Hybrid System Identification: Theory

and Algorithms for Learning Switching

Models Fabien Lauer
Visit to download the full and correct content document:
https://textbookfull.com/product/hybrid-system-identification-theory-and-algorithms-for
-learning-switching-models-fabien-lauer/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Hidden Semi-Markov models : theory, algorithms and

applications 1st Edition Yu

https://textbookfull.com/product/hidden-semi-markov-models-
theory-algorithms-and-applications-1st-edition-yu/

Nonlinear System Identification: From Classical

Approaches to Neural Networks, Fuzzy Models, and
Gaussian Processes 2nd Edition Oliver Nelles

https://textbookfull.com/product/nonlinear-system-identification-
from-classical-approaches-to-neural-networks-fuzzy-models-and-
gaussian-processes-2nd-edition-oliver-nelles/

Genetic Algorithms and Machine Learning for Programmers

Create AI Models and Evolve Solutions 1st Edition
Frances Buontempo

https://textbookfull.com/product/genetic-algorithms-and-machine-
learning-for-programmers-create-ai-models-and-evolve-
solutions-1st-edition-frances-buontempo/

Algorithms and Models for the Web Graph Anthony Bonato

https://textbookfull.com/product/algorithms-and-models-for-the-
web-graph-anthony-bonato/
Optimization of Complex Systems: Theory, Models,
Algorithms and Applications Hoai An Le Thi

https://textbookfull.com/product/optimization-of-complex-systems-
theory-models-algorithms-and-applications-hoai-an-le-thi/

Modeling and Control of Hybrid Propulsion System for

Ground Vehicles Yuan Zou

https://textbookfull.com/product/modeling-and-control-of-hybrid-
propulsion-system-for-ground-vehicles-yuan-zou/

Structural Health Monitoring System for Synthetic

Hybrid and Natural Fiber Composites Mohammad Jawaid

https://textbookfull.com/product/structural-health-monitoring-
system-for-synthetic-hybrid-and-natural-fiber-composites-
mohammad-jawaid/

Metaheuristic Algorithms for Image Segmentation Theory

and Applications Diego Oliva

https://textbookfull.com/product/metaheuristic-algorithms-for-
image-segmentation-theory-and-applications-diego-oliva/

Machine Learning Algorithms for Industrial Applications

Santosh Kumar Das

https://textbookfull.com/product/machine-learning-algorithms-for-
industrial-applications-santosh-kumar-das/
Lecture Notes in Control and Information Sciences 478

Fabien Lauer
Gérard Bloch

Hybrid
System
Identification
Theory and Algorithms for Learning
Switching Models
Lecture Notes in Control and Information
Sciences

Volume 478

Series editors
Frank Allgöwer, Stuttgart, Germany
Manfred Morari, Zürich, Switzerland

Series Advisory Board

P. Fleming, University of Sheffield, UK
P. Kokotovic, University of California, Santa Barbara, CA, USA
A. B. Kurzhanski, Moscow State University, Russia
H. Kwakernaak, University of Twente, Enschede, The Netherlands
A. Rantzer, Lund Institute of Technology, Sweden
J. N. Tsitsiklis, MIT, Cambridge, MA, USA
This series aims to report new developments in the fields of control and information
sciences—quickly, informally and at a high level. The type of material considered
for publication includes:
1. Preliminary drafts of monographs and advanced textbooks
2. Lectures on a new field, or presenting a new angle on a classical field
3. Research reports
4. Reports of meetings, provided they are
(a) of exceptional interest and
(b) devoted to a specific topic. The timeliness of subject material is very
important.

More information about this series at http://www.springer.com/series/642

Fabien Lauer Gérard Bloch
•

Hybrid System Identification

Theory and Algorithms for Learning
Switching Models

123
Fabien Lauer Gérard Bloch
LORIA CRAN
Université de Lorraine, CNRS Université de Lorraine, CNRS
Nancy, France Nancy, France

ISSN 0170-8643 ISSN 1610-7411 (electronic)

Lecture Notes in Control and Information Sciences
ISBN 978-3-030-00192-6 ISBN 978-3-030-00193-3 (eBook)
https://doi.org/10.1007/978-3-030-00193-3

Library of Congress Control Number: 2018954018

© Springer Nature Switzerland AG 2019

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. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

This book is about hybrid system identification, a field concerned with the theory
and practice of building mathematical models of dynamical systems switching
between different operating modes, from experimental observations. Mathematical
models have always been of primary importance in many branches of science and
especially in automatic control, where they constitute the first building block of
most analysis and control schemes. However, in many cases, such models cannot be
easily built from first principles or contain unknown parameters. System identifi-
cation precisely aims at estimating these parameters or, more generally, the models,
from experimental observations of the system behavior, usually taking the form of
input–output data. When the system under consideration involves multiple oper-
ating modes or discontinuities, this usually calls for the estimation of hybrid models
that can precisely take these particularities into account.
While system identification is now a well-established field, starting in the 1960s
with the availability of digital measurements, the subject of hybrid system identi-
fication has only been more recently developed, with most of the activity spanning
only the last 15 years or so. Nonetheless, it already served in many applications in
automotive control, computer vision, or even system biology, and numerous
advances were made in that period. Therefore, we felt it was time for a book
dedicated to this topic which could introduce new researchers to this passionating
field with many ramifications. Indeed, hybrid system identification lies at the
intersection of a number of disciplines, including control theory (for the analysis of
switched dynamical systems), system identification (for the estimation of models of
dynamical systems), and machine learning (for the classification of the data into the
different operating modes and the estimation of nonparametric models).
Optimization and computational complexity also play important roles here, perhaps
much more pronounced than in traditional system identification. This book covers
all the necessary background material on these fields in order to grasp the main
issues and understand the rationale behind the main methods of hybrid system
identification. After these overviews, we emphasize what makes the identification
of hybrid systems so different from classical system identification and how much
more complex the task can be. Then, practical methods leveraging the complexity

v
vi Preface

bottleneck are presented and the book ends with some more advanced material,
including a chapter on nonlinear hybrid systems with strong connections to
machine learning. Throughout the book, examples and figures are given to illustrate
the key points and we tried to keep the presentation accessible to most researchers
or graduate students in a control-related field, even when discussing, e.g., com-
putational complexity or machine learning topics.
It is worth noting that the keywords “hybrid system” may have multiple
meanings in science (and elsewhere), which can introduce some confusion
regarding the subject of this book. Here, we define hybrid systems as systems
switching between different operating modes, which amounts in technical terms to
systems combining continuous and discrete states, the discrete state indicating the
current operating mode. But, as an example, a growing community uses the key-
word “hybrid” to describe systems that combine continuous-time and discrete-time
behaviors, which is a completely different topic. We believe that the terminology
that we chose for this book is the dominant one in the control community. In
particular, it is in line with the terminology used for the names of Technical
Committees (TC) of the two major control societies (the IFAC TC 1.3 “Discrete
Event and Hybrid Systems” and the IEEE Control Systems Society TC “Hybrid
Systems”), relevant journals (e.g., “Nonlinear Analysis: Hybrid Systems”), and
conferences (e.g., “Hybrid Systems: Computation and Control” or “Analysis and
Design of Hybrid Systems”).
Besides, using “switched system” instead of “hybrid system” would have
introduced another confusion. Indeed, we here precisely divide “hybrid systems”
into the “switched” and “piecewise” subclasses, the first referring to arbitrarily
switched systems and the second one to systems switching according to a partition
of the regression/state space.
Our interest in hybrid system identification can be traced back to 2007, while
Fabien was in the middle of his Ph.D. supervised by Gérard, and we were devel-
oping machine learning techniques for system identification. At that time, Fabien
attended the 2nd HYCON Ph.D. School on Hybrid Systems in Sienna, Italy, and
came back with early thoughts on a new algorithm mixing ideas from support
vector regression and the algebraic approach of Sect. 6.1.1. Then, Fabien met with
René Vidal in 2008 while presenting this work at the HSCC conference in St.
Louis, MO, USA. This initiated our collaboration with the main contributor to the
algebraic approach, with whom we generalized the algorithm to what is now known
as the continuous optimization approach of Sect. 6.1.2 and drafted its nonlinear
extension presented in Sect. 9.1.
We further worked on this nonlinear extension with our former Ph.D. student Van
Luong Le, with whom we also developed a number of other techniques, including
the sparsity enhancing selective ‘1 -minimization sketched in Sect. 6.2.3.2,
the method to learn piecewise smooth models via convex optimization exposed in
Sect. 7.2 and the sparsity-based method for nonlinear hybrid systems described in
Sect. 9.2. The later is largely inspired by the work of Laurent Bako, who introduced
the error sparsification method of Sect. 6.2.3.
Preface vii

Beyond many interesting conversations on sparsity and hybrid system identifi-

cation, Laurent also deserves special thanks as this is with him that we first dis-
cussed the idea of writing a book on this topic in 2011 and drafted a preliminary
table of contents in 2013. However, we all had other duties and the project did not
go much forward during some time. In 2016, we picked it up again only the two of
us; this was easier since we were both working in Nancy. So we could work on a
more regular basis and ended up actually writing this book, after some rethinking
of the table of contents.
During these years, we benefited from interactions with other colleagues that
worked with us on hybrid system identification or related topics. In particular, we
had the chance to work with Hoai An Le Thi and Pham Dinh Tao on DC pro-
gramming techniques and with Henrik Ohlsson on sparse recovery issues.
Fabien also warmly thanks Christoph Schnörr for hosting him during his postdoc
at the University of Heidelberg and Yann Guermeur at the LORIA for providing
him the necessary means and freedom to work on this project.
Finally, writing such a book is rather time consuming, and we thank our fam-
ilies. Fabien thinks of Caroline, Adam and Noémie, and Gérard of Patricia, Antonin
and Etienne, the latter even if the book subject is far from laser–matter interactions
and Ultrafast Phenomena.

