Operations Research and Enterprise

Systems Third International Conference

ICORES 2014 Angers France March 6 8
2014 Revised Selected Papers 1st
Edition Eric Pinson
Visit to download the full and correct content document:
https://textbookfull.com/product/operations-research-and-enterprise-systems-third-int
ernational-conference-icores-2014-angers-france-march-6-8-2014-revised-selected-p
apers-1st-edition-eric-pinson/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Operations Research and Enterprise Systems 5th

International Conference ICORES 2016 Rome Italy
February 23 25 2016 Revised Selected Papers 1st Edition
Begoña Vitoriano
https://textbookfull.com/product/operations-research-and-
enterprise-systems-5th-international-conference-icores-2016-rome-
italy-february-23-25-2016-revised-selected-papers-1st-edition-
begona-vitoriano/

Operations Research and Enterprise Systems 8th

International Conference ICORES 2019 Prague Czech
Republic February 19 21 2019 Revised Selected Papers
Greg H. Parlier
https://textbookfull.com/product/operations-research-and-
enterprise-systems-8th-international-conference-
icores-2019-prague-czech-republic-february-19-21-2019-revised-
selected-papers-greg-h-parlier/

Engineering Multi Agent Systems Second International

Workshop EMAS 2014 Paris France May 5 6 2014 Revised
Selected Papers 1st Edition Fabiano Dalpiaz

https://textbookfull.com/product/engineering-multi-agent-systems-
second-international-workshop-emas-2014-paris-france-
may-5-6-2014-revised-selected-papers-1st-edition-fabiano-dalpiaz/

Formal Techniques for Safety Critical Systems Third

International Workshop FTSCS 2014 Luxembourg November 6
7 2014 Revised Selected Papers 1st Edition Cyrille
Artho
https://textbookfull.com/product/formal-techniques-for-safety-
critical-systems-third-international-workshop-
ftscs-2014-luxembourg-november-6-7-2014-revised-selected-
Multi Agent Based Simulation XV International Workshop
MABS 2014 Paris France May 5 6 2014 Revised Selected
Papers 1st Edition Francisco Grimaldo

https://textbookfull.com/product/multi-agent-based-simulation-xv-
international-workshop-mabs-2014-paris-france-
may-5-6-2014-revised-selected-papers-1st-edition-francisco-
grimaldo/

Swarm Intelligence Based Optimization First

International Conference ICSIBO 2014 Mulhouse France
May 13 14 2014 Revised Selected Papers 1st Edition
Patrick Siarry
https://textbookfull.com/product/swarm-intelligence-based-
optimization-first-international-conference-icsibo-2014-mulhouse-
france-may-13-14-2014-revised-selected-papers-1st-edition-
patrick-siarry/

Operations Research Proceedings 2014 Selected Papers of

the Annual International Conference of the German
Operations Research Society GOR RWTH Aachen University
Germany September 2 5 2014 1st Edition Marco Lübbecke
https://textbookfull.com/product/operations-research-
proceedings-2014-selected-papers-of-the-annual-international-
conference-of-the-german-operations-research-society-gor-rwth-
aachen-university-germany-september-2-5-2014-1st-edition/

Modelling and Simulation for Autonomous Systems First

International Workshop MESAS 2014 Rome Italy May 5 6
2014 Revised Selected Papers 1st Edition Jan Hodicky
(Eds.)
https://textbookfull.com/product/modelling-and-simulation-for-
autonomous-systems-first-international-workshop-mesas-2014-rome-
italy-may-5-6-2014-revised-selected-papers-1st-edition-jan-
hodicky-eds/

Scalable Information Systems 5th International

Conference INFOSCALE 2014 Seoul South Korea September
25 26 2014 Revised Selected Papers 1st Edition Jason J.
Jung
https://textbookfull.com/product/scalable-information-
systems-5th-international-conference-infoscale-2014-seoul-south-
korea-september-25-26-2014-revised-selected-papers-1st-edition-
Eric Pinson
Fernando Valente
Begoña Vitoriano (Eds.)

Communications in Computer and Information Science 509

Operations Research
and Enterprise Systems
Third International Conference, ICORES 2014
Angers, France, March 6–8, 2014
Revised Selected Papers

123
Communications
in Computer and Information Science 509

Editorial Board
Simone Diniz Junqueira Barbosa
Pontifical Catholic University of Rio de Janeiro (PUC-Rio),
Rio de Janeiro, Brazil
Phoebe Chen
La Trobe University, Melbourne, Australia
Alfredo Cuzzocrea
ICAR-CNR and University of Calabria, Cosenza, Italy
Xiaoyong Du
Renmin University of China, Beijing, China
Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Orhun Kara
TÜBİTAK BİLGEM and Middle East Technical University, Ankara, Turkey
Igor Kotenko
St. Petersburg Institute for Informatics and Automation
of the Russian Academy of Sciences, St. Petersburg, Russia
Krishna M. Sivalingam
Indian Institute of Technology Madras, Chennai, India
Dominik Ślęzak
University of Warsaw and Infobright, Warsaw, Poland
Takashi Washio
Osaka University, Osaka, Japan
Xiaokang Yang
Shanghai Jiao Tong University, Shangai, China
More information about this series at http://www.springer.com/series/7899
Eric Pinson Fernando Valente
•

Begoña Vitoriano (Eds.)

Operations Research
and Enterprise Systems
Third International Conference, ICORES 2014
Angers, France, March 6–8, 2014
Revised Selected Papers

123
Editors
Eric Pinson Begoña Vitoriano
IMA, LISA Complutense University
Angers Madrid
France Spain
Fernando Valente
Polytechnic Institute of Setúbal
Setúbal
Portugal

ISSN 1865-0929 ISSN 1865-0937 (electronic)

Communications in Computer and Information Science
ISBN 978-3-319-17508-9 ISBN 978-3-319-17509-6 (eBook)
DOI 10.1007/978-3-319-17509-6

Library of Congress Control Number: 2015937937

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015
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.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
 Preface

The present book includes extended and revised versions of a set of selected papers from
the Third International Conference on Operations Research and Enterprise Systems
(ICORES 2014), held in ESEO, Angers, Loire Valley, France, during March 6–8, 2014,
which was organized by the Institute for Systems and Technologies of Information,
Control, and Communication (INSTICC) and held in cooperation with the ACM Special
Interest Group on Applied Computing (ACM SIGAPP). ICORES is also technically
cosponsored by the Portuguese Association of Operational Research (Apdio).
The purpose of the International Conference on Operations Research and Enterprise
Systems (ICORES) is to bring together researchers, engineers, and practitioners
interested in the theory and applications in the advances and applications in the field of
operations research. Two simultaneous tracks were held, covering on one side domain-
independent methodologies and technologies and on the other side practical work
developed in specific application areas.
ICORE 2014 received 96 paper submissions from 31 countries, in all continents. To
evaluate each submission, a double-blind paper review was performed by the Program
Committee, whose members are highly qualified researchers in ICORES topic areas.
Based on the classifications provided, only 38 papers were selected for oral presentation
(20 full papers and 18 short papers) and 21 papers were selected for poster presentation.
The full paper acceptance ratio was about 21% and the total oral acceptance ratio
(including full papers and short papers) was 40%. These strict acceptance ratios show the
intention to preserve a high-quality forum which we expect to develop further next year.
We would like to highlight that ICORES 2014 included also three plenary keynote
lectures given by internationally distinguished researchers, namely: Nicolas Zufferey
(University of Geneva, Switzerland), Marie-Christine Costa (ENSTA - ParisTech,
France), and Paul Williams (London School of Economics, UK).
We would like to express our appreciation to all of them and in particular to those
who took the time to contribute with a paper to this book.
We must thank the authors, whose research and development efforts are recorded
here. We also thank the keynote speakers for their invaluable contribution and for
taking the time to synthesize and prepare their talks. Finally, special thanks to all the
members of the INSTICC team, whose collaboration was fundamental for the success
of this conference.

October 2014 Eric Pinson

Fernando Valente
Begoña Vitoriano
 Organization

Conference Co-chairs
Eric Pinson IMA, LISA, Angers, France
Fernando Valente Polytechnic Institute of Setúbal, Portugal

Program Chair
Begoña Vitoriano Complutense University, Spain

Organizing Committee
Helder Coelhas INSTICC, Portugal
Vera Coelho INSTICC, Portugal
Ana Guerreiro INSTICC, Portugal
André Lista INSTICC, Portugal
Andreia Moita INSTICC, Portugal
Raquel Pedrosa INSTICC, Portugal
Vitor Pedrosa INSTICC, Portugal
Cláudia Pinto INSTICC, Portugal
Susana Ribeiro INSTICC, Portugal
Sara Santiago INSTICC, Portugal
Mara Silva INSTICC, Portugal
José Varela INSTICC, Portugal
Pedro Varela INSTICC, Portugal

Program Committee
El-Houssaine Aghezzaf Ghent University, Faculty of Engineering
and Architecture, Belgium
Javier Alcaraz Universidad Miguel Hernández de Elche, Spain
Maria Teresa Almeida ISEG, UTL, Portugal
Lionel Amodeo University of Technology of Troyes, France
Ronald Askin Arizona State University, USA
Lyes Benyoucef Aix-Marseille University, France
Jean-Charles Billaut École Polytechnique de l’Université
François-Rabelais de Tours, France
Christian Blum IKERBASQUE and University of the Basque
Country, Spain
Ralf Borndörfer Zuse Institute Berlin, Germany
Endre Boros Rutgers University, USA
VIII Organization

