This page intentionally left blank

Constraint logic programming lies at the intersection of logic programming, optimisation

and articial intelligence. It has proved a successful tool for application in a variety of
areas including production planning, transportation scheduling, numerical analysis and
bioinformatics. Its migration from academic discipline to software development has been
due in part to the availability of software systems that realise the underlying methodology;
ECLi PSe is one of the leading such systems. It is exploited commercially by Cisco, and is
freely available and used for teaching and research in over 500 universities.
This book has a two-fold purpose. Its an introduction to constraint programming, appro-
priate for a one-semester course for upper undergraduate or graduate students of computer
science or for programmers wishing to master the practical aspects of constraint program-
ming. By the end of the book, the reader will be able to understand and write constraint
programs that solve complex problems.
Second, it provides a systematic introduction to the ECLi PSe system through carefully
chosen examples that guide the reader through the language, illustrate its power and versa-
tility, and clarify the gain achieved by this approach, which, ultimately allows the reader to
better understand the proper use of constraints in solving real-world problems.

Krzysztof R. Apt received his PhD in 1974 in mathematical logic from the University
of Warsaw in Poland. He is a senior researcher at Centrum voor Wiskunde en Informatica,
Amsterdam and Professor of Computer Science at the University of Amsterdam. He is the
author of three other books: Verication of Sequential and Concurrent Programs (with E.-R.
Olderog), From Logic Programming to Prolog, and Principles of Constraint Programming,
and has published 50 journal articles and 15 book chapters.
He is the founder and the rst editor-in-chief of the ACM Transactions on Computational
Logic, and past president of the Association for Logic Programming. He is a member of
the Academia Europaea (Mathematics and Informatics Section).
After completing a degree at Oxford in Mathematics and Philosophy, Mark Wallace
joined the UK computer company ICL, who funded his PhD at Southampton University,
which was published as a book: Communicating with Databases in Natural Language.
He has been involved in the ECLi PSe constraint programming language since its incep-
tion and has led several industrial research collaborations exploiting the power of constraint
programming with ECLi PSe . Currently he holds a chair in the Faculty of Information
Technology at the Monash University, Victoria, Australia and is involved in a major new
constraint programming initiative funded by National ICT Australia (NICTA), and in the
foundation of a Centre for Optimisation in Melbourne. He has published widely, chaired the
annual constraint programming conference, and is an editor for three international journals.
 Advance praise for Constraint Logic Programming using ECLi PSe
The strength of Constraint Logic Programming using ECLi PSe is that it simply and
gradually explains the relevant concepts, starting from scratch, up to the realisation of
complex programs. Numerous examples and ECLi PSe programs fully demonstrate the
elegance, simplicity, and usefulness of constraint logic programming and ECLi PSe .
The book is self-contained and may serve as a guide to writing constraint applications in
ECLi PSe , but also in other constraint programming systems. Hence, this is an indispensable
resource for graduate students, practioners, and researchers interested in problem solving
and modelling.
Eric Monfroy, Universite de Nantes

ECLi PSe is a exible, powerful and highly declarative constraint logic programming
platform that has evolved over the years to comprehensively support constraint programmers
in their quest for the best search and optimisation algorithms. However, the absence of a
book dedicated to ECLi PSe has presented those interested in this approach to programming
with a signicant learning hurdle. This book will greatly simplify the ECLi PSe learning
process and will consequently help ECLi PSe reach a much wider community.
Within the covers of this book readers will nd all the information they need to start
writing sophisticated programs in ECLi PSe . The authors rst introduce ECLi PSe s history,
and then walk the reader through the essentials of Prolog and Constraint Programming,
before going on to present the principal features of the language and its core libraries in a
clear and systematic manner.
Anyone learning to use ECLi PSe or seeking a course book to support teaching constraint
logic programming using the language will undoubtedly benet from this book.
Hani El-Sakkout, Cisco Systems, Boston, Massachusetts

It has been recognized for some years now within the Operations Research community
that Integer Programming is needed for its powerful algorithms, but that logic is a more
exible modelling tool. The case was made in most detail by John Hooker, in his book Logic-
Based Methods for Optimization: Combining Optimization and Constraint Satisfaction.
The ECLi PSe system is a highly successful embodiment of these ideas. It draws on
ideas coming from logic programming, constraint programming, and the Prolog language. I
strongly recommend this book as a systematic account of these topics. Moreover, it gives a
wealth of examples showing how to deploy the power thus made available via the ECLi PSe
system.
Maarten van Emden, University of Victoria, Canada

This is an impressive introduction to Constraint Logic Programming and the ECLi PSe
system by two pioneers in the theory and practice of CLP. This book represents a state-of-
the-art and comprehensive coverage of the methodology of CLP. It is essential reading for
new students, and an essential reference for practioners.
Joxan Jaffar, National University of Singapore
CONSTRAINT LOGIC PROGRAMMING
USING ECLi PSe

KRZYSZTOF R. APT AND MARK WALLACE

CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, So Paulo

Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521866286

Cambridge University Press 2007

This publication is in copyright. Subject to statutory exception and to the provision of

relevant collective licensing agreements, no reproduction of any part may take place
without the written permission of Cambridge University Press.
First published in print format 2006

ISBN-13 978-0-511-34874-7 eBook (EBL)

ISBN-10 0-511-34874-6 eBook (EBL)

ISBN-13 978-0-521-86628-6 hardback

ISBN-10 0-521-86628-6 hardback

Cambridge University Press has no responsibility for the persistence or accuracy of urls
for external or third-party internet websites referred to in this publication, and does not
guarantee that any content on such websites is, or will remain, accurate or appropriate.
 Contents

Introduction page ix

Part I: Logic programming paradigm 1

1 Logic programming and pure Prolog 3
1.1 Introduction 3
1.2 Syntax 4
1.3 The meaning of a program 7
1.4 Computing with equations 9
1.5 Prolog: the first steps 15
1.6 Two simple pure Prolog programs 23
1.7 Summary 26
1.8 Exercises 26
2 A reconstruction of pure Prolog 29
2.1 Introduction 29
2.2 The programming language L0 29
2.3 Translating pure Prolog into L0 33
2.4 Pure Prolog and declarative programming 35
2.5 L0 and declarative programming 36
2.6 Summary 37
2.7 Exercises 38

Part II: Elements of Prolog 39

3 Arithmetic in Prolog 41
3.1 Introduction 41
3.2 Arithmetic expressions and their evaluation 42
3.3 Arithmetic comparison predicates 47
3.4 Passive versus active constraints 51

v
vi Contents

3.5 Operators 52
3.6 Summary 55
3.7 Exercises 57
4 Control and meta-programming 59
4.1 More on Prolog syntax 59
4.2 Control 62
4.3 Meta-programming 69
4.4 Summary 74
4.5 Exercises 75
5 Manipulating structures 76
5.1 Structure inspection facilities 76
5.2 Structure comparison facilities 78
5.3 Structure decomposition and construction facilities 80
5.4 Summary 85
5.5 Exercises 85

Part III: Programming with passive constraints 87

6 Constraint programming: a primer 89
6.1 Introduction 89
6.2 Basic concepts 90
6.3 Example classes of constraints 91
6.4 Constraint satisfaction problems: examples 94
6.5 Constrained optimisation problems: examples 98
6.6 Solving CSPs and COPs 102
6.7 From CSPs and COPs to constraint programming 107
6.8 Summary 108
6.9 Exercises 108
7 Intermezzo: Iteration in ECLi PSe 110
7.1 Introduction 110
7.2 Iterating over lists and integer ranges 111
7.3 Iteration specifications in general 113
7.4 Arrays and iterations 121
7.5 fromto/4: the most general iterator 124
7.6 The n-queens problem 129
7.7 Summary 130
7.8 Exercises 132
8 Top-down search with passive constraints 133
8.1 Introduction 133
8.2 Solving CSPs using Prolog 134
 Contents vii

8.3 Backtracking search in Prolog 135

8.4 Incomplete search 139
8.5 Non-logical variables 146
8.6 Counting the number of backtracks 149
8.7 Summary 153
8.8 Exercises 153
9 The suspend library 155
9.1 Introduction 155
9.2 Solving CSPs and COPs using ECLi PSe 156
9.3 Introducing the suspend library 157
9.4 Core constraints 160
9.5 User defined suspensions 167
9.6 Generating CSPs 170
9.7 Using the suspend library 175
9.8 Summary 180
9.9 Exercises 181

Part IV: Programming with active constraints 183

10 Constraint propagation in ECLi PSe 185
10.1 Introduction 185
10.2 The sd library 186
10.3 The ic library 189
10.4 Disjunctive constraints and reification 200
10.5 Summary 203
10.6 Exercises 204
11 Top-down search with active constraints 205
11.1 Introduction 205
11.2 Backtrack-free search 206
11.3 Shallow backtracking search 208
11.4 Backtracking search 211
11.5 Variable ordering heuristics 215
11.6 Value ordering heuristics 218
11.7 Constructing specific search behaviour 220
11.8 The search/6 generic search predicate 223
11.9 Summary 227
11.10 Exercises 229
12 Optimisation with active constraints 230
12.1 The minimize/2 built-in 230
12.2 The knapsack problem 232
viii Contents

12.3 The coins problem 234

12.4 The currency design problem 235
12.5 Generating Sudoku puzzles 240
12.6 The bb min/3 built-in 246
12.7 When the number of variables is unknown 253
12.8 Summary 254
12.9 Exercises 255
13 Constraints on reals 256
13.1 Introduction 256
13.2 Three classes of problems 257
13.3 Constraint propagation 261
13.4 Splitting one domain 263
13.5 Search 264
13.6 The built-in search predicate locate 265
13.7 Shaving 267
13.8 Optimisation on continuous variables 272
13.9 Summary 275
13.10 Exercises 275
14 Linear constraints over continuous and integer variables 278
14.1 Introduction 278
14.2 The eplex library 279
14.3 Solving satisfiability problems using eplex 284
14.4 Repeated solver waking 285
14.5 The transportation problem 289
14.6 The linear facility location problem 294
14.7 The non-linear facility location problem 296
14.8 Summary 302
14.9 Exercises 303
Solutions to selected exercises 305
Bibliographic remarks 320
Bibliography 323
Index 327
 Introduction

The subject of this book is constraint logic programming, and we will

present it using the open source programming system ECLi PSe , available at
http://www.eclipse-clp.org. This approach to programming combines
two programming paradigms: logic programming and constraint program-
ming. So to explain it we first discuss the origins of these two programming
paradigms.1

Logic programming
Logic programming has roots in the influential approach to automated the-
orem proving based on the resolution method due to Alan Robinson. In his
fundamental paper, Robinson [1965], he introduced the resolution principle,
the notion of unification and a unification algorithm. Using his resolution
method one can prove theorems formulated as formulas of first-order logic,
so to get a Yes or No answer to a question. What is missing is the possi-
bility to compute answers to a question.
The appropriate step to overcome this limitation was suggested by Robert
Kowalski. In Kowalski [1974] he proposed a modified version of the resolu-
tion that deals with a a subset of first-order logic but allows one to generate
a substitution that satisfies the original formula. This substitution can then
be interpreted as a result of a computation. This approach became known
as logic programming . A number of other proposals aiming to achieve the
same goal, viz. to compute with the first-order logic, were proposed around
the same time, but logic programming turned out to be the simplest one
and most versatile.
In parallel, Alain Colmerauer with his colleagues worked on a program-
1 In what follows we refer to the final articles discussing the mentioned programming languages.
This explains the small discrepancies in the dateline.

ix
x Introduction

ming language for natural language processing based on automated theorem

proving. This ultimately led in 1973 to creation of Prolog . Kowalski and
Colmerauer with his team often interacted in the period 19711973, which
explains the relation between their contributions, see Colmerauer and Rous-
sel [1996]. Prolog can be seen as a practical realisation of the idea of logic
programming. It started as a programming language for applications in nat-
ural language processing, but soon after, thanks to contributions of several
researchers, it was successfully transformed into a general purpose program-
ming language.
Colmerauer, when experimenting with Prolog, realised some of its im-
portant limitations. For example, one could solve in it equations between
terms (by means of unification) but not equations between strings. Using
the current terminology, Prolog supports only one constraint solver. This
led Colmerauer to design a series of successors, Prolog II, Prolog III, and
Prolog IV. Each of them represents an advance in the logic programming
paradigm towards constraint programming. In particular Prolog III, see
Colmerauer [1990], included a support in the form of solving constraints
over Booleans, reals and lists and can be viewed as the first realisation of
constraint logic programming.

Constraint programming
Let us turn now our attention to constraint programming . The formal
concept of a constraint was used originally in physics and combinatorial
optimisation. It was first adopted in computer science by Ivan Sutherland in
Sutherland [1963] for describing his interactive drawing system Sketchpad.
In the seventies several experimental languages were proposed that used the
notion of constraints and relied on the concept of constraint solving. Also
in the seventies, in the field of artificial intelligence (AI), the concept of a
constraint satisfaction problem was formulated and used to describe
problems in computer vision. Further, starting with Montanari [1974] and
Mackworth [1977], the concept of constraint propagation was identified
as a crucial way of coping with the combinatorial explosion when solving
constraint satisfaction problems using top-down search.
Top-down search is a generic name for a set of search procedures in
which one attempts to construct a solution by systematically trying to ex-
tend a partial solution through the addition of constraints. In the simplest
case, each such constraint assigns a value to another variable. Common
to most top-down search procedures is backtracking , which can be traced
back to the nineteenth century. In turn, the branch and bound search, a
 Introduction xi

top-down search concerned with optimisation, was defined first in the con-
text of combinatorial optimisation.
In the eighties the first constraint programming languages of importance
were proposed and implemented. The most significant were the languages
based on the logic programming paradigm. They involve an extension of
logic programming by the notion of constraints. The main reason for the suc-
cess of this approach to constraint programming is that constraints and logic
programming predicates are both, mathematically, relations; backtracking
is automatically available; and the variables are viewed as unknowns in the
sense of algebra. The latter is in contrast to the imperative programming in
which the variables are viewed as changing, but each time known entities,
as in calculus.
The resulting paradigm is called constraint logic programming . As
mentioned above, Prolog III is an example of a programming language re-
alising this paradigm. The term was coined in the influential paper Jaffar
and Lassez [1987] that introduced the operational model and semantics for
this approach and formed a basis for the CLP(R) language that provided
support for solving constraints on reals, see Jaffar et al. [1992].
Another early constraint logic programming language is CHIP, see Dincbas
et al. [1988] and for a book coverage Van Hentenryck [1989]. CHIP incorpo-
rated the concept of a constraint satisfaction problem into the logic program-
ming paradigm by using constraint variables ranging over user-defined finite
domains. During the computation the values of the constraint variables are
not known, only their current domains. If a variable domain shrinks to
one value, then that is the final value of the variable. CHIP also relied on
top-down search techniques originally introduced in AI.
The language was developed at the European Computer-Industry Re-
search Centre (ECRC) in Munich. This brings us to the next stage in our
historical overview.

ECLi PSe
ECRC was set up in 1984 by three European companies to explore the devel-
opment of advanced reasoning techniques applicable to practical problems.
In particular three programming systems were designed and implemented.
One enabled complex problems to be solved on multiprocessor hardware,
and eventually on a network of machines. The second supported advanced
database techniques for intelligent processing in data-intensive applications.
The third system was CHIP. All three systems were built around a common
foundation of logic programming.
xii Introduction

In 1991 the three systems were merged and ECLi PSe was born. The con-
straint programming features of ECLi PSe were initially based on the CHIP
system, which was spun out from ECRC at that time in a separate company.
Over the next 15 years the constraint solvers and solver interfaces supported
by ECLi PSe have been continuously extended in response to users require-
ments.
The first released interface to an external state-of-the-art linear and mixed
integer programming package was in 1997. The integration of the finite do-
main solver and linear programming solver, supporting hybrid algorithms,
came in 2000. In 2001 the ic library was released. It supports constraints on
Booleans, integers and reals and meets the important demands of practical
use: it is sound, scalable, robust and orthogonal. ECLi PSe also includes as li-
braries some constraint logic programming languages, for example CLP(R),
that were developed separately. By contrast with the constraint solving fa-
cilities, the parallel programming and database facilities of ECLi PSe have
been much less used, and over the years some functionality has been dropped
from the system.
The ECLi PSe team was involved in a number of European research pro-
jects, especially the Esprit project CHIC Constraint Handling in Industry
and Commerce (19911994). Since the termination of ECRCs research ac-
tivities in 1996, ECLi PSe has actively been further developed at the Centre
for Planning and Resource Control at Imperial College in London (IC-Parc),
with funding from International Computers Ltd (ICL), the UK Engineering
and Physical Sciences Research Council, and the Esprit project CHIC-2
Creating Hybrid Algorithms for Industry and Commerce (19961999). The
Esprit projects played an important role in focussing ECLi PSe development.
In particular they emphasised the importance of the end user, and the time
and skills needed to learn constraint programming and to develop large scale
efficient and correct programs.
In 1999, the commercial rights to ECLi PSe were transferred to IC-Parcs
spin-off company Parc Technologies, which applied ECLi PSe in its optimi-
sation products and provided funding for its maintenance and continued
development. In August 2004, Parc Technologies, and with it the ECLi PSe
platform, was acquired by Cisco Systems.
ECLi PSe is in use at hundreds of institutions for teaching and research
all over the world, and continues to be freely available for education and
research purposes. It has been exploited in a variety of applications by
academics around the world, including production planning, transportation
scheduling, bioinformatics, optimisation of contracts, and many others. It
is also being used to develop commercial optimisation software for Cisco.
 Introduction xiii

Overview of the book

In this book we explain constraint logic programming using ECLi PSe . Since
ECLi PSe extends Prolog, we explain the latter first. The reader familiar
with Prolog may skip the first five chapters or treat them as a short reference
manual.
In Part I of the book we focus on the logic programming paradigm. We
begin by introducing in Chapter 1 a subset of Prolog called pure Prolog
and explaining the underlying computation model. In Chapter 2 we clarify
the programming features implicitly supported by pure Prolog by discussing
a toy procedural programming language and explaining how pure Prolog can
be translated into it.
Part II is devoted to a short exposition of Prolog. In Chapter 3 we
explain the Prolog approach to arithmetic. In particular, arithmetic con-
straints are allowed, but only for testing. We call constraints used in this
way passive. In Chapter 4 we discuss control in Prolog. Also, we explain
there how Prolog supports meta-programming, i.e., the possibility of writing
programs that use other programs as data. Next, in Chapter 5 we explain
how Prolog allows us to inspect, compare and decompose terms. This ex-
position of Prolog is incomplete: the standard Prolog has 102 built-ins and
we cover only those we need later.2
Part III begins with a primer on constraint programming, provided in
Chapter 6. We introduce here the concepts of constraints, constraint sat-
isfaction problems (CSPs) and constraint optimisation problems (COPs),
and discuss the basic techniques used to solve CSPs and COPs. Then in
Chapter 7 we return to ECLi PSe by introducing a large suite of iterators
that allow us to write programs without recursion. They are extensively used
in the remainder of the book. Next, in Chapter 8 we study the top-down
search in presence of passive constraints by considering search algorithms
for both complete and incomplete search.
One of the limitations of Prolog is that arithmetic constraints, when se-
lected too early cause a run-time error. In Chapter 9 we introduce the
suspend library of ECLi PSe that allows us to circumvent this problem by
automatically delaying various types of constraints, including arithmetic
constraints, until they reduce to tests.
Part IV forms the main part of the book. We begin by introducing in
Chapter 10 two ECLi PSe libraries, sd and ic, that treat constraints as
active constraints, by allowing their evaluation to affect their variables.

2 For those wishing to acquire the full knowledge of Prolog we recommend Bratko [2001] and
Sterling and Shapiro [1994].
xiv Introduction

This is achieved through the process of constraint propagation. In Chap-

ter 11 we explain how to combine the constraint propagation with ECLi PSe
facilities that support top-down search. Next, in Chapter 12 we discuss
the branch and bound library that allows us to solve COPs using ECLi PSe .
The last two chapters deal with constraints on continuous variables. In
Chapter 13 we explain how the approach to solving CSPs and COPs based
on the constraint propagation and top-down search can be modified to deal
with such constraints. Finally, in Chapter 14 we consider linear and non-
linear constraints over continuous and integer variables. In ECLi PSe they
are solved using the eplex library that provides an interface to a package
for solving mixed integer programming problems.

Resources on ECLi PSe

The main website of ECLi PSe is www.eclipse-clp.org. It provides a sub-
stantial body of documentation to aid users in getting the most out of the
system.
There are a tutorial and three manuals: a User Manual, a Reference
Manual and a Constraint Library Manual. There are documents about
ECLi PSe embedding and interfacing, and about applications development
in ECLi PSe . Also there are example ECLi PSe programs for solving a variety
of puzzles and applications in planning and scheduling. Finally, there is a
great deal of on-line help available, covering all (currently 2644) built-in
predicates.
The programs presented in this book can be found at www.cwi.nl/~apt/
eclipse.

