Informatics Networking and Intelligent Computing

Proceedings of the 2014 International Conference

on Informatics Networking and Intelligent
Computing INIC 2014 16 17 November 2014 Shenzhen
China 1st Edition Jiaxing Zhang (Editor) pdf
https://ebookgate.com/product/informatics-networking-and-
download
intelligent-computing-proceedings-of-the-2014-international-
conference-on-informatics-networking-and-intelligent-computing-
inic-2014-16-17-november-2014-shenzhen-china-1st-e/

Get Instant Ebook Downloads – Browse at https://ebookgate.com

 Editor
Zhang Informatics, Networking
and Intelligent Computing

Editor: Jiaxing Zhang

Informatics, Networking and Intelligent Computing collects contributions to the

and Intelligent Computing

Informatics, Networking
2014 International Conference on Informatics, Networking and Intelligent Computing
(Shenzhen, China 16-17 November 2014). The topics covered include:
- Computational intelligence
- Networking technology and engineering
- Systems and software engineering
- Information technology and engineering applications, and
- Signal and data processing

Informatics, Networking and Intelligent Computing not only provides the state-
of-the art in informatics and networking technology, but also offers innovative
ideas present and future problems. The book will be invaluable to academics and
professional involved in informatics, networking and intelligent computing.

an informa business
INFORMATICS, NETWORKING AND INTELLIGENT COMPUTING
This page intentionally left blank
PROCEEDINGS OF THE 2014 INTERNATIONAL CONFERENCE ON INFORMATICS,
NETWORKING AND INTELLIGENT COMPUTING (INIC 2014), 16–17 NOVEMBER 2014,
SHENZHEN, CHINA

Informatics, Networking and

Intelligent Computing

Editor

Jiaxing Zhang
Wuhan University, Wuhan, Hubei, China
CRC Press/Balkema is an imprint of the Taylor & Francis Group, an informa business
© 2015 Taylor & Francis Group, London, UK
Typeset by MPS Limited, Chennai, India
All rights reserved. No part of this publication or the information contained herein may be reproduced,
stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, by
photocopying, recording or otherwise, without written prior permission from the publishers.
Although all care is taken to ensure integrity and the quality of this publication and the information
herein, no responsibility is assumed by the publishers nor the author for any damage to the property or
persons as a result of operation or use of this publication and/or the information contained herein.
Published by: CRC Press/Balkema
P.O. Box 11320, 2301 EH Leiden, The Netherlands
e-mail: [email protected]
www.crcpress.com – www.taylorandfrancis.com
ISBN: 978-1-138-02678-0 (Hardback)
ISBN: 978-1-315-73453-8 (Ebook PDF)
 Informatics, Networking and Intelligent Computing – Zhang (Ed.)
© 2015 Taylor & Francis Group, London, ISBN: 978-1-138-02678-0

Table of contents

Preface IX
Organizing committee XI

Computational intelligence
Decomposition genetic algorithm for cellular network spatial optimization 3
M. Livschitz
A heating and cooling model for office buildings in Seattle 9
W.Q. Geng, Y. Fu & G.H. Wei
Multi-depth Deep Feature learning for face recognition 15
C.C. Zhang, X.F. Liang & T. Matsuyama
Research on camera calibration basing OpenCV 21
H.M. Nie
Designing fuzzy rule-based classifiers using a bee colony algorithm 25
I.A. Hodashinsky, R.V. Meshcheryakov & I.V. Gorbunov
City management based on Geospatial Business Intelligence (Geo-BI) 35
Y.L. Zhou & W.J. Qi
Research into the development mode of intelligent military logistics 39
F. Zhang, D.R. Ling & M. Wang
Incomplete big data imputation algorithm using optimized possibilistic c-means and
deep learning 43
H. Shen & E.S. Zhang
Human-machine interaction for an intelligent wheelchair, based on head poses 49
Y. Wang, N. Liu & Y. Luo
An optimization model in the design of a product process 55
T. Qi, S.P. Fang & C.Q. Liu

Networking technology and engineering

A new SIFT feature points restoration based on a watermarking scheme resilient to
geometrical attacks 61
O.J. Lou, S.H. Li, Z.X. Liu & S.T. Tang
Using CALL (Computer-assisted Language Learning) to achieve multidimensional
college English teaching 67
W. Liu
Reflections on multimedia teaching 71
W.G. Chang
Electromechanic installations vibration acceleration protection system 75
V.I. Erofeev, A.S. Plehov & D.U. Titov
The study of CBI theme-based teaching mode of college English from multiple intelligence
module perspective 79
W. Liu

V
The analysis of access control model based on Single Sign-on in SOA environment 83
G.Z. Wang, B. Zhang, X.F. Fei, Y. Liu, H.R. Gui & H.R. Xiong
An Android malware detection method using Dalvik instructions 89
K. Zhang, Q.S. Jiang, W. Zhang & X.F. Liao
Identification of spoofing based on a nonlinear model of an radio frequency power amplifier 95
Y.M. Gan & M.H. Sun
Computational model for mixed ownership duopoly competition in the electricity sector
with managerial incentives 101
V. Kalashnikov-Jr., A. Beda & L. Palacios-Pargas

Systems and software engineering

A software reliability testing theory and technology research 107
H.L. Sun, X.Z. Hou, K. Zh & H.F. Luo
Fingertips detection and tracking based on a Microsoft Kinect depth image 113
Z.X. Li, J. Liu, H.C. Wu & Z.M. Chen
A virtual dressing room approach based on Microsoft Kinect 117
J.F. Yao, L. Lysandra, L. Yang, B.R. Yang & Z.C. Huang
ASM (Active Shape Model) modeling of the human body and its application in virtual
fitting 123
X.Y. Xiong & X.J. Zhu
Building an orchestration architecture for cloud services: A case study of designing a platform
as a service (PaaS) runtime environment 127
P.C. Chen, Y.T. Huang, Y.C. Lee & C.C. Chu
Development of an MIPI (Mobile Industry Processor Interface) interface camera driver
based on WINCE (Windows Embedded Compact) 131
K. Xiao, L. Shan & Z.T. Li
Trends in the development of databases on statistics in the OECD, the EU and Russia 135
N. Chistyakova, V. Spitsin, J. Abushahmanova & N. Shabaldina
The effect of casting geometry on the thermal gradient in A201 aluminium alloy
plate castings 139
Y.S. Kuo & M.F. Lu
A research on multi-implementation game product-based learning for game development
specialty students 143
C. He
A network behaviour analyser: Automatic fingerprint extraction from functions of
mobile applications 147
P. Liu & C.Y. Wu

Information technology and engineering application

Design of dipole array antenna for a 2.4-GHz wireless local area network application 155
Y.Y. Lu & K.C. Liao
A Token-based Network Communication Library (TBNCL) in a private cloud storage system 159
Q. Wang, L. Li, Z.H. Guo, M. Lin & R. Pan
Analysis of phased array antenna’s vibration effects on the performance of shipborne MLS 163
H.S. Xie, P. Zhou, J.G. Wei, B.K. Luan & D. Wang
Application of PUS (Packet Utilization Standard) and XTCE (XML Telemetric and
Command Exchange) in satellite telemetry data exchange design and description 169
Y. Liu, J.Q. Li & Z.D. Li

VI
Information system designing for innovative development assessment of the efficiency
of the Association of Innovative Regions of Russia members 173
V.V. Spitsin, O.G. Berestneva, L.Y. Spitsina, A. Karasenko, D. Shashkov &
N.V. Shabaldina
Selected aspects of applying UWB (Ultra Wide Band) technology in transportation 177
M. Džunda, Z. Cséfalvay & N. Kotianová
Design of a wireless monitoring system for a Pleurotus eryngii cultivation environment 183
L. Zhao & X.J. Zhu
Research on Trellis Coded Modulation (TCM) in a wireless channel 189
X.M. Lu, F. Yang, Y. Song & J.T. He
Riemann waves and solitons in nonlinear Cosserat medium 193
V.I. Erofeev & A.O. Malkhanov
Research on the adaptability of SAR imaging algorithms for squint-looking 197
M.C. Yu
Improved factor analysis algorithm in factor spaces 201
H.D. Wang, Y. Shi, P.Z. Wang & H.T. Liu
Research on the efficacy evaluation algorithms of Earth observation satellite mission 207
H.F. Wang, Y.M. Liu & P. Wu
An image fusion algorithm based on NonsubSampled Contourlet Transform and Pulse
Coupled Neural Networks 211
G.Q. Chen, J. Duan, Z.Y. Geng & H. Cai
A cognitive global clock synchronization algorithm in Wireless Sensor Networks (WSNs) 215
B. Ahmad, S.W. Ma, L. Lin, J.J. Liu & C.F. Yang
A multi-drop distributed smart sensor network based on IEEE1451.3 219
H.W. Lu, L.H. Shang & M. Zhou
Solitary strain waves in the composite nonlinear elastic rod 225
N.I. Arkhipova & V.I. Erofeev
Semiconducting inverter generators with minimal losses 227
A.B. Daryenkov & V.I. Erofeev
Research into a virtual machine migration selection strategy 231
L. Sun & X.Y. Wu
An analysis of the influence of power converters on the operation of devices 235
A.I. Baykov, V.I. Erofeev & V.G. Titov

Signal and data processing

The classification of insect sounds by image feature matching based on spectrogram analysis 241
A.Q. Jia, B.R. Min & C.Y. Wei
Research on business model innovation method based on TRIZ and DEA 247
X. Liu, J.W. Ding & X.Q. Ren
Analytical solution for fuzzy heat equation based on generalized Hukuhara differentiability 251
T. Allahviranloo, Z. Gouyandeh & A. Armand
Identification of space contact for a dynamics medium 257
V.S. Deeva, M.S. Slobodyan, G.A. Elgina, S.M. Slobodyan & V.B. Lapshin
Membership functions of fuzzy sets in the diagnosis of structures pathology 261
G.G. Kashevarova, M.N. Fursov & Y.L. Tonkov
Global stock market index analysis based on complex networks and a multiple
regression model 265
Z.L. Zhang & S.J. Qiao

VII
A study of sign adjustment of complete network under the second structural theorem 269
H.Z. Deng, J. Wu, Y.J. Tan & P. Abell
Sybil detection and analysis of micro-blog Sina 273
R.F. Liu, Y.J. Zhao & R.S. Shi
A kinematics analysis of actions of a straddled Jaeger salto on uneven bars performed by
Shang Chunsong 279
L. Zhong, J.H. Zhou & T. Ouyang

Author index 283

VIII
 Informatics, Networking and Intelligent Computing – Zhang (Ed.)
© 2015 Taylor & Francis Group, London, ISBN: 978-1-138-02678-0

Preface

The 2014 International Conference on Informatics, Networking and Intelligent Computing (INIC2014) will be
held in Shenzhen, China on November 16–17, 2014. The main purpose of this conference is to provide a common
forum for experts and scholars of excellence in their domains from all over the world to present their latest and
inspiring works in the area of informatics, networking and intelligent computing.
The informatics is helpful for people to make full use of information technology and means to improve the
efficiency of work. In recent years, the networking technology has experienced a rapid development and it is
widely used in both our daily life and industrial manufacturing, such as games, education, entertainment, stocks
and bonds, financial transactions, architectural design, communication and so on. Any company or institution
with ambition cannot live without the latest high-tech products. At present, intelligent computing is one of the
most important methods of intelligent science and also is the current topic of the information technology. For
example, machine learning, data mining and intelligent control have become the hot topics of current research.
In general, informatics, networking and intelligent computing have become more and more essential to people’s
life and work.
INIC2014 has received a large number of papers and fifty-eight papers were finally accepted after reviewing.
These articles were divided into several sessions, such as computational intelligence, networking technology and
engineering, systems and software engineering, information technology and engineering application and signal
and data processing.
During the organization course, we have received much help from many people and institutions. Firstly, we
would like to show our thankfulness to the whole committee for their support and enthusiasm. Secondly, we
would like to thank the authors for their carefully writing. Lastly, the organizers of the conference and other
people who have helped us would also be appreciated for their kindness.
We wish all the attendees at INIC2014 can enjoy a scientific conference in Shenzhen, China. We really
hope that all our participants can exchange useful information and make amazing developments in informatics,
networking and intelligent computing after this conference.
INIC2014 Committee

IX
This page intentionally left blank
 Informatics, Networking and Intelligent Computing – Zhang (Ed.)
© 2015 Taylor & Francis Group, London, ISBN: 978-1-138-02678-0

Organizing committee

Honor Chair
E.P. Purushothaman, University of Science and Technology, India

General Chair
Jun Yeh, Tallinn University of Technology, Estonia
Y.H. Chang, Chihlee Institute of Technology, Taiwan

Program Chair
Tim Chou, Advanced Science and Industry Research Center, Hong Kong
W. K. Jain, Indian Institute of Technology, India

International Scientific Committee

M. Subramanyam, Anna University, India
Urmila Shrawankar, G.H. Raisoni College of Engineering, India
M.S. Chen, Da-Yeh University, Taiwan
I. Saha, Jadavpur University, India
X. Lee, Hong Kong Polytechnic University, Hong Kong
Antonio J. Tallón-Ballesteros, University of Seville, Spain
J. Xu, Northeast Dianli University, China
Q.B. Zeng, Shenzhen University, China
C.X. Pan, Harbin Engineering University, China
L.P. Chen, Huazhong University of Science and Technology, China
K.S. Rajesh, Defence University College, India
M.M. Kim, Chonbuk National University, Korea
X. Ma, University of Science and Technology of China, China
L. Tian, Huaqiao University, China
M.V. Raghavendra, Adama Science & Technology University, Ethiopia
J. Ye, Hunan University of Technology, China
Q. Yang, University of Science and Technology Beijing, China
Z.Y. Jiang, University of Wollongong, Australia

XI
This page intentionally left blank
Computational intelligence
This page intentionally left blank
 Informatics, Networking and Intelligent Computing – Zhang (Ed.)
© 2015 Taylor & Francis Group, London, ISBN: 978-1-138-02678-0

Decomposition genetic algorithm for cellular network spatial optimization

M. Livschitz
TEOCO, Fairfax, VA, USA

ABSTRACT: Spatial optimal planning of cellular networks is a fundamental problem in network design. This
paper suggests a new algorithmic approach for automatic cell planning and describes the multi layer’s decom-
position genetic algorithm, which significantly improves optimization convergence. Algorithm convergence is
compared with single layer genetic algorithms based on the cell planning of real cellular networks.

1 INTRODUCTION and future traffic load requirements. Based on traffic

demands for the area, propagation models and various
The cellular technologies of 3G and 4G cellular net- constraints and costs for base station locations, the
works use the same frequency for all users and are cell plan should define a set of locations for base sta-
projected to provide a wide variety of new services, tions and base-station configurations, including sector
based on high-data-rate wireless channels. Existing 3G numbers, directions, number of antennas per cell and
networks support 16.0 Mbps applications whilst 4G antenna directions. During the planning stage, the most
networks will support bit rates higher than 80 Mbps. vital issue is usually site location definition, when
The important aspect of these technologies is that these expanding existing networks, the most critical issue is
systems are most likely to be implemented supporting antenna configuration and direction. Design of most
high frequencies above 2 GHz. Such high frequencies 3G and 4G networks includes the installation of anten-
yield very quick signal degradation and strong diffrac- nas with remote changeable electrical tilt, allowing for
tion from small obstacles, forcing the reduction of quick implementation of cell planning results and sup-
cell size. On the other hand, high speed data require- porting quicker response times and shorter planning
ments support results in high Signal-to-Noise (SNR) cycles.
requirements. In order to decrease the amount of inter- Most 4G network deployments will be done by
ference and increase the quality of coverage, spatial existing cellular operators who have 2G and 3G net-
cell optimization and planning leads to higher antenna work infrastructure above existing infrastructure. This
tilt requirements and additional decrease of cell size, condition will result in site locations being more or less
resulting in increased site densities especially in urban predefined based on the existing network site loca-
areas. tions or among permitted shared locations and even
New applications and services lead to large dif- during the planning stage the key cell planning issue
ferences in traffic distributions and sector loads [7], is choosing antenna types and direction.
making the cell planning process more complicated. In this paper we suggest a new algorithm approach
Cell planning in modern cellular networks must have base in genetic algorithms (GA) for resolving the
the capability of responding to traffic distribution cell-planning problem and we illustrate algorithm
changes because this is the main way to improve net- efficiency by using some results of existing cellular
work efficiency. This is true for the planning stage, network optimization projects.
when the network is planning for a particular load as
well as after network deployment and traffic growth.
2 DECOMPOSITION GENETIC ALGORITHM
Spatial network planning is one of the most impor-
tant issues when building a new network or when
2.1 Decomposition of cell planning
expanding an existing network during traffic growth or
changing traffic patterns. Such planning will be much Let us consider cell-planning or budget cell-planning
more important and complicated for 4G networks. Pro- problems [3] of a cellular network with a given set
ducing a viable cell plan is a key ingredient in the of base stations Ib = {1, . . . , I}. The base-station con-
ability of an operator to provide QoS and coverage in figuration is defined by a vector of parameters Pp
a cost efficient manner. representing the typical antenna type (pattern), the
Cell planning typically includes planning a net- azimuth, tilt, and height together with cell power
work of base stations that provides a high level of definitions. If there is only one antenna per cell,
coverage in the service area with respect to current configuration of antenna i Pi = {p1 , . . . , p5 } is defined

3
by a vector of 5 parameters. Each parameter pp is con- between signals from all services and the interference
fined within a boundary and can be assigned one of from all antennas impacting the mobiles. Uplink and
the discrete values: downlink power control and handovers between cells
should also be taken into consideration by resolving
the sequence of system linear equations. As a result,
the evaluation could be quite expensive, especially
The boundaries of permissible values could be dif- for large network clusters with high antenna density.
ferent Thus tilt changes could be assigned any values The complexity of the Monte-Carlo evaluator is at
in a range with a predefined step (usually 1 deg), least O (I2 log(I)). This results in a very high impor-
azimuths have values according to a higher step and tance for improving convergence of GA for cellular
with a limitation on minimal change, while antenna cell-planning problems in wideband networks.
types and heights are assigned values from a list of This nature of 3G, 4G and beyond networks requires
available patterns or heights. reaching a higher isolation between remote cells from
The goal of cell planning F(K) is defined by com- the second tier and further. On the other hand, close
bining different key performance indicators (KPIs) cells have overlapping coverage areas supporting con-
representing network coverage quality or capacity tinuous coverage and have a very high impact on
K = {K1 , . . . ,Kk }, where each Kk represents a bad each other. Impacts from further cells depend on
level of KPI defined by a fuzzy threshold Tk [8] as site location, area topology, building height and other
follows: environmental properties. Some tall sites located on
mountain tops or building roofs influence distances
of 5–15 km, while other sites in urban areas could
be almost invisible even at distances less than 1 km.
Influences between cells depend on spatial network
configuration and can be changed or eliminated during
cellular network optimization activity.
Manual planning of wide areas is usually performed
by an interactive, iterative process that includes the
All KPIs are dependent on antenna configurations. following two phases:
Cell planning should improve defined KPIs • Planning of smaller clusters (local cluster
optimization);
• Synchronization between clusters and composing a
final solution
changing antenna configurations according to the pre-
defined constraints. In the case of budget cell-planning Antennas in small compact clusters should be con-
problems, cost will be assigned to each antenna change figured together, taking into consideration the strong
and an additional constraint on the total cost of network influences between them.
changes should be added. The synchronization phase should resolve in two
This means that the optimal cluster cell configu- types of interactions between clusters: impacts through
ration Po = {P1 ,…,PI } is defined by one value from cells on borders, which are impacted by two or more

I
5 clusters and long links from further sectors. There are
PipN available combinations. In most cases the usually not too many further impacts between antennas
i=1 p=1
from remote clusters. The process described above is
spatial cell planning problem is considered as NP-hard,
repeated till the network KPIs reach predefined values.
which means that finding an optimal solution for it
In case network planning should fail to reach
(within real networks) is not feasible in a reasonable
required KPIs for the cluster, new additional sec-
running time. Thus, much of the work is done with
tors, sites and locations will be recommended and the
genetic algorithms (GA) for the spatial cell planning
iterative process will continue.
problem [3, 6, 8, 9] and different heuristic solutions
We believe that problem-oriented heuristics should
for improving the GA convergence.
be used for the efficient optimization and improve-
Common GA algorithm schemes suppose there is a
ment of the optimization algorithm. The manual opti-
quick manner of goal function calculation and popula-
mization process scheme described above leads to
tion evaluation. For the cellular network optimization
the development of two-level GA for improving the
this means that there is an evaluator, allowing for an
convergence of spatial cellular network optimizations.
estimated network quality for multiple antenna spatial
locations and configurations.
2.2 Algorithm description
The network evaluator for wideband cellular net-
works should estimate network quality according to The main idea of the decomposition genetic algorithm
the planned traffic load. The best method for wide- is a reduction in the overall optimization problem
band network evaluation is based on the Monte-Carlo to multiple smaller optimization sub-problems with
techniques [8], which require simulation of uplinks consequent results composition.
and downlinks between all the simulated mobiles and Reduced optimization problems are resolved inde-
all serving cells. The main KPIs depend on the ratio pendently, bringing local improvement for sub-areas.

