of B (g) and D (g) from C (g) for Equations (13), (14), and
tational effort, the √
(16) is done every n generations in all simulations.

20 For example, invariance against rotation is lost when discrete recombination is applied.

Evolutionary Computation Volume 9, Number 2 181

N. Hansen and A. Ostermeier

• (15/2I , 100)-CORR-ES (correlated mutations), an ES with rotation angle adapta-

tion according to Section 3.1. The initialization of the rotation angles is randomly
uniform between −π and π, identical for all initial individuals.
• (2I , 10)-PATH-ES, the same strategy as the (2I , 10)-CMA-ES, while ccov is set to
zero. C (g) = I for all g = 0, 1, 2, . . ., and therefore, B (g) and B (g) D (g) can be set to
I. Equations (14) and (15) become superfluous, and only cumulative path length
control takes place.
• (2I , 10)-MUT-ES, an ES with mutative adaptation of one (global) step size and in-
termediate recombination on object and strategy parameters as in CORR-ES. Mu-
tation on the strategy parameter level is carried out as in Equation (1), where ξk is
√ 2
(0, 1/ 2n )-normally distributed.
In the beginning of Section 3, we formulated four fundamental demands on an
algorithm that adapt a general linear encoding in ESs. Now we pursue the question
whether the CMA-ES can satisfy these demands.
The invariance demand is met because the algorithm is formulated inherently in-
dependent of the coordinate system: All results of CMA-ES, PATH-ES, and MUT-ES
are independent of the basis o1 , . . . , on actually chosen (see Section 6), i.e., valid for
any orthonormal coordinate system. That means, these strategies are, in particular,
independent of rotation and reflection of the search space.
The performance demand is also satisfied. The (2I , 9)-CMA-ES performs on fsphere
between 1.5 (n large) and 3 (n = 5) times slower than the (1, 5)-ES with isotropic mu-
tation distribution and optimal step size. The applicability of the CMA algorithm is
independent of any reasonable choice of µ and λ.
The stationarity demand is respected because, under random selection, expecta-
tion of C (g+1) equals C (g) , and expectation of log σ (g+1) equals log σ (g) (Hansen, 1998).
Simulation runs on functions 3–5 in Section 7.2 will show that the adaptation de-
mand is met as well.
In Section 7.3, performance results on the unimodal test functions 1–9 are pre-
sented, and scaling properties are discussed. In Section 7.4, global search properties
are evaluated.
7.1 CPU-Times
We think strategy internal time consumption is of minor relevance. It seems needless to
say that results strongly depend on implementational skills and the programming lan-
guage chosen. Nevertheless, to complete the picture, we summarize CPU-times taken
by strategy internal operations of different (2I , 10)-ESs in Figure 7. The strategies were
coded in MATLAB and run uncompiled on a pentium 400 MHz processor. Differences
between MUT-ES and PATH-ES are presumably due to the more sophisticated recom-
bination operator in MUT-ES, which allows two parent recombination independent of
the parent number µ. MATLAB offers relatively efficient routines for the most time con-
suming computations in the CMA-ES. Therefore, even for n = 320, the CMA-ES needs
only 50ms CPU-time per generated offspring. This time can still be √reduced doing the
update of B (g) and D (g) every n/10 generations instead of every n generations. No
efficient MATLAB routines are available for the rotation procedure in the CORR-ES. To
make CPU-times comparable, this procedure was reimplemented in C and called from
MATLAB gaining a remarkable speed up. CORR-ES is still seven to ten times slower
than CMA-ES. This may be due to a trivial implementational matter and does not make
any difference for many non-linear, non-separable search problems.

182 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

CORR
CPU−Times [s]
−1
10
CMA

seconds
−2
10

−3
10
MUT
PATH

−4
10
5 10 20 40 80 160 320
dimension

Figure 7: Strategy internal CPU-times per function evaluation (i.e., per generated off-
spring) of different (2I , 10)-ESs in seconds. The strategies were coded in MATLAB
(compare text) and run on a pentium 400 MHz processor.