Nancy, France Fabien Lauer

July 2018 Gérard Bloch
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 What are Hybrid Systems? . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Dynamical Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Hybrid Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 What is System Identification? . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Outline of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 System Identification . . . . . . . . . . . . . . . . . . . . . . . ... . . . . . . . . . 15
2.1 Input–Output (I/O) Models . . . . . . . . . . . . . . . ... . . . . . . . . . 16
2.1.1 Models, Predictor, and Prediction Error .. . . . . . . . . . 16
2.1.2 Parameter Estimation . . . . . . . . . . . . . ... . . . . . . . . . 22
2.1.3 Optimization . . . . . . . . . . . . . . . . . . . ... . . . . . . . . . 24
2.2 State-Space (SS) Models . . . . . . . . . . . . . . . . ... . . . . . . . . . 28
2.2.1 Models and Properties . . . . . . . . . . . . ... . . . . . . . . . 28
2.2.2 Parameter Estimation . . . . . . . . . . . . . ... . . . . . . . . . 34
2.3 Recursive Identification . . . . . . . . . . . . . . . . . ... . . . . . . . . . 39
2.4 Nonlinear System Identification (★) . . . . . . . . ... . . . . . . . . . 41
2.4.1 Parametric Models . . . . . . . . . . . . . . ... . . . . . . . . . 41
2.4.2 Nonparametric Models . . . . . . . . . . . ... . . . . . . . . . 42
2.5 Model Selection and Assessment . . . . . . . . . . ... . . . . . . . . . 52
2.5.1 Model Assessment . . . . . . . . . . . . . . ... . . . . . . . . . 52
2.5.2 Model Selection . . . . . . . . . . . . . . . . ... . . . . . . . . . 53
Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... . . . . . . . . . 55
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... . . . . . . . . . 56

ix
x Contents

3 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.1 Discrimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.1.1 Binary Linear Classification . . . . . . . . . . . . . . . . . . . . 61
3.1.2 Multi-class Problems . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.1.3 Nonlinear Classification (★) . . . . . . . . . . . . . . . . . . . . 70
3.2 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4 Hybrid System Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.1 Hybrid System Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.1.1 State-Space Versus Input–Output Models . . . . . . . . . . 78
4.1.2 Linear Versus Nonlinear Submodels . . . . . . . . . . . . . . 79
4.1.3 Piecewise Smooth Versus Arbitrarily Switched
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...... 82
4.2 Identification Problems . . . . . . . . . . . . . . . . . . . . . . . . ...... 86
4.2.1 Hybrid System Identification with
Unknown Mode . . . . . . . . . . . . . . . . . . . . . . ...... 86
4.2.2 The Trade-Off Between the Number of Modes
and the Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2.3 Fixing the Number of Submodels . . . . . . . . . . . . . . . . 88
4.2.4 Fixing a Bound on the Error . . . . . . . . . . . . . . . . . . . 90
4.2.5 Hybrid Model Assessment . . . . . . . . . . . . . . . . . . . . . 93
4.3 Other Related Problems (★) . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3.1 Nonlinear System Identification . . . . . . . . . . . . . . . . . 95
4.3.2 Subspace Clustering . . . . . . . . . . . . . . . . . . . . . . . . . 96
Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5 Exact Methods for Hybrid System Identification . . . . . . . . . . . . . . 103
5.1 Straightforward Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.1.1 Switching Regression with Fixed s . . . . . . . . . . . . . . . 105
5.1.2 Bounded-Error Estimation . . . . . . . . . . . . . . . . . . . . . 106
5.1.3 Piecewise Affine Regression with Fixed s . . . . . . . . . . 108
5.2 Hardness Results (★) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.2.1 Basics in Computational Complexity . . . . . . . . . . . . . 109
5.2.2 Hardness of Switching Regression . . . . . . . . . . . . . . . 111
5.2.3 Hardness of PWA Regression . . . . . . . . . . . . . . . . . . 116
5.2.4 Hardness of Bounded-Error Estimation . . . . . . . . . . . . 118
5.3 Polynomial-Time Algorithms for Fixed Dimensions . . . . . . . . . 118
5.3.1 PWA Regression with Fixed s and d . . . . . . . . . . . . . 119
5.3.2 Switching Regression with Fixed s and d . . . . . . . . . . 122
5.3.3 Bounded-Error Estimation with Fixed d . . . . . . . . . . . 124
Contents xi

5.4 Global Optimization with Branch-and-Bound . . . . . . . . . . . . . 126

5.4.1 Switching Regression . . . . . . . . . . . . . . . . . . . . . . . . 127
5.4.2 Bounded-Error Estimation . . . . . . . . . . . . . . . . . . . . . 132
5.4.3 PWA Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.5 The Need for Approximation Schemes/Heuristics . . . . . . . . . . 137
Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6 Estimation of Switched Linear Models . . . . . . . . . . . . . . . . . . . . . 141
6.1 Fixed Number of Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.1.1 Algebraic Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.1.2 Continuous Optimization Approach . . . . . . . . . . . . . . 145
6.1.3 Block-Coordinate Descent Approach . . . . . . . . . . . . . 148
6.2 Free Number of Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.2.1 Bounded-Error Approach . . . . . . . . . . . . . . . . . . . . . . 151
6.2.2 Block-Coordinate Descent Approach . . . . . . . . . . . . . 154
6.2.3 Error Sparsification Method . . . . . . . . . . . . . . . . . . . . 155
6.2.4 Parameter Sparsification Method . . . . . . . . . . . . . . . . 159
Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
7 Estimation of Piecewise Affine Models . . . . . . . . . . . . . . . . . . . . . . 169
7.1 From Switched Affine Models to PWA Models . . . . . . . . . . . . 170
7.2 From Nonlinear Models to PWA Models . . . . . . . . . . . . . . . . 171
7.2.1 Limitation of the Classical
Regularization Schemes . . . . . . . . . . . . . . . . . . . . . . . 172
7.2.2 Local Regularization . . . . . . . . . . . . . . . . . . . . . . . . . 172
7.2.3 Learning Smooth Models of PWA Functions
by Convex Optimization . . . . . . . . . . . . . . . . . . . . . . 173
7.2.4 Recovering PWA Models . . . . . . . . . . . . . . . . . . . . . 175
7.3 From Local Models to PWA Models . . . . . . . . . . . . . . . . . . . 175
7.3.1 Fitting Local Models over Neighborhoods . . . . . . . . . 177
7.3.2 Controlling the Variations of the Local Models . . . . . . 178
7.3.3 Clustering the Local Models . . . . . . . . . . . . . . . . . . . 179
Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
8 Recursive and State-Space Identification of Hybrid Systems . . . . . 183
8.1 Input–Output Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
8.1.1 Parallel Identifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
8.1.2 Model Tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
8.2 State-Space Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
8.2.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.2.2 Off-Line Approaches . . . . . . . . . . . . . . . . . . . . . . . . . 194
xii Contents

Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
9 Nonlinear Hybrid System Identification (★) . . . . . . . . . . . . . . . . . . 205
9.1 Continuous Optimization Approach for Switched
Nonlinear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
9.1.1 Overall Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
9.1.2 A Few Methods for Structure Selection . . . . . . . . . . . 208
9.2 Error Sparsification Approach for Switched
Nonlinear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . .... . . 213
9.2.1 Choice of the Regularizer . . . . . . . . . . . . . . . .... . . 214
9.2.2 Sparsity Versus Compressibility . . . . . . . . . . . .... . . 215
9.3 Sum-of-Norms Approach for Piecewise Smooth Models ... . . 218
9.3.1 Controlling the Variations of Nonlinear
Local Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
9.3.2 Clustering Local Models in the RKHS . . . . . . . . . . . . 221
Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
10 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
10.1 Other Forms of Hybrid Models . . . . . . . . . . . . . . . . . . . . . . . 228
10.1.1 Input–Output Models . . . . . . . . . . . . . . . . . . . . . . . . . 228
10.1.2 Nonlinear Submodels . . . . . . . . . . . . . . . . . . . . . . . . 228
10.1.3 Continuous-Time Models . . . . . . . . . . . . . . . . . . . . . . 229
10.2 Computational Complexity and Exact Methods . . . . . . . . . . . . 229
10.3 Statistical Guarantees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
10.4 Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
10.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Appendix A: Basics of Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Appendix B: Basics of Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
Notations

General Notations

Scalars are written in default and lowercase letters, e.g., a, unless otherwise
specified.
Vectors are column vectors written in boldface and lowercase letters, e.g., a,
unless otherwise specified. For a vector a, the nth entry is denoted an . All
inequalities between vectors are meant entrywise.
Matrices are written in boldface and uppercase letters, e.g., A, unless otherwise
specified. For a matrix A, the entry in the nth row and the pth column is denoted
Anp .

Sets

N The set of natural numbers

Q The set of rational numbers
R The set of real numbers
An The set of n-dimensional vectors, with entries in the set A
An
p The set of n
p-dimensional matrices, with entries in the set A
½n The set of integers from 1 to n
YX The set of functions from the set X into the set Y
2 Set membership; e.g., x 2 A means that x is an element of A
[ Set union, e.g., A [ B ¼ fx : x 2 A or x 2 Bg
\ Set intersection, e.g., A \ B ¼ fx : x 2 A and x 2 Bg
;  Set inclusions, e.g., A  B , 8x 2 A, x 2 B, while A  B , A  B and
A 6¼ B
n Set difference, e.g., AnB ¼ fx 2 A : x 62 Bg
jAj Cardinality of the set A

xiii
xiv Notations

Vectors and Matrices

1 Vector or matrix of dimension clear from context or written in subscript with all
entries equal to 1
0 Vector or matrix of dimension clear from context or written in subscript with all
entries equal to 0
I Identity matrix of dimension clear from context or written in subscript with
main diagonal entries equal to 1 and all other entries equal to 0

Operations

a_ Time derivative of the vector a

a> ; A> Transpose of the vector a or matrix A; see (B.3)
jaj Absolute value of the scalar a
ðaÞ þ ; ðaÞ Positive and negative parts of a, ðaÞ þ ¼ maxf0; ag,
ðaÞ ¼ minf0; ag
ðaÞ þ ; ðaÞ Positive and negative parts of the vector a, defined entrywise
dae Ceiling function returning the smallest integer greater than or equal
to a
dn; p Kronecker delta which is 1 if n ¼ p and is 0 if n 6¼ p
detðAÞ Determinant of the (square) matrix A; see (B.23)
diagðaÞ Diagonal matrix with the entries of a on the diagonal; see (B.6)
eA Exponential of the (square) matrix A; see (B.10)
E Mathematical expectation; see (A.2), (A.3)
rf Gradient of the scalar-valued function f ; see (B.17)
r2 f Hessian of the scalar-valued function f ; see (B.18)
inf Infimum
Df Jacobian of the vector-valued function f ; see (B.18)
log Natural logarithm
max Maximum
min Minimum
OðgÞ f ¼ OðgÞ, asymptotic notation meaning that, for f : N ! N and
g : N ! N, 9n0 2 N; c 2 N, such that 8n  n0 , f ðnÞ  cgðnÞ
P Probability measure; see Appendix A.1
rankðAÞ Rank of the matrix A; see (B.22)
signðaÞ Sign of the number a, signðaÞ ¼ 1a  0  1a\0
sup Supremum
trðAÞ Trace of the (square) matrix A; see (B.4), (B.35)
vecðAÞ Vectorization of the matrix A; see (B.12)
Indicator function, is if the boolean expression A is true and 0
  otherwise  
n n n!
Binomial coefficient, ¼ p!ðnpÞ!
p p
Notations xv