Acknowledgements
Krzysztof Apt would like to thank his colleagues who during their stay at
CWI used ECLi PSe and helped him to understand it better. These are:
Sebastian Brand, Sandro Etalle, Eric Monfroy and Andrea Schaerf. Also,
he would like to warmly thank three people who have never used ECLi PSe
but without whose support this book would never have been written. These
are: Alma, Daniel and Ruth.
Mark Wallace offers this book as a testament both to the ECRC CHIP
team, who set the ball rolling, to Micha Meier at ECRC, and to the IC-
Parc ECLi PSe team, including Andy Cheadle, Andrew Eremin, Warwick
Harvey, Stefano Novello, Andrew Sadler, Kish Shen, and Helmut Simonis
of both the CHIP and ECLi PSe fame who have designed, built and
 Introduction xv

maintained the functionality described herein. Last but not least, he thanks
Joachim Schimpf, a great colleague, innovative thinker and brilliant software
engineer, who has shaped the whole ECLi PSe system.
Since dragging his reluctant family halfway round the world to Australia,
Mark has buried himself in his study and left them to it. With the comple-
tion of this book he looks forward to sharing some proper family weekends
with Duncan, Tessa and Toby and his long-suffering wonderful wife Ingrid.
The authors acknowledge helpful comments by Andrea Schaerf, Joachim
Schimpf, Maarten van Emden and Peter Zoeteweij. Also, they would like to
thank the School of Computing of the National University of Singapore for
making it possible for them to meet there on three occasions and to work
together on this book. The figures were kindly prepared by Ren Yuan using
the xfig drawing program.
To Alma, Daniel and Ruth and
to Duncan, Tessa, Toby and Ingrid
 Part I

Logic programming paradigm

 1

Logic programming and pure Prolog

1.1 Introduction 3
1.2 Syntax 4
1.3 The meaning of a program 7
1.4 Computing with equations 9
1.5 Prolog: the first steps 15
1.6 Two simple pure Prolog programs 23
1.7 Summary 26
1.8 Exercises 26

1.1 Introduction

L OGIC PROGRAMMING (LP in short) is a simple yet powerful for-

malism suitable for computing and for knowledge representation.
It provides a formal basis for Prolog and for constraint logic
programming . Other successful applications of LP include deductive
databases, an extension of relational databases by rules, a computational
approach to machine learning called inductive logic programming , and
a computational account of various forms of reasoning, in particular of non-
monotonic reasoning and meta-reasoning .
The logic programming paradigm substantially differs from other pro-
gramming paradigms. The reason for this is that it has its roots in auto-
mated theorem proving, from which it borrowed the notion of a deduction.
What is new is that in the process of deduction some values are computed.
When stripped to the bare essentials, this paradigm can be summarised by
the following three features:

any variable can stand for a number or a string, but also a list, a tree, a
record or even a procedure or program,

3
4 Logic programming and pure Prolog

during program execution variables are constrained, rather than being

assigned a value and updated,
program executions include choice points, where computation can take
alternative paths; if the constraints become inconsistent, the program
backtracks to the previous open choice point and tries another path.

In this chapter we discuss the logic programming framework and the cor-
responding small subset of Prolog, usually called pure Prolog . This will
allow us to set up a base over which we shall define in the successive chap-
ters a more realistic subset of Prolog supporting in particular arithmetic and
various control features. At a later stage we shall discuss various additions
to Prolog provided by ECLi PSe , including libraries that support constraint
programming.
We structure the chapter by focussing in turn on each of the above three
items. Also we clarify the intended meaning of pure Prolog programs.1
Consequently, we discuss in turn

the objects of computation and their syntax,

the meaning of pure Prolog programs,
the accumulation of constraints during program execution, and
the creation of choice points during program execution and backtracking.

1.2 Syntax
Syntactic conventions always play an important role in the discussion of
any programming paradigm and logic programming is no exception in this
matter. In this section we discuss the syntax of Prolog.
Full Prolog syntactic conventions are highly original and very powerful.
Their full impact is little known outside of the logic programming commu-
nity. We shall discuss these novel features in Chapters 3 and 4.
By the objects of computation we mean anything that can be denoted
by a Prolog variable. These are not only numbers, lists and so on, but also
compound structures and even other variables or programs.
Formally, the objects of computation are base terms, which consists of:

variables, denoted by strings starting with an upper case letter or

(the underscore), for example X3 is a variable,
numbers, which will be dealt with in Chapter 3,
1 To clarify the terminology: logic programs are pure Prolog programs written using the logical
and not Prolog notation. In what follows we rather focus on Prolog notation and, as a result,
on pure Prolog programs.
 1.2 Syntax 5

atoms, denoted by sequences of characters starting with a lower case

letter, for example x4 is an atom. Any sequence of characters put between
single quotes, even admitting spaces, is also an atom, for example My
Name,
and compound terms, which comprise a functor and a number of argu-
ments, each of which is itself a (base or compound) term.
By a constant we mean a number or an atom. Special case of terms are
ground terms which are terms in which no variable appears. In general,
the qualification ground, which we will also use for other syntactic objects,
simply means containing no variables.
In Chapter 4 we shall explain that thanks to the ambivalent syntax
facility of Prolog programs can also be viewed as compound terms. This
makes it possible to interpret programs as data, which is an important
feature of Prolog.
In the standard syntax for compound terms the functor is written first,
followed by the arguments separated by commas, and enclosed in round
brackets. For example the term with functor f and arguments a, b and c is
written f(a,b,c). Similarly, a more complex example with functor h and
three arguments:
(i) the variable A,
(ii) the compound term f(g,Twenty,X),
(iii) the variable X,
is written h(A,f(g,Twenty,X),X).
Some compound terms can be also written using an infix notation. Next,
we define goals, queries, clauses and programs. Here is a preliminary overview.
A program is made up of procedures.
A procedure is made up of clauses, each terminated by the period ..
A clause is either a fact or a rule.
There is no special procedure such as main and a user activates the pro-
gram by means of a query.
Now we present the details.
First we introduce an atomic goal which is the basic component from
which the clauses and queries are built. It has a predicate and a number
of arguments. The predicate has a predicate name and a predicate
arity . The predicate name is, syntactically, an atom. The arguments are
placed after the predicate name, separated by commas and surrounded by
round brackets. The number of arguments is the arity of the predicate.
6 Logic programming and pure Prolog

An example of an atomic goal is p(a,X). Here the predicate name is p

and its arity is 2. We often write p/n for the predicate with name p and
arity n.
A query (or a goal ) is a sequence of atomic goals terminated by the
period .. A query is called atomic if it consists of exactly one atomic
goal.
A rule comprises a head , which is an atomic goal, followed by :-,
followed by the body which is a non-empty query, and is terminated by
the period ..
An example of a rule is
p(b,Y) :- q(Y), r(Y,c).
A rule contributes to the definition of the predicate of its head, so this
example rule contributes to the definition of p/2. We also call this a
rule for p/2. Rule bodies may contain some atomic goals with a binary
predicate and two arguments written in the infix form, for example X =
a, whose predicate is =/2.
A fact is an atomic goal terminated by the period .. For example
p(a,b).
is a fact. This fact also contributes to the definition of p/2.
A sequence of clauses for the same predicate makes a procedure. The
procedure provides a definition for the predicate.
For example, here is a definition for the predicate p/2:
p(a,b).
p(b,Y) :- q(Y), r(Y,c).
A pure Prolog program is a finite sequence of procedures, for example
p(a,b).
p(b,Y) :- q(Y), r(Y,c).
q(a).
r(a,c).
So a program is simply a sequence of clauses.
In what follows, when discussing specific queries and rules in a running
text we shall drop the final period ..
Before leaving this discussion of the syntax of pure Prolog, we point out a
small but important feature of variable naming. If a variable appears more
than once in a query (or similarly, more than once in a program clause), then
 1.3 The meaning of a program 7

each occurrence denotes the same object of computation. However, Prolog

also allows so-called anonymous variables, written as (underscore).
These variables have a special interpretation, because each occurrence of
in a query or in a clause is interpreted as a different variable. That is
why we talk about the anonymous variables and not about the anonymous
variable. So by definition each anonymous variable occurs in a query or a
clause only once.
Anonymous variables form a simple and elegant device and, as we shall
see, their use increases the readability of programs in a remarkable way.
ECLi PSe and many modern versions of Prolog encourage the use of anony-
mous variables by issuing a warning if a non-anonymous variable is encoun-
tered that occurs only once in a clause. This warning can be suppressed
by using a normal variable that starts with the underscore symbol , for
example _X.
Prolog has several (about one hundred) built-in predicates, so predicates
with a predefined meaning. The clauses, the heads of which refer to these
built-in predicates, are ignored. This ensures that the built-in predicates
cannot be redefined. Thus one can rely on their prescribed meaning. In
ECLi PSe and several versions of Prolog a warning is issued in case an at-
tempt at redefining a built-in predicate is encountered.
So much about Prolog syntax for a moment. We shall return to it in
Chapters 3 and 4 where we shall discuss several novel and powerful features
of Prolog syntax.

1.3 The meaning of a program

Pure Prolog programs can be interpreted as statements in the first-order
logic. This interpretation makes it easy to understand the behaviour of a
program. In particular, it will help us to understand the results of evaluating
a query w.r.t. to a program. In the remainder of this section we assume that
the reader is familiar with the first-order logic.
Let us start by interpreting a simple program fact, such as p(a,b). It
contributes to the definition of the predicate p/2. Its arguments are two
atoms, a and b.
We interpret the predicate p/2 as a relation symbol p of arity 2. The
atoms a and b are interpreted as logical constants a and b. A logical con-
stant denotes a value. Since, in the interpretation of pure Prolog programs,
different constants denote different values, we can think of each constant
denoting itself. Consequently we interpret the fact p(a,b) as the atomic
formula p(a, b).
8 Logic programming and pure Prolog

The arguments of facts may also be variables and compound terms. Con-
sider for example the fact p(a,f(b)). The interpretation of the compound
term f(b) is a logical expression, in which the unary function f is applied
to the logical constant b. Under the interpretation of pure Prolog programs,
the denotations of any two distinct ground terms are themselves distinct.2
Consequently we can think of ground terms as denoting themselves, and so
we interpret the fact p(a,f(b)) as the atomic formula p(a, f (b)).
The next fact has a variable argument: p(a,Y). We view it as a statement
that for all ground terms t the atomic formula p(a, t) is true. So we interpret
it as the universally quantified formula Y. p(a, Y ).
With this interpretation there can be no use in writing the procedure
p(a,Y).
p(a,b).
because the second fact is already covered by the first, more general fact.
Finally we should mention that facts with no arguments are also admit-
ted. Accordingly we can assert the fact p. Its logical interpretation is the
proposition p.
In general, we interpret a fact by simply changing the font from teletype
to italic and by preceding it by the universal quantification of all variables
that appear in it.
The interpretation of a rule involves a logical implication. For example
the rule
p :- q.
states that if q is true then p is true.
As another example, consider the ground rule
p(a,b) :- q(a,f(c)), r(d).
Its interpretation is as follows. If q(a, f (c)) and r(d) are both true, then
p(a, b) is also true, i.e., q(a, f (c)) r(d) p(a, b).
Rules with variables need a little more thought. The rule
p(X) :- q.
states that if q is true, then p(t) is true for any ground term t. So logically
this rule is interpreted as q X. p(X). This is equivalent to the formula
X. (q p(X)).
If the variable in the head also appears in the body, the meaning is the
same. The rule
2 This will no longer hold for arithmetic expressions which will be covered in Chapter 3.
 1.4 Computing with equations 9

p(X) :- q(X).
states that for any ground term t, if q(t) is true, then p(t) is also true.
Therefore logically this rule is interpreted as X. (q(X) p(X)).
Finally, we consider rules in which variables appear in the body but not
in the head, for example
p(a) :- q(X).
This rule states that if we can find a ground term t for which q(t) is true,
then p(a) is true. Logically this rule is interpreted as X. (q(X) p(a)),
which is equivalent to the formula (X. q(X)) p(a).
Given an atomic goal A denote its interpretation by A. Any ground rule H
:- B1 , . . . , Bn is interpreted as the implication B1 . . . Bn H. In general,
all rules H :- B1 , . . . , Bn have the same, uniform, logical interpretation. If
V is the list of the variables appearing in the rule, its logical interpretation
is V. (B1 . . . Bn H).
This interpretation of , (as ) and :- (as ) leads to so-called declara-
tive interpretation of pure Prolog programs that focusses through their
translation to the first-order logic on their semantic meaning.
The computational interpretation of pure Prolog is usually called pro-
cedural interpretation. It will be discussed in the next section. In this
interpretation the comma , separating atomic goals in a query or in a body
of a rule is interpreted as the semicolon symbol ; of the imperative program-
ming and :- as (essentially) the separator between the procedure header
and body.

1.4 Computing with equations

We defined in Section 1.2 the computational objects over which computa-
tions of pure Prolog programs take place. The next step is to explain how
variables and computational objects become constrained to be equal to each
other, in the form of the answers. This is the closest that logic programming
comes to assigning values to variables.

1.4.1 Querying pure Prolog programs

The computation process of pure Prolog involves a program P against which
we pose a query Q. This can lead to a successful, failed or diverging com-
putation (which of course we wish to avoid).
A successful computation yields an answer, which specifies constraints on
the query variables under which the query is true. In this subsection we
10 Logic programming and pure Prolog

describe how these constraints are accumulated during query processing. In

the next subsection we will describe how the constraints are checked for
consistency and answers are extracted from them.
The constraints accumulated by Prolog are equations whose conjunction
logically entails the truth of the query. To clarify the discussion we will use
the following simple program:

p(X) :- q(X,a).
q(Y,Y).

Let us first discuss the answer to the atomic query

q(W,a).

The definition of the predicate q/2 comprises just the single fact q(Y,Y).
Clearly the query can only succeed if W = a.
Inside Prolog, however, this constraint is represented as an equation be-
tween two atomic goals: q(W,a) = q(Y1,Y1). The atomic goal q(W,a) at
the left-hand side of the equation is just the original query. For the fact
q(Y,Y), however, a new variable Y1 has been introduced. This is not im-
portant for the current example, but it is necessary in general because of
possible variable clashes. This complication is solved by using a different
variable each time. Accordingly, our first query succeeds under the con-
straint q(W,a) = q(Y1,Y1).
Now consider the query

p(a).

This time we need to use a rule instead of a fact. Again a new variable is
introduced for each use of the rule, so this first time it becomes:

p(X1) :- q(X1,a).

To answer this query, Prolog first adds the constraint p(a) = p(X1),
which constrains the query to match the definition of p/1. Further, the
query p(a) succeeds only if the body q(X1,a) succeeds, which it does un-
der the additional constraint q(X1,a) = q(Y1,Y1). The complete sequence
of constraints under which the query succeeds is therefore p(a) = p(X1),
q(X1,a) = q(Y1,Y1). Informally we can observe that these constraints hold
if all the variables take the value a.
Consider now the query:

p(b).
 1.4 Computing with equations 11

Reasoning as before, we find the query would succeed under the constraints:
p(b) = p(X1), q(X1,a) = q(Y1,Y1). In this case, however, there are no
possible values for the variables which would satisfy these constraints. Y1
would have to be equal both to a and to b, which is impossible. Consequently
the query fails.
Next consider a non-atomic query
p(a), q(W,a).
The execution of this query proceeds in two stages. First, as we already saw,
p(a) succeeds under the constraints p(a) = p(X1), q(X1,a) = q(Y1,Y1),
and secondly q(W,a) succeeds under the constraint q(W,a) = q(Y2,Y2).
The complete sequence of constraints is therefore: p(a) = p(X1), q(X1,a)
= q(Y1,Y1), q(W,a) = q(Y2,Y2). Informally these constraints are satis-
fied if all the variables take the value a.
A failing non-atomic query is:
p(W), q(W,b).
Indeed, this would succeed under the constraints p(W) = p(X1), q(X1,a)
= q(Y1,Y1), q(W,b) = q(Y2,Y2). However, for this to be satisfied W would
have to take both the value a (to satisfy the first two equations) and b (to
satisfy the last equation), so the constraints cannot be satisfied and the
query fails.
Operationally, the constraints are added to the sequence during compu-
tation, and tested for consistency immediately. Thus the query
p(b), q(W,b).
fails already during the evaluation of the first atomic query, because already
at this stage the accumulated constraints are inconsistent. Consequently the
second atomic query is not evaluated at all.

1.4.2 Most general unifiers

The Prolog system has a built-in algorithm which can detect whether a
sequence of equations between atomic goals is consistent or not. In general,
two outcomes are possible. Either the algorithm returns a failure, because
some atomic goals cannot be made equal (for example p(X,a) = p(b,X)),
or it yields a positive answer in the form of a substitution, which is a set
of equations of the form Variable = Term such that each variable occurs at
most once at a left-hand side of an equation. The resulting set of equations
is called a unifier .
12 Logic programming and pure Prolog

For example, the equation p(X,f(X,Y),c) = p(V,W,V) has a unifier {V =

c, X = c, W = f(c,Y)}. Also {V = c, X = c, W = f(c,a)} is a unifier
but the first one is more general in a sense that can be made precise. Such a
most general substitution that subsumes all others is called a most general
unifier (usually abbreviated to mgu). If an equation between two atomic
goals has a unifier, the atomic goals are called unifiable.
The problem of deciding whether a sequence of equations between atomic
goals has a unifier is called the unification problem. The following result
was established by A. Robinson.
Unification theorem An algorithm exists, called a unification algo-
rithm, that for a sequence of equations between atomic goals determines
whether they are consistent and if they do, it produces their mgu.
An answer to a query is then any most general unifier of the sequence
of constraints accumulated during query evaluation.
An elegant unification algorithm was introduced by M. Martelli and U.
Montanari. Given a term t we denote here the set of its variables by Var (t).

MartelliMontanari Algorithm
Non-deterministically choose from the set of equations an equation of a form
below and perform the associated action.

(1) f (s1 , ..., sn ) = f (t1 , ..., tn ) replace by the equations

s1 = t1 , ..., sn = tn ,
(2) f (s1 , ..., sn ) = g(t1 , ..., tm ) where f 6 g halt with failure,
(3) x = x delete the equation,
(4) t = x where t is not a variable replace by the equation x = t,
(5) x = t where x 6 Var (t) perform the substitution {x/t}
and x occurs elsewhere on all other equations
(6) x = t where x Var (t) and x 6 t halt with failure.

The algorithm starts with the original sequence of equations and termi-
nates when no action can be performed or when failure arises. In case of
success we obtain the desired mgu.
Note that in the borderline case, when n = 0, action (1) includes the case
c = c for every constant c which leads to deletion of such an equation. In
addition, action (2) includes the case of two different constants and the cases
where a constant is compared with a term that is neither a constant or a
variable.
 1.4 Computing with equations 13

To illustrate the operation of this algorithm consider some examples. To

facilitate the reading the selected equations are underlined.
(i) Consider the equation

p(k(Z, f(X,b,Z))) = p(k(h(X), f(g(a),Y,Z))).

Applying action (1) twice yields the set

{Z = h(X), f(X, b, Z) = f(g(a), Y, Z)}.

Choosing the second equation again action (1) applies and yields

{Z = h(X), X = g(a), b = Y, Z = Z}.

Choosing the third equation action (4) applies and yields

{Z = h(X), X = g(a), Y = b, Z = Z}.

Now, choosing the last equation action (3) applies and yields

{Z = h(X), X = g(a), Y = b}.

Finally, choosing the second equation action (5) applies and yields

{Z = h(g(a)), X = g(a), Y = b}.

At this stage no action applies, so {Z = h(g(a)), X = g(a), Y = b} is an mgu

of the atomic goals p(k(Z, f(X,b,Z))) and p(k(h(X), f(g(a),Y,Z))).

(ii) Consider now the equation

p(k(h(a), f(X,b,Z))) = p(k(h(b), f(g(a),Y,Z))).

Again applying action (1) twice yields the set

{h(a) = h(b), f(X, b, Z) = f(g(a), Y, Z)}.

Next, choosing the first equation action (1) applies again and yields

{a = b, f(X, b, Z) = f(g(a), Y, Z)}.

Choosing again the first equation action (2) applies and a failure arises. So
the atomic goals p(k(h(a), f(X,b,Z))) and p(k(h(b), f(g(a),Y,Z)))
are not unifiable.

(iii) Finally, consider the equation

p(k(Z, f(X,b,Z))) = p(k(h(X), f(g(Z),Y,Z))).

