You are on page 1of 6

Optimisation techniques for solving design problems in modern trombones

Alistair C.P. Braden


Room 4201, School of Physics, University of Edinburgh, EH9 3JZ, UK. e-mail: alistair.braden@ed.ac.uk

This paper presents further approaches to the problem of numerical shape optimisation of brass musical
instruments, including a flexible object-oriented technique that describes the internal geometry of the in-
strument with very few design variables. This technique, used here both with a genetic algorithm and the
Rosenbrock 0th-order search algorithm, gives a significant increase in optimisation speed. Results of an
experimental study of the input impedances of a wide variety of different trombones are presented, dis-
cussed, and used to identify design problems common in modern trombones. From this, target impedance
data for the optimisation are formulated and a preliminary attempt made at demonstrating the effectiveness
of optimisation techniques at solving ‘real-world’ instrument design problems.

1 Introduction rate, fast and easy to use, would be a valuable asset as


it would allow a great many potential designs to be eval-
The calculation of the input impedance of a brass instru- uated cheaply and quickly. While human expertise in de-
ment from its bore-profile is well-established in the lit- sign and testing will always be paramount, such a com-
erature. The inverse problem – finding the bore of an putational tool would be very useful to a designer.
instrument from its impedance – can be solved using op-
timisation techniques (i.e. trial and improvement). This
requires an optimisation algorithm, a suitable geometric 2 Optimisation Problem
representation of the instrument as a set of ‘design vari-
ables’ to be optimised, and an ‘objective function’ with 2.1 Objective Functions
which to judge the merit of each potential match.
To the author’s knowledge, two previous attempts have An objective function compares impedance properties of
been made at this inverse problem. The first, by Kausel each ‘tested’ instrument to target impedance data. As in
[1] represents the instrument using a piece-wise linear in- [2] we denote the instrument symbolically by α. Three
terpolation of points along the bore (‘point-wise’), and objective functions are used here in different combina-
optimises using either the 0th -order Rosenbrock search tions depending on the task, each giving scores between
algorithm, or (unsuccessfully) an incremental genetic al- 0 and 1 (a perfect match). Typical parameter values, de-
gorithm. The second, by Noreland [2], uses a point-wise termined from experience, are given. The first function is
representation of the second derivative of the bore radii, a windowed linear least-squares comparison of the entire
and the gradient-based Levenberg-Marquardt algorithm. (discrete) impedance magnitude curve:
Optimisation techniques of this kind can be applied either
to find an unknown bore from a known impedance curve n
(
(bore reconstruction) or to improve the design of a given 1X δZi2
if δZi < σ
J1 (α) = 1− σ2 (1)
bore such that its impedance matches desired character- n i=1 1 if δZi ≥ σ
istics (such as intonation or impedance envelope). Solu-
tion of these problems also requires an accurate physical where δZi = |Z(fi )−Z 0 (fi )|, i.e. the difference between
model, accurate measurements of the bore and impedance tested impedance magnitude Z and target Z 0 evaluated
of real instruments, and an understanding of the correla- at a series of frequencies fi , σ is the half-width of the
tion between impedance properties and playing charac- window (usually 1MΩ) and n the number of points in the
teristics. impedance curve. The second function is a windowed-
Instrument manufacturers have substantial accumulated Gaussian comparison, scoring only the peak-frequencies:
expertise regarding the design and construction of high-
quality instruments, and newly-designed models are cur-  µ ¶
rently being produced relatively frequently. However, the 
 −µφ δφ2i
1 − exp
1X 
k
design process is costly and time-consuming, as it gen- ξi2
J2 (α) = 1− if δφi < ξi
erally involves building prototypes which may or may k i=1 
 1 − exp(−µφ )

not be an improvement over existing models. A com- 1 if δφi ≥ ξi
putational design tool, provided it is sufficiently accu- (2)

