Adaptive and Intelligent Systems Third

International Conference ICAIS 2014

Bournemouth UK September 8 10 2014
Proceedings 1st Edition Abdelhamid
Bouchachia (Eds.)
Visit to download the full and correct content document:
https://textbookfull.com/product/adaptive-and-intelligent-systems-third-international-co
nference-icais-2014-bournemouth-uk-september-8-10-2014-proceedings-1st-edition-a
bdelhamid-bouchachia-eds/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Formal Modeling and Analysis of Timed Systems 12th

International Conference FORMATS 2014 Florence Italy
September 8 10 2014 Proceedings 1st Edition Axel Legay

https://textbookfull.com/product/formal-modeling-and-analysis-of-
timed-systems-12th-international-conference-
formats-2014-florence-italy-september-8-10-2014-proceedings-1st-
edition-axel-legay/

Belief Functions Theory and Applications Third

International Conference BELIEF 2014 Oxford UK
September 26 28 2014 Proceedings 1st Edition Fabio
Cuzzolin (Eds.)
https://textbookfull.com/product/belief-functions-theory-and-
applications-third-international-conference-belief-2014-oxford-
uk-september-26-28-2014-proceedings-1st-edition-fabio-cuzzolin-
eds/

Intelligent Data Engineering and Automated Learning

IDEAL 2014 15th International Conference Salamanca
Spain September 10 12 2014 Proceedings 1st Edition
Emilio Corchado
https://textbookfull.com/product/intelligent-data-engineering-
and-automated-learning-ideal-2014-15th-international-conference-
salamanca-spain-september-10-12-2014-proceedings-1st-edition-
emilio-corchado/

Intelligent Systems 2014 Proceedings of the 7th IEEE

International Conference Intelligent Systems IS 2014
September 24 26 2014 Warsaw Poland Volume 1
Mathematical Foundations Theory Analyses 1st Edition P.
Angelov
https://textbookfull.com/product/intelligent-
systems-2014-proceedings-of-the-7th-ieee-international-
conference-intelligent-systems-
Discovery Science 17th International Conference DS 2014
Bled Slovenia October 8 10 2014 Proceedings 1st Edition
Sašo Džeroski

https://textbookfull.com/product/discovery-science-17th-
international-conference-ds-2014-bled-slovenia-
october-8-10-2014-proceedings-1st-edition-saso-dzeroski/

Case Based Reasoning Research and Development 22nd

International Conference ICCBR 2014 Cork Ireland
September 29 2014 October 1 2014 Proceedings 1st
Edition Luc Lamontagne
https://textbookfull.com/product/case-based-reasoning-research-
and-development-22nd-international-conference-iccbr-2014-cork-
ireland-september-29-2014-october-1-2014-proceedings-1st-edition-
luc-lamontagne/

Intelligent Virtual Agents 14th International

Conference IVA 2014 Boston MA USA August 27 29 2014
Proceedings 1st Edition Timothy Bickmore

https://textbookfull.com/product/intelligent-virtual-agents-14th-
international-conference-iva-2014-boston-ma-usa-
august-27-29-2014-proceedings-1st-edition-timothy-bickmore/

Algorithmic Learning Theory 25th International

Conference ALT 2014 Bled Slovenia October 8 10 2014
Proceedings 1st Edition Peter Auer

https://textbookfull.com/product/algorithmic-learning-
theory-25th-international-conference-alt-2014-bled-slovenia-
october-8-10-2014-proceedings-1st-edition-peter-auer/

Computer Safety Reliability and Security 33rd

International Conference SAFECOMP 2014 Florence Italy
September 10 12 2014 Proceedings 1st Edition Andrea
Bondavalli
https://textbookfull.com/product/computer-safety-reliability-and-
security-33rd-international-conference-safecomp-2014-florence-
italy-september-10-12-2014-proceedings-1st-edition-andrea-
 Abdelhamid Bouchachia (Ed.)

Adaptive and
LNAI 8779

Intelligent Systems
Third International Conference, ICAIS 2014
Bournemouth, UK, September 8–10, 2014
Proceedings

123
Lecture Notes in Artificial Intelligence 8779

Subseries of Lecture Notes in Computer Science

LNAI Series Editors

Randy Goebel
University of Alberta, Edmonton, Canada
Yuzuru Tanaka
Hokkaido University, Sapporo, Japan
Wolfgang Wahlster
DFKI and Saarland University, Saarbrücken, Germany

LNAI Founding Series Editor

Joerg Siekmann
DFKI and Saarland University, Saarbrücken, Germany
Abdelhamid Bouchachia (Ed.)

Adaptive and
Intelligent Systems
Third International Conference, ICAIS 2014
Bournemouth, UK, September 8-10, 2014
Proceedings

13
Volume Editor
Abdelhamid Bouchachia
Bournemouth University
Faculty of Science and Technology
Fern Barrow
Poole, Dorset, BH12 5BB, UK
E-mail: abouchachia@bournemouth.ac.uk

ISSN 0302-9743 e-ISSN 1611-3349

ISBN 978-3-319-11297-8 e-ISBN 978-3-319-11298-5
DOI 10.1007/978-3-319-11298-5
Springer Cham Heidelberg New York Dordrecht London

Library of Congress Control Number: 2014947557

LNCS Sublibrary: SL 7 – Artificial Intelligence

© Springer International Publishing Switzerland 2014
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered and
executed on a computer system, for exclusive use by the purchaser of the work. Duplication of this publication
or parts thereof is permitted only under the provisions of the Copyright Law of the Publisher’s location,
in ist current version, and permission for use must always be obtained from Springer. Permissions for use
may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution
under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of publication,
neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or
omissions that may be made. The publisher makes no warranty, express or implied, with respect to the
material contained herein.
Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
 Preface

Adaptation plays a central role in dynamically changing systems. It is about

the ability of the system to “responsively” self-adjust upon change in the sur-
rounding environment. Like living creatures that have evolved over millions of
years developing ecological systems due to their self-adaptation and fitness ca-
pacity to the dynamic environment, systems undergo similar cycles to improve
or at least not weaken their performance when internal or external changes
take place. Internal or endogenous change bears on the physical structure of the
system (hardware and/or software components) due mainly to faults, knowl-
edge inconsistency, etc. it requires a certain number of adaptivity features such
as flexible deployment, self-testing, self-healing and self-correction. Extraneous
change touches on the environment implication such as the operational mode or
regime, non-stationarity of input, new knowledge facts, the need to cooperate
with other systems, etc. these two classes of change shed light on the research
avenues in smart and autonomous systems.

To meet the challenges of these systems, a sustainable effort needs to be de-

veloped: (1) adequate operational structures involving notions like self-healing,
self-testing, reconfiguration, etc.; (2) appropriate design concepts encompassing
self-x properties (self-organization, self-monitoring, etc.) to allow for autonomy
and optimal behaviour in the (dynamically changing) environment; (3) efficient
computational algorithms targeting dynamic setting, life-long learning, evolving
and adaptive behaviour and structure. For this three-fold development direc-
tion, various design and computational models stemming from machine learning,
statistics, metaphors of nature inspired from biological and cognitive plausibil-
ity, etc. can be of central relevance. Due to its versatility, online adaptation is
involved in various research areas which are covered by the International Con-
ference on Adaptive and Intelligent Systems (ICAIS): neutral networks, data
mining, pattern recognition, computational intelligence, smart and agent-based
systems, distributed systems, ubiquitous environments, Internet, system engi-
neering, hardware, etc. the ICAIS conference strives to sharpen its focus on
issues related to online adaptation and learning in dynamically evolving envi-
ronments and to expand the understanding of all these inter-related concepts.

ICAIS is a biennial event. After the very successful ICAIS 2009 and ICAIS
2011, the conference witnessed another success this year. ICAIS 2014 received
32 submissions from 11 countries around the world (Argentina, Austria, Brazil,
Canada, China, Czeck Republic, Germany, Poland, Portugal, Tunisia, UK) which
underwent a rigorous and tough peer-review by three reviewers. Based on the
referees’ reports, the Program Committee selected 19 full papers.
VI Preface

ICAIS 2014 was enhanced by 3 keynote speeches and 1 invited talk given
by internationally renowned researchers, presenting insights and ideas in the
research areas covered by the conference. The rest of the program featured in-
teresting sessions. Overall we had 4 sessions: 1- Advances in feature selection
consisting of 4 papers 2- Clustering and classification techniques consisting of 6
papers 3- Adaptive optimization consisting of 5 papers 4- Time series analysis 4
papers
Although this set of sessions does not cover the topics of the conference, it
is a very good mix proving fresh insight in some of the techniques related to
adaptive and intelligent systems.

This volume and the whole conference are of course the result of team
effort. My thanks go to the local organizers, particularly to Tobias Baker, the
Steering Committee, the publicity chair and especially the Program Committee
members and the reviewers for their cooperation in the shaping of the confer-
ence and running the refereeing process. I highly appreciate the hard work and
timely feedback of the referees who did an excellent job. I would like to thank
all the authors and participants for their interest in ICAIS 2014. Thank you
for your time and effort in preparing and submitting your papers and your pa-
tience throughout the long process. Your work is the backbone of this conference.

The support of the University of Bournemouth and the Faculty of Science

and Technology was crucial in making this event happen. I also thank the tech-
nical sponsors, IEEE Computational Society and International Neural Networks
Society for their trust in ICAIS, and Springer for enabling the publication of
this volume.

We worked hard to make ICAIS 2014 an enticing and informative event.