14 Logic programming and pure Prolog

Let us try to repeat the choices made in (i). Applying action (1) twice we
get the set
{Z = h(X), f(X, b, Z) = f(g(Z), Y, Z)}.
Next, choosing the second equation action (1) applies again and yields
{Z = h(X), X = g(Z), b = Y, Z = Z}.
Choosing the third equation action (4) applies and yields
{Z = h(X), X = g(Z), Y = b, Z = Z}.
Now, choosing the fourth equation action (3) applies and yields
{Z = h(X), X = g(Z), Y = b}.
Finally, choosing the second equation action (5) applies and yields
{Z = h(g(Z)), X = g(Z), Y = b}.
But now choosing the first equation action (6) applies and a failure arises.
Hence the atomic goals p(k(Z, f(X,b,Z))) and p(k(h(X),f(g(Z),Y,Z)))
are not unifiable.

1.4.3 Choice points and backtracking

Until now we did not fully explain what happens during a program execu-
tion when an equation, for instance q(b,a) = q(Y,Y), fails. When such
an equation is generated during the execution of a query, backtracking is
triggered, a process we first illustrate using the following simple program:
p(X) :- X = a.
p(X) :- X = b.
Each of the two rules in this program has a body consisting of an atomic
goal, with predicate =/2 written in the infix form. The behaviour of this
predicate is built into the underlying implementation and will be discussed
in the next section. Its logical interpretation is, of course, equality. When
we pose the query
p(a).
against this program, it therefore succeeds, returning the answer Yes.
Next, consider what happens when we pose the query
p(b).
 1.5 Prolog: the first steps 15

According to the computation process explained in Subsection 1.4.1, using

the first rule we accumulate two equations, p(b) = p(X1), X1 = a, which
are inconsistent. This might suggest that the answer to the query p(b) is
No. However, this disregards the fact that in the program the predicate
p/2 is defined by means of two clauses.
When the first equation, p(b) = p(X1), is considered a choice point
is created, which consists of the second clause, p(X1) :- X1 = b, and the
current state (in which values of all the variables are recorded). Now when
the failure arises during the execution that starts with the selection of the
first clause, the computation returns to the choice point, the state is restored
to the one stored and the second clause is selected. So all the substitutions
performed since the creation of this choice point are undone. Now we accu-
mulate the equations p(b) = p(X1), X1 = b, which are consistent. So the
answer is Yes.
Finally, when we pose the query
p(c).
against this program, we get the answer No because both X1 = a and X1
= b fail under the initial condition X1 = c.
In general, if a predicate is defined by n clauses with n > 1, then upon
selection of the first clause all the other clauses are kept in a choice point.
When during the execution which starts by selecting the first clause a failure
arises, the computation returns to the last created choice point and resumes
in the state stored there but with the next clause stored in this choice point.
If the last clause is selected, the choice point is removed. So when a failure
arises during the execution that involves the last clause, the computation
resumes at the one before last choice point, if such a choice point exists and
otherwise a failure results. This computation process is called backtrack-
ing .

1.5 Prolog: the first steps

1.5.1 Prolog: how to run it
It is useful at this stage to get an idea how one interacts with a Prolog system.
There are two ways of running a Prolog program using ECLi PSe . The first
is using tkeclipse, which provides an interactive graphical user interface
to the ECLi PSe compiler and system. The second is a more traditional
command line interface accessed by invoking just eclipse.
To start TkECLi PSe , either type the command tkeclipse at an operating
system command line prompt, or select TkECLi PSe from the program menu,
16 Logic programming and pure Prolog

on Windows. This will bring up the TkECLi PSe top level shown in Figure
1.1.

Fig. 1.1 Top level of TkECLi PSe

Help for TkECLi PSe and its component tools is available from the Help
menu in the TkECLi PSe window.
To write your programs you need an editor. From the editor, save your
program as a plain text file, and then you can compile the program into
ECLi PSe . To compile the program in TkECLi PSe , select the Compile option
from the File menu. This will bring up a file selection dialogue, from which
you select the file you want to compile and click on the Open button. This
will compile the file, and any other it depends on.
If you have edited a file and want to compile it again (together with any
other files that may have changed), you just click on the make button. Now
a query to the program can be submitted, by typing it into the Goal Entry
field.
 1.5 Prolog: the first steps 17

The idea is that the system evaluates the query with respect to the pro-
gram read-in and reports an answer. There are three possible outcomes.

The query succeeds.

Then an answer is printed followed by Yes. The answer corresponds
to the logical answer, so it shows for what values of the variables the query
becomes true.
Assuming the query has one or more variables, another answer can be
obtained by clicking on the more button.
The query fails or a request to produce the next answer fails.
Then the answer No is printed.
The query diverges.
You guessed it: then nothing happens. In this case, you can interrupt
the execution by clicking on the interrupt button.
In the command line interface the interaction starts by typing eclipse.
The ECLi PSe system replies by the prompt [eclipse 1]:. Now, assum-
ing that the program is stored in the file filename.ecl it can be read in by
typing [filename] (so without the suffix .ecl), followed by the period ..
(ECLi PSe also supports Prolog suffix .pl). Assuming that the program
is syntactically correct, the system replies with the answer Yes followed
by the prompt [eclipse 2]:. Now a query to the program can be sub-
mitted (so with the period . at its end). If the query succeeds, typing the
Enter key terminates the computation, whereas typing ; is interpreted as
the request to produce the next answer. Finally, typing halt. (which is
a propositional built-in) finishes the interaction with the system.
To illustrate these matters let us return to the program
p(X) :- q(X,a).
q(Y,Y).
discussed in Subsection 1.4.1. Assume it is stored in the file example1.ecl.
We have then the following interaction with ECLi PSe , where we reconsider
the previously discussed queries.
[eclipse 1]: [example1].
example1.ecl compiled traceable 72 bytes in 0.00 seconds

Yes (0.00s cpu)

[eclipse 2]: q(W,a).

W = a
18 Logic programming and pure Prolog

Yes (0.00s cpu)

[eclipse 3]: p(a).

Yes (0.00s cpu)

[eclipse 4]: p(b).

No (0.00s cpu)
[eclipse 5]: p(a), q(W,a).

W = a
Yes (0.00s cpu)
[eclipse 6]: p(W), q(W,b).

No (0.00s cpu)

1.5.2 Unification in Prolog

Armed with the knowledge of how to execute Prolog programs let us discuss
now how unification is available in Prolog. It is provided by means of the
already mentioned built-in =/2 defined internally by the single fact presented
in Figure 1.2.

% X = Y :- X and Y are unifiable.

X = X.

Fig. 1.2 The EQUALITY program

By ignoring the differences between the predicates and function symbols

we can equally well consider equations between a pair of atomic queries.
To illustrate the effect of =/2 let us return to the pairs of atomic queries
discussed in Subsection 1.4.2. We have then:

[eclipse 7]: p(k(Z, f(X,b,Z))) = p(k(h(X), f(g(a),Y,Z))).

Z = h(g(a))
X = g(a)
Y = b
Yes (0.00s cpu)
 1.5 Prolog: the first steps 19

For the second pair of atomic queries we get the same outcome as in
Subsection 1.4.2, as well:
[eclipse 8]: p(k(h(a), f(X,b,Z))) = p(k(h(b), f(g(a),Y,Z))).

No (0.00s cpu)
However, for the third pair of atomic queries we get a puzzling answer:
[eclipse 9]: p(k(Z, f(X,b,Z))) = p(k(h(X), f(g(Z),Y,Z))).

Z = h(g(h(g(h(g(h(g(h(g(h(g(h(g(h(g(...))))))))))))))))
X = g(h(g(h(g(h(g(h(g(h(g(h(g(h(g(h(...))))))))))))))))
Y = b
Yes (0.00s cpu)
The point is that for the efficiency reasons unification in Prolog is imple-
mented with the test x Var (t) (so whether x occurs in t) in the actions
(5) and (6) of the MartelliMontanari algorithm omitted. This test is
called an occur check . If x does occur in t, then instead of a failure a
circular binding is generated.
In ECLi PSe the situation is restored, by setting a specific flag on:
[eclipse 10]: set_flag(occur_check,on).

Yes (0.00s cpu)

[eclipse 11]: p(k(Z, f(X,b,Z))) = p(k(h(X), f(g(Z),Y,Z))).

No (0.00s cpu)
In practice it is very rare that this complication arises so the default is
that this flag is not set.
Let us return now to the program
p(X) :- X = a.
p(X) :- X = b.
discussed in Subsection 1.4.3 and consider the query p(X):
[eclipse 12]: p(X).

X = a
Yes (0.00s cpu, solution 1, maybe more) ? ;

X = b
20 Logic programming and pure Prolog

Yes (0.00s cpu, solution 2)

Here backtracking arises when the user requests more solutions by typing
; after a solution is displayed. The computation then returns to the last
created choice point. This leads to a second solution. In general, upon a
request for more solutions, a computation can yield another solution, but
does not have to, or it may diverge. If no choice point exists, the computation
terminates.

1.5.3 Comparing unification to assignment

To better understand the unification process it is useful to reflect for a mo-
ment on the similarities and differences between the unification and the
customary assignment statement present in the traditional programming
languages. In the examples below, following the computational interpreta-
tion of Prolog programs we interpret the comma , separating the atomic
goals in a query as the semicolon symbol ;.
To avoid confusion, when discussing the assignment statement we use
below the lower case letters to denote variables. Superficially, unification
seems to be very similar to the assignment statement. Indeed, the query
X = 1, Y = 2 yields the same result as the sequence of the assignments x
= 1; y = 2 and in both cases we can reverse the order of the equations,
respectively assignments, without affecting the outcome. However, this is
about all one can say about the similarities between these two concepts.
Here is a list of important differences.
Unification can succeed or fail, while assignment can succeed or abort (for
example when performing division by zero).
However, from a failure the computation can recover by backtracking
(we shall clarify this matter in the next section), while abortion is fatal.
Using unification we can assign terms to variables, for example f(a,b),
or even terms with variables, for example g(X,a).
Variables that appear in the values assigned to other variables are called
logical variables. The use of logical variables is a distinguishing feature
of the logic programming paradigm. It is a powerful mechanism that
allows us to compute a final object in stages, by passing a partially in-
stantiated output computed by one procedure call as an input to another
procedure call.
Once a ground term, in particular, a constant, is assigned to a variable by
means of unification, this variable cannot be modified anew. In contrast,
if the assignment is used, then one can modify the values of the variables.
 1.5 Prolog: the first steps 21

For example, the execution of the query X = 1, X = 2 yields a failure:

[eclipse 13]: X = 1, X = 2.

No (0.00s cpu)

while after performing the sequence of assignments x = 1; x = 2 the

value of x is 2.
This feature of the unification is sometimes called a one time assign-
ment.
The query of the form X = s, where X occurs in s but differs from s, always
fails, while the corresponding assignment x = s succeeds or aborts.
For example, the query X = X+1 fails, while the assignment x = x+1
leads to an increment of x. Further, the query X = X/0 fails, while the
assignment x = x/0 aborts.
Unification is symmetric, while assignment is not.
In particular, the queries X = Y and Y = X have always the same effect,
while the assignments x = y and y = x do not necessarily.
Unification, in contrast to the assignment, can be used to assign values to
several variables at the same time.
For example, the query f(b,X) = f(Y,a) assigns a to X and b to Y
(i.e., is equivalent to X = a, Y = b).

These differences between unification and the assignment statement reflect

the different views of variables. In logic programming and pure Prolog
variables are viewed as unknowns, very much like in mathematics. This is
in contrast to the imperative programming languages, such as C or Java, in
which variables represent known but varying quantities.

1.5.4 Lists in Prolog

A data structure that supports use of sequences with a single operation on
them an insertion of an element at the front is called a list. Lists form a
fundamental data structure and in Prolog special, built-in notational facil-
ities for them are available. In particular, the pair consisting of a constant
[] and a binary function symbol [.|..] is used to define them. Formally,
lists in Prolog are defined inductively as follows:

[] is a list, called the empty list.

if t is a list, then for any term h also [h | t] is a list; h is called its head
and t is called its tail .
22 Logic programming and pure Prolog

For example, [s(0)|[]] and [0|[X|[]]] are lists, whereas [0|s(0)] is

not, because s(0) is not a list. The tree depicted in Figure 1.3 represents
the list [a|[b|[c|[]]]].
[. ..]

a [. ..]

b [. ..]

c []

Fig. 1.3 A list

The list notation is not very readable and even short lists become difficult
to parse, as the depicted list [a|[b|[c|[]]]] shows.
So the following shorthands are carried out internally in Prolog for m 1
and n 0, also within the subterms:
[s0 |[s1 , ..., sm |t]] abbreviates to [s0 , s1 , ..., sm |t],
[s1 , ..., sm |[t1 , ..., tn ]] abbreviates to [s1 , ..., sm , t1 , ..., tn ].
Thus for example, [a|[b|c]] abbreviates to [a,b|c], and the depicted
list [a|[b|[c|[]]]] abbreviates to a more readable form, [a,b,c].
The following interaction with ECLi PSe shows that these simplifications
are also carried out internally.
[eclipse 14]: X = [a | [b | c]].

X = [a, b|c]
Yes (0.00s cpu)
[eclipse 15]: [a,b |c] = [a | [b | c]].

Yes (0.00s cpu)

[eclipse 16]: f(X) = f([a,b,c | [d,e,f]]).

X = [a, b, c, d, e, f]
Yes (0.00s cpu)
 1.6 Two simple pure Prolog programs 23

[eclipse 17]: [a,b,c] = [a | [b, c | []]].

Yes (0.00s cpu)

These abbreviations are widely used in Prolog programs. Also, to enhance
the readability, one uses in Prolog programs the names ending with s to
denote variables which are meant to be instantiated to lists.

1.6 Two simple pure Prolog programs

To illustrate some basic features of Prolog programming we now consider
two simple examples.

1.6.1 The BOOKS program

Consider the self-explanatory program given in Figure 1.4.

book(harry_potter, rowlings).
book(anna_karenina, tolstoy).
book(elements, euclid).
book(histories, herodotus).
book(constraint_logic_programming, apt).
book(constraint_logic_programming, wallace).

genre(harry_potter, fiction).
genre(anna_karenina, fiction).
genre(elements, science).
genre(histories, history).
genre(constraint_logic_programming, science).

author(Author, Genre) :- book(X, Author), genre(X, Genre).

Fig. 1.4 The BOOKS program

This program can be used both for testing:

[eclipse 18]: book(anna_karenina, tolstoy).

Yes (0.00s cpu)

[eclipse 19]: book(anna_karenina, rowlings).
24 Logic programming and pure Prolog

No (0.00s cpu)

and to compute one or more solutions, like in the following two queries:

[eclipse 20]: author(Author, history).

Author = herodotus
Yes (0.00s cpu, solution 1, maybe more) ? ;

No (0.00s cpu)

[eclipse 21]: book(constraint_logic_programming, Author).

Author = apt
Yes (0.00s cpu, solution 1, maybe more) ? ;

Author = wallace
Yes (0.00s cpu, solution 2)

Recall that ; is the users request for more solutions.

1.6.2 The APPEND program

To illustrate the use of lists we now discuss the perhaps most often cited Pro-
log program. This program concatenates two lists. The inductive definition
of concatenation is as follows:

the concatenation of the empty list [] and the list ys yields the list ys,
if the concatenation of the lists xs and ys equals zs, then the concatena-
tion of the lists [x | xs] and ys equals [x | zs].
This translates into the program given in Figure 1.5.3

% app(Xs, Ys, Zs) :- Zs is the result of

% concatenating the lists Xs and Ys.

app([], Ys, Ys).

app([X | Xs], Ys, [X | Zs]) :- app(Xs, Ys, Zs).

Fig. 1.5 The APPEND program

3 append/3 is actually a built-in so we denote the defined predicate by app/3.

 1.6 Two simple pure Prolog programs 25

So the app/3 predicate is defined by means of recursion, which is typical

for Prolog programming style.
The APPEND program can be used not only to concatenate the lists:

[eclipse 22]: app([mon, wed], [fri, sun], Zs).

Zs = [mon, wed, fri, sun]

Yes (0.00s cpu)

but also to split a list in all possible ways:

[eclipse 23]: app(Xs, Ys, [mon, wed, fri, sun]).

Xs = []
Ys = [mon, wed, fri, sun]
Yes (0.00s cpu, solution 1, maybe more) ? ;

Xs = [mon]
Ys = [wed, fri, sun]
Yes (0.00s cpu, solution 2, maybe more) ? ;

Xs = [mon, wed]
Ys = [fri, sun]
Yes (0.00s cpu, solution 3, maybe more) ? ;

Xs = [mon, wed, fri]

Ys = [sun]
Yes (0.00s cpu, solution 4, maybe more) ? ;

Xs = [mon, wed, fri, sun]

Ys = []
Yes (0.01s cpu, solution 5)

By combining these two ways of using APPEND we can easily generate

rotations of a list:

rotate(Xs, Ys) :- app(As, Bs, Xs), app(Bs, As, Ys).

The first call to app/3 generates upon backtracking all possible splits of its
last argument into two lists, while the second call concatenates these lists
in the reverse order. By imposing an additional condition on the output we
can then generate all rotations that satisfy some condition. For example, the
26 Logic programming and pure Prolog

following query generates in the variable Zs all rotations of a list of integers

that begin with 2:

[eclipse 24]: Zs = [2|_], rotate([1,2,2,3,1,2], Zs).

Zs = [2, 2, 3, 1, 2, 1]
Yes (0.00s cpu, solution 1, maybe more) ? ;

Zs = [2, 3, 1, 2, 1, 2]
Yes (0.00s cpu, solution 2, maybe more) ? ;

Zs = [2, 1, 2, 2, 3, 1]
Yes (0.00s cpu, solution 3, maybe more) ? ;

No (0.00s cpu)

1.7 Summary
In this chapter we introduced pure Prolog programs and discussed their
syntax, meaning and the computation process. In pure Prolog programs the
values are assigned to variables by means of the unification algorithm.
To explain the execution model of pure Prolog we discussed computing
with equations, choice points and backtracking. Finally, to clarify some
basic features of Prolog programming we introduced two simple pure Prolog
programs.

1.8 Exercises

Exercise 1.1 What is the ECLi PSe output to the following queries?

1. X = 7, X = 6.

2. X = 7, X = X+1.

3. X = X+1, X = 7.

4. X = 3, Y = X+1.

5. Y = X+1, X = 3.
 1.8 Exercises 27

Exercise 1.2 Consider the program BOOKS from Subsection 1.6.1. Using
X \= Y to test whether X is different from Y, what query do you use to find
the other author of the book by apt?

Exercise 1.3 Consider the following program:

q(X) :- Z = 1.
What is the ECLi PSe output to the following queries?

1. Z = 2, q(X).

2. Z = 2, q(Z).

3. Z = 2, q(X), Y = Z+1.

Exercise 1.4 Consider the following program:

p([prolog, works, for, nothing]).
p([prolog, is, pure, fantasy]).
p([prolog, works, for, solving(theories), only]).
p([prolog, works, for, solving(_), among, others]).
p([prolog, works, for, making(problems), and, others]).

q([programming, is, pure, fantasy]).

q([logic, works, in, some, cases]).
q([thanks, to, thierry, le, provost, for, this, example]).
q([prolog, works, for, solving(problems) | _]).
q([prolog, works, for, solving(_), between, others]).
What is the ECLi PSe output to the query p(Behold), q(Behold)?

Exercise 1.5 Reconsider the predicate rotate/2 defined on page 25. Ex-
plain why in the following interaction with ECLi PSe the first solution is
produced twice. (The query asks for all rotations of the list [1,2,2,3,1,2]
that begin with 1.)

[eclipse 1]: Zs = [1|_], rotate([1,2,2,3,1,2], Zs).

Zs = [1, 2, 2, 3, 1, 2]
Yes (0.00s cpu, solution 1, maybe more) ? ;

Zs = [1, 2, 1, 2, 2, 3]
28 Logic programming and pure Prolog

Yes (0.00s cpu, solution 2, maybe more) ? ;

Zs = [1, 2, 2, 3, 1, 2]
Yes (0.00s cpu, solution 3)
Suggest a modification of rotate/2, using \=, for which each solution is
produced once.
 2

A reconstruction of pure Prolog

2.1 Introduction 29
2.2 The programming language L0 29
2.3 Translating pure Prolog into L0 33
2.4 Pure Prolog and declarative programming 35
2.5 L0 and declarative programming 36
2.6 Summary 37
2.7 Exercises 38

2.1 Introduction

I N THE PREVIOUS chapter we discussed the basic aspects of the logic

programming paradigm by focussing on pure Prolog. To provide a
better insight into this paradigm we discuss in this chapter a translation
of pure Prolog programs into a more conventional programming language.
This translation allows us to clarify that pure Prolog implicitly supports
a number of programming features, including sequencing, procedures and
local variables.
Then we discuss the issue of declarative programming. We clarify in what
sense both pure Prolog and the small programming language here introduced
support declarative programming.

2.2 The programming language L0

In Subsection 1.5.3 we compared the equality interpreted as the unification
to the assignment statement. But, so interpreted, equality is not explicitly
present as a statement in logic programming and pure Prolog: it is hidden
in the computing process. Also the process of selection of the appropriate
clauses during an execution of a query is implicit and no distinction is made
between local and non-local variables.

