Full Ebook of 4Th Eai International Conference On Robotic Sensor Networks Shenglin Mu Online PDF All Chapter
Full Ebook of 4Th Eai International Conference On Robotic Sensor Networks Shenglin Mu Online PDF All Chapter
Full Ebook of 4Th Eai International Conference On Robotic Sensor Networks Shenglin Mu Online PDF All Chapter
https://ebookmeta.com/product/5th-eai-international-conference-
on-management-of-manufacturing-systems-1st-edition-lucia-
knapcikova/
https://ebookmeta.com/product/advances-of-science-and-
technology-9th-eai-international-conference-icast-2021-0013892/
https://ebookmeta.com/product/distributed-sensor-networks-second-
edition-sensor-networking-and-applications-s-sitharama-iyengar/
Proceedings of the 11th International Conference on
Computer Engineering and Networks 1st Edition Qi Liu
https://ebookmeta.com/product/proceedings-of-the-11th-
international-conference-on-computer-engineering-and-
networks-1st-edition-qi-liu/
https://ebookmeta.com/product/proceedings-of-the-joint-
international-conference-10th-textile-conference-and-4th-
conference-on-engineering-and-entrepreneurship-1st-edition-genti-
guxho/
https://ebookmeta.com/product/artificial-neural-networks-and-
machine-learning-icann-2021-30th-international-conference-on-
artificial-neural-networks-bratislava-slovakia-i-igor-farkas-
editor/
Shenglin Mu
Li Yujie
Huimin Lu Editors
4th EAI
International
Conference
on Robotic
Sensor Networks
EAI/Springer Innovations in Communication
and Computing
Series Editor
Imrich Chlamtac
European Alliance for Innovation
Ghent, Belgium
Editor’s Note
The impact of information technologies is creating a new world yet not fully
understood. The extent and speed of economic, life style and social changes
already perceived in everyday life is hard to estimate without understanding the
technological driving forces behind it. This series presents contributed volumes
featuring the latest research and development in the various information engineering
technologies that play a key role in this process.
The range of topics, focusing primarily on communications and computing engi-
neering include, but are not limited to, wireless networks; mobile communication;
design and learning; gaming; interaction; e-health and pervasive healthcare; energy
management; smart grids; internet of things; cognitive radio networks; computation;
cloud computing; ubiquitous connectivity, and in mode general smart living, smart
cities, Internet of Things and more. The series publishes a combination of expanded
papers selected from hosted and sponsored European Alliance for Innovation (EAI)
conferences that present cutting edge, global research as well as provide new
perspectives on traditional related engineering fields. This content, complemented
with open calls for contribution of book titles and individual chapters, together
maintain Springer’s and EAI’s high standards of academic excellence. The audi-
ence for the books consists of researchers, industry professionals, advanced level
students as well as practitioners in related fields of activity include information and
communication specialists, security experts, economists, urban planners, doctors,
and in general representatives in all those walks of life affected ad contributing to
the information revolution.
Indexing: This series is indexed in Scopus, Ei Compendex, and zbMATH.
About EAI
EAI is a grassroots member organization initiated through cooperation between
businesses, public, private and government organizations to address the global
challenges of Europe’s future competitiveness and link the European Research
community with its counterparts around the globe. EAI reaches out to hundreds of
thousands of individual subscribers on all continents and collaborates with an insti-
tutional member base including Fortune 500 companies, government organizations,
and educational institutions, provide a free research and innovation platform.
Through its open free membership model EAI promotes a new research and inno-
vation culture based on collaboration, connectivity and recognition of excellence by
community.
Huimin Lu
Kyushu Institute of Technology
Kitakyushu, Japan
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
v
vi Preface
Steering Committee
Organizing Committee
General Chair
Huimin Lu, Kyushu Institute of Technology, Japan
General Co-Chairs
Hao Gao, Nanjing University of Posts and Telecommunications, China
TPC Chair and Co-Chair
Li Yujie, Fukuoka University, Japan
Sponsorship and Exhibit Chair
Joze Guna, University of Ljubljana, Slovenia
Local Chair
Yin Zhang, Zhongnan University of Economics and Law, China
Workshops Chair
Guangxu Li, Tianjin Polytechnic University, China
Publicity & Social Media Chair
Junwu Zhu, Yangzhou University, China
Publications Chair
Shota Nakashima, Yamaguchi University, Japan
Web Chair
Shenglin Mu, Ehime University, Japan
Posters and PhD Track Chair
Xin Jin, Beijing Electronic Science and Technology Institute, China
Panels Chair
vii
viii Conference Organization
ix
x Contents
Phan Van Vinh, Tran Minh Thanh, Vo Duy Khang, Ngo Hoang An,
and Tran Duy Nhat
1 Introduction
Nowadays, Internet has become one of the most important aspects of our lives.
Everyone uses Internet for almost all of their life activities. Internet of Things
(IoT) is a network of devices such as cameras, vehicles, home appliances, and other
electronic devices which can connect, exchange, and transfer data together. One of
the most widely used IoT applications is surveillance and rescue system.
A surveillance system with CCTV (closed circuit television) cameras or other
IP cameras is used to monitor sensitive areas, such as outdoor areas (public places,
garden, front of a house, etc.) and indoor areas (warehouse, home, garage, etc.).
However, such cameras have some drawbacks and constraints such as they usually
have blind spots, and also to cover an entire area, many cameras are required to
reach every nook and corner. A rescue system helps save the life of victims who
get stuck in dangerous situations. In such case, a rescue team cannot cover the
entire affected area, especially the areas that are difficult to access. Moreover, it is
dangerous and risky to enter such areas without getting the prior information about
the number of victims and their condition and the environmental conditions of the
target area. Sometimes, due to lack of information, a rescuer may become a victim.
To overcome this problem, deploying a mobile robot into the given area is one of
the best solutions. The mobile robot mounted with video camera and sensors can go
to the affected area and send all real-time information from the target environment
to the rescue team. By analyzing the collected data, the rescue team can plan their
work more efficiently.
However, building such kind of mobile robot with all necessary features is very
expensive, complex, and time-consuming. One mobile robotic system for indoor
surveillance was proposed in [1]. This robot can handle basic problems related to
environment mapping, localization, and autonomous navigation. However, it cannot
detect the presence of people in the given area. The authors in [2] proposed a
surveillance robot that is able to avoid obstacles based on back propagation neural
network. To reduce the cost in building robots, [3] proposed a mobile robot based
on Android smartphone. A similar approach which uses Android smartphone with
Arduino microcontroller was also proposed [4]. However, due to the hardware and
software limitations, the above-mentioned approaches do not perform well in image
processing tasks, such as human detection. To handle this problem, a deep learning
algorithm with high performance of accuracy and speed can be considered, such
as Region Convolutional Neural Network (RCNN) [5], Fast RCNN [6], Multi-
scale Deep Convolutional Neural Network (MS CNN) [7], Faster RCNN [8], and
YOLO [9]. MobileNet-V2 [10] is the state-of-the-art object detection algorithm
with the best performance in accuracy. However, it needs to be run on computers
with powerful graphic cards compared to other algorithms.
