Professional Documents
Culture Documents
PDF Design Computing and Cognition 18 John S Gero Ebook Full Chapter
PDF Design Computing and Cognition 18 John S Gero Ebook Full Chapter
PDF Design Computing and Cognition 18 John S Gero Ebook Full Chapter
John S. Gero
Visit to download the full and correct content document:
https://textbookfull.com/product/design-computing-and-cognition-18-john-s-gero/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
https://textbookfull.com/product/making-sense-cognition-
computing-art-and-embodiment-simon-penny/
https://textbookfull.com/product/cognitive-computing-in-human-
cognition-perspectives-and-applications-pradeep-kumar-mallick/
https://textbookfull.com/product/on-being-and-cognition-
ordinatio-1-3-1st-edition-john-duns-scotus/
https://textbookfull.com/product/mass-spectrometry-based-
proteomics-methods-in-molecular-biology-2718-kris-gevaert/
Understanding the Brain From Cells to Behavior to
Cognition John E. Dowling
https://textbookfull.com/product/understanding-the-brain-from-
cells-to-behavior-to-cognition-john-e-dowling/
https://textbookfull.com/product/the-adaptable-mind-what-
neuroplasticity-and-neural-reuse-tells-us-about-language-and-
cognition-john-zerilli/
https://textbookfull.com/product/design-for-good-john-cary/
https://textbookfull.com/product/five-design-sheets-creative-
design-and-sketching-for-computing-and-visualisation-1st-edition-
jonathan-c-roberts/
https://textbookfull.com/product/adventure-diffusion-from-
meandering-molecules-to-the-spreading-of-plants-humans-and-ideas-
gero-vogl/
John S. Gero Editor
Design
Computing
and
Cognition '18
Design Computing and Cognition ’18
John S. Gero
Editor
Design Computing
and Cognition ’18
123
Editor
John S. Gero
Department of Computer Science
and School of Architecture
University of North Carolina at Charlotte
Charlotte, NC, USA
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
The Second World War demonstrated that scientific research could be readily
translated into technology. This opened up a much larger ambit of research than
what had existed previously because of the change in its perceived value. The
launch of the Sputnik satellite in 1957 by the Soviet Union prompted a reexami-
nation of research in the West resulting in the establishment of the National Science
Foundation and the Defense Advanced Research Agency in the US, which pro-
vided, and continue to provide, significant research funding to universities. Multiple
disciplines that had not seen research as part of their activities began to view
research, and in particular scientific research, as potentially beneficial. Design was
one of these disciplines although the transfer of research to industry is not as well
developed in design as in many other disciplines.
Design, when viewed from inside a discipline, often appears to be very narrowly
concentrated within that discipline with little in common with design in other
disciplines. This limited focus has made research into design less attractive as a
field in its own right rather than being connected to any particular discipline.
Further, the claim that the products of design are unique has been used to support
claims that design cannot be studied scientifically because of its inherent lack of
reproducibility. Science is built on the notion of regularities in phenomena with its
implication of reproducibility. The regularities in designs and designing come from
the structure of what has been designed and the processes used in their design,
which make it open to scientific study. This division between designs and designing
also matches computation well with its foundational ontology of representation and
process.
The scientific study of designs and designing is often conflated with the notion
that this makes designing scientific. Design science, a term coined by Buckminster
Fuller in 1957, has come to denote three different streams of thought. Nigel Cross
has called them scientific design, design science and a science of design. All three
are represented by research in this volume, although the primary focus is on design
science as the scientific study of design.
v
vi Preface
Design research treats designs and designing as a separate area from science and
the humanities to produce a third domain of human endeavor: designing is
designing. It borrows tools and techniques from other disciplines in the sciences
and the humanities but remains distinct from them. The three waves of approaches
to studying designing has been: formalization (through mathematics, logic and
artificial intelligence), understanding the designer’s mind while designing through
cognitive science, and most recently understanding the designer’s brain while
designing through cognitive neuroscience.
The papers in this volume are from the Eighth International Conference on
Design Computing and Cognition (DCC’18) held at the Polictecnico di Milano,
Lecco Campus, Italy. They represent the state of the art of research and develop-
ment in design computing and design cognition including the nascent area of design
cognitive neuroscience. They are of particular interest to design researchers,
developers and users of advanced computation in design as well as to design
educators. This volume contains knowledge about the cognitive behavior of
designers, which is valuable for those who need to gain a better understanding of
designing.
In these proceedings the papers are grouped under the following nine headings,
describing both advances in theory and application and demonstrating the depth and
breadth of design computing and design cognition:
New Design Methods
Design Cognition—Design Approaches
Design Synthesis
Design Theory
Design Cognition—Design Behaviors
Design Grammars
Design Processes
Design Modeling
Design and Visualization
A total of 103 full papers were submitted to the conference, from which 40 were
accepted and appear in these proceedings. Each paper was extensively reviewed by
at least three reviewers drawn from the international panel of reviewers listed on the
following pages. The reviewers’ recommendations were then assessed before the
final decision on each paper was taken. The authors improved their contributions
based on the advice of this community of reviewers prior to submitting the final
manuscript for publication. Thanks go to the reviewers, for the quality of these
papers depends on their efforts. Special thanks to Sarah Abdellahi who helped put
the volume together.
vii
viii List of Reviewers
xi
xii Contents
Christopher McComb
The design of a system commits a significant portion of the final cost of that sys-
tem. Many computational approaches have been developed to assist designers in the
analysis (e.g., computational fluid dynamics) and synthesis (e.g., topology optimiza-
tion) of engineered systems. However, many of these approaches are computationally
intensive, taking significant time to complete an analysis and even longer to itera-
tively synthesize a solution. The current work proposes a methodology for rapidly
evaluating and synthesizing engineered systems through the use of deep neural net-
works. The proposed methodology is applied to the analysis and synthesis of offshore
structures such as oil platforms. These structures are constructed in a marine envi-
ronment and are typically designed to achieve specific dynamics in response to a
known spectrum of ocean waves. Results show that deep learning can be used to
accurately and rapidly synthesize and analyze offshore structure.