Kronecker product; see (B.11)

h ; iv Inner product in the space V (V can be omitted when clear from
context); see Appendix B.2.3

Norms

kakp ‘p -norm of the vector a; see (B.1)

kak0 ‘0 -pseudo-norm of the vector a; see (B.2)
kAkF Frobenius norm of the matrix A; see (B.5)

Specific Notations

As a subscript, the letter j always refers to a mode or submodel index for hybrid
models or a category index in classification.
As a subscript, the letter k always refers to the discrete time or a data point index.
^
a Estimate of a
CE Classification error rate; see (4.27)
d Dimension of x, the regression vector or input vector of classifiers
D Data set of regressor–output pairs; see (2.24)
e, e Prediction error; see (2.6, 2.25, 2.50)
e; e Noise term in I/O models
f Real-valued function, model to be estimated
F Model space (a real-valued function class)
FIT Normalized criterion of fit; see (2.83)
/ Nonlinear mapping to feature space; see (2.75)
g Classifier, function with output in a finite set
g Parameter vector of a linear classifier
G Set of classifiers (an integer-valued function class with finite co-domain)
~gi Coefficients of the impulse response of a SISO system; see (2.1)
~hi Coefficients of the disturbance impulse response; see (2.2)
H Reproducing kernel Hilbert space; see Definition 2.2
j Index of mode, submodel, or category, with j 2 Q
J A cost function to be minimized
k Discrete time or data point index
K Kernel function; see Definition 2.1
K Kernel matrix with Kik ¼ Kðxi ; xk Þ; see (2.65)
‘ Loss function; see (2.27)
‚ Regularization hyperparameter
MNPE Mean normalized parametric error; see (4.26)
MSE Mean squared error; see (2.82)
xvi Notations

na Number of lagged outputs in the regression vector (SISO systems)

nb Number of lagged inputs in the regression vector (SISO systems)
nd Pure delay (SISO systems)
nx Number of states, dimension of the state vector
nu Number of inputs, dimension of the input vector
ny Number of outputs, dimension of the output vector
N Number of (training) data points
NPE Normalized parametric error; see (2.80)
X Continuous state-input domain
q Discrete state, index of mode or category, with q 2 Q
Q Set of mode or category indexes, of cardinality s (usually Q ¼ ½s)
q Forward shift operator
R Regularizer; see (2.63)
s Number of modes and submodels or number of categories in classification
t Continuous time
h Parameter vector
u; u System input
v Disturbance term in SISO I/O models
v Measurement noise term in SS models
w State noise term in SS models
x (continuous) State vector
x Regression vector or input vector of classifiers
X Observation matrix composed of regression vectors as
X ¼ ½ x1 . . . x k . . . x N  T
X Regression domain or input space for classification
X/ Feature space
y; y System output
Y Output domain
y Target vector defined as y ¼ ½ y1 . . . yk . . . yN T

Abbreviations

AFMM Adaptive Forgetting through Multiple Models (algorithm); see Sect. 8.

1.2.2
ARMAX AutoRegressive Moving Average with eXogenous input (model); see
(2.16)
ARX AutoRegressive with eXogenous (eXternal) input (model); see (2.9)
BJ Box–Jenkins (model); see (2.22)
FIR Finite Impulse Response (model); see (2.7)
FVS Feature Vector Selection; see Sect. 9.1.2.2
HDC Hybrid Decoupling Constraint; see (6.2)
Notations xvii

I/O Input–Output (model); see (1.2, 1.4)

KPCR Kernel Principal Component Regression; see Sect. 9.1.2.3
MCS Multi-level Coordinate Search (minimizing algorithm); see Sect. 6.1.2.1
ME Minimum-of-Errors (problem); see (4.21, 6.8)
MILP Mixed-Integer Linear Program; see (5.51)
MIN PFS Minimum Number of Feasible Subsystems Problem; see Sect. 6.2.1
MINLP Mixed-Integer NonLinear Program; see (4.19, 8.13)
MIQP Mixed-Integer Quadratic Program; see (5.50)
NLMS Normalized Least Mean Squares (algorithm); see (2.59)
OE Output Error (model); see (2.18)
pdf probability density function
PE Product-of-Errors (problem); see (6.9, 9.1)
PEM Prediction Error Method; see (2.26)
PWA PieceWise Affine (model); see Sect. 4.1.3.2
PWARX PieceWise ARX (model); see Sect. 4.1.3.2
PWS PieceWise Smooth (model); see Sect. 4.1.3.2
RKHS Reproducing Kernel Hilbert Space; see Definition 2.2
RLS Recursive Least Squares (algorithm); see (2.60)
s.t. subject to
SARX Switched ARX (model); see Sect. 4.1.3.1
SHS Stochastic Hybrid Systems; see Sect. 1.1
SISO Single-Input, Single-Output (model)
SS State-Space (model); see (1.1, 1.3, 2.34, 2.47)
SVD Singular Value Decomposition; see (B.2.8)
SVM Support Vector Machine; see Sect. 3.1.1.2
SVR Support Vector Regression; see (2.78)
vs. versus
w.r.t. with respect to
List of Figures

Fig. 1.1 Two examples of PWA function: a saturated affine function

and one with a dead zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Fig. 1.2 Roadmap of the book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Fig. 2.1 A one-dimensional convex function . . . . . . . . . . . . . . . . . . . . . . . 26
Fig. 2.2 Empirical and expected risks vs. the model complexity . . . . . . . . 44
Fig. 2.3 Effect of the regularization parameter ‚ on kernel
ridge regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 49
Fig. 2.4 The -insensitive loss function . . . . . . . . . . . . . . . . . . . . . . . . . .. 52
Fig. 3.1 Optimal separating hyperplane . . . . . . . . . . . . . . . . . . . . . . . . . .. 64
Fig. 3.2 Influence of C on the margin of the support vector machine
classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 65
Fig. 3.3 Multi-class problem with 3 categories . . . . . . . . . . . . . . . . . . . . .. 69
Fig. 3.4 K-means Algorithm 2 in action . . . . . . . . . . . . . . . . . . . . . . . . . .. 73
Fig. 4.1 Partition of the regression space into four convex polyhedra
in dimension 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 83
Fig. 4.2 A one-dimensional switching linear regression problem
with two modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 89
Fig. 4.3 A one-dimensional PWA regression problem
with two modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 91
Fig. 4.4 A one-dimensional switching regression problem with two
modes tackled by a greedy bounded-error approach . . . . . . . . . .. 92
Fig. 4.5 A one-dimensional PWA regression problem with two modes
tackled by a generic nonlinear regression method . . . . . . . . . . . .. 96
Fig. 4.6 Illustration of the difference between regression
and subspace estimation (a) . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 98
Fig. 4.7 Illustration of the difference between regression and subspace
estimation (b) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 98
Fig. 5.1 Reduction of the partition problem to a switching regression
problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Fig. 5.2 Reduction of the partition problem to a PWA
regression problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

xix
xx List of Figures

Fig. 5.3 One-dimensional example of binary PWA regression solved

exactly by Algorithm 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Fig. 5.4 Illustration of Proposition 5.1 for s ¼ 2 and d ¼ 1 . . . . . . . . . . . . 123
Fig. 5.5 Illustration of Lemma 5.2 for d ¼ 1 . . . . . . . . . . . . . . . . . . . . . . . 125
Fig. 5.6 Illustration of the splitting procedure when s ¼ 2 and d ¼ 1 . . . . 129
Fig. 5.7 Hinge functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Fig. 6.1 Illustration of the continuity of the pointwise minimum
of several continuous functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
Fig. 6.2 Probability of success of the k-LinReg Algorithm . . . . . . . . . . . . 151
Fig. 6.3 Exponentially decreasing -insensitive loss function . . . . . . . . . . . 153
Fig. 6.4 Basic idea of the error sparsification approach . . . . . . . . . . . . . . . 155
Fig. 7.1 A one-dimensional PWA regression problem tackled
via switching regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Fig. 7.2 Learning a PWA target function from a noisy data
set with local or global regularization . . . . . . . . . . . . . . . . . . . . . . 176
Fig. 7.3 Clustering local models for PWA regression. . . . . . . . . . . . . . . . . 180
Fig. 9.1 Illustration of the nonlinear error sparsification method
for Example 9.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Fig. 9.2 Iterations of the reweighting process in Example 9.2 . . . . . . . . . . 218
Fig. 9.3 Example of PWS regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Fig. 9.4 Clustering of the local nonlinear models
for PWS regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
List of Tables

Table 2.1 Common loss functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

Table 5.1 Computing time for the global optimization
of a switching linear model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Table 5.2 Comparison of the computing time of the polynomial-time
algorithm and the MIQP approach for PWA regression . . . . . . . 139

xxi
Chapter 1
Introduction

This book is about hybrid system identification, which is a topic at the intersection of
a number of fields, including hybrid dynamical systems, system identification, and
machine learning. This first chapter briefly introduces these subjects and presents a
number of applications, before giving an overview of the content of the book.

1.1 What are Hybrid Systems?

Hybrid dynamical systems—hybrid systems for short—are first and foremost dynam-
ical systems.1 A dynamical system can be broadly defined as a set of interacting
components and a “rule” determining its evolution during time, with descriptive
variables, internal (the “states”) or linking the set with its environment (the “inputs”
and “outputs”).

1.1.1 Dynamical Systems

Dynamical systems can be broken down into two classes, according to the nature
of the involved variables: continuous or discrete (dynamical) systems. For discrete
systems, more precisely discrete-state systems, the state variables are countable; that
is, each one takes a finite number of values, most often in a subset of N. In continuous
systems or analog systems, each variable can take an infinite number of values, most
often in R. The behavior of continuous systems is given by relationships, or “models,”
involving the derivatives (ordinary differential equations, in continuous time) or the
delayed values (difference equations, in discrete time) of the variables.