July 2014 Abdelhamid Bouchachia

 Organization
ICAIS 2014 is organized by the department of Computing, Faculty of Science &
Technology, Bournemouth University.
Executive Committee
Conference Chair
Abdelhamid Bouchachia
Advisory Committee
Nikola Kasabov
Xin Yao
Djamel Ziou
Plamen Angelov
Witold Pedrycz
Janusz Kacprzyk
Program Committee
Plamen Angelov, UK
Emili Balaguer-Ballester, UK
Albert Bifet, New Zealand
Nizar Bouguila, Canada
Ahmed Bouridane, UK
Martin Butz, Germany
Heloisa Camargo, Brazil
André de Carvalho, Brazil
Damien Fay, UK
Joao Gama, Portugal
Haibo He, USA
Eyke Huellermeier, Germany
Ata Kaban, UK
Frank Klawonn, Germany
Mario Koeppen, Japan
Lukasz Kurgan, USA
Bournemouth University, UK
Auckland University, New Zealand
University of Birmingham, UK
University of Sherbrooke, Canada
University of Lancaster, UK
University of Edmonton, Canada
Polish Academy of Sciences, Poland
Stéphane Lecoeuche, France
Ahmad Lotfi, UK
Edwin Lughofer, Austria
Trevor Martin, UK
Jose Antonio Iglesias Martı́nez, Spain
Christophe Marsala, France
Hammadi Nait-Charif, UK
Marcin Paprzycki, Poland
Jose de Jesus Rubio, Mexico
Ali Salah, Turkey
Daniel Sanchez, Spain
Moamar Sayed-Mouchweh, France
Igor Škrjanc, Slovenia
Vaclav Snasel, Czech Republic
Alexey Tsymbal, Germany
VIII Organization

Additional Reviewers

Heysem Kaya, Turkey Hande Alemdar, Turkey

Aljaz Osojnik, Slovenia Vânia G. Almeida, Portugal

Local Organizing Committee

Tobias Baker Bournemouth University, UK
Hammadi Nait-Charif Bournemouth University, UK
Emili Balaguer-Ballester Bournemouth University, UK
Damien Fay Bournemouth University, UK

Sponsoring Institutions
Bournemouth University, UK
Faculty of Science and Technology, Bournemouth University, UK
The International Neural Networks Society
Invited Talks
 Feature Extraction for Change Detection

Ludmila Kuncheva
School of Computer Science, Bangor University
Dean Street, Bangor, Gwynedd, LL57 1UT, United Kingdom
l.i.kuncheva@bangor.ac.uk

An online classifier is only accurate if the distribution of the incoming data is the
same as the distribution of the data the classifier was trained upon. Therefore,
detecting a change and retraining the classifier accordingly is an important task.
This talk will challenge the concept of change and change detectability from
unsupervised streaming data. What constitutes a change? The answer is uniquely
determined by the context. But in order to have change detection methods that
work across problems, the relevant features must be exposed. The talk will be
focused on feature extraction for the purposes of change detection.
 Distributed Data Stream Mining

João Gama
LIAAD - INESC Porto, Laboratory of Artificial Intelligence and Decision Support
University of Porto, Portugal
joao.jgama@gmail.com

The phenomenal growth of mobile and embedded devices coupled with their
ever-increasing computational and communications capacity presents an exciting
new opportunity for real-time, distributed intelligent data analysis in ubiquitous
environments. In these contexts centralized approaches have limitations due to
communication constraints, power consumption (e.g. in sensor networks), and
privacy concerns. Distributed online algorithms are highly needed to address the
above concerns. The focus of this talk is on distributed stream mining algorithms
that are highly scalable, computationally efficient and resource-aware. These
features enable the continued operation of data stream mining algorithms in
highly dynamic mobile and ubiquitous environments.
 Music Lessons and Other Exercises

Kurt Geihs
University of Kassel, EECS Department|Distributed Systems
Wilhelmshöher Allee 73, D-34121 Kassel, Germany
geihs@uni-kassel.de

For more than a decade researchers have explored software systems that dynami-
cally adapt their behavior at run-time in response to changes in their operational
environments, user preferences, and underlying computing infrastructure. Our
particular focus has been on context-aware, self-adaptive applications in ubiqui-
tous computing scenarios. In my presentation I will discuss lessons learned from
three projects in the realm of self-adaptive systems. Some of these insights re-
late to purely technical concerns. Others touch on socio-technical concerns that
substantially influence the users’ acceptance of self-adaptive applications. Our
experiments have shown that both kinds of concerns are important. We claim
that an interdisciplinary software development methodology is needed in order
to produce socially aware applications that are both acceptable and accepted.
 Adaptive Ambient Intelligence and Smart
Environments

Ahmad Lotfi
School of Science and Technology, Nottingham Trent University
Clifton Campus, Clifton Lane, Nottingham, NG11 8NS, United Kingdom
ahmad.lotfi@ntu.ac.uk

Ambient Intelligence refers to a digital environment that proactively supports

people in their daily lives. It is an emerging discipline that brings intelligence to
our living environments, makes those environments sensitive to us, and adapt
according to the user’s needs. By enriching an environment with appropriate
sensors and interconnected devices, the environment would be able to sense
changes and support decisions that benefit the users of that environment. Such
smart environments could help to reduce the energy consumption, increase user’s
comfort, improve security and productivity, etc. One specific area of interest is
the application of ambient intelligence in Ambient Assisted Living, where the
home environment provides assistance with daily living activities for people with
disabilities. In my presentation, I will provide a review of the technologies and
environments that comprises ambient intelligence, as well as how changes in the
environment are reflected in the overall design of an adaptive ambient intelligence
environment.
 Table of Contents

Advances in Feature Selection

A Hierarchical Infinite Generalized Dirichlet Mixture Model with
Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Wentao Fan, Hassen Sallay, Nizar Bouguila, and Sami Bourouis

Importance of Feature Selection for Recurrent Neural Network Based

Forecasting of Building Thermal Comfort . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Martin Macas, Fabio Moretti, Fiorella Lauro, Stefano Pizzuti,
Mauro Annunziato, Alessandro Fonti, Gabriele Comodi, and
Andrea Giantomassi

Classifying Textures with Only 10 Visual-Words Using Hidden Markov

Models with Dirichlet Mixtures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Elise Epaillard, Nizar Bouguila, and Djemel Ziou

A Novel Method for Scene Categorization Using an Improved Visual

Vocabulary Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Tarek Elguebaly and Nizar Bouguila

Clustering and Classification

Learning Sentiment from Students’ Feedback for Real-Time
Interventions in Classrooms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Nabeela Altrabsheh, Mihaela Cocea, and Sanaz Fallahkhair

Decremental Rough Possibilistic K-Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

Asma Ammar, Zied Elouedi, and Pawan Lingras

A Dynamic Bayesian Model of Homeostatic Control . . . . . . . . . . . . . . . . . . 60

Will Penny and Klaas Stephan

ICALA: Incremental Clustering and Associative Learning

Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Matthias U. Keysermann

Context Identification and Exploitation in Dynamic Data Mining –

An Application to Classifying Electricity Price Changes . . . . . . . . . . . . . . . 80
Lida Barakat
XVI Table of Contents

Adaptive Optimization
Towards a Better Understanding of the Local Attractor in Particle
Swarm Optimization: Speed and Solution Quality . . . . . . . . . . . . . . . . . . . . 90
Vanessa Lange, Manuel Schmitt, and Rolf Wanka

A Models@run.time Approach for Multi-objective Self-optimizing

Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Sebastian Götz, Thomas Kühn, Christian Piechnick,
Georg Püschel, and Uwe Aßmann

Reference Architecture for Self-adaptive Management in Wireless

Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Jesús M.T. Portocarrero, Flavia C. Delicato, Paulo F. Pires, and
Thais V. Batista

Monitoring of Quality of Service in Dynamically Adaptive Systems . . . . . 121

Sihem Loukil, Slim Kallel, Ismael Bouassida Rodriguez, and
Mohamed Jmaiel

Future Prospects of Human Interaction with Artificial Autonomous

Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Andrzej M.J. Skulimowski

Advances in Time Series Analysis

Empirical Identification of Non-stationary Dynamics in Time Series of
Recordings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Emili Balaguer-Ballester, Alejandro Tabas-Diaz, and Marcin Budka

Physical Time Series Prediction Using Dynamic Neural Network

Inspired by the Immune Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
Abir Jaafar Hussain, Haya Al-Askar, and Dhiya Al-Jumeily

Collaborative Wind Power Forecast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

Vânia Almeida and João Gama

QoS-Based Reputation Feedback Fusion under Unknown Correlation . . . 172

Mohamad Mehdi, Nizar Bouguila, and Jamal Bentahar

Multi-resident Activity Recognition Using Incremental Decision

Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Markus Prossegger and Abdelhamid Bouchachia

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

 A Hierarchical Infinite Generalized Dirichlet
Mixture Model with Feature Selection

Wentao Fan1 , Hassen Sallay2 , Nizar Bouguila3 , and Sami Bourouis4

1
Department of Computer Science and Technology, Huaqiao University, China
fwt@hqu.edu.cn
2
Al Imam Mohammad Ibn Saud Islamic University (IMSIU), Riyadh, Saudi Arabia
hmsallay@imamu.edu.sa
3
Concordia University, Montreal, QC, Canada
nizar.bouguila@concordia.ca
4
Taif University, Taif, Saudi Arabia
s.bourouis@tu.edu.sa

Abstract. We propose a nonparametric Bayesian approach, based on hi-

erarchical Dirichlet processes and generalized Dirichlet distributions,
for simultaneous clustering and feature selection. The resulting statistical
model is learned within a variational framework that we have developed.
The merits of the developed model are shown via extensive simulations
and experiments when applied to the challenging problem of images
categorization.

Keywords: Clustering, mixture model, feature selection, generalized

Dirichlet, variational inference, image databases.

1 Introduction
Dirichlet process (DP) has became as one of the most popular nonparametric
Bayesian techniques [12,10]. It can be viewed as a stochastic process and can
be considered as distribution over distributions, also. Recently, hierarchical DP
[20,19] has been developed as a hierarchical nonparametric Bayesian model and
has shown promising results to the problem of model-based clustering of grouped
data with sharing clusters. It is built on the DP and involves Bayesian hierar-
chy where the base measure for a DP is itself distributed according to a DP.
The hierarchical DP framework is particularly useful in problems involving the
modeling of grouped data where observations are organized into groups, and
allow these groups to remain statistically linked by sharing mixture components
[20,19]. As several other statistical approaches, existing hierarchical DP-based
models considers the Gaussian assumption. Unlike, these existing works, in this
paper, we focus on a specific form of hierarchical DP mixture model where each
observation within a group is drawn from a mixture of generalized Dirichlet (GD)
distributions. We are mainly motivated by the fact that the generalized Dirichlet
distribution has been shown to be efficient in modeling high-dimensional propor-
tional data (i.e. normalized histograms) in a variety of applications from different
disciplines (e.g. computer vision, data mining, pattern recognition) [5,6,8,9].

A. Bouchachia (Ed.): ICAIS 2014, LNAI 8779, pp. 1–10, 2014.


c Springer International Publishing Switzerland 2014
2 W. Fan et al.

Having the hierarchical infinite generalized Dirichlet mixture model in hand, a

principled variational approach is developed to learn its parameters. To validate
the overall statistical framework a challenging application, which has attracted
the attention of the computer vision community recently [2,16], namely the clas-
sification of images containing categories of animals is considered. The rest of
this paper is organized as follows. In the next section, we propose our model.
Section 3 develops our variational learning procedure. Section 4 is devoted to
the experimental results. Finally, we conclude the paper in Section 5.