Introduction
A significant amount of the final cost of a system is committed during design. Accord-
ing to the situated function–behavior–structure design framework, design consists of
navigating from the requirements for a solution to the documentation of that solution
[1, 2]. This process entails negotiating through several ontological categories, includ-
ing function, expected behavior, derived behavior, and structure. The focus of this
paper is on the tasks of analysis (deriving behavior from structure) and synthesis (gen-
erating a structure based on desired behavior). Many computational approaches have
been developed to assist designers in the analysis and synthesis of engineered sys-
tems (e.g., computational fluid dynamics and topology optimization, respectively).
C. McComb (B)
The Pennsylvania State University, University Park, PA, USA
e-mail: uum209@psu.edu
Background
Artificial neural networks (referred to in the remainder of this paper simply as neural
networks) are computational systems that are analogous to the biological neural
networks that make up nervous tissue and animal brains. Neural networks can be
trained to accomplish a variety of complex tasks, including regression, classification,
and feature extraction. Jain et al. provide a more detailed introduction to neural
networks [6]. Deep learning, which is the focus in this work, refers specifically to
neural networks that have more than one hidden layer.
Neural networks have shown significant success in two-dimensional image recog-
nition tasks. This success has led researchers to apply similar methodology to three-
dimensional recognition tasks [7], facilitated by recent advances in computing that
Toward the Rapid Design of Engineered Systems Through Deep … 5
enable such tasks to be performed at scale. Seminal dataset and classification efforts
include ObjectNet3D [8], ShapeNet [9], VoxNet [10], and PointNet [11]. The auto-
mated synthesis of three-dimensional objects is still a nascent field in machine learn-
ing. Most approaches focus on creating objects with a given form and category (e.g.,
[12, 13]), rather than attempting to derive a deeper relationship between desired
functionality and requisite form.
This work also makes use of autoencoders. These are specially designed neural
networks that take an input, map it into a space with reduced dimensionality, and then
output a reconstructed version of the input [14]. The two halves of the neural network
(the encoder and the decoder, respectively) can then be used for specific and useful
functions. The encoder can map an input into a reduced space, essentially performing
data compression, while the decoder can take compressed values and reconstruct an
output. This work uses variational autoencoders which map the input into a space
of latent variables so that the training data are normally distributed [15]. This is
accomplished by training the network with a loss function that measures reconstruc-
tion accuracy as well as how normally distributed the parameters in the maximally
compressed layer are (typically Kullback–Leibler divergence). This ensures that the
variables in the latent space are rich in information. Variational autoencoders have
been used to compress a wide variety of different data, including human faces [16],
handwritten numbers [17], and house numbers [18].
Neural networks of many varieties have been utilized in design and engineering
to accomplish various tasks. For instance, Tseng, Cagan, and Kotovsky utilized a
neural network to learn the preferences of a customer and then utilized that neural
network as the objective function for a genetic algorithm [19]. Dering and Tucker
utilized convolutional neural networks to predict the function of a product from its
form alone [20]. The utilization of deep learning, and specifically autoencoders,
also led to the creation of a computational framework that models the curiosity of
a given user in order to provide surprising examples [21]. Neural networks have
also been utilized to automatically predict quality defects in automotive parts [22]
and to support design for additive manufacturing [23–25]. These examples, while
not exhaustive, serve to highlight potential utility of neural networks for design and
the need for a standardized approach to implementing them. The current research
utilizes a generic, voxel-based approach for describing potential design solutions,
thus ensuring significant representation flexibility.
Offshore Structures
Offshore structures are comprised of buoys, drilling platforms, and wave energy
converters (WECs). WECs are an increasingly common type of offshore structure
that are designed to extract energy from ocean waves. WECs may serve an important
role in the future of humankind, since it is estimated that approximately 3.7 TW (3.7
trillion Watts) of power can be harvested from the world’s oceans [26]. However,
6 C. McComb
in order to access that power, several challenges in the design of WECs must be
overcome [27].
A substantial array of numerical methods have been developed for the simulation
of offshore structures, including analytical methods, empirical methods, Navier–S-
tokes equation methods, and boundary-integral equation methods [28]. Analytical
methods offer quick and rough estimates for devices with simple geometry, while
most empirical methods attempt to maintain simplicity while making use of experi-
mental values to increase accuracy. Navier–Stokes equation methods (NSEMs) can
resolve highly nonlinear phenomena, but generally do not permit closed-form solu-
tions, requiring the use of computational fluid dynamics.
Boundary-integral equation methods (BIEMs) are the focus of this work, as they
are the industry standard for design and analysis of offshore structures. BIEMs pro-
duce a potential flow solution in the frequency domain [28]. This means that the
outputs are given as spectra that indicate how much force, damping, or other quan-
tities are applied to an offshore structure for incoming ocean waves with varying
frequencies. Although they are far less computationally expensive than NSEMs,
producing a full BIEM solution for a model with a high mesh resolution can still
take hours. In addition, pure frequency-domain BIEMs are only weakly nonlinear
[28] which makes it impossible to directly implement nonlinear control strategies
within the simulation. One way to overcome this limitation is by numerically inte-
grating over several frequencies of the BIEM solution to yield a time-varying series
for different fluid phenomena [29]. These series can then be applied in an appropri-
ate 6 degree-of-freedom (6DOF) solver to produce a time-domain simulation, from
which important metrics such as average power production can be computed.
The application of the current work focuses on predicting frequency-varying wave
force spectra as a function of structure geometry (and vice versa). It should be noted
that a similar methodology could be applied directly to other frequency-varying fluid
phenomena that are produced by a BIEM solution. The results of the current work
could be integrated into software packages that utilize the BIEM + 6DOF approach
outlined above [30, 31]. This would enable rapid exploration of conceptual solutions,
either by human designers and engineers or by agent-based design algorithms [4, 5,
32].