1 Note however that static phenomena can also be modeled by resorting to hybrid system identifica-

tion methods as described in this book. In fact, due to the difficulty of hybrid system identification,
the dynamical aspect of hybrid systems is mostly ignored in many dedicated approaches.

© Springer Nature Switzerland AG 2019 1

F. Lauer and G. Bloch, Hybrid System Identification,
Lecture Notes in Control and Information Sciences 478,
https://doi.org/10.1007/978-3-030-00193-3_1
2 1 Introduction

Two general model forms of continuous systems can be chosen. The first one is
the so-called state-space (SS) form, which can be written in continuous time t ∈ R
as 
ẋ(t) = f(x(t), u(t))
y(t) = g(x(t), u(t)),

or in discrete time k ∈ N as

xk+1 = f(xk , uk )
(1.1)
yk = g(xk , uk ),

where xk ∈ Rn x , uk ∈ Rn u , and yk ∈ Rn y are, respectively, the continuous (internal)

“state,” input and output vectors at time k. Note that in almost all of what follows,
discrete-time systems are considered.
Another general form of discrete-time models is the input–output (I/O) form:

yk = f (x k ), (1.2)

where yk ∈ Rn y is the output vector, and the “regression vector” x k ∈ Rd , although

depending on the chosen subclass, typically includes lagged values of the input uk−i
and output yk−i . Note also that this regression vector x is not the state vector x and
must not be confused with it, particularly in Chap. 8, where both I/O and SS models
are dealt with.
In the preceding deterministic formulations, no noise is considered. We can take
into account uncertainty in the models, for instance measurement noise or modeling
errors, and this leads to the noisy variants of (1.1) and (1.2):

xk+1 = f(xk , uk , w k )
(1.3)
yk = g(xk , uk , v k ),

where wk ∈ Rn x and v k ∈ Rn y are, respectively, the state and measurement noise

terms, and
yk = f (x k , ε k ), (1.4)

where εk ∈ Rn y is the noise term.

1.1 What are Hybrid Systems? 3

1.1.2 Hybrid Systems

Hybrid systems include both interacting continuous and discrete dynamical behav-
iors.2 This results in systems that switch, according to the value of the discrete
variables, between different operating modes with continuous dynamics modeled as
in (1.3) or (1.4). Thus, a hybrid system can be formulated in SS form as

xk+1 = fqk (xk , uk , w k )
(1.5)
yk = gqk (xk , uk , v k ),

where xk , uk , yk , wk , v k are as in (1.3) and the new integer variable qk ∈ {1, . . . , s},
known as the discrete state or “mode,” selects which one among the s “submodels”
{(f j , g j )}sj=1 is active at time k. Note that we restrain the formulation here to a single
discrete variable qk . However, if the system involves multiple discrete variables,
each combination of values for these variables leads to a different mode and can be
encoded as a single value for qk .
For I/O models, this gives

yk = f qk (x k , ε k ), (1.6)

with yk , x k , ε k as in (1.4) and the discrete state qk ∈ {1, . . . , s} selecting the active
submodel among { f j }sj=1 at time k.

1.1.2.1 Switching Mechanism

We consider only two ways by which the value of the discrete state qk can change
during time. In the first one, qk is arbitrarily switched, both in its value and in the
change instant. This corresponds to a discrete state generated externally, for instance
by a supervisor, which forces the system to change its continuous behavior.
 state qk depends on the continuous
In the second way, for SS models, the discrete
x
state xk and input uk vectors, i.e., qk = j if k ∈ Ω j , where the Ω j ’s are regions
uk
that form a partition of the state-input domain Ω. Similarly, for I/O models, the
discrete state qk depends on the continuous regression vector x k , i.e., qk = j if
x k ∈ X j , where the X j ’s are regions that form a partition of the whole regression
space X . Such piecewise-defined models are often used to approximate a single
nonlinear continuous behavior by a collection of linear or affine submodels, each of
which is valid only in a particular region. These piecewise affine models, commonly
abbreviated as PWA, have the capability of universally approximating any nonlinear
dynamics, with arbitrary precision.

2 Notethat the keywords “hybrid systems” can have different meanings in different contexts. Our
motivations for using this terminology are detailed in the Preface.
4 1 Introduction

4 4

2 2

0 0

-2 -2

-4 -4
-4 -2 0 2 4 -4 -2 0 2 4

Fig. 1.1 Two examples of PWA function: a saturated affine function (left) and one with a dead
zone (right)

Figure 1.1 shows two examples of PWA functions of a scalar variable x with a
partition of X = [−4, 4] into three regions. The first PWA function,
⎧
⎪
⎨−3, if x < −1
f (x) = max {−3, min{2, 2x − 1}} = 2x − 1, if x ∈ −1, 23
⎪
⎩
2, if x > 23 ,

is a saturated affine function. Such saturations are encountered in many applications.

The second PWA function,

x, if |x| > 1
f (x) =
0, otherwise,

shows a dead zone in the interval [−1, 1] which introduces discontinuities. Such a
phenomenon occurs for instance in DC motors.

1.1.2.2 Other Hybrid Models

Many models of hybrid systems have been proposed in the literature, with a rather
confused terminology.
Linear complementarity systems are discussed in [1–3], mixed logical dynamical
(MDL) systems in [4], max-min-plus-scaling (MMPS) systems in [5]. Their equiv-
alence with PWA systems and the transformation from one model into the other are
exposed in [6]. Reference [7] regroups them in the class of discrete hybrid automata
(DHA) with an associated high-level modeling language and a set of tools for trans-
lating DHA into any of the former hybrid models.
Besides, stochastic hybrid systems (SHSs), where, particularly, the mode transi-
tions are triggered randomly or the new mode after transition is selected randomly,
1.1 What are Hybrid Systems? 5

are not treated in this book. One can refer to [8] for an introduction. Reference [9]
provides a comprehensive review of Markovian jump systems, where the transi-
tions between the submodels are determined by a Markov chain, and the references
therein include real-world applications, such as economic systems, flight systems,
power systems, communication systems, and networked control systems.

1.1.2.3 Connections with Other Models

In what precedes, hard switchings are considered between the submodels. There are
other possible ways to define transitions between submodels, as in linear parameter-
varying (LPV) systems, where the parameters of a linear model evolve (most often)
smoothly over time [10, 11]. In that sense, hybrid systems, as described here, can
be considered as particular LPV systems with abrupt and nonsmooth parameter
transitions.
There is also an important literature on multiple models. As in piecewise-defined
hybrid systems, in multiple models, the operating space of the system is split into a
finite number of regions, but these latter are possibly overlapping. Moreover, the cur-
rent model is interpolated from the set of the submodels associated to the regions. The
aggregation mechanism is based on a weighting function which quantifies the prox-
imity of the current data to the submodel regions. This includes various approaches,
such as fuzzy models, including Takagi–Sugeno models, or neuro-fuzzy models
[12–15]. In the machine learning community, the mixture of experts [16] provides
a model in which a gating network decides how to compute the output given those
of an ensemble of submodels. This gating network can be made to yield either hard
switchings between the submodels or combinations of their outputs.

1.2 What is System Identification?

System identification aims at building models of dynamical systems mainly from

experimental data. Some prior knowledge, particularly from first principles, can be of
help in selecting the input variables, choosing the model class and structure, or adding
constraints linked to the system physical behavior, to yield parametric models. Then,
system identification amounts to the estimation of a finite number of parameters in the
model. Alternatively, when lacking precise knowledge on the physical phenomena
involved, nonparametric models can be used to estimate unknown nonlinearities. In
this case, system identification includes the estimation of both the model structure
and its parameters from the data, which is a standard setting for machine learning
techniques.
In the parametric case, the general SS model (1.3) can be clarified as
6 1 Introduction

xk+1 = f(xk , uk , w k , θ )
(1.7)
yk = g(xk , uk , v k , θ ),

exhibiting the parameters of the model grouped in the vector θ . For the I/O form
(1.4), this reads:
yk = f (x k [θ], ε k , θ ), (1.8)

where the regression vector x k can itself depend on the parameter vector θ .
System identification then means selecting a particular structure for the model,
which includes the determination of the dimension n x of the state vector for SS
models, or the components and the size of the regression vector for I/O models,
and the estimation of the parameter vector θ from recorded input-output data. This
estimation is often performed by minimizing the prediction error, i.e., the difference
between the measurements yk and the values ŷk predicted by the model. We see
here that system identification is tightly connected to the field of optimization and
optimization problems will often be encountered throughout the book.
In the case of hybrid systems, the SS model (1.7) becomes

xk+1 = fqk (xk , uk , w k, θ qk )
(1.9)
yk = gqk (xk , uk , v k, θ qk ),

while the I/O model (1.8) becomes

yk = f qk (x k [θ qk ], ε k , θ qk ), (1.10)

and we have to determine the number s of submodels in addition to all the parameter
vectors θ j . We see that to do the job, we also have to determine by which one of
the submodels the different parts of the data are governed, which is a classification
problem. Indeed, it is a fundamental feature of hybrid system identification to tightly
combine data classification and parameter estimation.
Note that in order to avoid adding further complexity to the identification problem,
we often assume a fixed structure for the submodels, that is, particularly, same state
vector or parameter vector sizes.