Ahmed Bufardi École Polytechnique Fédérale de Lausanne,

Switzerland
Alfonso Mateos Caballero Universidad Politécnica de Madrid, Spain
Jacques Carlier Université de Technologie de Compiègne, France
José Manuel Vasconcelos Universidade do Minho, Portugal
Valério de Carvalho
Bo Chen University of Warwick, UK
John Chinneck Carleton University, Canada
Mikael Collan Lappeenranta University of Technology, Finland
Xavier Delorme Ecole Nationale Supérieure des mines de
Saint-Etienne, France
Marc Demange ESSEC Business School, France
Clarisse Dhaenens French National Institute for Research in Computer
Science and Control, France
Tadashi Dohi Hiroshima University, Japan
Nikolai Dokuchaev Curtin University, Australia
Christophe Duhamel Université Blaise Pascal, Clermont-Ferrand II,
France
Pankaj Dutta Indian Institute of Technology Bombay, India
Gintautas Dzemyda Vilnius University, Lithuania
Andrew Eberhard RMIT University, Australia
Ali Emrouznejad Aston University, UK
Nesim Erkip Bilkent University, Turkey
Gerd Finke Grenoble Institute of Technology G-SCOP
Laboratory, France
Jörg Fliege University of Southampton, UK
Bernard Fortz Université libre de Bruxelles, Belgium
Robert Fullér Óbuda University, Hungary
Heng-Soon Gan The University of Melbourne, Australia
Michel Gendreau École Polytechnique de Montréal, Canada
Giorgio Gnecco IMT - Institute for Advanced Studies - Lucca, Italy
Juan José Salazar Gonzalez Universidad de La Laguna, Spain
Armando Guarnaschelli INGAR CONICET-UTN, Argentina
Christelle Guéret University of Angers, France
Nalan Gulpinar The University of Warwick, UK
Jin-Kao Hao University of Angers, France
Mhand Hifi EPROAD EA 4669, University of Picardie Jules
Verne, France
Han Hoogeveen Universiteit Utrecht, The Netherlands
Johann Hurink University of Twente, The Netherlands
Josef Jablonsky University of Economics, Prague, Czech Republic
Joanna Józefowska Poznań University of Technology, Poland
Joaquim Júdice Instituto Telecomunicações, Portugal
Hans Kellerer University of Graz, Austria
Tamás Kis Hungarian Academy of Sciences, Hungary
Jesuk Ko Gwangju University, Korea
 Organization IX

Pradeep Kumar Indian Institute of Technology Roorkee, India

Philippe Lacomme Université Blaise Pascal, Clermont-Ferrand II,
France
Sotiria Lampoudi Liquid Robotics Inc., USA
Dario Landa-Silva University of Nottingham, UK
Pierre L’Ecuyer Universite de Montreal, Canada
Janny Leung The Chinese University of Hong Kong,
Hong Kong
Benjamin Lev Drexel University, USA
Shan Li Zicklin School of Business, Baruch College
(City University of New York), USA
Abdel Lisser The University of Paris-Sud 11, France
Pierre Lopez LAAS-CNRS, Université de Toulouse, France
Helena Ramalhinho Lourenço Universitat Pompeu Fabra, Spain
Prabhat Mahanti University of New Brunswick, Canada
Viliam Makis University of Toronto, Canada
Arnaud Malapert Université Nice Sophia Antipolis CNRS, France
Patrice Marcotte Université de Montréal, Canada
Concepción Maroto Universidad Politécnica de Valencia, Spain
Pedro Coimbra Martins Polytechnic Institute of Coimbra, Portugal
Ana Meca Universidad Miguel Hernández de Elche, Spain
Nimrod Megiddo IBM Almaden Research Center, USA
Marta Mesquita Universidade de Lisboa, Portugal
Michele Monaci Università degli Studi di Padova, Italy
Jairo R. Montoya-Torres Universidad de La Sabana, Colombia
Young Moon Syracuse University, USA
Dolores Romero Morales University of Oxford, UK
José Oliveira Universidade do Minho, Portugal
Linet Özdamar Yeditepe University, Turkey
Selin Özpeynirci Izmir University of Economics, Turkey
Sophie Parragh University of Vienna, Austria
Vangelis Paschos University Paris-Dauphine, France
Ulrich Pferschy University of Graz, Austria
Cynthia A. Phillips Sandia National Laboratories, USA
Diogo Pinheiro Brooklyn College of the City University
of New York, USA
Eric Pinson IMA, LISA, Angers, France
Selwyn Piramuthu University of Florida, USA
Jan Platos VSB - Technical University of Ostrava,
Czech Republic
Caroline Prodhon Charles Delaunay Institute, France
Günther Raidl Vienna University of Technology, Austria
Andres Ramos Universidad Pontificia Comillas, Spain
Celso Ribeiro Universidade Federal Fluminense, Brazil
Andre Rossi Université de Bretagne-Sud, France
Stefan Ruzika University of Koblenz and Landau, Germany
X Organization

Ahti Salo Aalto University, Finland

Marcello Sanguineti University of Genoa, Italy
Cem Saydam University of North Carolina at Charlotte, USA
Marc Sevaux Université de Bretagne-Sud, France
Patrick Siarry Paris-Est University, France
Francis Sourd SNCF, France
Kathryn Stecke University of Texas at Dallas, USA
Vadim Strijov Computer Center of the Russian Academy
of Sciences, Russian Federation
Thomas Stützle Université libre de Bruxelles, Belgium
David Sundaram University of Auckland, New Zealand
Jacques Teghem Faculté Polytechnique de Mons, Belgium
Alexis Tsoukiàs CNRS, France
Alexander Vasin Lomonosov Moscow State University,
Russian Federation
Begoña Vitoriano Complutense University, Spain
Maria Vlasiou Eindhoven University of Technology,
The Netherlands
Hsiao-Fan Wang National Tsing Hua University, Taiwan
Ling Wang Tsinghua University, China
Gerhard-Wilhelm Weber Middle East Technical University, Turkey
Dominique de Werra École Polytechnique Fédérale de Lausanne,
Switzerland
Marino Widmer University of Fribourg, Switzerland
Margaret Wiecek Clemson University, USA
Gerhard Woeginger Eindhoven University of Technology,
The Netherlands
Yiqiang Zhao Carleton University, Canada
Sanming Zhou University of Melbourne, Australia
Konstantinos Zografos Lancaster University Management School, UK

Auxiliary Reviewer
Pasi Luukka Lappeenranta University of Technology, Finland

Invited Speakers
Nicolas Zufferey University of Geneva, Switzerland
Marie-Christine Costa ENSTA - ParisTech, France
Paul Williams London School of Economics, UK
Jean Philippe Vial University of Geneva, Switzerland
 Contents

Invited Paper

Learning Tabu Search for Combinatorial Optimization. . . . . . . . . . . . . . . . . 3

Nicolas Zufferey and David Schindl

Methodologies and Technologies

Four Serious Problems and New Facts of the Discriminant Analysis . . . . . . . 15

Shuichi Shinmura

Simulated Annealing Algorithm for Job Shop Scheduling on Reliable

Real-Time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Daniil A. Zorin and Valery A. Kostenko

Nash Equilibria for Multi-agent Network Flow with Controllable Capacities . . . 47

Nadia Chaabane Fakhfakh, Cyril Briand, and Marie-José Huguet

A Fuzzy Model for Selecting Safeguards to Reduce Risks

in Information Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
E. Vicente, A. Mateos, and A. Jiménez-Martin

A Performance Improvement and Management Model for Small

and Medium Sized Enterprises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Madani Alomar and Zbigniew J. Pasek

Adjacency Variables Formulation for the Minimum Linear

Arrangement Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Serigne Gueye, Sophie Michel, and Mahdi Moeini

A Vessel Scheduling Problem with Special Cases . . . . . . . . . . . . . . . . . . . . 108

Selim Bora, Endre Boros, Lei Lei, W. Art Chovalitwongse, Gino J. Lim,
and Hamid R. Parsaei

Archimedean Copulas in Joint Chance-Constrained Programming . . . . . . . . . 126

Michal Houda and Abdel Lisser

Comparison of Stochastic Programming Approaches for Staffing

and Scheduling Call Centers with Uncertain Demand Forecasts . . . . . . . . . . 140
Mathilde Excoffier, Céline Gicquel, Oualid Jouini, and Abdel Lisser

A New Look at the Covariance Matrix Estimation in Evolution Strategies . . . 157

Silja Meyer-Nieberg and E. Kropat
XII Contents

A Relax and Fix Approach to Solve the Fixed Charge Network Design
Problem with User-Optimal Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Pedro Henrique González, Luidi Gelabert Simonetti,
Carlos Alberto de Jesus Martinhon, Edcarllos Santos,
and Philippe Yves Paul Michelon

Analysis of a Downward Substitution Strategy

in a Manufacturing/Remanufacturing System . . . . . . . . . . . . . . . . . . . . . . . 186
Fethullah Gocer, S. Sebnem Ahiska, and Russell E. King

Risk Tolerance Evaluation for an Oil and Gas Company

Using a Multi-criteria Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
António Quintino, João Carlos Lourenço, and Margarida Catalão-Lopes

A Mathematical Programming Model for the Real Time Traffic Management

of Railway Networks Under Disturbances . . . . . . . . . . . . . . . . . . . . . . . . . 215
Astrid Piconese, Thomas Bourdeaud’huy, Mariagrazia Dotoli,
and Slim Hammadi

Applications

Solving to Optimality a Discrete Lot-Sizing Problem Thanks to Multi-product

Multi-period Valid Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Céline Gicquel and Michel Minoux

A Quadratic Knapsack Model for Optimizing the Media Mix

of a Promotional Campaign . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
Ulrich Pferschy, Joachim Schauer, and Gerhild Maier

A Decomposition Approach to Solve Large-Scale Network Design Problems

in Cylinder Gas Distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Tejinder Pal Singh, Nicoleta Neagu, Michele Quattrone, and Philippe Briet

Elaboration of General Lower Bounds for the Total Completion Time

in Flowshop Scheduling Problems through MaxPlus Approach . . . . . . . . . . . 285
Nhat Vinh Vo, Pauline Fouillet, and Christophe Lenté

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301

Invited Paper
 Learning Tabu Search for Combinatorial
Optimization

Nicolas Zufferey1(B) and David Schindl2

1
Geneva School of Economics and Management, GSEM, University of Geneva,
Blvd du Pont-d’Arve 40, 1211 Geneva, Switzerland
[email protected]
2
Geneva School of Business Administration, Rte de Drize 7,
1227 Carouge, Switzerland
[email protected]

Abstract. In this paper, a new type of local search algorithm is pro-

posed, called Learning Tabu Search and denoted LTS. It is assumed that
any solution of the considered problem can be represented with a list
of characteristics. LTS involves a learning process relying on a trail sys-
tem. The trail system is based on the idea that if some combinations of
characteristics often belong to good solutions during the search process,
such combinations of characteristics should be favored when generating
new solutions. It will be showed that LTS obtained promising results on
a refueling problem in a railway network.