Methodology
The approach that is proposed in this work for rapidly evaluating and synthesizing
engineered systems can be broken into three steps. This process is depicted graph-
ically in Fig. 1. First, a diverse set of examples of a given system type must be
generated, and the performance of each example must be analyzed using current
methods (finite element analysis, computational fluid dynamics, experimental test-
ing, etc.). The second step entails training two autoencoders: one for the engineered
system and one for the performance assessment. The third and final step in the pro-
posed methodology is the recombination of performance and system autoencoders
Toward the Rapid Design of Engineered Systems Through Deep … 7
into two new, deep networks that are possible of accomplishing rapid analysis of
a system (encoding the system, decoding performance) as well as synthesis of a
system according to desired performance (encoding performance, and decoding sys-
tem). Both of these new networks should have one or more new layers that must
be trained between the encoder and the decoder, enabling a mapping between the
latent system space and the latent performance space (or vice versa). The use of
autoencoders is critical as it permits the learning of synthesis and analysis in the
latent space which has fewer dimensions (and thus less complexity) than the input
or output.
The current work shares the application of the above methodology to analysis
and synthesis of offshore structures. First, NEMOH, a BIEM solver, was used to
simulate thousands of different floating body geometries, deriving frequency-varying
response forces for each [35]. Next, the data generated in NEMOH was used to
train two variational autoencoders, one of which modeled key features of frequency-
varying response forces and the other modeled key geometric features of the input
geometries. Finally, these autoencoders were used to instantiate two networks: one
for predicting the force spectra of known geometries (analysis), and the other for
generating geometries for a known force spectrum (synthesis). All neural networks
were trained using the Keras neural network API [33] in conjunction with the Theano
library [34]. A full implementation of this work, including training data, is available
in the Python language under an MIT License.1
Data Generation
The dataset used in this work was generated by instantiating 5000 different common
shapes for offshore structures. These included wedges, hemispheres, cylinders, rect-
angular prisms, and cones. All shapes were generated to fit within a 10 m × 10 m ×
10 m bounding box. These offshore structure shapes were then analyzed using the
NEMOH BIEM solver [35], producing frequency-varying spectra describing the
1 https://github.com/HSDL/WAnet/releases/tag/v1.0-beta.
8 C. McComb
forces applied to the structure both parallel and perpendicular to the direction of the
incoming ocean waves (commonly referred to as heave and surge, respectively), as
well as a moment about the center of gravity of the body (referred to as pitch). An
example of these spectra is provided in Fig. 2a.
The geometry for the offshore structures was originally provided to NEMOH as
a mesh. The meshes were converted into a voxel-based format in order to make the
geometry data more accessible to the proposed neural network approach. Voxels
are a three-dimensional analogue of pixels. Specifically, the bounding box for the
offshore structure was discretized into a 32 × 32 × 32 grid, containing 32,768 voxels.
The voxel values in this grid were defined as 1 (if the structure occupied part of the
voxel) of 0 (if the structure did not occupy part of the voxel).
Thus, the final dataset consists of paired geometry-spectra observations. An exam-
ple of a paired observation is provided in Fig. 2. Plots of force response spectra and
voxelized geometries in the remainder of the paper will omit axis labels and scales
in the interest of clarity and concision. This dataset was randomly separated into a
training set (80% of the data, 4000 observations) and a testing set (20% of the data,
1000 observations). All accuracies reported in the remainder of this paper correspond
to measurements on the testing set.
Two variational autoencoders were trained based on the data generated with NEMOH.
The structure of these autoencoders is shown in Figs. 3 and 4. Both variational
autoencoders are designed to compress the input data into an N-dimensional latent
space, which describes the number of nodes in the smallest layer. The value of N is
identified through a parametric search, detailed in the results section of this paper.
Toward the Rapid Design of Engineered Systems Through Deep … 9
Both autoencoders were trained using the root mean square propagation
(RMSprop) algorithm [36]. The primary term in the loss function of the force spec-
tra autoencoder was based on mean squared error while the primary term for the
geometry autoencoder was based on binary cross-entropy. Both training algorithms
also included a term for Kullback–Leibler divergence [37] of the values in the latent
space (the innermost hidden layer) in the loss function. The computation of the loss
functions in this way is standard for variational autoencoders.
It should be noted that the dimensionality of the latent space for both the spectrum
autoencoder and the geometry autoencoder is described by a single variable, N,
despite the fact that the force spectrum is much simpler than the structure geometry.
This is an intentional decision, as equating the dimensionality of the spaces makes
it more likely that a one-to-one mapping can be found between them.
The network designed for evaluating offshore structures utilizes the geometry
encoder and the spectrum decoder (see Fig. 5). A layer of N nodes was included
between these two elements, and this interior layer was the only layer that was
trained. In essence, this network compresses the geometry of a structure into the N-
dimensional geometry latent space (using the geometry encoder), maps the geometry
latent space into the spectrum latent space (this is the trainable N-node layer), and then
reconstructs the full force spectra, producing the desired output (using the spectra
decoder).
The network designed for synthesis of offshore structure utilizes the spectrum
encoder and the geometry decoder with a trainable N-node layer in between the two
(see Fig. 6). This network compresses the spectra into the N-dimensional spectra
latent space (using the spectra encoder), maps that into the N-dimensional geometry
latent space (through trainable layer), and then reconstructs the full geometry (using
the geometry decoder).
Both analysis and the synthesis networks were trained using the RMSprop algo-
rithm [36]. The primary term in the loss function of the analysis network was based
on mean squared error (since the output was a set of real-valued curves) while the
primary term for the geometry autoencoder was based on binary cross-entropy (since
the output was a set of voxel data with values between 0 and 1).