4
 Figure 2. Cluster division into 4 sub-areas.

too low a number will not allow the rapid optimization

of sub-areas. For real cellular networks this number
Figure 1. Decomposition Genetic Algorithm. could be chosen considering the cluster division on the
RNC and is dependent on cluster size and the number
This allows easy calculation in the distribution of run- of antennas (I) used for cell planning.
ning each sub-area optimization in parallel on separate
computers. Local optimizations can be stopped after 2.2.2 Sub-area division
achieving an initial improvement even before the GA Cellular network cluster is automatically divided into
is converged, because the second optimization phase N sub-areas so that the site set will be divided into
will enable the solution to be finalized. Research on 
N
the N subset: In = Sn according to geographical
the definition of stop criteria for local optimizations n=1
and impact on convergence results will be performed criteria. Geographical areas should be compact areas
in the future. Following is a block diagram (Figure 1) including a few dozen cells (Figure 2).
describing decomposition genetic algorithm based on Subsets can overlap but should not be subsets of
an existing GA tool. other subsets: ∀n, k, Sn  ⊂ Sk . A site set can be divided
We consider decomposition genetic algorithm randomly, using technological criteria or based on dis-
(DGA) as an extension for any GA implementations tance criteria. Automatic division could be done using
not concerning the GA itself. There is only one excep- various algorithms.
tion. GA used by DGA should allow starting based Fuzzy logic cluster algorithm is used as a basic
on some particular solution. This limitation is usually division algorithm for DGA. Fuzzy logic clustering
realized for GA used for cell planning and especially associates with each element a set of membership
for budgeted cell planning, which optimize networks levels which indicate the strength of the association
starting from existing configurations. between that data element and a particular cluster.
It is used for assigning elements to one or more
2.2.1 Subset number definition spatial clusters. Fuzzy clustering creates N + 1 sub-
As a first step the number of subsets (N ), which will areas, including N non-overlapped sets and one border
define the number of sub-areas used for local opti- area. Border area includes elements which have sim-
mization, is chosen. The number of subsets is defined ilar membership for two or more clusters. It overlaps
by the sizes of small sub-areas or by taking into consid- with N or less clusters and used for the second phase
eration connectivity between different network parts. optimization by DGA.
As will be shown later, this parameter could signif-
icantly impact on convergence improvement, which 2.2.3 Sub-area goal function definitions
could be reached using the DGA. Too large a number Optimization criteria per sub-area are defined at this
will result in the division of clusters over too small an step. The optimization criteria should generally be the
area, which could not be optimized separately due to same for a whole cluster, but different sub-areas might
the high impact from antennas out of the area, while have different initial KPIs. The optimization criteria

5
are adjusted accordingly. This is an additional advan-
tage of DGA that wide network optimization might be
done based on different criteria per area.

2.2.4 GA sub-area optimizations

The following three steps should be performed for all
N sub-areas:
• Start from the initial network configuration, run
the GA for sub-area optimization according to the
defined criteria.
• Keep the optimized solution Pn = {Pj }, j ∈ Sn for
sub-area n. GA optimization will result in sub-area
antenna configuration.
• Sub-area optimizations are independent of each Figure 3. Optimization results for medium size cluster.
other and can be run in parallel.
Parallel calculation on multiple processes signifi-
cantly improves CPU utilization reduce overall time toolkits, supporting GA. The optimization tool uses
for the optimization. technological operators, which helps improve GA
convergence based on experience in radio access
2.2.5 Solution composing network optimization.
Solution composing creates the network configuration All comparisons were run within the same GA con-
for the whole cluster (Poin ), combining optimization figurations and show goal function improvement as
results for sub-areas. An overlapped areas between, a function of the evaluation number. For DGA goal
different networks is created based on border area function improvement the second phase is calculated
of fuzzy clustering. It is used for the global opti- relatively to the initial network quality. Plots depict
mization, synchronizing between different solutions. a graph of improvements on the baseline. TEOCO’s
These initial network configurations will show sig- optimization GA for the whole network and DGA
nificant improvements in most KPIs defined as the goal functions improvement is for the second phase
optimization goals. of decomposition GA. As a result of running the first
phase optimizations for all sub-areas decomposition
2.2.6 Second phase initialization GAs have some additional overhead. These optimiza-
The initial population for the second phase is created tion cycles for sub-areas are short and evaluations for
based on combined solutions of sub-areas (Poin ). The small sub areas are much quicker than for wide areas,
initial population should include one or more com- so this additional overhead is minor compared with an
posed solutions, but other population members are optimization run for a whole area and cannot change
built using mutation operators. the overall picture.

2.2.7 Second phase GA

The GA optimization is run for the whole boarding 3.1 Middle size cluster
area starting from the better initial population. This Optimization results of a cluster with 165 cells are
optimization should further improve the network, syn- shown in Figure 3. This network covers a small
chronizing configurations of cells located on borders town in a mountainous environment where impacts
between areas. It should also eliminate remote links between remote antennas are significant. Decompo-
between different areas. sition genetic algorithms show a little better goal
function improvement compared with the baseline
GA, but the conversion rate is about 4–5 times higher.
3 ALGORITHM VALIDATION Optimization results of DGA are impacted by a num-
ber of sub-areas N and their sizes used for the first
In this section some results of decomposition GA use phase.
will be presented. All the results of optimization runs Figure 4 depicts a dependence of maximal improve-
were made based on the optimization of real 3G net- ment in the convergence rate, which can be reached by
works – two markets with different properties and DGA on number N of the cluster division in sub-areas.
presenting different cases of cellular optimization. One Cluster division on only a few sub-areas (N is small)
market represents the optimization of a suburban area does not bring to the maximum convergence improve-
with middle site density. The second area is a typical ment, while cluster division over too many areas (N
high density urban market, representing cellular net- is big) will result also in a smaller improvement
work properties in large towns with very high traffic because the sub-areas would be smaller and borders
and high density of cellular sites. between sub-areas will have relatively higher weight.
All optimizations run were done by using TEOCO’s These results show that optimal cluster division could
optimization tool and TEOCO’s implementation of significantly improve algorithm efficiency.

6
 and estimate possible improvements. This approach
could be used for very general spatial problems, where
local influences are stronger than global. Different
algorithms of cluster division will also be compared.

5 CONCLUSIONS

Decomposition genetic algorithms show very signif-

icant convergence improvement compared with the
usual flat scheme of GA. DGA is able to achieve signif-
icantly better network improvement in a shorter time
for the huge spatial cell optimization and planning
problems. DGA is appropriated for parallel calcula-
tions on multi CPU computers.
Figure 4. Convergence rate for different divisions.

REFERENCES
[1] E. Amaldi, A. Capone, and F. Malucelli. Optimizing base
station siting in UMTS networks, In Proceedings of the
IEEE Vehicular Technology Conference, 4, 2001, 2828–
2832.
[2] D. Amzallag, J. Naor, and D. Raz. Cell planning of
4G cellular networks, In Proceedings of the 6th IEEE
International Conference on 3G & Beyond (3G’2005),
London, 2005, 501–506.
[3] D. Amzallag, M. Livschitz, J. Naor, D. Raz. Cell Planning
of 4G Cellular Networks: Algorithmic Techniques and
Results, In Proceedings of the 6th IEEE International
Conference on 3G & Beyond (3G’2005), London, 501–
506.
Figure 5. Optimization results for a large cluster. [4] F. Longoni and A. Länsisalmi and A. Toskala. Radio
Access Network Architechture, In H. Holma and A.
3.2 Large cluster optimization Toskala (editors), WCDMA for UMTS, John Wiley &
Sons, Third edition, 2004, 75–98.
Figure 5 depicts network improvement for a big cluster [5] C. Lee and H. G. Kang. Cell planning with capacity
with about 700 cells in a dense urban area. In this case expansion in mobile communications: A tabu search
DGA is able to reach more than three times higher an approach, IEEE Transactions on Vehicular Technology,
improvement in a shorter time. Base line GA was not 49, 2000, 1678–1691.
[6] K. Lieska, E. Laitinen, and J. Lähteenmäki. Radio cover-
able to reach a reasonable improvement in a feasible age optimization with genetic algorithms, In Proceedings
time for so large clusters. Modern cellular networks of the 9th IEEE International Symposium on Personal
of 3G and 4G have higher density and may have thou- Indoor and Mobile Radio Communications (PIMRC’98),
sands cells for which DGA approachis critical. Cluster 1998, 318–321.
division in a few sub-areas and the first optimization [7] M. Livschitz, D. Amzallag. High-Resolution Traffic Map
phase allowed the starting of a second phase of DGA of a CDMA Cellular Network, In Proceedings of the 6th
from a better initial point. INFORMS Telecommunication Conference, US, March
2002, 62–64.
[8] Maciej J. Nawrocki Undertanding UMTS Radio Net-
work. Modelling, planning and automated optimization,
4 FUTURE RESEARCH John Wiley & Sons, 2006.
[9] H. Lin, R. T. Juang, D. B. Lin, C.Y. Ke, andY. Wang. Cell
In the future we are going to investigate theoretical planning scheme for WCDMA systems using genetic
aspects of the DGA, research common properties of algorithms and measured background noise floor, IEEE
problems where improvement using DGA can be used Proceedings on Communications, 151, 2004, 595–600.

7
This page intentionally left blank
 Informatics, Networking and Intelligent Computing – Zhang (Ed.)
© 2015 Taylor & Francis Group, London, ISBN: 978-1-138-02678-0

A heating and cooling model for office buildings in Seattle

W.Q. Geng & Y. Fu

Department of Applied Computational Mathematical Science, University of Washington, Seattle, Washington, USA

G.H. Wei
School of Management, Dalian Jiaotong University, Dalian, Liaoning, China

ABSTRACT: To specify a model of heating and cooling systems inside office buildings, we focus on details
of difierent non-ignorable factors of standard office building in Seattle. As Seattle has distinctive climatic
variations, this is a good case for discussion. Based on an analysis of the temperature profile of Seattle, as
well as the application of heating and AC systems, we are able to build a model, which can be applied to real,
modern buildings in Seattle. Also, we consider the heat radiation from human bodies and electronics (lighting,
computers, and other sources) in an office building. By incorporating all the factors in the research, our goal is
to model the temperature curve that best fits the actual situation inside a building. Our model can be used to
evaluate and improve AC systems, thus making office buildings more comfortable for office workers, as well as
reducing the consumption of energy. For less energy consumption, we figure out that the minimum heat output
of heating and an AC system should be 10◦ F/h in order to make the temperature inside the building constant
within the comfort zone.

1 BACKGROUND 4. How do we minimize the use of energy while keep-

ing a comfortable temperature, for example, with
Office buildings are the places where people spend less running time?
most of their time besides their homes, because people
work in the office building for 8–10 hours per day. A
comfortable working environment can boost people’s 3 SIMPLIFICATIONS AND ASSUMPTIONS
efficiency and happiness. People tend to reduce energy
consumption in a more flexible way. In our model, a lot of assumptions are made to sim-
Therefore, our model aims to combine all factors plify the model, while keeping these assumptions still
about thermal comfort in the building, and then gets reasonable for building a credible result. Since every
rid of utilizing a traditional conditioner system because office building has differences in various aspects, such
it wastes a lot of energy. The goal is to design an auto- as thickness of walls, areas per person, and in AC sys-
matic thermal system inside the building, and to figure tems, we take average values or representative values
out the hourly out-door and in-door temperature curve, to build our model, while assuming that these data fit
the human body and electrics radiation in the office the office building in our model. Also, in the outside
building, and AC system. temperature model, we took temperatures over a few
days and assumed the temperature pattern during these
days would fit the whole year. Besides the simplifica-
2 PROBLEM DESCRIPTION tions in the variables we study, there are a lot of factors
that we did not incorporate in our model, for example,
This paper presents a heat dynamic model for calcula- the heat escaping from open windows and doors. We
tion of the indoor and outdoor temperature, heat flow, assumed that these unconsidered factors were negli-
and future energy needs for electrical heating in an gible. More detailed simplifications and assumptions
office building. are explained in the model section.
We consider the following questions:
1. How does the outside temperature influence the
4 A MATHEMATICAL MODEL
temperature inside the building?
2. How does the temperature inside the building
4.1 An overall model of the temperature inside a
change with inner heat radiation, heating, and an
building
AC system?
3. How do we modify the heating and theAC system to The temperature inside a building is based on the
make the temperature in the building fit the comfort heat conducted through the building’s walls, combined
zone during working hours? with the heat produced inside the building, as well as

9
the heating and the AC System. The conduction of
heat through walls follows Newton’s Law of Cooling/
Heating, which indicates that the rate of change of the
temperature inside the building is directly proportional
to the difference in temperature inside and outside. The
proportional constant k represents how fast the heat
conduction proceeds. The other factors will directly
affect the temperature inside the building. Let T (t) be
the temperature inside the building, and the follow-
ing equation can be used to describe the change of the
temperature:

Figure 1. Hourly Seattle temperature in one year.

in Seattle [1], we plotted the data to show the general

4.2 Temperature of Seattle trend of the temperature. To fit the curve into a func-
tion of t, we use curve fitting tools in MATLAB using
To build a model that shows the hourly temperature in five years of repeating data. Based on the shape of the
Seattle, firstly, our goal is to find the hourly tempera- plot, we choose the Fourier equation with two terms to
ture model that fits each day. Then we incorporate the fit the plot. The functions are in the format below:
daily model with the data that we collect through the
year, which is the minimum, maximum, and average
temperature of each month.
To generalize the model through one day, we collect
temperature data [9] during four days from 14th to
17th, Feb 2014 as plotted in blue line in the figure
below. The x-axis represents 96 hours in total in 4 days. To convert time from months to hours, t is multiplied
The y-axis represents the temperature. Based on the by 12/(365*24). According to the fitting data gener-
plot, the temperatures over the four days act as an ever- ated by MATLAB, ω = 0.5236 for all three curves.
repeating wave of peaks and valleys. Therefore, we For Mave (t), a0 = 51.95, a1 = −12.4, b1 = −1.227,
assume that the temperature is close to a trigonometry a2 = 1.163, b2 = 2.114. For Mmax (t), a0 = 59.9, a1 =
function of time and looks like: −14.73, b1 = −0.7251, a2 = 1.264, b2 = 2.858.
For Mmin (t), a0 = 44.01, a1 = −10.07, b1 = −1.738,
a2 = 1.058, b2 = 1.371. The end of the curve shows
some increase in order to fit the Fourier function.
M(t) denotes the temperature of Seattle during the Here it shows some error compared with the real
day. A is the amplitude of the sine curve. For each day, data plot, but we assume the fitting curve can rep-
the upper limit of amplitude is |Mmax − Mave |, while resent the hourly averages of maximum and minimum
the lower limit is the amplitude being |Mmin − Mave |. temperatures in Seattle.
φ is the phase shift of the curve. Based on this for- Since we have the temperature curve within one
mula, we have amplitude A in one cycle as a segmented day, as well as the Mmax (t), Mmin (t) and Mave (t) curve
function as shown below: throughout a whole year, we can put the two functions
together to generate the hourly temperature curve in
Seattle within a year as shown in figure 1.
The combined function of hourly temperature in
Seattle will be used as the outside temperature in the
Also, when t = 6 + φ, the temperature reaches heating and cooling of building model.
its maximum and when t = 18 + φ, the temperature
reaches its minimum. To find the phase shift that best
4.3 Heat from human body radiation and electrical
fits the curve, MATLAB is used to determine the
appliances
phase shift value with the highest correlated coef-
ficient, which gives the result of φ = 10.04. That In this section, our goal with this model is to find
indicates that the highest temperature happens around H(t), which denotes the change per hour in temperature
4 pm and the lowest temperature happens at 4 am. due to heat radiated by people and produced by lights,
The average minimum and maximum temperatures are computers, and other electrical appliances.
determined separately for each day which causes some Our model is built based on an office-like room that
discontinuous on the curve. has limited volume and is only used by one person. The
The second step of the model is to find Mmax , H(t) we generated from this model can be applied to
Mmin and Mave according to time t hourly throughout a the whole building.
whole year in Seattle. Based on monthly average, max- Firstly, the volume of the room is calculated through
imum, and minimum temperature during 2003–2013 the average area per person in the office buildings

10
by timing the average height of each floor. Then we as others in the offices, the heat output of iMac is
assume the room is filled with air, which is heated by 126 to 463 BTU/h, based on data from its official
human radiation and electrical appliances. Therefore, website [4]. In this case we will take the average
we have: value as 250 BTU/h ≈ 73.27 W. Finally, we need to
consider the heat produced by other electrical equip-
ment. This equipment is normally not in constant use,
such as printers, copiers, vacuums, etc. We assume
the equipment has a power of about 5 W. Adding
where Q(t) is the heat produced by human and elec- all the heat generated by these appliances, we have
trical appliances per hour. V denotes the volume of Qe = 10 W + 73.27 W + 5 W = 88.27 W ≈ 317770 J/h.
the room, which is also the volume of the air. ρ is the Combining the data above, it is easy to calculate
density of air, and c is the specific heat capacity of air. the heat produced by human radiation and electrical
Based on previous research, the working space appliances per hour:
in an office per person is about 185 square feet
[6]. The normal height of an office is 12 to
15 feet, and in this model we will consider 13
feet as the height of the room. So V = 185 ft2 ×
(1 m)3
13 ft × (3.2808.1 ft)3
≈ 68.102 m3 . Besides office areas, Because the model needs to fit the temperature
office buildings also have public areas such as lob- change for the whole day, so the working time of office
bies, elevators, and meeting rooms, which are not buildings needs to be considered. We assume that the
constantly occupied by people. We assume that the working time of the building is from 8 am. to 6 pm.
ratio of office areas to public areas is approximately The human body and the electrical appliances produce
1:1, and therefore the volume needed for each person to heat only during the working hours. Therefore, H(t)
be kept warm needs to be doubled, so V = 136.204 m3 . during the entire day is a segmented function:
ρ and c are all constants with value ρ = 1275g/m3 and
c = 1.007 J/(g K). Plug in all the constants:

4.4 Heat conduction through building roof and wall

The next step is to determine Q(t), which can be The temperature inside the building is affected by the
separated into two parts: the heat radiation from the outside temperature mostly through conduction in the
human body is Qh and the heat produced by electrical roof and walls. In our model, we assume the conduc-
appliances is Qe . In the room model, only one person tion of heat through the wall and roof of the building
is staying in the room. The heat radiation of a human is the only way that the heat is lost between the inside
body can be considered as black-body radiation [3], of a building and the outside in Seattle. As indicated
which power can be determined by Stefan-Boltzman in the general model, the change in temperature due to
Law: conduction per unit time C(t) follows Newton’s Law
of Cooling/Heating, which can be represented as:

A is the area of human skin, in this case it is assumed

to be 2 m2 . σ is a Stefan-Boltzman constant, includ- where M(t) is the temperature outside and T(t) is
ing the value of the emissivity of human clothes or the temperature inside the building. k is the con-
skin, which is 0.98 according to data from Infrared stant representing how quickly the heat is conducted
Services [7]. T is the surface temperature. Because through the walls which is arbitrary based on the mate-
of the existence of fabric clothes, the surface temper- rial and thickness of the walls. Let Qt (t) be the heat
ature at room temperature (T0 ≈ 20◦ C = 293.15K) is conducted through wall per hour, we have [8]:
approximately 28◦ C = 301.15K [5]. Plug in the con-
stants, the power of one human body radiation is
Qh = 93.324W = 335970 J/h.
The heat produced by electrical appliances is mainly
from lighting systems and computers. Since different Combine the two equations above and we can
types of lighting systems can produce significantly dif- determine k:
ferent amounts of heat, for example, a filament lamp
as opposed to a LED lamp, we make an assumption
that the total heat produced by light is 10 W in a one-
person room. Also, a large amount of heat comes from
a computer. Assuming that an office uses iMac com- U is the heat transfer coefficient that represents the
puter, which should produce similar amounts of heat heat conducted through certain materials per unit time

11
per unit area when the temperature is changed by1
degree. A is the area of the material, which is the wall
or the roof. ρ represents the density of air. V is the
volume of air, which is assumed to be the volume of
the building. c is the specific heat capacity of the air.
In the model, we assume an office building as
a cuboid with 40 floors, therefore the height is 13
ft./floor × 40 floor = 520 ft. = 158.5 m. And we also
assume the base of the building is square with an area
of 80 ft. × 80 ft. (24.38 m × 24.38 m). Then we can
calculate the area of walls and roof, as well as the
volume of the building:
A wall = 4 × 158.5 m × 24.38 m = 15456.92 mˆ2
A roof = 24.38 m × 24.38 m = 594.3844 mˆ2
24.38 m × 24.38 m × 158.5 m = 94209.9274 mˆ3
Since different materials have quite different val- Figure 2. Temperature inside the building without
ues of U, which will have a significant impacton the insulation.
conduction of heat, we will calculate two k values: ki
when the materials of the walls and roof contains an
insulation layer, which will decrease the rate of heat
conduction. And ku for the walls and roof without insu-
lation. Also, the roof and walls have different values
of U, therefore,

First, we calculate ku for the building without insu-

lation by checking the chart of common heat transfer
coefficients of some common building elements [8]:

Figure 3. Temperature inside the building with insulation.

Then, for the insulating materials [8]:

4.5 AC and heating system
Based on previous modelling, H(t), the change of
temperature due to heat generated inside the build-
ing by human radiation and electrical appliances has
a relatively high value that can eventually cause the
temperature inside the building to become too high
without any adjustment of the AC System. Also, the
temperature in winter is still too low fora comfortable
Since ki is only about one-third of ku , it is obvi- office environment.
ous that the change inside the building due to the The temperature difference within one day is nearly
heat conduction through walls and roofs will be much 20 degrees. Also, the temperature can be as low as
smaller when the insulation is presented. That is to ≈40◦ F in winter, and as high as about ≈105◦ F in sum-
say, the insulation materials can keep the temperature mer, which is extremely uncomfortable in an office.
inside the building steadier and less affected by the Therefore, AC and heating systems are used to adjust
temperature outside. The following figures show the the temperature inside the building into the comfort
temperature inside the building compared with that zone for people, which is between 22◦ C to 26◦ C
outside in figure 2 and without insulation in figure 3. (71.5◦ F to 78.8◦ F) as based on research [2].
It is obvious that the range of the inside tempera- In our model, we assume that the heating and the
ture is less with insulation. Therefore, the insulating AC systems have a constant power that is able to pro-
materials make it easier to maintain the temperature duce or transfer the same amount of heat outside the
inside the building within a comfortable range. In the building. Also, in order to constrain the temperature
following model, we assume the building has good within a comfort zone, therefore, when T(t) is higher
insulation ability. So we will use k = ki = 0.1892 h−1 than a comfort zone, the AC is on to decrease the tem-
as our proportional constant. perature. To the contrary, the heat is on when the T(t)

12
 In the process of AC system operation, we use
the segmented function to solve the general problem.
We estimate the general temperature difference within
one day according to the curve we plot. When T(t)
is higher than the comfort zone which we set up, AC
system tends to decrease the temperature, otherwise,
it increases the temperature. The segmented function
is set according to different temperature variation.

6 CONCLUSIONS

In a word, our model mainly achieves our aim to eval-

uate and improve the AC system. Our model combines
all factors about thermal comfort in the building, and
Figure 4. Temperature inside and outside office building avoids the disadvantages of traditional conditioner sys-
when m = 10. tems. We design an automatic thermal system inside
the building, and figure out the hourly out-door and
is lower than a comfort zone. In this way, U becomes in-door temperature curve, the radiation from the
a segmented function that: human body and from electrical devices in the office
building, and the AC system’s operation. In the
process, we use various mathematical methods:
curve – fitting tools in MATLAB, Fourier equa-
tion, and simple mathematical methods for differential
equations, applying Newton’s law of heating and cool-
where m is a positive constant that represents change ing, and setting segmented functions. We figure out the
rate of temperature per unit time due to the AC and final values that the minimum heating and AC system
heating system. By plotting different m values for the power need to keep T(t) steady within a comfort zone
system, higher m value is able to better control the of around 10◦ F/h. Under these circumstances, we can
temperature. Figure 4 shows T(t) when m = 10. consume less energy and achieve our primary goal.
For less energy consumption, we need to calculate
what is the minimum m that can make T(t) bound REFERENCES
within a comfort zone. From the figures above, we
can see that when m = 10◦ F/h, the curve for T(t) is [1] Climate Wizard. Seattle average min and max tem-
steady within the comfort zone. Therefore consider- perature from 2003 to 2013. http://climatewizard.org/,
ing the conservation of energy as well as the comfort 2014.
of the office, the heating and AC system for the office [2] W. Cui, G. Cao, J.H. Park, Q. Ouyang, and Y. Zhu.
Influence of indoor air temperature on human ther-
building need to be able to make temperature changes
mal comfort, motivation and performance. Building and
of about 10◦ F/h. Environment Building and Environment, 68(3):114–122,
2013. ID: 5136851685.
[3] James D. Hardy. The radiation of heat from the human
5 SOLUTION AND TECHNIQUES body: Iii. The human skin as a black-body radiator. The
Journal of Clinical Investigation, 13(4):615–620, 7 1934.
When we build the model of the yearly temperature [4] Apple Inc. iMac: Power consumption and thermal output.
change in Seattle, we use MATLAB and Fourier equa- http:// support.apple.com/kb/HT3559, 2013.
tions in mathematics. To fit the curve into the function [5] Bin Lee. Theoretical Prediction and Measurement of
the Fabric Surface Apparent Temperature in a Simulated
of t, we use curve – fitting tools in MATLAB by col-
Man/Fabric/Environment System. Melbourne: DSTO,
lecting five years of temperature data. Based on the 1999.
shape of the plot, we adopt Fourier equation with two [6] Norm Miller. Estimating office space per worker. Tech-
terms to fit the plot. Although some errors indeed exist nical report, Burnham-Moores Center for Real Estate,
in the figures, which we have plotted, compared to 2012.
the actual data, we assume that they can be ignored, [7] Infrared Services. Emissivity values for common mate-
and using the Fourier equation is a relatively accurate rials. http://infrared-thermography.com/material-1.htm,
method of curve fitting. 2007.
When considering the exchange of heat between [8] The Engineering ToolBox. Heat loss through build-
ing elements due to transmission. http://www.
the inside and the outside of the building, we apply
engineeringtoolbox.com/heat-loss-transmission-d_748.
Newton’s Law of Cooling/Heating, which indicates html, 2014.
that the rate of change of the temperature inside [9] Weather Spark. Feb 14 to Feb 17, 2014, hourly tem-
the building is directly proportional to the difference perature recorded at Boeing field/king county inter-
between the temperatures inside and outside. There- national airport, Seattle, WA. http://weatherspark.com/
fore we are able to build a differential equation and #!dashboard;ws=29735, 2014.
solve the problem as an exponential model.

13
This page intentionally left blank
 Informatics, Networking and Intelligent Computing – Zhang (Ed.)
© 2015 Taylor & Francis Group, London, ISBN: 978-1-138-02678-0

Multi-depth Deep Feature learning for face recognition

C.C. Zhang, X.F. Liang & T. Matsuyama

Department of Intelligence Science and Technology, Graduate School of Informatics, Kyoto University, Kyoto, Japan

ABSTRACT: Deep structure learning is a promising method for computer vision problems, such as face
recognition. This paper proposes a Multi-depth Deep Feature (MDDF) learning method to learn abstract features
which has the most information in the region of interest. It is unlike other deep structure learning methods that
evenly partition the raw data/images into patches of the same size and then recursively learn the features by
evenly aggregating these pieces of local information. MDDF does an uneven partition according to the density
of the discriminative power in the image and learns a data-driven deep structure that preserves more information
at the region of interest. In the cross-database experiments, MDDF shows a better performance over a canonical
deep structure learning (Deep-PCA) on face recognition. Moreover, MDDF achieves an accuracy, comparable
with other well-known face recognition methods.

Keywords: Deep structure learning, Multi-depth Deep Feature, fine-to-coarse Quad-tree partition

1 INTRODUCTION information. Therefore, it reports a better performance

on facial variations.
Appearance-based techniques have been extensively Most deep learning methods are inspired by Con-
studied, and acknowledged as one class of the most volutional Networks [5] and Deep Belief Networks
successful face recognition approaches. Principal [6]. Due to the complicity of connectionist networks
Component Analysis (PCA) [1] and Linear Discrim- in deep learning, many built in biases affect their
inant Analysis (LDA) [2] are two widely accepted performance. An interesting study, Deep PCA [7],
representatives in this framework. However, pose and separated out other properties in existing deep archi-
illumination changes, occlusion, and data missing tecture to build a simplified and non-connectionist
greatly challenge these techniques in the real-world framework. It demonstrated that the exploitation of
scenarios, because they have a weak ability to cope deep structure did increase performance. However,
with local facial variations. Multi-subregion fusion the conclusion in [7] needs further discussion. Deep
methods [3, 4] are proposed as a solution for this PCA partitioned the image into small patches of the
problem. They divide the face into a set of dis- same size, evenly extracted lower dimensional fea-
joint/overlapped subregions, perform recognition on tures, evenly aggregated these features from each
each subregion, and fuse the results. Experiments lower level, and finally formed an overall abstract fea-
shows that they provide considerable accuracy in well ture.The improved performance came from a complete
registered databases. The major reason for success, it hierarchy that assumed the information was evenly dis-
is argued, is that the sub-regions preserve well the local tributed. We argue that features are mostly distributed
information which is more robust for variations of the unevenly in visual data. That is the reason the Region
face. However, the size and shape of the sub-regions of Interest (ROI) must be cropped first in many appli-
have rather sensitive parameters that are often assigned cations. This also applies to face recognition. Both
by empirical experiences or multi-scale values. psychology [8] and biometrics [9] researches show the
Recently, deep learning is a promising method in periocular region offers advantages over other regions.
computer vision, and has been discussed on face recog- This finding indicates that evenly partitioned patches
nition. It attempts to replicate the mechanism of the do not reveal the fact of feature distribution on faces.
human visual cortex. General deep learning is an In this paper, we address the above problem and
unsupervised scheme, and can be regarded as a hier- propose a Multi-depth Deep Feature (MDDF) to learn
archy structure consisting of a low level of original abstract features. Unlike evenly partitioned images,
local information, and multiple higher levels of more a hierarchical Quad-tree is introduced to partition
abstract information that encodes the correlations of facial regions into uneven subregions following their
local ones in the level below. Since both local regions varied discriminative powers. A small subregion has
and the overall structure of these unites are involved, more dense discriminative information than a bigger
deep learning is able to extract both local and global one. The feature learned from different subregions,

15
therefore, preserves the corresponding local infor-
mation. The criterion of Quad-tree partition uses
LDA-motivated total variance, which ensures a robust
resistance to local noise and an efficiency of computa-
tion. In our framework, Quad-tree partition functions
as a convolutional neural network, but builds up an
incomplete hierarchical structure. Afterwards, aggre-
gation of these uneven features in the multi-depth
structure functions as a recursive neural network, and
outputs an abstract feature having more information at
the region of interest. Experiments on four challeng-
ing databases show that MDDF gains an advantage on
accuracy over Deep PCA. It also achieves comparable
accuracy with other well-known methods, using local
or global features. Figure 1. Illustration of the top-down Quad-tree parti-
tion and the bottom-up deep feature learning hierarchy.
PCA + LDA extract top vectors from each block to describe
2 HIERARCHICAL QUAD-TREE PARTITION local features. The green bounded node joins its children
nodes in the original order to preserve the global information.
Instead of dividing the face region into a uniform
grid, Quad-tree partitions the face region by means of
local discriminative variance. Larger partition means
the block has a lower feature density. By contrast, the
smaller partition means the block has a higher fea-
ture density. To make the partition more robust to local
noises, we consider the variance on all faces across the Figure 2. Fine-to-coarse Quad-tree partitions on Yale 2
entire database. Motivated by the idea of LDA which database.
encodes discriminative information by maximizing the
between-class scatter matrix Sb and minimizing the of Deep PCA tree structure when Tv = 0). The face
within-class scatter matrix Sw (see Eq. (1)), we define image is split into fewer and bigger blocks when Tv
a template face T by Eq. (2) to represent the distri- is large, but into more and smaller blocks when Tv is
bution of discriminative information for the database. small. Therefore, fine-to-coarse partition provides an
Thus, the total variance of the entire database is the opportunity to explore the effectiveness of varied deep
variance of the template: structures.

3 MULTI-DEPTH DEEP FEATURE LEARNING

where µ is the mean image of all classes, µi is the With a set of Quad-tree partitions, we are able to learn
mean image of class Xi , Ni is the number of samples features from a face. Deep feature learning produces a
in class Xi , and xk is the k-th sample of class Xi . bottom-up hierarchy of a feature representing face, in
We perform a top-down data driven Quad-tree parti- which the higher levels correspond to a shorter overall
tion on T to partition it into smaller blocks recursively, description of the face. It also encodes the correlation
according to a function doSplit(r), defined in Eq. (3). among the local patches. We create a hierarchy for
If the variance of a region r (starts from entire T ) is a face, based on the aforementioned Quad-tree parti-
higher than a threshold (Tv ∗ totalVar), then r is split tion using PCA + LDA. Figure 1 shows that a face is
into four sub-blocks with the same size. The partition partitioned into many blocks of varied sizes. Blocks
carries on under the criteria defined by the certifica- without a green ring are the leaf nodes at different
tion function in Eq. (3). Eventually, we have an uneven levels in the tree. These leaf nodes are used as the
partitioned face and an incomplete hierarchical struc- input for PCA + LDA, and select the top ki vectors as
ture of the face (see Fig. 1). Usually, it is rather difficult a feature basis, where ki is less than the corresponding
to find the best deep structure using one Tv . We, there- block size, i denotes the level index in the hierarchy.
fore, give a set of thresholds in an ascending order, and While the smaller i is, then the bigger ki becomes. Each
introduce fine-to-coarse face partitions (see Fig. 2). block is projected into a corresponding new basis, and
The leftmost partition is equivalent to the leaf nodes the four reduced-dimensionality neighbouring blocks

16
are then joined together back to their father node in
their original order. This process is repeated, using the
newly created layer as the data for the PCA + LDA and
join process to create the next layer up, until reaching
the root node.
As illustrated in Figure 1, the smallest blocks in the
Quad-tree partition measure 4 × 4 at level 3. We apply
PCA + LDA to these blocks, and extract the top 3 × 3 Figure 3. Template images on four databases: (a) ORL, (b)
Yale 2, (c) AR, (d) FERET.
vectors from each to describe the local feature. The
four neighbours are then joined in the inverse order
Table 1. Recognition accuracy of MDDF compared to other
of the partition, back to the father node at level 2, four reference methods.
where the original blocks have size of 8 × 8. At level
2, if a block was not further decomposed, PCA + LDA Database
extracts the top 6 × 6 vectors which are the same size
as the newly joined block from level 3. We recursively Method ORL Yale 2 AR FERET
apply PCA + LDA extraction, and join the neighbour
blocks to upper layer. Eventually, the procedure stops PCA + LDA 92.80 90.76 86.81 83.51
at root node. The last PCA + LDA selects about 30 MPCRC 91.50 92.80 88.60 73.64
top features as a vector. Obviously, these features pre- 30-Region 93.88 90.78 90.57 82.18
serve not only the global information thanks to the Deep PCA 92.33 91.11 84.67 85.18
MDDF 94.68 92.82 88.36 85.26
feature hierarchy, but also the local feature from blocks
partitioned by our Quad-tree partition.
and scarves). As in [4], a subset, with only illumi-
nation and expression changes, contains 50 male
subjects and 50 female subjects that are chosen.
4 EXPERIMENTS AND ANALYSIS
For each subject, we randomly choose 5 samples
for training and the left 9 images for test.
To demonstrate the effectiveness of MDDF, four public
[4] FERET database: 13539 images corresponding
and challenging databases were employed for eval-
to 1565 subjects. Images differ in facial expres-
uation: ORL [10], Extended Yale (Yale 2) [11], AR
sions, head position, lighting conditions, ethnicity,
[12], and FERET [13]. Face images in these databases
gender, and age. To evaluate the robustness of
are under variant conditions, including head poses
our methods with regard to facial expressions, we
(ORL, AR), illumination changes (ORL, Yale 2, AR,
worked with a subset of front faces labelled as Fa,
and FERET), facial expressions (ORL, AR, FERET),
Fb, where Fa represents regular facial expressions,
and facial details (e.g. with glasses or not: ORL, AR,
and Fb alternative facial expressions. All Fa are
FERET). Face images were cropped to 32 × 32. The
selected for training data, while Fb as test data.
template images created on each database are shown
in Figure 3. An example of a fine-to-coarse Quad-tree To verify the feature performance, we compared
partition is shown in Figure 2. Our method generated proposed MDDF with various methods using global
8–15 Quad-trees depending on the databases. These features, local features, and canonical deep fea-
Quad-trees are indexed according to the threshold Tvi tures, respectively. They are: (1) the conventional
in an ascending order. PCA + LDA method [2] which extracts a global fea-
ture vector from the whole face region; (2) MPCRC
[1] ORL database: 40 subjects, 10 samples per sub-
[4], which develops a multi-scale local patch-based
ject, with variances in facial expressions, with
method to alleviate the problem of sensitivity of path
open or closed eyes, with glasses or no glasses,
size; (3) a 30-region method [3], which defines the
scale changes (up to about 10 percent), and head
30 regions according to experimental experience; (4)
poses. 5 samples per subject are randomly selected
the Deep PCA [7], which integrates the discriminative
as training data while the left ones are selected as
information extracted from uniform local patches to
test data.
a global feature vector. Table 1 shows the comparison
[2] Extended Yale database (Yale 2): more than
results, and gives five observations:
20,000 single light source images of 38 subjects
with 576 viewing conditions (9 poses in 64 illumi- Observation (1): PCA+LDA extracts a global fea-
nation conditions). To evaluate the robustness of ture that focuses on the holistic information on
our method on the illumination changes, 5 samples the image. It could be regarded as a summary of
of the 64 illumination conditions are randomly faces, but ignores details such as local variations.
selected as training data, the remaining 59 images Thus, PCA + LDA has a rather average perfor-
as test data. mance on various databases, neither bad nor good.
[3] AR database: 4000 colour face images of 126 The local patch-based methods focus more on local
people (70 men and 56 women), including frontal variations, and perform better on Yale 2, AR and
views of faces, with different facial expressions, FERET databases which contains more local defor-
lighting conditions, and occlusions (sun glasses mations caused by facial expressions, illumination