Keywords: Tabu search · Combinatorial optimization · Learning process

1 Introduction
As exposed in [9], modern methods for solving complex optimization problems
are often divided into exact methods and metaheuristic methods. An exact
method guarantees that an optimal solution will be obtained in a finite amount
of time. Among the exact methods are branch-and-bound, dynamic program-
ming, Lagrangian relaxation based methods, and linear and integer program-
ming based methods [6]. However, for a large number of applications and most
real-life optimization problems, which are typically NP-hard [3], such methods
need a prohibitive amount of time to find an optimal solution. For these difficult
problems, it is preferable to quickly find a satisfying solution. If solution quality
is not a dominant concern, then a simple heuristic can be employed, while if
quality occupies a more critical role, then a more advanced metaheuristic proce-
dure is warranted. There are mainly two classes of metaheuristics: local search
and population based methods. The former type of algorithm works on a sin-
gle solution (e.g., descent local search, simulated annealing, tabu search, and
variable neighborhood search), while the latter makes a population of solutions
evolve (e.g., genetic algorithms, scatter search, ant colonies, adaptive memory
algorithms). At each iteration of a local search, a neighbor solution is generated

c Springer International Publishing Switzerland 2015
E. Pinson et al. (Eds.): ICORES 2014, CCIS 509, pp. 3–11, 2015.
DOI: 10.1007/978-3-319-17509-6 1
4 N. Zufferey and D. Schindl

from the current solution by performing a modification on the current solu-

tion, called a move. The reader interested in a recent book on metaheuristics is
referred to [4].
In this work is proposed a generalized version of the learning tabu search
(LTS) proposed in [8]. LTS is a new type of local search algorithm, easy to
adapt within the tabu search framework. It is assumed that any solution of the
considered problem can be represented with a list of characteristics. LTS involves
a learning process relying on a trail system. The trail system is based on the
idea that if some combinations of characteristics often belong to good solutions
during the search process, such combinations of characteristics should be favored
when generating new solutions.
This paper is organized as follows. In Sect. 2 is proposed and discussed a
generic version of LTS. In Sect. 3, a successful adaptation of LTS is described for
a refueling problem in a railway network. The paper ends up with a conclusion in
Sect. 4, where LTS is positioned according to the main existing metaheuristics.

2 Learning Tabu Search (LTS)

In this section are formally presented the general framework of local search
techniques, the descent local search, tabu search, and LTS, which is a tabu
search involving a learning process relying on a trail system.
Let f be an objective function which has to be minimized. At each step of
a local search, a neighbor solution s is generated from the current solution s by
performing a specific modification on s, called a move. All solutions obtained
from s by performing a move are called neighbor solutions of s. The set of all
the neighbor solutions of s is denoted N (s). First, a local search needs an ini-
tial solution s0 as input. Then, the algorithm generates a sequence of solutions
s1 , s2 , . . . in the search space such that sr+1 ∈ N (sr ). The process is stopped for
example when an optimal solution is found (if it is known), or when a time limit
is reached. Famous local search algorithms are: the descent method (where at
each step, the best move is performed and the process stops when a local opti-
mum is reached), simulated annealing, variable neighborhood search, and tabu
search. Note that to escape from a local optimum, in tabu search, when a move
is performed from a current solution sr to a neighbor solution sr+1 , it is forbid-
den to perform the reverse of that move during tab (parameter) iterations. Such
forbidden moves are called tabu moves. Formally, the solution sr+1 is computed
as sr+1 = arg min 
f (s), where N  (s) is a subset of N (s) containing solutions
s∈N (sr )
which can be obtained from s by performing a non tabu move. Many variants of
this basic tabu search algorithm can be found in [5]. More generally, the reader
is referred to [4] for a recent book on metaheuristics, and to [9] for guidelines to
efficiently design a metaheuristic.
A tabu search with a learning process relying on a trail system is now pro-
posed and denoted LTS. Let (P) be the problem under consideration. It is
assumed that C = {c1 , c2 , . . . , cN } is the set of all the possible characteristics
(i.e. specific features or attributes) of solutions of (P). Further, it is assumed that
 Learning Tabu Search for Combinatorial Optimization 5

any solution of (P) with n characteristics can be denoted s = {c(1) , c(2) , . . . , c(n) },
where c(i) is the ith characteristic of solution s. Therefore, with each solution
s can be associated a set IN (s) (resp. OU T (s)) of characteristics which belong
(resp. do not belong) to s. If s can be represented with a vector of size n, then c(i)
could simply be the ith component of the vector. At each iteration, in order to
generate a neighbor solution s of the current solution s, a basic move m consists
in one of the following options: (1) add a characteristic c to s (i.e. move c from
OU T (s) to IN (s), which is denoted m(c+ )), (2) drop a characteristic c from s
(i.e. move c from IN (s) to OU T (s), which is denoted m(c− )), (3) switch two
characteristics c and c between IN (s) and OU T (s) (i.e. move c from OU T (s)
to IN (s) and move c from IN (s) to OU T (s), which is denoted m(c ↔ c )).
The straightforward notation s = s + m can thus be used. Then, it is forbidden
(i.e. tabu) to perform the reverse move for tab (parameter) iterations. More pre-
cisely: (1) if m(c+ ) is performed, it is forbidden to perform m(c− ); (2) if m(c− )
is performed, it is forbidden to perform m(c+ ); (3) if m(c ↔ c ) is performed, it
is forbidden to perform m(c ↔ c).
The trail system relies on the idea that if some combinations of characteris-
tics often belong to good solutions during the search process, such combinations
of characteristics should be favored when generating new solutions. If one consid-
ers combinations of k characteristics, the trail tr(c(i1 ) , c(i2 ) , . . . , c(ik ) ) associated
with characteristics c(i1 ) , c(i2 ) , . . . , c(ik ) indicates if it is a good idea to have such k
characteristics together in a solution, according to the observation of the history
of the search.
Let Πkn (s) be the set of all possible combinations of k characteristics of a solu-
tion s containing n > k characteristics. Consider that s = s+m.  The trail T r(s )


associated with solution s can be defined as follows: (1) n
π∈Πk−1 (s) tr(c, π)

if m(c+ ) is considered; (2) − π∈Π n (s) tr(c, π) if m(c− ) is considered;
  k−1
 
(3) π∈Π n (s) tr(c, π) −
k−1 π∈Π n (s) tr(c , π) if m(c ↔ c ) is considered. For
k
instance, one can remark that if T r(s ) associated with a move m(c+ ) is large,
it means that c is in average attracted by the characteristics of the current solu-
tion s (because it was often observed in the past that the combination of c with
k − 1 characteristics of s leads averagely to good solutions).
The updating of the trail values is now discussed. Let ρ ∈ [0, 1] be a para-
meter representing an evaporation coefficient. A cycle of size I (parameter) is
defined as a sequence of I iterations of tabu search. Every I iterations, the
best solution ŝ of the last cycle is used to update the trail system as follows:
tr(c(i1 ) , c(i2 ) , . . . , c(ik ) ) = ρ · tr(c(i1 ) , c(i2 ) , . . . , c(ik ) ) + Δtr(c(i1 ) , c(i2 ) , . . . , c(ik ) ),
where Δtr(c(i1 ) , c(i2 ) , . . . , c(ik ) ) is a reinforcement term proportional to the qual-
ity of ŝ if the k characteristics c(i1 ) , c(i2 ) , . . . , c(ik ) appear in ŝ, and Δtr(c(i1 ) ,
c(i2 ) , . . . , c(ik ) ) = 0 otherwise. An alternative would be to say that Δtr(c(i1 ) ,
c(i2 ) , . . . , c(ik ) ) is proportional to: (1) the number of times that characteris-
tics c(i1 ) , c(i2 ) , . . . , c(ik ) jointly appear in the solutions visited within the cycle,
(2) the average quality of the solutions of the cycle having characteristics
c(i1 ) , c(i2 ) , . . . , c(ik ) together.
6 N. Zufferey and D. Schindl

At each iteration of LTS, it is proposed to first randomly choose a random

set A (resp. D and S) of non tabu neighbor solutions obtained with add (resp.
drop and switch) moves, such that |A| = |D| = |S|. Note that the size of A is
an important and sensitive parameter to tune. Then, let Aq (resp. Dq and Sq )
be the set containing the q (parameter) solutions of A (resp. D and S) with the
largest trail values. The performed move among Aq ∪Dq ∪Sq is then the best one
according to the objective function f of the considered problem. This technique
is particularly relevant if it is cumbersome to evaluate a solution with f , as f is
only used to evaluate a sample of solutions which rank highly according to the
trail function T r (which does not require a lot of computation). In other words,
at each iteration of LTS, the objective function f (or its associated incremental
computation) is only used to evaluate 3 · q good neighbor solutions according to
T r, which is likely to be more promising than to evaluate a random sample of
3 · q neighbor solutions.
In order to help to visit new regions of the solution space, the following
diversification mechanism (denoted DIV) is introduced, relying on a different

Algorithm 1. LTS: Learning Tabu Search.

Initialization

1. construct an initial solution s;

2. set s = s and f  = f (s);
3. set iter = 0 (iteration counter);
4. set DIV as closed (i.e. not active);

While a time limit is not reached, do:

1. from s, generate a set A (resp. D and S) of non tabu neighbor solutions obtained
with add (resp. drop and switch) moves, such that |A| = |D| = |S|;
2. if DIV is closed, identify the sets Aq ⊆ A (resp. Dq ⊆ D and Sq ⊆ S) containing
the solutions with the q largest trail values;
3. if DIV is open, identify the sets Aq ⊆ A (resp. Dq ⊆ D and Sq ⊆ S) containing
the solutions with the q smallest trail values;
4. select the neighbor solution: set s = arg  min f (s );
s ∈Aq ∪Dq ∪Sq
update the current solution: set s = s ;
5.
update the tabu status: the reverse move is forbidden for tab iterations;
6.
update the best encountered solution: if f (s) < f  , set s = s and f  = f (s);
7.
update the iteration counter: set iter = iter + 1;
8.
update the status of DIV:
9.
(a) open DIV if t1 iterations without improving s have been performed;
(b) close DIV if it has been performed during t2 consecutive iterations, or if s
has been improved within the current application of DIV;
10. if (iter mod I) = 0, update the trail system with the best solution ŝ among the
last I iterations;

Return s
 Learning Tabu Search for Combinatorial Optimization 7

use of the trail system. Generally, the trail system is used to favor good moves
which were often performed in the previous cycles. This is also the case for the
trail systems used in the various types of ant algorithms [10]. In contrast, to
diversify the search, it is proposed here to perform good moves which were not
often performed in the previous cycles. More precisely, assuming that a neighbor
solution s can be generated from the current solution s with move m (i.e.,
s = s + m), it is diversifying to perform m if the corresponding T r(s ) is small.
Each iteration of DIV is performed as above, but the set Aq (resp. Dq and Sq )
contains the q solutions of A (resp. D and S) with the smallest trail values.
DIV relies on two sensitive parameters t1 and t2 : it is triggered if t1 iterations
without improving s (the best encountered solution during the search) have
been performed, and it is performed until one of the following conditions is
satisfied: (1) s has been improved; (2) a sequence of t2 iterations of DIV have
been performed.
LTS can now be formulated in Algorithm 1.