Toward the Rapid Design of Engineered Systems Through Deep … 11
0.96
Spectrum Autoencoder
0.94 Geometry Autoencoder
Coefficient of Determination
Analysis Network
0.92 Synthesis Network
0.9
0.88
0.86
0.84
0.82
0.8
2 4 8 16 32
Dimensionality of Latent Space
Results and discussion are provided in three subsections. The first subsection details
a parametric study that was used to select the appropriate dimensionality for the
latent space. The second reports the results and examples for the autoencoders (both
geometry and spectrum). These autoencoders are critical as they permit the tasks
of synthesis and analysis to be learned in a space of reduced complexity. The third
subsection does the same for the recombined analysis and synthesis networks.
In order to determine the appropriate dimensionality for the latent space, a parametric
study was conducted. All four networks used here (the spectrum autoencoder, the
geometry autoencoder, the analysis autoencoder, and the synthesis autoencoder) were
trained for increasing values of the dimensionality of the latent space, N. The results
of this study are provided in Fig. 7. The horizontal axis shows dimensionality of the
latent space and the vertical axis shows network validation accuracy (specifically,
the percentage of variance explained by the trained network).
The data for the spectrum autoencoder is relatively flat, indicating that a small
number of latent dimensions are sufficient to accurately reconstruct that data. The
geometry autoencoder, in contrast, shows consistently increasing accuracy up to
approximately 16 dimensions, at which point it begins to decrease. The analysis
12 C. McComb
and synthesis networks (which make use of portions of the autoencoders) continue
to increase. However, training time increases substantially for larger latent spaces.
Based on this study, a latent space dimensionality of 16 was selected as all networks
are at or near maximum accuracy for this value.
Autoencoders
The force spectra autoencoder was trained for 100 epochs with a batch size of 100. The
final mean squared error on the testing dataset was 9.90 × 109 . The total variance of
the training data was 5.89 × 1010 yielding a coefficient of determination of 0.83. This
indicates that this autoencoder can account for approximately 83% of the variance
observed in the training data. Several randomly selected examples of original and
reconstructed spectra are provided in Fig. 8. Although the curves are not reconstructed
exactly, in all cases the reconstructed curves tend to share many similarities with the
original curves. These similarities include slope, range, and the location of maxima
and minima. However, some distinct differences become apparent for spectra that
have low values. For instances, in Fig. 8a, b, the green spectrum (corresponding
to pitch) is nearly flat in the original. The reconstructed version, however, shows
significantly higher values for that spectrum. A similar overestimation is observed
for the blue spectrum in Fig. 8d.
The geometry autoencoder was trained for 40 epochs with a batch size of 100.
The final binary cross-entropy on the testing dataset was 0.17. By comparing the
final binary cross-entropy of the model to the binary cross-entropy of a mean model
(where the value of every voxel is the average over all voxels in the dataset), it is pos-
This section reports the results of the neural networks designed to accomplish analysis
and synthesis—these are the ultimate objects of the current work. These networks
utilize portions of the autoencoders trained in the previous sections. Specifically,
the network trained to perform analysis utilizes the encoder for voxel geometry and
the decoder for force spectra. Conversely, the network trained for synthesis uses the
encoder for force spectra and the decoder for voxel geometry.
The analysis network was trained for 25 epochs with a batch size of 100. The final
mean squared error on the testing dataset was 7.49 × 109 , yielding a coefficient of
determination of 0.87. Figure 10 shows several examples for the analysis network.
From left to right, each example includes the geometry provided to the network as
an input, the true spectra (the set of spectra produced by the geometry in NEMOH),
and the predicted spectra (the output from the network). The characteristics of the
14 C. McComb
predicted spectra are similar in some ways to the reconstructed spectra in Fig. 8.
The analysis network correctly predicts qualitative aspects of the curves, accurately
producing curves with slopes, maxima/minima, and ranges that are similar to the true
spectra. However, like the autoencoder, the analysis network tends to overestimate
low values. This is evidenced in Fig. 10b, c. In addition, the true spectrum in Fig. 10c
shows a very specific cusp feature which does not appear in the predicted spectrum.
It is likely that cusp features of this type were rare in the training data, and thus are
filtered out by the spectra decoder.
The analysis network was trained for 25 epochs with a batch size of 100. The
final binary cross-entropy on the testing dataset was 0.45, yielding a coefficient of
determination of 0.90. Figure 11 shows several examples of the synthesis network.
From left to right, each example includes the set of spectra that was used as input,
the true geometry (the geometry originally used to produce the input spectra in
NEMOH), and the predicted geometry (the output from the network). In some of
these examples, the synthesized geometry shows distinct departures from the true
geometry. Sharp corners tend to be rounded off and flat faces bow outward slightly.
This is expected, since similar behavior was observed in the geometry autoencoder.
In addition, it appears that in Fig. 11c a cone-type geometry was synthesized
for what should have been a wedge. Similarly, in Fig. 11d, a square geometry was
synthesized in place of what should have been a cylinder. At this point, the reason
behind such idiosyncrasies is unclear. One possibility is that the departure from
expected performance is due to simple errors in the synthesis. On the other hand, the
synthesis network may have created a different geometry that provides force spectra
that are very similar to what was desired. This will be a subject of future work.
The analysis and synthesis networks potentially provide very real utility for
designers of offshore structure. The force spectra that are produced with the analysis
network are important for simulation of offshore structures. However, the produc-
tion of force spectra using BIEM methodology can take minutes for a simple mesh,
which precludes the direct use of the approach in many optimization algorithms
which might require tens of thousands of iterations. The use of the analysis net-
work as an approximate BIEM makes the direct use of optimization algorithms more
feasible.
Regarding the synthesis network, the ocean waves at a given location can be
characterized with a power-frequency spectrum similar to the force spectra computed
for the structure. If the peaks on the wave and device spectra align, then the device will
absorb significant energy from the waves; if the peaks do not align, energy absorption
is mitigated. Thus, many designers can estimate a desirable force spectrum for the
structure based on known characteristics of the installation location, and use the
synthesis network to directly generate a suitable geometry.