2 The Model

2.1 The Hierarchical Dirichlet Process Mixture Model

Let H be a distribution over some probability space Θ and γ be a positive real

number, then a random distribution G follows a DP with a base distribution H
and concentration parameter γ, denoted as G ∼ DP(γ, H), if

(G(A1 ), . . . , G(At )) ∼ Dir(γH(A1 ), . . . , γH(At )) (1)

where (A1 , . . . , At ) is the set of finite partitions of Θ, and Dir is a finite-

dimensional Dirichlet. Now, let us introduce the general setting of a two-level
hierarchical Dirichlet process: Assume that we have a grouped data set, in which
each group is associated with an infinite mixture model (a DP Gj ). This indexed
set of DPs {Gj } shares a base distribution G0 , which is itself distributed as a
DP:

G0 ∼ DP(γ, H) Gj ∼ DP(λ, G0 ) for each j, j ∈ {1, . . . , M } (2)

where j is an index for each group of data. In this work, we represent a hierarchi-
cal DP in a more intuitive and straightforward form through two stick-breaking
constructions which involves a global-level and a group-level construction [18,11].
In the global-level construction, the global measure G0 is distributed according
to a DP(γ, H) and can be described using a stick-breaking representation


k−1 ∞

ψk ∼ Beta(1, γ) Ωk ∼ H ψk = ψk (1 − ψs ) G0 = ψk δΩk (3)
s=1 k=1

where {Ωk } is a set of independent random variables distributed according to H,

δΩk is an atom
 at Ωk . The random variables ψk are the stick-breaking weights
that satisfy ∞ k=1 ψk = 1, and are obtained by recursively breaking a unit length
stick into an infinite number of pieces. In this work, we apply the conventional
stick-breaking representation [21] to construct each group-level DP Gj


t−1 ∞

  
πjt ∼ Beta(1, λ) jt ∼ G0 πjt = πjt (1 − πjs ) Gj = πjt δjt
s=1 t=1
 A Hierarchical Infinite Generalized Dirichlet Mixture Model 3

where δjt are group-level

∞ atoms at jt , {πjt } is a set of stick-breaking weights
which satisfies t=1 πjt = 1. Since jt is distributed according to the base
distribution G0 , it takes on the value Ωk with probability ψk . This can also
represented using a binary latent variable Wjtk as an indicator variable, such
that Wjtk ∈ {0, 1}, Wjtk = 1 if jt maps to the base-level atom Ωk which is
W
indexed by k; otherwise, Wjtk = 0. Then, we can have jt = Ωk jtk . As a result,
group-level atoms jt do not need to be explicitly represented which further
simplifies the inference process as it shall be clearer later. The indicator variable
W = (Wjt1 , Wjt2 , . . .) is distributed according to ψ in the form


M  ∞
∞ 
Wjtk
p(W |ψ) = ψk (4)
j=1 t=1 k=1

Since ψ is a function of ψ  according to a stick-breaking construction, then


M  ∞
∞  
  k−1 W
p(W |ψ  ) = ψk (1 − ψs ) jtk (5)
j=1 t=1 k=1 s=1

The prior distribution of ψ  is a specific Beta distribution

∞
 ∞

p(ψ  ) = Beta(1, γk ) = γk (1 − ψk )γk −1 (6)
k=1 k=1

Let i index the observations within each group j, we assume that each θji is a
factor corresponding to an observation Xji , and the factors θj = (θj1 , θj2 , . . .)
are distributed according to DP Gj , one for each j: θji |Gj ∼ Gj , Xji |θji ∼
F (θji ), where F (θji ) represents the distribution of the observation Xji given θji .
According to Eq.(2), the base distribution H of G0 provides the prior for θji .
This setting forms the definition of a hierarchical DP mixture model, where each
group is associated with a mixture model, and the components are shared among
these mixtures due to the sharing of atoms Ωk among all Gj . Furthermore, since
each θji is distributed according to Gj , it takes the value jt with probability
πjt . Next, we introduce a binary latent variable Zjit ∈ {0, 1} as an indicator
variable. That is, Zjit = 1 if θji is associated with component t and maps to the
Z
group-level atom jt ; otherwise, Zjit = 0. Therefore, we have θji = jtjit . Since
Z W Zjit
also maps to the global-level atom Ωk , we then have θji = jtjit = Ωk jtk
jt .
The indicator variable Z = (Zji1 , Zji2 , . . .) is distributed according to π as


M  ∞
N 
Z
p(Z|π) = πjtjit (7)
j=1 i=1 t=1

According to the stick-breaking construction, π is a function of π  . We then have


M  ∞
N  
t−1
p(Z|π ) = 
[πjt 
(1 − πjs )]Zjit (8)
j=1 i=1 t=1 s=1
4 W. Fan et al.

 ∞
M   ∞
M 
p(π  ) = Beta(1, λjt ) =  λjt −1
λjt (1 − πjt ) (9)
j=1 t=1 j=1 t=1

2.2 Hierarchical DP Mixture Model of GD Distributions with

Feature Selection

Assume that we have a D-dimensional vector Y = (Y1 , . . . , YD ) drawn from a

GD distribution with parameters α = (α1 , . . . , αD ) and β j = (β1 , . . . , βD )


D
Γ (αl + βl ) αl −1  
l
γ
GD(Y |α, β) = Yl 1− Yf l (10)
Γ (αl )Γ (βl )
l=1 f =1

D
where l=1 Yl < 1 and 0 < Yl < 1 for l = 1, . . . , D, αl > 0, βl > 0, γl =
βl − αl+1 − βl+1 for l = 1, . . . , D − 1, γD = βD − 1, and Γ (·) is the gamma
function. Based on an interesting mathematical property of the GD distribution
which is thoroughly discussed in [7], we can transform the original data point
Y using a geometric transformation into another D-dimensional data point X
D
with independent features in the form of GD(X|α, β) = l=1 Beta(Xl |αl , βl ),
l−1
where X = (X1 , . . . , XD ), X1 = Y1 and Xl = Yl /(1 − f =1 Yf ) for l > 1,
and Beta(Xl |αl , βl ) is a Beta distribution defined with parameters {αl , βl }. Now
let us consider a data set X that contains N random vectors separated into
M groups, then each vector X ji = (Xji1 , . . . , XjiD ) is represented in a D-
dimensional space and is drawn from a hierarchical infinite GD mixture model.
In practice, not all the features are important, some of them may be irrelevant
and may even degrade the clustering performance. Therefore, feature selection
technique is important and is adopted here as a tool to chooses the “best” fea-
ture subset. The most common feature selection technique, in the context of
unsupervised learning, defines an irrelevant feature as the one having a distribu-
tion independent from class labels [13]. Therefore, in our work the distribution
of each feature Xl can be defined by

p(Xjil ) = Beta(Xjil |αkl , βkl )φjil Beta(Xjil |αl , βl )1−φjil (11)

where φjil is a binary latent variable represents the feature relevance indicator,
such that φjil = 0 denotes the feature l of group j is irrelevant (i.e. noise)
and follows a Beta distribution: Beta(Xjil |αl , βl ); otherwise, the feature Xjil is
relevant. The prior distribution of φ is defined as


M 
N 
D
φ 1−φjil
p(φ|) = l1jil l2 (12)
j=1 i=1 l=1

where each φjil is a Bernoulli variable such that p(φjil = 1) = l1 and p(φjil =
0) = l2 . The vector  = (1 , . . . , D ) represents the features saliencies such that
 A Hierarchical Infinite Generalized Dirichlet Mixture Model 5

l = (l1 , l2 ) and l1 + l2 = 1. Furthermore, a Dirichlet distribution is chosen

over  as
D 
D
Γ (ξ1 + ξ2 ) ξ1 −1 ξ2 −1
p() = Dir(l |ξ) =   l2 (13)
Γ (ξ1 )Γ (ξ2 ) l1
l=1 l=1

In the next step, we need to introduce Gamma prior distributions for parameters
α, β, α and β .

3 Variational Model Learning

In order to simplify notations, we define Θ = (Ξ, Λ) as the set of latent and un-
known random variables, where Ξ = {Z, φ} and Λ = {W , , ψ, π , α, β, α , β  }.
Variational inference [1,3] is a deterministic approximation technique that is used
to find tractable approximations for posteriors of a variety of statistical models.
The goal is to find an approximation Q(Θ) to the true posterior distribution
p(Θ|X ) by maximizing the lower bound of ln p(X ):

L(Q) = Q(Θ) ln[p(X , Θ)/Q(Θ)]dΘ (14)

In this work, we adopt factorial approximation which is commonly used in

variational inference [3] to factorize Q(Θ) into disjoint tractable distributions.
Moreover, we apply a truncation technique as in [4] to truncate the variational
approximations of global- and group-level Dirichlet process at K and T . It is
noteworthy that the truncation levels K and T are variational parameters which
can be freely initialized and will be optimized automatically during the learning
process. By using truncated stick-breaking and factorization, the approximated
posterior distribution q(Θ) can be fully factorized into disjoint distributions as

q(Θ) = q(Z)q(W )q(φ)q(π  )q(ψ  )q(α)q(β)q(α )q(β  )q() (15)

In our work, variational inference is performed based on a natural gradient

method as introduced in [17]. The main idea is that, since our model has con-
jugate priors, the functional form of each factor in the variational posterior
distribution is known. Therefore, the lower bound L(q) can be considered as a
function of the parameters of these distributions by taking general parametric
forms of these distributions. The optimization of variational factors is then ob-
tained by maximizing the lower bound with respect to these parameters. In our
case, the functional form of each variational factor is the same as its conjugate
prior distribution, namely Discrete for Z and W , Bernoulli for φ, Dirichlet for ,
Beta for ψ  and π  , and Gamma for α, β, α and β . Therefore, the parametric
forms for these variational posterior distributions can be defined as the following


M 
N 
T
Z

M 
T 
K
W
q(Z) = ρjitjit q(W ) = ϑjtkjtk (16)
j=1 i=1 t=1 j=1 t=1 k=1
6 W. Fan et al.


M 
N 
D 
D
Dir(l |ξ ∗ )
φ
q(φ) = ϕjiljil (1 − ϕjil )
1−φjil
q() = (17)
j=1 i=1 l=1 l=1


M 
T 
K
q(π  ) = 
Beta(πjt |ajt , bjt ) q(ψ  ) = Beta(ψk |ck , dk ) (18)
j=1 t=1 k=1


K 
D 
K 
D
q(α) = G(αkl |ũkl , ṽkl ) q(β) = G(βkl |g̃kl , h̃kl ) (19)
k=1 l=1 k=1 l=1


D 
D
q(α ) = G(αl |ũl , ṽl ) q(β  ) = G(βl |g̃l , h̃l ) (20)
l=1 l=1

Consequently, the parameterized lower bound L(q) can be obtained by substi-

tuting Eqs.(16)∼(20) into Eq.(14). Maximizing this bound with respect to these
parameters then gives the required re-estimation equations. The variational in-
ference for our model can be performed then as an EM-like algorithm.

4 Experimental Results

In this part, we validate the proposed hierarchical infinite GD mixture model

with feature selection (referred to as HInGDFs) through a real-world application
concerning images categorization. In our case, we concentrate on the problem of
discriminating different images of breeds of cats and dogs, which is a specific type
of object categorization. This problem is extremely challenging since cats and
dogs are highly deformable and different breeds may differ only by a few subtle
phenotypic details [15]. In our experiments, we initialize the global truncation
level K to 800, and the group truncation level T to 100. The hyperparameters λjt
and γk are initialized to 0.05. The initial values of hyperparameters ukl , gkl , ul ,
vl for the conjugate Beta priors are set to 0.1, and vkl , hkl , vl , hl are set to 0.01.
The hyperparameters ξ1 and ξ2 of the feature saliency are both initialized to 0.5.
These specific choices were found convenient according to our experiments.