29
30 A reconstruction of pure Prolog

To make these and other aspects of pure Prolog explicit we now define
a translation of pure Prolog into a small programming language L0 that
does include equality as an explicit atomic statement. Additionally, this
language features a number of other programming constructs, including re-
cursive procedures and an alternative choice statement. This interpretation
of the pure Prolog programs within a more conventional programming lan-
guage provides an alternative account of their semantics, a matter to which
we return in Sections 2.4 and 2.5.
To start with, we assume the same set of terms as in the case of pure
Prolog. The L0 programming language comprises
four types of atomic statements:
the equality statement,
the skip statement,
the fail statement,
the procedure call statement,
the program composition, written as ;,
the alternative choice statement, written as

begin S1 orelse . . . orelse Sn end,

where n 1,
the block statement in which a sequence of local variables is introduced,
written as
begin new x1 , . . ., xn ; S end,
where n 1,
procedure declarations

p(x1 , . . ., xn ) : S,
where n 0, p is the procedure identifier, x1 , . . ., xn are different variables
called formal parameters, and S is the procedure body. If n = 0, we view
this declaration as p : S.
By a program we mean a sequence of procedure declarations followed by
a (main) statement. We now explain the meaning of the programs written
in this small language. All variables used range over the set of possible
values, which are terms. A state is an answer in the sense of Subsection
1.4.2 but now viewed as a finite mapping that assigns to each variable a
value which is a term. Each variable is interpreted in the context of the
current state. The statements transform the states, i.e., the answers.
 2.2 The programming language L0 31

The meaning of the introduced statements relies on the concepts of a

choice point and failure. The choice points are created by the alternative
choice statement
begin S1 orelse . . . orelse Sn end,
so we explain its meaning first.
For n = 1, that is for the statement begin S1 end, its meaning coincides
with the meaning of S1 . For n = 2 the statement begin S1 orelse S2 end
is executed as follows. First a choice point is created which consists of the
statement S2 and the current state. Then the statement S1 is executed.
If it succeeds, the next statement after begin S1 orelse S2 end (if any)
is executed. If the execution of S1 results in a failure (a contingency to
be discussed in a moment), then backtracking takes place. This means
that the execution returns to the last created choice point from which the
execution resumes. So, assuming that the execution of S1 did not create
any choice points, the execution continues with the S2 statement executed
in the state stored at this choice point. If this execution of S2 succeeds,
the next statement after begin S1 orelse S2 end (if any) is executed. If
this execution also results in a failure, the backtracking to the previous
choice point takes place. If no such choice point is present, the execution of
begin S1 orelse S2 end ends in a failure.
Finally, we interpret the statement begin S1 orelse . . . orelse Sn end
for n > 2 inductively as

begin S1 orelse begin S2 orelse . . . orelse Sn end end.

So the alternative choice statement can be seen as a means of providing the
automatic backtracking.
Next, the skip statement is simply a non-action statement. The fail
statement causes a failure, upon which the backtracking to the most recent
choice point takes place. If no such choice point is present, the execution
ends in a failure.
In turn, the equality statement s = t is executed by trying to unify the
terms s and t. So this statement can lead to either a success or a failure.
In the first case a substitution is generated that is an mgu of s and t. This
substitution is then combined with the current state (which is an answer)
to obtain a new state. In the second case a failure results, upon which the
backtracking to the last choice point takes place. If no such choice point is
present, the execution ends in a failure.
The program composition ; has the usual meaning. Next, the block state-
ment begin new x1 , . . ., xn ; S end is executed by renaming first x1 , . . ., xn to
32 A reconstruction of pure Prolog

the variables x01 , . . ., x0n that are not present in the current state. This yields
a renaming S 0 of S. Then S 0 is executed. If it succeeds, the equations in-
volving the variables x01 , . . ., x0n are removed from the state. If it fails, then
backtracking takes place.
Finally, a procedure call p(t1 , . . ., tn ), where p is declared by

p(x1 , . . ., xn ) : S

and n > 0, is interpreted by

renaming the formal parameters x1 , . . ., xn to variables x01 , . . ., x0n that do

not occur in p(t1 , . . ., tn ); this yields a renaming S 0 of the procedure body
S,
replacing each occurrence of xi in S 0 by ti ; this yields a statement S 00 ,
executing the statement S 00 .
If n = 0, the procedure call simply results in executing the procedure body
S. So the parameter passing is taken care of by means of a substitution,
which is typical of ALGOL 60 call-by-name parameter mechanism.
A procedure call p(t1 , . . ., tn ), where no declaration of p is present, yields
a failure. Note that in addition the failure can be also caused by the execu-
tion of the fail statement or by an unsuccessful execution of the equality
statement.
The following example clarifies the above explanations. As the variables
in pure Prolog start with the upper case letter, we adopt the same conven-
tion in the language L0 . In (iii) and (v) during the program execution a
backtracking takes place.

Example 2.1 Assume that the initial state is the empty answer, i.e.,
initially no variable has a value.
(i) The statement X = f(Y); Y = a successfully terminates in the state X
= f(a), Y = a.
(ii) The statement X = a; X = b fails.
(iii) The statement begin f(X) = g(b) orelse f(X) = f(a) end success-
fully terminates in the state X = a.
(iv) The statement X = a; begin new X; X = b end successfully termi-
nates in the state X = a.
(v) Assuming the procedure declaration

p(X,Y): begin X = a; Y = b orelse X = b; Y = a end

 2.3 Translating pure Prolog into L0 33

the procedure call p(b,a) successfully terminates in the empty state, while
the procedure call p(a,a) fails.

2.3 Translating pure Prolog into L0

We now explain how to translate pure Prolog into the L0 programming
language. We illustrate it first by means of an example. Consider the
following pure Prolog program NUM defining numerals:

num(0).
num(s(X)) :- num(X).

We translate it by first rewriting it to the following equivalent pure Prolog

program that uses the built-in =/2:

num(Y) :- Y = 0.
num(Y) :- Y = s(X), num(X).

Now both clauses have the same head, num(Y). In general, we shall rewrite
first a pure Prolog program in such a way that, for each predicate p, all
clauses that define p have the same head of the form p(x1 , . . ., xn ), where
x1 , . . ., xn are different variables.
Next, we translate the bodies of the above two clauses to the following two
statements: Y = 0 and begin new X; Y = s(X); num(X) end. The latter
reflects the fact that in the second clause X is a local variable, i.e., it does
not occur in the clause head. These two statements are then put together
by means of the alternative choice statement and the original program NUM
is translated to the following procedure declaration:

num(Y): begin
Y = 0
orelse
begin new X; Y = s(X); num(X) end
end

The reader can now check that given this procedure declaration the proce-
dure call num(s(s(X))) succeeds and assigns 0 to X, while the procedure call
num(s(f(X))) fails. This coincides with the outcomes of the corresponding
Prolog computations.
Let us define now the translation for an arbitrary pure Prolog program P .
It will be translated into a sequence of procedure declarations. We perform
successively the following steps, where X1 , . . ., Xn , . . . are new variables:
34 A reconstruction of pure Prolog

Step 1: Transform each fact p(t1 , . . ., tn ) of P to the clause

p(X1 , . . ., Xn ) :- X1 = t1 , . . ., Xn = tn .
Transform each rule p(t1 , . . ., tn ) :- B of P to the clause
p(X1 , . . ., Xn ) :- X1 = t1 , . . ., Xn = tn , B.
Step 2: Translate each clause body X1 = t1 , . . ., Xn = tn , A1 , . . ., Al to the
following program:

begin new Y1 , . . ., Ym ; X1 = t1 ; . . .; Xn = tn ; A1 ; . . .; Al end,

where Y1 , . . ., Ym are the variables that occur in t1 , . . ., tn , A1 , . . ., Al .
In the boundary case when the clause body is empty (so n = 0
and l = 0), translate it to the skip statement. If no variable occurs
in t1 , . . ., tn , A1 , . . ., Al (so m = 0), translate the clause body to
X1 = t1 ; . . .; Xn = tn ; A1 ; . . .; Al .
Step 3: Let
p(X) :- B1 .
...
p(X) :- Bk .
where k 1, be the definition of the predicate p obtained in step
1 and let S1 , . . ., Sk be the corresponding translations of the clause
bodies B1 , . . ., Bk into the program statements defined in step 2.
Replace the above definition of p by the following procedure dec-
laration:
p(X) : begin S1 orelse . . . orelse Sk end.
Step 4: For each predicate q that appears in the original pure Prolog pro-
gram P but not in a head of a clause in P (that is, the definition of
q is empty) add the following procedure declaration:

q(X) : fail.
To clarify the boundary cases of the above translation consider the fol-
lowing pure Prolog program:
p.
p :- r.
r :- q.
u.
 2.4 Pure Prolog and declarative programming 35

It translates to the following sequence of four procedure declarations:

p: begin skip orelse r end;

r: begin q end;
q: fail;
u: begin skip end;

2.4 Pure Prolog and declarative programming

As already stated in the previous chapter, there are two natural interpreta-
tions of a pure Prolog program. The first one, called a declarative interpre-
tation, is concerned with the question what is being computed. According to
this interpretation a pure Prolog program is simply viewed as a set of formu-
las. The second interpretation, called a procedural interpretation, explains
how the computation takes place. Informally, we can say that the declara-
tive interpretation is concerned with the meaning, whereas the procedural
interpretation is concerned with the method .
These two interpretations of pure Prolog programs are closely related to
each other by means of specific theorems that we omit here. The first inter-
pretation helps us to better understand the second and explains why pure
Prolog supports declarative programming . Loosely speaking, declarative
programming can be described as follows. Programs can be interpreted in
a natural way as formulas in some logic. Then the results of the program
computations follow logically from the program.
Now thinking declaratively is in general much easier than thinking pro-
cedurally. For example, in the case of pure Prolog programs, to understand
the declarative interpretation, we only need to know the semantics of the
first-order logic. In contrast, to understand the procedural interpretation we
need to be familiar with the notions of a most general unifier, substitutions
and their applications, and answers.
So programs written in a formalism supporting declarative programming
are often simpler to understand and to develop. In fact, in some situations
the specification of a problem in an appropriate format already forms the al-
gorithmic solution to the problem. In other words, declarative programming
makes it possible to write executable specifications.
It should be added however, that in practice the programs obtained in
this way are often inefficient. Moreover, the declarative interpretation does
not address the issue of termination. In summary, to write efficient Prolog
programs, a good understanding of the procedural interpretation is needed.
36 A reconstruction of pure Prolog

2.5 L0 and declarative programming

To better understand the idea of declarative programming we now show
that the programming language L0 supports declarative programming, as
well. So we explain how the programs written in it can be interpreted as
first-order formulas. To this end we successively interpret

the skip statement as true,

the fail statement as false,
each procedure call statement as an atomic formula,
the program composition ; as the conjunction ,
the alternative choice statement

begin S1 orelse . . . orelse Sn end

as the disjunction S1 . . . Sn ,
the block statement

begin new x1 , . . ., xn ; S end

as the existentially quantified formula x1 . . .xn . S,

the procedure declaration

p(x1 , . . ., xn ) : S

as the universally quantified formula x1 . . .xn . (S p(x1 , . . ., xn )).

In general, each program in L0 corresponds to a first-order formula that
can be viewed as its declarative interpretation. Now, consider an L0 pro-
gram P that consists of the main statement S and a sequence Proc of the
procedure declarations. Let S be the first-order formula corresponding to S
and Proc the set of first-order formulas corresponding to Proc. Suppose that
S succeeds and yields the state (i.e., the answer) E. One can prove that

E S then logically follows from Proc. In other words, the programming
language L0 supports declarative programming.
Note that if we added to L0 the customary assignment statement, then L0
would no longer support declarative programming, since there is no natural
way to interpret the assignment as a first-order formula.
It is instructive to compare the declarative interpretations of pure Prolog
and L0 as they differ. To clarify matters let us return to the pure Prolog
program NUM already discussed:

num(0).
num(s(X)) :- num(X).
 2.6 Summary 37

Its declarative interpretation is the set (or equivalently conjunction) of

the formulas
{num(0), X. (num(s(X)) num(X))}.
Next, consider the procedure declaration
num(Y): begin
Y = 0
orelse
begin new X; Y = s(X); num(X) end
end
obtained earlier by translating the above program NUM into L0 . It is inter-
preted as the following first-order formula:
Y. ((Y = 0 X. (Y = s(X) num(X))) num(Y )).
So these two interpretations differ. However, it can be shown (see the ref-
erence in Chapter 14.9) that they imply logically the same atomic formulas.
So there is no difference as far as the analysis of successful queries is con-
cerned. In some sense the declarative interpretation of L0 is more explicit
than that of pure Prolog.

2.6 Summary
In this chapter we defined the meaning of pure Prolog programs by trans-
lating them into a small programming language L0 . This translation re-
veals the programming features implicitly present in the logic programming
paradigm. In particular, it clarifies the following three points made in the
previous chapter:
computing takes place over the set of all terms,
values are assigned to variables by means of the most general unifiers,
the control is provided by a single mechanism: automatic backtracking .
Additionally, we note that
the predicates correspond to the procedure identifiers,
the variables present in the body of a clause but not in its head correspond
to local variables,
the definition of a predicate corresponds to a procedure declaration,
the comma , separating the atomic goals in a query or in a clause body
corresponds to the program composition ;.
38 A reconstruction of pure Prolog

So this translation shows that pure Prolog is a very succinct yet highly
expressive formalism.
Additionally, we clarified what it means that logic programming and pure
Prolog support declarative programming. Finally we showed that the lan-
guage L0 supports declarative programming, as well.

2.7 Exercises

Exercise 2.1 Assume the declaration of the procedure num introduced in

Section 2.3. What is the result of executing the statement
num(s(Y)); num(X); Y = X?

Exercise 2.2 Assume the procedure declaration

p(X,Y): begin
begin new X; X = b; Y = b end
orelse
X = b; Y = a
end
What is the outcome of executing the following procedure calls:
(i) p(a,a),
(ii) p(a,b),
(iii) p(b,a).

Exercise 2.3 Translate to the language L0 the program APPEND from Figure
1.5 on page 24.
 Part II

Elements of Prolog
 3

Arithmetic in Prolog

3.1 Introduction 41
3.2 Arithmetic expressions and their evaluation 42
3.3 Arithmetic comparison predicates 47
3.4 Passive versus active constraints 51
3.5 Operators 52
3.6 Summary 55
3.7 Exercises 57

3.1 Introduction

I N THIS CHAPTER we discuss Prolog support for arithmetic. In our

presentation we focus on the essentials. In particular ECLi PSe has
64 built-ins that support arithmetic and we discuss only sixteen. The
handling of arithmetic in Prolog boils down to two additions to pure Prolog.
First, a mechanism for an evaluation of ground arithmetic expressions is
offered. Second, a facility for a comparison of ground arithmetic expressions
is provided. The latter is the first step towards integrating constraints into
the logic programming paradigm. We shall discuss this matter in Section
3.4.
The handling of arithmetic in Prolog leads to a deviation from the com-
putational model based on logic programming in that it introduces the pos-
sibility of run-time errors. This compromises the declarative character of
the language. The problem is that, due to lack of types in Prolog, the facili-
ties that support arithmetic can be used with arbitrary terms, in particular
non-numeric ones.
On the other hand this lack of types turns out to be an advantage in the
case of meta-programming that will be discussed in the next chapter.1
1 Meta-programming usually stands for the possibility of writing programs that use other pro-
grams as data.

41
42 Arithmetic in Prolog

3.2 Arithmetic expressions and their evaluation

We begin the exposition with the arithmetic. In Prolog we have at our
disposal infinitely many integer constants:
0, 1, 1, 2, 2, . . .
and infinitely many floating point numbers, such as
0.0, 1.72, 3.141, 2.0, . . .
Floating point numbers are the computer representable reals and the num-
ber of digits allowed after the obligatory decimal point . depends on the
implementation. Of course, in each implementation of Prolog only (very
large) finite subsets of integer constants and floating point numbers are sup-
ported.
In what follows, by a number we mean either an integer constant or a
floating point number.
In addition, Prolog provides various operations on the numbers. In the
case of Prolog and ECLi PSe they include the following binary operations:
addition, written as +,
subtraction, written as -,
multiplication, written as *,
integer division, written as //, (rounding towards zero) or div (rounding
towards negative infinity),
remainder of the integer division, written as rem (rounding towards zero),
or mod (rounding towards negative infinity),2
exponentiation, written as ^,
and the following unary operations:
unary minus, written as -,
absolute value, written as abs.
We call the above operations arithmetic operators.
According to the usual notational convention of logic programming and
Prolog, function symbols and predicates are written in a prefix form, that
is in front of the arguments. In contrast, in accordance with the usage in
arithmetic, the binary arithmetic operators are written in the infix form,
that is between the arguments, while the unary arithmetic operators are
written in the prefix form. Moreover, unary minus of a number can be writ-
ten in the bracketless prefix form, that is without brackets surrounding
its argument.
2 x rem y is defined as x - y*(x//y).
 3.2 Arithmetic expressions and their evaluation 43

Recall that the integer division is defined as the integer part of the usual
division outcome and given two integers x and y such that y 6= 0, x mod
y is defined as x - y*(x//y). The use of the infix and bracketless prefix
form for arithmetic operators leads to well-known ambiguities. For example,
4+3*5 could be interpreted either as (4+3)*5 or 4+(3*5) and -3+4 could
be interpreted either as (-3)+4 or -(3+4). Further, 12//4//3 could be
interpreted either as (12//4)//3 or 12//(4//3), etc.
Such ambiguities are resolved in Prolog by stipulating the following bind-
ing order (written as ):
+, -(binary), -(unary) *, // div, rem, mod ^
and assuming that the binary arithmetic operators are left associative. That
is to say, + binds weaker than *, so 4+3*5 is interpreted as 4+(3*5).
In turn, 12//4*3 is interpreted as (12//4)*3, 12//4//3 is interpreted as
(12//4)//3, etc.
The arithmetic operators and the above set of numbers uniquely deter-
mine a set of terms. We call terms defined in this language arithmetic
expressions and introduce the abbreviation gae for ground arithmetic
expressions, i.e., arithmetic expressions without variables.
As in every programming language we would like to evaluate the ground
arithmetic expressions. For example, we would like to compute 3+4 in Pro-
log. In customary programming languages this is done by means of the
assignment statement. In Prolog this is done by means of a limited coun-
terpart of the assignment, the built-in is/2, written in the infix form. For
example, we have
[eclipse 1]: X is 3+4.

X = 7
Yes (0.00s cpu)
We call is/2 the arithmetic evaluator . The above example represents
the most common use of is/2. In general, s is t is executed by unifying
s with the value of the ground arithmetic expression t. If t is not a gae, a
run-time error arises. More precisely, the following possibilities arise.
t is a gae.
Let val(t) be the value of t.
s is a number identical to val(t).
Then the arithmetic evaluator succeeds and the empty substitution is
produced. For example,
44 Arithmetic in Prolog

[eclipse 2]: 7 is 3+4.

Yes (0.00s cpu)

s is a number different from val(t).
Then the arithmetic evaluator fails. For example,
[eclipse 3]: 8 is 3+4.

No (0.00s cpu)
s is a variable.
Then the arithmetic evaluator succeeds and the substitution s = val(t)
is produced. For example,
[eclipse 4]: X is 4+3*5.

X = 19
Yes (0.00s cpu)
s is not a number and not a variable.
Then a run-time error arises. For example,
[eclipse 5]: 3+4 is 3+4.
type error in +(3, 4, 3 + 4)
Abort
t is not a gae.
Then a run-time error arises. For example,
[eclipse 6]: X is Y+1.
instantiation fault in +(Y, 1, X)
Abort
[eclipse 7]: X is X+1.
instantiation fault in +(X, 1, X)
Abort
So the is/2 built-in significantly differs from the assignment statement.
In particular, as the last example shows, we cannot increment a variable
using is/2. Because of this, typical uses of counting in Prolog are rather
subtle and confusing for the beginners. Consider for instance the following
trivial problem:

Problem Write a Prolog program that computes the length of a list.

The correct solution is presented in Figure 3.1.3

3 length/2 is actually a built-in so we denote the defined predicate by len/2. The length/2 built-
 3.2 Arithmetic expressions and their evaluation 45

% len(List, N) :- N is the length of the list List.

len([], 0).
len([_ | Ts], N) :- len(Ts, M), N is M+1.

Fig. 3.1 The LENGTH program

The program employs recursion to perform the iteration. In Chapter 7

we shall see an alternative solution that involves iteration constructs that
are present in ECLi PSe . Note the use of the local variable M to compute
the intermediate result being the length of the list Ts. We have then for
example:

[eclipse 8]: len([a,b,c], X).

X = 3
Yes (0.00s cpu)

A naive solution in which instead of the second clause we use

len([_ | Ts], N+1) :- len(Ts, N).

generates unexpected results:

[eclipse 9]: len([a,b,c], X).