1.3 Applications

For about two decades, there has been increased research interest in the study of
hybrid systems [17], as they provide a suitable framework for modeling systems in a
wide range of engineering applications [18, 19], and designing hybrid control laws
[20]. In [21], a broader introduction to hybrid systems can be found. It is shown
that such systems can be encountered at various levels of detail and accuracy, and
in numerous application fields, such as chemical processes, process control systems,
1.3 Applications 7

robotic manufacturing systems, interconnected power systems, transportation and

intelligent highway systems, air traffic control, computer and communication net-
works. In the review of [22], a list of application domains for hybrid system identi-
fication is given, including computer vision, electromechanical/automotive systems,
systems biology and environmental systems.
There are many physical processes involving different dynamical behaviors. For
instance, threshold, dead zone, saturation, hysteresis phenomena are very common
and the processes including them can be better described by a hybrid model than by
a single dynamical model. Thus, most of the applications deal with piecewise affine
(PWA) systems where the piecewise character expresses the different behaviors and
the affine submodels are sufficient to describe each behavior.
In mechanical engineering, continuous motions may be interrupted by collisions.
Other examples are backslash in gears or motion systems with friction models that
distinguish between stick and slip modes. In [23, 24], the identification of the elec-
tronic component placement process in pick-and-place machines by PWA models is
described. In such a process, unilateral contact and saturation phenomena character-
ize the hybrid dynamics of the system. In [25, 26], a rotary actuator, equipped with a
speed reducer, which is commonly used in industrial robots and has strong nonlinear
friction, is modeled as a PWA system. Examples can also be easily given for elec-
tromechanical systems, as, in [27] or [28], a DC motor with saturation and dead zone.
An electronic throttle, that regulates air inflow into the engine of a car, is considered
in [29]. The model predictive control (MPC) design of this highly nonlinear compo-
nent (transmission friction and “limp-home” return spring) from a state-space PWA
model is described in [30]. PWA modeling of the nonlinear displacement from the
input voltage is also applied for a radio-frequency microelectromechanical system
(MEMS) switch in [31]. In fluid mechanics, the level in a tank with several sections
can be identified with respect to the flow rate with a PWA model, as in [32].
In electrical circuits, continuous phenomena, such as the charging of capacitors,
are interrupted by switches or diodes, or involve saturation and hysteresis, as for the
PWA modeling of an industrial current transformer in [33]. Applications in energy
management are described in [34], from DC-DC boost converter, represented by a
pair of affine continuous-time state-space equations, to a much more complicated
case, the control of an electrical power generation and distribution system. In [35],
the fault model of a track circuit in high-speed railway is identified to detect the
failure of a capacitor. In [36], an electromagnetic actuator is modeled by a state-
space piecewise linear model for fault detection and isolation by a piecewise linear
observer.
Basically, the operation of internal combustion engines is also hybrid, with the
well-known four-stroke cycle. Early, [37] described the modeling of an engine tur-
bocharger by hinging hyperplanes, which form a subclass of PWA models. In gear
shift control in automatic transmissions of cars, both continuous (throttle position,
car velocity) and discrete (gear ratio) variables are involved. Reference [38] intro-
duce various modeling and control challenges in power-train control, focusing on
fast force-transient control, cutoff control, and idle speed control. From that time,
numerous problems have been tackled in this field, as, for instance, in [7], the hybrid
8 1 Introduction

modeling of a car with robotized manual gear shift and cruise control, in [39], the
control of magnetic actuators or homogeneous charge compression ignition engine,
or, in [40], the piecewise affine modeling of NOx emission produced by a diesel
engine, or yet the driving behavior for automatic driving systems [41]. In [42], no
less than six PWA models are integrated in the complete model of a hybrid electric
vehicle (HEV) powertrain. The resulting high-fidelity model is then used for predic-
tion in a hybrid MPC control scheme. A recent comprehensive treatment of hybrid
system control in automotive engineering is given in [43].
Applications can also be found in the environmental field. For instance, in [44],
a sewer network is modeled by a mixed logical dynamical (MLD) model, a set of
linear dynamic state-space equations subject to linear mixed-integer inequalities, and
controlled by a model integer predictive control approach. In [45], an open-channel
system is modeled through a piecewise affine or switched model.
More generally, almost all control systems can also be considered as hybrid
systems. In chemical processes, the continuous evolution of chemical reactions is
controlled by valves and pumps. In thermal processes, a thermostat controlling the
temperature switches heating or cooling choices on or off, or links together different
temperature profiles, such as in [46], for the PWA modeling of an olive oil esteri-
fication reactor. Reference [47] describes the identification by hinging hyperplanes
and model predictive control of a laboratory water heater. The industrial control sys-
tems combine continuous control and logic, discrete-event controllers, that super-
vise high-level signals, trigger start-ups or shutdowns, alarms, or switch of operating
mode. Hybrid systems are natural models for computer-controlled systems since they
involve physical processes and computers. Besides, the communication between sys-
tems become more and more important and widespread. In the networked control
systems, the behavior of systems is influenced or controlled by events communicated
over the network. Such systems can be considered as complex hybrid systems.
In the slightly different but promising domain of computational biology, hybrid
systems are proposed as a framework for modeling biological networks [48], par-
ticularly gene regulatory networks [49–51], or segmenting and modeling bacterial
growth [52, 53]. Several biomedical applications of PWA system identification are
described in [54]. Note also the review [55] of applications of hybrid systems to
biological and medical systems. Besides, in computer vision, hybrid system identifi-
cation techniques can be used for segmenting sequences of images of moving scenes
(dynamic textures) and human gaits, as described in [56–58].

1.4 Outline of the Book

The content of this book can be divided into four major parts in an attempt to provide
(i) sufficient background knowledge on topics related to hybrid system identification,
(ii) a sound theoretical analysis of the main issues involved, (iii) an overview of the
relevant methods to tackle these issues in practice, and (iv) a broader view including
1.4 Outline of the Book 9