In this paper, we aim to design and implement a mobile robotic system with
all necessary features at low cost. The mobile robot system can provide live video
stream of the target area, detect human presence, and detect other environmental
related information such as humidity, temperature, CO, and poisonous gas. The pro-
posed system can be controlled and managed remotely via Web-based application.
The contribution of this paper is as follows: (1) to design and implement a mobile
robot prototype and test it in practical scenarios and (2) to implement a web-based
application to control and manage the mobile robot system.
A Mobile Robotic System for Rescue and Surveillance in Indoor Environment 3
In what follows, we present the principal design of the proposed approach in Sect.
2 and implementation of results in Sect. 3. Finally, we make concluding remarks in
Sect. 4.
In this section, we discuss the main design of the proposed system, including the
system architecture, hardware and software design, and the prototype system model
to help our readers understand clearly how our system works.
In this paper, we aim at designing a mobile robotic system to help rescue team
in their works in a timely and efficient manner. Therefore, the system architecture
should be simple, that is, easy to control and manage as shown in Fig. 1.
Control &
User/ User Interface Management
Application
Analytic/
Processing
Human Detection Web server Database server
Intel NCS2
Controller/ Raspberry
Gateway Pi
Sensors/
Actuators
In this section, we describe the hardware design of the proposed mobile robot. The
Raspberry Pi model is used as a main controller to connect to all other parts of the
robot, including sensors, LED, buzzer, DC motor, and Pi camera. Fig. 3 shows the
connection diagram of the required components on breadboard which is great for
prototyping circuits.
Table 2 presents all the required hardware components of our mobile robot
design.
Table 1 The specifications of some selected Raspberry Pi models
Raspberry Pi platform Raspberry Pi Zero Wireless Raspberry Pi 3 B Raspberry Pi 3 B+ Raspberry Pi 4B
Processor 1 GHz single-core ARM11 1.2 GHz, Quad-Core Cortex A53 1.4 GHz, Quad-Core Cortex A53 1.5 GHz, Quad-Core Cortex A72
RAM 512 MB 1 GB 1 GB 1/2/4 GB
USB 1 × micro USB 4× USB2.0 4× USB2.0 2× USB3.0 + 2× USB2.0
Ethernet – 10/100 Mbps 10/100 Mbps 1Gbps
Wi-Fi 802.11n 802.11n 2.4GHz and 5GHz 802.11 b/g/n/ac 2.4GHz and 5GHz 802.11 b/g/n/ac
Bluetooth 4.1 4.1 4.2 5.0
HDMI Mini-HDMI Yes Yes 2 × micro HDMI
GPIO 40 pins 40 pins 40 pins 40 pins
A Mobile Robotic System for Rescue and Surveillance in Indoor Environment
5
6 P. Van Vinh et al.
When completing the prototyping step with breadboard, we have to make the
robot work more permanent by designing the printed circuit board (PCB) of the
mobile robot. Figure 4 illustrates the schematic diagram of all required component
as follows:
• Power supply block: this includes the voltage reducer circuit with pulse power
supply which can have maximum output of 5 V-3 A from the input voltage of
5–24 V DC to supply power for other components (e.g., Raspberry PI, sensors,
and actuator).
A Mobile Robotic System for Rescue and Surveillance in Indoor Environment 7
• Motor driver block: this includes Dual H-bridge driver chip L298N which
can drive two DC motor and control the speed and direction of each motor
independently through PWM (pulse width modulation). Input voltage is 5–
30 V DC and maximum output is 2A for each bridge.
• Actuator block: this includes two servo motor, buzzer, and LED. The speed of
servo motor can be controlled through PWM.
• Sensor block: this includes DHT11, SRF04, and MQ2 sensor for collecting
environmental data.
The next step is to transfer the schematic diagram into a drawing of PCB (Fig.
5). The schematic will serve as a blueprint for laying out the traces and placing the
components on the PCB. In addition, the PCB editing software can import all of
the components, footprints, and wires into the PCB file, which will make the design
process easier.
In this case, we use double-layer PCB, where the entire bottom layer is covered
with a copper plane connected to ground. The thickness of power wires is 35 mil
while that of data wires is 10 mil (0.254 mm). The positive traces are routed on top
and connections to ground are made with through-holes or vias. Ground layers are
good for circuits that are prone to interference because the large area of copper acts
as a shield against electromagnetic fields. They also help dissipate the heat generated
by the components (Fig. 6).
Figure 7 shows the design of the web interface which is used to control the mobile
robot movement and visualize the collected data. Users must have an account to log
in into the system. There are two types of users: normal users who can only see
8 P. Van Vinh et al.
the monitoring data and live-stream image and admin users who have full privileges
to manage the whole system. The functionalities of some main pages in the web
interface are as follows:
• Robot Control page: this controls the movement of robot and pan/tilt camera,
observes the surrounding environment data and live stream from camera, records
video or take a photo and save in gallery, enables/disables sensors or object
detection feature, etc.
• Data page: this visualizes real-time environmental data acquisition in graph view
and tabular view and displays the recorded videos and photos.
• Location page: this shows the location of the robot on the map via Google Map.
The collected data are stored in a MySQL database which includes the following
table:
• User: stores the user information such as Id, name, role, e-mail, and contact
number.
• Monitor_data: stores the collected data, including temperature, humidity, and
MQ2.
• Obstacle_distance: stores distance value to obstacle.
• Photo: contains Id and name of a photo.
• Video: contains Id and name of a recorded video.
A Mobile Robotic System for Rescue and Surveillance in Indoor Environment 9
Fig. 5 The PCB layout of the designed mobile robot (2D view)
Image classification and object detection are the major applications of machine
learning. Image classification predicts the type or class of an object in an image,
whereas object detection locates the presence of defined objects with a bounding box
and assigns them a class label. Therefore, human detection is a specific application
10 P. Van Vinh et al.
START
No
Read Camera Exit with Error
Yes
Human Detection
Write Log to
Detect # of people Record Video
Database
FINISH
3 Experimental Results
To verify how the Intel NCS2 module can improve the performance of the edge
devices in image processing tasks, we conduct some experiments with the inference
benchmark testing. The experiments use SSD Mobilenet-V2 network model and
TensorFlow framework, frame size of 400×400 with object detection application.