557
Forum Acusticum 2005 Budapest Braden

where δφi = |φi − φ0i |, i.e. the difference between tested 3 Implementation
peak-frequency φi and target peak-frequency φ0i of peak
i, the half-width of the window ξi = νφ φ0i is dependent 3.1 Program Design
on target frequency and parameter νφ (typically 0.1), µφ
is a ‘strictness’ parameter (typically between 10 and 40) A primary design aim was to produce a computer optimi-
and k the number of impedance peaks being tested. sation ‘toolbox’ that is powerful, flexible and easy-to-use.
A very similar function J3 is also used for the peak- To this end, an application has been written allowing the
heights, using difference |Z(φi )−Z 0 (φ0i )| and parameters user to manipulate, modify and evaluate instruments and
νZ and µZ (typically 0.3 and 20 respectively). We then impedance curves. The choice of programming language
define an overall objective function: (C++) allows a modular, object-oriented approach, and is
very fast at run-time.
J(α) = j1 J1 (α) + j2 J2 (α) + j3 J3 (α) (3)
3.2 Instrument Representation
where j1...3 are weights, which can take any real value;
later we refer to the choice of objective function as e.g. Previous instrument representations have been point-wise
{0, 1, 0}, denoting values of j. We then seek α to max- (or ‘bottom-up’); while this is the most simple approach,
imise J(α). it requires N ∼ 100 (where N is the number of design
variables) to be effective and can lead to some unsatisfac-
2.2 Physical modelling tory ‘jagged’ or ‘wiggly’ shapes. The computation time
of optimisation algorithms is strongly dependent on N ;
Well-established in the literature is the one-dimensional if this value can be reduced without compromising effec-
transmission-line [3], where an axisymmetric horn is tiveness, the optimisation speed will be improved.
modelled as a concatenation of cylinders and truncated It is well-known that a single Bessel-horn (5) is a fair ap-
cones (piece-wise constant and piece-wise linear inter- proximation to a brass instrument bell; by using several,
polation); this is accurate for moderately-flaring horns. shorter, concatenated Bessel-horns we thus gain a much
We use this model here, and it should be noted that al- better approximation (piece-wise Bessel-horn interpola-
though more complex geometrical shapes are used in the tion). Bells can be described very closely by 5 Bessel-
optimiser, for the purposes of impedance calculation they horns of different flares, giving1 N = 12 and therefore a
are themselves interpolated linearly and the standard TL significant increase in optimisation speed.
method applied. No account is currently taken of bends We introduce an elegant ‘top-down’ approach, allow-
in the tubing. ing us to specify a prototype that clearly defines an in-
strument as having the features and geometry of e.g. a
2.3 Impedance Curves trombone but without the exact specifications. The opti-
miser is therefore constrained to consider only reasonable
Calculated impedance curves are commonly approxi- ‘trombone shapes’, considerably reducing the size of the
mated by a series of points at regular frequency inter- search space and therefore improving optimisation speed.
vals. Finer grids give more accurate peak locations but The instrument is represented as a flexible series of ‘el-
are computationally more expensive; an alternative is of- ements’, which are divided into two categories: primi-
fered here. A coarse grid is used, calculating impedance tive elements and list elements. Primitives (e.g. cylinder,
every 8.0Hz, to give a first approximation of the peak fre- cone, Bessel-horn) are defined in terms of their individ-
quencies; these are then refined recursively, to a given ual properties (length, radius etc.) instead of as a series
tolerance, using either the bisection method or inverse of points relative to the instrument’s coordinate origin;
parabolic interpolation [4], evaluating impedance at each elements can therefore be modified, added in or deleted
new frequency. Using this approach, an impedance curve without affecting the rest of the instrument. List elements
can be represented from 0-2000Hz with peaks known to allow a distinct part of the instrument to be treated as
a tolerance of 0.1Hz by ∼350 points; these peak data are a single element, allowing large sections of bore to be
then passed to (2). moved easily between instruments (e.g. a mouthpiece el-
ement, itself a list of cone elements).
This method is implemented in C++ as a recursive, poly-
morphic linked-list of elements. The element classes
contain member variables describing their geometry and
functions for evaluating their own impedance. This gives
1 One variable is the length of the sections, 6 describe the input and