7.2 Testing Adaptation

In this section, single simulation runs on the convex-quadratic functions 3–5 are shown
in comparison with runs on fsphere , revealing whether the adaptation demand can be
met.
We start with a discussion of the (15/2I , 100)-CORR-ES. In Figure 8, with respect
to the stop criterion, the best, middle, and worst out of 17 runs on fcigar , ftablet , and
felli are shown, where n = 5.21 Respective results for the (15I , 100)-CMA-ES are shown
for comparison. This selection scheme is not recommended for the CMA-ES, where
n = 5 (see Section 5.1, in particular Table 1 and 2). The recommended (4W , 8)-CMA-ES
performs roughly ten times faster (compare also Figure 10).
First, we compare the results between the left and right column in the figure. On
the left, the axis-parallel, and therefore completely separable, versions of the functions
are used (oi = ei and y = x, compare Section 6). Only axis-parallel mutation ellipsoids
are necessary to meet the adaptation demand. On the right, the basis o1 , . . . , on is
randomly chosen anew for each run. The CORR-ES performs roughly ten to forty times
slower on the non-axis-parallel oriented versions. In accordance with previous results
(Hansen et al., 1995; Holzheuer, 1996; Hansen, 2000), the CORR-ES strongly depends
on the orientation of the given coordinate system and largely exploits the separability
of the problem. In contrast, the CMA-ES performs identically on the left and right.
Only on the axis-parallel versions of felli and ftablet does the CORR-ES partly meet
the adaptation demand. At times, after an adaptation phase, progress rates are similar
to those on fsphere . On the non-axis-parallel functions, the progress rates are worse by
a factor between 100 (best run on felli ) and 6000 (worst run on fcigar ) compared to those
on fsphere . As Rudolph (1992) pointed out, the search problem on the strategy param-
eters is multimodal. Even after a supposed adaptation phase, on most test functions,
long phases with different progress rates can be observed. This suggests the hypothe-
sis that the ratio between the effect of the mutation (of the angles) and the width of the
21 With 17 runs, one would expect roughly 95% of any simulations to end up between the shown best and
worst run.

Evolutionary Computation Volume 9, Number 2 183

N. Hansen and A. Ostermeier

2 2
Cigar, Axis−Parallel Cigar, Non−Axis−Parallel
0 0

−2 −2

−4 −4

−6 −6

−8 −8

−10 −10
0 1000 2000 3000 4000 0 10000 20000 30000 40000
generations = feval/100 generations = feval/100
2 2
Tablet, Axis−Parallel Tablet, Non−Axis−Parallel
0 0

−2 −2

−4 −4

−6 −6

−8 −8

−10 −10
0 200 400 600 800 1000 0 2000 4000 6000 8000 10000
generations = feval/100 generations = feval/100
2 2
Ellipsoid, Axis−Parallel Ellipsoid, Non−Axis−Parallel
0 0

−2 −2

−4 −4

−6 −6

−8 −8

−10 −10
0 100 200 300 400 500 600 0 2000 4000 6000 8000 10000 12000
generations = feval/100 generations = feval/100

Figure 8: log10 (function value) vs. generation number with (15/2I , 100)-CORR-ES (5),
where n = 5. The best, middle, and worst out of 17 runs are shown. Bold symbols at
the lower edge indicate the mean generation number to reach f stop = 10−10 . In the left
column, oi = ei , that is, the functions are axis-parallel oriented and completely sepa-
rable. The range of the abscissa between left and right columns is enlarged by a factor
20 for felli (lower row) and 10 otherwise. For comparison are shown a single run on
fsphere (∗) and respective simulations with the (15I , 100)-CMA-ES (°) that do not reflect
the real performance of the CMA-ES. The default (4W , 8)-CMA-ES performs roughly
ten times faster. The CORR-ES does not meet the adaptation demand. Progress rates
on the non-separable functions are worse by a factor between 100 and 6000 compared
to those on fsphere .

184 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

10
8 6
Cigar
6 4
4 2
2 0
0 −2
−2 Tablet
−4
−4
−6
−6
−8 −8
−10 −10
−12 −12
0 100 200 300 400 0 200 400 600 800 1000
generations = feval/10 generations = feval/10

6 5
4
2 0
0 Sphere
Ellipsoid −5
−2
−4 −10
−6 Ellipsoid
−8 −15
−10
0 100 200 300 400 500 600 700 0 200 400 600 800 1000 1200 1400
generations = feval/10 generations = feval/10

Figure 9: Simulation runs with the (2I , 10)-CMA-ES, where n = 10. °: log10 (function
value). 2: log10 (smallest and largest variance of the mutation distribution), ∗:
log10 (function value) from a run on fsphere . The upper ten graphs (without symbol)
are the variances of the principal axes of the mutation distribution ellipsoid, sorted,
2
and multiplied with 105 /σ (g) on a logarithmic scale. Lower right: Simulation run on
−5
felli until function value 10 is reached and afterwards on const · fsphere . The CMA-
ES meets the adaptation demand. After an adaptation phase, in all cases, the progress
rates are identical with those on fsphere .

respective optimum differs significantly between different optima. Therefore, different

progress rates can be observed.
Finally, we stress reliability and replicability of the results with the CORR-ES. First,
the initialization of the angles is of major importance for the results. Initializing the
angles with zero, which is a global optimum for the axis-parallel functions, is almost
ten times slower on the axis-parallel felli . Also, initializing all individuals with different
angles is still considerably worse than the results shown. Second, the implemented
order of applied rotations influences the results especially on fcigar and ftablet . Using,
e.g., the reverse rotation order, results are significantly worse.
We continue with a discussion of the (2I , 10)-CMA-ES. Due to a small variance
between different simulation runs (compare Figures 8 and 10), it is sufficient here to
look at single runs on different functions. Figure 9 shows simulation runs on fcigar