4.1 Experimental Setting

We perform the categorization of cats and dogs using the proposed HInGDFs
model and the bag-of-visual words representation. Our methodology are summa-
rized as follows: First, we extract the 128-dimensional scale-invariant
feature transform (SIFT) [14] descriptors from each image using the Difference-
of-Gaussians (DoG) interest point detectors and then normalized. Then, these
extracted SIFT features are modeled using our HInGDFs. Specifically, each im-
age Ij is considered as a “group” and is therefore associated with a Dirichlet
process mixture (infinite mixture) model Gj . Thus, each extracted SIFT feature
vector Xji of the image Ij is supposed to be drawn from the infinite mixture
 A Hierarchical Infinite Generalized Dirichlet Mixture Model 7

model Gj , where the mixture components of Gj can be considered as “visual

words”. A global vocabulary is constructed and is shared among all groups (im-
ages) through the common global infinite mixture model G0 of our hierarchical
model. This setting matches the desired design of a hierarchical Dirichlet process
mixture model. It is noteworthy that an important step in image categorization
approaches with bag-of-visual words representation is the construction of visual
vocabulary. Nevertheless, most of the previously invented approaches have to
apply a separate vector quantization algorithm (such as K-means) to build a
visual vocabulary, where the vocabulary size is normally manually selected. In
our approach, the construction of the visual vocabulary is part of our mixture
framework, and therefore the size of the vocabulary (i.e., the number of mixture
components in the global-level mixture model) can be automatically inferred
from the data thanks to the property of nonparametric Bayesian model. Since
the goal of our experiment is to determine which image category (breeds of cats
and dogs) that a testing image Ij belongs to, we also need to introduce an in-
dicator variable Bjm associated with each image (or group) in our hierarchical
Dirichlet process mixture framework. Bjm means image Ij is from category m
and is drawn from another infinite mixture model which is truncated at level
J. This means that we need to add a new level of hierarchy to our hierarchical
infinite mixture model with a sharing vocabulary among all image categories. In
this experiment, we truncate J to 50 and initialize the hyperparameter of the
mixing probability of Bjm to 0.05. Finally, we assign a testing image to a given
category according to Bayes’ decision rule.

4.2 Data Set

In this work, we evaluate the effectiveness of the proposed approach for catego-
rizing cats and dogs using a publicly available database namely the Oxford-IIIT
Pet database [15]1 . It contains 7,349 images of cats and dogs, and is composed of
12 different breeds of cats and 25 different breeds of dogs. Each of these breeds
contains about 200 images. Some sample images of this database are displayed
in Fig. 1 (cats) and Fig. 2 (dogs). This database is randomly divided into two
partitions: one for training (to learn the model and build the visual vocabulary),
the other one for testing.

Results. In our experiments, we demonstrate the advantages of the proposed

HInGDFs approach by comparing its performance with two other mixture mod-
els including the hierarchical infinite GD mixture model without feature selection
(HInGD) and the hierarchical infinite Gaussian mixture model with feature se-
lection (HInGFs). To make a fair comparison, all of these models are learned
using variational inference and we evaluated the categorization performance by
running the approach 30 times. In the first part of our experiments, we measure
the performance of our approach for categorizing cats (12 breeds) and dogs (25
breeds), respectively. The average categorization performances of our approach
1
Database available at: http://www.robots.ox.ac.uk/∼vgg/data/pets/
8 W. Fan et al.

(a) (b) (c) (d) (e) (f) (g) (h)

Fig. 1. Sample cat images from the Oxford-IIIT Pet database. (a) Abyssinian, (b)
Bengal, (c) Birman, (d) Bombay, (e) British Shorthair, (f) Egyptian Mau, (g) Persian,
(h) Ragdoll.

(a) (b) (c) (d) (e) (f) (g) (h)

Fig. 2. Sample dog images from the Oxford-IIIT Pet database. (a) Basset hound, (b)
Boxer, (c) Chihuahua, (d) Havanese, (e) Keeshond, (f) Pug, (g) Samyod, (h) Shiba inu.

and the two other tested approaches are shown in Table 1. As we can see from
this table, our approach (HInGDFs) provided the highest categorization accu-
racy among all tested approaches. Moreover, HInGDFs outperformed HInGD
and HInGFs which demonstrates the advantage of incorporating a feature se-
lection scheme into our framework and verifies that the GD mixture model has
better modeling capability for proportional data than Gaussian. Next, we have
evaluated all approaches for the whole Oxford-IIIT Pet database (i.e., we do
not separate cat and dog images). The corresponding results are shown in Ta-
ble 1. Based on the obtained results, the proposed approach provides again the
best categorization accuracy rate (42.73%) as compared toHInGD (39.58%) and

Table 1. The average categorization performance (%) and the standard deviation
obtained over 30 runs using different approaches. The values in parenthesis are the
standard deviations of the corresponding quantities.

Method Cats Dogs Both

HInGDFs 56.38 (0.95) 44.23 (1.21) 42.94 (1.05)
HInGD 52.92 (1.46) 40.86 (1.07) 39.58 (1.19)
HInGFs 48.65 (1.39) 37.52 (1.13) 35.27 (0.93)
 A Hierarchical Infinite Generalized Dirichlet Mixture Model 9

0.9

0.8

Feature Saliency
0.7

0.6

0.5

0.4

0.3

0.2
0 20 40 60 80 100 120
Features

Fig. 3. Feature saliencies of different features calculated by HInGDFs

HInGFs (35.27%) approaches. Furthermore, the corresponding feature saliencies

of the 128-dimensional SIFT vectors obtained by the proposed are illustrated in
Fig. 3. According to this figure, it is clear that the different features do not con-
tribute equally in the categorization, since they have different relevance degrees.

5 Conclusion
In this paper, we have developed a nonparametric Bayesian approach for data
modeling, classification, and feature selection, using both hierarchical DP and
generalized Dirichlet distributions which allows to model data without the need
to define, a priori, the complexity of the entire model. In order to learn the
parameters of the proposed model we have derived an inference procedure that
relies on variational Bayes. The validation of the model has been based on a
challenging application namely images categorization. We are currently working
on the extension of our learning framework to online settings to handle the
problem of dynamic data modeling.

Acknowledgments. The second author would like to thank King Abdulaziz

City for Science and Technology (KACST), Kingdom of Saudi Arabia, for their
funding support under grant number 11-INF1787-08.

References
1. Attias, H.: A variational Bayes framework for graphical models. In: Proc. of Ad-
vances in Neural Information Processing Systems (NIPS), pp. 209–215 (1999)
2. Berg, T., Forsyth, D.: Animals on the web. In: 2006 IEEE Computer Society Con-
ference on Computer Vision and Pattern Recognition, vol. 2, pp. 1463–1470 (2006)
3. Bishop, C.M.: Pattern Recognition and Machine Learning. Springer (2006)
4. Blei, D.M., Jordan, M.I.: Variational inference for Dirichlet process mixtures.
Bayesian Analysis 1, 121–144 (2005)
10 W. Fan et al.

5. Bouguila, N., Ziou, D.: A hybrid SEM algorithm for high-dimensional unsupervised
learning using a finite generalized Dirichlet mixture. IEEE Transactions on Image
Processing 15(9), 2657–2668 (2006)
6. Bouguila, N., Ziou, D.: High-dimensional unsupervised selection and estimation of
a finite generalized Dirichlet mixture model based on minimum message length.
IEEE Transactions on Pattern Analysis and Machine Intelligence 29(10), 1716–
1731 (2007)
7. Boutemedjet, S., Bouguila, N., Ziou, D.: A hybrid feature extraction selection ap-
proach for high-dimensional non-Gaussian data clustering. IEEE Transactions on
Pattern Analysis and Machine Intelligence 31(8), 1429–1443 (2009)
8. Fan, W., Bouguila, N.: Variational learning of a Dirichlet process of generalized
Dirichlet distributions for simultaneous clustering and feature selection. Pattern
Recognition 46(10), 2754–2769 (2013)
9. Fan, W., Bouguila, N., Ziou, D.: Unsupervised hybrid feature extraction selection
for high-dimensional non-gaussian data clustering with variational inference. IEEE
Transactions on Knowlege and Data Engineering 25(7), 1670–1685 (2013)
10. Ferguson, T.S.: Bayesian Density Estimation by Mixtures of Normal Distributions.
Recent Advances in Statistics 24, 287–302 (1983)
11. Ishwaran, H., James, L.F.: Gibbs sampling methods for stick-breaking priors. Jour-
nal of the American Statistical Association 96, 161–173 (2001)
12. Korwar, R.M., Hollander, M.: Contributions to the theory of Dirichlet processes.
The Annals of Probability 1, 705–711 (1973)
13. Law, M.H.C., Figueiredo, M.A.T., Jain, A.K.: Simultaneous feature selection and
clustering using mixture models. IEEE Transactions on Pattern Analysis and Ma-
chine Intelligence 26(9), 1154–1166 (2004)
14. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International
Journal of Computer Vision 60(2), 91–110 (2004)
15. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.V.: Cats and dogs. In: IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3498–3505
(2012)
16. Ramanan, D., Forsyth, D.A.: Using temporal coherence to build models of animals.
In: Proc. of the 9th IEEE International Conference on Computer Vision (ICCV),
pp. 338–345. IEEE Computer Society (2003)
17. Sato, M.: Online model selection based on the variational Bayes. Neural Compu-
tation 13, 1649–1681 (2001)
18. Sethuraman, J.: A constructive definition of Dirichlet priors. Statistica Sinica 4,
639–650 (1994)
19. Teh, Y.W., Jordan, M.I.: Hierarchical Bayesian Nonparametric Models with Ap-
plications. In: Hjort, N., Holmes, C., Müller, P., Walker, S. (eds.) Bayesian Non-
parametrics: Principles and Practice. Cambridge University Press (2010)
20. Teh, Y.W., Jordan, M.I., Beal, M.J., Blei, D.M.: Hierarchical Dirichlet processes.
Journal of the American Statistical Association 101(476), 1566–1581 (2006)
21. Wang, C., Paisley, J.W., Blei, D.M.: Online variational inference for the hierarchical
Dirichlet process. Journal of Machine Learning Research - Proceedings Track 15,
752–760 (2011)
 Importance of Feature Selection
for Recurrent Neural Network
Based Forecasting of Building Thermal Comfort

Martin Macas1 , Fabio Moretti2 , Fiorella Lauro2 , Stefano Pizzuti2 ,

Mauro Annunziato2 , Alessandro Fonti3 , Gabriele Comodi3 ,
and Andrea Giantomassi4
1
Department of Cybernetics, Czech Technical University in Prague,
Prague, Czech Republic
mmacas@seznam.cz
2
Unità Tecnica Tecnologie Avanzate per l’Energia e l’Industria,
ENEA (Italian National Agency for New Technologies, Energy and Sustainable
Economic Development), Cassacia Research Center, Roma, Italy
3
Dipartimento di Ingegneria Industriale e Scienze Matematiche,
Università Politecnica delle Marche, Ancona, Italy
4
Dipartimento di Ingegneria dell’Informazione,
Università Politecnica delle Marche, Ancona, Italy

Abstract. The paper demonstrates the importance of feature selection

for recurrent neural network applied to problem of one hour ahead fore-
casting of thermal comfort for office building heated by gas. Although
the accuracy of the forecasting is similar for both the feed-forward and
the recurrent network, the removal of features leads to accuracy reduc-
tion much earlier for the feed-forward network. The recurrent network
can perform well even with less than 50% of features. This brings signifi-
cant benefits in scenarios, where the neural network is used as a blackbox
model of thermal comfort, which is called by an optimizer that minimizes
the deviance from a target value. The reduction of input dimensionality
can lead to reduction of costs related to measurement equipment, data
transfer and also computational demands of optimization.

