July 6, 2013
9:41
toneEvo IJBC rev1
International Journal of Bifurcation and Chaos
c World Scientific Publishing Company
Chaos and Music: from time series analysis to evolutionary
composition
Maximos A. Kaliakatsos-Papakostas
University of Patras, Department of Mathematics
Patras, GR-26110, Greece
maxk@math.upatras.gr
Michael G. Epitropakis
University of Stirling, School of Natural Sciences
Stirling FK9 4LA, Scotland UK
mge@cs.stir.ac.uk
Andreas Floros
Ionian University, Department of Audio and Visual Arts
Corfu, GR-49100, Greece
floros@ionio.gr
Michael N. Vrahatis
University of Patras, Department of Mathematics
Patras, GR-26110, Greece
vrahatis@math.upatras.gr
Received (to be inserted by publisher)
Music is an amalgam of logic and emotion, order and dissonance, along with many combinations
of contradicting notions which allude to deterministic chaos. Therefore, it comes as no surprise
that several research works have examined the utilization of dynamical systems for symbolic
music composition. The main motivation of the paper at hand is the analysis of the tonal composition potentialities of several discrete dynamical systems, in comparison to genuine human
compositions. Therefore, a set of human musical compositions is utilized to provide “compositional guidelines” to several dynamical systems, the parameters of which are properly adjusted
through evolutionary computation. This procedure exposes the extent to which a system is capable of composing tonal sequences that resemble human composition. In parallel, a time series
analysis on the genuine compositions is performed, which firstly provides an overview of their
dynamical characteristics and secondly, allows a comparative analysis with the dynamics of the
artificial compositions. The results expose the tonal composition capabilities of the examined
iterative maps, providing specific references to the tonal characteristics that they can capture.
Keywords: chaos and music, tonal time series, evolutionary composition
1
July 6, 2013
2
9:41
toneEvo IJBC rev1
Maximos A. Kaliakatsos-Papakostas et al.
1. Introduction
Music composition is a process that encompasses a combination of two contradicting forces: the determinism
imposed by music rules and the chaos that is subsumed in human creativity. Therefore, the utilization of
chaotic system for automatic music composition seems like a direction worthy of extensive exploration. In
fact, several research works have focused on the potential of using dynamical systems for music composition,
however, the compositional potentiality of dynamical systems remains relatively unexplored. Although
many methodologies have been proposed, the results that most works have presented do not provide
comparative analyses between the dynamical systems’ compositions and the ones composed by humans. As
a result, one may not come up with practical implications as to whether a human–like “chaotic composer”
can be created. If so, what would its dynamical and musical characteristics be? The aims of this paper
may be summarized as the examination of symbolic music compositional capabilities of several dynamical
systems, focusing on the tonal aspects of music.
Nonlinear characteristics are deeply related to the human perception of sound. Their existence in
natural sound has been studied for a large variety of phenomena, such as animal vocalization [Tokuda
et al., 2002] and birdsong production models [Amador & Mindlin, 2008]. Furthermore, the relation of
nonlinear dynamics with human speech has been investigated [Behrman, 1999], providing also significant
results in speech recognition [Jafari et al., 2010]. Additionally, theories that incorporate nonlinearities in
the mechanism though which humans perceive auditory events have provided great insights about sound
and human perception [Chialvo, 2003]. Consequently, it comes without saying that the ultimate human
cultural manifestation of sound – music – abounds in nonlinear structures on many levels: from the level of
instrumental timbre [Fletcher, 1994] and overall orchestration timbres [Voss & Clarke, 1978], to the level
of symbolic composition [Manaris et al., 2005].
Automatic music composition through dynamical systems has been a very popular technique which
inspired many researchers and artists, from the pioneering works of Pressing [Pressing, 1988], Bidlack [Bidlack, 1992], Herman [Hrman, 1993] and Harley [Harley, 1994], to more contemporary works. An interesting
approach for altering the tonal characteristics of a piece was proposed in [Dabby, 1996], where the pitch
space of the specific piece was “convolved” with a chaotic solution of a dynamical system. Thereafter, new
neighboring solutions provided novel tonal sequences that gradually diverged from the ones of the initial
piece. The mainstream compositional strategy however, is the utilization of well–known dynamical systems
like the Chua’s circuit [Chua et al., 1993] for the generation of sound [Choi, 1994] or music composition
[Bilotta et al., 2007; Rizzuti et al., 2009]. The latter works not only examine the composition of music with
the Chua’s circuit, but also evolve its parameters with genetic algorithms. The fitness of the evolutionary
process is provided by the pitch interval distribution of a target piece. Several dynamical systems have
been examined in [Coca et al., 2010] and the produced music was characterized according to several music
characteristics. A similar approach was followed in [Kaliakatsos-Papakostas et al., 2012b], but the musical
output in this case was examined in terms of its information complexity.
The paper at hand is motivated by the potentialities of dynamical systems for “symbolic” composition
of musical tonal sequences, reflected in numerous works that employ such systems for automatic composition. Since the focus is on the tonal aspect of music, the automatic composition methodology that is
followed disregards information on rhythm, intensities and timbre. To this end, two types of analysis are
applied, which incorporate the extraction of several tonal characteristics from genuine music masterpieces
and the composition of novel tonal sequences using chaotic systems and evolution. On the one hand, the
genuine compositions are examined in terms of their dynamical characteristics through a time series analysis approach which incorporates phase space reconstruction of the pieces’ attractors, the extraction of the
largest Lyapunov exponent and fractal dimension. A similar approach was presented in [Boon & Decroly,
1995], but the present paper provides an exhaustive experimental research on a large set of compositions,
combined with a comparative analysis on the time series characteristics between genuine and artificial compositions. On the other hand, several well–known dynamical systems are utilized to compose novel tonal
sequences, which are “trained” to share similar characteristics with the aforementioned pieces. In the latter
analysis, the parameters of these dynamical systems are optimized through evolutionary computation, so
that the generated novel pieces share similar characteristics to the respective “target” genuine ones.
July 6, 2013
9:41
toneEvo IJBC rev1
Chaos and Music: from time series analysis to evolutionary composition
3
These tonal characteristics are exhibited to highlight the exceptional characteristics of different music
styles and therefore constitute a qualitative measurement of the potentialities of each dynamical system
concerning music composition. Through an extensive experimental study, the tonal sequences produced by
some of the examined discrete dynamical systems are observed to encompass some tonal characteristics
that are present in genuine human compositions. However, through the phase space reconstruction that
is realized with the time series analysis of the genuine pieces, it is indicated that dynamical systems with
higher complexity and with a greater number of dimensions are required to compose human–like music.
These results are assessed through a comparative scrutinization between the genuine and the artificial
compositions, not only on the level of musical characteristics, but also on the level of their dynamical
behavior expressed with the largest Lyapunov exponent and fractal dimension.
The rest of the paper is organized as follows. Section 2 discusses the extraction of tonal time series
from genuine symbolic compositions in MIDI format and presents the tools for the time series analysis,
which allows the phase space reconstruction and the computation of the largest Lyapunov exponent and
fractal dimension. The utilized automatic music composition methodology is reviewed in detail in Section 3. Specifically, the evolutionary scheme that allows the tuning of the dynamical systems’s parameters
is presented and the music features that constitute the basis for the fitness function are throughly analyzed.
Section 3 also discusses the automatic tonal composition methodology which encompasses coarse compositional guidelines to the dynamical systems. A thorough experimental report is given in Section 4, which
incorporates the most important findings that were extracted by the application of the methodologies described in the preceding sections. The paper concludes in Section 5 with a summary and some pointers to
future work.
2. Dynamical properties of the genuine compositions through time series analysis
As mentioned previously, this work is targeted towards studying the tonal composition potentialities of
discrete dynamical systems, in the context of automatic music composition. Therefore, the output produced
in consecutive iterations of these dynamical systems, is mapped to tonal sequences through a straightforward procedure that is described below in detail. An initial approach to this subject is the study of tonal
sequences that derive from “genuine” music masterpieces, in terms of their dynamical characteristics. In
order to consider a wide spectrum of Western music, this study incorporates several genuine compositions
from J. S. Bach, Mozart, Beethoven and a collection of jazz standards composed by various artists. For the
rest of the paper, the entire jazz collection will be referred to as if it were composed by a single composer
for simplicity. The fact that all jazz compositions are considered per composer is not expected to affect
the accuracy of results, because the jazz compositional technique varies significantly from that of classical
music. Therefore, their inclusion in a single category is intended to examine the differences in dynamical
properties among the compositional styles of the collected music sets. The extraction of dynamical characteristics from the tonal sequences that derive from all the aforementioned compositions allows a first
cartographical overview of the underlying dynamical’s that are considered to produce these compositions.