3 LTS for a Refueling Problem in a Railway Network

The problem consists in optimizing the refueling costs of a fleet of locomotives
over a railway network [7]. It is assumed that there is only one source of fuel:
fueling trucks, located at yards. A solution of the problem has two important
components: choose the number of trucks contracted at each yard, and determine
the refueling plan of each locomotive (i.e. the quantity of fuel that must be
dispensed into each locomotive at every yard). Such components are respectively
called the truck assignment problem (TAP) and the fuel distribution problem
(FDP) in this paper. The constraints are the following: the capacity of the tank
of each locomotive is limited, as well as the maximum amount of fuel a truck can
provide the same day; a locomotive can not be refueled at its destination yard;
there is a maximum number of times (which is two) a train can stop to be refueled
(excluding the origin); it is forbidden to run out of fuel. The encountered costs
are the weekly operating cost of each fueling truck, the fuel price per gallon
associated with each yard (which can vary from yard to yard because of the
differences in distribution, marketing costs and other factors), and the fixed cost
associated with each refueling. A solution satisfying all the above constraints is
called feasible. The problem consists in finding a feasible solution minimizing the
sum of the costs.
If the number of trucks is known for every yard (i.e., if a solution of the
TAP is provided), it is possible to exactly solve the associated FDP with the
flow algorithm proposed in [7]. To tackle the TAP, various solution methods
have been proposed in [8]: a descent local search DLS, a regular tabu search
TS, and a learning tabu search LTS. The common features of DLS and TS are
the following. From a current solution s, a move simply consists in adding a
contracted truck to a yard (add move), or in removing a contracted truck from
a yard (drop move). When a move is performed, it is evaluated by the use of the
flow algorithm. During the evaluation, all the costs are considered: the refueling
8 N. Zufferey and D. Schindl

costs for the used gallons of fuel, the fixed refueling costs and the contracting
costs of the trucks. At each iteration, |A| add moves and |D| drop moves are
generated (with the tuning |A| = |B| = 10). The resulting descent local search
DLS for the TAP is presented in Algorithm 2. In contrast with DLS, TS does not
stop when it reaches a local optimum, but returns the best encountered solution
within a predefined time limit. When an add (resp. drop) move is performed on
yard yj , it is forbidden to consider yard yj for a drop (resp. add) move for tab
iterations (parameter tuned to 10).

Algorithm 2. DLS: Descent Local Search for the TAP.

Construct an initial solution s;

While a local optimum is not reached, do:

1. in solution s, randomly choose a set D containing yards for which drop moves are
allowed; for any yard yj ∈ D and from s, remove a truck from it, and apply the
flow algorithm to evaluate such a drop candidate move;
2. in solution s, randomly choose a set A of yards; for any yard yj ∈ A and from s,
add a truck to it, and apply the flow algorithm to evaluate such an add candidate
move;
3. from s, perform the best move among the | A ∪ D | above candidate moves, and
rename the resulting solution as s;

In order to adapt LTS to the considered problem, one mainly has to define
what is a characteristic and to set a learning process based on a trail system.
Let x and y be two yards. A characteristic is simply a pair (x, y) of yards such
that there is at least a truck on x and at least a truck on y. The trail tr(x, y)
associated with x and y aims to indicate if it is a good idea to have trucks on
both yards x and y in the same solution.
Every I (parameter tuned to 50) iterations of LTS, such trails are globally
updated with ŝ (the best solution of the cycle) as follows (with ρ tuned to 0.9):
tr(x, y) = ρ · tr(x, y) + Δtr(x, y), where Δtr(x, y) is the number of trucks on x
and y, computed only if ŝ has trucks on both x and y (it is 0 otherwise). A move
can be denoted by (x, s), indicating that a truck is added to or removed from
yard x, in the current solution s.  Let T r(x, s) be its associated trail value. It is
straightforward to set T r(x, s) = y∈s tr(x, y) if it is used as follows. If (x, s) is
an add move (i.e. add a truck to yard x of solution s), among the possible add
moves, it is interesting to select a move with a large T r(x, s) value (because the
history of the search seems to indicate that having trucks on yard x, as well as
on the yards which already contain trucks in the current solution s, is a good
idea). On the contrary, if (x, s) is a drop move (i.e. remove a truck from yard x
of solution s), among the drop moves, it is better to select a move with a small
T r(x, s) value.
The way to select a move at each iteration is now described. Remember that
in DLS and in TS, the performed move is the best among the ones in the set
 Learning Tabu Search for Combinatorial Optimization 9

| A ∪ D |, with |A| = |D| = 10. In LTS, two sets A and D of size 20 are
first randomly chosen. Then, let Aq (resp. Dq ) be the subset of A (resp. D)
containing the q (parameter tuned to 10) moves with the largest (resp. smallest)
trail values. Note that computing the trail value of a move is much quicker to
compute than the value of the resulting neighbor solution, as the flow algorithm
is requested for it. The performed move among Aq ∪ Dq is the best one according
to the objective function of the problem (i.e., the sum of the costs). Therefore,
for DLS, TS and LTS, the performed move has the best objective function value
among a sample of 20 evaluated solutions. This will allow to better measure
the impact of the trail system on the search. Note that if DIV is open (i.e., the
diversification mechanism is activated), Aq (resp. Dq ) is the subset of A (resp.
D) containing the q moves with the smallest (resp. largest) trail values.
Now are presented results on the 57 linear instances and the 15 non linear
instances instances associated with [8]. The non linear instances are based on the
following idea. If several trucks from the same company are contracted for the
same yard, the company is likely to propose discounted prices for that yard.
The proposed algorithms were tested on an Intel Quad-core i7 @ 3.4 GHz with
8 GB DDR3 of RAM memory. A common time limit T = 60 min is imposed
to each proposed method (i.e., DLS, TS and LTS). Note that DLS is restarted
from scratch each time a local optimum is found, in order to be able to apply
it during T minutes, and thus to perform a fair comparison with the two other
metaheuristics.
For DLS, TS and LTS, Fig. 1 compares the evolution of the average best
encountered solution value during 60 min on the linear instances (57 instances,
10 runs per instance). On the one hand, one can observe that LTS clearly out-
performs the other methods from the beginning to the end of the hour of compu-
tation. This obviously indicates that the learning process (i.e. the trail system)
introduced to TS to derive LTS is relevant. On the other hand, one can remark
that TS is better than DLS in the first 30 min, and then both methods have a
very comparable behavior.

Fig. 1. Evolution of DLS, TS and LTS on the linear instances.

10 N. Zufferey and D. Schindl

Fig. 2. Evolution of DLS, TS and LTS on the non linear instances.

Figure 2 is similar to Fig. 1 but is associated with the non linear instances.
The same observations as before can be made: LTS outperforms TS and DLS,
TS is better than DLS during half an hour, TS and DLS are comparable during
the second half hour.

4 Conclusion
In this paper, a new type of local search is presented, called Learning Tabu
Search and denoted LTS. It was showed that LTS was successfully adapted to
a refueling problem in a railway network, which was initially motivated by the
Railway Applications Section of INFORMS.
LTS involves a learning process relying on a trail system. The trail system
is based on the idea that if some combinations of characteristics often belong
to good solutions during the search process, such combinations of characteristics
should be favored when generating new solutions.
Even if the concept of trail system also exists in ant algorithms, it is managed
very differently in LTS. Recent overviews of ant algorithms can be found in [1]
and [2]. In contrast with ant algorithms: (1) LTS is a local search dealing with
a single solution, and not a method based on a population of solutions; (2) LTS
performs each decision quickly and with an aggressive manner (whereas in ant
algorithms, it is computationally cumbersome to select a move); (3) LTS uses
sequentially (instead of jointly) information based on the history of the search
(known in the ant community as the trail system) and on the short term profit
(known in the ant community as the heuristic information or the visibility or
the greedy force).
Within the local search frameworks, there already exist some mechanisms
favoring good moves (resp. attributes) which were frequently performed (resp.
encountered) in the past of the search process [4]. Such mechanisms are however
very differently managed in LTS, mainly because of the joint use of the follow-
ing two elements: (1) a trail system managed with an evaporation component
and a reinforcement component; (2) combinations of characteristics are handled
(instead of individually considering each characteristic).
 Learning Tabu Search for Combinatorial Optimization 11

Therefore, LTS can be clearly positioned according to the existing meta-

heuristics. Among the possible future works, it would be interesting to apply LTS
to other kinds of problems, for example to find out for each of them which value
of k is the most relevant.