Figure 9 shows the results from deep-learning inference benchmarks of some
selected Raspberry models and laptop or desktop computer with or without the Intel
NCS2 module attached. As we can see that with the Intel NCS2, the performance
of all Raspberry Pi models increases rapidly, especially for Raspberry Pi 4 with
high speed CPU and USB3.0 port supported. The Intel NCS2 also improves the
performance of other hardware platform, such as laptop or desktop computers which
do not have a graphic card inside. However, the upper bound benchmark of the Intel
NCS2 is about 20 FPS for this experiment.
Table 3 shows the full inference benchmarks and hardware utilization of some
selected platforms when running the object detection algorithm. From these results,
we can see that the Intel NCS2 not only improves the inference benchmark but also
reduces the CPU usage when running deep neural network applications. This allows
IoT-enabled devices to adapt well to a wide range of IoT applications.
12 P. Van Vinh et al.
Figure 10 presents the prototype of our proposed mobile robot. We deploy this robot
to the target environment and perform the experiment for functionality testing. We
can control the robot motion and pan/tilt camera for better video or photo capturing.
We can also get data from sensors to justify the environment.
When accessing to the system, the admin team can perform their operation with full
rights. At the Robot Control page (Fig. 11), we can see all collected data such as
live streaming, temperature, humidity, MQ2 level, or distance to obstacle. The alarm
message can be displayed when the monitoring data reach the threshold value.
At the Data page, all collected information from the target environment can be
observed in real time. Figure 12 shows the real-time monitoring data in graphs.
The detailed monitoring data can also be displayed as shown in Fig. 13.
When working in the target environment, the robot can take some photos or
record videos which are very helpful to the admin team in the future. An example
of recorded videos is presented in Fig. 14.
14 P. Van Vinh et al.
4 Concluding Remarks
In this paper, we have already applied the newest technology in IoT and computer
vision to design and implement a mobile robotic system which collects the ambient
data and then analyzes the collected data for making the right actions in a
timely manner. The proposed system is potentially helpful in providing the correct
information of target environment to the rescue team. Besides, it can detect the
objects, especially human being in the given area, which helps to save the human
lives quickly and efficiently in case of abnormal or dangerous situation. However,
to help our proposed system more practically and efficiently, some necessary
features need to be improved as follows. The indoor positioning system needs to
be implemented to get the correct position of the collected data (e.g., the real-time
and correct position when detecting human beings or emission of poisonous gas
in the target area). The prototype hardware model is also redesigned to make the
smaller and compact one which can be applied in the practical scenario.
References
1. Di Paola, D., Milella, A., Cicirelli, G., Distante, A.: An autonomous mobile robotic system for
surveillance of indoor environments. Int. J. Adv. Robot. Syst. 7, pp. 19–26 (2010)
2. Budiharto, W.: Intelligent surveillance robot with obstacle avoidance capabilities using neural
network. Comput. Intell. Neurosci. 2015, pp. 1–5 (2015)
3. Shah, M.S., Borole, P.B.: Surveillance and rescue robot using Android smartphone and the
Internet. In: International Conference on Communication and Signal Processing (ICCSP), pp.
1526–1530 (2016)
4. Azeta, J., Bolu, C.A., Hinvi, D., Abioye, A.A., Boyo, H., Anakhu, P., Onwordi, P.: An android
based mobile robot for monitoring and surveillance. Proc Manuf. 35, 1129–1134 (2019)
5. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object
detection and semantic segmentation. In: IEEE Conference on Computer Vision and Pattern
Recognition, pp. 580–587 (2014)
6. Girshick, R.: Fast R-CNN. In: IEEE International Conference on Computer Vision (ICCV), pp.
1440–1448 (2015)
7. Cai, Z., Fan, Q., Feris, R., Vasconcelos, N.: A unified multi-scale deep convolutional neural
network for fast object detection, In: European Conference on Computer Vision, pp. 354–370
(2016)
8. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: towards real-time object detection with
region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence,
pp. 1137–1149 (2017)
9. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: unified, real-time object
detection. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.
779–788 (2016)
10. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.: MobileNetV2: inverted residuals
and linear bottlenecks. In: IEEE/CVF Conference on Computer Vision and Pattern Recogni-
tion, pp. 4510–4520 (2018)
11. https://socialcompare.com/en/comparison/raspberrypi-models-comparison
Study on the Learning in Intelligent
Control Using Neural Networks Based
on Back-Propagation and Differential
Evolution
1 Introduction
Along with the development of computer science and electronic engineering, the
advancement of artificial intelligent technologies have been progressing in a great
pace in recent years. The most representative intelligent scheme Neural Networks
(NNs) have been introduced to many industrial fields, and various of applications
in our lives. Owing to the NNs’ excellent features, they are effectively applied in
image/speech processing and classification, emotion recognitions and so on. In the
applications of NNs, the most classic approach for the learning in them is the Back-
Propagation (BP) method, which is a gradient decent algorithm to obtain the weights
and biases in NNs. However, owing to the mechanism using cost function in the
BP learning, the differential information is necessary in the applications. For most
industrial applications, the differential information of the objectives are difficult to
be obtained [1].
Meanwhile, the development of evolutionary computation attracts a lot of
attentions in recent years, owing to their excellent features in complex optimization
problems. They are considered as good solutions to solve the limitations in the
learning of NNs. Genetic algorithm (GA), which is constructed according to the
evolutionary mechanisms of natural selection, is the most well-regarded algorithm
in the evolutionary computation [2]. Since GAs are with fine searching ability,
difficult to be trapped in local minima, and can be applied without considering
the differential information in the learning procedure, they have been applied in
NN for many applications [3]. However, because of the complexity in the encoding
manipulations, and the evolution mechanism in GAs, more effective algorithms with
simple manipulations and high efficiency are expected.
In previous research, the NN is proposed for the position control of ultrasonic
motor (USM). The traditional approach of NN is applied for the position control
of USMs [4]. The BP type NN is also constructed for the position control of
USM with speed compensation [5]. The proposed method was confirmed effective
and easy to be applied in the control of USM. To solve the problem of Jacobian
estimation, we introduced the particle swarm optimization (PSO) algorithms for the
NN type PID controller in the control of USM [6, 7]. Owing to the excellent features
of PSO in optimization both in continuous and discrete problems [8], the swarm
intelligence algorithm was introduced to the proposed scheme, and confirmed
effective in obtaining high accuracy in position and speed control of USM. However,
there are some concerns about the local minimum problem and ability of re-
convergence in PSO for the control process. Therefore, in this study, the algorithm
of differential evolution (DE) is introduced in an intelligent control application using
NN. According to the study from S. Das et al. [9], DE was proofed being with as
good convergence characteristics as PSO. Meanwhile, the hybridizations of both the
algorithms are excellent comparing with other soft computing tools. In the study of
[10], DE was investigated as a global optimization algorithm for feed-forward NN.