Keywords: Forecasting, thermal comfort, gas, heating, neural networks,

feature selection.

1 Introduction
Although artificial neural networks are very popular soft-computing techniques
used in industrial applications, recurrent neural networks are not used so often
like the feed-forward models. One possible cause is the fact that their training
is usually much more difficult and more complex recurrent models are more
sensitive to over-fitting. It can be therefore crucial to perform a proper selection
of network inputs, which can simplify the training and can lead to a better
generalization abilities [4], [2]. A proper input selection was observed to be very
important particularly in real-world applications [8].

A. Bouchachia (Ed.): ICAIS 2014, LNAI 8779, pp. 11–19, 2014.


c Springer International Publishing Switzerland 2014
12 M. Macas et al.

In this paper, a simple recurrent neural network model is adopted. The result-
ing network can be further used as a data-based black box model for optimization
of the building heating. At each hour, a building management system finds the
indoor air temperature set points that lead to a minimum output of the con-
sumption model and a proper level of thermal comfort. For this purpose, it is
crucial to reach a good prediction accuracy of both consumption and thermal
comfort. Since the cost function is highly nonlinear and multi-modal, popula-
tion based metaheuristical optimization can be used with advantage. Because
the neural network is used in optimization loop many times, it is also crucial to
have the network as smallest as possible. Moreover, if the optimization is per-
formed remotely, one must minimize the amount of data that are measured and
transferred from the building to the optimization agent. All these requirements
imply a critical need for a proper selection of features, which leads to reasonable
data acquisition requirements and proper prediction accuracy.
We focus on one hour ahead forecasting of thermal discomfort in particular
gas-heated office building. The comfort is assessed in terms of Predicted Mean
Vote model. First we demonstrate that accuracy of the recurrent model is higher
than the accuracy of feed-forward network. The main conclusion however is that
sensitivity based feature selection can help the recurrent network to reach much
higher reduction of the input dimensionality by simultaneously keeping a good
accuracy level. This phenomenon has been observed also in study [3] focused on
the prediction of heating gas consumption.
In Section 2, we briefly describe all the methods used in the experimental
part, which is described in section 3. Some final discussion and conclusions can
be found in Section 4.

Fig. 1. Outside of F40 building

Feature Selection for Recurrent Network based Thermal Comfort Forecasting 13

Fig. 2. Partitioning of F40 building zones

2 Data and Methods

2.1 Thermal Comfort Model

To asses a thermal comfort inside particular zones, we used Predicted Mean

Vote model (PMV). The model was developed by P. O. Fanger using heat bal-
ance equations and empirical studies about skin temperature to define comfort.
The model has two outputs. One is PMV value, which is a continuous variable
between −3 and +3. Its ideal value is zero. Positive PMV means too hot environ-
ment and the negative PMV corresponds to cold environment. The comfort zone
is for PMV between −0.5 and +0.5. The second output of PMV model is Pre-
dicted Percentage of Dissatisfied (PPD), which is always greater than 5%. In our
experiments we used implementation from Technical University of Eindhoven,
Netherland [10].

2.2 F40 Building Model

A real office building located at ENEA (Cassacia Research Centre, Rome, Italy)
was considered as a case study (see Figure 1). The building is composed of three
floors and a thermal subplant in the basement. The building is equipped with an
advanced monitoring system aimed at collecting data about the environmental
conditions and electrical and thermal energy consumption. In the building there
are 41 offices of different size with a floor area ranging from 14 to 36 m2 , 2
EDP rooms each of about 20 m2 , 4 Laboratories,1 Control Room and 2 Meeting
Rooms. Each office room has from 1 up to 2 occupants.
In order to simulate the variables of interest, a MATLAB Simulink simulator
based on HAMBASE model ([7], [9]) was developed. In particular, the build-
ing was divided into 15 different zones according to different thermal behavior
14 M. Macas et al.

depending to solar radiation exposure. Therefore each zone is modeled with sim-
ilar thermal and physical characteristics. Figure 2 shows the division of zones
for each floor. Each zone covers more rooms. Although there are 15 zones at
all, zones numbers 3, 8 and 13 correspond to corridors and do not have fan
coils. Below, these zones are called non-active zones while all the other zones
are called active zones. The simulator is used to generate data from which the
thermal comfort can be evaluated. Its inputs are indoor temperature set points
and external meteorological conditions. Its outputs include radiant temperature,
air temperature and relative humidity.
From those variables, the PPD can be computed using PMV model. Among
those taken from the building model,there are four variables, that were set to
their typical values - metabolism ( 70 [W m−2 ]), external work (0 [W m−2 ]),
clothing (1 [−]), and air velocity (0.1 [ms−1 ]). This simplification is used, because
we do not have any model of yet. This issue is out of scope of this paper and
will be solved in future work.