References
1. Blum, C.: Ant colony optimization: introduction and recent trends. Phys. Life Rev.
2(4), 353–373 (2005)
2. Dorigo, M., Birattari, M., Stuetzle, T.: Ant colony optimization—artificial ants as
a computational intelligence technique. IEEE Comput. Intell. Mag. 1(4), 28–39
(2006)
3. Garey, M., Johnson, D.S.: Computer and Intractability: A Guide to the Theory of
NP-Completeness. Freeman, San Francisco (1979)
4. Gendreau M., Potvin, J.-Y.: Handbook of metaheuristics. In: International Series
in Operations Research & Management Science, vol. 146, pp. 573–597. Springer,
New York (2010)
5. Glover, F., Laguna, M.: Tabu Search. Kluwer Academic Publishers, Boston (1997)
6. Nemhauser, G., Wolsey, L.: Integer and Combinatorial Optimization. Wiley,
New York (1988)
7. Schindl, D., Zufferey, N.: Solution methods for fuel supply of trains. Inf. Syst. Oper.
Res. 51(1), 22–29 (2013)
8. Schindl, D., Zufferey, N.: A learning tabu search for a truck allocation problem with
linear and nonlinear cost components. Nav. Res. Logistics 61(1), 42–45 (2015)
9. Zufferey, N.: Metaheuristics: some principles for an efficient design. Comput. Tech-
nol. Appl. 3(6), 446–462 (2012)
10. Zufferey, N.: Optimization by ant algorithms: possible roles for an individual ant.
Optim. Lett. 6(5), 963–973 (2012)
Methodologies and Technologies
 Four Serious Problems and New Facts
of the Discriminant Analysis

Shuichi Shinmura(&)

Faculty of Economics, Seikei University, Tokyo, Japan

[email protected]

Abstract. The discriminant analysis is essential knowledge in science, tech-

nology and industry. But, there are four serious problems. These are resolved by
Revised IP–OLDF and k–fold cross validation.

Keywords: Minimum number of misclassifications (MNM)
Revised IP-

OLDF
SVM
LDF
Logistic regression
k-fold cross validation

1 Introduction

Fisher [1] described the linear discriminant function (LDF), and founded the dis-
criminant theory. Following this, the quadratic discriminant function (QDF) and multi-
class discrimination using Mahalanobis distance were proposed. These functions are
based on the variance-covariance matrices, and are easily implemented in the statistical
software packages. They can be used in many applications. However, real data rarely
satisfy Fisher’s assumptions. Therefore, it is well known that logistic regression is
better than LDF and QDF, because it does not assume a specific theoretical distribution,
such as a normal distribution. In addition to this, the discriminant rule is very simple:
If yi*f(xi) > 0, xi is classified to class1/class2 correctly. If yi*f(xi) < 0, xi is misclas-
sified. There are four serious problems hidden in this simplistic scenario [22].
(1) Problem 1
We cannot properly discriminate between cases where xi lies on the discriminant hype-
plane (f(xi) = 0). This unresolved problem has been ignored until now. The proposed
Revised IP-OLDF is able to treat this problem appropriately. Indeed, except for
Revised IP-OLDF, no functions can correctly count the number of misclassifications
(NM). These functions should count the number of cases where f(xi) = 0, and display
this alongside the NM in the output.
(2) Problem 2
Fisher’s LDF and QDF cannot recognize linear separable data (where the Minimum
NM (MNM) = 0). This fact was first found when IP-OLDF was applied to Swiss bank
note data [3]. In this paper, the determination of pass/fail in exams is used because it is
trivially linear-separable and we can obtain it easily. We show that, in many cases, the
NMs of LDF and QDF are not zero. Next, 100 re-samples of these data are generated,
and the mean error rates are obtained by 100-fold cross validation. The mean error rates

© Springer International Publishing Switzerland 2015

E. Pinson et al. (Eds.): ICORES 2014, CCIS 509, pp. 15–30, 2015.
DOI: 10.1007/978-3-319-17509-6_2
16 S. Shinmura

of LDF are 6.23 % higher than that of Revised IP-OLDF in the validation samples of
Table 7.
(3) Problem 3
If the variance-covariance matrix is singular, Fisher’s LDF and QDF cannot be cal-
culated because the inverse matrices do not exist. The LDF and QDF of JMP [9] are
solved by the generalized inverse matrix technique. In addition to this, RDA [4] is used
if QDF causes serious trouble with dirty data. However, RDA and QDF do not work
properly for the special case in which the values of features belonging to one class are
constant. If users can choose proper options for a modified RDF developed for this
special case, it works better than QDF and LDF in Table 5.
(4) Problem 4
Some statisticians misunderstand that the discriminant analysis is the inferential sta-
tistical method as same as the regression analysis, because it is derived from Fisher’s
assumption. But there are no standard error (SE) of the discriminant coefficients or
error rate, and variable selection methods such as stepwise methods and statistics such
as Cp and AIC. In this paper, we propose “k-fold cross validation for small samples”
and new variable selection method, the minimum mean error rates of which is chosen
as the best model. In future works (Future1), generalization ability and 95 % confi-
dence intervals of all LDFs are proposed.
In this research, two Optimal LDFs (OLDFs) based on the MNM criterion are
proposed. The above three problems are solved by IP-OLDF and Revised IP-OLDF
completely. IP-OLDF [13–15] reveals the following properties.
Fact (1) Relation between LDFs and NMs. IP-OLDF is defined on the data and
discriminant coefficient spaces. Cases of xi correspond to linear hyper-planes
(Hi(b) = yi* (txib + 1) = 0) in the p-dimensional discriminant coefficient space that
divide the space into two half-planes: the plus half-plane (Hi(b) > 0) and minus half-
plane (Hi(b) < 0). Therefore, the coefficient space is divided into a finite convex
polyhedron by Hi(b). Interior point bj of the convex polyhedron corresponds to the
discriminant function fj(x) = tbjx + 1 on the data space that discriminates some cases
properly and misclassifies others. This means that each interior point bj has a unique
NM. The “Optimal Convex Polyhedron (OCP)” is defined as that with the MNM.
Revised IP-OLDF [16] can find the interior point of OCP directly, and solves the
unresolved problem (Problem 1) because there are no cases on the discriminant hyper-
plane (f(xi) = 0). If bj is on a vertex or edge of the convex polyhedron, however, the
unresolved problem cannot be avoided because there are some cases on f(xi) = 0.
Fact (2) Monotonous decrease of MNM (MNMp ≥ MNM(p+1)). Let MNMp be the
MNM of p features (independent variables). Let MNM (p+1) be the MNM of the (p + 1)
features formed by adding one feature to the original p features. MNM decreases
monotonously (MNMp ≥ MNM(p+1)), because OCP in the p-dimensional coefficient
space is a subset of the (p + 1)-dimensional coefficient space [18]. If MNMp = 0, all
MNMs including p features are zero. Swiss bank note data consists of genuine and
counterfeit bills with six features. IP-OLDF finds that this data is linear-separable
 Four Serious Problems and New Facts 17

according to two features (X4, X6). Therefore, 16 models including these two features
have MNMs = 0. Nevertheless, Fisher’s LDF and QDF cannot recognize that this data
is linear-separable, presenting a serious problem. In this paper, we show that Revised
IP-OLDF can resolve the above three problems, and is superior to Fisher’s LDF,
logistic regression, and Soft-margin SVM (S-SVM) [28] under 100-fold cross vali-
dation [20, 21] of the pass/fail determinations of exams [19] and their re-sampled data.

2 Discriminant Functions

2.1 Statistical Discriminant Functions

Fisher defined LDF to maximize the variance ratio (between/within classes) in Eq. (1).
This can be solved by non-linear programming (NLP).

MIN ¼ t bðm1  m2 Þt ðm1  m2 Þb=t b Rb; ð1Þ

If we accept Fisher’s assumption, the same LDF is obtained in Eq. (2). This equation
defines LDF explicitly, whereas Eq. (1) defines LDF implicitly. Therefore, statistical
software packages adopt this equation. Some statisticians misunderstand that dis-
criminant analysis is the same as regression analysis. Discriminant analysis is inde-
pendent of inferential statistics, because there are no SEs of the discriminant
coefficients and error rates (Problem 4). Therefore, the leave-one-out (LOO) method [6]
was proposed to choose the proper discriminant model.

Fisher's LDF : f ðxÞ ¼ t fx  ðm1 þ m2 Þ=2g R1 ðm1  m2 Þ ð2Þ

Most real data does not satisfy Fisher’s assumption. When the variance-covariance
matrices of two classes are not the same (Σ1 ≠ Σ2), the QDF defined in Eq. (3) can be
used. The Mahalanobis distance (Eq. (4)) is used for the discrimination of multi-
classes, and the Mahalanobis-Taguchi [25] method is applied in quality control.



QDF : f ðxÞ ¼ t x R1 1 t 1 t 1
2  R1 x=2 þ m1 R1  m2 R2 x þ c ð3Þ


D ¼ SQRT t
ðx  mÞR1 ðx  mÞ ð4Þ

These functions are applied in many areas, but cannot be calculated if some features
remain constant. There are three cases. First, some features that belong in both classes
are the same constant. Second, some features that belong in both classes are different
but constant. Third, some feature that belongs to one class is constant. Most statistical
software packages exclude all features in these three cases. On the other hand, JMP
enhances QDF using the generalized inverse matrix technique. This means that QDF
can treat the first and second cases correctly, but cannot handle the third case properly.
Recently, the logistic regression in Eq. (5) has been used instead of LDF and QDF
for two reasons. First, it is well known that the error rate of logistic regression is often
less than those of LDF and QDF, because it is derived from real data instead of some
normal distribution that is liberated from reality. Let ‘P’ be the probability of belonging
18 S. Shinmura

to a group of diseases. If the value of some feature is increasing/decreasing, ‘P’

increases from zero (normal group) to one (group of diseases). This representation is
very useful in medical diagnosis, as well as for ratings in real estate and bonds. On the
contrary, Fisher’s LDF assumes that cases near to the average of the diseases are
representative cases of the diseases group. Medical doctors never permit this claim.

LogðP=ð1  PÞÞ ¼ f ðxÞ ð5Þ

2.2 Before and After SVM

Stam [24] summarized Lp-norm discriminant methods until 1997, and answers the
question of “Why have statisticians rarely used Lp-norm methods?” He gives four
reasons: Communication, promotion and terminology; Software availability; Relative
accuracy of Lp-norm classification methods: Ad hoc studies; and the Accuracy of Lp-
norm classification methods: decision theoretic justification. While each of these reasons
is true, they are not important. The most important reason is that there is no comparison
between these methods with statistical discriminant functions, because discriminant
analysis was established by Fisher before mathematical programming (MP) approaches.
There are two types of MP applications. The first is modeling by MP, such as for
portfolio selection [26], and the second is catch-up modeling, such as for the regression
and discriminant analysis. Therefore, the latter type should be compared with preceding
results. No statisticians use Lp-norm methods, because there is no research indicating
that Lp-norm methods are superior to statistical methods. Liitschwager and Wang [7]
defined a model based on the MNM criterion. There are several mistakes, but the most
important one is the restriction on the discriminant coefficients. Only one discriminant
coefficient should be fixed to −1/1. There is no need to fix the other (k − 1) coefficients in
the range [−1, 1].
Vapnik proposed three different SVM models. The hard-margin SVM (H-SVM)
indicates the discrimination of linear separable data. H-SVM is defined to maximize the
distance of the “Support Vector (SV)” in order to obtain “good generalization” by NLP,
which is similar to “not overestimating the validation data in statistics.” H-SVM is
redefined to minimize (1/“distance of SV”) in Eq. (6). This is solved by quadratic
programming (QP), which can only be used for linear separable data. This may be why
investigation of linear separable data has been ignored. We statisticians misunderstand
that discrimination of linear separable data is very easy. In statistics, there was no
technical term for linear separable data. However, the condition “MNM = 0” is the
same as being linear-separable. Note that “NM = 0” does not imply the data is linear-
separable. It is unfortunate that there has been no research into linear separability.