X = 0 + 1 + 1 + 1
Yes (0.00s cpu)

The problem is that the gaes can be evaluated only using the is statement.
To compute the value of a gae defined inductively one needs to generate
new local variables (like M in the LENGTH program) to store in them the
intermediate results, and then use the is statement.
To save explicitly introducing such extra variables, ECLi PSe supports for
is/2 a functional syntax. Let us suppose p/2 is a predicate whose last
argument is numeric, viz.

p(a, 1).
p(b, 2).

in can be used both to compute the length of a list, and to build a list of different variables of
a specified length.
46 Arithmetic in Prolog

Then p/1 can be used as a function when it occurs on the right-hand side
of is, for example:
[eclipse 10]: X is p(a)+1.

X = 2
Yes (0.00s cpu)
In particular, both the len/2 predicate from the LENGTH program and the
length/2 built-in can be used as a function, for example:
[eclipse 11]: X is length([a,b,c]).

X = 3
Yes (0.00s cpu)
As another example of the use of an arithmetic evaluator consider a pred-
icate defining all the integer values in a given range. The appropriate defi-
nition is given in Figure 3.2.

% select_val(Min, Max, Val) :-

% Min, Max are gaes and Val is an integer
% between Min and Max inclusive.

select_val(Min, Max, Val) :- Min =< Max, Val is Min.

select_val(Min, Max, Val) :-
Min < Max, Min1 is Min+1,
select_val(Min1, Max, Val).

Fig. 3.2 The SELECT program

Note the use of a local variable Min1 in the arithmetic evaluator Min1 is
Min+1 to compute the increment of Min. In this program the appropriate
values are generated only upon demand, by means of backtracking. We
shall see the usefulness of such a process in Section 7.6. We have for example:
[eclipse 12]: select_val(10, 13, Val).

Val = 10
Yes (0.00s cpu, solution 1, maybe more) ? ;

Val = 11
 3.3 Arithmetic comparison predicates 47

Yes (0.00s cpu, solution 2, maybe more) ? ;

Val = 12
Yes (0.00s cpu, solution 3, maybe more) ? ;

Val = 13
Yes (0.00s cpu, solution 4, maybe more) ? ;

No (0.00s cpu)

Note that, unlike in the customary programming languages, using this

definition we can recognise as well as generate values. For example we have

[eclipse 13]: select_val(10, 13, 9).

No (0.00s cpu)
[eclipse 14]: select_val(10, 13, X), select_val(12, 15, X).

X = 12
Yes (0.00s cpu, solution 1, maybe more) ? ;

X = 13
Yes (0.00s cpu, solution 2, maybe more) ? ;

No (0.01s cpu)

3.3 Arithmetic comparison predicates

Prolog allows us not only to compute the values of gaes but also to compare
them by means of the following six arithmetic comparison predicates
(in short comparison predicates):

less than, written as <,

less than or equal, written as =<,
equality, written as =:=,
disequality, written as =\=,
greater than or equal, written as >=,
greater than, written as >.

The equality predicate =:=/2 between gaes should not be confused with
the is unifiable with predicate =/2 discussed in Section 1.4.
48 Arithmetic in Prolog

The comparison predicates work on gaes and produce the expected out-
come. So for instance >/2 compares the values of two gaes and succeeds if
the value of the first argument is larger than the value of the second and
fails otherwise. Thus, for example
[eclipse 15]: 5*2 > 3+4.

Yes (0.00s cpu)

[eclipse 16]: 7 > 3+4.

No (0.00s cpu)
However, when one of the arguments of the comparison predicates is not
a gae, a run-time error arises. In particular, =:=/2 cannot be used to assign
a value to a variable. For example, we have
[eclipse 17]: X =:= 5.
instantiation fault in X =:= 5
Abort
As a simple example of the use of the comparison predicates consider the
program in Figure 3.3 that checks whether a list is an -ordered list of
numbers.

% ordered(Xs) :- Xs is an =<-ordered list of numbers.

ordered([]).
ordered([H | Ts]) :- ordered(H, Ts).

% ordered(H, Ts) :- [H|Ts] is an =<-ordered list of numbers.

ordered(_, []).
ordered(H, [Y | Ts]) :- H =< Y, ordered(Y, Ts).

Fig. 3.3 The ORDERED program

We now have
[eclipse 18]: ordered([1,1.5,2,3]).

Yes (0.00s cpu)

but also
 3.3 Arithmetic comparison predicates 49

[eclipse 19]: ordered([1,X,3]).

instantiation fault in 1 =< X
Abort

Here a run-time error took place because during the computation the com-
parison predicate =</2 was applied to an argument that is not a number.
On the other hand, we also have

[eclipse 20]: ordered([a]).

Yes (0.00s cpu)

which shows that programs that use the arithmetic comparison predicates
can be quite subtle. In this case, the ORDERED program correctly rejects (by
means of a run-time error) some queries, such as ordered([1,X,3]) but
incorrectly accepts some other queries, such as ordered([X]).
A possible remedy is to use the number/1 built-in of Prolog that tests
whether its argument is a number and rewrite the program as follows.

ordered([]).
ordered([H | Ts]) :-
number(H),
ordered(H, Ts).

ordered(_, []).
ordered(H, [Y | Ts]) :-
number(Y),
H =< Y, ordered(Y, Ts).

However, we get then the uninformative No answer to such queries as

ordered([X]) or ordered([1,X,3]). We shall explain in the next chapter
how more informative answers can be generated.
As another example of a Prolog program that employs the comparison
predicates consider the Prolog version of the well-known quicksort proce-
dure proposed in Hoare [1962]. According to this sorting procedure, a list is
first partitioned into two sublists using an element X of it, one consisting of
elements smaller than X and the other consisting of elements larger than or
equal to X. Then each sublist is quicksorted and the resulting sorted sublists
are appended with the element X put in the middle. This can be expressed
in Prolog by means of the program in Figure 3.4, where X is chosen to be
the first element of the given list.
50 Arithmetic in Prolog

% qs(Xs, Ys) :-
% Ys is an =<-ordered permutation of the list Xs.

qs([], []).
qs([X | Xs], Ys) :-
part(X, Xs, Littles, Bigs),
qs(Littles, Ls),
qs(Bigs, Bs),
app(Ls, [X | Bs], Ys).

% part(X, Xs, Ls, Bs) :-

% Ls is a list of elements of Xs which are < X,
% Bs is a list of elements of Xs which are >= X.

part(_, [], [], []).

part(X, [Y | Xs], [Y | Ls], Bs) :-
X > Y,
part(X, Xs, Ls, Bs).
part(X, [Y | Xs], Ls, [Y | Bs]) :-
X =< Y,
part(X, Xs, Ls, Bs).
augmented by the APPEND program of Figure 1.5.

Fig. 3.4 The QUICKSORT program

In the part/4 predicate the first two arguments are used as inputs and
the last two as outputs. Note that the list constructors are used both in the
input positions and in the output positions.
We then have

[eclipse 21]: qs([7,9,8,1,5], Ys).

Ys = [1, 5, 7, 8, 9]
More (0.00s cpu) ? ;

No (0.00s cpu)

However, we also have

[eclipse 22]: qs([3,X,0,1], Ys).

 3.4 Passive versus active constraints 51

instantiation fault in 3 > X

Abort

because during the computation the comparison predicate >/2 is applied to

non-gae arguments.
In the next chapter we shall return to this program to discuss a typical
source of inefficiency in Prolog programs and a way of avoiding it.

3.4 Passive versus active constraints

Atomic goals that employ arithmetic comparison predicates, for example

3*X < Y+2, go beyond the syntax of pure Prolog. Like the goals that refer
to the =/2 built-in, for example f(a,X) = f(Y,b), they impose a restriction
on their variables. The important difference is that at present the first type
of goals can be successfully processed only when all its variables are ground.
Both atomic goals can be viewed as constraints in the sense that they refer
to some predefined predicate (here < and =), and hence restrict the set of
values the variables can be assigned to. (A formal account of constraints will
be given in Chapter 6.) We call the atomic goals that use the comparison
predicates arithmetic constraints.
In what follows we distinguish between active constraints and passive
constraints. This distinction refers to the way the constraints are used
during the computation process. The evaluation of the former ones can affect
the variables present in them, while the evaluation of the latter ones cannot.
So f(a,X) = f(Y,b) is an active constraint, while at present 3*X < Y*Y+2
is a passive constraint.
In other words at this stage we can use arithmetic constraints only for
testing. Moreover, these constraints can cause a run-time error, when their
variables are not ground. In particular, the query ordered([1,X,3]) men-
tioned on page 49 yields a run-time error. The query ordered([X,2,1])
also yields a run-time error, even though it would fail for any value of X.
Later, in Chapter 9 we shall discuss an ECLi PSe library that allows us
to use arithmetic constraints as passive constraints without causing a run-
time error, so the last query can correctly fail instead of causing an er-
ror. Still later, in Chapter 10, we shall discuss an ECLi PSe library that al-
lows us to express arithmetic constraints as active constraints, so the query
ordered([2,X,2]) can succeed, instantiating X to 2.
52 Arithmetic in Prolog

3.5 Operators
Arithmetic operators are examples of function symbols that are written in
the infix form (for example +) or bracketless prefix form (the unary -).
These forms are equivalent to the standard, prefix, form. For example in
Prolog and ECLi PSe the expressions +(1,2) and 1+2 are indistinguishable:

[eclipse 23]: 1+2 = +(1,2).

Yes (0.00s cpu)

Prolog provides a means to declare a symbol so that it can be used in a

form alternative to the standard one. Function symbols and predicates that
are declared this way are called operators.

3.5.1 Binary operators: infix notation

Let us take as an example the predicate before/2. We wish to be able to
write time1 before time2 instead of before(time1, time2). To allow it
we use an operator declaration, which is written as follows:

:- op(700, xfx, before).

This declares before/2 as an infix binary operator. The first two arguments,
700 and xfx, will be explained later.
We can now write 10 before 11 anywhere we could previously write
before(10, 11). For example we could write the following program:

10 before 11.
9 before 10.
X before Z :- X before Y, Y before Z.
and ask

[eclipse 24]: 9 before X.

X = 10
Yes (0.00s cpu, solution 1, maybe more) ? ;

X = 11
Yes (0.00s cpu, solution 2, maybe more) ?
% Asking for more here would start an infinite loop!

This query can still be written in the standard syntax, as before(9, X).
 3.5 Operators 53

3.5.2 Binary operators: binding order

The drawback of using infix operators is the resulting ambiguity. To see
this suppose we also want to write the time using an infix operator, thus
10:00. When we write 10:00 before 11:00 we intend this to be the same
as before(:(10,00), :(11,00)). However, this expression could also be
parsed in a different way that is equivalent to the expression

:(10, :(before(00,11), 00)).

One way to guide the parser is to introduce brackets, thus

(10:00) before (11:00).

Now there is no ambiguity. The last expression is parsed then as

before(:(10,00), :(11,00)).

To achieve the other form we would have to bracket the original expression
thus:
10:((00 before 11):00).

This second bracketing specifies that ((00 before 11):00) is a single com-
pound term. This term has the function symbol :/2 and two arguments,
(00 before 11) and 00.
If such brackets were always required, however, then it would make arith-
metic expressions quite unreadable. For example, we would have to write
(2*3)-(4//((3*Y)+(2*Z))) instead of 2 * 3 - 4 // (3 * Y + 2 * Z).
In order to guide the parser without the need for extra brackets, we asso-
ciate a priority with each operator. In the declaration

:- op(700, xfx, before).

the priority is 700. In Prolog :/2 is already predeclared as an operator

with the priority 600. A lower priority number indicates a tighter bind-
ing, so :/2 binds tighter than before/2. Consequently, in the context
T1:T2 before T3 the parser implicitly inserts the bracketing resulting in
(T1:T2) before T3.
Thus we can safely write a more natural version of our program as follows:

10:00 before 11:00.

9:00 before 10:00.
X before Z :- X before Y, Y before Z.
54 Arithmetic in Prolog

3.5.3 Binary operators: associativity

Operator priorities can be used to disambiguate terms involving different
operators, but not multiple occurrences of the same operator. Consider the
term 1+2+3+4. Since +/2 is associative this could be written equivalently as
((1+2)+3)+4 or 1+((2+3)+4) or in several other ways. The second argu-
ment of the operator declaration
:- op(700, xfx, before).
is used to handle terms like this. In fact xfx is the strictest form of infix
binary operator and it does not allow such ambiguity. We shall return to
this form in a moment.
To allow terms such as 1+2+3+4 the forms yfx and xfy are provided.
They denote left associativity and right associativity. The operator +/2 is
predeclared in Prolog as follows:
:- op(500, yfx, +).
With this declaration the term 1+2+3+4 is successfully parsed and is equiv-
alent to the term ((1+2)+3)+4. The binary - is also declared as left asso-
ciative:4
:- op(500, yfx, -).
We can also declare a right associative operator
:- op(500, xfy, ++).
and then the term 1++2++3++4 is parsed as 1++(2++(3++4)).
In turn, of/2 is declared as a non-associative operator:
:- op(650, xfx, of).
This rules out expressions such as a of b of c.
Similarly, on account of the declaration
:- op(700, xfx, before).
the expression t1 before t2 before t3 is rejected by the parser.
ECLi PSe also allows prefix binary operators to be declared using the forms
fxx and fxy. For example, given the declaration
:- op(800, fxx, bin).
4 This is regrettable since ambiguous terms such as 1-2+3 are then accepted, and as explained
below are parsed as (1-2)+3. It would be better to use the declaration
:-op(500, xfx, -).
and require such terms as 1-2+3 to be explicitly bracketed by the user.
 3.6 Summary 55

ECLi PSe will successfully parse the term bin 2 2 as bin(2,2).

3.5.4 Unary operators

In addition to binary operators, Prolog also admits unary prefix or postfix
operators. For example the unary - is predeclared as a prefix operator, so
we can write -1 + 2. The operator declaration for this prefix unary operator
-/1 is:
:- op(500, fx, -).
Because the priority of before is 700, the operator -/1 binds tighter than
before. Consequently the expression
- T1 before T2
is parsed like
(- T1) before T2.
In addition to the form fx also the forms xf, fy and yf are allowed.
Together with the forms for binary operators they allow for ambiguous sit-
uations with operators with the same priority. In the specification of an
operator (e.g. xf, yfx, xfy), x represents an argument whose priority
must be lower than that of the operator and y represents an argument
whose priority must be lower than or equal to that of the operator.
For example, +/2 has the same priority as -/1. Because the declaration
of -/1 specifies fx and not fy, the term following - cannot have the form
X+Y. Consequently
- X + Y
is parsed like
(- X) + Y.
Moreover, the form fx of -/1 ensures that the expression
- - Y
is illegal.

3.6 Summary
In this chapter we discussed Prolog facilities for arithmetic. They consist of
the is/2 built-in that allows us to evaluate ground arithmetic expressions,
56 Arithmetic in Prolog

the customary six comparison operators, written as

<, =<, =:=, =\=, >=, >,
the number/1 built-in that tests whether its argument is a number,
the possibility of declaring a function symbol or a predicate in an infix
or postfix form by means of an operator declaration. These declarations
involve the following forms for the binary function symbols:

xfx (no associativity),

xfy (right associativity),
yfx (left associativity),
fxx,
fxy,
and for the unary function symbols:
fx,
fy,
xf,
yf.
The arithmetic operators mentioned in Section 3.2 are predeclared as

:- op(500, yfx, [+, -]).

:- op(500, fx, -).
:- op(400, yfx, [*, //, div, rem, mod]).
:- op(200, xfy, ^).

This ensures the binding order and the associativity information stated
in Section 3.2.
In turn, the comparison predicates introduced in Section 3.3 are prede-
clared as

:- op(700, xfx, [<, =<, =:=, =\=, >=, >]).

When discussing the comparison predicates we introduced the distinction

between passive and active constraints to which we shall return in later
chapters.
Finally, we would like to point out that the possibility of declaring a
function symbol or a predicate in an infix or postfix form allows us to
view programs as data and consequently to realise meta-programming in
a straightforward way. We shall explain this in detail in the next chapter.
 3.7 Exercises 57

3.7 Exercises

Exercise 3.1 What is the ECLi PSe output to the following queries?
1. X is 7, X is 6.

2. X is 7, X is X+1.

3. X is X+1, X is 7.

4. X is 3, Y is X+1.

5. Y is X+1, X is 3.

Exercise 3.2 Consider the following program:

q(X) :- Z is 1.

inc(X) :- Z is X+1.
What is the ECLi PSe output to the following queries?
1. Z is 2, q(X).

2. Z is 2, q(Z).

3. Z is 2, q(X), Y is Z+1.

4. Z = 2, inc(X), X = 2.

5. X = 2, inc(X), Z = 2.

Exercise 3.3 Consider the following two programs:

% min(X, Y, Z) :- Z is the minimum of the gaes X and Y.

min(X, Y, X) :- X < Y.
min(X, Y, Y) :- X >= Y.
and
% min(X, Y, Z) :- Z is the minimum of the gaes X and Y.

min(X, Y, Z) :- X < Y, Z is X.
58 Arithmetic in Prolog

min(X, Y, Z) :- X >= Y, Z is Y.
Find a query for which they yield different answers.

Exercise 3.4 Assume the following declarations:

:- op(500, yfx, have).
:- op(400, yfx, and).
What are the interpretations of the following terms (that is, what is their
representation in the customary first-order logic syntax):
1. she and he have luck,
2. they have time and money and luck.

Exercise 3.5 Redefine the len/2 predicate from Figure 3.1 on page 45 as
an infix operator has length/2, using an appropriate operator declaration.
 4

Control and meta-programming

4.1 More on Prolog syntax 59

4.2 Control 62
4.3 Meta-programming 69
4.4 Summary 74
4.5 Exercises 75

S O FAR, IT looks that Prolog syntax is just a syntactic sugar of the

first-order logic syntax. This is, however, not the case and Prologs
novel use of syntax turns it into a powerful language. This is achieved
by a combination of three Prolog features: operators, ambivalent syntax and
meta-variables. We continue our presentation with the latter two.

4.1 More on Prolog syntax

4.1.1 Ambivalent syntax
The power of the operators in Prolog becomes apparent only when they are
combined with two other syntactic features of Prolog, ambivalent syntax
and meta-variables.
To explain the first one let us recall that in the first-order logic one as-
sumes that the sets of used function symbols and predicates are disjoint. In
Prolog this assumption is relaxed and there is no difference between func-
tion symbols and predicates. A function symbol or a predicate f of arity n
is referred to as f/n and is called a functor . Once there is no difference
between function symbols and predicates the difference between terms and
atomic goals also disappears. A term or an atomic goal is called in Prolog
a structure.
Moreover, also in contrast to first-order logic, the same name can be used

59
60 Control and meta-programming

to denote function symbols or predicates of different arity. For example, it

is perfectly legal to use in a Prolog program functors p/1 and p/2 and build
syntactically legal structures such as p(p(a,b),c,p(X)). Whether such a
structure corresponds to a term or to an atomic goal depends on the context
in which it is used. For example, in the clause
p(p(a,b), c, p(X)) :- q(a,b,X).
the above structure is used as an atomic goal and in the clause
q(X) :- r(p(p(a,b),c,p(X)).
it is used as a term. In both cases we have syntactically legal Prolog clauses.
In the presence of the ambivalent syntax we need to modify slightly the
MartelliMontanari unification algorithm of Section 1.4 by using in-
stead of action (2) the following one:

(20 ) f (s1 , ..., sn ) = g(t1 , ..., tm ) halt with failure,

where f 6 g or n 6= m.

This takes care of equations such as f(s) = f(t,u) which are now syntac-
tically allowed. (Here s,t,u are structures.)
In the previous chapter we explained using the example of the before/2
predicate that binary predicates can be turned into operators. Now that we
have also introduced the ambivalent syntax only one small step is needed to
realise meta-programming in Prolog. Namely, it suffices to identify queries
and clauses with the structures. To this end we just view the comma ,
separating the atomic goals of a query as a right-associative, infix operator
and :- as an infix operator. To ensure the correct interpretation of the
resulting structures the following declarations are internally stipulated:

:- op(1000, xfy, ,).

:- op(1200, xfx, :-).
So when referring to the first operator we use quotes and write ,/2. .
To see how we can now view a clause as a structure consider for instance
the clause
qs([X | Xs], Ys) :-
part(X, Xs, Littles, Bigs),
qs(Littles, Ls),
qs(Bigs, Bs),
app(Ls, [X | Bs], Ys)
 4.1 More on Prolog syntax 61

of the QUICKSORT program. The reader can easily check that it is a structure
with the priority 1200 and interpretation

:-(qs([X | Xs], Ys),

(part(X, Xs, Littles, Bigs),
(qs(Littles, Ls),
(qs(Bigs, Bs),
app(Ls, [X | Bs], Ys)
)
)
)
).