17
 changes, and occlusions, etc. This implies the sig-
nificant advantage of the subregion-based features
over the holistic-based features in dealing with local
deformations.
Observation (2): the MPCRC method outperforms
PCA + LDA on Yale 2 and AR databases obviously
due to the robustness of the patches in local vari-
ations, but degrades significantly on the FERET
database where only a Single Sample Per Person
(SSPP) is collected as gallery data. Under the SSPP,
the MPCRC degenerates into the original patch-
based method PCRC which has no collaboration
with multiple patches in variant scales. Since the
performance of local patch-based methods is very
sensitive to patch size, they suffer from severe
degradation in performance under inappropriate
patch size. That is why many local patch-based
methods cooperate with global features to over-
come this problem. This motivates the development
of our multi-depth deep structure learning to asso-
ciate local features with a global representation of Figure 4. Recognition accuracy against a Quad-tree index
an entire image structure. during a fine-to-coarse partition on four databases: (a) ORL,
Observation (3): a 30-Region method is composed of (b) Yale 2, (c) AR, and (d) FERET.
30 subregions which have large overlaps between
each other. All these subregions are empirically Quad-tree partitions the face into small patches with
designed according to facial structures after being the same size. We can see that varied deep struc-
well registrated. Particularly, one ‘subregion’ actu- ture features perform quite a large range. The best
ally is the entire facial region. We can think of it as performance is usually achieved at a certain parti-
a brute-force integrating global and local features. tion, depending on the database, but this does not
Therefore, it obtains a higher performance on the come from the Deep PCA in most cases. Because
AR database being well-registered and having only our Quad-Tree-based partition is processed on the
frontal features. It also performs well on an ORL template image, which is obtained from Sb and
database because of the effectiveness of its subre- Sw , and deemed as a summary of the database.
gions. However, the performance degrades in not This data-driven strategy makes our method adapt
well registered databases, such asYale 2 and FERET. to databases, and generates the most appropriate
The reason might be that these sub-regions do not partition for deep structure learning automatically.
fit with the data. This suggests that the design of the Our method benefits from the data-driven image
sub-regions must adapt to diverse data. partition-based deep structure learning method and
Observation (4): Deep PCA improves the perfor- it can be widely applied to diverse databases, espe-
mance of conventional PCA + LDA methods on cially for those with a large number of variations.
Yale 2 and FERET databases, which shows the effec- It must be pointed out that MDDF performs worse
tiveness of deep structure learning for coping with than the 30-region method but has a better perfor-
local deformations. However, it performs worse on mance than the Deep PCA method on AR database.
an AR database. The reason might be that Deep We argue the reason that MDDF learns a multi-
PCA assumes that the discriminative information depth structure feature from the template face of the
is evenly distributed. It partitions the face into a database. But sub-regions in 30 regions are empir-
uniform grid and evenly aggregates features from ically designed for well-registered databases, and
the lower level. This result indicates the importance ARs, but not for others.
of designing a data-driven deep structure learning
method.
Observation (5): the proposed MDDF achieves the
best performance in most experiments. The multi- 5 CONCLUSION
depth deep feature learning is developed based on a
fine-to-coarse Quad-tree partition. To explore how This paper proposes a novel deep structure learning
varied deep structure affects the effectiveness of method Multi-depth Deep Feature (MDDF) for face
deep features, we defined a set of thresholds for recognition. It unevenly partitions face images accord-
Quad-tree partition. Figure 4 plots the recognition ing to the density of the discriminative power in the
accuracy against the Quad-tree index (the indices local regions, and learns a data-driven deep struc-
are corresponding to the thresholds Tvi in ascending ture that preserves more information at the region of
order). Please note that the leftmost dot indicates interest. The comparison with diverse methods using
the accuracy of Deep PCA because Tv = 0 and global, local features, and the canonical deep structure

18
feature ‘Deep PCA’ shows the comparable perfor- [4] P. Zhu, L. Zhang, Q. Hu, and S. Shiu, “Multi-scale patch
mance of MDDF in four challenging databases. We based collaborative representation for face recognition
can conclude that the MDDF has the most accurate with margin distribution optimization,” in ECCV, 2012.
description of image structure for recognition. In cur- [5] Y. Bengio and Y. LeCun, “Scaling learning algorithms
towards AI,” 2007.
rent work, the dimension of extracted features from [6] G. E. Hinton and S. Osindero, “A fast learning algo-
bigger blocks, which are not further partitioned, is the rithm for deep belief nets,” Neural Computation, vol.
same as the one aggregated from four smaller blocks 18, pp. 1527–1554, 2006.
at a lower level. In feature research we will explore [7] B. Mitchell and J. Sheppard, “Deep structure learning:
what feature dimensions of these bigger blocks would beyond connectionist approaches,” in IEEE ICMLA,
induce an optimal performance. 2012.
[8] F. Berisha, A. Johnston, and P. W. McOwan, “Identify-
ing regions that carry the best information about global
ACKNOWLEDGEMENT facial configurations,” Journal of Vision, vol. 11, no. 10,
pp. 1–8, 2010.
[9] U. Park, R. Jillela, and A. Ross, “Periocular biometrics
This work is supported by: Japan Society for the Pro- in the visible spectrum,” IEEE Trans. on Info. Forensics
motion of Science, Scientific Research KAKENHI for and Security, no. 6, pp. 96–106, 2011.
Grant-in-Aid for Young Scientists (ID: 25730113). [10] Ferdinando S. Samaria and Andy C. Harter, “Parame-
terization of a stochastic model for human face iden-
tification,” in Proceedings of 2nd IEEE Workshop on
REFERENCES Applications of Computer Vision, pp. 138–142, 1994.
[11] A. S. Georghiades, P. N. Belhumeur, and D. J. Kriegman,
[1] M. Turk and A. Pentland, “Eigenfaces for recognition,” “From few to many: Illumination cone models for face
J. Cognitive Neuroscience, vol. 3, no. 1, 1991. recognition under variable lighting and pose,” IEEE
[2] P. N. Belhumeur, J. P. Hespanha, and D. J. Kriegman, Transactions on PAMI, vol. 23, no. 6, pp. 643–660,
“Eigen faces vs. Fisher faces: Recognition using class 2001.
specific linear projection,” IEEE Trans. PAMI, vol. 20, [12] A. M. Martinez and R. Benavente, “The AR face
no. 7, pp. 711–720, 1997. database,” CVC Technical Report, 1998.
[3] L. Spreeuwers, “Fast and accurate 3D face recognition [13] P. Jonathon Phillips, H. Moon, S. A. Rizvi, and P. J.
using registration to an intrinsic coordinate system and Rauss, “The FERET evaluation methodology for face
fusion of multiple region classifiers,” IJCV, vol. 93, recognition algorithms,” IEEE Transactions on PAMI,
no. 3, pp. 389–414, 2011. vol. 22, no. 10, pp. 1090–1104, 2000.

19
This page intentionally left blank
 Informatics, Networking and Intelligent Computing – Zhang (Ed.)
© 2015 Taylor & Francis Group, London, ISBN: 978-1-138-02678-0

Research on camera calibration basing OpenCV

H.M. Nie
Heilongjiang University of Science and Technology, China

ABSTRACT: According to the principle of camera calibration, it puts forward a calibration technique based on
OpenCV camera, and with the help of open-source computer vision library, OpenCV completes the calibration in
VS 2008 development platform. The experiment proves that the calibration procedure based on OpenCV camera
has the following advantages: calibration accuracy, high computation efficiency, good cross platform portability,
which can be effectively applied in the field of computer vision system.

Keywords: OpenCV; Camera calibration

1 INTRODUCTION The camera coordinate system: the camera itself

constitutes an object image coordinate system, and the
People usually complete the reconstruction of natural coordinates of a target point in the coordinate system
objects by imitating biological method, using different is expressed by (Xc , Yc , with Zc ).
imaging system to replace the visual system, and using The world coordinate system: we live in a three-
computers to replace the human brain to complete the dimensional world coordinate, and coordinates of a
three-dimensional reconstruction. That is to use two target point in the coordinate system is expressed by
cameras in different view point at the same time to cap- (Xw , Yw , with Zw ).
ture image information, through the camera calibration
matrix operations to obtain three-dimensional coor-
dinates, and then synthesize these three-dimensional 2.2 Linear camera model
coordinate information to establish the object model
in three-dimensional space, in order to make it more In the case of no distortion, setting the camera imag-
convenient and accurate to observe and operate on the ing is an ideal pinhole imaging; Pinhole camera model
object in all directions. OpenCV is an open-source is shown in Figure 1. Figure 1, 0 point as the cen-
computer vision library which is developed by Intel. It ter of projection is the coordinate origin. And 01
consists of a series of C and C++ functions, which point which is the intersection of the optical axis and
realizes the computer vision and image processing the image plane, is the center point of the imaging
of many common algorithms, including object track- plane.
ing, image processing, pattern recognition, motion Q point coordinates of (Xc , Yc , Zc ) in Figure 1 are
analysis, structure analysis, camera calibration and the coordinates Q in the camera coordinate system.
3D reconstruction algorithm. The camera calibration Q point coordinates (Xw , Yw , Zw ) are the coordinates
module of OpenCV which provides a good interface Q point in the world coordinate system. q coordinates
for users, which supports windows, Linux platform (Xu ,Yu ) are the coordinates of the q point in the image.
to improve the development efficiency, and enhances In Figure 1, f is the distance from center of projection
the portability of program, can be used for the actual to the image plane, the focal distance.
development project. According to the myopia of triangle principle, it
comes to the following formula:

2 CALIBRATION PRINCIPLE

2.1 The definition of the concept

The image coordinates: image formed from objects
projecting onto the two-dimensional plane, the estab-
lishment of coordinate system, and the coordinate of a In the real production process of camera, 01 as the
target point in the coordinate system are expressed by image center may not fall on the optical axis. The opti-
(Xu , Yu ). cal axis in a plane coordinate sets off (X0 , Y0 ), so X0

21
 In Formula (10), R is 3 × 3 rotation matrix; T is the
translation matrix 3 × 1

Formula (11, 12) are the camera external parameters.

Figure 1. Pinhole camera. 2.3 Nonlinear camera model

and Y0 are the offset 01. Therefore, formula 1 and 2 The intensity of camera must be increased for fast
are changed into the following: imaging. The usual practice is to add the lens in front
of the camera, which will cause of image distortion.
Therefore we need to correct the distortion introduced
by the camera.
There are two forms of distortion caused by lens:
the radial distortion and tangential distortion. Each
point in the imaging sensor plane can be presented by
(x, y) Cartesian coordinates. It also can be presented
Also in the practical production process in camera,
in polar coordinates (r, t), the vector representation,
pixel sensor cannot be processed into a square, usually
where r is a vector of length; t is the vector of horizontal
rectangular. So we let the pixel sensor length and width
angle; and the center is located in the center of the sen-
be defined as x and y. Therefore we get the following
sor. Radial distortion is the change of vector endpoint
formulas:
length direction (r direction). Tangential distortion
is the angle change of vector endpoint extension of
tangent direction (t direction). Radial distortions will
not distort when r equates 0, so we use the first few
items of the Taylor series expansion to quantitatively
describe. That is:
Adding the Formula (3) and Formula (5)

Adding the Formula (4) and Formula (6)

The tangential distortion is due to the fact that we
introduce two parameters p1 , p2 to describe

Representing the Formula (7, 8) by the homogeneous

coordinate

So calibration of camera is actually the measure

of internal and external parameters and distortion
parameters of the camera.

According to Formula (9), we can complete the conver-

sion between image coordinate and camera coordinate. 3 THE CALIBRATION METHOD OF OPENCV
We also need to converse the camera coordinate
system into the world coordinate system. We apply planar checkerboard calibration template in
OpenCV camera calibration, through the free mobile
camera or template, grasping planar calibration tem-
plate images at different angles, to achieve the cali-
bration of camera. The checkerboard calibration plate
has length of 20 mm, 9 rows and 13 columns, a total
of 96 corners of the chessboard as template to grab a
different angle images using at least squares calibra-
tion calculation. Then it put multiple calibrated images

22
 cvSeqPush to save the sub-pixel coordinates into
the coordinate sequence.
(6) to substitute the sub-pixel coordinates of corner
points and corner points in the world coordinate
value into the cvCalibrateCamera2 (), and to get
the camera internal and external parameters and
distortion parameters.
(7) to release memory space of function allocated and
to prevent memory leaking.

4 THE IMPLEMENTATION OF
PROGRAMMING

Programs are programming, debugging and testing

by using VS 2008 in Windows XP. Due to space
limitations, variable definitions and initializations are
omitted. Here are the key codes:
CvSize CBoardSize=cvSize(rCount, cCount) ; //The
calibration board size
if((srcimage=cvLoadImage(filename,1))==0)
//Load the image
continue;
//Failed to load, to continue loading the next picture
cvCvtCoor(srcimage,grayimage,CV_BGR2GRAY);
//Change color image into gray image
CvPoint2D32f * pCorners = ( CvPoint2D32f * )
( rCount* cCount* sizeof(CvPoint2D32f)) ;
// store the detected corner
Int iCount; //Save the detected corner points
reult=cvFindChessboardCorners(grayimage,
CBoardSize, pCorners,& iCount,
CV_CALIB_CB_ADAPTIVE_THRESH);
// Obtain the angle point
for( int i = 0; i < iCount; i ++ )
{
cvFindCornerSubPix (grayimage, pCorners, iCount,
cvSize( 11, 11), cvSize(- 1, -1), cvTermCriteria(
CV_TERMCRIT_EPS+
CV_TERMCRIT_ITER, 30, 0.1) ) ;
Figure 2. Calibration process. // gain exactly each corner coordinates
cvSeqPush( pSeq, pCorners) ;
//Store obtained coordinates in the sequence
into the same directory, and make it ready for the cal- }
ibration program reading. The calibration process is cvCalibrateCamera2 ( pObj, pIP, pPC, iS, pIc, pDn,
shown in Figure 2. pRn, pTn, 0) ;
// Obtain correct data
(1) to obtain calibration image directory file list. pObj is the world coordinates of corners.
(2) through the imread () to load the image. This pIP is the image coordinates of corners.
function supports the commonly used file format. pPC is the number of different image corner points.
(3) to call cvFindChessboardCorners () function to iS is the image size.
find chessboard corner, and succeed in returning pIc is the camera intrinsic matrix.
non-zero and failed returns zero. Function suc- pDn is the camera distortion coefficient.
ceeds and returns corner coordinates. Call this pRn is a camera rotation vector.
function, if the image is a true color image it must pTn is the camera translation vector.
be converted to gray scale image. If the image is
gray scale images it does not to be converted.
(4) to use cvCreateMemStorage () and cvCreateSeq () 5 EXPERIMENTAL DATA
to create the stored angular coordinate sequence.
(5) to call cvFindCornerSubPix () to obtain the sub- According to the calibration principle above, we have
pixel coordinates of corner values, and to use developed an experimental program, the program

23
Table 1. Comparison of camera parameters. computer vision. While in these studies, it is neces-
sary to determine the geometrical relationship between
Camera Imaging Matlab the corresponding points in visual images and in the real
parameters results calibration results world. The purpose of Camera calibration is to estab-
lish the correspondence between 3D world coordinate
fx 677.761683 677.911558
and 2D image coordinate. The camera calibration
fy 699.952933 703.2464865
x0 337.5338665 338.213853 procedure developed with OpenCV, has advantages
y0 285.015533 285.95167 of accurate calibration results, high calculating effi-
k1 −0.3013115 −0.298848 ciency, good cross-platform portability, which can
k2 0.13680118 0.135938 be applied in the field of computer vision system
p1 0.000446 0.000386 effectively.
p2 −0.00133 −0.001355

*Note Parameter k3 is ignored in Matlab, so it is not given in REFERENCES

this paper.
[1] Cheng Jianpu, Xiang Huiyu. A Camera Calibration of
Vision Measurement of Body Panel Based on OpenCV
results are saved as a text file. After several debugging, [J]. Mechanical design and manufacturing. 2010, 11:
finally the program is stable, accurate, and success- 198–200.
fully finds focus, setting a group of 20 pieces of [2] Feng Saleisi, Digital Image Process (Matlab Vision) [M].
Beijing: Electronic Industry Press, 2005.
1280 * 1024 pixel picture, when using 1.2 s, which can [3] Sun Jie, Zhu Shiqiang, Lai Xiaobo. A Method of Vision
satisfy the actual demand. At the same time in order Based Navigation Camera and Effective Calibration [J].
to verify the accuracy of the data, we apply Matlab to Computer Engineering 2010, 36(21): 212–213.
calibrate the 20 image for comparison. Parameters of [4] Tian Yuan-yuan, Tan Qing-chang. Study of CCD laser-
program calibration number camera and parameters of range finderbased on static image [J]. Micro—Computer
Matlab camera are shown in Table 1. Information, 2007, 11(31): 96–98.
[5] Weixin Kang, Zhou Sheng. Trailing Vehicle Video
Retrieval basedon Kalman and OpenCV [R]. Proceed-
6 CONCLUSION ings of 2010 3rd International Conference on Future
BioMedical Information Engineering (Volume 2), 2010.
The visual measurement and 3D reconstruction are
important parts in the current study on application of

24
 Informatics, Networking and Intelligent Computing – Zhang (Ed.)
© 2015 Taylor & Francis Group, London, ISBN: 978-1-138-02678-0

Designing fuzzy rule-based classifiers using a bee colony algorithm

I.A. Hodashinsky, R.V. Meshcheryakov & I.V. Gorbunov

Tomsk State University of Control Systems and Radioelectronics, Tomsk, Tomsk Oblast, Russia

ABSTRACT: This paper proposes a fuzzy approach which enables one to build intelligent decision-making
systems. We present new learning strategies to derive fuzzy classification rules from data. The training procedure
is based on a bee colony algorithm. We observe four components of the algorithm: initialization, work of scout
bee, rules of the antecedent’s generation, and weight tune. The evaluation of the fuzzy rule-based classifier
performance, adjusted according to given algorithms, is finally applied to well-known classification problems,
such as bupa, iris, glass, new thyroid, and wine. The comparison with the work of such algorithms as Ant
Miner, Core, Hider, Sgerd, and Target is made in training and in testing random selections. We describe the
implementation of the fuzzy rule-based classifier to forecast the efficiency of the non-medical treatment of
patients rehabilitated in the Federal State Institution of Tomsk Scientific-Research Institute of Health-Resort
and Physiotherapy Federal Medical and Biological Agency of Russia.