Furthermore, this analysis allows a comparison in the dynamical characteristics of the genuine and the
artificial compositions.
The composition process that is assumed throughout the paper regards the symbolic level, i.e. only
information that is interpretable in a score is considered, without any timbre-related information. Therefore,
the aforementioned genuine compositions are collected in MIDI format, which is a protocol that includes the
most viable symbolic information of music content, allowing a quite accurate interpretation to music score.
Additionally, in order to have comparable results among the time series extracted from the genuine pieces,
only piano executions were considered. Specifically, the dataset comprises piano sonatas from Beethoven and
Mozart, the “Well Tempered Clavier” of Bach and piano transcriptions of several well-known jazz standards.
This fact “neutralizes” the effect of instrument–imposed constraints, like tonal range and polyphony, that
may affect the format of the tonal sequences. To this end, the tonal sequence of a piece is obtained by
the serial concatenation of all its tones in the order they appear, disregarding inter–onset distance and
duration. In the case of polyphonic events, i.e. music events that incorporate multiple simultaneous tones,
the order of tones is considered from the lowest to the highest tone of the polyphonic cluster. Although
July 6, 2013
4
9:41
toneEvo IJBC rev1
Maximos A. Kaliakatsos-Papakostas et al.
this approach might be considered simplistic, it reflects the qualitative demeanor of this study1 .
2.1. Phase space reconstruction
Based on Takens’s [Takens, 1981] theorem, if a time series, {xi }i∈{0,1,2,...,n} , derives from a dynamical system, it encompasses all the viable characteristics needed to construct a topologically equivalent dynamical
system. Thereby, the dynamical behavior of the initial system can be “reconstructed” in a topologically
consistent manner, even if the system’s dimensionality is higher than the observed time series’. This procedure is called “phase space reconstruction” and is realized by estimating the “novelty” introduced in
segments of the time series (to compute a proper time delay) and then computing the dimensionality of the
dynamical system’s space. The phase space of the dynamical system is then reconstructed by embedding
replicates of the time series, shifted by a proper index expressed by an optimal time lag, in each dimension
of the reproduced system. The resulting dynamical system is considered to produce the following vector
sequence, y(i) = (xi , xi+τ , xi+2τ , . . . , xi+(µ−1)τ ). A brief review of this methodology is discussed in the
following paragraphs, while for a thorougher review the reader is referred to [Lai & Ye, 2003; Kodba et al.,
2005]. In the paper at hand, the software implementation presented in [Kugiumtzis & Tsimpiris, 2010] was
utilized. The solicitation of further dynamical information is realized through two more measures, namely
the Largest Lyapunov Exponent (LLE) and the Fractal Dimension(FD) as described later. The latter two
dynamical measures are not necessary for the phase space reconstruction, they are rather used as measures
of comparison between the genuine and the automatically produced compositions.
2.1.1. Embedding delay
The rationale behind determining a proper time delay is the allocation of a time series segmentation
to equal segments of length τ , where each segment incorporates information about the other. Such a
segmentation exposes the interrelations that the underlying dynamical system imposes to each segment
of the time series. The time delay (or segment length) value, τ , should be chosen so that the variables
xi , xi+τ , xi+2τ , . . . , xi+(q−1)τ , where q = ⌊l/τ ⌋ and l is the length of the time series, are as independent as
possible. To this end, Fraser and Swinney [Fraser & Swinney, 1986] proposed the examination of the mutual
information between all segments of the time series, for various τ values. The computation of mutual information is performed with the MATS [Kugiumtzis & Tsimpiris, 2010] toolbox in MATLAB, in which a partitioning is considered that produces a grid in the time series domain interval [min{xi }, max{xi }]i∈{0,1,2,...,n} ,
with k partitions of length k/(max{xi } − min{xi })i∈{0,1,2,...,n} . Given the aforementioned domain partitioning, the mutual information for a time delay τ is given by [Kodba et al., 2005]
I(τ ) = −
k
k
Pn,m (τ ) (l − τ )
1
,
Pn,m (τ ) log
l − τ n=1 m=1
Pn Pm
(1)
where l is the length of the time series, Pn and Pm is the probability that a time series value is in the
n-th and m-th grid position respectively and Pn,m (τ ) is the joint probability that xi is in the n-th grid
position and xi+τ is in the m-th grid position. The first minimizer of I(τ ) constitutes the proper time delay
(segmentation length), to obtain the dynamical behavior as has hitherto been discussed.
2.1.2. Embedding dimension
The computation of the embedding dimensions, i.e. the number of dimensions in which the attractor of
the assumed dynamical system lies, is vitally important for the phase space reconstruction. Thereby, the
attractor of the underlying dynamical system is assumed to fold and unfold smoothly when the embedding
changes from an integer value to the next one, without sudden irregularities in its structure. A popular
1
Several other considerations of polyphonic events were considered, like from lowest to highest tone, selecting only the highest
tone, and random selection of tone order in polyphonic events, and the results were similar, and therefore omitted in the
present study.
July 6, 2013
9:41
toneEvo IJBC rev1
Chaos and Music: from time series analysis to evolutionary composition
5
method for estimating a proper embedding dimension for the dynamical system’s attractor, denoted by
µ, is the false nearest neighbor [Kennel et al., 1992] method. With this method, the unfolding smoothness
of the attractor is examined in spaces with increasing dimension (µ values). Thus, for each reconstructed
point in the examined embedding dimensions, µ, it is investigated whether its nearest neighbor remains
below a distance threshold when the embedding dimension is increased to µ + 1. If this does not hold, then
these points are called false neighbors. The embedding dimension that is chosen as appropriate, is the one
that incorporates a small percentage of false neighbors when the dimension is increased. In the context
of the presented work, the satisfactory percentage of false nearest neighbors was set to zero. Therefore,
the proper embedding dimension was chosen to be the smallest one that incorporates no false nearest
neighbors.
2.2. Largest Lyapunov exponent
The sensitivity of a dynamical system to initial conditions is quantified by the Lyapunov exponents, which
roughly discuss the expansion or contraction rate of trajectories with initial conditions over a small µdimensional sphere. While the system’s iterations progress, the sphere evolves into an ellipsoid the principal directions of which expand, contract or remain unchanged, as measured by positive, negative or zero
Lyapunov exponents. In directions which incorporate negative Lyapunov exponents, the attractor’s projection is a fixed point, while zero Lyapunov exponents denote directions of limit circles. Positive Lyapunov
exponents signify the existence of chaos. The set of Lyapunov exponents to all directions is called the
Lyapunov spectrum and it is denoted by (λ1 , λ2 , . . . , λµ ), but random initial condition vectors are expected
to converge to the most unstable manifold (with probability 1 [Rosenstein et al., 1993]). Therefore, the
computation of the largest Lyapunov exponent (LLE), λ1 , is sufficient to indicate the existence of chaos in
the examined system.
Although there are numerical methods for computing the entire Lyapunov spectrum [Wolf et al.,
1985] from time series, the acquisition of the largest exponent is needed to define the attractor’s chaotic
potentiality. A well–known methodology for evaluating the LLE of the reconstructed attractor from a time
series, is due to Rosenstein [Rosenstein et al., 1993] and is the numerical method utilized in the paper at
hand. A prerequisite for Rosenstein’s method is to reconstruct the attractor’s phase space based on the
time series observation, thus obtaining the time delay (τ ) and embedding dimension (µ) as described in
Section 2.1. After the reconstruction, the nearest neighbor (yı̂ ) of each point (yi ) in the trajectory is located
and their distance in this initial iteration is computed as
di (0) = min yi − yı̂ 2 .
yı̂
(2)
The LLE is then estimated as the mean rate of nearest neighbor separation throughout their successive
iterations. By definition, therefore, it holds that
di (t) = di (0) eλ1 (t) ,
(3)
where λ1 (t) is the LLE estimation at time step t. Taking the logarithm of both sides in the equation, we
have
ln(di (t)) = ln(di (0)) + λ1 (t).