Now that we have identified queries and clauses with structures we can
naturally view programs as lists of clauses and pass them as arguments.
However, it is much more convenient to be able to access the considered
program directly. To this end Prolog provides a built-in that we shall discuss
in Section 4.3.

4.1.2 Meta-variables
Another unusual syntactic feature of Prolog is that it permits the use of vari-
ables in the positions of atomic goals, both in the queries and in the clause
bodies. Such a use of a variable is called a meta-variable. Computation
in the presence of the meta-variables is defined as for pure Prolog programs
with the exception that the mgus employed now also bind the meta-variables.
So for example, given the QUICKSORT program and the clause

solve(Q) :- Q.

the query solve(qs([7,9,8,1,5], Ys)) resolves in one step to the already

considered (in Section 3.3) query qs([7,9,8,1,5], Ys) that succeeds and
produces the answer Ys = [1,5,7,8,9].
Prolog requires that the meta-variables become instantiated before they
are selected. Otherwise a run-time error arises. For example, the query
solve(Q) ends up in a run-time error.
Meta-variables are useful in a number of ways as we shall see in the next
two sections.
62 Control and meta-programming

4.2 Control
So far the only control mechanism we introduced is the automatic back-
tracking. Prolog provides several control statements that allow one to write
more efficient and simpler programs.

if-then-else The familiar if B then S else T statement is present in Prolog

and is written, in a somewhat awkward way, as
B -> S ; T.
It has the expected meaning: if B succeeds, then the execution proceeds with
S and if B fails, then the execution proceeds with T. There is no separate
provision for the if B then S statement. So it is written as if B then S else
true, so as
B -> S ; true.
Further, B -> S ; fail abbreviates to B -> S.
The evaluation of B can lead to the side effects if at the moment of its
evaluation it contains variables. For example we have
[eclipse 1]: X = a -> Y = a ; Y = b.

X = a
Y = a
Yes (0.00s cpu)
As a more meaningful example consider the problem of modifying the
ORDERED program of Figure 3.3 so that in the case of a wrong input an
appropriate information is printed. To this end we use the number/1 built-
in which tests whether its argument is a number and the write/1 built-in
which writes its argument (which can be a compound term, like wrong
input:H) below). This modification is presented in Figure 4.1.
We get then as desired
[eclipse 2]: ordered([1,a,0]).
wrong input : a
Yes (0.00s cpu)

Disjunction Next, we discuss two control mechanisms of Prolog inspired

by the first-order logic. Both are powerful additions to Prolog. We shall
discuss their merits later.
Disjunction is a Prologs built-in written as an infix operator ;/2. The
query Q ; R is executed in a completely analogous way as the alternative
 4.2 Control 63

% ordered(Xs) :- Xs is an =<-ordered list of numbers.

ordered([]).
ordered([H | Ts]) :-
( number(H) ->
ordered(H, Ts)
;
write(wrong input:H)
).

% ordered(H, Ts) :- [H|Ts] is an =<-ordered list of numbers.

ordered(_, []).
ordered(H, [Y | Ts]) :-
( number(Y) ->
H =< Y, ordered(Y, Ts)
;
write(wrong input:Y)
).

Fig. 4.1 Another ORDERED program

choice statement discussed in Section 2.2. So the explanation of its mean-

ing involves the already discussed there concept of a choice point. First, a
choice point is created which consists of the query Q and the current sub-
stitution (state). Then the query Q executed. If it succeeds, the computed
substitution is displayed. When a request for more solutions is issued, the
backtracking to the last created choice point takes place. The same happens
when Q fails but in that case of course no substitution is displayed. If R suc-
ceeds, the new computed substitution is displayed. If R fails a backtracking
takes place to the last choice point created before the execution of the query
Q ; R. If no such choice point is present, a failure results. So for example

[eclipse 3]: X = a ; X = b.

X = a
More (0.00s cpu) ? ;

X = b
64 Control and meta-programming

Yes (0.00s cpu)

[eclipse 4]: X = a, (X = b ; Y = c).

X = a
Y = c
Yes (0.00s cpu)

Internally the disjunction is defined in a remarkably concise way thanks

to the automatic backtracking and the presence of meta-variables. We just
need the following declaration and clauses that employ the meta-variables:

:- op(1100, xfy, ;).

X ; _ :- X.
_ ; Y :- Y.

Negation In turn, negation is a unary functor not/1 written in the brack-

etless prefix form. To use a catchy phrase, not turns failure into a success.
More precisely, the query not Q succeeds if and only if the query Q fails.
Thus we have

[eclipse 5]: not a = b.

Yes (0.00s cpu)

[eclipse 6]: not a = a.

No (0.00s cpu)

Also note that we have

[eclipse 7]: not X = 1.

No (0.00s cpu)

The reason is that the query not X = 1 fails since the query X = 1 succeeds.
However, there exists an instance of X that is different than 1, for example
2. So the use of not in the presence of variables can lead to counterintuitive
answers. From the declarative point of view it is preferable to use it only
as a test, so when the query in its scope, like a = b in not a = b, has no
variables at the moment of the execution.
We shall discuss the internal definition of not once we introduce the so-
called cut functor.
 4.2 Control 65

Once Next, Prolog provides the functor once/1. The query once Q is exe-
cuted the same way as Q is with the exception that once a solution is found,
the other solutions are discarded. So for example
[eclipse 8]: once (X = a ; X = b).

X = a
Yes (0.00s cpu)

[eclipse 9]: once(a = b).

No (0.00s cpu)
Again, we shall discuss the internal definition of once after we have in-
troduced the cut functor.

Cut The most controversial control facility of Prolog is the nullary functor
called cut and denoted by !/0. It is used to make the programs more
efficient by pruning the computation tree over which the search for solutions
takes place. Informally, the cut means commit to the current choice and
discard other solutions, though the precise meaning is more complicated.
To explain its effect, we explain first the implementation of two already
introduced built-ins using cut.
The once/1 functor is internally defined by means of the following decla-
ration and the single clause:
:- op(900, fy, once).
once Q :- Q, !.
So once Q is executed by executing Q first. If Q succeeds for the first time
the cut functor ! is executed. This has the effect that the execution
of once Q successfully terminates without any attempt to explore whether
other solutions to Q exist. If Q fails, then once Q fails, as well.
The implementation of the negation functor not/1 is equally short though
much less easy to understand:
:- op(900, fy, not).
not Q :- Q, !, fail.
not _ .
Here fail/0 is a Prolog built-in with the empty definition. This cryptic
two line program employs several discussed features of Prolog. Consider the
call not Q, where Q is a query. If Q succeeds, then the cut ! is performed.
66 Control and meta-programming

This has the effect that all alternative ways of computing Q are discarded and
also the second clause is discarded. Next, the built-in fail is executed and a
failure arises. Since the only alternative clause was just discarded, the query
not Q fails. If on the other hand the query Q fails, then the backtracking
takes place and the second clause, not , is selected. It immediately succeeds
with the empty substitution and so the initial query not Q succeeds with
the empty substitution.
After these two examples of the uses of cut let us provide its definition in
full generality. Consider the following definition of a predicate p:
p(s1 ) :- A1 .
...
p(si ) :- B,!,C.
...
p(sk ) :- Ak .
Here the ith clause contains the cut !. Suppose now that during the
execution of a query a call p(t) is encountered and eventually the ith clause
is used and the indicated occurrence of the cut is executed. Then this
occurrence of ! succeeds immediately, but additionally

(i) all alternative ways of computing B are discarded, and

(ii) all computations of p(t) using the i + 1th to kth clause for p are
discarded as backtrackable alternatives to the current selection of the
ith clause.

The cut was introduced in Prolog to improve the efficiency of the pro-
grams. As an example let us return to the QUICKSORT program defined in
Figure 3.4 of Section 3.3 and consider the execution of a typical query such as
qs([7,9,8,1,5], Ys). To see that the resulting computation is inefficient
let us focus on the definition of the part/4 predicate:

part(_, [], [], []).

part(X, [Y | Xs], [Y | Ls], Bs) :-
X > Y,
part(X, Xs, Ls, Bs).
part(X, [Y | Xs], Ls, [Y | Bs]) :-
X =< Y,
part(X, Xs, Ls, Bs).

The above query qs([7,9,8,1,5], Ys) leads to the call

part(7, [9,8,1,5], Littles, Bigs).

 4.2 Control 67

Now the execution of the second clause above fails when 7 is compared with
9 and subsequently the last, third, clause is tried. At this moment 7 is again
compared with 9. The same redundancy occurs later when 1 is compared
with 5. To avoid such inefficiencies the definition of part/4 can be rewritten
using the cut as follows:
part(_, [], [], []).
part(X, [Y | Xs], [Y | Ls], Bs) :-
X > Y, !,
part(X, Xs, Ls, Bs).
part(X, [Y | Xs], Ls, [Y | Bs]) :-
part(X, Xs, Ls, Bs).
So the cut is introduced after the test X > Y whereas the test X =< Y is
deleted. The reader can easily check that the cut has the desired effect here.
The cut can also be used to ensure that only one solution is produced.
Consider for example the MEMBER program defined in Figure 4.2.1

% mem(Element, List) :- Element is an element

% of the list List.

mem(X, [X | _]).
mem(X, [_ | Xs]) :- mem(X, Xs).

Fig. 4.2 The MEMBER program

This program can be used both to check whether an element is a member

of a list or to generate all members of a list:
[eclipse 10]: mem(wed, [jan, wed, fri]).

More (0.00s cpu) ? ;

No (0.00s cpu)

[eclipse 11]: mem(X, [jan, wed, fri]).

X = jan
More (0.00s cpu) ? ;

1 member/2 is actually a built-in so we denote the defined predicate by mem/2.

68 Control and meta-programming

X = wed
More (0.00s cpu) ? ;

X = fri
More (0.00s cpu) ? ;

No (0.00s cpu)

To ensure that only one solution is produced we modify the MEMBER pro-
gram by adding to the first clause the cut:

mem_check(X, [X | _]) :- !.
mem_check(X, [_ | Xs]) :- mem_check(X, Xs).

After this change we now have

[eclipse 12]: mem_check(wed, [jan, wed, fri]).

Yes (0.00s cpu)

[eclipse 13]: mem_check(X, [jan, wed, fri]).

X = jan
Yes (0.00s cpu)

Another use of the cut will be illustrated in the next section. In general,
the cut is a powerful mechanism inherent to Prolog. Some of its uses can
be replaced by employing other built-ins that are internally defined using
the cut. For example, instead of the just discussed revision of the MEMBER
program we could use the once/1 built-in and simply run the queries of
the form once(mem(s,t)). Further, notice that the not/1 built-in can be
defined in a much more intuitive way using the if-then-else statement as
follows:

not Q :- (Q -> fail ; true).

The above clause simply formalises the intuitive meaning of negation, namely
that it reverses failure and success. What is happening here is that the cut is
absorbed in the if-then-else statement, which itself can be defined using
the cut as follows:

if_then_else(B, Q, R) :- B,!,Q.
if_then_else(B, Q, R) :- R.
 4.3 Meta-programming 69

On the other hand, several natural uses of the cut can be modelled by
means of other built-ins only at the cost of, sometimes considerable, rewrit-
ing of the program that can affect the readability. For example, in the case
of the part/4 predicate we can use the if-then-else operator and rewrite
the second clause as
part(X, [Y | Xs], Ls, Bs) :-
( X > Y ->
Ls = [Y | L1s], part(X, Xs, L1s, Bs)
;
Bs = [Y | B1s], part(X, Xs, Ls, B1s)
).
The effect is the same as when using the cut but we had to introduce the
explicit calls to the unification functor =/2.
Non-trivial complications arise during the interaction of the cut with other
Prolog built-ins. In some situations cut is not allowed. For example, no
cuts are allowed in the query B in the context of the if-then-else call
B -> S ; T. Moreover, once the cut is used, the clause order becomes of
crucial importance. For example, if we ignore the efficiency considerations,
in the original QUICKSORT program the order of the clauses defining the
part/4 predicate was immaterial. This is no longer the case in the revised
definition of part/4 that uses the cut. In short, the cut has to be used with
considerable care.

4.3 Meta-programming
4.3.1 Accessing programs directly
We explained in Section 4.1 how Prolog programs can be conveniently passed
as data, in the form of a list of clauses. But it is much more convenient to be
able to access the considered program directly. To this end Prolog provides
the clause/2 built-in.
In its call the first argument has to be a term that is neither a variable nor
a number. This makes it possible to determine the functor to which the call
refers to. So for example each of the calls clause(X, Y) and clause(100,
Y) yields a run-time error, while the call clause(p(X), Y) may succeed or
fail.
Given a call clause(head, body), first the term head is unified with
a head of a clause present in the considered program. If no such clause
exists, the call clause(head, body) fails. Otherwise, the first such clause
is picked and the term head :- body is unified with this clause. (It is
70 Control and meta-programming

assumed that true is the body of a unit clause. (true/0 is a Prolog built-in
that immediately succeeds.) Upon backtracking successive choices for head
are considered and the corresponding alternative solutions are generated.
As an example of the use of clause/2 assume that the file with the
QUICKSORT program of Figure 3.4 from page 50 is read in. Then we have
the following interaction:2

[eclipse 14]: clause(qs(U, V), Z).

U = []
V = []
Z = true
Yes (0.00s cpu, solution 1, maybe more) ? ;

U = [X|Xs]
V = V
Z = part(X, Xs, Littles, Bigs), qs(Littles, Ls),
qs(Bigs, Bs), app(Ls, [X|Bs], V)
Yes (0.00s cpu, solution 2)

which shows how easily we can dissect the definition of the qs/2 predicate
into parts that can be further analysed. Note three references to the ,/2
operator in the structure assigned to Z.
To be able to access the definition of a functor p/n by means of the
clause/2 built-in, p/n has to be declared as a dynamic predicate. For
example, the above interaction with the QUICKSORT program requires the
presence of the declaration
:- dynamic qs/2.

4.3.2 Meta-interpreters
Using the clause/2 built-in it is remarkably easy to write meta-interpreters,
so programs that execute other programs. As an example consider the prob-
lem of writing in Prolog an interpreter for pure Prolog. The program is
presented in Figure 4.3. It is strikingly concise and intuitive.
The first clause states that the built-in true/0 succeeds immediately. The
second clause states that a query of the form A,B succeeds if both A and B
2 In ECLi PSe the values of all variables are displayed. This explains the superfluous equation V
= V. In particular, the empty substitution is displayed as a sequence of such equalities for each
variable present in the query.
 4.3 Meta-programming 71

% solve(X) :-
% The query X succeeds for the pure Prolog
% program accessible by clause/2.

solve(true) :- !.
solve((A,B)) :- !, solve(A), solve(B).
solve(H) :- clause(H, Body), solve(Body).

Fig. 4.3 The META INTERPRETER program

succeed. Here A,B is a structure the functor of which is ,/2. Recall from
page 60 that ,/2 was declared by

:- op(1000, xfy, ,).

So ,/2 is right associative and in A,B the first argument, A, denotes an

atomic goal and the second argument, B, a sequence of the remaining goals.
The additional pair of brackets around A,B is neccessary to ensure that A,B
is passed as one argument to solve.
Finally, the last clause states that an atomic query H succeeds if there ex-
ists a clause of the form H :- Body such that the query Body succeeds. The
cuts are used here to enforce the distinction by cases: either the argument
of solve is true or else a non-atomic query or else an atomic one.
To illustrate the behaviour of the META INTERPRETER program assume
that APPEND is a part of the considered program. We then have

[eclipse 15]: solve((app(Xs, [_, _|Ys], [mon, wed, fri, sun]),

app(Xs, Ys, Zs))).

Xs = []
Ys = [fri, sun]
Zs = [fri, sun]
More (0.00s cpu) ? ;

Xs = [mon]
Ys = [sun]
Zs = [mon, sun]
More (0.00s cpu) ? ;

Xs = [mon, wed]
72 Control and meta-programming

Ys = []
Zs = [mon, wed]
More (0.00s cpu) ? ;

No (0.00s cpu)

So we get the same results as when executing the query

app(Xs, [_, _|Ys], [mon, wed, fri, sun]), app(Xs, Ys, Zs)

directly.
The META INTERPRETER program forms a basis for several non-trivial meta-
programs written in Prolog, including debuggers, tracers, interpreters and
partial evaluators. As an example of a simple extension of this program con-
sider the problem of enhancing the answers to the queries with the length of
the derivation. The appropriate modification of solve/1 is straightforward
and is presented in Figure 4.4.

% solve(X, N) :-
% The query X succeeds for the pure Prolog
% program accessible by clause/2 in N steps.

solve(true, 0) :- !.
solve((A, B), N) :- !, solve(A, N1), solve(B, N2),
N is N1+N2.
solve(H, N) :- clause(H, Body), solve(Body, M),
N is M+1.

Fig. 4.4 The SOLVE program

For the same query as above we now get the following answers:

[eclipse 16]: solve((app(Xs, [_, _|Ys], [mon, wed, fri, sun]),

app(Xs, Ys, Zs)), N).

Xs = []
Ys = [fri, sun]
Zs = [fri, sun]
N = 2
More (0.00s cpu) ? ;
 4.3 Meta-programming 73

Xs = [mon]
Ys = [sun]
Zs = [mon, sun]
N = 4
More (0.00s cpu) ? ;

Xs = [mon, wed]
Ys = []
Zs = [mon, wed]
N = 6
More (0.00s cpu) ? ;

No (0.00s cpu)

By using meta-variables we can easily extend the above program SOLVE

so that it also deals with the programs that use the comparison predicates.
To this end it suffices to modify it as in Figure 4.5.

solve(true, 0) :- !.
solve(A, 1) :- arithmetic(A), !, A.
solve((A, B), N) :- !, solve(A, N1), solve(B, N2),
N is N1+N2.
solve(H, N) :- clause(H, Body), solve(Body, M),
N is M+1.

arithmetic(_ < _).

arithmetic(_ =< _).
arithmetic(_ =:= _).
arithmetic(_ =\= _).
arithmetic(_ >= _).
arithmetic(_ > _).

Fig. 4.5 The SOLVE2 program

In this program the calls of the arithmetic atomic queries are just shifted
to the system level. In other words, these calls are executed directly by the
underlying Prolog system. We now have

[eclipse 17]: solve(qs([7,9,8,1,5], Ys), N).

74 Control and meta-programming

Ys = [1, 5, 7, 8, 9]
N = 36
More (0.00s cpu) ? ;

No (0.00s cpu)
So it takes 36 derivation steps to compute the answer to the query
qs([7,9,8,1,5], Ys).

4.4 Summary
In this chapter we discussed in turn:
ambivalent syntax and meta-variables,
various control constructs,
meta-programming.
In the process we introduced the following built-ins:
the if-then-else statement written in Prolog as B -> S ; T,
write/1,
disjunction: ;/2,
negation: not/1,
fail/0,
once/1,
cut: !/0,
clause/2.
In particular we showed how the syntactic features of Prolog (and the
absence of types) make it easy to write various meta-programs. The am-
bivalent syntax goes beyond the first-order logic syntax. It is useful to realise
that there is an alternative syntactic reading of Prolog programs in which
the first-order logic syntax is respected and the ambivalent syntax is not
used. Namely, after interpreting the operators used as the customary prefix
operators we can view each Prolog clause as a construct of the form
s :- t1 , . . ., tn ,
where s is a term that is neither a variable or a number and each ti is a
term that is not a number.
This syntactic reading respects the first-order logic but has one important
disadvantage. Namely, the declarative interpretation of the programs that
relied on the use of the predicates is then lost. The ambivalent syntax can
 4.5 Exercises 75

be viewed as an attempt to restore the declarative interpretation in the

presence of the operators.

4.5 Exercises

Exercise 4.1 Redefine the len/2 predicate from Figure 3.1 on page 45 so
that it can be used both to compute the length of a list, and to build a list
of different variables of a specified length, like the length/2 built-in. Hint.
Use the integer/1 built-in which tests whether its argument is an integer.

Exercise 4.2 Suppose we maintain sets as lists. Consider the following two
ways of defining the result of adding an element to a set:
% add(X, Xs, Ys) :- The set Ys is the result of adding
% the element X to the set Ys.

add(X, Xs, Ys) :- member(X, Xs), !, Xs = Ys.

add(X, Xs, [X | Xs]).
and
add(X, Xs, Xs) :- member(X, Xs), !.
add(X, Xs, [X | Xs]).
Find a query for which they yield different answers.

Exercise 4.3 Propose an implementation of a predicate while(B, S)

which is to simulate the customary while B do S statement of imperative
programming.
 5

Manipulating structures

5.1 Structure inspection facilities 76

5.2 Structure comparison facilities 78
5.3 Structure decomposition and construction facilities 80
5.4 Summary 85
5.5 Exercises 85

W E MENTIONED EARLY on that Prolog computations take place

over the set of all terms. One of the useful features of Prolog is
that it allows us to access the elements of this data by providing
various built-ins for inspecting, comparing and decomposing terms. On
the other hand, as we shall see, these features compromise the declarative
character of the language
Because of the ambivalent syntax of Prolog (discussed in Section 4.1) we
shall rather talk about structures instead of terms. These built-ins can be
naturally divided into three categories.