Evolutionary Computation Volume 9, Number 2 185

N. Hansen and A. Ostermeier

(upper left) and ftablet (upper right), where n = 10, compared to a run on fsphere . The
adaptation process can be clearly observed through the variances of the principle axes
of the mutation distribution, where global step size is disregarded (upper ten graphs).
When the variances correspond to the scaling in the objective function, progress rates
identical to those on fsphere are realized (notice again that the basis o1 , . . . , on is chosen
randomly here). The shorter adaptation time on fcigar is due to the cumulation, which
detects the parallel correlations of single steps very efficiently.
In the lower left of Figure 9, a simulation run on felli is shown. Similar to ftablet ,
but more pronounced, local adaptation occurs repeatedly on this function. Progress
increases and decreases a few times together with the step size. When the adaptation
is completed, the variances are evenly spread in a range of 106 . In the lower right of
Figure 9, the objective function is switched from felli to (some multiple of) fsphere , after
reaching function value 10−5 . The mutation distribution adapts from an ellipsoid-like
shape back to the isotropic distribution. Adaptation time and graphs of function value
and step size are very similar to those on felli . In fact, from a theoretical point of view,
the algorithm must show exactly the same behavior in both adaptation cases (apart
from stochastic influences). Again, after the adaptation phase, progress rates identical
to those on fsphere are achieved in both cases.
Concluding these results, the adaptation demand is satisfied by the CMA-ES: Any
convex-quadratic function is rescaled into the sphere function.

7.3 Testing Convergence Speed and Scaling

In this section, we investigate the number of function evaluations to reach f stop on
test functions 1–9, where n = 5; 20; 80; 320. The (2I , 10)-CMA-ES is compared to the
(2I , 10)-PATH-ES, where only global step size adaptation takes place and which usually
(slightly) outperforms the (2I , 10)-MUT-ES. Depending on CPU-time resources, up to
50 simulation runs are evaluated. Figure 10 shows the results on the convex-quadratic
test functions 1–5 (first row). On fsphere , both strategies perform almost identically,
while on fSchwefel , the results are still comparable (upper left). The difference on fsphere
is due to the decreasing variance of the covariance matrix supporting the adaptation of
the step size σ that is somewhat too slow (see above discussion of damping parameter
dσ in Section 5.1). If the axis ratio between the longest and shortest axis is 1 : 1000,
as on fcigar , ftablet , and felli (upper right), CMA-ES outperforms PATH-ES by a factor
between 7 (ftablet , n = 320) and almost 60 000 (fcigar ). Only on ftablet for n > 30 does
the factor fall below 100.
The lower row of Figure 10 shows the results on functions 6–9. The CMA-ES works
quite well even on different kinds of ridge-like topologies. On fRosen , the CMA-ES
is 7–80 times faster than PATH-ES; on fparabR , the factor becomes 2000 (lower left).
Results for the PATH-ES on fsharpR and fdiff pow are omitted (lower right). On the latter,
f stop is reached after more than 1010 function evaluations (n = 5). On the former, the
PATH-ES does not reach f stop because step size converges to zero (as with mutative
step size control). Setting the minimum step size to 10−10 , far more than 1010 function
evaluations are needed to reach f stop .
We take an interesting look at the scaling properties with n (compare the sloping
grids in the figure). As with simple ESs in general, PATH-ES scales linearly on most
functions. There are two exceptions. First, on fSchwefel the PATH-ES scales quadrat-
ically. The axis ratio between longest and shortest principal axis of this function in-
creases with increasing n. That is, not only problem dimension but also “problem diffi-
culty” increases. Therefore, a simple ES scales worse than linear. This favors CMA-ES

186 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

2 9 3
10 5
6 2
10 8
10
function evaluations

function evaluations
7 4
5 10
10 5
6
4
1 10
5
4 10 3
10
4 3 cigar
1 sphere 10
4 4 tablet
3 2 schwefel
10 3 5 ellipsoid
10
4 10 30 100 300 4 10 30 100 300
dimension dimension
8 6 8
10 10
8
7 7
10 10 9
function evaluations

function evaluations

8
7
6 6
10 10

5 5
10 10
6
4 4
10 10
6 parabolic ridge 7 sharp ridge
3 8 rosenbrock 3 9 different powers
10 10
4 10 30 100 300 4 10 30 100 300
dimension dimension

Figure 10: Function evaluations to reach f stop with the (2I , 10)-CMA-ES (solid graphs)
and the (2I , 10)-PATH-ES (dashed graphs) vs. dimension n on the test functions 1–9,
where n = 5; 20; 80; 320. Shown are mean values and standard deviations. Sloping
dotted lines indicate const·n = [10; 100; 1000; . . .]·n and const·n2 = [1/100; 1; 100; . . .]·n2
function evaluations. For the PATH-ES the missing result on fcigar (n = 320) could not
be obtained in reasonable time (upper right) and results on fsharpR and fdiff pow are
far above the shown area (lower right, compare text). On fsphere (upper left), both
strategies perform almost identically.