MIN ¼ jjbjj2 =2; yi  ðt xi b þ b0 Þ  1; b : p  discriminant coefficients:

ð6Þ
yi ¼ 1=  1 for xi 2 class1=class2: xi : p  featuresðindependent variablesÞ:

Real data are rarely linear-separable. Therefore, S-SVM has been defined in Eq. (7).
S-SVM permits certain cases that are not discriminated by SV (yi*(txib + b0) < 1).
 Four Serious Problems and New Facts 19

The second objective is to minimize the summation of distances of misclassified cases

(Σei) from SV. These two objects are combined by defining some “penalty c.” The
Markowitz portfolio model to minimize risk and maximize return is the same as S-SVM.
However, the return is incorporated as a constraint, and the objective function minimizes
risk. The decision maker chooses a solution on the efficient frontier. On the contrary,
S-SVM does not have a rule to determine c properly; nevertheless, it can be solved by an
optimization solver. (Kernel-SVM is omitted from the research.)

MIN ¼ jjbjj2 =2 þ c  Rei ; yi  ðt xi b þ b0 Þ  1  ei ; ð7Þ

c: penalty c to combine two objectives. ei: non-negative value.

2.3 IP-OLDF and Revised IP-OLDF

Shinmura and Miyake [12] developed the heuristic algorithm of OLDF based on the MNM
criterion. This solves the five features (5-features) model of Cephalo Pelvic Disproportion
(CDP) data that consisted of two groups having 19 features. SAS was introduced into
Japan in 1978, and three technical reports about the generalized inverse matrix, the sweep
operator [5], and SAS regression applications [8] are related to this research. LINDO was
introduced to Japan in 1983. Several regression models are formulated by MP [10], e.g.,
least-squares problems can be solved by QP, and LAV (Least Absolute Value) regression
is solved by LP. Without a survey of previous research, the formulation of IP-OLDF can be
defined as in Eq. (8). This notation is defined on p-dimensional coefficient space, because
the constant is fixed to 1. In pattern recognition, the constant is a free variable. In this case,
the model is defined on (p + 1)-coefficient space, and we cannot elicit the same deep
knowledge as with IP-OLDF. This difference is very important. IP-OLDF is defined on
both p-dimensional data and coefficient spaces. We can understand the relation between
the NM and LDF clearly. The linear equation Hi(b) = yi* (txib + 1) = 0 divides
p-dimensional space into plus and minus half-planes (Hi(b) > 0, Hi(b) < 0). If bj is in the
plus half-plane, fj(x) = yi*(tbjx + 1) discriminates xi correctly, because fj(xi) = yi*
(tbjxi + 1) = yi *(txibj + 1) > 0. On the contrary, if bj is included in the minus half-plane,
fj(x) cannot discriminate xi correctly, because fj(xi) = yi*(tbjxi + 1) = yi *(txibj + 1) < 0. The
n linear equations Hi(b) divide the coefficient space into a finite number of convex
polyhedrons. Each interior point of a convex polyhedron has a unique NM that is equal to
the number of minus half-planes. We define the OCP as that for which NM is equal to
MNM. If xi is classified correctly, ei = 0 and Hi(b) ≥ 0 in Eq. (8). If there are p cases on f(xi)
=0, we can obtain the exact MNM. However, if there are over (p+1) cases on f(xi)=0, this
causes the unresolved problem. If xi is misclassified, ei = 1 and Hi(b) ≥ −10000. This
means that IP-OLDF chooses the discriminant hyper-plane Hi(b) = 0 for correctly clas-
sified cases, and Hi(b) = −10000 for misclassified cases according to a 0/1 decision
variable. if IP–OLDF chooses a vertex having p cases, it chooses the OCP correctly. If it
chooses a vertex having over (p+1) cases, it may not choose the OCP. In addition to this
defect, IP-OLDF must be solved for the three cases where the constant is equal to 1, 0, −1,
because we cannot determine the sign of yi in advance. Combinations of yi = 1/−1 for xi ∊
class1/class2 are decided by the data, not the analyst.
20 S. Shinmura

MIN ¼ Rei ; Hi ðbÞ   M  ei ; M : 10; 000 ðBig M constantÞ: ð8Þ

The Revised IP-OLDF in Eq. (9) can find the true MNM, because it can directly find
the interior point of the OCP. This means there are no cases where yi*(txib + b0) = 0.
If xi is discriminated correctly, ei = 0 and yi*(txib + b0) ≥ 1. If xi is misclassified, ei = 1
and yi*(txib + b0) ≥ −9999. It is expected that all misclassified cases will be extracted
to alternative SVs, such as yi*(txib + b0) = −9999. Therefore, the discriminant scores of
misclassified cases become large and negative, and there are no cases where yi*
(txib + b0) = 0. This means that b is interior point of OCP defined by IP-OLDF.

MIN ¼ Rei ; yi  ð t xi b þ b0 Þ  1  M  e i ; b0 : free decision variable: ð9Þ

If ei is a non-negative real variable, we utilize Revised LP-OLDF, which is an L1-

norm LDF. Its elapsed runtime is faster than that of Revised IP-OLDF. If we choose a
large positive number as the penalty c of S-SVM, the result is almost the same as that
given by Revised LP-OLDF, because the role of the first term of the objective value in
Eq. (7) is ignored. Revised IPLP-OLDF is a combined model of Revised LP-OLDF and
Revised IP-OLDF. In the first step, Revised LP-OLDF is applied for all cases, and ei is
fixed to 0 for cases that are discriminated correctly by Revised LP-OLDF. In the second
step, Revised IP-OLDF is applied for misclassified cases in the first step. Therefore,
Revised IPLP-OLDF can obtain an estimate of MNM faster than Revised IP-OLDF
[17, 23], but it is unknown to be free from the unresolved problem.

3 The Unresolved Problem (Problem 1)

3.1 Perception Gap of This Problem

About the unresolved problem, there are several understandings. Most researchers treat
the cases xi on f(xi) = 0 in class1. There is no explanation of why it makes sense. Some
statisticians explain that it is decided stochastically, because the statistics is a sturdy of
probability. This explanation seems theoretically at first glance, but it is nonsense by
two reasons. Statistical software adopt the former decision rule because many papers
and researchers adopt this rule. In the medical diagnosis, medical doctors strive to
judge the patients near by the discriminant hyper-plane. If they know second expla-
nation, they are deeply disappointed in the discriminant analysis. Until now, all LDFs
such as Fisher’s LDF, logistic regression, H-SVM and S-SVM cannot treat this
problem properly. IP-OLDF reveals that only interior points of convex polyhedron can
resolve this problem. It can find the vertex of true OCP if data is general position and it
stop the optimization choosing p cases on the discriminant hyper-plane. But, it may not
find the true MNM if data is not general position and it choose over (p + 1) cases on the
discriminant hyper-plane. Revised IP-OLDF can find the interior point of the OCP
directly. We cannot judge whether other LDFs choose the interior point, edge or vertex
of the convex polyhedron. This is confirmed by checking the number of cases xi that
satisfy |f(xi)| ≤ 10−6 if we consider | f(xi)| ≤ 10−6 is zero. If this number is zero, this
Another random document with
no related content on Scribd:
brizna, vna miserable ocasion
para le despegar de sí. Dize que
por tener grande edad le perdio el
respeto que le deuia como a
señor. O que le trata mal sus
hijos; o que quiere mandar más
que él; y si eres moço leuantate
que te le quieres echar con la hija,
o con la muger; o que te hallaron
hablando con vna donzella de
casa en vn rincon. De manera
que nunca les falta con que
infame y miserablemente los
echar, y avn sin el salario que
siruio, y donde penso el
desuenturado del sieruo que auia
proueydo a la pobreza y
neçesidad en que pudiera venir
se ofreçio de su voluntad a la
causa y ocasion de muy mayor,
pues echado de aquellas agenas
casas viene forçado al hospital.
Alli viejos los tales y enfermos y
miserables los dan de comer y
beber y sepoltura por limosna y
amor de Dios. Resta agora,
Miçilo, que quieras considerar
como cuerdo y auisado animo
todo lo que te he representado
aqui, porque todo lo esperimenté
y passó por mí. No çeues ni
engañes tu entendimiento con la
vanidad de las cosas desta vida,
que talmente suelen engañar, y
mira bien que Dios y naturaleza a
todos crian y produçen con
habilidad y estado de poder gozar
de lo bueno que ella crió, si por
nuestro apetito, oçio y miseria no
lo venimos a perder, y de aqui
adelante contentate con el estado
que tienes, que no es çierto digno
de menospreçiar.
Miçilo.—¡O gallo
bienauenturado! que
bienauenturado me has hecho oy,
pues me has auisado de tan gran
bien; yo te prometo nunca serte
ingrato a benefiçio de tanto valor.
Solo te ruego no me quieras
desamparar que no podre viuir sin
ti; y porque es venido el dia
huelga, que quiero abrir la tienda
por vender algun par de çapatos
de que nos podamos mantener
oy.