Toward the Rapid Design of Engineered Systems Through Deep … 15
Conclusions
The design of modern systems and products typically involves intensive computa-
tional analysis. For domains such as the design of offshore structures, these analyses
can be particularly time-consuming. Standard methods for evaluating and synthe-
sizing WECs and other offshore structure are too computationally expensive to effi-
ciently implement within modern optimization and design algorithms. This work pre-
sented an autoencoder-based methodology for rapidly synthesizing and evaluating
engineered systems in a space of reduced complexity, and applied that methodology
to the synthesis and analysis of offshore structures.
The first step in the proposed methodology is the generation of data consisting of
paired system design and performance information. In the offshore structure applica-
tion of this paper, this consisted of voxel-based geometry paired with force spectra.
The second step is the creation of two autoencoders that can compress and recon-
struct both the system design and the performance information. The autoencoder for
the force spectra achieved an overall reconstructive accuracy of 0.83, and provided
strong qualitative reconstruction of the inputs (matching approximate range and loca-
tion of maxima). The autoencoder for the voxelized geometry achieved an accuracy
of 0.95 showing a strong ability to reconstruct common offshore structures, albeit
with a propensity for rounding sharp corners. The third step of the methodology is
the construction of networks for synthesis and analysis by reusing portions of the
autoencoders. The analysis network (predicting force spectra based on geometry)
achieved an accuracy of 0.87 and the synthesis network (predicting geometry based
on design spectra) achieved an accuracy of 0.90. These results demonstrate that the
proposed deep learning methodology is a promising means for accomplishing the
rapid design of engineered systems.
Future work should investigate methods for increasing the accuracy of the autoen-
coders used here, as they are likely the limiting factor in the final accuracy of the
analysis and synthesis networks. It may be possible to increase autoencoder accuracy
through the use of convolutional layers [10] or the incorporation of generative adver-
sarial network constructs [38]. In addition, the inclusion of eXplainable Artificial
Intelligence (XAI) concepts [39–42] in conjunction with convolutional layers could
provide designers with voxelized features that are aligned with high-performance
solutions. Furthermore, although the geometries constructed by the synthesis net-
work only differ slightly from the true geometries, the actual performance of the
synthesized geometries is unknown. Future work should use NEMOH or another
BIEM tool to directly evaluate the actual performance of synthesized geometries. In
a similar vein, mapping differences between predicted and actual performance could
indicate regions of the space that are particularly high performance.
Extensions of this work should also test the proposed methodology in other
domains. As noted in the background section of this paper, machine learning for three-
dimensional data is still nascent, particularly for synthesis tasks (typically referred
to in machine learning as “generative” algorithms). Engineering design provides a
large quantity of structured, three-dimensional data in the form of CAD files and pro-
Another random document with
no related content on Scribd:
He put his finger on the crazy device perched up independently in
the left-hand corner; and then came down to the lines again.
“Let that be for the moment,” said he. “It don’t much signify, after
all. How do these notes go? that’s the main question. Read ’em off.”
I spelt them out, following his finger: “b a c e f d e c a d e c.”
“That’s a good boy,” he said. “And now, what are these things
beyond, that have run off the lines, so to speak?”
“What are they? Why, I don’t see what they can be but notes.”
“Exactly. Five notes.”
I stared at the bundle in my hand, and then up at Chaunt.
“O-o-o-o!” I exclaimed.
He uttered a loud ironic laugh. “Well,” he said: “what does ‘b a c e f
d e c a d e c’ spell?”
I scratched my nose. “You tell me, please.”
“O Jerusalem!” he cried, and took his pencil to the line, thus: b a c
| e f | d e | c a d e | c—
“Well?” he said again.
I shook my head.
He positively stamped. “Listen here,” he cried: “ ‘bac ef de cad-e
c’—don’t you see?”
“No.”
“O, you ineffable ass! ‘Back of the caddy’ (that’s to say the tea-
caddy; there it is), ‘see’—see what? What follows? Why, five notes,
don’t they? ‘Back of the caddy see five notes’—and there they are.”
I sank in a heap in my chair. Light had dawned on me. “And you
found ’em there, I suppose?” I murmured—“behind a false back or
something?”
He nodded. “You’re getting on.”
“And, please, what’s the thing at the top?” I continued faintly. “Let
me get it all over at once.”
“Ah!” he said: “there’s a trifle more ingenuity in that, perhaps. What
is it, to begin with? A demisemiquaver balanced on the top of an M
Y, eh?”
“So it appears to me.”
“To any one. Don’t be frightened. Try it every way round, and
conclude with this: ‘On the top of M Y’—that is to say, ‘on M Y,’ which
is my, ‘a demisemiquaver’: or, shorn of all superfluities (he pencilled
it down), thus: ‘on my demisemiquaver.’ Now apply the same
process.”
I looked; pondered; felt myself instantly and brilliantly inspired;
seized the pencil from him, and ticked off the measurements:—
“On my demise | mi | q | u | av | er.”
“Exactly,” said Chaunt, rising with the air of an at-length-released
martyr, and proceeding to roll another cigarette: “ ‘On my demise, my
cue you have here.’ ’Pon my word, without irreverence, it’s worthier
of the composer of ‘Say, den, Julius, whar yo’ walkin’ roun’?’ than of
the author of ‘Some Unnoticed Sides of Bacon.’ But all one can say
is that he adapted himself to the intellectual measure of his legatee.
Have you got a match?”
I must end, I am really ashamed to say, with this. Anyhow, in one
way my uncle was triumphant: I was convinced, at last and at least,
of a value in cryptograms.