some advanced topics and open issues. However, it shall be clear that this does not
constitute an exhaustive review of the literature on this rapidly growing field.
Figure 1.2 gives a graphical representation of the connections between the chapters
of this book, whose contents are detailed below. Each chapter ends with bibliographic
Notes. Throughout the book, technical sections that can be skipped upon first reading
are marked with ().
Introductory material is given in Chaps. 2–4 with the first ones exposing the
necessary background in related fields and Chap. 4 actually presenting hybrid system
identification. Computational complexity will be introduced only in Chap. 5. Indeed,
the involved technicalities can seem far from standard concerns of a researcher in
control theory, and the general idea of computational complexity issues related to
hybrid system identification can be grasped without all the details.
Chapter 2 starts with an overview of classical system identification, common
models (state-space, input–output, linear, nonlinear, etc.) and standard estimation
methods for parametric and nonparametric models. As most estimation methods
considered in this book heavily rely on optimization, this field is also briefly intro-
duced in Sect. 2.1.3.
Then, Chap. 3 introduces classification, which constitutes the second pillar of
hybrid system identification. Two settings are distinguished depending on whether
the learning problem is supervised or not. The first case corresponds to the estimation
of the partition of the regression or state space used for determining the (current)
mode for piecewise-defined systems, while the second case is tightly connected to
hybrid system identification itself, as it amounts to assign data to the different modes.
Finally, Chap. 4 presents hybrid system identification. This chapter details the
various formulations of the hybrid system models. In particular, the first distinc-
tion, which has been introduced above, is the one between state-space models and
input–output models, these latter being considered in the majority of the works. The
second concerns the nature of the submodels. The preceding formulations were very
general, but linear models have been extensively investigated with strong theoretical
results. So, linear submodels are mainly considered in the hybrid system literature.
They are presented versus nonlinear ones. Finally, the third opposition, piecewise
systems/arbitrarily switched systems, is introduced. The second part of the chapter
poses the hybrid system identification problem and the inherent trade-off between
the number of modes and the error of the model. This leads to different variants of
the problems, depending on the assumptions and on which part of the trade-off is
focused on. The chapter ends with a brief discussion on other problems studied in
other fields that are related to hybrid system identification.
The following three chapters focus on linear hybrid systems in I/O form.
Theoretical material is exposed in Chap. 5, which focuses on algorithmic com-
plexity and global optimization issues. The chapter starts with a brief introduction
to computational complexity. Then, we analyze the opportunity of solving exactly
the optimization problems set up in the previous chapter. In particular, it is shown
that these problems are N P-hard, i.e., that we cannot hope to solve them in all cir-
cumstances. Yet, the chapter also details algorithms that can be guaranteed to yield
exact (or sufficiently close to exact) solutions in reasonable time for low-dimensional
Another random document with
no related content on Scribd:
am the one!” to be able to see him, and know that at last he had got
rid of that job at Vanderlynden’s. But nothing happened. It was
always just going to happen.
At length the Division moved right up into the coal-fields and sat
down by a slag heap near Béthune. Then Colonel Birchin called to
him one morning across the office: “I say, Dormer, I’ve got the
whereabouts of that fellow Chirnside. He’s near Rheims. You’ll have
to go.”
Dormer went. For two whole days he travelled across civilian France.
France of the small farm, the small town, and the small villa. Far
beyond the zone of the English Army, far beyond the zone of any
army, he passed by Creil to Paris, and from Paris on again into a
country of vine-clad hills above a river. He was in a part where he
had never been as a soldier, never gone for one of those brief
holidays to Switzerland he had sometimes taken. It caused him
much amusement to think of the regular Calais—Bâle express of
pre-War days. If they would only run that train now, how it would
have to zig-zag over trenches, and lines of communication.
He entered the zone of a French Army. On all sides, in the towns
and villages, in the camps and manœuvre areas, he saw blue-
coated men, and stared at them, with the same fascinated interest
as he now felt, in spite of himself, in spite of any habit or tradition or
inclination, in his own khaki variety. These fellows carried more on
their backs, had far less transport. His general impression was of
something grimmer, more like purgatory, than that which English
troops gave him. The physical effort of the individual was greater, his
food, pay and accommodation less. And there was none of that
extraordinary volunteering spirit of the Kitchener Armies, the spirit
which said: “Lumme, boys, here’s a war. Let’s have a go at it!” The
French had most of them been conscripted, had known that such a
thing might, probably would happen to them, had been prepared for
it for years. They had not the advantage of being able to say to
themselves: “Well, I jolly well asked for it. Now I’ve got it!”
A saturnine fate brooded over them. He noticed it in the railway and
other officials he met. They were so much more official. R.T.O.’s and
A.P.M.’s—or the equivalent of them, he supposed—who surveyed
his credentials, and passed him on to the place where he was going,
did so with the cynical ghost of amusement, as who should say:
“Aha! This is you. You’re going there, are you? You might as well go
anywhere else.”
Eventually, in a stony village, beneath a pine-clad ridge, he found the
familiar khaki and brass, the good nature and amateurishness of his
own sort. He stepped out of the train and across a platform and with
a curious pang, almost of home-sickness, found himself in England.
Here was the superior corporal in slacks from the orderly room. Here
were the faultless riding horses, being exercised. There was nothing
like them in all the blue-coated armies through which he had passed.
The Commandant to whom he reported, treated him partly as an
officer reporting, partly as a nephew, asked amused questions about
billets in Flanders, who was doing such-and-such a job with Corps,
what were the prospects of leave, and above all, did Dormer play
bridge? He did. Ah! Then the main necessities of modern warfare
were satisfied.
And as he found his billet and changed his clothes, Dormer reflected
how right it all was. What was the good of being officious and ill-
tempered? What was the good of being energetic even? Here we all
were, mixed up in this inferno. The most sensible, probably the most
efficient thing to do, was to forget it every night for a couple of hours,
and start fresh in the morning. Chirnside was away with his
detachment, but would be back shortly. In the meantime the
Commandant hoped Dormer would join his Mess. The billet was
comfortable and Dormer made no objection. On the contrary, he
settled down for a day or two with perfect equanimity. It was always
a day or two nearer the inevitable end of the War, which must come
sometime, a day or two without risk, and actually without discomfort.
What more could one ask?
The Commandant, Major Bone, was a fine-looking man, past middle
age, with beautiful grey hair and blue eyes with a twinkle. His height
and carriage, a certain hard-wearing and inexpensive precision
about his uniform, suggested an ex-guards Sergeant-major. It was
obvious that he had spent all his life in the army, took little notice of
anything that went on outside it, and felt no qualms as to a future
which would be provided for by it. He was one of those men with
whom it was impossible to quarrel, and Dormer pleased him in the
matter of blankets. The Major offered some of those necessities to
Dormer, who was obliged to reply that he had six and feared his
valise would hold no more. He had won the old man’s heart.
The Major had fixed his billet in a little house belonging to the
representative of some firm auxiliary to the wine trade. The little
office had become his office. Orders, nominal rolls, lists of billets and
maps hung over the advertisements of champagne, and
photographs of Ay and Epernay. On the other side of the hall, the
little dining-room suited the Major admirably, as his Mess. It had just
that substantial stuffiness that he considered good taste. The chairs
and table were heavy, the former upholstered in hot crimson, as was
the settee. Upon the mantelpiece, and upon pedestals disposed
wherever there was room and sometimes where there was not, were
bronze female figures named upon their bases “Peace,” “Chastity,”
“The Spirit of the Air.” Dormer did not admire them. They were nude.
As if this were not enough they had their arms either before them or
behind them, never at their sides, which seemed to him to aggravate
the matter. Together with a capacious sideboard, full of glass and
china, couronnes de noces and plated ware, all securely locked in,
these decorations made it almost impossible to move, once the
company was seated at table.
Indeed, during the winter, the Major complained he had been in the
position of having one place frozen at the door, and one roasted next
the Salamander anthracite stove. But with the milder weather, things
were better, for the two big casement windows could be opened, and
filled the room with sweet country air in a moment; they gave on to
the street which was merely a village street, and across the road,
over the wall was a vineyard. The Mess consisted of the Major,
Doctor, Ordnance Officer, and Chirnside, whose place Dormer
temporarily took. There they were a happy little family, removed far
from the vexations attending larger and smaller formations, isolated,
with their own privileges, leave list, and railway vouchers, as pretty a
corner as could be found in all that slow-moving mass of discomfort
and ill-ease that was the War.
On the third day, Dormer’s conscience made him inquire how long
Chirnside would be. “Not long,” was the reply. “You can hear what’s
going on?” He could indeed. For two days the earth and air had been
atremble with the bombardment. French people in the village, and
the French soldiers about the place had a sort of cocksure way of
saying “Ça chauffe?” Indeed, the offensive had been widely
advertised and great things were expected of it.
Then finally Chirnside did return. Dormer had been doing small jobs
for the Major all day, because idleness irked him, and on coming
back to change, found a grizzled oldish man, thin and quiet, a slightly
different edition of the Major, the same seniority, the same ranker
traditions, but memories of India and Egypt instead of Kensington
and Windsor. Dormer listened quietly while the two old soldiers
discussed the offensive. There was no doubt that it was an
enormous and costly failure. That hardly impressed him. He was
used to and expected it. But he had never before seen an offensive
from outside. He had always been in them, and too tired and short of
sleep, by the time they failed, to consider the matter deeply. But this
time he listened to the conversation of the two old men with wonder
mixed with a curious repulsion. They were hard working, hospitable,
but they had the trained indifference of the regular soldier that
seemed to him to be so ominous. In the regular army, where every
one shared it, where it was part of a philosophy of life derived from
the actual conditions, and deliberately adopted like a uniform, all
very well. But no one knew better than Dormer that none of the
armies of 1917 contained any appreciable percentage of regulars,
but were, on the other hand, composed of people who had all sorts
of feelings to be considered and who had not the slightest intention
of spending their lives in the army. Not for the first time did he
wonder how long they would stand it.
The Doctor and Ordnance Officer being busy sorting casualties and
replacing stores, there was no bridge that evening and he was able
to approach Chirnside as to the object of his journey. The old man
heard him with a sort of quizzical interest, but was evidently inclined
to be helpful, twisted his grey moustache points and let his ivory-
yellow eyelids droop over his rather prominent eyes.
“Spanish Farm. April 1916. Oh, aye!”
“Could you recall an incident that occurred there. Damage to a little
chapel in the corner of the pasture where the roads met. A driver
wanted to shelter his mules and broke into the place?”
Chirnside thought hard, looking straight at Dormer. It was obvious to
Dormer that the old man was thinking, with army instinct, “Here,
what’s this I’m getting involved in? No you don’t,” and hastened to
reassure him.
“It’s like this. The case has become unfortunately notorious. The
French have taken it up very strongly. You know what these things
are, once they become official test cases. We’ve got to make an
arrest and probably pay compensation as well, but at present our
people at Base are sticking out for treating it as a matter of
discipline. The unit was the 469 T.M.B., but there have been so
many casualties that no one can tell me the name of the driver who
did it.”
Dormer was thinking: “There, that’s the umpteenth time I’ve told the
yarn, and what good is it?” When suddenly he had a stroke of
genius:
“Of course, they’ve got hold of your name.”
It succeeded remarkably well. A sort of habitual stiffening was
obvious in the Army-worn old face in front of him. Chirnside shifted
his legs.
“I can’t tell y’much about it. I don’t know the chap’s name or number,
and I expect all the rolls are destroyed. Anyway he might not be on
them, for he wasna’ a driver!”
Chirnside was relapsing into his native Scotch, but Dormer didn’t
notice. He had got a clue.
“What was he then?”
“He had been servant to young Fairfield, who was killed.”
“You don’t remember Fairfield’s regiment. That might help us?”
“No, I don’t, and it wouldn’t help you, for he came out to Trench
Mortars, and not with his own crowd. This servant of his he picked
up at Base, or from some employment company.”
“What on earth was he doing with those mules?”
“What could you do with ’em? The driver was killed and the limber
smashed to matchwood. The feller had nothing to do, so he did that!”
“You don’t remember what happened to him after that?”
“Um—I think he went as young Andrews’ servant.”
“Ah! What did he come from?”