Fin del deçimo nono canto del

gallo.
 NOTAS:
[1102] G., canto del gallo.
[1103] G., poseen.
[1104] G., morir.
[1105] G., acordarsseme.
[1106] G., el daño que despues de tragado el çeuo en el anzuelo
está, y teniendo la meluca en la boca para la tragar no te la hago
echar fuera antes que prendiendo la punta en tu paladar bomites
la sangre y vida con dolor.
[1107] G., el veneno.
[1108] G., no ha de aprouechar mi.
[1109] G., fuerça.
[1110] G., ¿como podre yo aprobar vuestra opinion?
[1111] G., se suple.
[1112] G., pleytos.
[1113] G., los.
[1114] G., poder soliçitar.
[1115] G., los sienten.
[1116] G., y casa.
[1117] G., trabajo, y por el consiguiente más miseria y
enfermedad que lleuan.
[1118] G., deuen desear aquella vida, por solo el deleyte y
contentamiento que da vibir en aquellas anchas y espaçiosas
casas, habitaçion de dioses y de sola persona Real y inçitados de
aquellas grandes esperanças que prometen aquellos poderosos
señores con su real y generosa conuersaçion.
[1119] G., por gozar solamente de aquellos marauillosos tesoros,
aparadores de oro.
[1120] G., al cunplimiento de vuestra neçesidad.
[1121] G., hallara.
[1122] G., propias.
[1123] G., pobres.
[1124] G., día.
[1125] G., grande.
[1126] G., en tal vida.
[1127] G., a la.
[1128] G., acaben.
[1129] G., porque se acuerde de tu.
[1130] G., dignidad.
[1131] G., que vas a.
[1132] G., patrimonio.
[1133] G., seruiçio.
[1134] G., a tu amo.
[1135] G., que te aconteçe que preguntandote el señor que
hombre fue el rey Tholomeo, respondas tu que fue hermano y
marido de Clopatra; o otra cosa que va muy lexos de la intinçion
de tu señor.
[1136] G., dizen que es temor.
[1137] G., de tu habilidad, persona y linaje.
[1138] G., y esta pesquisa que de ti.
[1139] G., a tu.
[1140] G., que digan que se sirue.
[1141] G., sabios y cuerdos.
[1142] G., oculto y sonoliento.
[1143] G., tu saber, cordura y discreçion.
[1144] G., trihunfador.
[1145] G., mereçes, no de roble o arrayan como los otros en la
Olimpia.
[1146] G., cosas.
[1147] G., alguna.
[1148] G., de.
[1149] G., ante.
[1150] G., vna gran.
[1151] G., quanto a grandes salarios.
[1152] G., con.
[1153] G., inuidiosos.
[1154] G., pues.
[1155] G., puede.
[1156] G., osas.
[1157] G., les tienen.
[1158] G., çinquenta.
[1159] G., la.
[1160] G., que te manda tu señora.
[1161] G., que yo tanto amo.
[1162] G., confiar.
 ARGUMENTO
DEL VIGESSIMO Y
VLTIMO CANTO

En este vigessimo canto el auctor

representa a Demophon, el
qual viniendo vn dia a casa de
Miçilo su vezino a le visitar le
halló triste y afligido por la
muerte de su gallo, y
procurando dexarle consolado
se vuelue a su casa.