2.3 Data

A potential remote control agent would be based on a simple data-driven black

box model (here it is a surrogate of the simulator). Metaheuristical algorithms
can use such model to optimise temperature set point of the zones so as to keep
a predefined comfort level or to minimise thermal consumption and maximise
user’s comfort. Optimized temperature set points would be then applied to the
simulator in order to evaluate the resulting energy savings and comfort. The
optimization issue is not covered by this paper, which is devoted only to black
box modeling. We used the simulator with the following settings.
To obtain valid and reliable results, we simulated four heating seasons -
2005/2006, 2006/2007, 2007/2008 and 2008/2009. Each data set consists of 75
days which corresponds to 75 × 24 = 1800 hourly data instances. The data from
the first heating season are called the training data and are used for both the
training and feature selection. The data from 2006/2007 are used for selection
of the best number of inputs. And the data from 2007/2008 and 2008/2009 are
used for the final validation of methods.
The behavior of supply water temperature set point was controlled by a sim-
ple weather compensation rule. To excite the dynamics of the system in a proper
degree, we also added a random component. The value of the set point is Gaus-
sian random number with standard deviation 10◦ C and mean equal to 70 − 2Te ,
where Te is the external temperature. If the generated number is out of feasibil-
ity interval 35; 70◦ C, the value of water temperature set point is replaced by
uniformly distributed random number from this feasibility interval. The behav-
ior of inside air temperature set points differs for daytime and nighttime hours.
Between 6 a.m. and 8 p.m., they are also Gaussian random numbers with mean
21◦ C and standard deviation 1◦ C. Moreover, there is a saturation under 19◦ C
and above 23◦ C. Between 8 p.m. and 6 a.m., there is a nighttime regime and
all the set points are 17◦ C.
Another random document with
no related content on Scribd:
miles long and 69 broad, average depth 492 feet; temperature of
water 48°; annual rainfall about 78 inches. Besides the bleak table-
land west of the Cordillera Real, the Department comprises many
beautiful valleys east of the mountains. It contains, with the capital
La Paz, several important though not large cities: Corocoro, a famed
copper centre; Sorata, north, a delightful spot, at an altitude of 8000
feet, a centre for the rubber industry and for mining; Achacachi, a big
Indian town near the Lake; Copacabana, a very sacred religious
shrine, also on the Lake; Coroicas, east in the Yungas Valley. Actual
or potential wealth of almost every kind may be found here: large
flocks of llamas, alpacas, and some vicuñas; sheep and goats, herds
of cattle and horses; vegetable products, coca, cotton, coffee,
cereals, etc.; great riches in minerals, gold, silver, tin, copper, rock
crystal, berenguela, a fine native marble, all these in fine quality and
immense quantity. The wonderfully grand and varied scenery of this
Department can be duplicated or rivalled in few quarters of the
globe.
Oruro, directly south of La Paz, between Potosí and Chile, is the
smallest of the Departments, entirely on the plateau, thus having a
much smaller range of altitude and of climate. Its wealth is in
minerals, especially silver and tin, though gold, bismuth, borax, and
sulphur are found. Alpaca, chinchilla, and wool are other products.
The capital, Oruro, 127 miles south of La Paz, while for years the
terminus of the railway from Antofagasta, became an important
commercial and industrial centre; in addition it is one of the busiest
mining districts. Half of the mining men are foreigners; a large
proportion of the ordinary population is Indian. There is also
agreeable society, with people quite up to date in Paris fashions. The
only other city in the Department is Challapata.
Cochabamba, south of El Beni, east of La Paz and west of Santa
Cruz, is sometimes called the granary of the Republic. It is a
particularly delightful region, half temperate and half semitropical,
with altitude ranging from 973 to 16,777 feet. It possesses immense
plains with innumerable herds of cattle, magnificent fertile valleys,
rich forests, and minerals.
 Cochabamba, the capital, is the city next in size to La Paz, 276
miles distant; with an altitude of 8387 feet it enjoys a delightful
climate, an average temperature of 66°, and abundant rain. It has a
good club, an American Institute (a school for boys), and the least
illiteracy. The recent arrival of the railroad insures its rapid growth.
Potosí is another plateau Department, both east and south of
Oruro, bordering also on Chuquisaca and Tarija, and with Argentina
south and Chile west. It has for the most part a rather cold,
disagreeable climate, the altitude ranging from above 20,000 to a
little over 6000 feet.
The name Potosí, for centuries famous as almost a synonym for
silver, belongs also to the capital city, at a height of 13,388 feet. In
1650, when New York was a small village, it had 160,000
inhabitants. The city has an excellent mint, built in 1562, also a
superb old cathedral, a good library, and a museum. Above the town
32 artificial lakes were constructed of which 22 are still good. The
temperature varies from 9° to 59°; it is said that they have the four
seasons in one day. Even here potatoes, barley, and beans will
grow, but living is very dear. The town has foundries, engineering
shops, and 28 smelters. The Department still contains enormous
quantities of silver and tin in almost every variety of combination, and
more towns that are rich mining centres than any other Department
except La Paz. Uyuni is a town of 6000 inhabitants, Colquechaca
has a population of 8000 miners; Tupiza, soon to be reached by the
railway, is a pretty city in a pleasant valley, 66 miles from the
Argentine border.
Chuquisaca, east of Potosí and south of Cochabamba and Santa
Cruz, is much warmer, the altitude ranging from 255 to 13,450 feet.
Here also are mines of the usual varieties and petroleum too, tropical
fruits and vegetables of course, and plains with vast herds and
flocks.
Sucre, the legal capital of the Republic, altitude 9328 feet, is
situated on an arm of the Pilcomayo River. Social distinctions are
here more marked, and there is an exclusive circle agreeable to
those of the right type. Sucre, Cochabamba, and La Paz are called
the most cultured cities. More than 300 miles from La Paz in a
straight line and, as it is on the east side of the Cordillera, much
farther on horseback, which until lately was the only way to journey
thither, except for some miles of staging, Sucre is now more
accessible, as will be indicated later.
Tarija, south of Chuquisaca and southwest of Potosí, is with
Potosí the most southern of the Departments, now smaller than
formerly, since the Gran Chaco has been made a separate Territory.
Like Chuquisaca it is on the east side of the Cordillera, the greatest
altitude being 12,874 feet. The climate obviously varies from
temperate on the west to hot in the lowlands, the productions varying
similarly. There are possibilities for all varieties of grain and fruit, as
well as for stock raising; the western part lies in the silver region,
having gold and copper as well.
The capital, Tarija, 80 miles east of Tupiza, is 250 miles from
Sucre and 600 by rail from La Paz, but a few days must still be spent
in stage or on horseback. Prospects for speedy rail connection are
good when rapid development will follow.
Santa Cruz, north of Chuquisaca and east of Cochabamba is the
largest Department of all, and the richest in wild animals and
vegetation. This Department, altitude 377-9754 feet, does not reach
the great mountains, but far to the east it has a little range of its own,
the highest summits of which are just below 4000 feet. Here are
forests and prairies, cattle, tropical fruit and vegetables; cotton grows
profusely, rice, manioc, coffee, copal, rubber, what you will. Here too
is petroleum, iron, quicksilver. The white population is said to be of
purer Spanish blood than in most of the Departments, comparatively
unmixed; but as the climate is hot the people are naturally indolent.
The capital, Santa Cruz, altitude 1400 feet, 550 miles from La Paz
by way of Cochabamba, and 330 from Sucre, was founded, only
think! in 1545. It is on two of the main routes running from Bolivia to
Argentina and Paraguay. With a population of 20,000, it is quite a
city, having a variety of local industries, and being on a favored route
for the export of rubber, sugar, and coffee to Europe; but the
enormous cost and difficulty of transportation retards its progress.
Three hundred and ninety miles east near the Paraguay River is
Puerto Suarez, two hours or 11 miles by cart road from Corumbá,
Brazil, to which point steamers come from Buenos Aires. A railway
too practically reaches Corumbá from São Paulo.
El Beni, the last of the Departments, occupies the northeast
portion of Bolivia, being divided from La Paz by the Beni River. With
an area of 100,000 square miles the variation in altitude is slight,
465-2800 feet; hence everywhere a hot climate. It has hardly 40,000
inhabitants, mostly Indians. Nearly all the whites are engaged in the
rubber industry, though stock raising and agriculture are slightly
practised in the higher sections.
The capital, Trinidad, 400 miles from La Paz, with a population of
5000, is the least important of the various Department capitals,
though containing many stores and trading concerns for the sale of
imported products. Some smaller places are of greater commercial
consequence and better known, on account of their location.
Colonias. Of the Territories of Bolivia Colonias of the Northwest is
west of El Beni, and north of La Paz. Since the settlement of the
boundary dispute between Peru and Bolivia, the area of the territory
is given as 81,600 miles. It is similar in character to El Beni, a great
wilderness, full of water courses and forests, a land of wonderful
vegetation; population 40,000, besides 15,000-20,000 wild Indians,
not to mention animals, mostly wild, gorgeous birds, snakes and
insects. The swamp fevers, it is said, may be avoided by healthy
persons who use proper precautions and live on the highlands where
forest clearings permit the free course of the wind. Rain falls from
December to May, but May, June, and July are agreeable, with night
temperatures 60°-70°. In the warmest months the range is 76°-90° in
the hottest part of the day. In September cool south winds bring
sudden changes often producing colds.
The capital, Riberalta, 920 miles from La Paz, at the confluence of
the Madre de Dios and the Beni, stands on a high bank 65 feet
above the water, which the main street faces, an unusually desirable
location. Its population, perhaps 4000, is increasing rapidly with the
number of trading houses and rubber exporters who own small
steamers for river traffic.
Villa Bella, 125 miles below Riberalta, at the junction of the Beni
and Mamoré, is on comparatively high land, facing the Brazilian
town, Villa Murtinho on the other side of the Mamoré. On account of
the favorable location for export, rubber and timber property in the
vicinity is especially valuable. The climate is better than in many
other sections of the tropical lowlands.
El Oriente. This Territory, uncertain as to area and boundary, is
lowland southeast of Santa Cruz, extending to the Paraguay River,
north of the region claimed by Paraguay.
The capital, Puerto Suarez, is a small place on a bayou of the
Paraguay River facing Corumbá.
Gran Chaco. South of Santa Cruz and southeast of Tarija, of
which it was formerly a part, is the Territory of Gran Chaco, where
the boundary with Paraguay is still undetermined. Fifty-seven
thousand square miles is given as the area. Here are low lying
plains, some covered with water in the rainy season (December to
March), one or two feet deep; also open forests with palms, willows,
cedar, carobs, and quebracho; and hilly sections occupied by wild
Indians.
Yacuiba, formerly the capital, which is now transferred north to
Villa Montes, is a town 200 miles southeast of Tarija, and still the
most important place in the territory. In addition to forest riches, are
those of the plains, now harboring wild cattle, and undoubtedly
suitable for stock raising.
 CHAPTER XXVI
BOLIVIA: PORTS AND TRANSPORTATION

La Paz. Bolivia unhappily has no seaport, but lake and river ports
are better than none; of these she has several. Nevertheless La Paz
is the chief port of entry with the principal Custom House, the
terminus of the three railways leading from the Pacific.

Railways

Route from Mollendo. In 1903 not a single railway reached La

Paz, but not long after, the road, which then left one on the great
plateau, arrived at the edge of the alto, whence by electric power the
cars are brought down the steep walls of the cañon to the
picturesque city. This route is the conclusion of the road from
Mollendo, by all means the most desirable for the traveler coming
from the north. The inconvenience suffered by many persons going
from the coast directly to the high table-land, and the real danger to
those with weak hearts, may be lessened or entirely obviated by
pausing two or three days in the pleasant city of Arequipa. Even the
compulsory stay of a single night may render the journey to Puno
innocuous to the person in good condition. Having crossed the
divide, altitude 14,688 feet, one arrives at nightfall at Puno, on the
shore of Lake Titicaca, altitude 12,500 feet. The boat is there taken
for the night journey of 130 miles to Guaqui, the Bolivian port at the
south end of the lake, a 14 hour sail which no one should miss; for in
the morning an inspiring vision may be had, hardly equalled on the
globe, of the magnificent Cordillera Real, a hundred mile stretch of
rugged snowclad peaks from Illampu to Illimani with Huaina Potosí
halfway. A splendid view of Illampu is had on the left, during the
three hours’ ride, 56 miles, over the almost desert plateau, and after
passing its edge a matchless view of the cañon, the city, 1000 feet
below, and Illimani towering far above.
 Arica-La Paz Railway. The second route is from the port of Arica,
formerly in Peru, but now in the possession of Chile, which country in
accordance with the peace terms with Bolivia of 1883 constructed
the Arica-La Paz Railway as far as the frontier, the most difficult
portion, while Bolivia laid the rails for 145 miles within her own
boundaries. A branch 5 miles long leads to the important mining
town of Corocoro. This is by far the shortest route from the coast,
only 281 miles, but as we have seen, this is a disadvantage to
travelers rather than a benefit; nor is there a gain in time to one
coming from the north, as the steamer arrives at Arica, a day later
than at Mollendo. As one may come up from Mollendo, including the
night at Arequipa, in 48 hours, one thus arrives in La Paz a few
hours before one who travels in 22 hours from Arica. The journey
from Mollendo is 534 miles. For the more rapid climb from Arica
there are 28 miles of rack railway with a maximum grade of six per
cent. The equipment of the road has been inadequate for the
business, but more has been ordered; larger port works at Arica are
needed and expected. The highest point on the railway is 13,986
feet near the Bolivian boundary, nearly 700 feet lower than on the
road from Mollendo; but being reached in less than half as many
hours, and also at night when the effect of the altitude is always
greater, even the most robust are likely to suffer from a headache if
not worse, as is the case with persons coming up from Antofagasta.
Going down by either route there is usually no trouble, except for
those who have been a long time at the higher altitude.
Bolivia-Antofagasta Railway. The journey to La Paz from
Antofagasta, on the Chilian coast 400 miles below Arica, occupies
two days for the 718 miles. This, the earliest of Bolivian railways,
with Oruro for years as the terminus, in 1911 was continued to La
Paz, i.e., to Viacha (15 miles from La Paz), where it meets the road
from Guaqui, the two forming a long link in the Pan American chain.
The later road now has its own line into the city. At Uyuni, where
there is a Custom House and a change of cars, the road turns off
towards Chile. From Ollague on the Chilian border, a 58-mile branch
runs to the Collahuasi copper mines, among the richest in the world.
A Bolivian extension is now being made along the plateau to connect
with the Argentine road from La Quiaca to Buenos Aires. Already
Atocha has been reached; 60 miles remain to Tupiza, and 66 more
to La Quiaca, the terminus of the Argentine railway to the border. An
automobile service is operated in the dry or winter season from
Atocha to La Quiaca, making the journey in eight hours. The rest of
the year one goes on muleback. Thus by the weekly train from La
Quiaca to Buenos Aires in 48 hours, the entire journey, 2000 miles,
may be made from La Paz much more quickly than by way of Chile,
in case one does not wish to go there. This route also serves when
the Trans-Andine is blocked by the winter snow.
Other Argentine connections are planned. A branch railway is to
be built from La Quiaca or possibly from Orán to Tarija, about 82
miles; meanwhile it is expected to construct an automobile road for
the service. A railway extension from the Argentine boundary farther
east, hoped for in the near future, is from Embarcación to Yacuiba,
thence to Santa Cruz, thus opening the cattle region of the Chaco, of
Chuquisaca, and Santa Cruz for the export of cattle through Buenos
Aires.
Several Branches. Farther north important branches to the
interior are already in operation. The first, opened in 1912, was from
the station, Rio Mulato, 130 miles south of Oruro, to far famed
Potosí. This section, 108 miles long, climbs to a height of 15,814
feet, the second highest altitude reached by rail, following the
Morococha branch of the Central Peruvian. The cost was over
$8,000,000, the maximum grade three per cent. The trip occupies
eight hours; there is weekly service. The Oruro-Cochabamba Line
was completed in 1917; its length 132 miles, its cost above
$10,000,000. It crosses the Cordillera at about 14,000 feet. This road
is equally important, opening up a fine agricultural and well
populated district. Its fine scenic attractions should allure the tourist.
Food is very expensive on the plateau, but with improved access to
the warmer sections, lower prices and greater variety of fruits and
vegetables may be expected above.
A branch on the plateau long in operation is a 24 mile line from
Uyuni to the Huanchaca silver mines, the road owned by the Mining
Company. Another branch is now under construction from
Machacamarca to Uncia, 57 miles, built by Simon Patiño, the
wealthy owner of extremely rich tin mines at Uncia. Thirty-eight miles
are in operation.
New Lines. Under construction are two very important lines: one
the Potosí-Sucre Railway, 105 miles, with 40 in operation; the other,
the La Paz-Yungas Line of very difficult construction, passing over
the mountain range near La Paz at an altitude of 15,338 feet and
going down into the deep valley. One hundred and thirty-seven miles
have been surveyed. The passage has been accomplished, the road
opened to Unduavi, 16 miles, from which point it will be continued
down the valley to Coroico, later to Rurenabaque, and some time to
Puerto Brais on the Beni, the latter port being open to navigation all
the year. Already a delightful road for tourist travel and useful for
bringing fruit, vegetables, and other products to the La Paz market,
has been constructed. With the completion of the road to Puerto
Brais, a more speedy route for travel and freight to Europe will be
attained. On account of the scarcity and high cost of coal this road is
operated by electricity, for which the great fall of small streams gives
ample power. On account of transportation difficulties and increased
cost the completion of the Potosí-Sucre Railway has been delayed,
but when this is realized the road will probably be continued
northeast to Santa Cruz. From this city a road to Puerto Suarez on
the Paraguay, now talked of, will come later. Bolivia has about 1400
miles of railway in service. A road to Santa Cruz from Cochabamba
instead of Sucre is advocated by many.