The good features of DE, such as no learning rate setting requirement and offering
high quality solutions were confirmed. Comparing with the traditional GAs, DE
Study on NN Based on BP and DE 19
algorithms are with superior features such as simpler construction, high efficiency,
and excellent convergence features. In the NN scheme, the DE algorithm is applied
to update the weight for learning, without considering the differential limitation
as in BP algorithms, or the encoding manipulations as in GAs. The proposed DE
type NN is applied to an intelligent PID control scheme for USM to confirm the
effectiveness. The intelligent automation with the proposed method is expected to
contribute in the medical and welfare applications utilizing the excellent features of
USMs [11].
The chapter is organized as follows. In Sect. 2, the general introduction about
NN, DE, and the proposed DE type NN scheme is given. In Sect. 3, the simulation
study of the proposed intelligent scheme using the DE type NN is introduced.
Section 4 gives an experimental study of the proposed method on an existent USM
servo system. The conclusions are given in Sect. 5.
NNs are machine learning methods that imitate the mechanism of the human brain.
Owing to the excellent characteristics of them, they have been studied and applied
in various kinds of fields as the most representative of artificial intelligence [12].
Figure 1 shows an example of an NN consisting of three layers of neurons. inputi
is the input signal and the neurons contained the signals consist the first layer. wij
is the coupling weight from the first layer to the second layer (hidden layer), Hj
is the signal activated according to the sum of the weighted input signals. wij is
the coupling weight from the second layer to the third layer (output layer). Om
is the output signal activated according to the sum of the weighted input signals
i m output1
input1
input2 output2
input3
outputm
inputi
wij wjm
20 S. Mu et al.
from hidden layer. Theoretically, the weights in general schemes are initialized
with random values. However, it is necessary to adjust it to a value suitable for
the problem to be dealt with. The BP algorithm is often used for optimization.
However, some limitations in traditional methods prevent the BP type NN from
wider applications. Therefore, in this study, the DE algorithm is introduced to the
learning of NN.
Xi = [x1 , x2 , . . . , xD ] (1)
Fig. 2 Flowchar of DE
Start
algorithm
Randomly initialization
Mutation
Crossover
Selection
No
Stop criterion?
Yes
End
Study on NN Based on BP and DE 21
where rand(0, 1) is the random number within the range of (0, 1).
2.2.2 Mutation
• DE/rand/2/bin
• DE/current-to/1/bin
• DE/best/1/bin
r1, r2, r3, r4, r5 are the randomly chosen integers with different value from each
other and i, within the range of [1, N(number of searching points)].
2.2.3 Crossover
In crossover, the donor vector is applied to be combined with the target vector. There
are mainly two kinds of crossover methods (exponential and binomial) applied in
DE algorithms. In this study, we employ the binomial crossover, which applies
22 S. Mu et al.
where rand(0, 1) is a real number random number in [0, 1], and jrand is an integer
random number in 1, . . . , D.
2.2.4 Selection
In this study, the DE algorithm is introduced to update the weights in the learning
of NN. The method can be constructed with simple structure, easy to be applied,
without considering the differential information.
In the proposed scheme, the vector of DE is designed as the real value of weights
in NN. The target vector in DE can be expressed as
The weights in the NNs are updated by the mutation, crossover, and selection
procedures in the DE algorithm to obtain the optimal solutions in NN’s learning.
3 Simulation Study
To confirm the effectiveness of the proposed scheme, the simulation study was
implemented employing the Rosenbrock function for optimization using NN. The
Rosenbrock function is a convex function with two input values, defined as x and y
as follows.
Study on NN Based on BP and DE 23
The function can be plotted as in Fig. 3. In the study, the NN with the topologic
structure shown in Fig. 4 was introduced to find the minimum of the Rosenbrock
function. The NN is constructed to find the optimal (x, y) to minimize the
Rosenbrock function.
The weights in the NN is designed to be updated by the DE introduced in Sect.
2. The weights are initialized by the random real number within the range of (−1.0,
1.0). The mutation strategies were investigated in the simulation to confirm the
characteristics in DE. The parameter F which is the scaling factor is set as 0.5.
The crossover probability Cr is set as 0.8. The number of searching unit in DE is
set as 10. To investigate the effectiveness of the proposed scheme comparing with
the conventional method, we also implement the NN’s learning using traditional BP
algorithm. The weights of NN in BP is updated as follows. The weights between the
output layer and the hidden layer can be updated as
∂f
wj m = wj m − η (13)
∂wj m
where η is the learning rate in NN, and wjm is the weights between the hidden layer
and the output layer. Then, the weights wij between the input layer and the hidden
layer can be updated as
∂f ∂Om ∂Hj
wij = wij − η (14)
∂Om ∂Hj ∂wj m
By applying both the methods of DE type NN and BP type NN, the NNs are trained
to converge to the optimal values with proper weights. Outputs of the BP and DE
type NNs are shown in Fig. 5. The figure shows that the outputs of both NNs
converged to the optimal (x,y) to minimize the value of the function in hundreds
of iterations. Comparing with the convergence of the BP type NN, the DE type NN
showed better convergence characteristics. The errors of the minimum value and the
100000
90000
80000
70000
60000
50000
40000
30000
20000
10000
0
4
2
0
−4 −2 −2
y
0 2 −4
x 4
24 S. Mu et al.
0.0 y
Ii Om
Hj
0.5
1
Error
0.5
NNs’ outputs are shown in Fig. 6. The variation of the errors shows the same trends
as we see in Fig. 5. The variations of weights in the BP and DE type NNs are shown
in Fig. 7. The DE type NN with the mutation strategy of DE/best/1/bin. It can be
Study on NN Based on BP and DE 25
Weights
0
-5
0 250 500 750 1000
Iteration
0.5
0 25 50 75 100
Iteration
seen that the DE update the NN’s weights in iterations. The weights can converge
to certain constant values in short time. The values of weights converged quick in
around 100 iterations. Comparing the proposed DE type NN with the traditional BP
type NN, the proposed scheme was with an excellent convergence features.
