An Intelligent Design Explorer for New Violin Shapes

Hao Wang and Lutz Hamel

Department of Computer Science and Statistics, University of Rhode Island, Kingston, U.S.A.

Keywords:

Evolutionary Computation, Evolutionary Strategies, Creative Evolutionary Systems, Art and Design.

Abstract:

This work focuses on a design tool, the violin design explorer, for classic violin outline design. Our design

tool employs an evolution strategies algorithm and can be viewed as a creative evolutionary system. With this

system, users without any design knowledge can create their favorite violin outline through a set of simple

choices. The underlying violin designs are implemented using the Digital Amati geometry engine. In order to

validate the design tool we conducted an anonymous experiment. Four groups of participants were asked to

design their favorite violin outline. Design patterns that emerged within these four participant groups during

the experiment seem to validate that our design tool works as we envisioned. We view this implementation of

the designer as the ﬁrst stage of a comprehensive violin design tool.

1 INTRODUCTION

The origins of evolutionary computation can be traced

back to the late 1950s and started to receive signif-

icant attention during the 1980s (B

ack et al., 2018).

Pioneers have conducted research using evolutionary

computation, especially in the ﬁeld of design, mu-

sic, and art. Here we explore whether we can use

evolutionary computation in the area of violin outline

design. Our design tool, the violin design explorer,

employs an evolution strategies algorithm and can be

viewed as a creative evolutionary system. The vio-

lin design explorer aims to help users without design

experience to directly integrate their imagination and

creativity into violin outline design and create novel

solutions. The explorer abstracts each violin outline

into a unique genotype representation suitable for de-

sign representation within the Digital Amati geometry

engine (Digital Amati, 2018) which also handles vio-

lin design parameters and constraints. Instead of us-

ing a ﬁtness function the violin design explorer uses

human selection to determine the parent populations

of the evolutionary computation. Besides allowing

users without design background or knowledge to un-

dertake non-trivial design tasks another advantage of

creative evolutionary systems is that they allow the

user to sidestep the limitations of “conventional wis-

dom” and “design ﬁxation” (Menges and Ahlquist,

2011).

We validated the design tool by conducting an

anonymous experiment. Four groups of participants

were asked to design their favorite violin outline. De-

sign patterns that emerged within these four partici-

pant groups during the experiment seem to validate

that our design tool works as we envisioned.

We view this implementation of the designer as

the ﬁrst stage of a comprehensive violin design tool.

A next logical step would be to extend the designer to

embody acoustic characteristics besides the pure aes-

thetic consideration currently implemented through

Digital Amati. A good starting point for developing

such an extension would be the research by Carleen

Hutchins (Hutchins, 1981).

The remainder of the paper is structured as fol-

lows. Section 2 discusses related work. In Section 3

we describe the main internal structures of our design

tool. Section 4 describes our experiment and we close

the paper with Section 5 where we discuss conclu-

sions and further work.

2 RELATED WORK

2.1 Creative Evolutionary Systems

A creative evolutionary system is a computer system

that makes use of some aspect of evolutionary compu-

tation to explore creative solutions (Bentley, 1999a).

This kind of system is designed to:

• Aid our own creative processes, and/or

• to generate results to problems that traditionally

Wang, H. and Hamel, L.

An Intelligent Design Explorer for New Violin Shapes.

DOI: 10.5220/0008353102290236

In Proceedings of the 11th International Joint Conference on Computational Intelligence (IJCCI 2019), pages 229-236

ISBN: 978-989-758-384-1

229

required creative people to ﬁnd solutions (Bentley

and Corne, 2002).

There are ﬁve elements that make up a creative

evolutionary system framework:

• An evolutionary algorithm

• A genetic representation

• An embryogeny

• A phenotype representation

• Fitness function(s) and/or processing of user in-

put (Bentley and Corne, 2002)

A creative evolutionary system uses some kind of

evolutionary algorithm to create new offspring. The

genotype representation is a series of elements ab-

stracted from real-world objects. The evolutionary al-

gorithm modiﬁes the genotype representation through

recombination and mutation processes to generate a

new genotype representation. Once the new geno-

type is generated, an embryogeny process decodes the

genotype representation to construct the correspond-

ing phenotype. Finally, a ﬁtness function or human

selection process will help users select an optimal so-

lution to create the solutions that make up the next

generation. The difference between a creative evo-

lutionary system and standalone evolutionary algo-

rithms is that an evolutionary algorithm is used to

search for an optimal solution whereas a creative evo-

lutionary system acts as an explorer to provide inspi-

ration and investigate creative solutions.