1 INTRODUCTION successfully been proven in many real life applica-

tions where domain knowledge is imprecise or inexact.
Typical problems of decision-making theory is the A common problem with the implementation of such
selection of one or more of the best objects (options, systems, however, is the acquisition of the production
alternatives), and the ordering or ranking of objects rules on which decision-making is based.
is based on their properties and classification by a The fuzzy rule-based classification systems on
set of categories (Petrovsky, 2007, Meyer & Roubens, the ground of IF-THEN rules find wide implemen-
2005). Many real life decision-making problems can tations in the solutions of management problems,
be treated as a classification which is a decision- decision making, learning, adaptation, generalization,
making process that involves making choices. Many and reasoning. Fuzzy rules can deal with imprecise
applications like pattern recognition, disease diagno- knowledge and the uncertainty of information and
sis, and credit scoring can be considered as classifica- strengthen the knowledge representation power. The
tion problems. Classification consists of predicting the main advantage of such classifiers in comparison with
value of a (categorical) attribute (the class) based on classifiers such as the black box (neural networks) is
the values of other attributes (the predicting attributes). their interpretability. Fuzzy rules represent knowledge
The classifier can be considered as a mapper from acquired from empirical observations and experience.
the feature domain into the class labels domain. A set The use of descriptive linguistic variables in fuzzy
of different types of classifiers exist which use dif- rules provides the additional advantage of represent-
ferent approaches to perform this mapping such as: ing this knowledge in a form that is easy for humans
1) linear discriminate analysis; 2) Bayesian classifiers; to comprehend and validate.
3) decision trees; 4) neural networks-based classi- In the past, fuzzy classifiers were created by fuzzy
fiers; 5) support vector machines; 6) fuzzy rule-based rules based on apriori knowledge and expert knowl-
classifiers, and etc. (Angelov & Zhou, 2008). edge, but in many applications, it is difficult to obtain
It would be helpful if we could design a classi- fuzzy rules without apriori data. In recent years, the
fier based on linguistic interpretable rules, because so-called data-driven approaches have become dom-
it expresses the behaviour of the system in a human- inant in the fuzzy systems design area (Angelov &
readable form. One of the approaches to solve this Lughofer, 2008), (Angelov & Zhou, 2008), &
classification problem is to formulate a solution using (Lavygina & Hodashinsky, 2011).
a fuzzy rule-based classifier. The generally accepted problem of fuzzy sys-
Fuzzy set theory, fuzzy logic, and fuzzy reasoning tems tuning is a rule-base formation. For generat-
methods contributes to the development of alterna- ing fuzzy classification rules, in (Ishibuchi H. &
tive models for uncertainty that are of interest for Nojima Y. 1999), (Ho, Chen, Ho, & Chen, 2004),
building alternative approaches to classical decision (Mansoori, Zolghadri, & Katebi, 2008), and (Chang &
theory (Blanning, 1996), (Radojević & Suknović, Lilly, 2004), a genetic and evolutional algorithm
2008), and (Ye, 2012). Fuzzy rule-based systems have are proposed. Some studies have attempted to solve

25
 Table 1. Numerical benchmark functions.

Minimum
Function Ranges value

sin2 ( x12 + x22 ) − 0.5
f1 (x) = 0.5 + −100 ≤ xi ≤ 100 f1 (0) = 0
(1 + 0.001(x12 + x22 ))2

n
f2 (x) = xi2 −100 ≤ xi ≤ 100 f2 (0) = 0
n
i=1  n  
1 
xi − 100
f3 (x) = (xi − 100)2 − cos √ +1 −600 ≤ xi ≤ 600 f3 (100) = 0
4000 i=1 i=1 i

n
f4 (x) = (xi2 − 10cos(2πxi ) + 10) −5.12 ≤ xi ≤ 5.12 f4 (0) = 0
i=1

n−1
f5 (x) = (100(xi+1 − xi2 )2 + (xi − 1)2 ) −50 ≤ xi ≤ 50 f5 (1) = 0
i=1

the classification problem by employing hybrid (Karaboga, 2005), which is a recent swarm intelli-
approaches on the grounds of decision trees and evo- gence based approach to solve nonlinear and complex
lutional algorithms (Pulkkinen & Koivisto, 2008), and optimization problems. It is as simple as PSO and dif-
machine learning methods, Learning Classifier Sys- ferential evolutionalgorithms, and uses only common
tems, in particular, and the given approach is based on control parameters such as colony size and maximum
reward training and genetic algorithms (Ishibuchi & cycle number.
Nojima, 2007). In studies (Karaboga & Basturk,2008), (Karaboga &
An even more recent approach is that of Swarm Akay, 2009), the performance of the ABC algorithm
Intelligence. The two most widely used swarm intelli- is compared with that of differential evolution (DE),
gence algorithms are Ant Colony Optimisation (ACO) PSO and Evolutionary Algorithms (EA) for multi-
(Dorigo, Maniezzo, & Colorni, 1996) and Particle dimensional and multimodal numeric problems. Clas-
Swarm Optimisation (PSO) (Kennedy & Ebenhart, sical benchmark functions are presented in Table 1.
1995). In (Casillas et al., 2005), an approach to In experiments, f1 (x) has 2 parameters; f2 (x) has 5
the fuzzy rule learning problem ACO algorithms parameters and f3 (x), f4 (x) and f5 (x) functions have
is proposed. This learning task is formulated as a 50 parameters. Parameter ranges, formulations, and
combinatorial optimization problem, and the features global optimum values of these functions are given in
related to ACO algorithms are introduced. In (Abadeh Table 1 (Karaboga & Basturk, 2008).
et al., 2008) an evolutionary algorithm to induct fuzzy The mean and the standard deviations of the func-
classification rules is proposed which uses an ant tion values obtained by DE, PSO, EA, and ABC
colony optimization based local searcher to improve algorithms are given in Table 2 (Karaboga & Basturk,
the quality of the final fuzzy classification system. 2008).
The proposed local search procedure is used in the Simulation results show that the ABC algorithm
structure of a Michigan based evolutionary fuzzy performs better than the above mentioned algorithms
system. and can be efficiently employed to solve the multi-
Another category for fuzzy rule-based classifiers modal engineering problems.
design is PSO. The article (Elragal, 2008) discusses The ABC has a lot of advantages in memory, local
a method for improving accuracy of fuzzy-rule-based search, solution improvement mechanism, and so on,
classifiers by using particle swarm optimization. In and it is able to get excellent performance on opti-
this work, two different fuzzy classifiers are consid- mization problems (Karaboga & Akay, 2009), (Zhao
ered. The first classifier is based on the Mamdani et al., 2010). In recent years, the ABC algorithm
fuzzy inference system while the second one is based has been successfully used to solve hard combina-
on the Takagi-Sugeno fuzzy inference system. The tional optimization problems including traveling the
parameters of the proposed fuzzy classifiers include salesman problem (Li, Cheng, Tan & Niu, 2012),
antecedent, consequent parameters, and the structure the quadratic knapsack problem (Pulikanti & Singh,
of fuzzy rules, optimized by using PSO. 2009), and the leaf-constrained minimum spanning
In our study we implement the novel bee colony a tree problem (Singh, 2009). In the article (Singh,
algorithm to identify the structure and parameters of 2009) by comparing the approach against genetic algo-
the fuzzy rule-based classifier. There are many widely rithm, ant-colony optimization algorithms, and tabu
known algorithms based on the behaviour of honey searches, Singh reported that the ABC outperformed
bees in nature. These algorithms can be divided into the other approaches in terms of best and average
two categories and they correspond to the behaviour solution qualities and the computational time.
of bees while gathering food and mating. Our study is The ABC algorithm has been applied to various
based on the Artificial Bee Colony (ABC) algorithm problem domains including the training of artificial

26
Table 2. The results obtained by DE, PSO, EA, and ABC The fuzzy classification is described by the func-
algorithms. tion f: → [0,1]m , which refers the classified object
to each class with the definite grade of member-
DE PSO EA ABC ship being calculated in the following way: βj (x) =

n
f1 (x) = 0 0 ± 0 0.00453 ± 0±0 0±0 Aki (xk ) · CFi , j = 1, 2, . . . , m.
0.0009 Rij k=1
f2 (x) = 0 0±0 2.5113E-8 ± 0 0±0 0±0 The classifier is the class being defined in the
f3 (x) = 0 0±0 1.5490 ± 0.00624 ± 0±0 following way: class = cj∗ , j ∗ = arg max βj .
0.067 0.00138 1<j<m
f4 (x) = 0 0±0 13.1162 ± 32.6679 ± 0±0 The fuzzy classifier can be presented as a function
1.4482 1.9402 c = f(x, θ, CF), where θ – the rule base containing
f5 (x) = 0 35.3176 ± 5142.45 ± 79.818 ± 0.1331 ± rules of type (1).
0.2744 2929.47 10.4477 0.2622 Let us assume that the multitude of teaching data
(observation table) is given by {(xp ; cp ), p = 1, . . . , z},
and let us define the following unit function:
neural networks (Karaboga, Akay B, & Ozturk, 2007),
(Kumbhar & Krishnan, 2011), the design of a digi-
tal infinite impulse response filter (Karaboga, 2009),
solving software testing (Suri & Snehlata, 2011), and
the prediction of the tertiary structures of proteins then the computational criteria of classification system
(Benitez & Lopes, 2010). However, it has not yet been accuracy can be defined in the following way:
used to tune a fuzzy rule-based classifier.
The main motivation of this paper is to present a
novel classifier design. The idea is to use the improved
bee colony algorithms for finding more accuracy in
real life data classification.
The paper is organized as follows: Section 2 The problem of identification turns into the problem
introduces the main theoretical aspects of the fuzzy of maximum search for the given function in multi-
rule-based classification systems; Section 3 briefly dimensional space, coordinates of which correspond
describes the artificial bee colony technique; Section to the fuzzy system parameters. To optimize θ, it is
4 describes a modified Bees algorithm and how it was advisable to use the bee colony algorithm which is set
used in this paper and Section 5 shows test results in to generate and change the rule base. And to set CF, it
the well-known data set and results in the forecasting is advisable to use the modified bee colony algorithm
of non-medical treatment efficiency. The final section which relies on the characteristics of realization of the
offers the conclusions. bee dance operation.

2 FUZZY RULE-BASED CLASSIFIER

3 THE ARTIFICIAL BEE COLONY
Fuzzy classification is a decision process based on ALGORITHM
fuzzy logic. Fuzzy rule-based classifiers consist of
classification rules. Fuzzy rules are a collection of lin- Artificial Bee Colony is a novel optimization algo-
guistic statements that describe how a fuzzy inference rithm inspired by the natural behaviour of honey bees
system should make a decision regarding that an input in their search of the best food sources. This tech-
is classification (Fernandez & Herrera, 2012). This nique was proposed by Karaboga, (2005) and further
classifier determines a mapping from a given input to improved by Karaboga & Basturk, (2008).
an output representing the class: Rij : The ABC algorithm works with a colony of artifi-
cial bees. The colony consists of three groups of bees:
IF x1 = A1i AN, AND xk = Aki AND
employed bees (workers), on lookers, and scouts. The
AND xn = Ani THEN class = cj , w = CF i (1)
position of a food source represents a possible solution
where x = (x1 , x2 , x3 , . . . , xn ) – vector of decision vari- to the optimization problem and the nectar amount of a
ables (attributes or features); Aki – the fuzzy term food source corresponds to the quality (fitness) of the
which characterizes k-marker in i-rule (i ∈ [1, R]), R – associated solution. The number of the employed bees
the number of rules; cj – identifier of j-class, j ∈ [1, m]; or the onlooker bees is equal to the number of solutions
CF i – rule weighing coefficient or degree of belief of in the population. At the first step, the ABC gener-
i-rule, CF i ∈ [0, 1]. ates a randomly distributed initial population of SN
The task of traditional classification can be solutions (food source positions), where SN denotes
described by the function f : → {0, 1}m , where the size of population. Each solution xi (i= 1, 2, . . .,
f (x) = (c1 , c2 , . . . , cm ) and also ci = 1, and cj = 0 ( j ∈ SN) is a D-dimensional vector. Here, D is the num-
[1, m], i = j), and the object set by vector x belongs to ber of optimization parameters. After initialization, the
ci class. population of the positions (solutions) is subjected to

27
repeated cycles, C= 1, 2, . . ., MCN, of the search pro- 4.1 The algorithm of the rule base initialisation on
cesses of the employed bees, the onlooker bees and the ground of the training data set
scout bees.
This algorithm is aimed to form a primary rule base
The main steps of the algorithm are given below
which has one rule for each class. In addition, the algo-
(Akay & Karaboga, 2009), (Karaboga & Akay, 2009):
rithm eliminates those which are not covered with any
1: Initialize the population of solutions xij , i = 1, . . . , terms of change in the input variables.
SN and j = 1, 2, . . . , D Input: The number of mclasses, the observation
2: Evaluate the population table {xpi }, Type – the type of membership function.
3: C = 1 Output: The primary rule base of θ classifier.
4: repeat{Employed Bees’ Phase} Algorithm: 1: θ = {∅}
5: Produce new solutions vi for the employed bees 2: For each q class Do:
by vij = xij + φij (xij − xkj ), 2.1: For each i marker Do:
where k ∈ {1, 2, . . . , SN } and j = 1, 2, . . . , D are 2.1.1: Searching for min classqi = min(xpi )
p
randomly chosen indices. Although k is deter-
2.1.2: Searching for max classqi = max(xp,i )
mined randomly, it has to be different from i. And p
φij is a random number between [−1, 1] which 2.1.3: Term Aiq of the type Type creation, laying the
controls the production of neighbouring food interval [min classqi , max classqi ]
sources around xij and represents the comparison End do (i);
of two food positions as seen by a bee. 2.2: The rule creation type 1 with class = cq , w = 1
6: Apply the greedy selection process between vi and and antecedent contains Aiq for each xi
xi {Onlooker Bees’ Phase} 2.3: θ := θ ∪ {Rq }
7: Calculate the probability values pi for the solutions End do (q);
xi by: 3: With each I marker Do:
3.1: Check the existence of the areas to ensure that
is not covered by area term
3.2: IF not, covered places are found THEN :
3.2.1: The closest term to the left removes its right
where fit i is the fitness value of the solution i border according to the size of the empty distance
which is proportional to the nectar amount of the 3.2.2: The closest term to the right removes its right
food source in the position i and SN is the number border according to the size of the empty distance
of food sources which is equal to the number of End IF;
employed bees or onlooker bees. End do (i).
8: Produce the new solutions vi for the onlookers Figure 1a) represents the results of the work of the
from the solutions where xi is selected depending initialisation algorithm after step 2. Three classes set
on pi and begins to evaluate them on the xi marker by triangular membership functions
9: Apply the greedy selection process between vi and are considered. The run of xi input marker – [mini ,
xi {Scout Bees’ Phase} maxi ]. Not the whole run is laid by terms, but the
10: Determine the abandoned solution for the scout, areas [max class3i , min class1i ] and [max class1i , maxi ]
if it exists, and replace it with a new randomly are not covered. Figure 1b) represents the outcome of
produced solution xi by work of the algorithm for xi marker.

11: Memorize the best solution achieved so far 4.2 The work of the scout bee algorithm
12: C = C + 1 In this algorithm scout-bees create one fuzzy rule for
13: until C = MCN the chosen class. The rule generated by the algorithm
contains fuzzy terms with a previously defined view
of membership function. In actual fact, the algorithm
4 THE ARTIFICIAL BEE COLONY realizes the methodology of random search of the
ALGORITHM fuzzy rule.
Input: The number of n features, mini – the mini-
The design process of a fuzzy rule-based system mum value of i feature, maxi – the maximum value of
from a given input-output data set can be presented i feature, Type – the type of the membership function,
as a structure- and parameter-optimization problem q – the current class.
(Evsukoff et al., 2009). Output: The rule of the fuzzy classifier R.
Some algorithms take part in the fuzzy classifier Algorithm:
tuning: the algorithm of the rule-base initialisation on 1: For each i marker Do:
the ground of a training data set, the base algorithm of 1.1: Set the term Aiz at random, corresponding to
a rule generation (it includes the algorithm of work of the type of the membership function Type, in such a
scout bee), and the modified algorithm of a rule weigh- way that the left border Aiz > mini and the right border
ing coefficient setting. The step-by-step submission of Aiz < maxi .
the above mentioned algorithms is given below. End do (i);

28
 REllite – the multitude from l the closest to
BestScout in terms of the result from the multitude
RScouts;
RElliteWork j – the multitude from o rules received
as a result of perfecting of the rule REllitej by working
bees, 1 < j < l;
ABase – initial classification accuracy at the
current iteration;
ABestScout – the increase of the classifier work
precision at the cost of inclusion of BestScout into
the rule base;
AScoutsi – the increase of the classifier work preci-
sion at the cost of inclusion of RScoutsi into rule base
Figure 1. The work of the initialization theory RBase, 1 < i < s;
demonstration. ABestScoutWork i – the increase of the classifier
work precision at the cost of inclusion of RBestScout-
Work i into the base RBase, 1 < i < o;
2: Create the rule type 1 with class = cq , w = 1 and AElliteWork ji – the increase of the classifier work
antecedent contains Aiz for each xi . precision at the cost of inclusion of RBestScout-
Work ji into the base RBase, 1 < j < l, 1 < i < o;
4.3 The base algorithm of the bee colony for CountAddRulle – the number of the rules added into
generation of the fuzzy classifier rules rule base.
Algorithm: 1: θ ∗ = θ
The given algorithm serves for the fuzzy classifier rule
2: Calculate ABase = E(θ, {1})
base formation and aims to receive the initial rule base
3: Scout bees generate s rules RScoutsi , 1 < i < s
which is surely better than the random completing.
according to the algorithm of scout bee work
The algorithm joins two conceptions of decision
4: For each i – scout calculate: AScoutsi =
making: scout bees using random searching method-
E(θ ∪ RScoutsi ,{1}) – Abase
ology realise the above given algorithm of work gen-
5: Find the BestScout rule which satisfies the
erating new solutions and working bees, realising the
following condition ABestScout = max(AScouts)
idea of local searches, set antecedents (IF-parts) and
6: REllite = ∅
consequents (THEN-parts) of the rule.
7: Form the multitude of REllite rules in
To make the algorithm less cumbersome, one can
the following way REllite ∈ RScoutsi , where i
make a decision to shorten the number of improved
satisfies the condition min (ABestScout-AScoutsi ) and
rules on the ground of their utility value, defined by 1<i<s
the incremental rate growth of the number of train- RScoutsi ∈RElliteWork,
/ 1<i<s
ing samples, as correctly classified objects. The given 8: Each i – working bee from BestScout forms the
solution is the analogue of the bee “dance” in nature. vector RBestScoutWorki . 1 < i < o using the method
After that, the best solution from the multitude of the of the local search
best solutions in each stage within the limits of the 9: For each i – working bee calculate:
given iteration is picked and this very solution will be ABestScoutWork i = E(θ ∪ RBest_Scout_Worki ,
added into rule base. {1}) – Abase;
In the current version the application field of the 10: Each i – working bee from REllitej forms
algorithm is limited by sets of data which contain the vector RElliteWorkji , 1 < j < l, 1 < i < o using the
integral-valued and real-valued markers cancelling out method of the local search
nominal markers. 11: For each j – rule from RElliteDo:
Input: s – the number of scout bees, o – the number 11.1: For each i – working bee calculate:
of working bees, l – the number of rules at best under ABestScoutWork ji = E(θ ∪ RElliteWorkji , {1}) –
investigation, z – the number of the rules generated Abase;
by the algorithm, θ – initial rule base according to the End do (j);
algorithm of the initialisation of the fuzzy classifier 12: Place rule into θ* the base corresponding
rule base. the condition max(AScout BestScout , ABestScoutWork,
Output: θ∗ – the effective base of the fuzzy classifier AElliteWork);
rules. 13: Countaddrulle := Countaddrulle + 1, if Coun-
Algorithm variables: taddrulle = z, that EXIT, in other words step 2.
RScouts – the multitude of the rules received by
scout bees;
4.4 The modified bee colony algorithm tooptimize
BestScout – the best rule according to the value of
the weight
utility from the multitude RScouts;
RBestScoutsWork – the multitude from the o rules The above mentioned algorithm is used to optimize the
received by working bees with perfecting the rule CF vector. The algorithm is modified because there is
BestScout; no single understanding of how bees are attracted to