output radii of the sections and 5 give their flare coefficients

558
Forum Acusticum 2005 Budapest Braden

tremendous flexibility and it is very easy for a user to set II gave very poor results. Although the GA performance
up an instrument with any desired geometry, including is a significant improvement on [1], it is still neither sta-
fully point-wise instruments. It can, in principle, be used ble nor capable enough to be practical. As we shall see,
with any suitable optimisation algorithm. the GA is again outperformed by Rosenbrock; given that
GA’s generally perform best in low-N problems, such as
that presented here, we can conclude that further use of
3.3 Optimisation Algorithms genetic algorithms for for brass-instrument optimisation
is not worthwhile.
Two different optimisers are used here: a simple ge-
netic algorithm [5] and the Rosenbrock [6, 7] algorithm.
In each case the optimiser is essentially seeking a set 4.2 Rosenbrock
{α1 . . . αN } of design variables to maximise J(α). The
choice of instrument representation allows us to specify Experiment I was repeated using the Rosenbrock algo-
that single dimensions (e.g. an input radius) or entire sec- rithm, with a 4-Bessel bell (N = 11). Results were
tions (e.g. a mouthpiece) are to be held constant. Each in every way superior to the GA: a near-perfect match
design variable can be given upper and lower bounds and of both impedance and bore was found, with total re-
a starting value. No constraints are made on the smooth- peatability, in 62s. Experiment II was then attempted:
ness of joins between elements (notably Bessel-horns) due to the kinks in the target’s cylindrical section (which
and therefore some unsatisfactory shapes remain in the were intentionally not accounted for in the reconstruc-
search space; any such constraints would compromise tion prototype), a close impedance match was outside the
impedance match in favour of smoothness, and are there- search space and therefore impossible, thus making the
fore undesirable anyway. problem intrinsically more difficult. In spite of this, an
The computations were performed on a 2.6GHz Windows equally-good match of the bore (excepting the kinks in
PC. the cylinder) was found – the closest impedance match
was still very close to the correct bore-profile, suggesting
that this optimiser is capable of reconstructing an instru-
ment’s bell without any knowledge of its valve sections.
4 Bore Reconstruction Results
The problem of reconstructing trombones was then at-
To test the efficacy of the optimisation techniques, a diffi- tempted. Prototypes based on two virtual trombones (A,
cult problem was set. A virtual2 trumpet had its cylin- a large-bore tenor, and B, a bass) were relatively com-
drical section smoothed to remove ‘kinks’ due to the plex (N = 24 and N = 18) due to a series of alter-
valves, tuning-slide etc. The optimisers were asked to nating cylindrical and conical sections required between
reconstruct the bore of this instrument from its calculated the main slide and the bell. The virtual instruments were
impedance (experiment I). The kinks in the cylindrical successfully reconstructed as above, with mouthpiece E
section were then replaced, altering the target impedance attached, both with and without the kinks in the target.
(expt. II) accordingly. The instrument had a reconstruc- These reconstructions, which took 91-129s and required
tion prototype of a cylinder attached to a multi-Bessel 520-618 J-evaluations, are a very close match, differing
bell; the total bell length was a variable with each Bessel- in radius from the target virtual instruments generally by
section an equal fraction of this length. Wide bounds <0.2mm (Figures 1 & 2), excepting the kinks and slight
were set on all variables (generally ±10%, wider for flare discrepancies in the extreme bell region. While the two
coefficients), and objective function {1, 1, 0} was used. prototypes model tenor and bass trombones respectively,
they are still capable of reconstructing the other instru-
ment, though admittedly somewhat less effectively.
4.1 Genetic Algorithm
After considerable experimentation, the best results were
found by optimising a 4-Bessel bell (N = 11) for two
50-individual generations, then, for each member of the
population, splitting each Bessel-section into two equal
parts, giving an 8-Bessel bell (N = 19), and optimising
for a further 18 generations. Results were inconsistent,
not always producing good matches when the experiment
was repeated. Experiment I (at best) gave a good match Figure 1: Target A (solid) and reconstructed A1 (dotted)
to both impedance and bore in ∼5 minutes; experiment bore profile (kinks detail inset). Units /cm.
2 We define ‘virtual’ to denote a theoretical instrument based on

physical measurements on the bore-profile of a real-world instrument.

559
Forum Acusticum 2005 Budapest Braden

Figure 2: Target B (solid) and reconstructed B1 (dotted)


bore profile (kinks detail inset).