2.2 Evolutionary Computation in

Engineering Design

As one of the most well known evolutionary algo-

rithms, the genetic algorithm (Kramer, 2017), is now

a well-established technique in engineering design

applications. For example, in 1992 John Frazer and

his students used a genetic algorithm in yacht hull de-

sign (Frazer, 2002). Their program was intended to

explore a range of widely different alternative solu-

tions to yacht hull design (Figure 1). The program

starts by generating a randomized initial population

of yacht hull designs. All designs are evaluated by

a ﬁtness function which considers both objective and

subjective factors. The objective engineering factors

include parameters such as stability, center of buoy-

ancy, wetted surface area, prismatic coefﬁcient, and

blocking. The subjective factors (i.e. designer crite-

ria) mainly considered aesthetic appearance, historic

tradition, and allusion of form.

On the basis of the ﬁtness function, the designs

with the highest ﬁtness scores are selected and be-

come the parent population. The crossover and mu-

Figure 1: Yacht hull designs (Frazer, 2002).

tation processes are then applied to the parent popula-

tion and produce new offspring. These processes con-

tinue until a satisfactory design emerges. The yacht

hulls design program produces a range of choices for

clients and provides a series of new ideas in yacht hull

designs to explore.

2.3 Evolutionary Computation in Art

Compared with engineering design, music and art are

more subjective, fuzzy and difﬁcult to analyze. But it

can still be evolved through evolutionary computing.

The algorithmic art assembly (Algorithmic Art As-

sembly, 2019) hosted in San Francisco in 2019 show-

cased a diverse range of artists who are using algo-

rithmic tools and processes in their works. Although

the algorithmic tools are not necessarily evolutionary

in nature it points to a new approach of using com-

putational tools in music and art generation. Darwin-

Tunes (MacCallum et al., 2012) is a computational

engine to evolve music. As shown in Figure 2 it uses a

genetic algorithm to maintain a population of tree-like

digital genomes, each of which encodes a computer

program representing a song. DarwinTunes uses re-

combination and mutation to randomly create songs

with novel musical motifs, rhythms, and harmonies.

Instead of using a ﬁtness function, user selection via

a web interface is the only way to help DarwinTunes

select the parent individuals for the next generation.

One of the best known “evolutionary artists”,

Steven Rooke, uses genetic programming algorithms

to evolve some astonishing pieces of art (Rooke,

2002). The image evolution system begins with a

series of randomly generated tree structure genomes.

Based on these genomes the computer generates cor-

responding images according to speciﬁc mapping

rules. In each generation, the image evolution sys-

tem determines the parent genomes by random selec-

tion or user selection. The parent genomes generate

new offspring through crossover and mutation steps.

Figure 3 shows some of the images generated by the

image evolution system.

ECTA 2019 - 11th International Conference on Evolutionary Computation Theory and Applications

230

Figure 2: The DarwinTunes program (MacCallum et al.,

2012).

Figure 3: Some creative outcomes of Rooke’s image evolu-

tion system (Rooke, 2002).

3 THE DESIGNER

Our violin design explorer is conceived as a creative

evolutionary system which can help users without

formal design background create their favorite violin

outline design.

3.1 The Digital Amati Project

Digital Amati (Digital Amati, 2018) developed by

Harry Mairson is a software package based on Eu-

clidean geometry and uses a “geometry engine” lan-

guage (AmatiML) to design classical stringed instru-

ments. The AmatiML language deﬁnes a series of ge-

ometric elements such as points, line segments, and

circles; and uses a series of inscribed circles and re-

verse curves formed by these geometric elements to

draw a violin outline (Figure 4).