29
the source of the nectar. Although it is known that this
correspondence has a stochastic nature, it is connected
with the quality of the source of nectar. In the modi-
fied algorithm the enlistment is performed according
to a genetic algorithm breeding method; to choose the
solution an annealing imitation method is used.
Input:The size of hive U , the rule base θ, the number
of iterations Iter, required accuracy E, initial temper-
ature T0 , cooling index α, the percentage of scouts
from a hive P_S, g parameter, the view of population
formation algorithm Alg.
Output: CF – optimal rule weighing coefficient of
the fuzzy classifier.
Algorithm variables:
Iter – the number of current iteration;
best – the best decision vector;
BS – accident decision vectors;
W – block of decision vectors of working bees;
NW – decision vectors formed on the ground of the
whole block of working bees;
Figure 2. Wine data: a) accuracy vs. number of rules;
NB – decision vectors formed on the ground of the b) accuracy vs. size of hive for training and test patterns.
vector best;
F – storage of all decision vectors;
15: Formation of scouts |BS|:=|F| – |W|, if |BS|
Fj .co – standardized estimated accuracy for j vector
more than P_S *U /100%, then |BS|:=P_S *U /100%
of the storage F.
16: If the number of iterations is exceeded or
Algorithm: 1: Iter = 1, g = 5, |BS| = P_S *U /100%,
the required accuracy E is gained, then CF =
T = T 0, W = ∅, CF = ∅
Wk k = argmax(Wi .accuracy), 1 < i < |W|, EXIT, in
2: The creation of random BS decision vectors for
other way Iter: = Iter + 1, T : = αT , move to step 2.
each of scouts
3: Calculation of classification accuracy BSj . accu-
racy = E(θ, BSj )
5 BENCHMARK AND COMPARISON OF
4: Definition of the best decision best = argmax(BSk .
RESULTS
accuracy, Wq . accuracy)
5: Run the cycle of j from 1 to |BS|
The proposed fuzzy classifier is tested with the
5.1: IF exp(−|BSj . accuracy–best. accuracy |/T ) <
benchmark classification problems: Bupa, Iris, Glass,
rand* g THEN include j decision into working bees
New thyroid, and Wine. They are publicly avail-
massive W
able on the KEEL-dataset repository web page
6: best is being included into working bees
(http://www.keel.es). Ten-fold cross-validation is
massive W
employed to examine the generalization ability of the
7: Formation of new decisions NW on the ground
proposed approach to classifying the dataset. In the
of W. Run the cycle of j from 1 to |W|:
ten-fold cross-validation procedure, the datasets are
7.1: NWj = Wj ± rand(Wj – best)*g;
separated into ten subsets of almost the same size.
8: Calculation of accuracy NWj . accuracy=
Nine subsets are used as training patterns for designing
E(θ, NWj )
the fuzzy rule-based classifier. The remaining single
9: Formation of new decisions NB on the ground of
subset is used as the test set to evaluate the con-
best. Run the cycle of j from 1 to |W|:
structed fuzzy classifier. This procedure is performed
9.1: NBj : = best ± rand (Wj – best)*g;
ten times after that roles of subsets are exchanged with
10: Calculation of accuracy NBj . accuracy =
each other so that every subset is used as the test set.
E(θ, NBj )
The computer simulation implemented the previously
11: Calculation of decisions normalized accuracy
mentioned procedure ten times and calculated the aver-
for all foragers F=NW + NB + W
age accuracy, the corresponding standard deviation,
12: Fj .co := Fj . accuracy/ i Fi . accuracy
and the average number of rules associated with the
13: Formation of W. It will include decisions
generated fuzzy rule-based classifiers.
from F,
⎧ We examine the relationship between the accuracy
⎨ > 0.8, decision j is included 3 times and the number of rules, accuracy, and size of the
Fj .co > 0.4, decision j is included 2 times hive. On Wine dataset, the performance for varying
⎩ > 0.05, decision j is included 1 time number of fuzzy rules and the size of the hives are
shown in Fig. 2, on Glass dataset – in Fig. 3.
14: Formation of a new population with the num- From the figures, we can see that the accuracy of
ber equals the hive size U , according to the given classification on learn patterns and test patterns is fur-
algorithm Alg ther improved by increasing the rules for each class

30
 Table 3. Average results of training accuracy (%) obtained.

New
Data set Bupa Iris Glass thyroid Wine

Pfc* 79.55 98.59 71.92 98.66 98.25

AM 80.38 97.26 81.48 99.17 99.69
CORE 61.93 95.48 54.26 92.66 99.06
HIDER 73.37 97.48 90.09 95.97 97.19
SGERD 59.13 97.33 53.84 90.23 91.76
TARGET 68.86 93.50 45.07 88.05 85.19

*Proposed fuzzy classifier

Table 4. Standard deviations of training accuracy (%)

obtained.

New
Data set Bupa Iris Glass thyroid Wine

Pfc* 1.72 0.42 2.17 0.61 0.71

Figure 3. Glass data: a) accuracy vs. number of rules; AM 3.25 0.74 6.59 0.58 0.58
b) accuracy vs. size of hive for training and test patterns. CORE 0.89 1.42 1.90 1.19 0.42
HIDER 2.70 0.36 1.64 0.83 0.98
generated. In addition, in other figures, we can see SGERD 0.68 0.36 2.96 0.87 1.31
TARGET 0.89 2.42 0.90 2.19 1.58
that the size of the hive equals 40 good compromises
between the accuracy and the compact spread of the *Proposed fuzzy classifier
result.
The following five algorithms are used as reference
algorithms (Alcala-Fdez et al., 2011). Table 5. Average results of test accuracy (%) obtained.
Ant-Miner (AM) – an ant colony system based on
a heuristic function based on the entropy measure for New
each attribute-value (Parpinelli et al., 2002); Data set Bupa Iris Glass thyroid Wine
CO-Evolutionary Rule Extractor (CORE) – aco-
evolutionary method which is employed as a fitness Pfc* 63.06 88.00 65.38 90.69 85.30
measure combination of the true positive rate and the AM 57.25 96.00 53.74 90.76 92.06
false positive rate (Tan, Yu, & Ang, 2006); CORE 61.97 92.67 45.74 90.76 94.87
Hierarchical Decision Rules (HIDER) – a method HIDER 65.83 96.67 64.35 90.28 82.61
which iteratively creates rules that cover randomly SGERD 57.89 96.67 48.33 88.44 87.09
TARGET 65.97 92.93 44.11 86.79 82.24
selected examples of the training set (Aguilar-Ruiz,
Riquelme, &Toro (2003)); Steady-State GeneticAlgo- *Proposed fuzzy classifier
rithm for Extracting Fuzzy Classification Rules from
Data (SGERD) – a steady-state GA which generates
a pre-specified number of rules per class following a Table 6. Standard deviations of test accuracy (%) obtained.
GCCL approach (Mansoori et al., 2008);
Tree Analysis with Randomly Generated and New
Evolved Trees (TARGET) methodology – a genetic Data set Bupa Iris Glass thyroid Wine
algorithm where each chromosome represents a com-
plete decision tree (Gray & Fan, 2008). Pfc* 10.34 4.22 6.89 3.84 5.80
AM 7.71 3.27 12.92 6.85 6.37
Tables 3–6 compare the results obtained herein with
CORE 4.77 4.67 9.36 5.00 4.79
those in five reference algorithms (Alcala-Fdez et al., HIDER 10.04 3.33 12.20 7.30 6.25
2011). SGERD 3.41 3.33 5.37 6.83 6.57
It is obvious from Tables 3–4 that the liaison of TARGET 1.41 4.33 5.37 5.83 7.57
the proposed fuzzy classifier gives results according
to the training samples, ranking the first place (iris), *Proposed fuzzy classifier.
the second (bupa, new thyroid) and the third (glass)
in the rating among the given algorithms. This points
to the good associative quality of the algorithm. The conception included in it, as is witnessed by a root-
decision depends on the training sample. However, if mean-square error value.
one takes only one sample, the algorithm gives two According to Tables 5–6, we can estimate the
or three variants of the rule base which are different prognostic properties of the solutions provided by
in terms of quality because of the random searching the algorithms. The best decision is reached by the

31
algorithm of the complicated sample – glass, the sec- Table 7. Classification results of training and test
ond place is won by bupa, the third by thyroid, wine, accuracy (%).
and easily the worst result is obtained by iris. The bad
result of the iris sample is connected with the fact that Pfc*** PSO ACO
our algorithms do not mark the important markers for
ID* KD** Training Test Training Test Training Test
objects, unlike Hider, Sgerd, but they treat all markers
as if they were of the same importance. KZ, IS
The developed algorithms are implemented for tun- 1 90.95 75.33 72.24 71.43 78.53 72.72
ing the fuzzy classifier to non-pharmacological treat- 2 89.61 70.94 69.65 60.28 79.57 74.61
ment efficiency forecasting. The forecasting is based 3 91.48 81.43 68.89 66.43 76.04 74.20
on the analysis of retrospective data before and after 4 87.01 74.29 63.89 56.71 67.68 65.31
the treatment of rehabilitated patients. The patients 5 85.59 70.00 59.50 56.11 67.74 61.80
are prescribed of five kinds of treatment. The empiric KZ. TSH. TST
base for forecasting is medical data of patients being 1 81.25 69.5 67.45 59.30 75.32 65.95
2 85.72 71.36 71.34 67.75 76.75 73.11
rehabilitated in Federal State Institution of Tomsk
3 82.5 67.16 70.15 62.13 78.17 65.51
Scientific-Research Institute of Health-Resort and 4 80 72.51 61.43 61.14 74.26 69.43
Physiotherapy Federal Medical and Biological Agency 5 80 62.57 60.82 57.33 71.38 61.48
of Russia.
Index of the functional stress of organism *ID – Input Data
FNO = Index_AG/Index_LRP, where index_AG is **KT – Kind of treatment
adaptive hormone index: ratio of glucocorticoid con- ***Pfc – Proposed fuzzy classifier
centration (GC) to insulin (IS) in blood serum;
index_LRP is index of lipid reserve for peroxidation,
is calculated on the basis of data totality.
The patient undergoes the medical test after treat- According to Table 7, the proposed fuzzy classifier
ment course and then FNO index is calculated. At that gives results better thatACO and PSO, except one case.
increase of FNO index value in dynamics gives evi- In this case, result of the ACO better in the testing
dence of strengthening degree of functional stress of selections of the second kind of treatment.
organism and decrease to give evidence of disturbed The trained fuzzy system is used to choose the most
function normalization. effective kind of treatment. For this purpose, anew
FNO_koef = FNObefore/FNOafter serves as a patient’s data is sent to the system input, and then the
prediction value.The value of this index gives evidence system gives back the foreseeable class of the patient’s
of treatment effectiveness. If FNO_koef > 1, it means state change from each kind of treatment.
that patient has improvement after passing of treatment
course (class 1), otherwise – notable improvement is
not observed (class 2). 6 CONCLUSION
Fuzzy system permits to give predictions of treat-
ment performance of new incoming patients with Transformation of experimental information in a fuzzy
one or another complex after training on the existing knowledge base can be a useful method of data
precedents (training choice). processing in medicine, banking, management, and
As input variables we selected in the first experi- other areas where decision-makers, rather than those
ment KZ – glucocorticoid, IS – insulin; in the second working in strict quantitative relationships, prefer
one: KZ – glucocorticoid, TSH – thyroid-stimulating transparent and easily interpreted verbal rules.
hormone, TST – testosterone. A modified version of the ABC algorithm for fuzzy
If to be used for training all having sample, it can rule-based classifiers’ design has been introduced.
come out with the problem of the so called ‘over fit- Four components of the algorithm are observed: ini-
ting’. In other words, classifiers will be tuned only on tialization, work of scout bee, rule antecedent gener-
this sample and will not necessary be effective work- alisation, and scales setting.
ing with other data. To overcome this problem and to No Free Lunch Theorem shows that in the absence
obtain a fuzzy system of high generalized ability we of assumptions we should not prefer any classifica-
used in our work the method of cross-validation. For tion algorithm over another (Wolpert & Macready,
each: from five complexes, each table of observations 1997). The given comparisons of the developed algo-
is divided into training and the testing selections in the rithm with its analogues demonstrate that the proposed
ratio 80:20. algorithms are feasible and effective in solving com-
Before an application, our proposed fuzzy clas- plex classification problems.
sifier uses two different methods: a Particle Swarm The algorithms are checked against the real data
Optimization algorithm (PSO) and an Ant Colony with the implementation of the fuzzy classifier to
Optimization algorithm (ACO) with Single tone forecast the efficiency of non-medical treatment.
approximation. The offered algorithms can be implemented to
The average results from 25 experiments with the solve questions in the area of data mining.
same input parameters are compared with algorithms This project is done with financial support from the
that were used before and are given in Table 7. Russian Foundation for Basic Research (project No.
32
Other documents randomly have
different content
 CHAPTER IV
HUNTING IN THE CATTLE COUNTRY; THE
PRONGBUCK

The prongbuck is the most characteristic and distinctive of

American game animals. Zoologically speaking, its position is
unique. It is the only hollow-horned ruminant which sheds its horns,
or rather the horn sheaths. We speak of it as an antelope, and it does
of course represent on our prairies the antelopes of the Old World;
but it stands apart from all other horned animals. Its place in the
natural world is almost as lonely as that of the giraffe. In all its ways
and habits it differs as much from deer and elk as from goat and
sheep. Now that the buffalo has gone, it is the only game really at
home on the wide plains. It is a striking-looking little creature, with
its prominent eyes, single-pronged horns, and the sharply contrasted
white, brown and reddish of its coat. The brittle hair is stiff, coarse
and springy; on the rump it is brilliantly white, and is erected when
the animal is alarmed or excited, so as to be very conspicuous. In
marked contrast to deer, antelope never seek to elude observation;
all they care for is to be able themselves to see. As they have good
noses and wonderful eyes, and as they live by preference where there
is little or no cover, shots at them are usually obtained at far longer
range than is the case with other game; and yet, as they are easily
seen, and often stand looking at the hunter just barely within very
long rifle-range, they are always tempting their pursuer to the
expenditure of cartridges. More shots are wasted at antelope than at
any other game. They would be even harder to secure were it not that
they are subject to fits of panic folly, or excessive curiosity, which
occasionally put them fairly at the mercy of the rifle-bearing hunter.
In the old days the prongbuck was found as soon as the westward-
moving traveller left the green bottom-lands of the Mississippi, and
from thence across to the dry, open valleys of California, and
northward to Canada and southward into Mexico. It has everywhere
been gradually thinned out, and has vanished altogether from what
were formerly the extreme easterly and westerly limits of its range.
The rates of extermination of the different kinds of big game have
been very unequal in different localities. Each kind of big game has
had its own peculiar habitat in which it throve best, and each has also
been found more or less plentifully in other regions where the
circumstances were less favorable; and in these comparatively
unfavorable regions it early tends to disappear before the advance of
man. In consequence, where the ranges of the different game
animals overlap and are intertwined, one will disappear first in one
locality, and another will disappear first where the conditions are
different. Thus the whitetail deer had thrust forward along the very
narrow river bottoms into the domain of the mule-deer and the
prongbuck among the foothills of the Rocky Mountains, and in these
places it was exterminated from the narrow strips which it inhabited
long before the mule-deer vanished from the high hills, or the
prongbuck from the great open plains. But along great portions of
the Missouri there are plenty of whitetails yet left in the river
bottoms, while the mule-deer that once dwelt in the broken hills
behind them, and the prongbuck which lived on the prairie just back
of these bluffs, have both disappeared. In the same way the mule-
deer and the prongbuck are often found almost intermingled through
large regions in which plains, hills, and mountains alternate. If such
a region is mainly mountainous, but contains a few valleys and table-
lands, the prongbuck is sure to vanish from the latter before the
mule-deer vanishes from the broken country. But if the region is one
primarily of plains, with here and there rows of rocky hills in which
the mule-deer is found, the latter is killed off long before the
prongbuck can be hunted out of the great open stretches. The same is
true of the pronghorn and the wapiti. The size and value of the wapiti
make it an object of eager persecution on the part of hunters. But as
it can live in the forest-clad fastnesses of the Rockies, into which
settlement does not go, it outlasts over great regions the pronghorn,
whose abode is easily penetrated by sheep and cattle men. Under
anything like even conditions, however, the prongbuck, of course,
outlasts the wapiti. This was the case on the Little Missouri. On that
stream the bighorn also outlasted the wapiti. In 1881 wapiti were still
much more plentiful than bighorns. Within the next decade they had
almost totally disappeared, while the bighorn was still to be found; I
shot one and saw others in 1893, at which time I had not authentic
information of a single wapiti remaining anywhere on the river in my
neighborhood, although it is possible that one or two still lurked in
some out-of-the-way recess. In Colorado at one time the bighorn was
nearly exterminated, while the wapiti still withstood the havoc made
among its huge herds; then followed a period in which the rapidity of
destruction of the wapiti increased far beyond that of the bighorn.
I mention these facts partly because they are of interest in
themselves, but chiefly because they tend to explain the widely
different opinions expressed by competent observers about what
superficially seem to be similar facts. It cannot be too often repeated
that allowance must be made for the individual variability in the
traits and characters of animals of the same species, and especially of
the same species under different circumstances and in different
localities; and allowance must also be made for the variability of the
individual factor in the observers themselves. Many seemingly
contradictory observations of the habits of deer, wapiti, and
prongbuck will be found in books by the best hunters. Take such
questions as the keenness of sight of the deer as compared with the
prongbuck, and of the pugnacity of the wapiti, both actual and
relative, and a wide difference of opinion will be found in three such
standard works as Dodge’s “The Hunting-grounds of the Great
West,” Caton’s “Deer and Antelope of America,” and the
contributions of Mr. Grinnell to the “Century Book of Sports.”
Sometimes the difference will be in mere matters of opinion, as, for
instance, in the belief as to the relative worth of the sport furnished
by the chase of the different creatures; but sometimes there is a
direct conflict of fact. Colonel Dodge, for instance, has put it upon
record that the wapiti is an exceedingly gentle animal, less dangerous
than a whitetail or blacktail buck in a close encounter, and that the
bulls hardly ever fight among themselves. My own experience leads
me to traverse in the most emphatic manner every one of these
conclusions, and all hunters whom I have met feel exactly as I do; yet
no one would question for a moment Colonel Dodge’s general
competency as an observer. In the same way Mr. Grinnell has a high
opinion of the deer’s keenness of sight. Judge Caton absolutely
disagrees with him, and my own experience tends to agree with that
of the Judge—at least to the extent of placing the deer’s vision far
below that of the prongbuck and even that of the bighorn, and only
on a par with that of the wapiti. Yet Mr. Grinnell is an unusually
competent observer, whose opinion on any such subject is entitled to
unqualified respect.
Difference in habits may be due simply to difference of locality, or
to the need of adaptation to new conditions. The prongbuck’s habits
about migration offer examples of the former kind of difference. Over
portions of its range the prongbuck is not migratory at all. In other
parts the migrations are purely local. In yet other regions the
migrations are continued for great distances, immense multitudes of
the animals going to and fro in the spring and fall along well-beaten
tracks. I know of one place in New Mexico where the pronghorn
herds are tenants of certain great plains throughout the entire year. I
know another region in northwestern Colorado where the very few
prongbucks still left, though they shift from valley to valley, yet
spend the whole year in the same stretch of rolling, barren country.
On the Little Missouri, however, during the eighties and early
nineties, there was a very distinct though usually local migration.
Before the Black Hills had been settled they were famous wintering
places for the antelope, which swarmed from great distances to them
when cold weather approached; those which had summered east of
the Big Missouri actually swam the river in great herds, on their
journey to the Hills. The old hunters around my ranch insisted that
formerly the prongbuck had for the most part travelled from the
Little Missouri Bad Lands into the Black Hills for the winter.
 ON THE LITTLE MISSOURI