scaling slightly better than PATH-ES on fSchwefel . Second and more surprisingly, per-
formance of the PATH-ES on ftablet is almost independent of n. This effect is due to
the cumulative path length control that adapts here better (i.e., larger) step sizes with
higher dimension.
In general, we expected the CMA-ES to scale quadratically with n. The number
of free strategy parameters to be adapted is (n2 + n)/2. fRosen and fdiff pow meet our
expectations quite well. On fsphere , no adaptation is necessary, and CMA-ES scales
linearly with n, as to be expected of any ES.

Evolutionary Computation Volume 9, Number 2 187

N. Hansen and A. Ostermeier

In contrast, the perfect linear scaling on fcigar and fparabR , even though desirable,
comes as a surprise to us. On both functions one long axis has to be adapted. We
found the cumulation to be responsible for this excellent scaling. Without cumulation,
the scaling on fcigar is similar to the scaling on ftablet . As mentioned above, the cu-
mulation especially supports the adaptation of long axes. On ftablet , fsharpR , felli , and
fSchwefel , increasingly ordered, the scaling is between n1.6 and n1.8 . Where the CMA-ES
scales worse than the PATH-ES, performances align with n → ∞ because the progress
surpasses the adaptation procedure with n → ∞.
In summary, the CMA-ES substantially outperforms the PATH-ES in dimensions
up to 320 on all tested functions –– apart from fsphere , where both strategies perform
almost identically. The CMA-ES always scales between n and n2 : Exactly linearly for
the adaptation of long axes and if no adaptation takes place, nearly quadratically if a
continuously changing topology demands persistent adaptation.

7.4 Testing Global Search Performance

In evolutionary computation, the aspect of global search is often emphasized. In con-
trast, we interpret ESs –– not taking into account the initial search phase –– as local search
strategies. The population occupies a comparatively small area, and the horizon be-
yond this area is limited by the actual step size: Steps larger than some multiple of
the distribution variance do virtually not appear.22 In our opinion, even the so-called
premature convergence is often due to the lack of local convergence speed. When de-
veloping adaptation mechanisms, we mainly strove to address local convergence speed
and did not consider global search performance.
Consequently, even though the general judgment of global search performance is
problematic, it is sometimes argued that adaptation to the local topology of the ob-
jective function spoils global search properties of an algorithm. With respect to the
CMA-ES, we discuss this objection now: There is good reason, and some evidence, that
even the opposite is a more appropriate point of view: The local adaptation mechanism
of the CMA-ES improves global search properties.
We compare different (2I , 10)-ESs on the generalized Rastrigin function, where n =
20. Increasing population size improves the performance on this function. In contrast,
the differences between the strategies compared are not affected. Note that a smaller
initial step size worsens the performance as interpreted in Section 6. The CMA-ES is
compared to the MUT-ES, where mutative control of only global step size takes place.
PATH-ES performs similar to MUT-ES on the investigated Rastrigin functions.
Figure 11 shows 30 simulation runs on fRast with the CMA-ES (left) and the MUT-
ES (right). Behavior of both strategies is very similar. They get trapped into local min-
ima with function values between 30 and 100 within about 2000 and 3000 function
evaluations.
Figure 12 shows 30 simulation runs on the scaled Rastrigin function fRast10 that
should be regarded as a more realistic multimodal test problem than fRast . Even though
the mis-scaling between longest and shortest axis is only of a factor ten, the CMA-ES
(left) and the MUT-ES (right) perform quite differently here. Function values obtained
with MUT-ES are worse by a factor of 20 than those obtained with CMA-ES. In addition,
these values are finally reached after about 5 · 105 function evaluations compared to
22 In a pure mathematical sense, this is, of course, wrong. But, in contrast to any theoretical consideration,
in practical applications, the finite time horizon is too short to wait for those events to occur. Even if a
distribution is chosen that facilitates large steps more often, due to a search space volume phenomenon,
these steps will virtually never produce better points if n exceeds, say, ten.

188 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

7 7
10 10
CMA-ES on fRast MUT-ES on fRast
6 6
10 10

5 5
10 10
function value

function value
4 4
10 10

3 3
10 10

2 2
10 10

1 1
10 10
0 200 400 600 800 1000 0 200 400 600 800 1000
generations generations

Figure 11: 30 simulation runs on the generalized Rastrigin function fRast with the
(2I , 10)-CMA-ES (left) and the (2I , 10)-MUT-ES (right), where n = 20. Both strategies
perform very similar.