In the violin design explorer all violin drawing

work is handled by the Digital Amati software.

Figure 4: A violin outline drawn by Digital Amati.

Figure 5: The scope of points.

3.2 Representation

3.2.1 Genetic Representation

In the violin design explorer the genotype represen-

tation consists of an array of point locations (details

follow below). Each array corresponds to a single vi-

olin outline. Each element in the genotype represen-

tation array has a scope. For example, consider points

a and b that are the centers of two circles (Figure 5).

Point g is the intersection of these two circles. Here

we can interpret point g as the apex of the violin’s

upper left sharp corner. As the positions of points a

and b change the sharp corner of the violin changes.

When the positions of points a and b are far enough

apart the intersection (point g) of the two circles dis-

appears. Therefore, point g has a scope attached in

order to ensure that the two circles intersect. A scope

can be interpreted as a range of values that satisﬁes

certain criteria.

Interestingly, in Digital Amati the scopes for some

points can be discontinuous. For example, the scope

of point P (see Table 1) is 6-9 and 11-13 written in

interval notation as [6,9] and [11,13]. The violin de-

sign explorer manages these discontinuities by using

ﬁve different models or search spaces for violin out-

lines. Table 1 summarizes the different models and

their point scopes used by the violin explorer.

In the following we brieﬂy highlight the function

An Intelligent Design Explorer for New Violin Shapes

231

Table 1: The violin explorer genotype representation models.

Model Genotype Scopes

Model 1 [q,O,e,c] q:[1,8] O:[4,5] e:[6,10] c:[2,6]

Model 2 [q,O,e,c,P] q:[2,4] O:[3,5] c:[2,6] P:[6,9] e:[8,10]

Model 3 [q,O,e,c,P] q:[2,4] O:[3,5] c:[2,6] P:[11,13] e:[8,10]

Model 4 [q,O,e,c,P] q:[5,8] O:[3,5] c:[2,6] P:[6,9] e:[8,10]

Model 5 [q,O,e,c,P] q:[5,8] O:[3,5] c:[2,6] P:[11,13] e:[8,10]

Figure 6: The effect of the point q value.

Figure 7: The effect of the point O value.

of each point in the genotype representation. The

value of point q controls the width of the upper bouts

of a violin. As the value of point q increases, the up-

per bouts of the violin design will become wider. As

shown in Figure 6, the point q value of the left design

is 2 and the right design is 8.

The value of point O controls the smoothness of

the upper bouts of a violin. When the value of point O

increases, the upper bout of a violin is smoother. For

example, there are three upper bouts design in Fig-

ure 7. The point O value of the left design is 3. The

point O value of the middle design is 5. The point O

value of the right design is 6.

The value of point e controls the width of the mid-

dle bouts of a violin. As the value of point e increases,

the middle bouts of the violin will become narrower.

As shown in Figure 8, the point e value of the left de-

sign is 6 and the point e value of the right design is

10.

The value of point c controls the shape of the

lower bouts. As the value of point c increases, the cor-

ner of the violin’s lower bouts become much sharper.

As shown in Figure 9, the point c value of the left de-

sign is 2 and the point c value of the right design is

The value of point P also affects the shape of a

violin’s lower bouts. As the value of point P increases,

the violin’s lower bouts become taller and rounder. As

shown in Figure 10, the point P value of the left design

is 6 and the point P value of the right design is 8.

Figure 8: The effect of the point e value.

Figure 9: The effect of the point c value.

3.2.2 Embryogeny

An embryogeny is a mapping process from geno-

type to phenotype (Bentley and Corne, 2002). The

embryogeny process of the violin explorer is imple-

mented through the Digital Amati software package.

The Digital Amati package uses the AmatiML lan-

guage to deﬁne a series of geometric elements such as

points, lines, and circles; and use a series of inscribed

circles and reverse curves formed by these geometric

elements to draw a violin outline. Each component in

the genotype is used to construct part of a particular

violin outline as described in in the previous section.

3.2.3 Phenotype Representation

A phenotype representation in the violin design

explorer is a violin outline drawn by the Digi-

tal Amati software using the Retroﬁt and Batik li-

braries (Retroﬁt 2, 2019; Apache

Batik SVG

Toolkit, 2019), as shown in Figure 11.