To confirm the conference characteristics of DE with different mutation strate-
gies, the simulation study was implemented in the optimization of Rosenbrock func-
tion using the proposed NN. The DEs with the mutation strategies of DE/best/1/bin,
DE/current to/1/bin, DE/rand/1/bin and DE/rand/2/bin were investigated [13]. The
variation of the evaluation value was studied as Fig. 8 shows. According to the
simulation results, it is clear to see that among the DEs with the strategies, the
DE/best/1/bin strategy converged to quite low evaluation value in only a few
iterations. The DE/best/1/bin strategy is with great convergence characteristics to
approach the optimal. Therefore, the DE/best/1/bin is considered as the best solution
for experimental study.
26 S. Mu et al.
Gains
r(k) + e(k) u(k) y(k)
PID USM
−
DE
z −1
NN
z −2
4 Experimental Study
Based on previous research in [7], an intelligent control method using PID control
combined with DE type NN is applied as the controller for USM as shown in Fig. 9.
The NN scheme we propose is a fixed structure as we see in Fig. 10. In the proposed
intelligent scheme, DE type NN is employed for updating the PID gains. The error
signal of [e(k), e(k − 1), e(k − 2)] is employed as input. The scheme is designed for
tuning PID gains automatically to minimize the error in USM control. The output
of the NN unit is the variation of PID gains [ΔKP (k), ΔKI (k), ΔKD (k)]. The output
of neurons in hidden-layer is expressed as Hj (k), which can be estimated by Eq. 15.
1
fs = − 0.5 (17)
1 + e−x
The weights between three layers, expressed by wij (k) and wjm (k), are updated by
BP and DE algorithms. Then, the output of PID controller u(k) can be calculated as
In our experimental study, the sin wave was applied as the input signal. The
response of the proposed method is shown in Fig. 11. There was no visible error
in the response of sin wave input. Figure 12 shows the variation of the PID gains
updated by the proposed scheme using DE type NN. The gains oscillated in the
Study on NN Based on BP and DE 27
i m ΔKp (k)
e(k)
-45
0 4 8
Time[sec]
initial phase in a short period. Then, they converged to certain values quickly. The
convergence and accuracy of the proposed method is confirmed according to the
experimental results.
5 Conclusions
In this chapter, an intelligent PID control scheme using the DE type NN is proposed.
In the proposed scheme, an NN method is employed for optimizing the gains in PID
controller for the control of ultrasonic motor. The weights of the NN are designed to
be updated by DE algorithm. By employing simulation study, the DE algorithm was
28 S. Mu et al.
Gain
5
0
0 4 8
Time [sec]
confirmed effective in the learning of NN. According to the simulation results, the
mutation strategies were investigated. The DE/best/1/bin was the most effectiveness
in the optimization process. By the experimental study, the DE type NN was
confirmed effectiveness in the PID control of the USM. The proposed method is able
to be applied for achieving high control performance in compensating characteristic
changes USM. Meanwhile, its application is without any consideration of Jacobian
estimation in NN’s leering. By employing the method, USMs are expected more in
various applications, such as the meal-assistance robots [14], human support robots,
even some humanoid robots [15], especially for the applications in medical and
welfare fields in the future.
References
1. Tsoukalas, L.H., Uhrig, R.E.: Fuzzy and Neural Approaches in Engineering. Wiley, New York
(1997)
2. Srinivas, M., Patnaik, L.M.: Genetic algorithms: a survey. Computer 27(6), 17–26 (1994)
3. Wang, L.: A hybrid genetic algorithm – neural network strategy for simulation optimization.
Appl. Math. Comput. 170(2), 1329–1343 (2005)
4. Senjyu, T., Miyazato, H., Uezato, K.: Position control of ultrasonic motor using neural network
(Japanese). Trans. Inst. Electr. Eng. Jpn. D 116, 1059–1066 (1996)
5. Oka, M., Tanaka, K., Uchibori, A., Naganawa, A., Morioka, H., Wakasa, Y.: Precise position
control of the ultrasonic motor using the speed compensation type NN controller (Japanese). J.
Jpn. Soc. Appl. Electromagn. Mech. (JSAEM) 70(6), 1715–1721 (2004)
6. Mu, S., Kanya, T.: , Intelligent IMC-PID control using PSO for ultrasonic motor. Int. J. Eng.
Innov. Manag. 1(1), 69–76 (2011)
7. Mu, S., Tanaka, K., Nakashima, S., Alrijadjis, D.: Real-time PID controller using neural
network combined with PSO for ultrasonic motor. ICIC Exp. Lett. 8(11), 2993–2999 (2014)
8. Jordehi, A.R., Jasni, J.: Particle swarm optimisation for discrete optimisation problems: a
review. Artif. Intell. Rev. 43, 243–258 (2015)
Study on NN Based on BP and DE 29
9. Das, S., Abraham, A., Konar, A.: Particle swarm optimization and differential evolution
algorithms: technical analysis, applications and hybridization perspectives. Stud. Comput.
Intell. 116, 1–38 (2008)
10. Ilonen, J., Kamarainen, J., Lampinena, J.: Differential evolution training algorithm for feed-
forward neural networks. Neural Process. Lett. 17, 93–105 (2003)
11. C. Zhao, Ultrasonic Motor - Technologies and Applications. Science Press/Springer Bei-
jing/Berlin (2011)
12. Basheer, I.A., Hajmeer, M.: Artificial neural networks: fundamentals, computing, design, and
application. J. Microbiol. Methods 43, 3–31 (2000)
13. Islam, S.M., Das, S., Ghosh, S., Roy, S., Suganthan, P.N.: An adaptive differential evolution
algorithm with novel mutation and crossover strategies for global numerical optimization.
IEEE Trans. Syst. Man Cybern. B (Cybernetics) 42(2), 482–500 (2012)
14. Tanaka, K., Kodani, K., Oka, M., Nishimura, Y., Farida, F.A., Mu, S.: Meal assistance robot
with ultrasonic motors. Int. J. AEM 36, 177–181 (2011)
15. Zhou, C., Wang, X., Li, Z., Tsagarakis, N.: Overview of gait synthesis for the humanoid
COMAN. J. Bionic Eng. 14(1), 15–25 (2017)
Word Sense Disambiguation Based
on Graph and Knowledge Base
Fanqing Meng
F. Meng ()
School of Information Science and Engineering, Zaozhuang University, Zaozhuang, China
e-mail: [email protected]
by researchers. This paper attempts to use HowNet and the Chinese Semantic
Dictionary as the knowledge base for Chinese WSD tasks.
1 Related Work
Most WSD methods based on graph are inspired by lexical chain. The lexical
chain means some semantic related words in a given text that form a sequence
and the words that are linked together through lexical semantic relationships.