(4)
The average LLE is finally approximated by the gradient of the linear regression among all neighboring
pairs
b(t) = ln(di (t))i ,
(5)
where ·i denotes the mean value over all possible i indexes.
2.3. Fractal dimension
Fractal dimension (FD) is a concept that describes the complexity of geometric objects as a ratio of change
in detail to the change in scale [Mandelbrot, 1982]. Unlike the topological dimension, FD may have non
July 6, 2013
6
9:41
toneEvo IJBC rev1
Maximos A. Kaliakatsos-Papakostas et al.
integer values and a popular method for its computation is the well–known box counting method [Alevizos &
Vrahatis, 2010]. For the computation of the FD from time series data however, the box counting method may
not be straightforwardly applicable. Several methods have been proposed for the FD computation of time
series, among which are the correlation dimension method [Grassberger & Procaccia, 1983], the methods of
Kantz [Kantz, 1988], Higuchi [Higuchi, 1988], Petrosian [Petrosian, 1995] and Sevcik [Sevcik, 1998]. Many
works have compared the above methodologies either by their effectiveness on experimental data (for
example for seizure detection in EEG), and/or on functions with theoretically computable FD values (like
the Weierstrass cosine functions [Tricot, 1994]) [Esteller et al., 2001; Goh et al., 2005; Raghavendra & Dutt,
2010; Ahmadi & Amirfattahi, 2010; Polychronaki et al., 2010]. The results yielded by these works provide
uncertain and contradicting results about the suitability of each method for different tasks, reflecting the
fact that the computation of the FD of time series is case dependent.
The method used for the results presented in this paper is the method of Sevcik [Sevcik, 1998], since
it is among the most recently developed, most easily implementable and faster methods. This method is
based on the Hausdorff dimension, which is computed for n–dimensional objects as
− log(N (ε))
,
ε→0
log(ε)
DH = lim
(6)
where N (ε) is the number of n–dimensional open balls with radius ε needed to completely cover the set
under examination. If a length L is assumed for the set that comprises the curve under examination, then
this curve can be covered by at least N (ε) = ⌈L/(2ε)⌉ balls of radius ε. Sevcik proposed the consideration
of the timeseries as a two dimensional object, laying on the plane defined by time (x–axis) and the timeseries’ values (y–axis). Therefore, the balls that are utilized for the examination of the curve coverage, are
considered as two dimensional balls. With these fact under consideration, the computation of DH may be
written as [Raghavendra & Dutt, 2010]
log(L) − log 2
.
ε→0
log (ε)
DH = 1 − lim
(7)
The method proposed by Sevcik utilizes the latter expression of the Hausdorff dimension to derive the counterpart computation for time series. Thereby, if a single dimensional time series is considered, a normalization of both the time series values (yi ) and indexes (xi ) is considered, with yi∗ = (yi − ymin )/(ymax − ymin )
and x∗i = (xi − xmin )/(xmax − xmin ) respectively, where ymin , ymax , xmin and xmax are the minimum and
maximum values and indexes respectively. By considering a time series with N observations, an N × N
grid is constructed on the the normalized xy time series plane, and the fractal dimension is approximated
as [Raghavendra & Dutt, 2010]
Ds = 1 +
log(L) − log (2)
,
log (2(N − 1))
(8)
where L is the length of the normalized time series.
3. Automatic composition of tonal sequences with specified characteristics
Automatic symbolic music composition encompasses a great variety of methodologies that produce novel
music content in the form of tones, onsets, note durations and timbre among others. The specific subdomain
that the paper at hand discusses could be characterized as “supervised” algorithmic composition, in a
sense that the compositional algorithm is “forged” to compose music which complies with certain musical
characteristics. Under this perspective, the composition process is integrated with a supervised training
process, in which the parameters of the automatic music generation algorithm are properly adjusted, to
allow the composition of music that is circumscribed within a musical area of predefined characteristics. In
the context of the presented work, the algorithm that generates notes comprises an iterative map among the
ones demonstrated in Table 1 and a typical iteration–to–tones interpretation methodology that is described
in Section 3.3.
The tunable parameters of the automatic composition algorithm are the parameters of the maps, within
the ranges demonstrated in the third column of Table 1. It has to be noticed that the initial conditions of
July 6, 2013
9:41
toneEvo IJBC rev1
Chaos and Music: from time series analysis to evolutionary composition
7
a map is also considered as a parameter in the present work. Parameter values within the selected ranges,
allow the respective maps to expose their dynamic potentialities, from fixed points and periodic orbits
to chaos. These parameters are properly adjusted so that the generated tones comply with some musical
criteria that are defined by a set of music features. The adjustment process is realized with the Differential
Evolution (DE) algorithm [Storn & Price, 1997; Price et al., 2005], which evolves populations of parameter
values towards values that are better fitted to the problem at hand, i.e. they allow the respective iterative
systems to compose music that complies with the target music features. The target music features are
extracted from each piece in the set of available genuine pieces. Under the guidance provided by a target
genuine piece in the form of music features, each map generates a new composition which is expected to
have similar features to the ones provided by the target piece. The rhythmic and orchestration part of
music are not considered and therefore, the notes produced by the dynamical systems are “attached” to
the respective rhythmic values of the target pieces.
Table 1.
logistic
tent
circle
Henon
general
quadratic
Ikeda
The iterative maps used as tone generators, their parameters and the parameters’ considered ranges.
xn+1 =
4rxn (1 − xn )
2rxn , if xn ∈ [0, 12 )
xn+1 =
2r(1 − xn ), if xn ∈ [ 21 , 1]
K
xn+1 = (xn + Ω − 2π
sin(2πxn ))mod1
2
xn+1 = α − xn + βyn
yn+1 = xn
2
xn+1 = α0 + α1 xn + α2 x2n + α3 + α4 yn + α5 yn
2
yn+1 = α6 + α7 yn + α8 yn + α9 + α10 xn + α11 x2n
xn+1 = α + β (xn cos(θ) + yn sin(θ))
, θ = c − x2 +yd 2 +1
n
n
yn+1 = β (xn sin(θ) − yn cos(θ))
r ∈ [0, 1], x0 ∈ [0, 1]
r ∈ [0, 1], x0 ∈ [0, 1]
(Ω, K) ∈ [0, 10]2 , x0 ∈ [0, 1]
α ∈ [0, 2], β ∈ [−1, 1], (x0 , y0 ) ∈
[−1, 1]2
(α0 , α1 , . . . , α11 )
∈
[− 23 , 32 ]12 ,
2
(x0 , y0 ) ∈ [−1, 1]
α ∈ [0, 10], β ∈ [ 12 , 1], c ∈ [2, 5],
d ∈ [35, 50], (x0 , y0 ) ∈ [−1, 1]2
3.1. Evolving dynamical systems to composers
The appropriateness of a map’s parameters, considering the established evolutionary nomenclature, is
called “fitness” and it is a value that describes the “distance” between the features of the target and the
artificial composition. A proper distance measure should consider all features equally important, to drive
the evolution of new dynamical system parameters towards ones that compose better music, according to
every feature impartially. Since the range of the features is not known a priory, or it could be any value
in (−∞, ∞), a straightforward normalization of the feature values in [−1, 1] is not possible. To this end,
the mean relative distance, denoted dMRD measure could provide a good impartial approximation of the
feature differences, since it offers an estimation about the percentage of the difference between the features
of the genuine and the artificial pieces. Considering two vectors of features, fg and fa , their mean relative
distance is computed as [Kaliakatsos-Papakostas et al., 2013]
dMRD =
k
1 |fg (i) − fa (i)|
,
g (i), fa (i)})
k
max({
f
i=1
(9)
where k is the length of the feature vectors and denotes the number of features, the index i refers to the
i–th vector element, or the i–th feature and max(a, b) returns the maximum among the numbers a and b.