7
10
CMA-ES on fRast10 MUT-ES on fRast10
6 6
10 10

5
10
function value

function value

4 4
10 10

3
10

2 2
10 10

1
10
0 200 400 600 800 1000 0 200 400 600 800 0 5 10
1000 105
generations generations x 10
4

Figure 12: 30 simulation runs on the scaled Rastrigin function fRast10 , maximal axis
ratio 1 : 10, with the (2I , 10)-CMA-ES (left) and the (2I , 10)-MUT-ES (right), where
n = 20. Function values reached with MUT-ES are worse by a factor 20.

5 · 103 function evaluations with the CMA-ES.

Figure 13 shows 30 simulation runs on fRast1000 . The variance of function values
reached with the CMA-ES (left) is larger than on fRast and fRast10 . Nevertheless, about
half of the simulation runs end up in local optima with function values less than 100.
This takes about 3 · 104 function evaluations, i.e., about ten times longer than on fRast
or fRast10 . The adaptation time is now a decisive factor. With the MUT-ES, obtained
function values are worse by a factor 10 000 than those obtained with the CMA-ES,
and simulation roughly needs ten times the number of function evaluations (strictly
speaking, even after 106 function evaluations, final function values are not yet reached).
We found very similar results to those presented here on the Rastrigin func-
tions in earlier investigation on a more complex, constrained multimodal test problem
(EVOTECH-7, 1997).
Even though at first glance these results are surprising, there is a simple expla-
nation why global search properties are improved. While the shape of the mutation
distribution becomes suitably adapted to the topology of the objective function, the
step size is adjusted much larger than without adaptation. A larger step size improves

Evolutionary Computation Volume 9, Number 2 189

N. Hansen and A. Ostermeier

10 10
10 10

CMA-ES on fRast1000 MUT-ES on fRast1000

8 8
10 10
function value

function value
6 6
10 10

4 4
10 10

2
2
10 10

0 5 105
0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 10 4
generations generations x 10

Figure 13: 30 simulation runs on the scaled Rastrigin function fRast1000 , maximal axis
ratio 1 : 1000, with the (2I , 10)-CMA-ES (left) and the (2I , 10)-MUT-ES (right), where
n = 20. Function values reached with the MUT-ES are worse by a factor 10 000.

the global search performance of a local search procedure. The effect of the distribu-
tion adaptation on the step size can be clearly observed in Figure 9. Variances in all
directions continually increase with the ongoing adaptation to the topology on fcigar
between generations 200 and 330 (shown are the smallest and the largest variance) and
several times on felli .
Concluding this section, we observe the local adaptation in the CMA-ES to come
along with an increasing step size that is more likely to improve than to spoil global
search performance of a local search algorithm.

8 Conclusion
For the application to real-world search problems, the evolution strategy (ES) is a rele-
vant search strategy if neither derivatives of the objective function are at hand, nor dif-
ferentiability and numerical accuracy can be assumed. If the search problem is expected
to be significantly non-separable, an ES with covariance matrix adaptation (CMA-ES),
as put forward in this paper, is preferable to any ES with only global or individual step
size adaptation.
We believe that there are principle limitations to the possibilities of self-adaptation
in evolution strategies: To reliably adapt a significant change of the mutation distri-
bution shape, at least roughly 10n function evaluations have to be done (where n is
the problem dimension). For a real-world search problem, it seems unrealistic to ex-
pect the adaptation to improve strategy behavior before, say, 30n function evaluations
(apart from global step size adaptation). A complete adaptation can take even 100n2
function evaluations (compare Section 7.3). Therefore, to get always the most out of
adaptation, CPU resources should allow roughly between 100 and 200(n + 3)2 function
evaluations.
One main reason for the general robustness of ESs is that the selection scheme
is solely based on a ranking of the population. The CMA-ES preserves this selection
related robustness because no additional selection information is used. Furthermore,
the CMA-ES preserves all invariance properties against transformations of the search
space and of the objective function value, which facilitates any simple (1, λ)-ES with
isotropic mutation distribution. Apart from initialization of object and strategy param-
eters, the CMA-ES yields an additional invariance against any linear transformation

190 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

of the search space –– in contrast to any evolution strategy to our knowledge (besides
Hansen et al. (1995) and Hansen and Ostermeier (1999)). Invariance properties are of
major importance for the evaluation of any search strategy (compare Section 6).
The step from an ES with isotropic mutation distribution to an ES facilitating
the adaptation of arbitrary normal mutation distributions, if successfully taken, can
be compared with the step from a simple deterministic gradient strategy to a quasi-
Newton method. The former follows the local gradient, which in a certain sense, also
does an ES with isotropic mutation distribution. The latter approximates the inverse
Hessian matrix in an iterative sequence without acquiring additional information on
the search space. This is exactly what the CMA does for evolution strategies.
In simulations, the CMA-ES reliably approximates the inverse Hessian matrix of
different objective functions. In addition, there are reported successful applications of
the CMA-ES to real-world search problems (Alvers, 1998; Holste, 1998; Meyer, 1998;
Lutz and Wagner, 1998a; Lutz and Wagner, 1998b; Olhofer et al., 2000; Bergener et al.,
2001; Cerveri et al., 2001; Igel and von Seelen, 2001; Igel et al., 2001). Consequently,
comparable to quasi-Newton methods, we expect this algorithm, or at least some quite
similar method, based on its superior performance to become state-of-the-art for the
application of ESs to real-world search problems.