Other Forms of Transportation

Roads. Automobile roads are planned and being constructed to fill

as far as possible present requirements, but in the rainy months,
December to May, traffic is generally suspended. From Cochabamba
to Sucre there is motor service for freight and passenger traffic, a
more direct route than that from Potosí for persons in La Paz. From
Potosí to Sucre a motor bus makes the journey in 12 hours. Another
motor car road soon to be available is from Riberalta to Villa Church
(65 miles), opposite Guajará-Mirím in Brazil, the terminus of the
Madeira-Mamoré Railway. From Cochabamba autos run to Totora,
also from La Paz to Achacachi, and over a spur of Illampu down to
Sorata.
Aside from railways land transportation is chiefly confined to
burros, mules, and llamas; in some places there are mule wagons or
ox-carts, mules generally carrying the traveler. The cart roads,
constructed and maintained by the Government or the Municipality,
are mostly very poor; yet such as they are they extend 2000 miles.
They are often if not usually impassable in the wet season. Llamas
are more extensively employed than in any other country. In the
plateau region with their 100 pound burden they wander leisurely
along, cropping the coarse grass, tended by their Indian owners or
drivers, who are in no more of a hurry than they. Along difficult
mountain slopes are paths constructed by the Indians, skirting
precipices and climbing steep inclines, where llamas, mules, and
burros bear loads of minerals, coca, coffee, or other products or
imports, from one section to another. About 5000 mules and burros
pass over the La Paz-Yungas road daily. Except in the dry season,
May to October, travel either by mule or by stage may anywhere be
difficult or impossible on account of rivers becoming impassable by
reason of freshets or roads because of mud. The importance of
proper packing according to directions of goods which may be
shipped to the interior, or indeed to any destination, cannot be
overestimated.
Lake Transport. Aside from the railways transportation in Bolivia
cannot be called excellent either on land or water, though there is fair
steamboat service on Lake Titicaca. This with Lake Aullagas (Poopo)
and the Desaguadero River are the only waterways of the table-land
where navigation exists. On the larger lake the Railway Company
has steamers of 1500 tons or less, five in all, connecting with the
trains, and some touching at small ports on the islands and around
the Lake. Other boats are used, among them Indian balsas, built
entirely of reeds, sails and all. In the use of these the Indians are
very expert. The Desaguadero River is navigable to Lake Poopo for
steamers of 500 tons. Poopo is navigable for boats of shallow draft.
River Transport. Of great and increasing importance are the
rivers of the eastern slope, both those of the Amazon Basin, and to a
less degree those of the Paraguay, which river for some distance is
practically the eastern boundary of Bolivia. Its chief tributaries are
the Pilcomayo and the Bermejo. Bolivian statistics give the total
length of all their navigable rivers as nearly 12,000 miles, but these
figures include sections not conceded as within the boundaries of the
Republic as well as rivers on the boundary. Half of the distance
mentioned, 6000 miles, is said to be navigable by steamers of 4-6
feet draft.
In the Amazon Basin steamboat navigation is possible on the Acre
from the Upper Purús, the Abuna, Orton, Madre de Dios, Beni,
Madidi, Mamoré, Itenez or Guaporé Rivers, with some of their
tributaries. Of the above, the Beni and the Mamoré are the most
important. On the latter, regular steam service has recently been
inaugurated from Cuatro Ojos in Santa Cruz to Guajará-Mirím (a
distance of nearly 1000 miles), the terminus of the Madeira-Mamoré
Railway. From the Department Cochabamba, boats descend on the
rivers Chaparé, Chimoré, and Iabaro to the Mamoré. The Guaporé,
though easily navigable for 1000 miles, flowing through an
uninhabited region, is little used.
Although most of the streams are more or less obstructed by
rapids, rocks, and masses of tree trunks, which last might easily be
cleared away, they are nevertheless useful for navigation. In spite of
difficulties and dangers much transport with canoes and rafts is
effected by the Indians, who are expert boatmen. On the Beni, rafts
are used from Puerto Pando 155 miles to Rurenabaque; from there
steamboats serve 573 miles to Riberalta; the Esperanza Rapids
below the town making their further passage impracticable. Hence
the journey to Villa Bella opposite Villa Murtinho on the Madeira-
Mamoré Railway or to Villa Church, higher up, opposite Guajará-
Mirím must be made by land; therefore the motor road is being
constructed.
In high water the Madre de Dios is navigable by steamers far up
its tributaries, at other times by callapos which consist of two or three
balsas fastened side by side. The balsa here is a small raft made of
several logs. It is usually 22 to 26 feet long, 5-6 wide, carries 750
pounds, and is managed by three boatmen. A callapo or monteria
may carry 3400 pounds with a crew of from 3 to 15 highly skilful
men. They make 9 or 10 miles daily, navigating 10-12 hours. With
the return of normal conditions the development of this region will be
hastened, and greater benefits will be realized from the opening of
the Madeira-Mamoré Railway which occurred in 1912.
 CHAPTER XXVII
BOLIVIA: RESOURCES AND INDUSTRIES