A detailed description of the music features that comprise the music vector is provided in Section 3.2.
Each iterative map listed in Table 1 is employed as a music composer that generates tonal sequences
with a methodology described in Section 3.3. The characteristics of the music it composes rely on its
parameters’ values. Therefore, a method that extensively searches for proper parameter values is required,
in order to explore the map’s compositional potentialities comprehensively. As mentioned earlier, this
method is the DE algorithm. With this method, a population of parameter values (called individuals) is
randomly initialized, and the system is allowed to compose music with each one of them. Afterwards, these
parameters are altered with the application of some evolutionary operators, which create new individuals
July 6, 2013
8
9:41
toneEvo IJBC rev1
Maximos A. Kaliakatsos-Papakostas et al.
(i.e. sets of parameter values). If the new individual produces “better music” than its ancestor, then it is
selected to belong to the new generation of individuals, else its ancestor passes on to the new generation.
This procedure continuous iteratively, creating populations of parameter values with better fitness, i.e.
they produce better music. For the presented results, a population size of 50 individuals was utilized and
evolved for 50 generations. The range of the parameter values throughout all iterations, were forced to
remain within the limits demarcated in the third column of Table 1 respectively for each map.
3.2. Tonal features in short music segments
When considering automatic music composition towards the direction provided by some music features,
these features should encompass “fundamental” music attributes, which describe as accurately as possible
the desired music output. Furthermore, features that describe musical characteristics accurately should
yield clearer comparison results, allowing safer determination about which map exhibits better compositional capabilities. The features that have been selected for the reported results are shown in Table 2.
Since the purview of this paper is the tonal musical domain (not rhythm or orchestration), the features
are focused on tonal aspects, which concern statistics about the transition of notes. These features do not
incorporate statistics about the tonal constitution of pieces, i.e. statistics that reflect the key of composition, since such features would not contribute to the assessment of the automatic composition system’s
adaptability.
Table 2.
The considered tonal features.
name
tonal range
tonal gradient
tonal jumps mean
tonal jumps standard deviation
acronym
range
grad
pdM
description
difference between maximum and minimum pitch values
gradient of the interpolating line through pitch values
mean value of consecutive pitch value differences
pdS
standard deviation of pitch value differences
ascending profile
asc
descending profile
desc
constant profile
const
ratio of ascending intervals over the total number of intervals
ratio of descending intervals over the total number of
intervals
ratio of constant intervals over the total number of intervals
To obtain an insight about the musical descriptiveness of the seven aforementioned music features, a
two–sided Wilcoxon [Wilcoxon, 1945] rank sum test is applied on the features extracted by short segmentations of each piece for each composer. Through this test, the statistical significance of the difference in each
feature’s distribution for each composer is examined. Specifically, for each pair of composers we employ
the rank sum test to each respective feature pair, to obtain the probability that these two features belong
to continuous distributions with equal medians, a fact that would not make them descriptive. Formally,
the null hypothesis for each feature pair is that they are independent samples from identical continuous
distributions with equal medians. If the null hypothesis is rejected at the 5% significance level for a pair
of composers and a pair of features, then these features are indicated to encompass significant different
statistical information about each composer.
The results of this test are demonstrated in Table 3. Each cell in this table concerns the results about
each composer pair. The sequence of binary digits within each cell denotes the rejection of the null hypothesis with 1 and contrarily with 0, for the respective feature. For example, for the Bach–Mozart pair, the null
hypothesis is rejected to all but the third and sixth feature. Since Table 3 exhibits a statistical significance
in the distributions of most pairs of respective features for each composer, it is clearly indicated that these
features encompass viable musical information. Therefore, within the extent of the presented statistical
analysis, the results that are reported in Section 4 could be trusted as indicative of the compositional
capabilities of each dynamical system.
July 6, 2013
9:41
toneEvo IJBC rev1
Chaos and Music: from time series analysis to evolutionary composition
9
Table 3. Statistical significance in the difference of each
features’ distributions among the pieces of each composer.
Each quadruple of 1 and 0 digits denotes the significance
or not respectively, of distribution differences between
composer pairs for the respective feature. For example,
the sequence 1101111 for the Mozart–Beethoven pair denotes that the distributions all all features except the
third are significantly different.
Bach
Mozart
Beethoven
Bach
—
—
—
Mozart
1101101
—
—
Beethoven
1111111
1101111
—
Jazz
1111111
1111111
1111111
3.3. Automatic tonal composition methodology
Since the main subject of examination is the tonal compositional potentialities of iterative dynamical
systems, the automatic composition process is directed solely towards tonal composition. The rhythmic
elements are considered to be borrowed by the target piece, thus creating a composition with identical
rhythm. The specific scope of the presented tonal composition methodology, assumes the existence of a
target genuine music piece which provides the tonal compositional guidelines for the generative system. It
is important that these guidelines encompass the necessary amount of information, which may be assumed
by general admissions. For instance, although it is a tempting conjecture, one may hardly presume that
a dynamical system would form musical structures that resemble human compositions in every hierarchical level, e.g. key structure, chord progressions or note motifs. The fact that musical knowledge is not
directly inherited to dynamical systems has been addressed by several means, like the imposition of key
constraints [Coca et al., 2010; Kaliakatsos-Papakostas et al., 2012b] (i.e. eligible notes on certain key) or
tonal restrictions imposed by a target piece itself [Dabby, 1996].
A tone generation approach similar to the one presented here has been utilized for the generation of
tonal sequences, in the context of an automatic intelligent music improviser [Kaliakatsos-Papakostas et al.,
2012a]. The compositional methodology used in the current work is adjusted to receive tonal guidelines
from the target piece, while the above cited system receives guidelines from a human improviser. The target
piece provides guidelines that encompass music knowledge at an elementary level, within short segments
throughout its duration. These guidelines incorporate tonal range, chord information and pitch class profile
(PCP) complexity expressed by the Shannon information entropy [Shannon, 2001] (SIE). These information
features characterize a list of allowed notes, which are eligible to be “played” by the dynamical system
through a straightforward mapping procedure. An overview of the system, as described in this paragraph,
is illustrated in Figure 1, while a detailed analysis is provided in the next two paragraphs.
The target piece provides some intrinsic musical guidelines to the dynamical system, which characterize
the target piece’s music content in short time segments, through the formulation of a note list for each
segment. The note list comprises notes within the range of each respective segment, which also belong to
certain pitch class values. Initially, the target genuine piece is segmented in short intervals, according to
its tempo, and the chord of each segment is recognized with a typical template based technique [Oudre
et al., 2010], considering several chord templates with up to five voices. Since the available pieces are in
MIDI format, a segmentation in certain measure subdivisions is possible. To this end, a segment length
of 2 beats was considered the most appropriate compromise between selecting a too small, or a too large
segmentation length. The main concern that the discussed methodology faces is the detection of chords
within each segment. Therefore, too small a segment may not provide a sufficient amount of notes for
successful chord recognition. Contrarily, larger segments would most likely incorporate more than one
chords, leading, in the best case, to the disregard of all except one chord in the segment.
At first, the note list of each segment includes the notes that pertain in the recognized chord, but additional notes may be required to capture the tonal constitution of a segment. No matter what segmentation
length would be chosen, it would still be possible to find a segment which incorporates more than one
chords, or a segment with no chord (e.g. chromatic phrase). To this end, the supplementary information
July 6, 2013
10
9:41
toneEvo IJBC rev1
Maximos A. Kaliakatsos-Papakostas et al.
analyze
MIDI target genuine piece with
N note events
compose
run dynamical system for N
iterations
segmentation
for each segment
for each segment
identify main
chord
estimate PCP
SIE
map iterations to notes
based on note list
assign notes to the
genuine's onsets
construct note list
Fig. 1.
Flow diagram of the automatic composition process.
provided by the SIE of the PCP is also considered. In each segment, a value is obtained from the SIE of the
PCP, which indicates the information complexity of tones in this segment. If this complexity is “reachable”
by utilizing only notes of the recognized chord, then the notes that comprise the chord are sufficient and no
additional notes are needed to form the final note list of the segment. On the contrary, if the SIE threshold
of the recognized chord is exceeded, the most prominent notes of the PCP are added incrementally to
the final note list, except from the ones that are already added from the recognized chord, until the SIE
threshold is covered. The final list of available notes comprises notes within the range dictated by the
respective segment and belong to the recognized chord, with possible additional notes imposed by the SIE
complexity threshold. After the note list has been constructed, the iterations of the dynamical system are
normalized within [0, 1] and are then linearly mapped to the indexes of each segment’s note list. The tonal
sequences that are produced, are adhered to the rhythmic note events of the genuine composition.