THE UNLUCKIEST MAN IN THE
WORLD
He was nicknamed, ironically, Carabas—a sort of French equivalent
for Fortunatus—the only title by which I ever knew him. Perhaps the
underlying sympathy which impelled the jest reconciled him to its
mockery; for there is, after all, an acute distinction in being the
unluckiest man in the world. Somebody says somewhere that it is
better to “lead” in hell than be a super in heaven. There came a time,
I think, when Carabas would have resented good fortune as an
outrage. It would have broken his record, and made him
commonplace at a blow. As with Hawthorne’s young woman who
was bred and throve on poisons, a normal dietary would have been
fatal to him. Carabas was nurtured on ill-luck.
I made his acquaintance at Verey’s in Montreux. It was for ever
Carabas here and Carabas there, and, sometimes, in badinage, M.
le Marquis; for the fellow was always in huge request for his
capability and good-humour. There was a great deal of
commiseration being shown for him when I first arrived. Latterly he
had drawn a prize ticket—for thirty thousand francs, I think it was—in
some State lottery. But, alas! a few days before the declaration of the
winning numbers, he had parted with his voucher for a trifle over cost
price. We got up a consolation subscription for him in the hotel—
relatively, quite a respectable little sum—which, with effusive thanks,
he deposited in the Bureau de Secours Mutuels. The bank stopped
payment almost at once, and Carabas lost his nest-egg, with a
prospect of future “calls” from the parent cuckoo.
After that, we abandoned him to his Nemesis. We had recognized
finally, I suppose, that vails to him meant nothing but tips to his evil
destiny, to whom, as to a rapacious head-waiter, they all accrued.
And so he himself was convinced with us. He showed himself neither
surprised nor aggrieved; but remained the sunniest fatalist, with just
a touch of wistfulness, which Nature had ever produced out of a
union between Candour and Philosophy.
I don’t know what his official position was. I don’t think he knew
himself. He wore a plain peaked cap, and a sleeved waistcoat with
brass buttons, and was, loosely, jackal to the tremendous concierge
whose bullion took the costly glass tabernacle in the hall with
splendour. Carabas himself was not at all a figure of splendour. He
was small, and placable in expression, with smiling cheeks, mobile
lips, pencilled over by a tiny black moustache, and strength visible in
nothing but his eyes. They were his vouchers of distinction above the
common brand.
One thing certain about him was that he was an accomplished
linguist; a second, that, for all his unspoiledness, he had a large
experience of man, and (notably) womankind; a third, that his
courage was equal to his good temper; a fourth, that, with every
natural claim to consideration, his pride halted at no service, whether
of skill or complaisance, which an unscrupulous management could
exact of him; a fifth and last, that he permitted his employers so to
presume upon his reputation for successlessness, as to accept from
them, in reward for his many accomplishments, wages which would
have been cheap to inefficiency. His own material welfare, indeed,
seemed always the thing remotest from his interests. To be helpful to
others was the sum of his morality.
I never could satisfy myself as to his nationality. Once—as one
might ask him anything without offence—I put the question to him. To
my secret surprise, he seemed to hesitate a perceptible moment
before he answered, with a smiling shrug of his shoulders—
“Cosmopolitan, monsieur; a foundling of Fortune.”
“We should do very well, then,” I answered, “to claim you for
England.”
Was it fancy on my part that his pleasant face paled a little? “As to
that,” he said, “I know nothing.”
“You have never been in England?”
He made no reply, but began bustling over some incoming
luggage, calling to the porters at the lift; and in a moment he left me.
The next day he was taken ill. The reversion of a service of raw
oysters, supplied to the guests at table d’hôte, had found its way to
the supper-table of the staff. Carabas detested oysters, but his
gallantry to the fair sex was proverbial, and Ninette, the prettiest of
filles de cuisine, sat next to him. She extracted a single “bivalve”
from her half-dozen, and put it on his plate, moueing at him
ravishingly.
“Love conquers everything, M. Carabas,” she said; “even the
antipathies of the stomach. I will not believe in your protestations
unless you eat this for my sake.”
He swallowed it at a gulp, and—it was a bad oyster, the only
doubtful one in the whole consignment. Later, he was very sick; and
afterwards ill for four days. Ninette cried, and then laughed, and
congratulated herself on her escape. But as for the hotel, it was
disconsolate in the temporary loss of its Carabas.
For my part, I was even particularly conscious of a vague
discomfort in his absence. Somehow a certain personal
responsibility which I had undertaken seemed to weigh upon me the
more heavily for it. It was not that Carabas could have lightened, by
any conceivable means, my burden. It was just a sense of moral
support withdrawn at a critical moment. It was as if the knees of my
conscience were weak, owing to something having gone wrong with
my backbone. But I will explain.
Mr. G——, a very famous lawyer in our own country, had brought
his family, a son and daughter, to holiday in Lucerne. The boy was a
conceited and susceptible youth; to the lady I was—engaged.
There seems no reason why impressionability should spell
obstinacy; yet very often it does. Young Miller (so I will call him)
having invited himself, at the Schweitzerhof, into the toils of a siren—
a patently showy and dubious one—resisted all the efforts of his
family to help him out. Baffled, but resolute, the father thereupon
shifted the scene to Montreux, where they were no sooner arrived
than he was summoned home on business at a moment’s notice. In
the meanwhile, to me (hastily called from Paris, where it had been
arranged I was to join the party on its homeward journey), was
assigned the unenviable and impossible task of safeguarding the
family interests. Miller had positively refused to accompany his father
home, then or thereafter, until his absurd “honour,” as he called his
fatuity, was vindicated. It would never do to abandon the wretched
infant in the wilderness. He had his independence, and was a
desirable parti. Hence my promotion to an utterly fictitious authority.
I knew, naturally, how it would be; and so it turned out. The head
was no sooner withdrawn, than Mademoiselle Celestine—privately
advised, of course, of the fact—arrived at Verey’s. Here, then, was
defiance unequivocal—naked and unashamed, I might have said,
and been nearer the truth of the case. For mademoiselle’s charms
were opulent, and she made no secret of them. One would have
thought a schoolboy might have seen through that rouge and
enamel, through the crude pencilling on those eyelashes, through all
that self-advertising display. I will not dwell upon its details, because
their possessor made, after all, only a summer nightmare for us, and
was early discomfited. She served, at best, for foil to a brighter soul;
and such is her present use in the context.