“Andrews? Gunner, he was!”
“Thanks. That may help. You saw the row when the Mayor of the
village came to certify the damage?”
“Aye, there was some blethers about the business. You couldna’
wonder. The old feller was got up like a Tattie Bogle. The men had
had no rest, and were going straight back to the line. They marched
all right, but you couldn’t keep them from calling names at such a
Guy—young troops like that!”
“You couldn’t describe Andrews’ servant to me?”
“No. He looked ordinary!”
A mistake, of course, no use to ask old Chirnside things like that. A
third of a century in the army had long ago drilled out of him any sort
of imagination he might ever have had. He was just doing a
handsome thing by a brother officer in remembering at all. His
instinct was obviously to know nothing about it. But, piqued by the
novelty of Commissioned rank, he went on: “Yes, I can tell you
something. That feller had a grievance. I remember something
turning up in one of his letters, when we censored ’em. Lucky spot
when you think how most of the censoring was done.”
“I should think so. What was it?”
“Couldn’t say now. Grievance of some sort. Didn’t like the army, or
the War, or something.”
Dormer sat down and wrote out the information obtained and made
his preparations to rejoin the Division. The Major said: “Oh, no hurry,
stop another day, now you’re here!” And all that evening, as he
thought and wrote, and tried to believe this fatal business a step
nearer completion, he heard the two old soldiers, like two good-
natured old women, gossiping. Each expected the other to know
every camp or barrack in which he had lain, each named this or that
chance acquaintance, made any time those thirty years, anywhere in
the world, as though the other must know him also. Often this was
the case, in which they both exclaimed together, “Ah, nice feller,
wasn’t he?” Or, if it were not the case, the other would rejoin, “No,
but I knew So-and-so, of the sappers,” and probably the second shot
would hit the mark. It could hardly fail to do so in the old close
borough of the Regular Army. And then they would exclaim in unison
again.
Dormer was as impressed as he ever was by any member of the
Professional Army. They knew how to do it. He would never know.
The army was their God and King, their family and business. In a
neat circle they went, grinding out the necessary days to their
pensions. The present state of Europe, while verbally regretted or
wondered at, did not scratch the surface of their minds. How could
it? It had been a golden opportunity for them. It made the difference
to them and to any human wife or family they might have accreted,
between retiring on Commissioned pay-scale, or taking a pub or
caretaker’s place, as the ex-Sergeant-major they would otherwise
have been. But there was charm in their utter simplicity. Nothing
brutal, very little that was vain, and some nicely acquired manners.
The offensive of the French Army, in the machinery of which they
had their places, moved them not at all. Chirnside casually
mentioned that he gathered it had been a big failure. Dormer
expected to hear him recite some devastating tale of misdirected
barrage, horrible casualties or choked communications. Nothing so
graphic reached him. The old man had simply attended to his job,
and when he found that the troops were returning to the same billets,
drew his own conclusions. That was all. Dormer was horrified, but no
one could be horrified long with Chirnside. Of course, he didn’t mind
how long the War went on.
Having completed his preparations, Dormer went up to his little room
and was soon asleep. He was in fine condition and thoroughly
comfortable, and was astonished after what appeared to be a very
short interval, to find himself wide awake. There was no mistaking
the reason. It was the row in the street. He pulled on his British
Warm and went to look. It was quite dark, but he could make out a
confused crowd surging from side to side of the little street, could
see bayonets gleaming, and could hear a clamour of which he could
not make out a word. It was like nothing he had ever heard in the
War, it recalled only election time in his native city, the same aimless
shuffling feet, the same confusion of tongues, the same
effervescence, except that he had instinct enough to know from the
tones of the voices that they were raised in lamentation, not triumph.
He was extremely puzzled what to do, but clear that no initiative lay
with him. For ten minutes he waited, but the situation did not change.
He opened his door very quietly. Not a sound from the Major. From
Chirnside, opposite, heavy regular breathing. Above, in the attics, the
low cockney brevity of soldier servants discussing something with
the detachment of their kind. Reassured, he closed the door, and got
back into his blankets. The noise was irritating but monotonous. He
fell asleep. He next awoke to the knocking of his servant bringing his
morning tea, and clean boots.
“What was all that row in the night?”
“Niggers, sir.”
“What do you mean?”
“French coloured troops, sir. They got it in the neck seemingly. They
don’t half jabber.”
Major Bone was more fully informed. There was no doubt that the
French had had a nasty knock. Black troops were coming back just
anyhow, out of hand, not actually dangerous, the old soldier allowed
it to be inferred, but a nuisance. What struck him most forcibly was
the dislocation of the supply services. Defeat he accepted, but not
unpunctuality.
“These Africans are besieging the station, trying to board the trains,
and get taken back to Africa. I can’t get hold of an officer, but
Madame says they’re all killed. She’s in an awful state. I don’t
suppose you’ll get away to-day!”
He was right enough. Dormer’s servant shortly returned, humping
the valise. The station was closed, the rolling stock had been
removed. The black troops were swarming everywhere, collapsing
for want of food and sleep, disorganized and incoherent. Dormer
went out shortly after and verified the state of affairs. He was not
molested, so far had the breakdown gone, but was the object of what
appeared to him most uncomplimentary allusions, but all in pidgin-
French, too colonial for his fair, but limited, knowledge of the
language. There was clearly nothing to be done, so far as transport
went, that day, and he resigned himself to spending his time in the
little Mess.
The Doctor and Ordnance Officer appeared at dinner with reassuring
news. The failure of the offensive had been bad, but the French had
never really lost control and were getting their people in hand
immediately. There was a rumour that a General who tried to restore
order had been thrown into the river, but it might be only a tale. Major
Bone was contemptuous of the whole thing. Do—what could they do,
a lot of silly blacks? The French would cut off their rations and
reduce them to order in no time. Thus the old soldier. But he did not
prevent Dormer going to bed with a heavy heart. To him it was not so
much a French offensive that had failed. It was another Allied effort,
gone for nothing. His life training in apprehension made him paint the
future in the gloomiest colours. Where would fresh men be obtained
from? Whence would come the spirit—what they called morale in
military circles—to make another attempt? If neither men nor morale
were forthcoming, would the War drag out to a stalemate Peace? He
had no extravagant theories for or against such an ending to it. To
him it meant simply a bad bargain, with another war to make a better
one looming close behind it. And his recent military training had also
received an unaccustomed shock. A new army enlistment, he had
seen nothing of the retreat from Mons, and was far from being able
to picture March or April 1918, still twelve months in the future.
For the first time in his life he had seen panic, confusion, rout. True,
it was already stopped, but that did not expunge from his mind the
sight, the noise, the smell even, of that crowd of black soldiers who
had suddenly ceased to be soldiers, numbers standing in line, and
had so dramatically re-become men. The staring eyeballs, the
physical collapse, the officer-less medley of uncertain movement
were all new to him, and all most distressing. Of course, the fellows
were mere blacks, not the best material, and had probably been
mishandled. But under a more prolonged strain, might not the same
thing happen to others? The Germans were the least susceptible he
judged, the Russians most. What would he not see, some day, if the
War dragged on?
Whatever narrow unimaginative future his unadventurous mind
conjured up, his far stronger faculty for getting on with the matter in
hand soon obliterated. He was no visionary. Contemplation was not
in him. Directly the trains were running he left that cosy little Mess of
Major Bone’s to rejoin. He left off thinking about the War, and took up
his job where he had, for a moment, allowed it to lie, disregarded
under the stress of new events and strange emotions.
As the train moved on and on through French lines of
communication he was wondering again about the fellow who had
done the trick at Vanderlynden’s, of how he was to be found, of how
the whole thing would frame itself. These French chaps, whose
transport he saw each side of him, Army Corps after Army Corps.
Biggish men, several of them, in a round-shouldered fashion, due
partly to their countrified occupation, partly to their uniform, with its
overcoat and cross-straps. Browner skinned, darker of hair and eye
than our men, they confirmed his long-established ideas about them,
essentially a Southern people, whose minds and bodies were formed
by Biscayan and Mediterranean influences. They would not be
sentimental about mules, he would wager. On the other hand, they
would not laugh at a Mayor. They did not laugh much as a rule, they
frowned, stared, or talked rapidly with gestures, and then if they did
laugh, it was uproariously, brutally, at some one’s misfortunes. Satire
they understood. But they missed entirely the gentle nag, nag, nag of
ridicule, that he used to hear from his own platoon or company,
covering every unfamiliar object in that foreign land, because it was
not up to the standard of the upper-middle-classes. To the French,
life was a hard affair, diversified by the points at which one was less
unfortunate than one’s neighbour.
To the English, life was the niceness of a small class, diversified by
the nastiness of everything else, and the nastiness was endlessly
diverting. For the French were mere men, in their own estimation.
Not so the poorer English of the towns. They were gentlemen. If they
lapsed (and naturally they lapsed most of the time) they were comic
to each other, to themselves even. How well he remembered, on the
march, when the battalion had just landed, passing through a village
where certain humble articles of domestic use were standing outside
the cottage doors, waiting to be emptied. A suppressed titter had run
all along the column.
A Frenchman would never have thought them funny, unless they fell
out of a first-floor window on to some head and hurt it. Again, to a
Frenchman, Mayor and Priest, Garde Champêtre and Suisse were
officials, men plus authority and therefore respectworthy. To
Englishmen, they were officials, therefore not gentlemen, therefore
ridiculous. If a big landowner, or member of Parliament, or railway
director had walked into Vanderlynden’s pasture, just as 469 T.M.B.
fell in for their weary march back to the line, would they have
laughed? Not they. But then those members of England’s upper
classes would not have worn tricolour sashes to enforce authority.
So there you were. With this philosophic reflection he fell asleep.
Dormer returned to an army which was at its brightest. It had held
the initiative in the matter of offensives for over a year and a half,
and if no decision had been come to, a wide stretch of ground had
been won, and hope on the whole was high. From time to time there
were rumours of a queer state of things in Russia, but it was far off
and uncertain. The matter of the moment was Messines, the famous
ridge which had been lost at the very beginning of the War and
which was now to be regained. In this affair Dormer found himself
busily engaged. Here were no waste downs of the Somme, but some
of the most fertile land in the world.
Among other matters confronting the Generals was the problem of
how to keep civilians from rushing back to cultivate land of which
they had been deprived for three years. The day came, the explosion
of the great mines, so Dormer was told, was heard in London. If he
did not hear it, it was because a well-directed long-range artillery
bombardment, complicated by a bombing that was German and
German only in its thoroughness, deafened and bewildered him, took
his sleep, killed his servant, and stampeded the horses of all the
divisional ammunition columns near him, so that his tent was
trampled down, his belongings reduced to a state hardly
distinguishable from the surrounding soil. However, the blow, such
as it was, was successful. Irish and Scotch, Colonial and London
divisions took that battered hillock that had defied them so long, and
Dormer in spite of all his experience could not help thinking: “Oh,
come, now we are really getting on.”
But nothing happened. Dormer heard various reasons given for this,
and twice as many surmises made about it, but well aware how
much importance to attach to the talk that floated round Divisional
Offices and Messes, relied upon his own experience and arithmetic.
According to him, nothing could happen, because each offensive
needed months of preparation. Months of preparation made possible
a few weeks of activity. A few weeks of activity gained a few square
miles of ground. Then more months of preparation, grotesquely
costly, and obvious to every one for a hundred miles, so that the
enemy had just as long to prepare, made possible a few more
weeks’ activity and the gain of a few miles more.
This was inevitable in highly organized mechanical war, fought by
fairly matched armies, on a restricted field, between the sea and the
neutral countries. He admitted it. But then came his lifelong habit of
reducing the matter to figures. He roughed out the area between the
“front” of that date and the Rhine, supposing for the sake of
argument that we went no farther, and divided this by the area
gained, on an average, at the Somme, Vimy and Messines. The
result he multiplied by the time taken to prepare and fight those
offensives, averaged again. The result he got was that, allowing for
no setbacks, and providing the pace could be maintained, we should
arrive at the Rhine in one hundred and eighty years.
For the only time in his life Dormer wished he were something other
than Dormer. For a few moments after arriving at his conclusion, he
desired to be a person of power and influence, some one who could