5 Impedance Measurements
5.1 Overview
Figure 3: Measured EFP for trombones B (dashed), C
In order to perform intonation optimisation on instru-
(solid) and D (dotted), normalised by 4th mode.
ments we must first gain an understanding of a good tar-
get. To this end, the input impedances of a variety of
modern trombones were measured using the BIAS equip- We give here a brief summary of a larger study conducted
ment [8]. The frequency of each peak was then recorded on many more trombones than can be shown here. Fig-
and plotted in terms of its equivalent fundamental pitch ure 3 gives EFP plots for three typical trombones, nor-
(EFP) [9]. This calculates, for each peak fn , the funda- malised by 4th mode (thus supressing the absolute into-
mental frequency of which it is an exact nth harmonic nation of the instruments for easier comparison). C and
and the pitch of this equivalent fundamental relative to an D are large-bore tenor trombones measured with mouth-
arbitrary reference frequency F ; this gives a measure, in piece E, and B is a bass trombone measured with mouth-
cents, of the harmonic alignment of the peak frequencies: piece G. B and C are commonly described by profes-
µ ¶ sional players as being of first-rate quality. We see that
1200 fn the assumption that brass instruments should have exactly
E(fn ) = log (4)
log(2) nF harmonically-aligned peaks is somewhat naïve: Large-
bore tenor and bass trombones have a general right-hand
When plotting EFP for several different instruments on diagonal EFP trend – i.e. higher peaks are tuned to a pro-
the same axes, it is often convenient to normalise by gressively sharper effective fundamental (medium-bore
choosing F = f4 /4 for each instrument, thus shifting all trombones and French horns [9] generally have a vertical
plots such that each peak 4 lies on the vertical axis (ef- EFP trend). This suggests that exact harmonic alignment
fectively retuning the instruments so that each 4th mode may not be desirable for this class of instrument. How-
coincides). The shape of the plots is independent of F . ever, bass trombones tend to have peak 2 more closely
aligned with peak 4 (i.e. the regime for B[2 or the ‘lower
5.2 Results & Discussion register’) than for tenor trombones, suggesting that close
alignment may yet be desirable in certain circumstances.
It has been known since Benade [10] that each sounded Clearly the issue of the harmonic alignment of peaks is
note consists of a ‘regime’ of several harmonically- not straightforward, and there are several factors in ef-
related peaks; for example B[2 played on a trombone ex- fect. Further work is needed to determine the sensitivity
cites peaks 2, 4, 6 etc. It should be noted that some peaks of players to changes in alignment, and the dependence of
(e.g. 4, 6 & 8) are part of the regimes of several sounded the performance of the instrument on both the alignment
notes, whereas others (those of prime index) contribute and the impedance envelope.
only to one regime. Consider the regimes of B[3 (i.e. peaks 4, 8, 12 etc.) &
It is generally accepted by the trombone community that D4 (peaks 5, 10, 15 etc.) of instrument D – peak 8 has an
the material composition (e.g. alloy and thickness of ma- equivalent fundamental that is +56c sharper than that of
terial) of a brass instrument affects its performance; while peak 4, and peak 12 is likewise +57c sharper; similarly
there currently no conclusive proof, recent work [11, 12] peak 10 is −12c and peak 15 −9c compared to peak 5.
is suggestive. We should therefore accept that there may It is reasonable to suggest that a player moving from the
be objective aspects of instrument performance which regime of peak 4 (+56, +57) to that of peak 5 (−12, −9)
cannot be accounted for in impedance measurements by with instrument D will therefore have to make a larger
narrow-band excitation of comparatively low amplitude, embouchure adjustment than they would with C (+37,
such as those presented here. +42 and −2, +20) as the two regimes differ much more

560
Forum Acusticum 2005 Budapest Braden

in EFP. Players may be sensitive to this difference, and


further investigation may prove valuable.

6 Intonation Optimisation
Instrument A, with (fixed) virtual mouthpiece E, was set
as a starting instrument. Two experiments using {0, 1, 0} Figure 6: Bore profiles for B2 (solid) and B3 (dotted).
were performed, firstly to shift peaks 2 & 3 by +10c (giv-
ing result A2) and secondly to replicate the measured EFP
plot (Figure 3, solid) of instrument C (A3). Both tasks
were largely successful (Figures 4 & 5); the first caused
a shift of higher peaks and the second produced a good,
but not perfect, match of the target EFP (even though this
required some peaks to be shifted by >50c).
Two experiments were performed on B, again with
mouthpiece E, to shift peak 5 by −15c, firstly with all
other peaks held constant (B2), and secondly to also shift
peaks 2 & 3 by +10c (B3). Both experiments largely suc-
ceeded (Figures 6 & 7) at a cost of slight shifts in other
peaks.
For both instruments, smooth bore-profiles were pro-
duced and the impedance peak magnitudes were largely
unaltered. Optimisations using {0,1,1}, keeping the im-
Figure 7: EFP for B (solid), B2 (dashed) and B3 (dotted).
pedance heights constant, proved less successful (but
were found to be effective when optimising Horns).
These results demonstrate that the optimiser is capable
of modifying instruments to give a good match of speci-
fied theoretical impedance characteristics, whether shift-
ing single or many peaks, while maintaining a smooth,
æsthetically-pleasing bore profile.