From the outset there was no finesse, no pretence of propitiation
in her tactics. She understood that it was a matter of now or never
with her quarry, and aimed to bring him down sitting. A woman, even
the best of her sex, never gives “law” in these matters. She goes out
to kill.
The two together formed an opposition camp—quite flagrantly, out
in the sunlight. I thought sometimes the boy looked unhappy; but the
witch would never let me have him to myself, and I could not
manœuvre her from under his guns. I would never have scrupled to
roll her in the mud, could I once have got her alone. But she was too
cunning for that; and, as for her companion, his warfare was, after
all, an honourable warfare. And all the time I had my own particular
Campaspe to safeguard, to console, to squire through the odious
notoriety which her brother’s infatuation had conferred upon us all.
It was Carabas, of course, who in the end procured us a way, his
own, out of the difficulty. The scandal being common property, there
was no need for him to affect an ignorance of it. Yet we never knew,
until the moment of his decision, how it had been occupying his mind
from the beginning, or how, quietly and unobtrusively, he had been
studying to qualify himself as our advocate. “Our advocate,” I say;
but I knew his brief was for the bright eyes of Campaspe. He struck
for the credit of the hotel, he declared; and mam’selle was
associated with the best of that. Anyhow he struck, and daringly.
He had risen from his bed on the fourth day, as smiling, as
complaisant as ever. His presence, like a genial thaw, ameliorated
the little winter of our discontent. We greeted his reappearance with
effusion, and dated, from the moment of it, our restoration to the
social sanities.
It was a dusk, warm evening. The peaks of the Dent du Midi,
thrust into a dewy sky, had been slowly cooling from pink to pearl-
ash, like ingots of white-hot steel. Everything seemed one harmony
of colour, except our thoughts, Campaspe’s and mine, as we strolled
in the deserted garden. The Celestine and her victim had been out
boating on the lake. We met them, unexpectedly returning.
Mademoiselle was eating cherries out of a bag, and daintily spitting
the stones right and left as she advanced. I don’t know how we
should have faced the contretemps; I had no time, indeed, at the
moment, to form a decision, before Carabas came softly and swiftly
from a leafy ambush, and took command of the occasion.
We all, I believe, instinctively recognized it for a critical one.
Mademoiselle’s bosom, though she laughed musically (she had
managed to preserve, it must be owned, the unspoiled voice of a
séductrice) began to rise and fall in spasms. The portier addressed
her without a moment’s hesitation.
“I take the liberty to inform madame that she is in danger.”
She gave a little gasp.
“But is this comedy or melodrama?” she cried vehemently.
“That is,” said Carabas, “as madame shall decide. I have the plot
up my sleeve.”
“The plot!” she echoed, and fell staring at him; and then furiously
from him to us.
“Go on,” she said. “I know very well who has instigated you to
this.”
She checked herself, and, smiling, put out a hand towards her
companion, as if to ask, or give, reassurance. But I noticed, already
to my satisfaction, that the boy did not respond. As for us, we were in
complete darkness.
“I obey, madame,” said Carabas. “This plot is told in a word. There
was once in Paris a certain notorious courtisane et joueuse. Will
madame desire her name?—à bon entendeur demi-mot. One night
this lady’s husband, a Corsican, from whom she was separated on
an honourable allowance, visited, purely by accident, her
establishment. There was a fine scene, and he wounded her
severely. She was forced by the police to prosecute him, and the
jury, amidst the plaudits of the public, gave their verdict—against
madame. But, triumphant there, the husband’s vengeance was
whetted rather than assuaged. He would throw himself upon the
suffrages of his countrymen in a more drastic vindication of his
honour. She had disguised herself—her name—had fled. He
devoted himself to the business of pursuit. At length he believed he
had traced her to an hotel in Territet.”
Carabas shrugged his shoulders and his lips, stuck out his arms at
right angles with his body, stiff from the elbow, and came to a
significant stop. I declare I pitied the adventuress. Every expression
but that of panic seemed eliminated from her face at a touch. She
looked old and haggard; and then, as if conscious of her self-
betrayal, collapsed in a moment, dropping her bag of cherries.
“I am not very well,” she stammered; “the night air tries me.” She
turned lividly upon the portier: “Par pitié, monsieur! C’est pour me
prevenir que vous etes venu, non pour me trahir?”
Without waiting for his answer, she gathered herself together,
literally, folding her train about her arm; made a desperate effort at
self-command; wrenched out a smile, and went off, quavering a little
airy chansonnette. But, after a few steps, despite her royal amplitude
she was running. Carabas, very pale but self-possessed, picked up
the bag, found one cherry in it, put it in his mouth abstractedly, and—
“My God!” cried Miller hoarsely.
Carabas jumped, and gulped.
“A thousand devils!” he cried. “You made me swallow the stone,
monsieur.”
The boy was in a fever of agitation.
“Is she really that—that sort?” he said.
My Campaspe fell upon his neck.
“O, my dear, O, my dear, I am so sorry!” she sobbed.
He put her roughly, but not unkindly, away.
“I’m—I’m going back to England—to the governor,” he said.
“Carabas,” I demanded privately, as we returned to the hotel, “is it
a fact that——?”
“The husband is here? No, monsieur; it is not a fact.”
“But——”
“It was a cause célèbre. I was confident I recognized madame
from the published prints. For the rest, it was just a chance shot; but
it hit the mark.”
“Carabas, you are wonderful; and we shall not forget.”