Acknowledgments
This work was supported by the Deutsche Forschungsgemeinschaft under grant
Re 215/12-1 and the Bundesministerium für Bildung und Forschung under grant
01 IB 404 A. We gratefully thank Iván Santibáñez-Koref for many helpful discussions
and persistent support of our work. In addition, we thank Christian Igel and all re-
sponding users of the CMA-ES who gave us helpful suggestions from many different
points of view.

A CMA-ES in MATLAB
% CMA-ES for non-linear function minimization
% See also [Link]
function xmin=cmaes
% Set dimension, fitness fct, stop criteria, start values...
N=10; strfitnessfct = ’cigar’;
maxeval = 300*(N+2)ˆ2; stopfitness = 1e-10; % stop criteria
xmeanw = ones(N, 1); % object parameter start point (weighted mean)
sigma = 1.0; minsigma = 1e-15; % step size, minimal step size

% Parameter setting: selection,

lambda = 4 + floor(3*log(N)); mu = floor(lambda/2);
arweights = log((lambda+1)/2) - log(1:mu)’; % for recombination
% parameter setting: adaptation
cc = 4/(N+4); ccov = 2/(N+2ˆ0.5)ˆ2;
cs = 4/(N+4); damp = 1/cs + 1;

% Initialize dynamic strategy parameters and constants

B = eye(N); D = eye(N); BD = B*D; C = BD*transpose(BD);
pc = zeros(N,1); ps = zeros(N,1);
cw = sum(arweights)/norm(arweights);
chiN = Nˆ0.5*(1-1/(4*N)+1/(21*Nˆ2));

% Generation loop
counteval = 0; arfitness(1) = 2*abs(stopfitness)+1;
while arfitness(1) > stopfitness & counteval < maxeval

Evolutionary Computation Volume 9, Number 2 191

N. Hansen and A. Ostermeier

% Generate and evaluate lambda offspring

for k=1:lambda
% repeat the next two lines until arx(:,k) is feasible
arz(:,k) = randn(N,1);
arx(:,k) = xmeanw + sigma * (BD * arz(:,k)); % Eq.(13)
arfitness(k) = feval(strfitnessfct, arx(:,k));
counteval = counteval+1;
end

% Sort by fitness and compute weighted mean

[arfitness, arindex] = sort(arfitness); % minimization
xmeanw = arx(:,arindex(1:mu))*arweights/sum(arweights);
zmeanw = arz(:,arindex(1:mu))*arweights/sum(arweights);

% Adapt covariance matrix

pc = (1-cc)*pc + (sqrt(cc*(2-cc))*cw) * (BD*zmeanw); % Eq.(14)
C = (1-ccov)*C + ccov*pc*transpose(pc); % Eq.(15)
% adapt sigma
ps = (1-cs)*ps + (sqrt(cs*(2-cs))*cw) * (B*zmeanw); % Eq.(16)
sigma = sigma * exp((norm(ps)-chiN)/chiN/damp); % Eq.(17)

% Update B and D from C

if mod(counteval/lambda, N/10) < 1
C=triu(C)+transpose(triu(C,1)); % enforce symmetry
[B,D] = eig(C);
% limit condition of C to 1e14 + 1
if max(diag(D)) > 1e14*min(diag(D))
tmp = max(diag(D))/1e14 - min(diag(D));
C = C + tmp*eye(N); D = D + tmp*eye(N);
end
D = diag(sqrt(diag(D))); % D contains standard
deviations now
BD = B*D; % for speed up only
end % if mod

% Adjust minimal step size

if sigma*min(diag(D)) < minsigma ...
| arfitness(1) == arfitness(min(mu+1,lambda)) ...
| xmeanw == xmeanw ...
+ 0.2*sigma*BD(:,1+floor(mod(counteval/lambda,N)))
sigma = 1.4*sigma;
end
end % while, end generation loop

disp([num2str(counteval) ’: ’ num2str(arfitness(1))]);
xmin = arx(:, arindex(1)); % return best point of last generation

function f=cigar(x)
f = x(1)ˆ2 + 1e6*sum(x(2:end).ˆ2);

References
Alvers, M. (1998). Zur Anwendung von Optimierungsstrategien auf Potentialfeldmodelle. Berliner ge-
owissenschaftliche Abhandlungen, Reihe B: Geophysik. Selbstverlag Fachbereich Geowis-
senschaften, Freie Universität Berlin, Germany.