7 Conclusions & Further Work


Figure 4: Bore profiles for A2 (solid) and A3 (dotted).
An object-oriented method for representing the geometry
of a brass instrument with very few design variables is
presented. This method is used both with a genetic algo-
rithm and with the Rosenbrock algorithm, the latter being
considerably more effective. An optimisation applica-
tion has been written which can reconstruct smooth bore
profiles from given calculated (theoretical) impedance
curves to a very good accuracy, even if certain ‘kinks’ ex-
ist in the target bore. Results of an extensive experimen-
tal survey of the impedances of modern trombones are
briefly summarised. Two ‘virtual’ instruments based on
modern trombones are optimised to shift their impedance
peaks to match target criteria; while there is generally a
cost of other peaks also being somewhat shifted, the mod-
ified bore profiles are smooth.
The optimisation technique is proven to be successful
Figure 5: EFP for A (solid), A2 (dashed) and A3 (dotted). both at reconstructing and optimising instruments based
on theoretical impedance curves. However, significant
discrepancies exist between the calculated impedance

561
Forum Acusticum 2005 Budapest Braden

for virtual (target) instruments and their corresponding [2] D. Noreland, ‘A gradient based optimisation al-
experimentally-measured impedance. Causes include the gorithm for the design of brass-wind instruments’,
inaccuracy inherent in the plane-wave approximation, PhD thesis, Uppsala University (2003)
and difficulty getting complete and accurate measure-
ments of the bore of real instruments. As a consequence, [3] D. Mapes-Riordan, ‘Horn Modeling with Coni-
the changes in calculated impedance due to alterations in cal and Cylindrical Transmission-Line Elements’,
the bore may not accurately represent the corresponding J.Audio Eng. Soc., 41, p.471-483 (1993)
measured impedance if the changes were implemented in [4] W.H. Press, ‘Numerical Recipes in C++’,
practice; this difficulty does not, however, alter the ef- Cambridge University Press, p.406-407, ISBN
fectiveness of the optimisation technique itself. Once the 0521750334 (2002)
optimiser is coupled with a sufficiently accurate physical
model, it is reasonable to expect its results to be of gen- [5] M. Wall, ‘GALib’, C/C++ Genetic Algorithm Li-
uine practical use for brass instrument design. brary, http://lancet.mit.edu/ga
In order to improve the modelling, one must either take [6] F. Van den Bergen, C/C++ Rosenbrock package,
account of higher modes in the transmission-line [13] http://iridia.ulb.ac.be/~fvandenb/
or use a (hybrid) FEM approach [14]. Improvements optimization/rosenbrock.html
may also be found by accounting for bends in the tub-
ing – manufacturers describe bend shapes as having a [7] H.H. Rosenbrock, ‘An automatic method for finding
noticeable effect. Optimisation speed may be improved the greatest or least value of a function’, Comp. J.,
by combining the above technique with the gradient- 3, p.175-184 (1960)
based methods of [2]. Further, more detailed, studies of
impedance curves for trombones and other instruments, [8] ‘Brass Instrument Analysis System’
coupled with rigorous playing-tests, will give useful in- http://www.bias.at
sight into desirable impedance characteristics. [9] J.P. Chick, C. Lumb & D.M. Campbell, ‘Pas-
sive acoustic characteristics and intonation prob-
lems of modern orchestral horns’, Proceedings of
Appendix ISMA2004, Nara, Japan (2004)

Bessel-horns are defined as: [10] A.H. Benade, ‘Fundamentals of Musical


Acoustics’, Oxford University Press, p402-405
(1976)
r(x) = b(x + x0 )−γ (5)
[11] J.W. Whitehouse, ‘A study of the wall vibrations
Where r is the radius, x the distance along the instrument, excited during the playing of lip-reed instruments’,
γ the flare coefficient and b, x0 parameters usually de- PhD thesis, Open University (2003)
fined in terms of specified length and input/output radii.
[12] T.R. Moore, E.T. Shirley, I.E.W. Codrey & A.E.
Daniels, ‘The effects of bell vibrations on the sound
Ackowledgements of the modern trumpet’, ACUSTICA, 91, p.578-589
(2005)
I offer sincere thanks to Murray Campbell and John [13] V. Pagneux, N. Amir & J. Kergomard ‘A study
Chick for their invaluable help and discussion throughout of wave propagation in varying cross-section
this work, to Arnold Myers for providing bore measure- waveguides by modal decomposition. Part I: Theory
ments, to the IWK for provision of the BIAS equipment, and Validation’, J. Acoust Soc. Am., 100, p. 2034-
and to Michael Rath Brass Instruments for their insight 2048 (1996)
and kind provision of excellent instruments to measure.
This work was supported by EPSRC. [14] D. Noreland, ‘A Numerical Method for Acoustic
Waves in Horns’, ACUSTICA, 88, p.576-586,
(2002)
References
[1] W. Kausel, ‘Computer optimization of brass wind
instruments’, Diderot Forum on Mathematics and
Music, p.227, OECG: ISBN 3-85403-133-5 (1999)

562

You might also like