These phenotype representations are treated as so-

lutions to violin outline designs and are evaluated by

the users of the system.

3.3 The Core Evolutionary Algorithm

The violin design explorer uses a standard evolution-

ary strategies algorithm (Beyer, 2013; Hansen et al.,

2015), as outlined in the following:

Step 1: The violin design explorer is initialized by

randomly generating 10 violin outlines. These vi-

ECTA 2019 - 11th International Conference on Evolutionary Computation Theory and Applications

232

Figure 10: The effect of the point P value.

Figure 11: The phenotype representation of the violin ex-

plorer.

olin outlines are treated as the original population

pool.

Step 2: Users are asked to select two of the most sat-

isfying designs from the population pool.

Step 3: These two designs will become the parent

population. Offspring are introduced through re-

combination. All offspring are placed into the

population pool until it reaches the maximum

population; in our case 10 violin outlines.

Step 4: Mutation is applied to each single violin out-

line in the population pool based on values in

strategy parameters of the system. The strategy

parameters themselves are also mutated within the

overall mutation process.

Step 5: Once recombination and mutation have gen-

erated enough new individuals to ﬁll the popula-

tion pool, users will be asked to select two of the

most satisfying designs from the population pool.

These two designs will be treated as the parent

population for the next generation.

Step 6: Repeat Step 3 through Step 5 to generate new

generations of violin designs.

This evolutionary process terminates after it

reaches a predeﬁned number of generations or when

users are satisﬁed with the current design. In our case

Figure 12: Recombination process.

we limit the number of generations to 50 across all 5

models.

3.3.1 Evolutionary Strategies Scheme

As we have seen, the violin design explorer uses a

(µ, λ)-ES scheme to create new individuals in each

generation. Here µ is the number of parents and λ

is the number of offspring (Beyer, 2013). We chose

the values µ = 2 and λ = 10 according to the recom-

mendations by Hanson et al. (Hansen et al., 2015).

According to Hanson et al. the optimal truncation

ratio is µ/λ = 0.27 with a corresponding maximized

progress per offspring of 0.202. Our choice of the

values for µ = 2 and λ = 10 gets us close to the opti-

mal truncation ratio, keeps things straight forward in

terms of computation, and yet still maintain a man-

ageable number of designs for the user to process.

3.3.2 Recombination

Recombination is applied at each generation cy-

cle which recombines parent individuals in different

ways to create new offspring. The violin explorer

deﬁnes a breakpoint integer bp which is randomly

generated and designed to determine the position to

break a genotype into two pieces. For example, when

bp = 3 both parents are divided into two pieces from

the 3rd position (Figure 12) and the resulting pieces

are recombined to form offspring.

According to our overall algorithm the recombi-

nation process will repeat until sufﬁcient offspring are

generated.

3.3.3 Mutation

Mutation plays an important role in evolutionary strat-

egy in order to enhance the overall search for pos-

sible solutions (Bentley, 1999b). The mutation pro-

cess in the violin design explorer involves making

small changes to each genotype element. The value of

change is determined by the mutation strength param-

eter. The mutation strength parameter itself is an array

that is randomly generated when the original popula-

tion is initialized. This array has the same length as

An Intelligent Design Explorer for New Violin Shapes

233

the genotype representation array and each array ele-

ment within the mutation strength array corresponds

to a position in the genotype representation. The mu-

tation strength of each element is used to produce ran-

dom changes in the corresponding genotype element

according to a Gaussian distribution with a mean of

zero. During the evolutionary process the mutation

strength is reduced in order to foster a better conver-

gence behavior.

3.3.4 Solution Evaluation

In our currently implementation violin outline design

is considered an aesthetic problem. It is almost an

impossible task to establish a ﬁtness function to cap-

ture aesthetics. Many factors such as a user’s age and

background, as well as current fashion trends may be-

come key factors inﬂuencing violin design. Therefore

the violin design explorer uses human interaction pro-

cesses instead of using a ﬁtness function to evaluate

violin outline designs.