Mining

The mining industry, at present the most important in Bolivia, is

likely to continue the leader for an indefinite period, although with
easy communication and large population in the lower districts it may
ultimately have a rival in forestal and agricultural products, or in
cattle, certain to be at least a very valuable adjunct. The mineral
riches of Bolivia may equal if they do not surpass those of Peru,
though except for silver and tin they are probably less known. They
include almost every variety of the precious metals, with others not
so classed. To mention a few of these, there are gold, silver, tin,
copper, bismuth, lead, antimony, tungsten, platinum, zinc, petroleum,
with fine marbles, alabaster, malachite, opals, emeralds, jasper,
borax, salt, etc. There are thousands of known lodes, but
comparatively few are worked. The statement of the scientist,
Raimondi, that the plateau of Bolivia is a table of silver supported on
columns of gold is declared by Walle to be no exaggeration. A few of
the mining belts where some work has been carried on will be
mentioned.
Gold. At the present time little is done in the way of gold mining,
tin and copper being more fashionable. Another reason for inactivity
in this line is that the Bolivians possess the majority of the more
favorably situated holdings, which they refuse to part with except at
prohibitive prices or possibly at all, hoping some day themselves to
be able to operate them. It is true that in the early colonial days
enormous quantities of gold and silver both were produced and
exported, although obtained by crude mining methods. It is thought
that with modern machinery excellent if not better results may yet be
obtained. In the 210 years previous to 1750 more than 12 billion
dollars worth of gold was produced in the country. There has been a
falling off since then, but one family in the last century obtained over
$3,000,000 from their property. Between 1868 and 1900 over
$120,000,000 is believed to have been produced. In many sections
gold is known to exist in abundance and with further exploration it
will doubtless be discovered in others. It is found in alluvial deposits,
also in veins or lodes of quartz, from which the deposits are the
washings. Veins of antimony which are common in Bolivia contain
gold in chemical or mechanical mixture.
There are three regions where gold is found, all of considerable
extent: the first and best known crosses the provinces of the
Department of La Paz, chiefly on the east slopes of the Cordillera
Real, continues through that of Cochabamba and runs out in Santa
Cruz towards the Rio Paraguay. In La Paz are the well known
deposits of the Tipuani River, of Chuquiaguillo, and many others.
The second region begins in the southwest corner of the country,
and passing south of Tupiza turns north through Potosí towards
Santa Cruz. The third is in the northwest part of the Republic, joining
the similar section in Peru. Although said to be the richest of all, it is
practically unexploited and unexplored.
While the opinion is held that the lack of means of communication
is all that prevents a large production of gold in Bolivia, I believe that
to Americans this is a smaller drawback than the distance of the
whole country from the United States; this objection will have less
weight in the future. Certainly the hardships of those regions and the
difficulty in reaching many of them is slight indeed in comparison
with the trials experienced by early Alaskan miners. The Tipuani, for
instance, is within a day’s horseback ride from the pretty town of
Sorata, population 8000. Chuquiaguillo, now owned by an American
company, is within easy walking distance of La Paz. Others are more
remote, most of them on the eastern slope of the Cordillera Real, to
be reached over passes of 15,000 feet or more, yet within a few
days’ ride of civilization. The single region of the Tipuani is 120 miles
long.
Besides the opportunities for placer mining there are strata formed
of sand, clay, and stones, from 50 to 330 feet deep. Sometimes in
these there are veins with gold in thin flakes 98 per cent pure. Some
workings are open, others in shafts and galleries like an ordinary
mine. However, difficulty is experienced in taking machinery over the
poor trails, and also in obtaining labor. In May, 1904, a nugget was
found at Chuquiaguillo containing 47 ounces of pure gold, nearly
$1000 worth. One hundred miners are employed here at 50 cents to
$1.50 a day. In the Province of Velasco, Department of Santa Cruz,
is a region considered by Walle more accessible than others, to
which one would come from the Atlantic to Corumbá on the
Paraguay River or perhaps by rail from São Paulo, and enter Bolivia
across the plains. It is thought that Bolivia may become one of the
leading gold producers of the world.
Silver. The silver mines are better known and are now worked on
a larger scale than the gold, although 10,000 lodes are practically
abandoned or operated slightly. The suspension was not because of
a scarcity of silver but for lack of capital, means of transport, and
suitable machinery; also its lower price. A few of the largest mines
have continued to be regularly worked. With the recently higher price
of silver, greater activity has prevailed than a few years earlier. The
richest of the ores contain from 10 to 50 per cent of silver and even
80; more has from 1 to 10 per cent.
The Department of Potosí is world renowned for its silver. In the
first 40 years of colonial production more than $70,000,000 was
taken from the Cerro Potosí, a sugar loaf 16,000 feet above the sea,
in which 5000 mines have been opened. Up to the present time a
billion dollars, one writer says four billion, have been realized from
the silver of Bolivia. In neighboring Provinces of the same
Department, Potosí, are many other mines, most of which now
produce tin as well. The ores of Potosí which originally contained 60
per cent or more of silver are now poorer with more iron pyrites. The
present queen of South American silver mines, the Huanchaca-
Pulacayo, east of Uyuni, consists of a dozen groups over 8500
acres, a bed of fabulous richness. In the 28 years before 1901 it
produced 4250 tons of silver. The majority of the stock is held in
France. Some years ago the Pulacayo mines were flooded with hot
water, but they have now been drained and are again operated. On
the property is a town of 10,000 people, hundreds of whom are
employed in the works, including women who first try out the ore. In
1918 $2,622,000 worth of silver was produced.
Oruro is next to Potosí in production of silver, and in tin it stands
first. The richest silver ores, as at Huanchaca, are so exported, the
poorer are treated on the spot. Other rich mines are well known.
Tin mining is a more recent industry, dating from 1895. Already it
has become the leading export of the country, which now provides
more than one-third of the world’s supply. It was well known to the
Spanish colonists as much of it was mingled with the silver, but they
rejected it as rubbish, using it to fill in depressions. Large profits
have been recovered from these fillings and from the dumps. While
tin is produced in many other parts of the world, the Bolivian lodes
are of unusual extent and richness. Their production is now
exceeded only by that of the Straits Settlements. Since the
development of the tin plate industry and its use in other ways the
demand for tin and the price have enormously increased; from $350
a ton in 1898 to $900 before the war. In 1912 tin was produced,
barilla, 60 per cent pure, 37,700 tons, worth $18,000,000; in 1918
the value of tin exported was $45,364,000.
All along the eastern part of the plateau from Lake Titicaca to the
southern boundary, tin is found in thick unbroken lodes. The three
principal districts are La Paz, Oruro, Potosí. In La Paz the best
known mines are on the slopes of the splendid mountain, Huaina
Potosí in the Cordillera Real, though many other mines are worked.
Those in the Department of Oruro are more important, producing
one-third of the output. The lodes vary in thickness from a few inches
to 10, 15, or 20 feet. Rich pockets are found 30-60 feet in diameter,
and veins with stannic oxide fragments, running from 50 to 100 per
cent. Cassiterite, tin stone, or tin with stannic oxide, is frequent with
55-60 per cent tin; this is sent to Europe as extracted. The
percentage of tin in a lode is very variable, often 6-8 per cent,
sometimes 15. Veins of tin on the Cerro Potosí penetrate the hill
parallel to veins of silver; some are united; in others tin, silver, and
copper alternate or are in union. The tin is generally in the upper part
nearer the surface. Sulphides are found at Oruro, but oxides are
more frequent. All are at an altitude of about 10,000-17,000 feet. The
large Socavón Company employs 1000 hands, has a concentrating
plant, and puts out six tons of silver and 250 tons of tin yearly. Many
other mines export barilla, some of it with 70 per cent of tin. A
Bolivian, Simon J. Patiño, from being an ordinary laborer has risen to
be the Tin King, both banker and mine owner, working his mines by
modern methods and electric machinery. His mines at Uncia are
among the richest in the world, producing about 30 per cent of the
Bolivian total. The Llallagua mines, the richest of all, have still
greater production.
On Huaina Potosí are placers containing with tin, gold in flakes,
bismuth, and oxide of iron; more in other locations. The Company
owning the Chlorolque Mountain and other properties, with tin as the
principal product, mines also silver, bismuth, copper, and wolfram; it
paid a 30 per cent dividend in 1917. Capital only is needed to exploit
on a much larger scale these rich and available deposits. Three
hundred to four hundred dollars a ton hardly paid the cost of
production, but with prices three times as much it is a highly
profitable business. The Guggenheims have recently purchased
three tin mines located at Caracoles on the Quimsa Cruz Range,
south of Illimani, to which they have built an automobile road 150
miles from Eucalyptus, a station on the main line between La Paz
and Oruro. The mines are at an altitude of 16,000 to 18,000 feet, the
quarters at 15,000 to 16,000. The tin runs to 10-15 per cent. It is
thought that the property will rival the Uncia. The International Mining
Company has a tin property in the Yungas Province 52 miles from La
Paz at a height of 8000 feet, a rather more comfortable altitude.
Copper. Of late the copper deposits in Peru and Chile have
received more attention than those of Bolivia, but this country also
has extensive formations from Lipez at the south where white copper
is found, along the East Cordillera, past Potosí and Oruro to
Corocoro in the West, ending in the Nudo of Apolobamba. Veins
occur in all the buttresses of the Andes.
Copper is found chiefly in its native state, the deposits of Corocoro
being especially famous. The copper here occurs in powder, plates,
or nodules, in beds of reddish sandstone. One veta, or mineralized
bed, is 1000 yards deep, extending several miles; another, a branch
formation above this surrounds it on the southwest and east,
occupying a great part of the Desaguadero Valley. Both formations
are arseniates. One branch 2000 yards thick extends 6 miles; in
another direction a branch runs 35 miles from Corocoro to
Chacarilla, then branching south and east. Copper generally occurs
in small irregular grains with from 70 to 92 per cent copper. The
native copper, the great wealth of the mines, here ranges from
microscopic grains to plates and arborescent forms called charquis.
At La Charcarilla large plates are found 3¹⁄₂ inches thick. Ores with
15-20 per cent copper are usually neglected. As there are no good
kilns, barilla is exported 80 per cent pure, a quintal, 220 pounds,
costing $3-4 for production. Except from Corocoro the present output
is insignificant. For a time these mines were abandoned owing to the
low price of copper and the difficulty of transport. It was formerly
necessary to send the barilla by carts, mules, or llamas to the port of
Arica, or else to the Desaguadero River, to be carried in boats to
Lake Titicaca and Puno, thence by rail to Mollendo. With the great
increase in the price of copper and the opening of the Arica-La Paz
Railway with a 5 mile branch to Corocoro, these difficulties were
removed and active work progressed with a broad field for
enlargement. Copper formation here is said to resemble greatly
those in Northern Michigan, such as the Copper Range, and the
Calumet and Hecla. In 1917 over 82,000,000 pounds were exported,
worth above $4,000,000.
Lead is found in the region of La Quiaca and some is exported by
way of Argentina. Galena, an ore of lead and silver, is widespread.
Bismuth. Of less used minerals bismuth may be mentioned, in the
production of which Bolivia leads the world and controls the supply.
At present, chiefly used as a drug on account of its high price, it
might be more largely employed in alloys. Bismuth ore, as
carbonates, oxides, and arsenical sulphates, is so abundant in many
parts of Bolivia that if bismuth sold at 20 cents a pound a fair profit
would be realized. With the price for years $3500 a ton, it would
seem a profitable business to engage in. Over one million pounds
were exported in 1918 worth one and a half million dollars.
 Antimony also is found in large quantities. Nearly 18,000 tons
were exported in 1916, worth approximately $5,000,000, but with
high prices and abnormal demand due to the war. Less than one per
cent of the quantity was exported in 1912. The ores are rich, often
running 70 per cent pure metal.
Zinc in the form of sulphides or blends is abundant, largely in
connection with silver. Such ores are sent to Europe, there to be
separated, the scarcity and almost prohibitive price of coal rendering
such a course desirable for various products, to avoid the excessive
cost of smelting; at the same time the high cost of transportation
limits the export.
Wolfram, an iron-manganese tungstate, is found near veins of tin
and much is exported, the production being especially in the Santa
Cruz district, at the northeast of Cochabamba. Some lodes contain
no less than 50 per cent of tungstic acid, the ores concentrated for
export running higher. About 7¹⁄₂ million pounds were exported in
1918 valued at nearly $3,708,000.
Coal. The lack of good coal is the chief fault in nature’s economy
in Bolivia. Its usual cost in La Paz is $40-$50 a ton, at times $75. It is
therefore used as little as possible, although some is imported. While
Bolivia has been said to possess no coal at all, a lignite of fair quality
has been discovered in the peninsula of Copacabana, Lake Titicaca,
some of which has been used by the Guaqui-La Paz Railway to
develop electric power; but not in serious fashion, experimental
merely. Other seams are found near the Ayoapó station, which a
company has been organized to exploit. Good samples have come
from Tacora on the Arica Railway near the frontier, the coal said to
be equal to that of Coronel in Chile, the seams rich and thick. If so
their exploitation will immensely advance the industries of the
country. Good coal is reported across the mountains from Sorata
and lignite is found near Cochabamba.
On account of the enormous cost of coal, various substitutes are
in use. On the plateau there is practically no wood. For heating,
foreigners generally employ oil stoves, the natives nothing at all,
though the mercury may range from 40° to 50° in the house on a
winter’s day. In La Paz for cooking purposes a substance called
taquia is employed. It looks like pecan nuts, but is the dung of
llamas, which is collected by Indians on the plateau or elsewhere. It
sells generally at about $4 a ton, but has lately risen to $7. It is the
principal fuel used at the mines of Corocoro. In some places as at
Potosí, yareta is used, a vegetable or shrub which grows in a dense
mass to a height of several feet with a greater diameter. The green
mass has a woody fibre, is highly resinous, and burns like peat,
giving a good amount of heat. It grows at altitudes between 13,000
and 16,000 feet. Another woody shrub, a species of broom called
tola, is used, with another, quenua, which like charcoal gives much
heat with little ash. Extensive peat beds exist near La Paz, estimated
to yield 30,000,000 tons of briquettes with a fuel value about half that
of coal. The discovery and wide exploitation of coal and petroleum
would be an immense boon both to industry and to the
householders.
Petroleum is now believed to exist in large quantities, sufficient to
solve the fuel problem. The best known fields are in the southeast
where seepages occur and a belt is indicated extending northwest
and southeast 150 miles from the Argentine border. Concessions
covering millions of acres have been granted to American
companies. The zone traverses the Provinces of Santa Cruz,
Chuquisaca, and Tarija. Geological investigations indicate that a
petroleum belt or basin exists along the foothills of the east side of
the East Cordillera. There are indications in the Beni district a
considerable distance north of La Paz, and beyond the Madidi in
Colonias. A probable thickness of over-lying strata in southeast
Bolivia is 160-660 feet. If so development costs will be diminished.
The great difficulty is the inaccessibility of the locations. On this
account few wells have yet been sunk. To pipe the oil up to the
plateau would be an expensive, some say an impracticable,
proposition. At present approached only by cart roads or caminos,
the fields will be rendered more accessible by the railways planned
from Puerto Suarez or Pacheco and from Argentina to Santa Cruz.
The oils vary in quality, some having an asphalt base with 4 per cent
of gasoline, up to .81 specific gravity with 40 per cent of gasoline.
The heavier is in the lower sands. Along the eastern base of the