Galley et al. put forward a method of WSD based on lexical chain [7], which is
mainly divided into two steps: first, when constructing the graph model, senses
of the target ambiguous word are added to the graph, the words in the text are
processed sequentially, and the senses of each word are compared with the words
that have been processed. If there is some semantic relationships between them,
then the relationship is used as a link edge in the graph, and the weight is arranged
according to the semantic relation and the distance between the words. Second, in
the disambiguation stage, the weight of each word sense and context-linked word
node is computed, and the word sense with the largest weight is selected as the
correct word sense of the target word. This method achieves 62.1% disambiguation
accuracy on the noun dataset of SemCor.
Mihalcea proposes a disambiguation method based on the PageRank algorithm
[8], which takes the senses of each word in the text as the vertex and constructs the
disambiguation graph by using the semantic relation between the senses of words.
In addition to the semantic relationships in WordNet, it includes a coordinating
relation to link the sense of the same upper concept. In this paper, the PageRank
algorithm is used to calculate the importance of each sense node iteratively, and
the maximum score of the senses of ambiguous words is chosen as its final sense.
Navigli et al. propose the structured semantic interconnect algorithm (SSI) [9].
With the help of knowledge resources such as WordNet, the algorithm structurally
expresses the senses of contextual words; it constructs grammar rules by manual
means to describe the interconnection patterns in semantic concept mappings; it
performs disambiguation iteratively, giving priority to ambiguous words with less
difficulty. According to the semantic interconnection mode of words that have
been determined, the disambiguation words are processed. This method achieves
the best results in the international semantic evaluation Senseval-3 and SemEval-
2007. Agirre et al. propose the Personalized PageRank algorithm for WSD [10, 11]
which is similar in principle to Mihalcea’s method [8] and the difference is that
it has some improvements in the importance of degree computation, to modify the
importance of some word nodes. Nieto Pina et al. present a graph-based WD method
for Swedish, and it trains embeddings by using PageRank on the graph of lexical
knowledge base (LKB) [12]. Janz et al. propose a WSD method for English and
Polish based on Monte Carlo approximation and PageRank [13]. It utilizes the local
semantic relatedness by an extensive knowledge base. Scozzafava et al. propose a
knowledge-based multilingual WSD system, and it performs WSD using PageRank
Word Sense Disambiguation Based on Graph and Knowledge Base 33
and syntagmatic information [14]. These methods mainly focus on the English WSD
task, but the Chinese semantic computing resources are relatively scarce, which
limits the popularization and application of these methods in Chinese WSD tasks.
Yang et al. propose a network graph disambiguation method based on word
distance [15]. The method performs word sense disambiguation according to the
distance between the vertices of the words in the network graph. The vertices of
words closer to ambiguous words are strengthened, and the vertices of distant words
are weakened, that is, words with close distance have a strong recommendation
effect on ambiguous words, and words with far distance have weak recommenda-
tion. Lu et al. proposes a graph model disambiguation method based on domain
knowledge [16]. It applies the domain knowledge to the graph model and improves
the importance of degree scores of the sense nodes by improving the various graph
scoring methods and then selects the correct sense. Although these methods improve
the WSD effect, they do not explore and utilize the existing Chinese word sense
computing resources.
2 Method
HowNet reveals the relations between concepts and the attributes of concepts
[17], such as hypernym, hyponymy, synonymy, and meronymy. The traditional
HowNet-based WSD method has the problem of insufficient knowledge utilization
of HowNet. It does not fully exploit the disambiguation knowledge existing in
HowNet. For example, the example sentences in HowNet have a strong ability to
distinguish word sense. By digging deep into the semantic knowledge contained
in HowNet, it can be used as the basis for WSD, which will contribute to the
improvement of WSD performance.
This paper proposes a WSD method based on graph and knowledge base. It ana-
lyzes the ambiguous sentences, obtains the context-dependent tuples, and obtains
the dependency graph to construct the context disambiguation graph. In HowNet,
the target ambiguous example sentences are obtained, and the dependency parsing
is performed to get the dependent tuple set. Combining with context-dependent
tuples, construct dependent disambiguation graph, fuse context disambiguation
graph and dependent disambiguation graph, use graph scoring algorithm to score
the importance of degree of word sense vertices, and select the one which gets the
biggest score in the senses of target ambiguous word as the correct word sense.
2.1 HowNet
HowNet is a common-sense knowledge base that Professor Dong spends more than
10 years organizing and constructing [17]. It builds a network of knowledge base
based on the relation between the concepts of Chinese and English, and the attributes
34 F. Meng
The Chinese Semantic Dictionary (CSD) is a semantic knowledge base for natural
language processing, and it provides a huge amount of semantic information, which
Word Sense Disambiguation Based on Graph and Knowledge Base 35
includes more than 66,539 entries, Chinese words and their English counterparts
[18]. Examples sentences in CSD can be extracted as a supplement for HowNet.
The disambiguation framework of the WSD method based on graph and knowledge
base proposed in this paper is shown in Fig. 1. Obtain dependency parsing of
ambiguous sentence, obtain context-dependent tuples, get dependency graph, and
construct context disambiguation graph; obtain example sentences in HowNet,
perform dependency parsing, get dependent tuple sets, and combine context-
dependent tuples to construct a dependency disambiguation graph; fuse the context
disambiguation graph and the dependency disambiguation graph to perform word
sense disambiguation. The specific description is as follows:
1. Obtain dependency parsing of ambiguous sentences, obtain context-dependent
tuple sets, and get a context-dependent graph; according to the shortest path
length of the context words in the dependency graph from the target ambiguous
word, the context words are extracted as context knowledge.
2. Use HowNet to label the senses of context knowledge and target ambiguous
word, which may exist in HowNet, and compute the similarity among them.
3. Take the senses of the word in the context knowledge as the vertex and the target
ambiguous word as the vertices, the semantic relation between concepts as the
edge, the concept similarity as the weight of the edge, to construct the context
disambiguation graph.
4. Obtain the example sentences of the senses of the target ambiguous word
in HowNet (and CSD), perform dependency parsing to obtain the example
sentence-dependent tuple set, and construct the dependency disambiguation
graph in combination with the context tuple set.
.
Dependency Content Content HowNet similarity
Sentence
parsing dependency graph knowledge computation
Correct Merge
HowNet, CSD Graph score
sense disambiguation graph
5.
He arrived at the Café de Paris about half past eight, keyed up and
exalted. The afternoon seemed to have added an inch to his stature.
The most charming woman in Monte!... And he had kissed her....
Having a little time to spare he went into the Rooms. The impulse he
knew so well was luring him on.