4 EXPERIMENT

The primary objective of our experiment is to vali-

date the violin design explorer. Four distinct groups of

participants were asked to design their favorite violin

outline. We were working under the hypothesis that if

we see distinct design patterns in the four participant

groups then our violin design explorer works as envi-

sioned. The design patterns that emerged within these

four participant groups during the experiment seem to

validate that our design tool works as we envisioned.

4.1 Experiment Setup

Our experiment population consisted of 100 volun-

teers of different ages and backgrounds. The volun-

teers were asked two questions to determine their ex-

perimental grouping:

• Which age group are you in? “18 to 40” or “older

than 40”?

• Do you have an artistic background such as paint-

ing or playing musical instruments?

Based on the participants’ answers to these two ques-

tions they were divided into four experimental groups

each with 25 participants as shown in Table 2.

During the experiment volunteers used the violin

design explorer to design their favorite violin outlines.

Our analysis of the data often makes reference to the

“classic” or “original” violin design. By this we mean

a design as established by the old Italian masters in

Figure 13: The designs of group 1.

Figure 14: The designs of group 2.

the 1600’s. Figure 11 shows an image of this refer-

ence design.

4.2 Results

The volunteers in group 1 were between 18 and 40

years old and had an artistic background. The 25 vio-

lin designs of that group are shown in Figure 13.

The volunteers in group 2 were between 18 and

40 years old and had no artistic background. The 25

violin designs in that group are shown in Figure 14.

The volunteers in group 3 were older than 40 years

and had an artistic background. The 25 violin designs

in that group are shown in Figure 15.

The volunteers in group 4 were older than 40 years

and had no artistic background. The 25 violin designs

in that group are shown in Figure 16.

We averaged the genotype representations within

ECTA 2019 - 11th International Conference on Evolutionary Computation Theory and Applications

234

Table 2: Four experimental groups.

Group 1 Participants are between the ages of 18 to 40 and have an artistic background.

Group 2 Participants are between the ages of 18 to 40 but do not have any artistic background.

Group 3 Participants are older than 40 and have an artistic background.

Group 4 Participants are older than 40 but do not have any artistic background.

Figure 15: The designs of group 3.

Figure 16: The designs of group 4.

Figure 17: The average designs for the four groups.

each experimental group in order to compute the “av-

erage” design for each group and these designs are

shown in Figure 17. A visual inspection of these av-

erage designs reveals that the average design of group

3 (older than 40 with an artistic background) is most

similar to the classic violin design of the old Italian

Table 3: The design variance in each group.

Group Design Variance

1 1.93

2 2.29

3 1.41

4 1.94

Table 4: Design deviation from the classic design.

Group Design Deviation

1 1.94

2 2.37

3 1.35

4 2.20

masters in Figure 11.

An interesting metric to consider is the design

variance in each group. The design variance is com-

puted by computing the parameter variances in each

group and then averaging those variances for each

group. The design variance for each group is shown

in Table 3. We observe that groups 1 and 3 have the

lowest design variances indicating that users with an

artistic background have similar notions of optimal

designs. The lowest design variance is in group 3 indi-

cating the older users with an artistic background tend

to have very similar ideas about violin designs. The

highest design variance is in group 2 indicating that

younger users without any artistic background are the

most likely to experiment with violin designs.

The design deviation from the classic design in

Figure 11 in each group can also be considered as a

metric. We computed the design deviation by com-

puting the standard deviation of the designs in each

group from the classic design. Table 4 shows the

design deviations in each group. Perhaps not unex-

pected we ﬁnd that groups 1 and 3 have the lowest

deviation from the classic design. We also ﬁnd that

group 3 (older users with an artistic background) have

the overall lowest deviation from the classic design

validating our casual observation from earlier that the

designs in this group are most similar to the classic

violin design. Finally, we ﬁnd that users in group 2

(younger users without an artistic background) have

the highest design deviation further validating the fact

that users in this group are the most likely to experi-

ment with violin designs.

Given these not unexpected design patterns within

An Intelligent Design Explorer for New Violin Shapes

235

each of the groups we feel that this validates the work-

ings of our violin design explorer.