4. Results
The results focus on three inquiries. At first, the dynamical properties of the genuine compositions are
assessed, providing information about the attractor characteristics of each composer’s tonal sequences.
Secondly, results are reported on the adaptivity of each dynamical system to the tonal characteristics
that are imposed by each composer. This analysis incorporates a thorough examination of the musical
characteristics that each system is able to reproduce. Finally, a comparative analysis is performed on the
dynamical characteristics of the genuine and the artificial compositions.
4.1. Time series analysis on the genuine target compositions
This section presents the findings yielded by the application of the methods presented in Section 2. Specifically, the note sequence of each genuine piece is considered to form a time series, by which the information
of the attractor reconstruction are assessed, namely the time delay (τ ) and embedding dimension (µ),
together with the Largest Lyapunov Exponent (LLE) and the Fractal Dimension (FD). This analysis is
supplementary to the main perspective of the paper, which is the examination of the compositional capabilities of several dynamical systems. Therefore, this analysis firstly aims to provide some descriptive
statistical information about the attractor characteristics for different composers through the τ , µ, LLE
and FD values assessed by every piece. Secondly, the LLE and FD values offer the opportunity to perform
a comparative analysis, presented in Section 4.3, between the respective values extracted from the pieces
that were artificially produced by the dynamical systems.
The attractor characteristics are shown in the box plots in Figure 2. Although the values presented
July 6, 2013
9:41
toneEvo IJBC rev1
Chaos and Music: from time series analysis to evolutionary composition
20
100
18
90
80
16
70
14
60
12
µ
τ
11
50
10
40
8
30
6
20
4
Bach
Mozart
Beethoven
composer
Bach
Jazz
(a) τ
Mozart
Beethoven
composer
Jazz
(b) µ
1.7
1
1.68
fractal dim.
Lyapunov
0.8
0.6
0.4
1.66
1.64
1.62
1.6
1.58
1.56
0.2
1.54
0
1.52
Bach
Mozart
Beethoven
composer
Jazz
(c) LLE
Fig. 2.
Bach
Mozart
Beethoven
composer
Jazz
(d) FD
Estimated τ , µ, LLE and FD values for the tonal time series of each piece in the dataset.
therein constitute numerical approximations, it is clearly observed that there are differences of these characteristics between pairs of composers. To examine which time series characteristics are significantly different
between which pairs of composers, we perform a two–sided Wilcoxon [Wilcoxon, 1945] rank sum test,
in a similar fashion as with the music features in Section 3.2. For the currently examined case, the null
hypothesis is that the observations deriving from a time series’ values from a composer pair, are samples
from identical continuous distributions with equal medians. The results for these tests are demonstrated in
Table 4, while it is reminded that the examined pieces are all piano compositions or piano transcriptions
(in the case of the jazz standards). Therein, each four–tuple of digits denotes whether the null hypothesis
is rejected for the distributions of the respective time series values, for the respective composer of each row
and column. Specifically, the digit 1 shows the rejection of the null hypothesis on the significance level of
5% (thus the results are statistically significant), while the digit 0 the opposite. The display order of each
digit corresponds to the distributions of τ , µ, LLE and FD values. It should be noted that for every pair of
composers, there is at least one pair of distributions that is significantly different, while the opposite holds
for the LLE distributions.
4.2. Adaptivity of each dynamical system
The inquiries discussed in this paragraph incorporate the adaptivity of each dynamical system to the
specified composition tasks. The adaptivity is measured by the fitness value of the best individual yielded
by the evolutionary process as described in Section 3.1. Table 5 presents the mean value and the standard
deviation of the fitness values provided by the best individual in each composition task. It is reminded that
each individual represents a set of parameter values (including initial iteration) for each map, as displayed
in the third column of Table 1, and that a composition task is the composition of novel tonal content that
encompasses the tonal characteristic of a target piece. From Table 5 one may first notice that the worst
July 6, 2013
12
9:41
toneEvo IJBC rev1
Maximos A. Kaliakatsos-Papakostas et al.
Table 4. Statistical significance of differences between τ , µ, LLE and FD distributions among the
pieces of each composer. Each quadruple of 1 and
0 digits denotes the rejection or not respectively, of
the null hypothesis for the distribution of τ , µ, LLE
and FD respectively between composer pairs. For example, the sequence 1101 for the Bach–Mozart pair
denotes that the differences in τ , µ and FD distributions are statistically significant, while the LLE are
not.
Bach
Mozart
Beethoven
Bach
—
—
—
Mozart
1101
—
—
Beethoven
1100
0001
—
Jazz
0101
1101
1101
performance for all composers is provided by the tent map, while the general quadratic map has produced
individuals with the best mean fitness, for the compositions of all composers except from Beethoven. The
best mean performance for Beethoven was achieved by the circle map. Nevertheless, the mean performance
of the general quadratic, the circle and the Henon maps among all composers are comparable. A thorougher
quantification of the significance of this similarity is performed below.
Table 5. Mean and standard deviation (in parentheses) of the best individuals’ fitness values for all composers and maps. The best mean fitness value for each composer is demonstrated in boldface typesetting.
Bach
Mozart
Beethoven
Jazz
logistic
0.237 (0.017)
0.279 (0.021)
0.261 (0.021)
0.276 (0.015)
tent
0.268 (0.017)
0.306 (0.015)
0.277 (0.020)
0.295 (0.015)
circle
0.213 (0.023)
0.257 (0.022)
0.255 (0.016)
0.264 (0.020)
Henon
0.217 (0.017)
0.260 (0.024)
0.256 (0.018)
0.263 (0.016)
gen. quad.
0.212 (0.021)
0.254 (0.024)
0.257 (0.022)
0.258 (0.016)
Ikeda
0.229 (0.020)
0.274 (0.023)
0.268 (0.018)
0.271 (0.013)
Figure 3 depicts the distribution of the training errors among the examined music features with error
bars, for the best and worst performing dynamical systems. This figure signifies that the main differences of
performance in these systems lies on the first and the last three features, namely the range and percentages
of ascending, descending and constant intervals. An additional fact that should be noticed is the arithmetic
value of the mean fitness errors. For almost any measurement this value lies between 0.2 and 0.3, while
even the best mean value is above 0.2. By considering the fitness measurement method, the mean relative
distance (MRD) as defined in Section 3.1, one may assume that the best artificial compositions are expected
to be 20% different than the original ones, if such a lax quantitative conclusion may be reached. When
consulting the findings in Figure 3 however, it is indicated that there is a malapportionment of the errors
among different features. Specifically, the grad feature, which is the gradient of the line that interpolates
the notes in a segment, presents errors around 100%.
This big difference in the gradients may be explained by consulting Figure 4, which illustrates the mean
and standard deviation of the artificial pieces’ tonal features. The values of the gradients are exhibited to be
small in magnitude, a fact that also holds for the respective feature in the genuine compositions. Therefore,
an endogenous imbalance of the MRD is exposed when considering small magnitudes. For example, if the
target gradient in a segment is 0.001 and the dynamical system generates music that presents a gradient
of 0.02 in the respective segment, then the MRD value is 0.95, denoting 95% error. This measurement is in
contradiction with the intuitive approach that both gradients are small in magnitude. The errors in the rest
of the features remain in levels compared with the overall fitness (around 0.2 or 20%), with an exception
in the range feature where better performance was reached for both the best and the worst performing
maps, although their difference is noticeable. Better performance for this feature is expected, since the
range is considered in the construction of the note list, as discussed in Section 3.3. Therefore, accurate
measurements for this feature could be assessed if the iterations of the dynamical systems approached their
entire range in every segment.