Arnold, D. (2000). Personal communication.

192 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

Bäck, T. and Schwefel, H.-P. (1993). An overview of evolutionary algorithms for parameter opti-
mization. Evolutionary Computation, 1(1):1–23.

Bäck, T., Hammel, U., and Schwefel, H.-P. (1997). Evolutionary computation: Comments on the
history and current state. IEEE Transactions on Evolutionary Computation, 1(1):3–17.

Bergener, T., Bruckhoff, C., and Igel, C. (2001). Parameter optimization for visual obstacle de-
tection using a derandomized evolution strategy. In Blanc-Talon, J. and Popesc, D., editors,
Imaging and Vision Systems: Theory, Assessment and Applications, Advances in Computation:
Theory and Practice. NOVA Science Books, Huntington, New York.

Beyer, H.-G. (1995). Toward a theory of evolution strategies: On the benefit of sex - the (µ/µ, λ)-
theory. Evolutionary Computation, 3(1):81–110.

Beyer, H.-G. (1996a). On the asymptotic behavior of multirecombinant evolution strategies. In

Voigt, H.-M. et al., editors, Proceedings of PPSN IV, Parallel Problem Solving from Nature, pages
122–133, Springer, Berlin, Germany.

Beyer, H.-G. (1996b). Toward a theory of evolution strategies: Self-adaptation. Evolutionary Com-
putation, 3(3):311–347.

Beyer, H.-G. (1998). Mutate large, but inherit small! In Eiben, A. et al., editors, Proceedings of
PPSN V, Parallel Problem Solving from Nature, pages 109–118, Springer, Berlin, Germany.

Beyer, H.-G. and Deb, K. (2000). On the desired behaviors of self-adaptive evolutionary algo-
rithms. In Schoenauer, M. et al., editors, Proceedings of PPSN VI, Parallel Problem Solving from
Nature, pages 59–68, Springer, Berlin, Germany.

Cerveri, P., Pedotti, A., and Borghese, N. (2001). Enhanced evolution strategies: A novel approach
to stereo-camera calibration. IEEE Transactions on Evolutionary Computation, in press.

EVOTECH-7 (1997). Evotech—Einsatz der Evolutionsstrategie in Wissenschaft und Technik,

7. Zwischenbericht. Interim report of the Fachgebiet Bionik und Evolutionstechnik der Tech-
nischen Universität Berlin under grant 01 IB 404 A of the Bundesminister für Bildung, Wis-
senschaft, Forschung und Technologie.

Ghozeil, A. and Fogel, D. B. (1996). A preliminary investigation into directed mutations in evo-
lutionary algorithms. In Voigt, H.-M. et al., editors, Proceedings of PPSN IV, Parallel Problem
Solving from Nature, pages 329–335, Springer, Berlin, Germany.

Hansen, N. (1998). Verallgemeinerte individuelle Schrittweitenregelung in der Evolutionsstrategie.

Eine Untersuchung zur entstochastisierten, koordinatensystemunabhängigen Adaptation der Mu-
tationsverteilung. Mensch und Buch Verlag, Berlin, Germany. ISBN 3-933346-29-0.

Hansen, N. (2000). Invariance, self-adaptation and correlated mutations in evolution strategies.

In Schoenauer, M. et al., editors, Proceedings of PPSN VI, Parallel Problem Solving from Nature,
pages 355–364, Springer, Berlin, Germany.

Hansen, N. and Ostermeier, A. (1996). Adapting arbitrary normal mutation distributions in evo-
lution strategies: The covariance matrix adaptation. In Proceedings of the 1996 IEEE Interna-
tional Conference on Evolutionary Computation, pages 312–317, IEEE Press, Piscataway, New
Jersey.

Hansen, N. and Ostermeier, A. (1997). Convergence properties of evolution strategies with the
derandomized covariance matrix adaptation: The (µ/µI , λ)-CMA-ES. In Zimmermann, H.-
J., editor, Proceedings of EUFIT’97, Fifth European Congress on Intelligent Techniques and Soft
Computing, pages 650–654, Verlag Mainz, Aachen, Germany.

Hansen, N., Ostermeier, A., and Gawelczyk, A. (1995). On the adaptation of arbitrary normal
mutation distributions in evolution strategies: The generating set adaptation. In Eshelman,
L., editor, Proceedings of the Sixth International Conference on Genetic Algorithms, pages 57–64,
Morgan Kaufmann, San Francisco, California.

Evolutionary Computation Volume 9, Number 2 193

N. Hansen and A. Ostermeier

Herdy, M. (1993). The number of offspring as strategy parameter in hierarchically organized

evolution strategies. SIGBIO Newsletter, 13(2):2–7.