5.1 Structure inspection facilities

Here are the most important built-ins belonging to this category.
var/1, which tests whether the structure is a variable:
[eclipse 1]: var(X).

X = X
Yes (0.00s cpu)
[eclipse 2]: var(f(X)).

No (0.00s cpu)

76
 5.1 Structure inspection facilities 77

nonvar/1, which tests whether the structure is not a variable:

[eclipse 3]: nonvar(f(X)), nonvar(a).

X = X
Yes (0.00s cpu)

ground/1, which tests whether the structure is ground (has no variables):

[eclipse 4]: ground(f(a,b,c)).

Yes (0.00s cpu)

[eclipse 5]: ground(f(a,b,X)).

No (0.00s cpu)

compound/1, which tests whether the structure is compound:

[eclipse 6]: compound(f(a,b,c)).

Yes (0.00s cpu)

[eclipse 7]: compound(a).

No (0.00s cpu)

number/1, which tests whether the structure is a number (already men-

tioned in Section 4.2):

[eclipse 8]: number(10), number(-8), number(3.14).

Yes (0.00s cpu)

atom/1, which tests whether the structure is a non-numeric constant (an

atom in Prologs terminology):

[eclipse 9]: atom(a).

Yes (0.00s cpu)

[eclipse 10]: atom(10).

No (0.00s cpu)

atomic/1, which tests whether the structure is a constant, so an atom or

a number:
78 Manipulating structures

[eclipse 11]: atomic(a), atomic(3.14).

Yes (0.00s cpu)

To see why these built-ins affect the declarative reading of the programs
consider a query of the form A, A. In the case of pure Prolog its effect is
the same as A. In fact, the second call of A becomes then a test, like in the
query f(X,a) = f(b,Y), f(X,a) = f(b,Y). In the presence of the var/1
built-in things change:
[eclipse 12]: var(X), X = 1.

X = 1
Yes (0.00s cpu)
[eclipse 13]: var(X), X = 1, var(X), X = 1.

No (0.00s cpu)
Also the order of various simple atomic goals now becomes crucial: the
query var(X), X = 1 succeeds but the query X = 1, var(X) fails.

5.2 Structure comparison facilities

Next, the following built-ins allows us to compare structures.
==/2, which tests whether two structures are literally identical:
[eclipse 14]: f(X) == f(X).

X = X
Yes (0.00s cpu)
[eclipse 15]: f(X) == f(Y).

No (0.00s cpu)
\==/2, which tests whether two structures are not literally identical:
[eclipse 16]: f(X) \== f(X).

No (0.00s cpu)
[eclipse 17]: f(X) \== f(Y).

X = X
Y = Y
 5.2 Structure comparison facilities 79

Yes (0.00s cpu)

So these two built-ins are used as infix operators. In ECLi PSe they are
defined internally as follows:

:- op(700, xfx, [==, \==]).

As an example of the use of the built-ins introduced so far consider the

following seemingly trivial problem.

Problem Write a Prolog program that tests whether a term is a list.

A natural idea is to mimic the definition of a list:

list([]) :- !.
list([_ | Ts]) :- list(Ts).

(The cut prevents backtracking in the case of success.) Unfortunately, this

program is incorrect. Indeed, the query list(X) then succeeds instead of
failing:

[eclipse 18]: list(X).

X = []
Yes (0.00s cpu)

The obvious correction consists of using in the first clause the ==/2 built-in:

list(X) :- X == [], !.
list([_ | Ts]) :- list(Ts).

Now, however, the query list(X) diverges. The reason is that during its
execution a backtracking takes place and then the second clause is repeatedly
used. The right remedy turns out to add an appropriate test in the second
clause. The resulting program is given in Figure 5.1.

% list(X) :- X is a list.

list(X) :- X == [], !.
list([_ | Ts]) :- nonvar(Ts), list(Ts).

Fig. 5.1 The LIST program

80 Manipulating structures

5.3 Structure decomposition and construction facilities

The built-ins in this category allow us to decompose the structures. To
explain their meaning we use the following terminology. Given a structure
f (s1 , . . ., sn ), where n 0, we say that f is its leading symbol, n is its arity,
s1 , . . ., sn are its arguments and si is its ith argument (i [1..n]). (So if
n = 0 the structure has no arguments.) In addition, we call a structure of
the form f (x1 , . . ., xn ), where n 0 and x1 , . . ., xn are different variables, a
most general structure of arity n. When n = 0 the most general structure
is simply a constant.
The following three built-ins are most relevant.

functor/3 The functor/3 built-in either extracts from a structure the

leading symbol and its arity or constructs a most general structure of the
given arity with the given function symbol as the leading symbol.
More precisely, consider the query functor(t, f, n). Two cases arise.
t is a non-variable structure.
Let f1 be the leading symbol of t and n1 the arity of t. Then the query
functor(t, f, n) is executed by unifying the pair (f,n) with (f1,n1).
For example,
[eclipse 19]: functor(f(a,b), F, N).

F = f
N = 2
Yes (0.00s cpu)
[eclipse 20]: functor(c, F, N).

F = c
N = 0
Yes (0.00s cpu)
In the presence of the arithmetic operators functor/3 deals correctly
with the infix notation:
[eclipse 21]: functor(3*4+5, F, N).

F = +
N = 2
Yes (0.00s cpu)
Also, functor/3 allows us to reveal the internal representation of lists:
 5.3 Structure decomposition and construction facilities 81

[eclipse 22]: functor([], F, N).

F = []
N = 0
Yes (0.00s cpu)
[eclipse 23]: functor([a | b], F, N).

F = .
N = 2
Yes (0.00s cpu)
Indeed, the lists are represented internally using the functor ./2 and the
constant []. For example we have:

[eclipse 24]: X = .(a, .(b, [])).

X = [a, b]
Yes (0.00s cpu)
t is a variable.
Then f has to be a non-numeric constant (an atom) and n a natural
number. Then the query of functor(t, f, n) is executed by instantiat-
ing t to a most general structure of arity n the leading symbol of which
is f. For example:

[eclipse 25]: functor(T, f, 3).

T = f(_163, _164, _165)

Yes (0.00s cpu)
[eclipse 26]: functor(T, f, 0).

T = f
Yes (0.00s cpu)
[eclipse 27]: functor(T, ., 2).

T = [_162|_163]
Yes (0.00s cpu)

All other uses of functor/3 lead to a run-time error.

arg/3 The arg/3 built-in extracts a specific argument from a structure.

Consider the query arg(n, t, a). Then n has to be a natural number
82 Manipulating structures

and t a compound structure. The query arg(n, t, a) is then executed by

unifying the nth argument of t with a. For example:
[eclipse 28]: arg(2, f(a, g(a, X), c), A).

X = X
A = g(a, X)
Yes (0.00s cpu)
[eclipse 29]: arg(2, [a,b,c], A).

A = [b, c]
Yes (0.00s cpu)
As an example of the use of these built-ins consider the following problem.

Problem Given a term produce its list of variables (including repetitions).

The desired program is given in Figure 5.2.

This procedure can be applied to an arbitrary structure, for example a
clause:
[eclipse 30]: vars(
app([X|Xs],Ys,[X|Zs]) :- app(Xs,Ys,Zs), List
).

X = X
Xs = Xs
Ys = Ys
Zs = Zs
List = [X, Xs, Ys, X, Zs, Xs, Ys, Zs]
Yes (0.00s cpu)
Of course, by a minor modification of the above program we can compute
in the same way the list of variables of a term, without repetitions (see
Exercise 5.2).

=../2 The =../2 built-in (for historic reasons pronounced univ) either
creates a list which consists of the leading symbol of the structure followed
by its arguments or constructs a structure from a list that starts with a
function symbol and the tail of which is a list of structure arguments.
It is internally defined as an infix operator with the following declaration:
 5.3 Structure decomposition and construction facilities 83

% vars(Term, List) :-
% List is the list of variables of the term Term.

vars(Term, []) :- atomic(Term), !.

vars(Term, [Term]) :- var(Term), !.
vars(Term, List) :-
compound(Term),
functor(Term,_,K),
args(K, Term, List).

% args(K, Term, List) :-

% List is the list of variables occurring
% in the first K arguments of the term Term.

args(0, _, []) :- !.
args(K, Term, List) :-
K > 0,
K1 is K-1,
args(K1, Term, L1s),
arg(K,Term,A),
vars(A, L2s),
app(L1s, L2s, List).
augmented by the APPEND program of Figure 1.5.

Fig. 5.2 The VARIABLES program

:- op(700, xfx, =..).

More precisely, consider the query s =.. t. Two cases arise.
s is a non-variable structure, say f(s1 , . . ., sn ). Then the query s =.. t
is executed by unifying the list [f, s1 , . . ., sn ] with t. For example:
[eclipse 31]: f(a, g(X)) =.. List.

X = X
List = [f, a, g(X)]
Yes (0.00s cpu)
s is a variable.
Then t has to be a list, say [f, s1 , . . ., sn ] the head of which (so f) is
84 Manipulating structures

a non-numeric constant (an atom). Then the query s =.. t is executed

by instantiating s to the structure f(s1 , . . ., sn ). For example:

[eclipse 32]: T =.. [app, [a,b], [c,d], Zs].

T = app([a, b], [c, d], Zs)

Zs = Zs
Yes (0.00s cpu)

All other uses of =../2 lead to a run-time error.

So using =../2 we can construct terms and pass them as arguments. More
interestingly, we can construct queries and execute them using the meta-
variable facility. This way it is possible to realise higher-order programming
in Prolog in the sense that predicates can be passed as arguments. To
illustrate this point consider the program MAP given in Figure 5.3.

% map(P, Xs, Ys) :-

% The list Ys is the result of applying P
% elementwise to the list Xs.

map(_, [], []).

map(P, [X | Xs] , [Y | Ys]) :-
apply(P, [X, Y]),
map(P, Xs, Ys).

% apply(P, [X1, ..., Xn]) :-

% Execute the atomic query P(X1, ..., Xn).

apply(P, Xs) :- Query =.. [P|Xs], Query.

Fig. 5.3 The MAP program

In the last clause =../2 is used to construct an atomic query. Note the
use of the meta-variable Query. MAP is Prologs counterpart of a higher-order
functional program and it behaves in the expected way. For example, given
the clause
increment(X, Y) :- Y is X+1.
we get
 5.4 Summary 85

[eclipse 33]: map(increment, [1,2,3,4], Ys).

Ys = [2, 3, 4, 5]
Yes (0.00s cpu)

5.4 Summary
In this chapter we discussed Prolog built-ins that allow us to manipulate
structures. We discussed built-ins that allow us

to inspect structures:
var/1,
nonvar/1,
ground/1,
compound/1,
number/1,
atom/1,
atomic/1,
to compare structures:
==/2,
\==/2,
to decompose and construct structures:
functor/3,
arg/3,
../2.

Using these built-ins we can analyse structures and also realise higher-order
programming in Prolog.

5.5 Exercises

Exercise 5.1 Write an alternative version of the program VARIABLES from

Figure 5.2 using =../2, without any arithmetic operations.

Exercise 5.2 Modify the program VARIABLES from Figure 5.2 to compute
the list of variables of a term, without repetitions.
86 Manipulating structures

Exercise 5.3 Write a program that implements the substitution operation.

The call subs(Input, X, Term, Output) should perform the substitution
of the variable X by the term Term in the input term Input. The result
should be produced in the last argument.
 Part III

Programming with passive constraints

 6

Constraint programming: a primer

6.1 Introduction 89
6.2 Basic concepts 90
6.3 Example classes of constraints 91
6.4 Constraint satisfaction problems: examples 94
6.5 Constrained optimisation problems 98
6.6 Solving CSPs and COPs 102
6.7 From CSPs and COPs to constraint programming 107
6.8 Summary 108
6.9 Exercises 108

6.1 Introduction

C ONSTRAINT PROGRAMMING IS an alternative approach to pro-

gramming in which the programming process is limited to a gener-
ation of requirements (constraints) and a solution of these require-
ments by means of general or domain specific methods. In the forthcoming
chapters we explain how constraint programming is realised in ECLi PSe .
To this end we shall explain in turn:
what constraint satisfaction problems and constrained optimisation prob-
lems are and how they are solved using constraint programming methods,
how to generate these problems in ECLi PSe ,
what support ECLi PSe provides for general and domain specific methods,
how to solve constraint satisfaction problems and constrained optimisation
problems in ECLi PSe .
In this chapter we focus on the first topic. Informally, a constraint sat-
isfaction problem consists of a set of constraints, each on a sequence of
variables, and the task is to find a solution to it. In turn, a constrained op-
timisation problem is a constraint satisfaction problem equipped with a cost

89
90 Constraint programming: a primer

function and the task is to find an optimal solution w.r.t. the cost function.
In the area of constraint programming various methods and techniques were
developed to solve both kind of problems and we shall explain briefly their
basics.

6.2 Basic concepts

Syntactically, a constraint is an atomic formula in some, usually first-order,
language. We already encountered in Chapter 3 constraints that involve
arithmetic comparison operators, so x < y, x y, x = y, x 6= y, x
y, x > y. But several other natural examples exist and we shall discuss
them below. Depending on the number of variables used in a constraint we
speak about a unary , binary or n-ary constraint. We assume that each
constraint has at least one variable.
When discussing the constraints we have in mind their specific inter-
pretation determined by the (non-empty) domains of the variables and the
meaning assigned to the used function symbols and predicates. For example,
if we interpret the variables over the set of integers Z, then x < y determines
the set {(a, b) | a < b, a, b Z}, since we implicitly interpret < as the less
than relation. In turn, if we interpret the variables over the set of reals R,
then x < y determines the set {(a, b) | a < b, a, b R}.
In general, each interpretation associates with a constraint c a subset of
the Cartesian product of the domains of its variables. We refer to such an
interpretation of a constraint c as a [c]. From the context it will be always
clear over which domains the variables range and which interpretation of
the used function symbols and predicates we have in mind.
An assignment of values to the constraint variables satisfies a con-
straint (is a solution to a constraint) if the used sequence of the values
belongs to its interpretation. For example, if we interpret the variables over
the set of integers, then the assignment {(x, 2), (y, 5)} satisfies the con-
straint x2 < y. We say that a constraint is satisfiable if some assignment
of values to its variables satisfies it and unsatisfiable if no such assignment
exists. For example, if we interpret the variables over the set of reals, the
constraint x > x + y is satisfiable, whereas if we interpret the variables over
the set of natural numbers, it is unsatisfiable.
In turn, a constraint is solved if all assignments of values to its variables
satisfy it. For example, if we interpret the variables over the set of natural
numbers, the constraint x + y 0 is solved.
By a constraint satisfaction problem, in short a CSP, we mean a
finite sequence of variables, each ranging over a possibly different domain,
 6.3 Example classes of constraints 91

and a finite set of constraints, each on a subsequence of the considered

variables. The variables used in a CSP are called decision variables. We
always consider the CSPs in the context of an interpretation. By a solution
to a CSP we mean an assignment of values to its variables that satisfies
each constraint. We say that a CSP is consistent (or feasible) if it has a
solution and inconsistent (or infeasible) if it does not.
Further, we call a CSP solved if each of its constraints is solved. So
a CSP is solved if each assignment of values to its variables is a solution.
Finally, we say that a CSP is failed if one its domains is empty or one of its
constraints is unsatisfiable. Clearly, if a CSP is failed then it is inconsistent,
but the reverse implication does not hold. In practice it is much easier to
check whether a CSP is failed than whether it is inconsistent.
More formally, consider a CSP involving a sequence of variables X :=
x1 , . . ., xn with respective domains D1 , . . ., Dn , and a finite set C of con-
straints, each on a subsequence of X. We write it as hC ; x1 D1 , . . ., xn
Dn i. By a solution to hC ; x1 D1 , . . ., xn Dn i we mean an assignment
{(x1 , d1 ), . . ., (xn , dn )}, where each di is an element of Di , such that for each
constraint c C on the variables xk1 , . . ., xkm we have (dk1 , . . ., dkm ) [c],
where [c] is the interpretation of c.
Given two CSPs hC ; x1 D1 , . . ., xn Dn i and hC 0 ; x1 D10 , . . ., xn
Dn0 i on the same sequence of variables we say that they are equivalent if
they have the same set of solutions.

6.3 Example classes of constraints

Let us discuss now the most common classes of constraints we shall deal with.
Each of them is uniquely determined by the considered function symbols and
predicates. In each case we have a specific interpretation in mind which is
clear from the context.

Equality and disequality constraints

This is the simplest meaningful class of constraints. We have no function
symbols and just two predicates, equality, =, and disequality, 6=. The
variables are interpreted over an arbitrary domain.
The constraints can be either unary, like x = x, or binary, like x 6= y.
Further, note that the constraint x = x is solved, while the constraint x 6= x
is unsatisfiable. In turn, the constraint x = y is satisfiable iff the domains
of x and y are not disjoint, and the constraint x 6= y is satisfiable iff either
92 Constraint programming: a primer

the domains of x and y are disjoint or at least one of them is not a singleton
set.
As an example of a CSP involving these constraints take
hx = y, y 6= z, z 6= u ; x {a, b, c}, y {a, b, d}, z {a, b}, u {b}i.
It is straightforward to check that {(x, b), (y, b), (z, a), (u, b)} is its unique
solution.

Boolean constraints
Boolean constraints are formed using specific function symbols called con-
nectives. We adopt the unary connective (negation) and the binary
connectives (conjunction) and (disjunction). We also allow two
constants, true and false. The resulting terms are called Boolean ex-
pressions. The binary connectives are written using the infix notation and
when applied to a variable is written without the brackets. By a Boolean
constraint we mean a formula of the form s = t, where s, t are Boolean ex-
pressions. So the equality symbol = is the only predicate in the language.
Each Boolean expression s can be viewed as a shorthand for the Boolean
constraint s = true.
The variables are interpreted over the set of truth values, identified with
{0, 1}, and are called Boolean variables. In turn, the connectives are
interpreted using the standard truth tables. This interpretation allows us to
interpret each Boolean constraint c on the Boolean variables x1 , . . ., xn as a
subset of {0, 1}n .
For example, the interpretation of the Boolean constraint (x) (y z),
that is (x) (y z) = true, is the set
{(0, 0, 1), (0, 1, 0), (0, 1, 1)},
where we assume the alphabetic ordering of the variables.

Linear constraints
Next, we define linear constraints. We shall interpret them either over the
set of integers or over the set of reals, or over a subset of one of these sets
(usually an interval). To start with, we assume a fixed set of numeric
constants representing either integers or reals.
As in the case of Boolean constraints we introduce the constraints in two
steps. By a linear expression we mean a term formed in the alphabet
that contains
 6.3 Example classes of constraints 93

the assumed numeric constants,

for each numeric constant r the unary function symbol r (representing
the multiplication by r),
two binary function symbols, + and , both written in the infix nota-
tion.

We also assume that + and are left associative and have the same
binding strength.
By a linear constraint we mean a formula of the form

s op t,

where s and t are linear expressions and op {<, , =, 6=, , >}.

For example, 4 x + 3 (y x + 7) is a linear expression and

4x+3yx5z+7y

is a linear constraint.
When linear constraints are interpreted over the set of reals, it is custom-
ary to admit only , = and as the comparison operators.

Arithmetic constraints
Finally, we introduce arithmetic constraints. The only difference between
them and linear constraints lies in the use of the multiplication symbol. In
the case of linear constraints the multiplication was allowed only if one of
its operands (the first one) was a numeric constant. To fit this restriction
into the syntax of first-order logic we introduced the unary multiplication
symbols r. Now we dispense with these unary function symbols and simply
allow the multiplication as a binary function symbol.
We define the arithmetic expressions and arithmetic constraints
analogously to the linear expressions and linear constraints, but now using
the binary multiplication symbol . In particular, each linear constraint is
an arithmetic constraint.
For example 4 x + x y x + 7 is an arithmetic expression and

4 x + x y x y (z + 5) 3 u

is an arithmetic constraint.
94 Constraint programming: a primer

6.4 Constraint satisfaction problems: examples

To clarify the introduced concepts and classes of constraints we now dis-
cuss some well-known CSPs. Our interest in introducing these examples is
twofold.
We want to clarify the process of formalising a problem as a CSP.
This process is usually called modelling . In general more than one
representation of a problem as a CSP exists and the selection of an appro-
priate representation can have a dramatic impact on the problem solving
process.
In the subsequent chapters we shall explain how the CSPs introduced here
are modelled and solved in ECLi PSe .