July 6, 2013
9:41
toneEvo IJBC rev1
Chaos and Music: from time series analysis to evolutionary composition
13
1.6
Bach gen. quad
Mozart gen. quad
Beethoven gen. quad
Jazz gen. quad
Bach tent
Mozart tent
Beethoven tent
Jazz tent
1.4
1.2
error
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
Fig. 3.
range
grad
pdM
pdS
feature
asc
desc
const
The errors of each feature for the best (general quadratic) and the worst (tent) overall fitted maps.
Figure 4 also demonstrates that the worst performing map generates tones which efface the characteristic tonal marks of each composer, as imprinted in the collected features. Contrarily, the best performing
map preserves, at some extent, the uniqueness of features for each composer. Therefore, the implications
of fitness difference are also indicated to entail a “homogenization” in the qualitative and discriminative
characteristics of the generated music. This raises a question of which map captures each music characteristic, expressed by the examined features, more successfully. This inquiry is scrutinized in Table 6 with
a comparative analysis between all pairs of maps, for each feature and for every composer. Therein, the
dynamical system of each row is compared to the system of each column according to their performance in
each feature, for each composer. If the map of a row performs significantly better than the map of a column
for a specific feature and composer, then a “+” sign appears in the respective row, column, composer and
feature order. The results for each composer appear in each line in the respective cell (with the order
being (Bach, Mozart, Beethoven and Jazz)), while the feature are presented from left to right in the order
appeared in Table 2. If the row map is outperformed significantly, this is denoted with a “−” sign, while
a “=” sign denotes that there is no significant difference.
The significance in error differences is measured with the two–sided Wilcoxon rank sum test, as previously, with the null hypothesis being that the errors produced by a map for every composer and each
feature belong to identical continuous distributions with equal medians. The rejection of the null hypothesis
in the 5% significance level for each respective pair of error distributions is denoted with a “+” or “−”
and the contrary with the “=” symbol. The findings in Table 2 demonstrate that there is an overall feature superiority for the better performing maps, with some exceptions mainly encountered for the second
and third features (the tonal range and mean value of consecutive pitch differences). This fact probably
signifies that the dynamical systems with the “weakest” compositional skills were the ones that could not
produce descending, ascending and constant patterns that resembled the genuine compositions. Therefore,
the evolutionary process endued these maps with the comparative advantage to compose music with more
accurate overall tonal gradient and between–pitch distances.
4.3. Dynamical properties of the genuine and the artificial compositions
As mentioned in Section 2, a comparison of the dynamical properties of the artificial compositions is also
examined. The dynamical characteristics incorporate the LLE and the FD. The phase space reconstruction
is not necessary since its properties are already known through the analytic form of the dynamical systems. Furthermore, since the dynamical systems are known, different approaches could be followed for the
computation of the LLE and FD (probably including analytic computation where permitted). Nonetheless, since a comparative analysis is pursued, the same numerical approaches were considered as for the
genuine compositions. Therefore, the tonal parts of the artificial compositions were considered themselves
as time series, with their attractor characteristics (time delay and embedding dimension) being provided
by the respective dynamical system that produced them. Thought these time series, the LLE and FD were
extracted and the results, together with the fitness values, are illustrated with box plots in Table 5. These
July 6, 2013
14
9:41
toneEvo IJBC rev1
Maximos A. Kaliakatsos-Papakostas et al.
Bach
0.7
norm. value
Mozart
0.6
Beethoven
0.5
Jazz
0.4
0.3
0.2
0.1
0
−0.1
range
grad
pdM
pdS
feature
asc
desc
const
(a) features of genuine compositions
0.7
0.6
norm. value
0.5
0.4
0.3
0.2
0.1
0
−0.1
range
grad
pdM
pdS
feature
asc
desc
const
(b) features of compositions produced by the general quadratic map
0.8
0.7
norm. value
0.6
0.5
0.4
0.3
0.2
0.1
0
−0.1
−0.2
range
grad
pdM
pdS
feature
asc
desc
const
(c) features of compositions produced by the tent map
Fig. 4. Error bars of normalized feature distributions for each composer, for geniune composition (a), for artificial ones
produced by the general quadratic map (b) and by the tent map (c).
results are grouped in two manners, according to composer and map.
The finding in Figure 5 (c) and (e) demonstrate the distributions of the LLE and FD for all the artificial
compositions of all maps. In comparison to the respective distributions in Figure 2 (c) and (d), the LLE
and FD distributions for the artificial piece do not differentiate for different composers. Consulting the
distributions of compositions for all composers per map, in Figure 5 (d) and (f), one may notice that the
LLE and FD distributions for the artificial pieces are vastly different for each map. Additionally, the fitness
distributions per map, as shown in Figure 5 (b), seem to follow the behavior of the LLE distributions per
map illustrated in Figure 5 (d). This observation is evaluated by the strong correlation (0.83) of the fitness
and LLE values of the best individuals. This fact indicates that better fitness is expected to emerge from
maps with positive LLE value closer to zero. Therefore, someone could arguably notice that there might
July 6, 2013
9:41
toneEvo IJBC rev1
Chaos and Music: from time series analysis to evolutionary composition
15
Table 6. Statistical significance of the compositional superiority of each map according to each feature and composer.
Each cell in the table incorporates the comparison of the maps in the respective row and column. A significant superiority
of the row map per feature and composer is denoted by the “+” sign, the opposite with the “−” sign, while no significant
difference with the “=” symbol. Each row of symbols in a cell represents the pieces of each composer.
logistic
tent
+− = + + ++
+−−++++
+=−++++
+−−=++−
circle
− = − + −− =
− = − + −− =
− = −+ = − =
−−−=−−−
− = −−−−−
− = +−−−−
− = ++−−−
−=+=−−−
Henon
−=+=−−−
− = + = −− =
− === −− =
−−+−−−−
−++−−−−
−++−−−−
−=+=−−−
−++−−−−
− + +− == −
− = +−++−
− = +− == −
− + +− == −
logistic
—
tent
—
—
circle
—
—
—
Henon
—
—
—
—
gen. quad.
—
—
—
—
gen. quad.
−=+−−−−
− = + − −− =
− + + = −− =
−=−−−−−
−++−−−−
−++−−−−
−++−−−−
−++−−−−
− + +− == −
−=+−−−−
− = +− == −
−+ = − − −−
+ = − ====
+ = − − −− =
+ = + == − =
+ = − = −− =
—
Ikeda
− = + + −− =
− = + + −− =
− + + = −− =
−−+−−−−
−++−−−−
−++−−−−
− = +−−−−
−++−−−−
+ + + ====
+ = + = ++ =
+ = + − −− =
− = +−−−−
+ = ++ == +
+ = + + −− =
+ = + = −−+
+=+=−=+
+ = ++ == +
+ = + + ++ =
+ = + === +
+ = + ====
be a threshold of “musical chaos” above but near zero LLE values. Such an argument however, needs more
extensive scrutinization.
The mean values and standard deviations of the LLE and FD distributions for genuine and artificial
compositions for each composer and map are shown in Tables 7 and 8 respectively. The LLEs of the genuine
piece are exhibited to be closer to zero than any other LLE value of the artificial pieces in Table 7. The
LLEs that correspond to the best mean fitness, as demonstrated in Table 5, are the smallest among all
other artificial pieces, except from the piece of Beethoven. Similarly, the FD values of the genuine pieces is
considerably lower than the FDs of the artificially composed pieces. In this case, the artificial pieces with
the best fitness per composer are not the ones with the smallest FD value. It is therefore indicated, that
there is no immediate connection between the FD value, at least as expressed by the computation of the
time series algorithm, and the compositional capabilities of the dynamical systems.
Table 7. Mean and standard deviation (in parentheses) of the best individual’s largest Lyapunov exponent values for
all composers and maps. The values that correspond to best mean fitness value in Table 5 are demonstrated in boldface
typesetting.
Bach
Mozart
Beethoven
Jazz
genuine
0.043 (0.105)
0.021 (0.031)
0.053 (0.163)
0.146 (0.302)
logistic
0.373 (0.211)
0.305 (0.189)
0.345 (0.214)
0.311 (0.191)
tent
0.627 (0.016)
0.629 (0.021)
0.626 (0.015)
0.624 (0.019)
circle
0.374 (0.457)
0.370 (0.496)
0.250 (0.296)
0.533 (0.537)
Henon
0.076 (0.096)
0.082 (0.105)
0.089 (0.147)
0.057 (0.091)
gen. quad.