5 CONCLUSIONS AND FURTHER

WORK

Here we discussed an evolutionary design tool, the

violin design explorer, for violin outline design. With

this system, users without any formal design back-

ground can create their favorite violin outline through

a set of simple choices. Our design tool employs an

evolution strategies algorithm and can be viewed as

a creative evolutionary system. The genotype repre-

sentation of the system directly supports violin out-

line manifestation through the Digital Amati software

library. In order to validate the design tool we con-

ducted an anonymous experiment. Four groups of

participants were asked to design their favorite vio-

lin outline. Design patterns that emerged within each

of these four participant groups during the experiment

seem to validate that our design tool works as we en-

visioned.

Currently we partition the violin outline search

space into ﬁve parts via the ﬁve models driven by pa-

rameter scopes. We would like to explore other ways

of structuring the search space and therefore provide

potential solutions to the users in different ways. Fur-

thermore, in our current implementation we use the

default violin template and its associated parameters

provided in Digital Amati as a starting point for the

evolution of new outlines. This completely constrains

the user to stay within the somewhat classical look

of a violin with upper, lower, and “C” bouts, etc.

It would be interesting to experiment with different

kinds of templates as starting points providing the

user with a larger range of possible outline designs.

Here we only explored one aspect of violin de-

sign – the aesthetics of the violin outline. Our goal is

to provide comprehensive interactive design tool for

a fully functional violin. In order to accomplish this

we will need to address body cavity design as well

as neck, sound hole and bridge designs. This means

we have to develop models for these violin parts in or-

der to provide constrained parameter spaces that make

sense for violin design similar to the scopes we saw

in the outline design. A good starting point for devel-

oping these kind of models is the research by Carleen

Hutchins (Hutchins, 1981).

Ultimately we would like to be able to generate

designs that are sufﬁciently detailed so that they can

be directly submitted to a CNC (Xu and Newman,

2006) or other milling machine for construction.

REFERENCES

Algorithmic Art Assembly (2019). https://aaassembly.org.

Apache

Batik SVG Toolkit (2019). https://xmlgraphics.

apache.org/batik.

ack, T., Fogel, D. B., and Michalewicz, Z., editors (2018).

Evolutionary computation 1: Basic algorithms and

operators. CRC Press.

Bentley, P. (1999a). Evolutionary design by computers.

Morgan Kaufmann.

Bentley, P. (1999b). An introduction to evolutionary de-

sign by computers. Evolutionary design by computers,

pages 1–73.

Bentley, P. J. and Corne, D. W. (2002). An introduction to

creative evolutionary systems. In Creative evolution-

ary systems, pages 1–75. Elsevier.

Beyer, H.-G. (2013). The theory of evolution strategies.

Springer Science & Business Media.

Digital Amati (2018). http://www.digitalamati.org.

Frazer, J. (2002). Creative design and the generative evo-

lutionary paradigm. In Creative evolutionary systems,

pages 253–274. Elsevier.

Hansen, N., Arnold, D. V., and Auger, A. (2015). Evolution

strategies. In Springer handbook of computational in-

telligence, pages 871–898. Springer.

Hutchins, C. M. (1981). The acoustics of violin plates. Sci-

entiﬁc American, 245(4):170–187.

Kramer, O. (2017). Genetic algorithm essentials, volume

679. Springer.

MacCallum, R. M., Mauch, M., Burt, A., and Leroi,

A. M. (2012). Evolution of music by public choice.

Proceedings of the National Academy of Sciences,

109(30):12081–12086.

Menges, A. and Ahlquist, S., editors (2011). Computational

design thinking: computation design thinking. John

Wiley & Sons.

Retroﬁt 2 (2019). https://github.com/square/retroﬁt.

Rooke, S. (2002). Eons of genetically evolved algorithmic

images. In Creative evolutionary systems, pages 339–

365. Elsevier.

Xu, X. W. and Newman, S. T. (2006). Making cnc machine

tools more open, interoperable and intelligent—a re-

view of the technologies. Computers in Industry,

57(2):141–152.

ECTA 2019 - 11th International Conference on Evolutionary Computation Theory and Applications

236