Miller was as good as his word. With characteristic disregard for
any but his own interests, he was gone the next morning, without
sign or message, leaving us to wobble in his backwash of scandal,
and to get out of it as best we could. His flight, of course, threw open
all the doors of gossip. My business in Paris being unfinished, I had
to go; but first I did my best to provide against unpleasantnesses by
confiding Campaspe to the care of the least slanderous dame de
compagnie I could find. I am afraid, nevertheless, she had but a poor
time of it.
A week later I received a letter from Mr. G——, who in the interval
had returned to Montreux.
“All is happily over with all,” he wrote: “with the exception, that is to
say, of poor Carabas, who is to undergo an operation for
appendicitis. It appears that a cherry-stone, which he swallowed
unwittingly, did the business. The management (OWLS!) demur to
the expense. I have insisted (FOOLS!) upon undertaking it upon my
own account. We owe much to him; and so do they (IDIOTS!). But
they don’t understand how to pay your debts is very often the best
foresight.”
It was a case of pitch and toss. For days, it appeared, Carabas’s
life hung in the balance. In the meanwhile, I was enabled to rejoin
Mr. G—— and his daughter at Montreux, and to take my share in the
nursing. Between gratitude and indignation, we rather claimed
Carabas among us. Campaspe the poor fellow simply adored. Once,
when he fancied himself losing hold, he confided to her, while we
stood by, some main incidents in his life. I retail them here, in an
abbreviated form.
CARABAS’S STORY
“There is no doubt,” he said “that as truly as some men are born
without a palate, so some men are born without luck. It is no use
trying to remedy the deficiency; it is well, rather, to study to reconcile
oneself to it.
“I was born in an English village, of naturalized Huguenot parents.
When I was nineteen, I fell passionately in love. I had for a rival a
youth very strong and unscrupulous. One day he persuaded me to
bathe with him in the river, then swollen with floods. In mid-stream he
pounced upon me, and strove to bear me under. I struggled
desperately—it was of no avail. Death thundered in my ears; the
water enwrapped and proceeded to swallow me. The last thing I saw
was a figure gesticulating and shouting on the bridge a little way
above; then consciousness fled, and I sank. I came to myself,
stranded somewhere in a dark channel. A mad face was bending
over me. I knew it—it was that of the miller. I had been carried into
his race, and, just short of the wheel, he had caught and dragged me
to shore. He was a drunkard, of that I was aware; and he was now
quite demented.
“ ‘Mordieu!’ he said, ‘I see what you’ve come for, and the devil
shan’t call twice for his own.’
“I understood instantly. He meant himself to go with me into the
water—to join issues with the devil who had called for him, and have
a fine frolic into eternity with his visitor. Terror lent me strength. I
caught at a post, and, as he leaned down, shot my whole body at
him like a spring. He went over with a splash, and I heard the wheel
hitch, then begin to turn again, chewing its prey. O, my friends, what
a situation! I lay like one damned, a thousand dreadful reflections
mastering me. I should be accused, if caught, of murdering this man.
That terror quite devoured the other, and increased with every
moment that I lay. Darkness came upon me, and then I rose and
fled. I thought of nothing but to escape; and so, stealing always by
night, I reached London.
“Now, I will tell you the irony of this destiny. Many weeks later I
read, by chance, in a newspaper, how my rival had been granted
your Royal Humane Society’s testimonial, on the evidence of a
casual spectator, for a brave but unsuccessful attempt to save me
from drowning; and how the little pretty romance had terminated with
his marriage to the admiring object of our two regards. So I was
dead; and, as long as I lay in my nameless grave—for my body, it
appeared, had never been recovered—the ghost of my fear was laid.
I do not complain, therefore. Yet—ah, mademoiselle, most
condescending of sympathizers!—she had been very dear to me.”
Here Carabas found it necessary to console my Campaspe before
he could go on.
“I obtained work—under an assumed name, of course—and for
many years found at least a living in that immense capital. I had an
aptitude for languages, which was my great good fortune; yet
prosperity never more than looked at me through the window. What
then? I could keep body and soul together. Ill-luck is too mean a
spirit for Death to patronize. Many a time has the great Angel turned
his back disdainfully on the other’s spiteful hints. He will not claim
me, I believe, until he sees him asleep, or tired of persecuting me.
“One day I was travelling on your underground railway. I had for
companion in my compartment a single individual. He jumped out at
the Blackfriars station, leaving a handbag on the seat. At the
moment the train moved off I noticed this, seized it, and leaned with
it out of the window, with a purpose to shout to its owner. I saw him
in the distance, hurriedly returning. The train gathered speed; I saw
he could never reach me in time, and I flung the bag upon the
platform. Instantly I perceived him leap, and jerk his arm across his
eyes; and on the same moment a terrible explosion occurred.
“Stunned, but unhurt, I had fallen back, when, in a flash, the full
horror of my situation burst upon me. It was the time of the dynamite
scares, and—ah, mon Dieu, mam’selle! your quick wit has already
perceived my misfortune.
“The train had stopped; the place was full of smoke; the hubbub of
a great tumult sounded in my ears. The owner of the infernal
machine was certainly destroyed in his own trap; I, at the same time,
had as certainly thrown the bag. No evidence to exonerate me was
now possible. Without an instant’s consideration, I opened the door
upon the line, slipped out, closed it, and raced for my life through the
smoke to the next station. I was successful in gaining its platform
without exciting observation. News of the catastrophe had already
been passed on, so that I was able, mingling with a frenzied crowd,
to make my way to the streets. But panic was in my feet, and all
reason had fled from my brain. I felt only that to remain in London
would be to find myself, sooner or later, the most execrated of
human monstrosities, on the scaffold. There and then I effaced
myself for the second time, hurried to the docks, and procured a post
as steward on an out-going steamer. I have never been in England
since. I now give monsieur the explanation he once asked for,
secure in the thought that, as ill-luck has at last conceded to me the
ministrations of this dear angel of a mademoiselle, his persecution
must be nearing its end before the approach of the only foe he
dreads. I leave it to monsieur, if he likes, to vindicate my name.”