When I was ranching on that river, however, this custom no longer

obtained, for the Black Hills were too well settled, and the herds of
prongbuck that wintered there were steadily diminishing in
numbers. At that time, from 1883 to 1896, the seasonal change in
habits, and shift of position, of the prongbucks were well marked. As
soon as the new grass sprang they appeared in great numbers upon
the plains. They were especially fond of the green, tender blades that
came up where the country had been burned over. If the region had
been devastated by prairie fires in the fall, the next spring it was
certain to contain hundreds and thousands of prongbucks. All
through the summer they remained out on these great open plains,
coming to drink at the little pools in the creek beds, and living where
there was no shelter of any kind. As winter approached they began to
gather in bands. Some of these bands apparently had regular
wintering places to the south of us, in Pretty Buttes and beyond; and
close to my ranch, at the crossing of the creek called Beaver, there
were certain trails which these antelope regularly travelled,
northward in the spring and southward in the fall. But other bands
would seek out places in the Bad Lands near by, gathering together
on some succession of plateaus which were protected by neighboring
hills from the deep drifts of snow. Here they passed the winter, on
short commons, it is true (they graze, not browsing like deer), but
without danger of perishing in the snowdrifts. On the other hand, if
the skin-hunters discovered such a wintering place, they were able to
butcher practically the entire band, if they so desired, as the
prongbucks were always most reluctant to leave such a chosen
ground.
Normally the prongbuck avoids both broken ground and timber. It
is a queer animal, with keen senses, but with streaks of utter folly in
its character. Time and again I have known bands rush right by me,
when I happened to surprise them feeding near timber or hills, and
got between them and the open plains. The animals could have
escaped without the least difficulty if they had been willing to go into
the broken country, or through even a few rods of trees and brush;
and yet they preferred to rush madly by me at close range, in order to
get out to their favorite haunts. But nowadays there are certain
localities where the prongbucks spend a large part of their time in
the timber or in rough, hilly country, feeding and bringing up their
young in such localities.
Typically, however, the prongbuck is preeminently a beast of the
great open plains, eating their harsh, dry pasturage, and trusting to
its own keen senses and speed for its safety. All the deer are fond of
skulking; the whitetail preeminently so. The prongbuck, on the
contrary, never endeavors to elude observation. Its sole aim is to be
able to see its enemies, and it cares nothing whatever about its
enemies seeing it. Its coloring is very conspicuous, and is rendered
still more so by its habit of erecting the white hair on its rump. It has
a very erect carriage, and when it thinks itself in danger it always
endeavors to get on some crest or low hill from which it can look all
about. The big bulging eyes, situated at the base of the horns, scan
the horizon far and near like twin telescopes. They pick out an object
at such a distance that it would entirely escape the notice of a deer.
When suspicious, they have a habit of barking, uttering a sound
something like “kau,” and repeating it again and again, as they walk
up and down, endeavoring to find out if danger lurks in the unusual
object. They are extremely curious, and in the old days it was often
possible to lure them toward the hunter by waving a red
handkerchief to and fro on a stick, or even by lying on one’s back and
kicking the legs. Nowadays, however, there are very few localities
indeed in which they are sufficiently unsophisticated to make it
worth while trying these time-honored tricks of the long-vanished
trappers and hunters.
 Along the Little Missouri the fawns, sometimes one and sometimes
two in number, were dropped in May or early in June. At that time
the antelope were usually found in herds which the mother did not
leave until she was about to give birth to the fawn. During the first
few days the fawn’s safety is to be found only in its not attracting
attention. During this time it normally lies perfectly flat on the
ground, with its head outstretched, and makes no effort to escape.
While out on the spring round-up I have come across many of these
fawns. Once, in company with several cowboys, I was riding behind a
bunch of cattle which, as we hurried them, spread out in open order
ahead of us. Happening to cast down my eyes I saw an antelope fawn
directly ahead of me. The bunch of cattle had passed all around it,
but it made not the slightest sign, not even when I halted, got off my
pony, and took it up in my arms. It was useless to take it to camp and
try to rear it, and so I speedily put it down again. Scanning the
neighborhood, I saw the doe hanging about some half a mile off, and
when I looked back from the next divide I could see her gradually
drawing near to the fawn.
If taken when very young, antelope make cunning and amusing
pets, and I have often seen them around the ranches. There was one
in the ranch of a Mrs. Blank who had a station on the Deadwood
stage line some eighteen years ago. She was a great worker in
buckskin, and I got her to make me the buckskin shirt I still use.
There was an antelope fawn that lived at the house, wandering
wherever it wished; but it would not permit me to touch it. As I sat
inside the house it would come in and hop up on a chair, looking at
me sharply all the while. No matter how cautiously I approached, I
could never put my hand upon it, as at the last moment it would
spring off literally as quick as a bird would fly. One of my neighbors
on the Little Missouri, Mr. Howard Eaton, had at one time upon his
ranch three little antelope whose foster-mother was a sheep, and
who were really absurdly tame. I was fond of patting them and of
giving them crusts, and the result was that they followed me about so
closely that I had to be always on the lookout to see that I did not
injure them. They were on excellent terms with the dogs, and were
very playful. It was a comic sight to see them skipping and hopping
about the old ewe when anything happened to alarm her and she
started off at a clumsy waddle. Nothing could surpass the tameness
of the antelope that are now under Mr. Hornaday’s care at the Bronx
Zoological Garden in New York. The last time that I visited the
garden some repairs were being made inside the antelope enclosure,
and a dozen workmen had gone in to make them. The antelope
regarded the workmen with a friendliness and curiosity untempered
by the slightest touch of apprehension. When the men took off their
coats the little creatures would nose them over to see if they
contained anything edible, and they would come close up and watch
the men plying the pick with the utmost interest. Mr. Hornaday took
us inside, and they all came up in the most friendly manner. One or
two of the bucks would put their heads against our legs and try to
push us around, but not roughly. Mr. Hornaday told me that he was
having great difficulty, exactly as with the mule-deer, in
acclimatizing the antelope, especially as the food was so different
from what they were accustomed to in their native haunts.
The wild fawns are able to run well a few days after they are born.
They then accompany the mother everywhere. Sometimes she joins a
band of others; more often she stays alone with her fawn, and
perhaps one of the young of the previous year, until the rut begins.
Of all game the prongbuck seems to me the most excitable during the
rut. The males run the does much as do the bucks of the mule and
whitetail deer. If there are no does present, I have sometimes
watched a buck run to and fro by himself. The first time I saw this I
was greatly interested, and could form no idea of what the buck was
doing. He was by a creek bed in a slight depression or shallow valley,
and was grazing uneasily. After a little while he suddenly started and
ran just as hard as he could, off in a straight direction, nearly away
from me. I thought that somehow or other he had discovered my
presence; but he suddenly wheeled and came back to the original
place, still running at his utmost speed. Then he halted, moved about
with the white hairs on his rump outspread, and again dashed off at
full speed, halted, wheeled, and came back. Two or three times he did
this, and let me get very close to him before he discovered me. I was
too much interested in what he was doing to desire to shoot him.
In September, sometimes not earlier than October, the big bucks
begin to gather the does into harems. Each buck is then constantly
on the watch to protect his harem from outsiders, and steal another
doe if he can get a chance. I have seen a comparatively young buck
who had appropriated a doe, hustle her hastily out of the country as
soon as he saw another antelope in the neighborhood; while, on the
other hand, a big buck, already with a good herd of does, will do his
best to appropriate any other that comes in sight. The bucks fight
fearlessly but harmlessly among themselves, locking their horns and
then pushing as hard as they can.
Although their horns are not very formidable weapons, they are
bold little creatures, and if given a chance will stand at bay before
either hound or coyote. A doe will fight most gallantly for her fawn,
and is an overmatch for a single coyote, but of course she can do but
little against a large wolf. The wolves are occasionally very
destructive to the herds. The cougar, however, which is a much worse
foe than the wolf to deer and mountain sheep, can but rarely molest
the prongbuck, owing to the nature of the latter’s haunts. Eagles, on
occasion, take the fawns, as they do those of deer.
I have always been fond of the chase of the prongbuck. While I
lived on my ranch on the Little Missouri it was, next to the mule-
deer, the game which I most often followed, and on the long wagon
strips which I occasionally took from my ranch to the Black Hills, to
the Big Horn Mountains, or into eastern Montana, prongbuck
venison was our usual fresh meat, save when we could kill prairie-
chickens and ducks with our rifles, which was not always feasible. In
my mind the prongbuck is always associated with the open prairies
during the spring, summer, or early fall. It has happened that I have
generally pursued the bighorn in bitter weather; and when we laid in
our stock of winter meat, mule-deer was our usual game. Though I
have shot prongbuck in winter, I never liked to do so, as I felt the
animals were then having a sufficiently hard struggle for existence
anyhow. But in the spring the meat of the prongbuck was better than
that of any other game, and, moreover, there was not the least
danger of mistaking the sexes, and killing a doe accidentally, and
accordingly I rarely killed anything but pronghorns at that season. In
those days we never got any fresh meat, whether on the ranch or
while on the round-up or on a wagon trip, unless we shot it, and salt
pork became a most monotonous diet after a time.
Occasionally I killed the prongbuck in a day’s hunt from my ranch.
If I started with the intention of prongbuck hunting, I always went on
horseback; but twice I killed them on foot when I happened to run
across them by accident while looking for mule-deer. I shall always
remember one of these occasions. I was alone in the Elkhorn ranch-
house at the time, my foreman and the only cow-puncher who was
not on the round-up having driven to Medora, some forty miles
away, in order to bring down the foreman’s wife and sister, who were
going to spend the summer with him. It was the fourth day of his
absence. I expected him in the evening and wanted to have fresh
meat, and so after dinner I shouldered my rifle and strolled off
through the hills. It was too early in the day to expect to see
anything, and my intention was simply to walk out until I was five or
six miles from the ranch, and then work carefully home through a
likely country toward sunset, as by this arrangement I would be in a
good game region at the very time that the animals were likely to stir
abroad. It was a glaring, late-spring day, and in the hot sun of mid-
afternoon I had no idea that anything would be moving, and was not
keeping a very sharp lookout. After an hour or two’s steady tramping
I came into a long, narrow valley, bare of trees and brushwood, and
strolled along it, following a cattle trail that led up the middle. The
hills rose steeply into a ridge crest on each side, sheer clay shoulders
breaking the mat of buffalo-grass which elsewhere covered the sides
of the valley as well as the bottom. It was very hot and still, and I was
paying but little attention to my surroundings, when my eye caught a
sudden movement on the ridge crest to my right, and, dropping on
one knee as I wheeled around, I saw the head and neck of a
prongbuck rising above the crest. The animal was not above a
hundred yards off, and stood motionless as it stared at me. At the
crack of the rifle the head disappeared; but as I sprang clear of the
smoke I saw a cloud of dust rise on the other side of the ridge crest,
and felt convinced that the quarry had fallen. I was right. On
climbing the ridge crest I found that on the other side it sank
abruptly in a low cliff of clay, and at the foot of this, thirty feet under
me, the prongbuck lay with its neck broken. After dressing it I
shouldered the body entire, thinking that I should like to impress the
new-comers by the sight of so tangible a proof of my hunting prowess
as whole prongbuck hanging up in the cottonwoods by the house. As
it was a well-grown buck the walk home under the hot sun was one of
genuine toil.
The spot where I ran across this prongbuck was miles away from
the nearest plains, and it was very unusual to see one in such rough
country. In fact, the occurrence was wholly exceptional; just as I once
saw three bighorn rams, which usually keep to the roughest country,
deliberately crossing the river bottom below my ranch, and going for
half a mile through the thick cottonwood timber. Occasionally,
however, parties of prongbuck came down the creek bottoms to the
river. Once I struck a couple of young bucks in the bottom of a creek
which led to the Chimney Butte ranch-house, and stalked them
without difficulty; for as prongbuck make no effort to hide, if there is
good cover even their sharp eyes do not avail them. On another
occasion several does and fawns, which we did not molest, spent
some time on what we called “the corral bottom,” which was two or
three miles above the ranch-house. In the middle of this bottom we
had built a corral for better convenience in branding the calves when
the round-up came near our ranch—as the bottom on which the
ranch-house stood was so thickly wooded as to make it difficult to
work cattle thereon. The does and fawns hung around the corral
bottom for some little time, and showed themselves very curious and
by no means shy.
When I went from the ranch for a day’s prongbuck hunting of set
purpose, I always rode a stout horse and started by dawn. The
prongbucks are almost the only game that can be hunted as well
during the heat of the day as at any other time. They occasionally lie
down for two or three hours about noon in some hollow where they
cannot be seen, but usually there is no place where they are sure they
can escape observation even when resting; and when this is the case
they choose a somewhat conspicuous station and trust to their own
powers of observation, exactly as they do when feeding. There is
therefore no necessity, as with deer, of trying to strike them at dawn
or dusk. The reason why I left the ranch before sunrise and often
came back long after dark was because I had to ride at least a dozen
miles to get out to the ground and a dozen to get back, and if after
industrious walking I failed at first to find my game, I would often
take the horse again and ride for an hour or two to get into new
country. Prongbuck water once a day, often travelling great distances
to or from some little pool or spring. Of course, if possible, I liked to
leave the horse by such a pool or spring. On the great plains to which
I used to make these excursions there was plenty of water in early
spring, and it would often run, here and there, in the upper courses
of some of the creeks—which, however, usually contained running
water only when there had been a cloudburst or freshet. As the
season wore on the country became drier and drier. Water would
remain only in an occasional deep hole, and few springs were left in
which there was so much as a trickle. In a strange country I could not
tell where these water-holes were, but in the neighborhood of the
ranch I of course knew where I was likely to find them. Often,
however, I was disappointed; and more than once after travelling
many miles to where I hoped to find water, there would be nothing
but sun-cracked mud, and the horse and I would have eighteen hours
of thirst in consequence. A ranch horse, however, is accustomed to
such incidents, and of course when a man spends half the time
riding, it is merely a matter of slight inconvenience to go so long
without a drink.
Nevertheless, if I did reach a spring, it turned the expedition into
pleasure instead of toil. Even in the hot weather the ride toward the
plains over the hills was very lovely. It was beautiful to see the red
dawn quicken from the first glimmering gray in the east, and then to
watch the crimson bars glint on the tops of the fantastically shaped
barren hills when the sun flamed, burning and splendid, above the
horizon. In the early morning the level beams threw into sharp relief
the strangely carved and channelled cliff walls of the buttes. There
was rarely a cloud to dim the serene blue of the sky. By the time the
heat had grown heavy I had usually reached the spring or pool,
where I unsaddled the horse, watered him, and picketed him out to
graze. Then, under the hot sun, I would stride off for the hunting
proper. On such occasions I never went to where the prairie was
absolutely flat. There were always gently rolling stretches broken by
shallow watercourses, slight divides, and even low mounds,
sometimes topped with strangely shaped masses of red scoria or with
petrified trees. My object, of course, was, either with my unaided
eyes or with the help of my glasses, to catch sight of the prongbucks
before they saw me. I speedily found, by the way, that if they were
too plentiful this was almost impossible. The more abundant deer are
in a given locality the more apt one is to run across them, and of
course if the country is sufficiently broken, the same is true of
prongbucks; but where it is very flat and there are many different
bands in sight at the same time, it is practically impossible to keep
out of sight of all of them, and as they are also all in sight of one
another, if one flees the others are certain to take the alarm. Under
such circumstances I have usually found that the only pronghorns I
got were obtained by accident, so to speak; that is, by some of them
unexpectedly running my way, or by my happening to come across
them in some nook where I could not see them, or they me.
Prongbucks are very fast runners indeed, even faster than deer.
They vary greatly in speed, however, precisely as is the case with
deer; in fact, I think that the average hunter makes altogether too
little account of this individual variation among different animals of
the same kind. Under the same conditions different deer and
antelope vary in speed and wariness, exactly as bears and cougars
vary in cunning and ferocity. When in perfect condition a full-grown
buck antelope, from its strength and size, is faster and more
enduring than an old doe; but a fat buck, before the rut has begun,
will often be pulled down by a couple of good greyhounds much more
speedily than a flying yearling or two-year-old doe. Under favorable
circumstances, when the antelope was jumped near by, I have seen
one overhauled and seized by a first-class greyhound; and, on the
other hand, I have more than once seen a pronghorn run away from
a whole pack of just as good dogs. With a fair start, and on good
ground, a thoroughbred horse, even though handicapped by the
weight of a rider, will run down an antelope; but this is a feat which
should rarely be attempted, because such a race, even when carried
to a successful issue, is productive of the utmost distress to the steed.
Ordinary horses will sometimes run down an antelope which is
slower than the average. I once had on my ranch an under-sized old
Indian pony named White Eye, which, when it was fairly roused,
showed a remarkable turn of speed, and had great endurance. One
morning on the round-up, when for some reason we did not work the
cattle, I actually ran down an antelope in fair chase on this old pony.
It was a nursing doe, and I came over the crest of the hill, between
forty and fifty yards away from it. As it wheeled to start back, the old
cayuse pricked up his ears with great interest, and the moment I gave
him a sign was after it like a shot. Whether, being a cow-pony, he
started to run it just as if it were a calf or a yearling trying to break
out of the herd, or whether he was overcome by dim reminiscences of
buffalo-hunting in his Indian youth, I know not. At any rate, after the
doe he went, and in a minute or two I found I was drawing up to her.
I had a revolver, but of course did not wish to kill her, and so got my
rope ready to try to take her alive. She ran frantically, but the old
pony, bending level to the ground, kept up his racing lope and closed
right in beside her. As I came up she fairly bleated. An expert with
the rope would have captured her with the utmost ease; but I missed,
sending the coil across her shoulders. She again gave an agonized
bleat, or bark, and wheeled around like a shot. The cow-pony
stopped almost, but not quite, as fast, and she got a slight start, and
it was some little time before I overhauled her again. When I did I
repeated the performance, and this time when she wheeled she
succeeded in getting on some ground where I could not follow, and I
was thrown out.
Normally, a horseman without greyhounds can hope for nothing
more than to get within fair shooting range; and this only by taking
advantage of the prongbucks’ peculiarity of running straight ahead in
the direction in which they are pointed, when once they have settled
into their pace. Usually, as soon as they see a hunter they run
straight away from him; but sometimes they make their flight at an
angle, and as they do not like to change their course when once
started, it is thus possible, with a good horse, to cut them off from
the point toward which they are headed, and get a reasonably close
shot.
I have done a good deal of coursing with greyhounds at one time or
another, but always with scratch packs. There are a few ranchmen
who keep leashes of greyhounds of pure blood, bred and trained to
antelope coursing, and who do their coursing scientifically, carrying
the dogs out to the hunting-grounds in wagons and exercising every
care in the sport; but these men are rare. The average man who
dwells where antelope are sufficiently abundant to make coursing a
success, simply follows the pursuit at odd moments, with whatever
long-legged dogs he and his neighbors happen to have; and his
methods of coursing are apt to be as rough as his outfit. My own
coursing was precisely of this character. At different times I had on
my ranch one or two high-classed greyhounds and Scotch
deerhounds, with which we coursed deer and antelope, as well as
jack-rabbits, foxes, and coyotes; and we usually had with them one or
two ordinary hounds, and various half-bred dogs. I must add,
however, that some of the latter were very good. I can recall in
particular one fawn-colored beast, a cross between a greyhound and
a foxhound, which ran nearly as fast as the former, though it
occasionally yelped in shrill tones. It could also trail well, and was
thoroughly game; on one occasion it ran down and killed a coyote
single-handed.
On going out with these dogs, I rarely chose a day when I was
actually in need of fresh meat. If this was the case, I usually went
alone with the rifle; but if one or two other men were at the ranch,
and we wanted a morning’s fun, we would often summon the dogs,
mount our horses, and go trooping out to the antelope-ground. As
there was good deer-country between the ranch bottom and the
plains where we found the prongbuck, it not infrequently happened
that we had a chase after blacktail or whitetail on the way. Moreover,
when we got out to the ground, before sighting antelope, it frequently
happened that the dogs would jump a jack-rabbit or a fox, and away
the whole set would go after it, streaking through the short grass,
sometimes catching their prey in a few hundred yards, and
sometimes having to run a mile or so. In consequence, by the time
we reached the regular hunting-ground the dogs were apt to have
lost a good deal of their freshness. We would get them in behind the
horses and creep cautiously along, trying to find some solitary
prongbuck in a suitable place, where we could bring up the dogs
from behind a hillock and give them a fair start. Usually we failed to
get the dogs near enough for a good start; and in most cases their
chases after unwounded prongbuck resulted in the quarry running
clean away from them. Thus the odds were greatly against them; but,
on the other hand, we helped them wherever possible with the rifle.
We usually rode well scattered out, and if one of us put up an
antelope, or had a chance at one when driven by the dogs, he always
fired, and the pack were saved from the ill effects of total
discouragement by so often getting these wounded beasts. It was
astonishing to see how fast an antelope with a broken leg could run.
If such a beast had a good start, and especially if the dogs were tired,
it would often lead them a hard chase, and the dogs would be utterly
exhausted after it had been killed; so that we would have to let them
lie where they were for a long time before trying to lead them down
to some stream-bed. If possible, we carried water for them in
canteens.
There were red-letter days, however, on which our dogs fairly ran
down and killed unwounded antelope—days when the weather was
cool, and when it happened that we got our dogs out to the ground
without their being tired by previous runs, and found our quarry
soon, and in favorable places for slipping the hounds. I remember
one such chase in particular. We had at the time a mixed pack, in
which there was only one dog of my own, the others being
contributed from various sources. It included two greyhounds, a
rough-coated deerhound, a foxhound, and the fawn-colored cross-
bred mentioned above.
We rode out in the early morning, the dogs trotting behind us;
and, coming to a low tract of rolling hills, just at the edge of the great
prairie, we separated and rode over the crest of the nearest ridge.
Just as we topped it, a fine buck leaped up from a hollow a hundred
yards off, and turned to look at us for a moment. All the dogs were
instantly spinning toward him down the grassy slope. He apparently
saw those at the right, and, turning, raced away from us in a diagonal
line, so that the left-hand greyhound, which ran cunning and tried to
cut him off, was very soon almost alongside. He saw her, however—
she was a very fast bitch—just in time, and, wheeling, altered his
course to the right. As he reached the edge of the prairie, this
alteration nearly brought him in contact with the cross-bred, which
had obtained a rather poor start, on the extreme right of the line.
Around went the buck again, evidently panic-struck and puzzled to
the last degree, and started straight off across the prairie, the dogs
literally at his heels, and we, urging our horses with whip and spur,
but a couple of hundred yards behind. For half a mile the pace was
tremendous, when one of the greyhounds made a spring at his ear,
but failing to make good his hold, was thrown off. However, it halted
the buck for a moment, and made him turn quarter round, and in a
second the deerhound had seized him by the flank and thrown him,
and all the dogs piled on top, never allowing him to rise.
Later we again put up a buck not far off. At first it went slowly, and
the dogs hauled up on it; but when they got pretty close, it seemed to
see them, and letting itself out, went clean away from them almost
without effort.
Once or twice we came upon bands of antelope, and the hounds
would immediately take after them. I was always rather sorry for
this, however, because the frightened animals, as is generally the
case when beasts are in a herd, seemed to impede one another, and
the chase usually ended by the dogs seizing a doe, for it was of course
impossible to direct them to any particular beast.
It will be seen that with us coursing was a homely sport.
Nevertheless we had good fun, and I shall always have enjoyable
memories of the rapid gallops across the prairie, on the trail of a
flying prongbuck.
Usually my pronghorn hunting has been done while I have been
off with a wagon on a trip intended primarily for the chase, or else
while travelling for some other purpose.