Hildebrand, L., Reusch, B., and Fathi, M. (1999). Directed mutation—a new self adaptation for
evolution strategies. In Angeline, P., editor, Proceedings of the 1999 Congress on Evolutionary
Computation CEC99, pages 1550–1557, IEEE Press, Piscataway, New Jersey.

Holste, D. (1998). Modellkalibrierung am Beispiel von Kläranlagenmodellen. In Hafner, S., edi-

tor, Industrielle Anwendungen Evolutionärer Algorithmen, chapter 4, pages 37–44, Oldenbourg
Verlag, München, Germany.

Holzheuer, C. (1996). Analyse der Adaptation von Verteilungsparametern in der Evolutions-

strategie. Diploma thesis, Fachgebiet Bionik und Evolutionstechnik der Technischen Uni-
versität Berlin, Berlin, Germany.

Igel, C. and von Seelen, W. (2001). Design of a field model for early vision: A case study of
evolutionary algorithms in neuroscience. In 28th Goettingen Neurobiology Conference. In press.

Igel, C., Erlhagen, W., and Jancke, D. (2001). Optimization of neural fields models. Neurocomput-
ing, 36(1–4):225–233.

Lutz, T. and Wagner, S. (1998a). Drag reduction and shape optimization of airship bodies. Journal
of Aircraft, 35(3):345–351.

Lutz, T. and Wagner, S. (1998b). Numerical shape optimization of natural laminar flow bodies.
In Proceedings of 21st ICAS Congress, International Council of the Aeronautical Sciences and
the American Institute of Aeronautics, Paper No. ICAS-98-2,9,4.

Meyer, M. (1998). Parameteroptimierung dynamischer Systeme mit der Evolutionsstrategie.

Diploma thesis, Fachgebiet Bionik und Evolutionstechnik der Technischen Universität
Berlin, Berlin, Germany.

Olhofer, M., Arima, T., Sonoda, T., and Sendhoff, B. (2000). Optimisation of a stator blade used
in a transonic compressor cascade with evolution strategies. In Parmee, I., editor, Adaptive
Computing in Design and Manufacture (ACDM), pages 45–54. Springer Verlag, Berlin, Ger-
many.

Ostermeier, A. (1992). An evolution strategy with momentum adaptation of the random number
distribution. In Männer, R. and Manderick, B., editors, Parallel Problem Solving from Nature,
2, pages 197–206, North Holland, Amsterdam, The Netherlands.

Ostermeier, A. (1997). Schrittweitenadaptation in der Evolutionsstrategie mit einem entstochastisierten

Ansatz. Ph.D. thesis, Technische Universität Berlin, Berlin, Germany.

Ostermeier, A. and Hansen, N. (1999). An evolution strategy with coordinate system invariant
adaptation of arbitrary normal mutation distributions within the concept of mutative strat-
egy parameter control. In Banzhaf, W. et al., editors, Proceedings of the Genetic and Evolution-
ary Computation Conference, GECCO-99, pages 902–909, Morgan Kaufmann, San Francisco,
California.

Ostermeier, A., Gawelczyk, A., and Hansen, N. (1994a). A derandomized approach to self-
adaptation of evolution strategies. Evolutionary Computation, 2(4):369–380.

Ostermeier, A., Gawelczyk, A., and Hansen, N. (1994b). Step-size adaptation based on non-local
use of selection information. In Davidor, Y. et al., editors, Proceedings of PPSN IV, Parallel
Problem Solving from Nature, pages 189–198, Springer, Berlin, Germany.

Press, W., Teukolsky, S., Vetterling, W., and Flannery, B. (1992). Numerical Recipes in C: The Art of
Scientific Computing. Second Edition. Cambridge University Press, Cambridge, England.

Rechenberg, I. (1973). Evolutionsstrategie, Optimierung technischer Systeme nach Prinzipien der biolo-
gischen Evolution. Frommann-Holzboog, Stuttgart, Germany.

194 Evolutionary Computation Volume 9, Number 2

 Derandomized Self-Adaptation

Rechenberg, I. (1994). Evolutionsstrategie ’94. Frommann-Holzboog, Stuttgart, Germany.

Rechenberg, I. (1998). Personal communication.

Rudolph, G. (1992). On correlated mutations in evolution strategies. In Männer, R. and Man-

derick, B., editors, Parallel Problem Solving from Nature, 2, pages 105–114, North-Holland,
Amsterdam, The Netherlands.
Schwefel, H.-P. (1981). Numerical Optimization of Computer Models. Wiley, Chichester, England.

Schwefel, H.-P. (1995). Evolution and Optimum Seeking. Sixth-Generation Computer Technology
Series. John Wiley and Sons, New York, New York.

Whitley, D., Mathias, K., Rana, S., and Dzubera, J. (1996). Evaluating evolutionary algorithms.
Artificial Intelligence, 85:245–276.

Evolutionary Computation Volume 9, Number 2 195