0.070 (0.107)
0.029 (0.060)
0.074 (0.109)
0.022 (0.031)
Ikeda
0.287 (0.169)
0.327 (0.197)
0.269 (0.231)
0.364 (0.202)
5. Conclusions
In this paper we presented a thorough study on the tonal composition potentiality of several discrete
dynamical systems. Several genuine music masterpieces composed by J. S. Bach, Mozart, Beethoven and
various jazz musicians, were utilized as music reference, providing compositional “guidelines” to the dynam-
July 6, 2013
toneEvo IJBC rev1
Maximos A. Kaliakatsos-Papakostas et al.
0.34
0.34
0.32
0.32
0.3
0.3
0.28
0.28
fitness
fitness
16
9:41
0.26
0.24
0.22
0.26
0.24
0.22
0.2
0.2
0.18
0.18
0.16
0.16
Bach
Mozart
Beethoven
composer
Jazz
logistic tent
1.8
1.6
1.6
1.4
1.4
1.2
1.2
1
0.8
0.6
Ikeda
1
0.8
0.6
0.4
0.4
0.2
0.2
0
0
Bach
Mozart
Beethoven
composer
Jazz
logistic tent
(c) Lyapunov per composer
circle Henon
map
quad.
Ikeda
(d) Lyapunov per map
1.76
1.76
1.74
1.74
fractal dim.
fractal dim.
quad.
(b) fitness per map
1.8
Lyapunov
Lyapunov
(a) fitness per composer
circle Henon
map
1.72
1.7
1.68
1.66
1.72
1.7
1.68
1.66
1.64
Bach
Mozart
Beethoven
composer
Jazz
1.64
logistic tent
(e) FD per composer
Fig. 5.
circle Henon
map
quad.
Ikeda
(f) FD per map
Fitness, LLE and FD per composer and per map for the artificial compositions.
Table 8. Mean and standard deviation (in parentheses) of the best individual’s fractal dimension values for all composers
and maps. The values that correspond to best mean fitness value in Table 5 are demonstrated in boldface typesetting.
Bach
Mozart
Beethoven
Jazz
genuine
1.637 (0.027)
1.657 (0.024)
1.633 (0.041)
1.583 (0.031)
logistic
1.721 (0.013)
1.721 (0.015)
1.724 (0.016)
1.723 (0.015)
tent
1.685 (0.004)
1.684 (0.005)
1.685 (0.004)
1.685 (0.005)
circle
1.701 (0.021)
1.692 (0.024)
1.683 (0.024)
1.698 (0.021)
Henon
1.730 (0.017)
1.724 (0.019)
1.712 (0.020)
1.719 (0.018)
gen. quad.
1.714 (0.022)
1.706 (0.019)
1.709 (0.023)
1.699 (0.019)
Ikeda
1.732 (0.008)
1.730 (0.013)
1.722 (0.023)
1.724 (0.024)
ical systems. Through an evolutionary approach, the parameters of the discussed systems were optimized
so that the tonal sequences that each system composes, resemble the ones of the genuine pieces, according
to some tonal features which encompass essential musical information. Firstly, a time series analysis was
performed on the tonal sequences provided by the genuine composition that allowed an approximate phase
July 6, 2013
9:41
toneEvo IJBC rev1
REFERENCES
17
space reconstruction of their attractor. Consequently, their Largest Lyapunov Exponent (LLE) and Fractal
Dimension (FD) could be assessed, allowing a comparative analysis of their dynamical characteristics and
the respective characteristics of the artificial pieces. The resulting artificial compositions were scrutinized in
several aspects, which incorporated their tonal capabilities and dynamical characteristics for each composer
and according to each tonal feature. Thereby, a thorough survey on the tonal compositional capabilities of
each map was allowed, together with several comparative studies that provide insights about the qualitative
music characteristics that may be pursued by utilizing dynamical systems as automatic music composers.
Through the time series analysis performed on the genuine musical masterpieces, an interesting finding
is that the attractor characteristics of each composer differ significantly. The dynamical characteristics
of these attractors however, could not be accurately reproduced by the examined dynamical systems.
This is probably due to the fact that the examined dynamical systems are quite “simple” in terms of
their dimensionality and time delay, as demonstrated by the respective phase space reconstruction values.
However, some important insights were provided about the connection of qualitative music features and
dynamical characteristics through the LLE value. Specifically, it was indicated that dynamical systems with
smaller positive LLEs produced compositions which were more similar to the genuine pieces. Additionally,
it may be assumed that iterative maps which performed better were the ones that could reproduce tonal
motif–like structure, which is expressed through the features that incorporated percentages of ascending,
descending or constant pitch intervals within short segments of music. Contrarily, the worst performing
maps, were better at capturing less refined musical features, like absolute pitch differences and pitch–change
gradient.
The fact that the reconstructed phase space of the original compositions incorporated a great number
of dimensions , indicates that the appropriate dynamical systems should probably also incorporate more
dimensions. Thereby, the task at hand could be approached by a similar evolutionary strategy, with iterative
dynamical systems of higher dimensionality. The equations that constitute these systems however could
incorporate a large number of parameters. For example, the general quadratic map, which constitutes
a general form of a two dimensional quadratic system incorporates 12 parameters, considering also the
initial iterations. The n–dimensional version of the general quadratic map, for example, incorporates 2n +
2n2 parameters and therefore their number becomes overwhelming even for a relatively small number of
dimensions. Future work could also include a similar time series analysis on pieces with more simple tonal
structure, like dances or even contemporary popular songs. Thereby, the connections between the perceived
tonal complexity and the dynamical characteristics, e.g. the fractal dimension, could be directly examined.
Finally, future work should also incorporate the examination of the compositional capabilities of several
dynamical systems under “special” parameter value combinations, which provoke “special” dynamical
behavior. For instance, the compositional characteristics of the logistic function could be examined when
its parameter is set at the Feigenbaum point, provoking “weakly chaotic” dynamics, through a similar
comparison with genuine human compositions.
Acknowledgments
We would like to thank Professor Tassos C. Bountis for the useful discussions over several aspects of the
paper’s topics. This research has been partially co-financed by the European Union (European Social Fund
ESF) and Greek national funds through the Operational Program “Education and Lifelong Learning” of
the National Strategic Reference Framework (NSRF) - Research Funding Program: Thales. Investing in
knowledge society through the European Social Fund.
References
Ahmadi, B. & Amirfattahi, R. [2010] “Comparison of correlation dimension and fractal dimension in
estimating BIS index,” Wireless Sensor Network 2.
Alevizos, P. D. & Vrahatis, M. N. [2010] “Optimal dynamic box–counting algorithms,” International Journal of Bifurcation and Chaos 20, 4067–4077.
July 6, 2013
18
9:41
toneEvo IJBC rev1
REFERENCES
Amador, A. & Mindlin, G. B. [2008] “Beyond harmonic sounds in a simple model for birdsong production,”
Chaos: An Interdisciplinary Journal of Nonlinear Science 18, 043123–043123–6.
Behrman, A. [1999] “Global and local dimensions of vocal dynamics,” The Journal of the Acoustical Society
of America 105, 432–443.
Bidlack, R. [1992] “Chaotic systems as simple (But complex) compositional algorithms,” Computer Music
Journal 16, 33–47.
Bilotta, E., Pantano, P., Cupellini, E. & Rizzuti, C. [2007] “Evolutionary methods for melodic sequences
generation from non-linear dynamic systems,” Applications of Evolutionary Computing, no. 4448 Lecture Notes in Computer Science (Springer Berlin Heidelberg), pp. 585–592.
Boon, J. P. & Decroly, O. [1995] “Dynamical systems theory for music dynamics,” Chaos: An Interdisciplinary Journal of Nonlinear Science 5, 501–508.
Chialvo, D. R. [2003] “How we hear what is not there: A neural mechanism for the missing fundamental
illusion,” Chaos: An Interdisciplinary Journal of Nonlinear Science 13, 1226–1230.