“If I give her a mille,” he thought, “why not make the Casino pay for
it?”
At the first table he threw a maximum on red and won. Rather sourly
they paid him six thousand francs.
“Five thousand for me,” he said, “and a thousand for Mrs. B.”
He found her exquisitely dressed and impatiently awaiting him. His
luck had added to his excitement. He was feeling capable of any
folly. He folded up a mille note very neatly and slipped it into her bag.
Then somewhat to his chagrin she excused herself.
“I’m awfully sorry. I have an engagement at the club. Probably see
you to-morrow.”
She left him feeling rather chilled and sober.
“Well,” he consoled himself, “perhaps it is just as well. In any case I
now have twenty-five thousand francs.”
He did not play for the next few days. His success rather dismayed
him; his nerve was gone. Then on two successive days a chance
appeared which was too favourable to be resisted; it was what he
called a sure-shot. As a matter of fact, it really came off in five cases
out of six, but only by watching and waiting could one get on to it. A
long run on one of the simple chances had to be followed by a
certain combination of both chances. In each case when he played it
he won; his capital had now increased to thirty-five thousand francs.
He began to feel what he had never felt in his life before,—secure.
To a rich man that sum was nothing; to him, everything. But he was
becoming increasingly nervous. He must not lose this money. He
must be more prudent than ever. If he lost it, it would discourage him
utterly; he would never have the heart to begin over again. Yet he
must make the sixty thousand on which he had set his heart. Only
twenty-five thousand more. Four wins of maximums would practically
do it. Well, he must pluck up his courage and try again. He must
nerve himself for the final struggle.
CHAPTER EIGHT
THE BIG FIGHT
DURING his long spells of waiting, he became more and more
intimately acquainted with the world of the Casino,—from the
prowling seeker of the sure-shot to the holiday plunger, from the
philosophic veteran with pencil and note-book to the nouveau riche
spraying the table with louis. He came to know the eternal types, the
avid-eyed old women, the blowsy, brazen matron, the cocotte
throwing money away with cynical contempt, the young girl from the
convent risking her first five franc piece.
Then the unending Casino comedies. For instance, the Honeymoon
Couple comedy. The first day she hangs on his arm while he plays.
The second day he hangs on her arm while she plays. The third day
they separate a little and try tiny flutters of their own. The fourth day
they get seats at the same table and advise each other as to their
play. The fifth day they get seats at separate tables, and each plays
as if the other did not exist. The sixth day she is begging him for
money, and he is refusing. The seventh day they moon round
without playing, he moody, she sulky, very near to a quarrel. Then on
the eighth day they disappear, perhaps never to be the same again.
Then there is the woman who talks to the croupier, fawning on him
and asking him how she ought to play. Sometimes the knight of the
rateau hazards a guess. If it comes off, tant mieux. Perhaps, it
means a tip; if it doesn’t, tant pis. No man is infallible.
These types repeat themselves endlessly; but among them from
time to time appear strange original characters piquing the curiosity
of the public. The Casino is like a stage where they enter, play their
parts, and make their exit. Old and young, good and bad, rich and
poor,—they come and go; they lose or win; they sidle across the
glossy floor under the great white dome; they smirk and posture,
wrangle and vapour. Beefy Englishman and desiccated Yank,
flatulent Frenchman and oily Italian, morose Spaniard and bovine
Swede; Jap, Chinaman, Rasta and Levantine Jew—they mix amid
the throng that surges around the whirring wheels, and their strange
tongues mingle in one confused babble.
So fantastic did it appear to Hugh that at times he rubbed his eyes
and wondered if it could be real. How he wished he were a writer. If
only he could see into their hearts, know their histories, pick their
brains, what books he could write, a library of books, a document of
humankind that would outweigh the works of Balzac and Zola
combined.
He became better acquainted with daily frequenters of the Casino
and watched them with unceasing amusement. There was one, a
Greek, a grossly fat man with three chins and a promontory of
greasy waistcoat. On his pudgy hand he wore rubies as large as
walnuts; and as he walked from table to table, laying mille placques
on the dozens he never ceased to perspire. The tall woman in grey
also attracted him. She was always gloved, always veiled. She
played with persistency her game on number one, losing or winning
with apparent indifference. Who was she? She went from table to
table with the strangeness and mystery of a specter.
His attention was drawn irresistibly to the very tall man with the
spade-shaped beard, who, he had learned, was a Brazilian diplomat.
His name was Doctor Bergius and he was said to be of mixed
Spanish and German parentage. From the moment he had entered
the Rooms, the doctor had become a dominating personality. He
was as straight as a shoot of bamboo, with a high carriage and an
eye of piercing command. His long nose resembled the beak of a
vulture of the Andes; his brow retreated from his piercing black eyes
and his skin was as coppery as that of an Indian. He always dressed
with immaculate care. He never played, but looking gravely on, with
his hands behind his back, appeared only remotely interested in the
game.
Then there was another man who attracted Hugh, partly because he
was so graceful and handsome, partly because he had once broken
the bank. He was an Italian called Castelli, of medium height and
well-shaped figure. He had the olive skin, dark velvety eyes, and the
perfect features often found in men who prey on women. He always
had women with him. He played a dashing game, superbly
indifferent to loss or gain.
Hugh was conscious of an atmosphere of unceasing suspicion.
Every day hundreds of false louis were foisted on the bank, and no
one knew how it was done. It was impossible to check them. Most of
the players were too lazy to cash in before leaving the Casino and
the chips continued to circulate in town as freely as money. The
Casino inspectors were up on their toes, every one was being
watched.
One day there came to him the mood for which he was waiting, the
conquering mood. His nerves were of steel, and he felt that he could
win a fortune or lose all without turning a hair. He had in his pockets
seven packets of five thousand francs each. He sat down at the
suicide table and began to play.
It was strange how confident he was. Something was fighting on his
side. He could not be beaten. Then to his dismay, he lost the first
coup, but won the three following. Again he lost a coup, but gained
another two. And so it went on. The scales dipped, now for, now
against him. On the whole, however, fortune favoured him, and he
steadily drew ahead.
At last the great battle was on. He felt inspired. Silently he sat,
watchful, emotionless as a wax figure. His eyes became opaque;
and a crease of concentration came between his brows.
“A big game,” the rumour spread. “Come and see.”
The circle of watchers deepened and their interest increased. But to
Hugh they were only a dull blur of meaningless faces. No, not all, for
he saw one that thrilled him for a moment. It was that of Mrs.
Belmire. Then her too he forgot in the stress of the struggle.