Demophon. Miçilo.
Demophon.—¡O Miçilo! vezino y
amigo mio, ¿qué es la causa que
ansi te tiene atormentado por
cuydado y miserable
aconteçimiento? veote triste,
flaco, amarillo con representaçion
de philosopho, el rostro lançado
en la tierra, pasearte por este
lugar obscuro dexado tu contino
offiçio de çapateria en que tan a
la contina te solias ocupar con
eterno trabajo, ¿consumes agora
el tiempo en sospiros? Nuestra
igual edad, vezindad y amistad te
obliga a fiar de mí tus tan
miserables cuydados; porque ya
que no esperes de mí que
cunpliese tus faltas ayudarte he
con consejo; y si todo esto no
estimares, bastarte ha saber que
mitiga mucho el dolor comunicar
la pena, prinçipalmente
contandose a quien en alguna
manera por propria la sienta.
¿Qué es de tu belleza y alegria,
desemboltura y comunicaçion con
que a todos tus amigos y vezinos
te solias dar de noche y de dia en
çenas y combites y fuera dellos?
ya son pasados muchos dias que
te veo recogido en soledad en tu
casa que ni me quieres ver ni
hablar, ni visitar como solias.
Miçilo.—¡O mi Demophon! mi
muy caro hermano y amigo. Solo
esto quiero que como tal amigo
de mí sepas, que no sin gran
razon en mí ay tan gran muestra
de mal. Prinçipalmente quando
tienes de mí bien entendido que
no qualquiera cosa haze en mí
tan notable mudança, pues has
visto en mí auer disimulado en
varios tienpos notables toques de
fortuna y infortunios tan graues
que a muy esforçados varones
huuieran puesto en ruyna, y yo
con igual rostro los he sabido
passar. Avnque comunmente se
suele dezir que al pobre no ay
infortunio, que aunque esto sea
ansi verdad no dexamos de sentir
en nuestro estado humilde lo que
al anima le da a entender su
natural. Ansi que tengo por çierto,
Demophon, que no ay igual dolor
de perdida ni miseria que con
gran distançia se compare con el
mio.
Demophon.—Mientras más me le
has encareçido más me has
augmentado la piedad y miseria
que tengo de tu mal; de donde
naçe en mí mayor deseo de lo
saber. Por tanto no reserues en tu
pecho tesoro tan perjudiçial, que
no hay peor espeçie de auariçia
que de dolor. Por çierto en poco
cargo eres a naturaleza pues
pribandote del oro y riquezas, de
pasiones y miserias fue contigo
tan liberal que en abundançia te
las comunicó. Dime porqué ansi
te dueles, que no podré consentir
lo passes con silençio y
disimulaçion.
Miçilo.—Quiero que ante todas
las cosas sepas, ¡o Demophon!
que no es la que me fatiga falta
de dineros para que con tus
tesoros me ayas de remediar, ni
de salud para que con medicos
me la ayas de restituir. Ni tanpoco
me aflixo por mengua que me
hagan las tus vasijas, ni aparatos
y arreos de tapetes y alhajas con
que en abundançia te sueles
seruir. Pero faltame de mi casa vn
amigo, vn conpañero de mis
miserias y trabajos y tan igual que
era otro yo; con el qual poseya yo
todos los tesoros y riquezas que
en el mundo ay; faltame, en
conclusion, vna cosa, Demophon,
que con ningun poder ni fuerças
tuyas la puedes suplir: por lo qual
me escuso de te la dezir, y a ti de
la saber.
Demophon.—No en vano suelen
dezir, que al pobre es proprio el
filisofar, como agora tú; yo no
creo que has aprendido esa
retorica en las scuelas de
Athenas, con que agora de nueuo
me encareçes tu dolor: ni sé qué
maestro has tenido della de poco
acá.
Miçilo.—Ese maestro se me
murio, cuya muerte es causa de
mi dolor.
Demophon.—¿Quien fue?[1163].
Miçilo.—Sabras, amigo, que yo
tenia vn gallo que por mi casa
andaua estos dias en conpañia
destas mis pocas gallinas que las
albergaua y recogia y defendia
como verdadero marido y varon.
Suçedio que este dia de
carnestolendas que passó, vnas
mugeres desta nuestra vezindad,
con temeraria libertad, habiendo
solamente cuenta, y
pareçiendoles que era el dia
priuillegiado me entraron mi casa
estando yo ausente, que
cautelosamente aguardaron que
fuesse ansi, y tomaron mi gallo y
lleuaronle al campo, y con gran
grita y alarido le corrieron
arroxandosele las vnas a las
otras: y como quien dize[1164],
daca el gallo, toma el gallo, les
quedauan las plumas en la mano.
En fin fue pelado y desnudo de su
adornado y hermoso vestido; y no
contentas con esto, rendiendosele
el desuenturado sin poderles
huyr, confiandose de su
inoçençia: pensando que no
pasara adelante su tirania y[1165]
crueldad, subjetandoseles con
humildad, pensando que por esta
via las pudiera conuençer y se les
pudiera escapar, sacaron de sus
estuches cuchillos, y sin tener
respecto alguno a su inoçençia le
cortaron su dorada y hermosa
çeruiz, y de comun acuerdo
hiçieron çena opulenta dél.
Demophon.—Pues ¿por faltarte
vn gallo te afliges tanto que estás
por desesperar? Calla que yo lo
quiero remediar con embiarte otro
gallo criado en mi casa, que creo
que hará tanta ventaja al tuyo
quanta haze mi despensa a la
tuya para le mantener.
Miçilo.—¡O Demophon! quanto
viues engañado en pensar que mi
gallo perdido con qualquiera otro
gallo se podria satisfazer.
Demophon.—¿Pues qué tenia
más?
Miçilo.—Oyeme, que te quiero
hazer saber que no sin causa me
has hallado philosopho rectorico
oy.
Demophon.—Dimelo.
Miçilo.—Sabras que aquel gallo
era Pythagoras el philosopho,
eloquentissimo varon, si le has
oydo dezir.
Demophon.—Pythagoras,
muchas vezes le oy dezir. Pero
dime ¿cómo quieres que entienda
que el gallo era Pythagoras: que
me pones en confusion?
Miçilo.—Porque si oyste dezir de
aquel sapientissimo philosopho,
tambien oyrias dezir de su
opinion.
Demophon.—¿Quál fue?
Miçilo.—Este afirmó que las
animas passauan de vn cuerpo a
otro. De manera que dixo que
muriendo vno de nosotros luego
desanparando nuestra alma este
nuestro cuerpo en que vibio se
passa a otro cuerpo de nueuo a
viuir: y no sienpre a cuerpo de
honbre. Pero aconteçe que el que
agora fue rey passar[1166] a
cuerpo de vn puerco, vaca ó leon,
como sus hados y suçeso[1167] lo
permiten, sin el alma lo poder
evitar; y ansi el alma de
Pythagoras despues aca que
naçio auia viuido en diuersos
cuerpos, y agora viuia en el
cuerpo de aquel gallo que tenia
yo aqui.
Demophon.—Esa manera de
dezir ya la oy que la afirmaua él.
Pero era un mentiroso,
prestigioso y embaydor, y tanbien
como el era efficaz en el persuadir
y aquella gente de su tienpo era
sinple y ruda, façilmente les hazia
creer qualquiera cosa que él
quisiesse soñar.
Miçilo.—Çierto es yo que ansi
como lo dezia era verdad.
Demophon.—¿Como ansi?
Miçilo.—Porque en aquel gallo
me habló y me mostró en muchos
dias ser él.
Demophon.—¿Que te habló?
Cosa me cuentas digna de
admiraçion. En tanta manera me
marauillo de[1168] lo que dices por
cosa nueua que sino huuiera
conoçido tu bondad y sinçera
condiçion pensara yo agora que
estauas fuera de seso y que
como loco deuaneas. O que
teniendome en poco pensauas
con semejantes sueños vurlar de
mí. Pero por Dios te conjuro ¡o
Miçilo! y por nuestra amistad, la
qual por ser antigua entre
nos[1169] tiene muestra de
deydad, me digas muy en
particular todo lo que en la verdad
es.
Miçilo.—¡O Demophon! que sin
lagrimas no te lo puedo dezir,
porque sé yo solo lo mucho que
perdi. Auianme tanto faboreçido
los hados que no creo que en el
mundo haya sido honbre tan feliz
como yo. Pero pareçeme que
este fabor fue para escarneçer de
mí, pues me comunicaron tan
gran bien con tanta breuedad,
que no parece sino que como
anguila se me deleznó.
Solamente me pareçe que
entendí mientra le tuue en le
apretar en el puño para le poseer,
y quando pense que le tenía con
alguna seguridad se me fue.
Tanbien sospecho que los hados
me quisieron tentar si cabia en mí
tanto bien, y por mi mala suerte
no fue dél mereçedor; y porque
veas si tengo razon de lo
encareçer, sabras que en él tenía
yo toda la consolaçion y
bienauenturança que en el mundo
se podia tener. Con él pasaua yo
mis trabajos de noche y de dia: no
auia cosa que yo quisiesse saber
o auer que no se me diesse a
medida de mi voluntad. El me
mostró la vida de todos quantos
en el mundo ay: lo bueno y malo
que tiene la vida del rey y del
çiudadano, del cauallero, del
mercader y del labrador. El me
mostró quanto en el çielo y el
infierno ay, porque me mostró a
Dios y todo lo que gozan los
bienauenturados allá. En
conclusion ¡o Demophon! yo perdi
vn tesoro que ningun poderoso
señor en el mundo más no pudo
poseer.
Demophon.—Por çierto tengo, ¡o
Miçilo! sentir con mucha razon el
gran mal que te han hecho esas
mugeres en pribarte de tanto
bien, quando queriendo satisfazer
a sus vanos apetitos, çelebrando
sus lasçiuas y adulteras fiestas no
perdonan cosa dedicada ni
reseruada por ningun varon, con
tanto que executen su voluntad.
No miraron que tú no eras honbre
con quien tal dia se suelen
festejar, y que por tu edad no
entras en cuenta de los que
çelebran semejantes fiestas. Que
los moços ricos subjetos al tirano
y lasçiuo[1170] amor, enpleados en
las contentar no les pueden negar
cosa que haga a su querer, y ansi
por[1171] los entretener les
demandan en tales dias cosas
curiosas, en el cumplimiento de
las quales conoçen ellas su
mayor y más fiel enamorado y
seruidor; y ansi agora dandoles a
entender que para su laçiuia no
los han menester en el tienpo que
entra[1172] de la quaresma,
mostrando gran voluntad de se
contener pelan aquellos gallos en
lugar de la juuentud; mostrando
menospreçiar su gallardia por ser
tienpo santo el que entra, y que
no se quieren dellos en este
tienpo seruir; y ansi, burlando
dellos, pelan aquellos gallos en su
lugar, dando a entender que los
tengan en poco, pues pelados de
toda su pluma y hazienda en el
tienpo pasado que les fue
disimulado el luxuriar, ya,
recogiendose a la santidad, los
dexan[1173]; ¡o animal tirano y
ingrato a todo bien!; que en todas
sus obras se preçian mostrar su
mala condiçion. ¿Y no vian que tú
no estauas en edad para vurlar de
ti?
Miçilo.—Y avn por conocer yo
bien esa verdad ni me casé, ni las
quise ver; y avn no me puedo
escapar de su tirania, que
escripto me dizen que está que
no ay honbre a quien no alcançe
siquiera la sombra de su veneno y
maldiçion. Solamente me lastima
pensar que ya que me auian de
herir no fue de llaga que se
pudiesse remediar. Quitaronme
mi consejero, mi consuelo y mi
bien. Avn pluguiesse a Dios que
en este tienpo tan santo se
recogiessen de veras y sin alguna
fiçion[1174] tratassen de veras la
virtud. Ayunar, no beber, ni comer
con tanta disoluçion, no se
afeytar, ni vestirse tan
profanamente, ni vurlar, ni mofar
como en otro qualquiera tienpo
comun[1175]. Pero vemos que sin
alguna rienda viben el dia de
quaresma como qualquiera otro.
Son sus fiestas las que aborreçe
Dios, porque no son sino para le
ofender.
Demophon.—Por çierto, Miçilo,
espantado estoy de ver la vurla
destas vanas mugeres; con
quantas inuençiones[1176] passan
su tienpo, y quantas astuçias
vsan para sacar dineros de sus
amantes. Prinçipalmente en estos
pueblos grandes de villas y
çiudades; porque estas cosas no
las saben los aldeanos[1177], ni ha
llegado del todo la maliçia
humana por allá. Por çierto cosas
ay de gran donayre que se
inuentan en estos pueblos
grandes[1178]; con las quales los
inuentores dellas entretienen sus
cosas, y hazen sus hechos[1179]
por su proprio fin de cada qual y
interes; por çierto que me tienen
de cada dia en más admiraçion.
Prinçipalmente en este pueblo
donde ay tanta concurrencia de
gentes, o por causa de corte Real
o por[1180] chançelleria; porque la
diuersidad de estrangeros haze
dar en cosas, y inuentar donayres
que confunden el ingenio auerlas
solamente de notar. Quantas
maneras de santidades fingidas,
romerias, bendiçiones y
peregrinaçiones. Tanto hospital,
colejios de santos y santas; casas
de niños y niñas é hospitales de
viejos. Tanta cofradia de
disçiplinantes de la cruz y de la
pasion, y proçesiones. Tanto
pedigueño de limosnas, que más
son los que piden que son los
pobres que lo[1181] quieren[1182]
reçebir.
Miçilo.—Por çierto, Demophon,
tú tienes mucha razon y vna de
las cosas de que yo estoy más
confuso es de ver que en este
nuestro lugar, siendo tan noble y
el más prinçipal de nuestra
Castilla, donde[1183] ay más
letrados y honbres más agudos
en la conuersaçion y cosas del
mundo y cortesanía, y en estas
flaquezas y engaños que se
ofreçen[1184], son todos en vn
común más façilmente arroxados
y derrocados que en todos
quantos en otros pueblos ay; y
avn engañados para lo aprobar,
auctorizar y seguir[1185]. Que se
atreua vn honbre a entrar aqui en
este pueblo donde está la flor de
cordura y agudeça y discreçion, y
que debajo de vn habito religioso
engañe a todo estado eclesiastico
y seglar, diziendo que hará boluer
los rios atras, y hará cuaxar el
mar, y que forçará los demonios
que en los infiernos estan, y que
hará[1186] parir quantas[1187]
mugeres son, quanto quiera que
de su naturaleza sean esteriles y
que no puedan conçebir[1188], y
que en esto vengan a caer todos
los más prinçipales y generosos
prinçipes y señores, y se le
vengan a rendir quantas dueñas y
donzellas viben en este
lugar[1189]. Que se sufra vibir en
este pueblo vn honbre que debajo
de nonbre de Juan de Dios, no se
le çierre puerta de ningun Señor
ni letrado, ni se le niegue cosa
alguna que quiera demandar, y
después le quemen públicamente
por sometico engañador. Pues,
¿no se ha disimulado tanbien un
clerigo que auia sido primero
frayle veynte años, al qual por
tener muestra de gran santidad le
fue encargado aquel colegio de
niñas? tal sea su salud qual dellas
cuenta dio. ¿En que está esto,
amigo?
Demophon.—A tu gallo quisiera
yo, Miçilo que lo huuieras
preguntado antes que a mí
porque él te supiera mejor
satisfazer. Pero para mi bien creo
que en alguna manera deuo de
açertar; que creo que de los
grandes pecados que ay en este
lugar[1190] viene esta comun
confusion, o çeguedad. Que como
no hay en este pueblo más
prinçipal ni más comun que
pecados y ofensas de Dios;
pleytos, hurtos, vsuras, mohatras,
juegos, blasfemias, symonias,
trapazas y engaños, y despues
desto una puteria general, la qual
ni tiene punto, suelo, ni fin. Que ni
se reserua dia, ni fiesta,
quaresma, ni avn Semana Santa
ni pasqua en que se çese[1191] de
exerçitar como offiçio conueniente
a la republica, permitido y
aprobado por neçesario en la ley,
en pena deste mal nos çiega Dios
nuestros entendimientos, orejas y
ojos, para que auisandonos no
entendamos, y oyendo no
oyamos, y con ojos[1192] seamos
como çiegos que palpamos la
pared. En tanta manera somos
traydos en çeguedad que
estamos rendidos al engaño muy
antes que se ofrezca el
engañador. Hanos hecho Dios
escarnio, mofa y risa a los muy
chicos[1193] niños de muy tierna
edad. ¿En qué lugar por pequeño
que sea se consentira, o
disimulará lo mucho, ni lo muy
poco que se disimula y sufre
aqui? ¿Dónde hay tanto juez sin
justiçia como aqui? ¿Dónde tanto
letrado sin letras como aqui?
¿Dónde tanto executor sin que se
castigue[1194] la maldad? ¿Dónde
tanto escribano, ni más comun el
borron? Que no ay honbre de
gouierno en este pueblo que trate
más que su proprio interes, y
como más se auentajará. Por esto
permite Dios que vengan vnos
zarlos, o falsos prophetas que con
embaymientos, aparençias y
falsas demostraçiones nos hagan
entender qualquiera cosa que nos
quieran fingir. Y lo que peor es,
que quiere Dios que despues
sintamos más la risa que el
interes en que nos engañó.
Miçilo.—Pues avn no pienses,
Demophon, que la vanidad y
perdiçion destas liuianas mugeres
se le ha de passar a Dios sin
castigo; que yo te oso afirmar por
cosa muy çierta y que no faltará.
Que por ver Dios su disoluçion,