CAMPING ON THE ANTELOPE GROUNDS

All life in the wilderness is so pleasant that the temptation is to

consider each particular variety, while one is enjoying it, as better
than any other. A canoe trip through the great forests, a trip with a
pack-train among the mountains, a trip on snowshoes through the
silent, mysterious fairyland of the woods in winter—each has its
peculiar charm. To some men the sunny monotony of the great
plains is wearisome; personally there are few things I have enjoyed
more than journeying over them where the game was at all plentiful.
Sometimes I have gone off for three or four days alone on horseback,
with a slicker or oilskin coat behind the saddle, and some salt and
hardtack as my sole provisions. But for comfort on a trip of any
length it was always desirable to have a wagon. My regular outfit
consisted of a wagon and team driven by one man who cooked,
together with another man and four riding ponies, two of which we
rode, while the other two were driven loose or led behind the wagon.
While it is eminently desirable that a hunter should be able to rough
it, and should be entirely willing to put up with the bare minimum of
necessities, and to undergo great fatigue and hardship, it is yet not at
all necessary that he should refrain from comfort of a wholesome
sort when it is obtainable. By taking the wagon we could carry a tent
to put up if there was foul weather. I had a change of clothes to put
on if I was wet, two or three books to read—and nothing adds more
to the enjoyment of a hunting trip—as well as plenty of food; while
having two men made me entirely foot-loose as regards camp, so that
I could hunt whenever I pleased, and, if I came in tired, I simply
rested, instead of spending two or three hours in pitching camp,
cooking, tethering horses, and doing the innumerable other little
things which in the aggregate amount to so much.
On such a trip, when we got into unknown country, it was of
course very necessary to stay near the wagon, especially if we had to
hunt for water. But if we knew the country at all, we would decide in
the morning about where the camp was to be made in the afternoon,
and then I would lope off on my own account, while the wagon
lumbered slowly across the rough prairie sward straight toward its
destination. Sometimes I took the spare man with me, and
sometimes not. It was convenient to have him, for there are
continually small emergencies in which it is well to be with a
companion. For instance, if one jumps off for a sudden shot, there is
always a slight possibility that any but a thoroughly trained horse
will get frightened and gallop away. On some of my horses I could
absolutely depend, but there were others, and very good ones too,
which would on rare occasions fail me; and few things are more
disheartening than a long stern chase after one’s steed under such
circumstances, with the unpleasant possibility of seeing him leave
the country entirely and strike out for the ranch fifty or sixty miles
distant. If there is a companion with one, all danger of this is over.
Moreover, in galloping at full speed after the game it is impossible
now and then to avoid a tumble, as the horse may put his leg into a
prairie-dog hole or badger burrow; and on such occasions a
companion may come in very handily. On the other hand, there is so
great a charm in absolute solitude, in the wild, lonely freedom of the
great plains, that often I would make some excuse and go off entirely
by myself.
Such rides had a fascination of their own. Hour after hour the wiry
pony shuffled onward across the sea of short, matted grass. On every
side the plains stretched seemingly limitless. Sometimes there would
be no object to break the horizon; sometimes across a score of miles
there would loom through the clear air the fantastic outlines of a
chain of buttes, rising grim and barren. Occasionally there might be a
slightly marked watercourse, every drop of moisture long dried; and
usually there would not be as much as the smallest sage-brush
anywhere in sight. As the sun rose higher and higher the shadows of
horse and rider shortened, and the beams were reflected from the
short, bleached blades until in the hot air all the landscape afar off
seemed to dance and waver. Often on such trips days went by
without our coming across another human being, and the loneliness
and vastness of the country seemed as unbroken as if the old
vanished days had returned—the days of the wild wilderness
wanderers, and the teeming myriads of game they followed, and the
scarcely wilder savages against whom they warred.
Now and then prongbuck would appear, singly or in bands; and
their sharp bark of alarm or curiosity would come to me through the
still, hot air over great distances, as they stood with head erect
looking at me, the white patches on their rumps shining in the sun,
and the bands and markings on their heads and necks showing as if
they were in livery. Scan the country as carefully as I would, they
were far more apt to see me than I was them, and once they had seen
me, it was normally hopeless to expect to get them. But their strange
freakishness of nature frequently offsets the keenness of their senses.
At least half of the prongbucks which I shot were obtained, not by
stalking, but by coming across them purely through their own fault.
Though the prairie seemed level, there was really a constant series of
undulations, shallow and of varying width. Now and then as I topped
some slight rise I would catch a glimpse of a little band of
pronghorns feeding, and would slip off my horse before they could
see me. A hasty determination as to where the best chance of
approaching them lay would be followed by a half-hour’s laborious
crawl, a good part of the time flat on my face. They might discover
me when I was still too far for a shot; or by taking advantage of every
little inequality I might get within long range before they got a
glimpse of me, and then in a reasonable proportion of cases I would
bag my buck. At other times the buck would come to me. Perhaps
one would suddenly appear over a divide himself, and his curiosity
would cause him to stand motionless long enough to give me a shot;
while on other occasions I have known one which was out of range to
linger around, shifting his position as I shifted mine, until by some
sudden gallop or twist I was able to get close enough to empty my
magazine at him.
When the shadows had lengthened, but before any coolness had
come into the air, I would head for the appointed camping-place.
Sometimes this would be on the brink of some desolate little pool
under a low, treeless butte, or out on the open prairie where the only
wood was what we had brought with us. At other times I would find
the wagon drawn up on the edge of some shrunken plains river,
under a line of great cottonwoods with splintered branches and
glossy leaves that rustled all day long. Such a camp was always
comfortable, for there was an abundance of wood for the fire, plenty
of water, and thick feed in which the horses grazed—one or two being
picketed and the others feeding loose until night came on. If I had
killed a prongbuck, steaks were speedily sizzling in the frying-pan
over the hot coals. If I had failed to get anything, I would often walk a
mile or two down or up the river to see if I could not kill a couple of
prairie-chickens or ducks. If the evening was at all cool, we built a
fire as darkness fell, and sat around it, while the leaping flames lit up
the trunks of the cottonwoods and gleamed on the pools of water in
the half-dry river bed. Then I would wrap myself in my blanket and
lie looking up at the brilliant stars until I fell asleep.
In both 1893 and 1894 I made trips to a vast tract of rolling prairie
land, some fifty miles from my ranch, where I had for many years
enjoyed the keen pleasure of hunting the prongbuck. In 1893 the
prong-horned bands were as plentiful in this district as I have ever
seen them anywhere. Lambert was with me; and in a week’s trip,
including the journey out and back, we easily shot all the antelope we
felt we had any right to kill; for we only shot to get meat, or an
unusually fine head. Lambert did most of the shooting; and I have
never seen a professional hunter do better in stalking antelope on the
open prairie. I myself fired at only two antelope, both of which had
already been missed. In each case a hard run and much firing at long
ranges, together with in one case some skilful manœuvring, got me
my game; yet one buck cost ten cartridges and the other eight. In
1894 I had exactly the reverse experience. I killed five antelope for
thirty-six shots, but each one that I killed was killed with the first
bullet, and in not one case where I missed the first time did I hit with
any subsequent shot. These five antelope were killed at an average
distance of about 150 yards. Those that I missed were, of course,
much farther off on an average, and I usually emptied my magazine
at each. The number of cartridges spent would seem extraordinary to
a tyro; and an unusually skilful shot, or else a very timid shot who
fears to take risks, will of course make a better showing per head
killed; but I doubt if men with experience in antelope hunting, who
keep an accurate account of the cartridges they expend, will see
anything much out of the way in the performance.
During the years I have hunted in the West I have always, where
possible, kept a record of the number of cartridges expended for
every head of game killed, and of the distances at which it was shot. I
have found that with bison, bear, moose, elk, caribou, bighorn and
white goat, where the animals shot at were mostly of large size and
usually stationary, and where the mountainous or wooded country
gave chance for a close approach, the average distance at which I
have killed the game has been eighty yards, and the average number
of cartridges expended per head slain, three; one of these
representing the death-shot, and the others standing either for
misses outright, of which there were not many, or else for wounding
game which escaped, or which I afterward overtook, or for stopping
cripples or charging beasts. I have killed but two peccaries, using but
one cartridge for each; they were close up. My experiences with
cougar have already been narrated. At wolves and coyotes I have
generally had to take running shots at very long range, and I have
shot but two—one of each—for fifty cartridges. Blacktail deer I have
generally shot at about ninety yards, at an expenditure of about four
cartridges apiece. Whitetail I have killed at shorter range; but the
shots were generally running, often taken under difficult
circumstances, so that my expenditure of cartridges was rather
larger. Antelope, on the other hand, I have on the average shot at a
little short of 150 yards, and they have cost me about nine cartridges
apiece. This, of course, as I have explained above, does not mean that
I have missed eight out of nine antelope, for often the entire nine
cartridges would be spent at an antelope which I eventually got. It
merely means that, counting all the shots of every description fired at
antelope, I had one head to show for each nine cartridges expended.
Thus, the first antelope I shot in 1893 cost me ten cartridges, of
which three hit him, while the seven that missed were fired at over
400 yards’ distance while he was running. We saw him while we
were with the wagon. As we had many miles to go before sunset, we
cared nothing about frightening other game, and, as we had no fresh
meat, it was worth while to take some chances to procure it. When I
first fired, the prongbuck had already been shot at and was in full
flight. He was beyond all reasonable range, but some of our bullets
went over him and he began to turn. By running to one side I got a
shot at him at a little over 400 paces, as he slowed to a walk,
bewildered by the firing, and the bullet broke his hip. I missed him
two or three times as he plunged off, and then by hard running down
a watercourse got a shot at 180 paces and broke his shoulder, and
broke his neck with another bullet when I came up.
This one was shot while going out to the hunting-ground. While
there Lambert killed four others. I did not fire again until on our
return, when I killed another buck one day while we were riding with
the wagon. The day was gray and overcast. There were slight flurries
of snow, and the cold wind chilled us as it blew across the endless
reaches of sad-colored prairie. Behind us loomed Sentinel Butte, and
all around the rolling surface was broken by chains of hills, by
patches of bad lands, or by isolated, saddle-shaped mounds. The
ranch wagon jolted over the uneven sward, and plunged in and out of
the dry beds of the occasional water courses; for we were following
no road, but merely striking northward across the prairie toward the
P. K. ranch. We went at a good pace, for the afternoon was bleak, the
wagon was lightly loaded, and the Sheriff of the county, whose
deputy I had been, and who was serving for the nonce as our
teamster and cook, kept the two gaunt, wild-looking horses trotting
steadily. Lambert and I rode to one side on our unkempt cow-ponies,
our rifles slung across the saddle bows.
Our stock of fresh meat was getting low and we were anxious to
shoot something; but in the early hours of the afternoon we saw no
game. Small parties of horned larks ran along the ground ahead of
the wagon, twittering plaintively as they rose, and now and then
flocks of long-spurs flew hither and thither; but of larger life we saw
nothing, save occasional bands of range horses. The drought had
been severe and we were far from the river, so that we saw no horned
stock. Horses can travel much farther to water than cattle, and, when
the springs dry up, they stay much farther out on the prairie.
At last we did see a band of four antelope, lying in the middle of a
wide plain, but they saw us before we saw them, and the ground was
so barren of cover that it was impossible to get near them. Moreover,
they were very shy and ran almost as soon as we got our eyes on
them. For an hour or two after this we jogged along without seeing
anything, while the gray clouds piled up in the west and the
afternoon began to darken; then, just after passing Saddle Butte, we
struck a rough prairie road, which we knew led to the P. K. ranch—a
road very faint in places, while in others the wheels had sunk deep in
the ground and made long, parallel ruts.
Almost immediately after striking this road, on topping a small
rise, we discovered a young prongbuck standing off a couple of
hundred yards to one side, gazing at the wagon with that absorbed
curiosity which in this game so often conquers its extreme wariness
and timidity, to a certain extent offsetting the advantage conferred
upon it by its marvellous vision. The little antelope stood broadside
on, gazing at us out of its great bulging eyes, the sharply contrasted
browns and whites of its coat showing plainly. Lambert and I leaped
off our horses immediately, and I knelt and pulled trigger; but the
cartridge snapped, and the little buck, wheeling round, cantered off,
the white hairs on its rump standing erect. There was a strong cross-
wind, almost a gale, blowing, and Lambert’s bullet went just behind
him; off he went at a canter, which changed to a breakneck gallop, as
we again fired; and he went out of sight unharmed, over the crest of
the rising ground in front. We ran after him as hard as we could pelt
up the hill, into a slight valley, and then up another rise, and again
got a glimpse of him standing, but this time farther off than before;
and again our shots went wild.
However, the antelope changed its racing gallop to a canter while
still in sight, going slower and slower, and, what was rather curious,
it did not seem much frightened. We were naturally a good deal
chagrined at our shooting and wished to retrieve ourselves, if
Welcome to Our Bookstore - The Ultimate Destination for Book Lovers
Are you passionate about books and eager to explore new worlds of
knowledge? At our website, we offer a vast collection of books that
cater to every interest and age group. From classic literature to
specialized publications, self-help books, and children’s stories, we
have it all! Each book is a gateway to new adventures, helping you
expand your knowledge and nourish your soul
Experience Convenient and Enjoyable Book Shopping Our website is more
than just an online bookstore—it’s a bridge connecting readers to the
timeless values of culture and wisdom. With a sleek and user-friendly
interface and a smart search system, you can find your favorite books
quickly and easily. Enjoy special promotions, fast home delivery, and
a seamless shopping experience that saves you time and enhances your
love for reading.
Let us accompany you on the journey of exploring knowledge and
personal growth!

ebookgate.com