say with weight that the thing ought to stop here and now. But this
very unusual impulse did not last long with him.
All that remained of Belgium and wide tracts of French Flanders
adjoining it, became one huge ant-heap. Never had there been such
a concentration, Corps next to Corps, Services mosaiced between
Services, twenty thousand men upon roads, no one could count how
many handling munitions, as, from Ypres to the sea, the great
offensive of 1917 slowly germinated.
Dormer was soon caught up and landed in the old familiar blackly-
manured soil of the Salient. He was not disgusted or surprised. He
was becoming increasingly conscious of a sensation of going round
and round. Now, too, that troops were always pouring along a road
before him, he had again the feeling that his head was an empty
chamber, round which was painted a frieze, men, men, mules, men,
limbers, guns, men, lorries, ambulances, men, men, men. It might be
just worry and overwork, it might be that he was again forced to
share his limited accommodation with Kavanagh. They were in a
dug-out on the canal bank, just by one of those fatal causeways built
to make the passage of the canal a certainty, instead of the gamble it
had been in the days of the pontoon bridges. The passage became,
like everything else in the War, a certainty for the Germans as much
as for the Allies. The place was registered with the utmost precision
and hit at all times of the day and night. It probably cost far more
than the taking of any trench.
Amid the earth-shaking explosions that seldom ceased for long, in
the twilight of that narrow cavern in the mud, Kavanagh was as
unquenchable as he ever had been on the high and airy downs of
the Somme. During the daylight, when nothing could be done
outside, he bent over his map of cables while Dormer perfected his
plan for getting first-line transport past that infernal canal. He
purposed to send an N.C.O. a good two miles back, with small
square pieces of card, on which were written 9.0 p.m., 9.5 p.m., and
so on, the times being those at which the unit so instructed was to
arrive beside his dug-out. He thought rather well of this idea, no
jamming and confusion, and if the enemy made a lucky hit, there
would be fewer casualties and less to clear away. In the middle of his
calculations he heard
 “Why, soldiers, why
Should we be melancholy,
Whose duty ’tis to die!”
He could not resist saying:
“If you must make that d——d noise, I wish you’d put some sense
into it.”
“Sense. I was trying to cheer you up!”
“‘Duty ’tis to die’ is jolly cheering, and quite untrue.”
“Oh, is it? What is our duty then?”
“Our duty is to live if we possibly can. And I mean to do it. It’s the
people who keep alive who will win the War.”
“According to that, all one has got to do is to get to Blighty, or
preferably the United States, and stay there?”
“Not a bit. You exaggerate so. All I said was, that it is foolish to make
it a duty to become a casualty.”
“Dormer, I shall never get you to see things in the proper light. You’re
like a lamb trying to leap with joy, and never able to get its hind legs
off the ground.”
“This is all rot. What connection is there between lambs and leaping,
and our jobs? Mine is to see that various people and things are in
the position where they will be wanted, at the moment at which they
will have most effect in winning the War. Yours is to see that they can
speak and be spoken to when required.”
“Lovely, lovely! What a teacher you would have made.”
“I had a better job.”
“There is no better job, except perhaps the one we are doing. I do
admire your descriptions of them. All you want is to put in a personal
allegorical note. You might condense the whole thing by saying that
you will be Minerva if I will be Mercury. Yep?”
“Whatever are you talking about?”
“Yours to see that all is in order. That is a matter of reason. You are
the Goddess. I am merely a lesser God. Mercury was God of
Communications. I wonder whether they’d let me design a cap
badge for signallers. Mercury playing on a buzzer. You may have
your Owl!”
“Oh, shut up.”
“I fear I must, the bugle calls, and I must follow, or my watch shows it
is time I was looking after my chaps. But you’ve had a brilliant idea,
Dormer.”
“I?”
“You’ve had the idea of fighting the War allegorically. Wisdom and
Light we are. That would do away with half the horror. So long!”
Then queerly, instead of feeling relieved from an annoyance, Dormer
felt more despondent than ever. What could it be? Was the fellow
right? Surely not! All that nonsense! And yet—and yet what would
not he, Dormer, conscious of his own probity, have given to be
conscious instead, of Kavanagh’s lightness of heart? That very
probity drove him out in the all-too-late summer dusk to see that
everything was going right. Yes, here they were; details of transport,
parties to dig, parties to carry, details of services, engineers of all
their various grades. Punctual, incredibly docile, honest English in
their gestureless manner of getting on with the job. They took care of
their mules, look at these beasts pulling as though they were English
too (instead of the Argentine crossbreds he knew them to be), not
because it was a duty, although it was, and not because the mule
was a miracle, like a tank or an aeroplane, but just because it was a
mule, that meant, to English soldiers, and to English soldiers only, a
fellow-creature, a human being. On they went, reporting to him, and
pushing on, sometimes with a hurried question as to map square, or
other crucial uncertain detail, sometimes with only a grunt. That
endless procession had not been in progress many minutes before,
amid the considerable and gently growing shell-fire, there came a
bang that seemed to go right through his head. He knew from old
trench experience what it was. Nothing but a gun pointing straight at
you could make that particular hrrmph.
He set his feet, not a moment too soon. It was a five-nine, the sort
the French called “Grande Vitesse.” A whirlwind, a small special
whirlwind pointed like an arrow, hit the causeway so that it shook and
then went up with a wheel of splintered bits. He was glad he had
devised his patent card system. The units were not too close
together. He had time to shout to the next, “Come on, you’ve two
minutes to get over!” and over they went, as if the Devil were after
them, instead of a lump of Krupp steel fitted with lethal chemicals.
They were hardly over before the second came, whump! To say that
Dormer was frightened, was to fail to describe the matter. He was
stiffened all over, his hair stood up, his heart thumped so that it hurt
him, his feet were stone cold, but he knew his job and did it.
The next lot to come was a whole field company to do some special
duty, and although he hurried them, the tail of the brown column was
still high and exposed when the shell came. They ducked and darted
into any cover that was available, and he heard his voice, as the
voice of some one far away speaking to a public meeting, like a
voice on the wireless, saying:
“Come on. Get out of that and come on. If I can stand here, surely
you can get out of it.”
They did so. Behind them came a special party to dig in the
Meteorological Officer. What a menagerie it was! Every trade, every
nation too, Chinese, Zulu, West Indian, Egyptian. He did not blame
the Germans who had chalked in blue on the bare back of a
Portuguese, whom they captured and stripped, “The Monkey House
is full,” before they drove him back into English lines.
Even truer did Dormer find it when he had to go back for any reason,
to Corps H.Q. or beyond. French and Belgians he knew, he had
found them in the trenches beside him years before. Portuguese he
had become accustomed to, Americans he looked forward to with
anticipation. But farther back, he found Chinese, Africans of all
descriptions, Indians, East and West, while the French, in addition to
their black troops, had Spanish and Italian labour.
It did not please his parochial mind. He felt increasingly that there
was something wrong when you had to drag in all these coloured
people from every remote quarter of the globe, without even the
excuse the French had, that they were “Colonials.” But no one could
tell, least of all Dormer himself, whether his feelings were the result
of a strong belief in the Colour Bar, or whether it were merely the
futility of it all. For in spite of the omnium gatherum of race, tongue
and religion, the offensive failed. As a matter of routine, the weather
broke on Z day. Within forty-eight hours it was obvious that the affair
had stuck. Apart from a feeling of the hand of Fate in it, a sinister
feeling of great incomprehensible forces working out his destiny for
him, without his having the least power to influence the matter for
better, for worse, which was so desolating to his pre-War habit of
mind, where a certain line of unostentatious virtue had always
carried a reward that could be reckoned on with the greatest
exactitude, there were other disturbing elements in the situation.
Of course the Bosche was ready. He was bound to be ready, couldn’t
avoid it. He had immensely thickened his depth of defence, which
was now composed not of the old obvious trenches full of men, all of
which could be blown to pieces, but of small isolated turrets of ferro-
concrete, where two or three machine gunners (and who made
better machine gunners than the careful Germans) could hold an
army at bay, until dislodged by a direct hit by a shell of six-inch
calibre or over, or laboriously smoke-screened and bombed out, at
the rate of perhaps a mile a day, on good days. He saw his
computation of one hundred and eighty years altogether insufficient
for getting to the Rhine. Moreover, for such work this medley of
nations was of no good at all. It reminded him of a book by Anatole
France he had been compelled by a friend to read, wherein a great
conqueror enlisted in his army all the men of his nation, then all the
men of the neighbouring nations, then all the savages at the end of
the earth, and finally the baboons and other combatant animals. That
was all very well. That was just story telling. But it horrified Dormer
all the more to see such story telling coming true before his eyes. As
coloured-labour company after coloured-labour company filed past
his tent, guttural and straggling, he was able to pull himself together,
and see that it was not true after all.
These people, little better than beasts, uglier in some cases and far
more troublesome, were no good. They couldn’t fight. You couldn’t
trust them to stand the shelling or to obey an order. Then just as he
was feeling rather relieved, he saw the logical result of his
conclusion. All the fighting would have to be done by those very men
who had volunteered or been conscripted and who had been so
generously wasted ever since. They were sticking it, and sticking it
well, but this new offensive that had just opened promised to try
them pretty high. Would they stick that? Would the day ever come
when he would see them a mere mob, like those French black troops
he had seen in May? Perhaps peace would be made. Such is the
eternal hopefulness of men, that he even hoped, against all previous
experience. That quenchless gleam common to all human souls, one
of the basic things that makes war so long, and peace, where it is so
much less necessary, just that much less attractive, added to work
for fifteen hours a day, kept Dormer sane and healthy for weeks, in
spite of worsening conditions, and the steady increase in enemy
shelling. It was with a return of that uncanny feeling of being haunted
that he found himself called up to Divisional Head-quarters. He knew
quite well what it was, but he had relied on the difficulty of finding
Andrews, on the tremendous strain of this most costly and urgent of
all offensives, to keep the matter out of his path, or rather to keep
him out of its path, for he had long dropped into the habit of feeling
himself as in a nightmare, pursued by something he could not see or
even imagine, but which was certainly sinister and personally fatal to
him.
When he got to the office his feeling of nightmarishness was rather
aggravated than allayed. Colonel Birchin was talking to the A.D.M.S.
The fact was that the A.D.M.S. was a new one, patently a Doctor
who had been fetched out from Doctoring, had been found capable
of organization and had been shoved into the job vice some one else
gone higher up. Beside him Colonel Birchin shone, as it were, with
the glamour of another world. Dormer had seen him in camp and
hut, and château and Mairie for a year and a half, just like that,
handsome and sleek, filling his plain but choice khaki with a
distinction that no foreign officer could gain from all the blues and
reds and yellows and greens and blacks, varnished belts and metal
ornaments of other armies. And in that moment of sharpened nerves
and unusual power of vision Dormer seemed to see why. Colonel
Birchin was not an officer of a national army in the sense that any
French, German, Italian or Russian Colonel was. There was nothing
of the brute and nothing of the strategian about those nice manners,
that so easily and completely excluded everything that was—what?
Unmilitary? Hardly. There was nothing consciously, offensively
military about the Colonel, “regular” or professional soldier that he
was. He would never have swaggered in Alsace, massacred in
Tripoli, Dreyfused in France. He would never have found it
necessary. For Colonel Birchin was not a state official. He was an
officer of the Watch, the small band of paid soldiers that Stuart and
subsequent kings kept to defend themselves from mobs, national
armies and other inconvenients. Colonel Birchin might write himself
as of “The Herefordshire Regiment,” but it made no difference. His
chief, inherited, and most pronounced quality was that he was a
courtier. He represented the King. Preferably, at home, of course,
where one could live in all that thick middle-class comfort that had
ousted the old land-owning seignorial dignity and semi-starvation.
But upon occasion, Colonel Birchin could betake himself to Africa,
India, and now even to this France, sure that even in this most
tedious and unpleasant of wars, he would be properly fed and
housed.
So here he was, representing the King even more exactly than
before he was seconded from the King’s Own Herefordshire
Regiment. He spoke and looked, in fact, rather as if he were the
King. Ignorant and unused to the immense transport, the
complicated lists of highly scientific equipment, he judged rightly
enough that his one safe line was to represent authority, and see
that these semi-civilians who did understand such things got on with
the War. So he listened in a gentlemanly way to the A.D.M.S. (who
wore beard and pince-nez) explaining at great length a difficult
alternative as to the siting of Forward Dressing Stations, and
contributed:
“You do what is best, Doctor, and we shall back you up!”