Choi, I. [1994] “Interactive exploration of a chaotic oscillator for generating musical signals in real-time
concert performance,” Journal of the Franklin Institute 331, 785–818.
Chua, L., Wu, C., Huang, A. & Zhong, G.-Q. [1993] “A universal circuit for studying and generating
chaos. i. routes to chaos,” IEEE Transactions on Circuits and Systems I: Fundamental Theory and
Applications 40, 732 –744.
Coca, A. E., Tost, G. O. & Zhao, L. [2010] “Characterizing chaotic melodies in automatic music composition,” Chaos: An Interdisciplinary Journal of Nonlinear Science 20.
Dabby, D. S. [1996] “Musical variations from a chaotic mapping,” Chaos: An Interdisciplinary Journal of
Nonlinear Science 6, 95–107.
Esteller, R., Vachtsevanos, G., Echauz, J. & Litt, B. [2001] “A comparison of waveform fractal dimension
algorithms,” IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications
48, 177 –183.
Fletcher, N. H. [1994] “Nonlinear dynamics and chaos in musical instruments,” Complexity International
1.
Fraser, A. & Swinney, H. [1986] “Independent coordinates for strange attractors from mutual information,”
Physical Review A 33, 1134–1140.
Goh, C., Hamadicharef, B., Henderson, G. T. & Ifeachor, E. C. [2005] “Comparison of fractal dimension
algorithms for the computation of EEG biomarkers for dementia,” Proceedings of the 2nd International Conference on Computational Intelligence in Medicine and Healthcare (CIMED2005) (Lisbon,
Portugal), pp. 464–471.
Grassberger, P. & Procaccia, I. [1983] “Measuring the strangeness of strange attractors,” Physica D: Nonlinear Phenomena 9, 189–208.
Harley, J. [1994] “Algorithms adapted from chaos theory: Compositional considerations,” Proceedings of
the 1993 International Computer Music Conference (ICMA 1994) (San Francisco, USA), pp. 209–212.
Higuchi, T. [1988] “Approach to an irregular time series on the basis of the fractal theory,” Phys. D 31,
277–283.
Hrman, H. [1993] “Deterministic chaos, iterative models, dynamical systems and their application in algorithmic composition,” Proceedings of the 1993 International Computer Music Conference (ICMA
1993) (San Francisco, USA), pp. 194–197.
Jafari, A., Almasganj, F. & Bidhendi, M. N. [2010] “Statistical modeling of speech poincar sections in
combination of frequency analysis to improve speech recognition performance,” Chaos: An Interdisciplinary Journal of Nonlinear Science 20, 033106–033106–11.
Kaliakatsos-Papakostas, M., Floros, A. & Vrahatis, M. N. [2013] “evodrummer: Deriving rhythmic patterns
through interactive genetic algorithms,” Evolutionary and Biologically Inspired Music, Sound, Art and
Design (Springer Berlin Heidelberg), pp. 25–36.
Kaliakatsos-Papakostas, M. A., Floros, A. & Vrahatis, M. N. [2012a] “Intelligent real-time music accompaniment for constraint-free improvisation,” Proceedings of the 24th IEEE International Conference
on Tools with Artificial Intelligence (ICTAI 2012) (Piraeus, Athens, Greece), p. (to appear).
Kaliakatsos-Papakostas, M. A., Floros, A. & Vrahatis, M. N. [2012b] “Music synthesis based on nonlinear
July 6, 2013
9:41
toneEvo IJBC rev1
REFERENCES
19
dynamics,” In Proceedings of Bridges 2012, Mathematics, Music, Art, Architecture, Culture (Baltimore, USA), p. (To appear).
Kantz, M. J. [1988] “Fractals and the analysis of waveforms,” Computers in Biology and Medicine 18,
145–156.
Kennel, M. B., Brown, R. & Abarbanel, H. D. I. [1992] “Determining embedding dimension for phase-space
reconstruction using a geometrical construction,” Physical Review A 45, 3403–3411.
Kodba, S., Perc, M. & Marhl, M. [2005] “Detecting chaos from a time series,” European Journal of Physics
26, 205–215.
Kugiumtzis, D. & Tsimpiris, A. [2010] “Measures of analysis of time series (mats): A matlab toolkit for
computation of multiple measures on time series data bases,” Journal of Statistical Software 33, 1–30.
Lai, Y.-C. & Ye, N. [2003] “Recent developments in chaotic time series analysis,” International Journal of
Bifurcation and Chaos 13, 1383–1422.
Manaris, B., Romero, J., Machado, P., Krehbiel, D., Hirzel, T., Pharr, W. & Davis, R. B. [2005] “Zipf’s law,
music classification, and aesthetics,” Computer Music Journal 29, 55–69, ArticleType: research-article
/ Full publication date: Spring, 2005 / Copyright 2005 The MIT Press.
Mandelbrot, B. B. [1982] The Fractal Geometry of Nature (W. H. Freeman and Company), ISBN
0716711869.
Oudre, L., Févotte, C. & Grenier, Y. [2010] “Probabilistic framework for template-based chord recognition,”
Multimedia Signal Processing (MMSP), 2010 IEEE International (Saint-Malo, France), pp. 183 –187.
Petrosian, A. [1995] “Kolmogorov complexity of finite sequences and recognition of different preictal eeg
patterns,” Proceedings of the Eighth Annual IEEE Symposium on Computer-Based Medical Systems,
CBMS ’95 (IEEE Computer Society, Washington, DC, USA), pp. 212–217.
Polychronaki, G. E., Ktonas, P. Y., Gatzonis, S., Siatouni, A., Asvestas, P. A., Tsekou, H., Sakas, D. &
Nikita, K. S. [2010] “Comparison of fractal dimension estimation algorithms for epileptic seizure onset
detection,” Journal of Neural Engineering 7, 046007.
Pressing, J. [1988] “Nonlinear maps as generators of musical design,” Computer Music Journal 12, 35–46.
Price, K., Storn, R. M. & Lampinen, J. A. [2005] Differential Evolution: A Practical Approach to Global
Optimization (Natural Computing Series) (Springer-Verlag New York, Inc., Secaucus, NJ, USA), ISBN
3540209506.
Raghavendra, B. S. & Dutt, N. D. [2010] “Computing fractal dimension of signals using multiresolution
box-counting method,” International Journal of Information and Mathematical Sciences 1, 50–65.
Rizzuti, C., Bilotta, E. & Pantano, P. [2009] “A GA-Based control strategy to create music with a
chaotic system,” Applications of Evolutionary Computing, no. 5484 Lecture Notes in Computer Science
(Springer Berlin Heidelberg), pp. 585–590.
Rosenstein, M. T., Collins, J. J. & De Luca, C. J. [1993] “A practical method for calculating largest
lyapunov exponents from small data sets,” Physica D: Nonlinear Phenomena 65, 117–134.
Sevcik, C. [1998] “A procedure to estimate the fractal dimension of waveforms,” Complexity International
5.
Shannon, C. E. [2001] “A mathematical theory of communication,” ACM SIGMOBILE Mobile Computing
and Communications Review 5, 3–55.
Storn, R. & Price, K. [1997] “Differential evolution – a simple and efficient adaptive scheme for global
optimization over continuous spaces,” Journal of Global Optimization 11, 341–359.
Takens, F. [1981] “Detecting strange attractors in turbulence,” Dynamical Systems and Turbulence, Warwick 1980 (Springer Berlin Heidelberg), pp. 366–381.
Tokuda, I., Riede, T., Neubauer, J., Owren, M. J. & Herzel, H. [2002] “Nonlinear analysis of irregular
animal vocalizations,” The Journal of the Acoustical Society of America 111, 2908–2919.
Tricot, C. [1994] Curves and Fractal Dimension (Springer).
Voss, R. F. & Clarke, J. [1978] “1/f noise in music: Music from 1/f noise,” The Journal of the Acoustical
Society of America 63, 258–263.
Wilcoxon, F. [1945] “Individual comparisons by ranking methods,” Biometrics Bulletin 1, 80–83.
Wolf, A., Swift, J. B., Swinney, H. L. & Vastano, J. A. [1985] “Determining lyapunov exponents from a
time series,” Physica D: Nonlinear Phenomena 16, 285–317.