And it was in reality a struggle. The bank had abandoned its attitude
of disdain, and was directly fighting him. No longer was he a petty
“piker” but a foeman worth while. He felt that the eight silent men in
black who ran the table were concentrating their wills against his.
Psychology was coming into play. He willed that that capricious little
ball should go one way; these eight willed it to take another. He
would beat them; he would make the ball go where he wanted it to
go.
He waited until it was spun; and then, acting from sheer impulse,
threw a packet of notes on one of the simple chances. Sometimes
he even threw on two chances. The battle swayed. He advanced, he
retreated, but only to advance again.
As the spirit of the fight glowed in him, his play increased in
boldness. If he fell he would fall gloriously. For once he had the
centre of the stage. He would be worthy of the part, of the audience.
He began to play on all three simple chances. He won ... and won
again. Fortune favoured the brave. Hurrah for the big game! Again a
triple shot. Ha! he had lost one of the three that time.... Well, the next
time. The grey croupier who was paying did not look at all pleased.
No wonder! Handing out mille notes continually from the little grilled
box. And the chef de parti was scowling at the little chap who twirled
the ball. Hugh won another coup, another fifteen thousand.
“Messieurs, le boule passe.”
The cowards! They were changing the croupier who threw the ball.
That one was too unlucky. What a stack of notes he had. Must be
nearly a hundred thousand. Ah, the devil! this new man was beating
him. He had lost ... lost ... lost. His stack was diminishing. His luck
had turned. He heard people asking: “Why doesn’t he stop?” No, he
would win back again ... win back. He was dazed. He scarcely knew
what he was doing. He was reaching out to play another time when a
hand gripped his arm.
“You darned young fool, quit now. Quit while you’re still to the good.”
It was Mr. Gimp. The American dragged him away from the table,
followed by the admiring gaze of a hundred eyes. Then in a corner
they counted the gains. With his own capital he had fifty-five
thousand francs.
He sat on a leather seat in a stupor. He wanted to smoke a cigarette,
to go home, but did not have energy enough. Then, when he was
finally starting the stranger’s door swung open and Professor Durand
made his first entrance into the Rooms.
CHAPTER NINE
THE PROFESSOR BEGINS
1.
2.
3.
4.
On the fourth day the professor again made his appearance in the
afternoon; and again repeated his performance of the second day.
The directors were in a state of consternation. For the first time they
began to think that here was that inconceivable thing, a man with an
invincible system. All Monte Carlo was talking of it. They awaited
events with the greatest anxiety.
CHAPTER TEN
THE COMPACT
1.
AS the following day was Sunday the professor did not play. Instead
he went to early mass accompanied by Margot, to whom he had
taken a fancy, and in the afternoon the two went for a walk together
in the direction of Menton. They asked Hugh to accompany them but
he refused.
To tell the truth he had other fish to fry. Since his success, he had
blossomed out in many ways. Some time before he had been to a
good tailor and had several suits made. He had developed a
fondness for the American Bar of the Café de Paris, and had
become increasingly intimate with Mrs. Belmire.
At first he had merely admired her immensely; he was flattered by
her interest in him and had found a kind of heady pleasure, like
champagne, in her society. In the end he had become really
fascinated. Then one day something happened that acted on him
like a spur. He was walking along the road to Beaulieu when a big
carmine car swung round a bend. He recognized Vulning at the
wheel. He thought the lady with him seemed familiar, but it was not
until she looked back laughingly and waved her hand that he saw it
was Mrs. Belmire.
If it had been any other man he would not have minded. He knew
she went about with Fetterstein and one or two others. But
Vulning!... He disliked Vulning. He realized with a shock that he was
jealously infatuated with the lady.
On this particular Sunday afternoon he had arranged to motor with
her to Brodighera. He was happy in the thought that he would have
her all to himself for four radiant hours.
They had not been long on the road when she said sweetly:
“You poor boy! I believe we nearly ran you down the other day. You
looked ever so startled. Paul always drives so recklessly. He gives
me all kinds of thrills.”
“Why do you go out with him then?” Hugh snapped.
She looked at him curiously. There was a gleam of amused
satisfaction in her eye.
“Why shouldn’t I? I love thrills. Besides he’s an old friend. Good
gracious! I don’t believe you like him.”
“Why should I?”
“Oh, he’s not a bad sort. A bit blasé and all that sort of thing. But ...
well, you don’t know him.”
“We have met.”
“Have you really? I asked him after we had passed if he knew you,
and he said he didn’t. He meets so many I suppose he forgot. He’s
much sought after.”
Hugh grew more savage. Again he snapped. “I don’t believe he’s the
man a woman should be seen with often if she values her
reputation.”
“Her reputation!”
Hugh thought she was going to be angry. She stared at him a
moment, then suddenly said: “Oh you funny boy! You nice boy! I like
you so much for that. But ... I do believe you’re jealous of Paul
Vulning.”
“Why should I be?”
“Because he’s so like you. You’ve got the same regular features, the
same fair hair. You’re both tall and rather slight. But you needn’t
resent all that. Paul is a rather soiled man of the world; you are a
nice clean boy who might be fresh from Oxford.”
She looked at him thoughtfully.
“I’d like to keep you as you are, but it might be better for yourself if
you were a little more sophisticated. I have half a mind to take you in
hand, to educate you in the ways of the world. May I?”
“I don’t know what you mean.”
“No, I didn’t think you would. You are a little—may I say it?—naïve
sometimes.... I say, don’t you admire my new blouse? I just got it this
morning. The embroidery’s so pretty, I think.”
Something in the design struck Hugh as familiar. Yes, it was the work
he had seen Margot sitting up so late to finish.
“You look so serious. Don’t you like it?”
“Yes, it’s lovely.”
“I think so too. By the way, why don’t you begin to gamble again?
You play so well and you are so lucky. You must join the Club. You
might make lots of money. Talking about horrid money reminds me ...
mine hasn’t arrived and I’m rather low again. You may lend me
another mille if you like.”
On the way back they overtook an old man and a girl. It was Margot
and the professor. Margot’s arms were full of flowers. Hugh waved to
them, and the professor waved back. Margot only stared. Hugh felt
uncomfortable. He was sorry he had encountered them.
When he reached home the room was bright with blossoms. Margot
was very quiet. She did not refer to their meeting and neither did he.
But for some reason for which he could not account, he did not care
to meet her eyes.
2.
On his way to the Casino, the next day, he missed the usual crowd
at the edge of the “Cheese,” and hunting up Mr. Tope inquired after
Mrs. Fitzoswald.
“Haven’t you heard? The Casino capitulated. They had the Major
released and paid the fares of the whole family to England. They all
went off this morning.”