6.4.1 Map colouring problems

These are problems that are specified by
a finite set of regions,
a (smaller) set of colours,
a neighbour relation between pairs of regions.
Given a set of regions and of colours and a neighbour relation the problem
is to associate a colour with each region so that no two neighbours have the
same colour.
To model such a problem as a CSP we simply use disequality constraints.
We represent each region i by a variable xi . The domain of each variable
is the set of assumed colours. Then for each pair of regions (i, j) that are
neighbours we introduce the constraint xi 6= xj .
To be more specific suppose that
there are four regions 1, 2, 3, 4,
there are three colours, red, yellow, blue,
The neighbour relation consists of the following set of pairs:
{(1, 2), (1, 3), (1, 4), (2, 3), (2, 4)},
see Figure 6.1.
Then we have four variables x1 , x2 , x3 , x4 , each ranging over the set
{red, yellow, blue} and the following set of disequality constraints:
{x1 6= x2 , x1 6= x3 , x1 6= x4 , x2 6= x3 , x2 6= x4 }.
This specific problem has six solutions:
 6.4 Constraint satisfaction problems: examples 95

3 2 4

Fig. 6.1 An example map

{(x1 , red), (x2 , yellow), (x3 , blue), (x4 , blue)},

{(x1 , red), (x2 , blue), (x3 , yellow), (x4 , yellow)},
{(x1 , yellow), (x2 , red), (x3 , blue), (x4 , blue)},
{(x1 , yellow), (x2 , blue), (x3 , red), (x4 , red)},
{(x1 , blue), (x2 , red), (x3 , yellow), (x4 , yellow)},
{(x1 , blue), (x2 , yellow), (x3 , red), (x4 , red)}.

6.4.2 SEND + MORE = MONEY puzzle

This is a classic example of a so-called cryptarithmetic problem. These
are mathematical puzzles in which the digits are replaced by letters of the
alphabet or other symbols. In the problem under consideration we are asked
to replace each letter by a different digit so that the above sum, that is

SEND
+ MORE
MONEY
is correct.
The problem can be formulated using linear constraints in a number of
different ways. The variables are S, E, N, D, M, O, R, Y . The domain of
each variable consists of the integer interval [0..9].

Representation 1 We use the equality constraint

1000 S + 100 E + 10 N + D
+ 1000 M + 100 O + 10 R + E
= 10000 M + 1000 O + 100 N + 10 E + Y,
96 Constraint programming: a primer

two disequality constraints:

S 6= 0, M 6= 0,
and 28 disequality constraints x 6= y for x, y ranging over the set
{S, E, N, D, M, O, R, Y } with x preceding y in, say, the presented order.

Representation 2 Another natural modelling consists of additionally in-

troducing per column a carry variable ranging over {0, 1} and using instead
of the above single equality constraint five simpler ones, one for each column:

D+E = 10 C1 + Y,
C1 + N + R = 10 C2 + E,
C2 + E + O = 10 C3 + N,
C3 + S + M = 10 C4 + O,
C4 = M.
Here, C1 , . . ., C4 are the carry variables.

Representation 3 We replace the 28 disequality constraints x 6= y by a sin-

gle constraint that stipulates that the variables are pairwise different. Given
a sequence of variables x1 , . . ., xn , where n 2, with respective domains
D1 , . . ., Dn we introduce an n-ary all different constraint interpreted as
follows:
[all different(x1 , . . ., xn )] := {(d1 , . . ., dn ) | di 6= dj for i 6= j}.

Then we can replace the above 28 disequality constraints by a single con-

straint
all different(S, E, N, D, M, O, R, Y ).

Representation 4 Yet another way to eliminate the above 28 disequality

constraints is by following a standard method used in the area of Integer Lin-
ear Programming, the objective of which is to find optimal integer solutions
to linear constraints over the reals.
For each variable x from {S, E, N, D, M, O, R, Y } and a digit d we intro-
duce a 0/1 variable isx,d . The constraint that each variable from
{S, E, N, D, M, O, R, Y } takes one value is expressed as the set of eight con-
straints
X9
isx,d = 1,
d=0
 6.4 Constraint satisfaction problems: examples 97

where x {S, E, N, D, M, O, R, Y }. In turn, the all different constraint

is expressed as the set of ten constraints
X
isx,d 1,
x{S,E,N,D,M,O,R,Y }

where d {0, 1, . . ., 9}. The 0/1 variables isx,d relate back to the original
variables used in the equality constraint(s) via eight constraints
9
X
d isx,d = x,
d=0

where x {S, E, N, D, M, O, R, Y }.
The disadvantage of this approach is of course that we need to introduce
80 new variables.
The problem under consideration has a unique solution depicted by the
following sum:
9567
+ 1085
10652
That is, the assignment

{(S, 9), (E, 5), (N, 6), (D, 7), (M, 1), (O, 0), (R, 8), (Y, 2)}

is a unique solution of the original CSP and each other discussed represen-
tation of the problem has a unique solution, as well.

6.4.3 The n-queens problem

This is probably the most well-known CSP. The problem is to place n queens
on the nn chess board, where n 3, so that they do not attack each other.
We discuss two representations of this problem as a CSP.

Representation using 0/1 variables This representation uses n2 0/1

variables xi,j , where i [1..n] and j [1..n], each of them representing one
field of the chess board. The appropriate constraints can then be written
using sums of relevant variables.
The following constraints formalise the problem:
Pn
x = 1 for i [1..n] (exactly one queen per row),
Pnj=1 i,j
j=1 xj,i = 1 for i [1..n] (exactly one queen per column),
98 Constraint programming: a primer

xi,j + xk,l 1 for i, j, k, l [1..n] such that i 6= k and |i k| = |j l|

(at most one queen per diagonal).

The condition i 6= k in the last item ensures that the fields represented by
xi,j and xk,l are different.
The obvious disadvantage of this representation is the large number of
variables it involves.

Representation using linear constraints A more economic representa-

tion of this problem as a CSP uses n variables, x1 , . . ., xn , each with the
domain [1..n]. The idea is that xi denotes the position of the queen placed
in the ith column of the chess board.
This representation implies that no two queens are placed in the same
column, so we only need to take care of the remaining requirements. They
can be formulated as the following disequality constraints for i [1..n] and
j [1..i 1]:

xi 6= xj (at most one queen per row),

xi xj 6= i j (at most one queen per SW-NE diagonal),
xi xj 6= j i (at most one queen per SE-NW diagonal).

Using the all different constraint introduced in the third representa-

tion of the SEND + MORE = MONEY problem we can replace the first
set of n(n1)
2 disequalities by a single constraint

all different(x1 , . . ., xn ).

6.5 Constrained optimisation problems: examples

The general task of constrained optimisation is to find optimal solutions
to a set of constraints subject to some cost function cost. More precisely,
consider a CSP P := hC ; x1 D1 , . . ., xn Dn i together with a function

cost : D1 Dn R

from the set of all sequences in D1 Dn to the set R of real num-

bers. We are interested in finding a solution {(x1 , d1 ), . . ., (xn , dn )} to P for
which the value cost(d1 , . . ., dn ) is optimal. In what follows we assume that
optimal means minimal. We call then the pair (P, cost) a constrained
optimisation problem, in short a COP. The area of constrained optimi-
sation is huge and a number of books have been written on this subject.
Here we limit ourselves to a presentation of a few simple examples.
 6.5 Constrained optimisation problems: examples 99

6.5.1 The knapsack problem

We are given n objects, each with a volume and a value, and a knapsack of a
fixed volume. The problem is to find a collection of the objects with maximal
total value that fits in the knapsack. More formally, we have n objects with
volumes a1 , ..., an and values b1 , ..., bn and the knapsack volume v. Further,
we have n variables, x1 , ..., xn , each with the domain {0, 1}. The inclusion
of the object i in a collection is modelled by setting the value of xi to 1. The
requirement that the collection fits in the knapsack translates to the linear
constraint
X n
ai x i v
i=1

on the variables x1 , ..., xn . We seek a solution to this constraint for which

the sum
X n
bi xi
i=1

is maximal. Since in our setup we seek solutions with a minimal value of

the cost function, we use ni=1 bi xi as the cost function.
P

6.5.2 A coins problem

Consider the following problem.
What is the minimum number of coins that allows one to pay exactly any amount
smaller than one euro? Recall that there are six different euro cent coins, of de-
nomination 1, 2, 5, 10, 20, 50.

We formulate it as a COP. First, we determine the variables and their

domains. Since the question refers to the number of coins, it is natural to
adopt the variables that represent the selected amounts of each coin. This
brings us to six variables that we denote by x1 , x2 , x5 , x10 , x20 , x50 . The idea
is that each variable xi denotes the selected amount of coins of denomination
i, so we use [0..99] as the domain for each variable.
Further, we introduce for each i [1..99] six variables xi1 , xi2 , xi5 , xi10 , xi20 ,
i
x50 with respectively the same domains as x1 , x2 , x5 , x10 , x20 and x50 to
formulate a constraint stating that the amount of i euro cents can be paid
exactly using the selected coins. So we have in total 600 variables.
Next, we state the appropriate constraints. For each i [1..99] we use the
linear constraint
xi1 + 2xi2 + 5xi5 + 10xi10 + 20xi20 + 50xi50 = i
100 Constraint programming: a primer

to state that the amount of i euro cents can be paid using xi1 coins of 1 euro
cent, xi2 coins of 2 euro cents, xi5 coins of 5 euro cents, xi10 coins of 10 euro
cents, xi20 coins of 20 euro cents and xi50 coins of 50 euro cents.
Additionally, we add the constraints stating that for each i [1..99] the
amounts xij are respectively smaller than xj . More precisely, for each i
[1..99] and j {1, 2, 5, 10, 20, 50} we use the following constraint:
xij xj .
These 594 inequality constraints ensure that each amount of i euro cents
can indeed be paid using the collection represented by the xj variables.
Finally, for each solution to the just formulated CSP the sum
x1 + x2 + x5 + x10 + x20 + x50
represents the total amount of coins that we would like to minimise. So it
is the appropriate cost function.

6.5.3 A currency design problem

Suppose now that we have the freedom of choosing the values of six coins.
What is then the minimum number of coins that allows one to pay exactly
any amount smaller than one euro?
To formulate this problem as a COP it suffices to modify the above linear
equality constraints but this leads to the arithmetic constraints. Indeed, we
just replace each previous linear constraint
xi1 + 2xi2 + 5xi5 + 10xi10 + 20xi20 + 50xi50 = i
by the arithmetic constraint
v1 xi1 + v2 xi2 + v3 xi3 + v4 xi4 + v5 xi5 + v6 xi6 = i,
where v1 , . . ., v6 are the values of the coins. Additionally, we assume that
the vi variables range over [1..99] and are ordered, so stipulate for i [1..5]
the constraint
vi < vi+1 .
The rest of the representation is the same.

6.5.4 The facility location problem

This problem is concerned with a choice of a set of facility locations to best
serve a given set of customers with a given set of demands.
 6.5 Constrained optimisation problems: examples 101

More precisely, we are given a number of customers and a number of

locations for a facility, say a warehouse. There is only one product provided.
Each customer i is characterized by
its demand (in the provided product) demi .
In turn, each warehouse location (from now on, just a warehouse) j is char-
acterized by
its setup cost scostj , and
its capacity capj .
We also have for each warehouse j and each client i the transportation
costs tcosti,j of the product from warehouse j to client i, per unit of supplied
quantity.
We assume that the total capacity of all the warehouses exceeds (or equals)
the total demand from all the customers, i.e.,
X X
capj demi .
j i

Further, we assume that each customer may be served by any set of ware-
houses. We want to choose a set of warehouses from which the product will
be provided to the customers so that
the demands of all customers are satisfied,
the total cost, which equals the sum of the setup costs of the selected
warehouses plus the resulting transportation costs, is minimal.
In Figure 6.2 we illustrate the situation for three warehouses and four
customers. With each line connecting a warehouse and a customer a trans-
portation cost is associated.
To formalise this problem as a COP we use two kinds of decision variables.
For each warehouse j there is a Boolean variable Bj reflecting whether the
warehouse j is open (i.e., chosen) or not. Next, for each warehouse j and
customer i there is a continuous variable (i.e., a variable ranging over the
set of reals) Si,j saying how much of the customer i demand is supplied by
the warehouse j.
Then for each warehouse j there is a capacity constraint
X
capj Si,j ,
i

and for each customer i there is a demand constraint

X
Si,j demi .
j
102 Constraint programming: a primer

customer 1
dem1
cost 1
,1
t
warehouse 1
scost1 tcost2,1
cap1
customer 2
dem2
warehouse 2
scost2 tco
cap2 st
3,1

customer 3
warehouse 3 dem3
scost3
tc
os
cap3
t 4,
1

customer 4
dem4

Fig. 6.2 A facility location problem

We also state that the supplies from the warehouses that are not selected
equal zero, i.e., for each warehouse j

Bj = 0 i : Si,j = 0.

Finally, the cost function is the sum of the setup costs plus the trans-
portation costs:
X X
(Bj scostj + Si,j tcosti,j ).
j i

6.6 Solving CSPs and COPs

The definition of constraint satisfaction problems and constrained optimi-
sation problems is extremely general and it is of course preferable to solve
each such problem by domain specific methods. For example, when we deal
with systems of linear equations over reals, it is natural to use the methods
 6.6 Solving CSPs and COPs 103

developed in the area of Linear Algebra and when we deal with systems of
linear inequalities over reals, it is preferable to use the methods developed
in the area of Linear Programming.
Still, in many situations the constraints used do not adhere to any simple
syntax restrictions or are a mixture of several types of constraints. Then it
is useful to rely on general methods. Constraint programming developed a
collection of such methods. They allow us to solve problems formulated as
CSPs and COPs. These methods are based on search. In general two forms
of search are used.

6.6.1 Local search

Local search is a generic name for a whole class of search algorithms. Their
purpose is to find a solution to a CSP or to a COP by starting with an initial
assignment defined on all the variables (called in this context a state) and
trying to improve its quality iteratively, by small (so local) changes, called
moves. The quality of a state is defined by means of a cost function. A
simple example of a cost function is the number of constraints violated by
the state. Then the quality of a solution is 0.
The basic ingredient of the local search is the notion of a neighbourhood ,
which is a function that assigns to each state s (i.e., an assignment) a set
of states, that are called the neighbours of s. An execution of a local
search algorithm starts in an initial state, obtained by another technique
or randomly, and enters a loop in which repeatedly a move is performed
from a state to one of its neighbours. The final state is either a solution
to the considered CSP (respectively, an optimal solution to a COP) or a
stop state. In general the local search is incomplete so a stop state only
indicates that no solution has been found so far. Specific forms of local
search are obtained by specifying the way the moves are chosen and the way
the search is stopped.

6.6.2 Top-down search

In this book we focus on a more common form of search, usually called
top-down search. This form of search is combined with a branching (or
splitting ) strategy. Additionally, a crucial technique for constraint pro-
gramming of constraint propagation is used.
Intuitively, branching allows us to split a given CSP into two or more
CSPs the union of which is equivalent to the initial CSP. In turn, constraint
propagation allows us to transform a given CSP into one that is equivalent
104 Constraint programming: a primer

but simpler. The terms in italic can be made precise. Constraint prop-
agation is alternated with branching. This yields a tree, called a search
tree, over which a top-down search takes place. The leaves of the tree are
CSPs that are either obviously inconsistent (usually with an empty variable
domain) or solved. An important aspect of the top-down search is that the
search tree is generated on the fly.
An example of a search tree in which we refer explicitly to constraint
propagation and branching is depicted in Figure 6.3.

constraint propagation

branching

constraint propagation

branching

constraint propagation

Fig. 6.3 A search tree for a CSP

The standard form of a top-down search is the usual backtracking search.

It starts at the root of the tree and proceeds by descending to its first de-
scendant. This process continues as long as a node is not a leaf. If a leaf
is encountered, the search proceeds by moving back to the closest ancestor
node of the leaf that has another descendant or else to the root. If such an
ancestor exists, its next descendant is selected. This process continues until
the control is back at the root node and all of its descendants have been
visited. In Figure 6.4 we depict the backtracking search on a simple tree,
where the arrows indicate the order in which the nodes are visited.
The most common form of backtracking search proceeds by ordering the
decision variables according to some variable choice heuristics and by
branching through splitting the domain of a variable. This heuristic may
reorder the variable statically, so only once, at the beginning of the search,
or dynamically, during the search.
An example of branching is labelling that splits a finite domain of a
variable into the singleton domains. This corresponds to a systematic ex-
 6.6 Solving CSPs and COPs 105

Fig. 6.4 Backtracking search

ploration of all assignments of a value to this variable. The ordering of

the values in each domain is then done according to some value choice
heuristics. Another natural form of branching is bisection which is ap-
propriate when the domains are intervals or finite sets. In this case the
variable domain is split in two halves by choosing an appropriate mid-point.
This form of backtracking search is usually combined with some form of
constraint propagation that involves removal from the domains of variables
some values that do not participate in any solution. As an example consider
the constraint x < y, where the domain of x is [10..1000] and the domain of
y is [1..50]. Then the constraint propagation allows us to reduce the domain
of x to [10..49] and the domain of y to [11..50]. This domain reduction
can trigger a sequence of further domain reductions concerned with other
constraints in which, if the domain of x or y is further reduced, the constraint
x < y is again selected.
The presence of constraint propagation can be sometimes exploited by
adding to the initial CSP implied constraints. These constraints do not
alter the set of solutions but can lead to a smaller search tree. As an example
consider the CSP

hx3 + x < 520, x3 + y > 550 ; x [1..1000], y [1..50]i.

If the constraint propagation is available only for linear constraints, when

solving this CSP by backtracking search first a branching will take place. But
in presence of the implied constraint x < y first the constraint propagation
will take place and, as explained above, the domains of x and y will get
reduced to [10..49] and [11..50] before any branching takes place.
The backtracking search is an example of a complete search. In this form
of search all possible assignments of values are explored. In the incomplete
search one focusses on the most promising assignments selected according
to some heuristics. Various forms of incomplete top-down search can be
described by allocating credit to each value choice, giving more credit to
more promising values. This credit is then available for the search subtree
106 Constraint programming: a primer

which remains after choosing that value for that variable. In Section 8.4 we
shall discuss in more detail this form of search and program it in Prolog.

6.6.3 Branch and bound search

To solve COPs one takes into account the cost function. The appropri-
ate modification of the backtracking search is called branch and bound
search. Informally, it can be explained as follows. Suppose that we are
interested in finding a solution with the minimal value of the cost func-
tion. During the search one maintains the currently best value of the
cost function in a variable bound. This variable is initialised to . Each
time a solution with a smaller cost is found, this value is recorded in bound.
There are many variations on the branch and bound algorithm. One
consideration is what to do after a solution with a new best cost is found.
The simplest approach is to restart the computation with the bound variable
initialised to this new best cost.
A less naive approach is to continue the search for better solutions without
restarting. In this case the constraint cost(x1 , . . ., xn ) < bound is used and
each time a solution with a new best cost is found, this cost is dynamically
imposed through the constraint cost(x1 , . . ., xn ) < bound. The constraint
propagation triggered by this constraint leads to a pruning of the search
tree by identifying the nodes under which no solution with a smaller cost
can be present.
In Figure 6.5 we depict a possible branch and bound search on the tree
already considered in Figure 6.4. The parts of the tree that are ignored
during the search are represented using the dotted lines.

Fig. 6.5 Branch and bound search

Sometimes it is better to give up on the completeness and search for a

solution the cost of which is close to the optimum. This can be achieved
by imposing each time a tighter bound on the bound variable, for example
requiring the next solution to be a fixed amount better than the current
best. If this fixed amount is, say, 10, then when the branch and bound
 6.7 From CSPs and COPs to constraint programming 107

search terminates, the solution returned is guaranteed to be within 10 of

the optimum. Alternatively the improvement can be expressed as a fraction
of the current best bound. If the fraction is, say 0.1, and a solution with cost
100 had just been found, then the new bound would be 90. Then, once the
branch and bound search terminates, the solution returned is guaranteed to
be within 10% off the optimum.
A more radical variant is to exploit the full range of the cost variable from
its upper (worst) bound to its lower (best) bound. When a solution is found,
this yields a new upper bound on the cost variable. However, the algorithm
selects the mid-point between the current upper and lower bounds, and seeks
a solution with cost better than this. If there is no solution, the lower bound
of the cost is updated, and the new mid-point is selected. Search stops when
the upper and lower bounds are sufficiently close. This dichotomic branching
approach is only rarely the most efficient.

6.7 From CSPs and COPs to constraint programming

Now that we have explained the basic techniques used to solve constraint
satisfaction problems and constrained optimisation problems it is clear how
one can try to solve them in any programming language. The distinguish-
ing feature of constraint programming languages is that they provide
a substantial support for generating and solving CSPs and COPs. First,
note that in the traditional programming languages such as C or Java, the
variables are varying but known entities. So these variables cannot be used
to represent the CSP variables that are similar to unknowns in the sense
of algebra. Hence to represent CSPs one needs to represent the variables
differently, for example by means of objects.
In constraint programming languages and systems the representation of
the CSP variables and their domains is straightforward thanks to the built-in
facilities. As a result the programmer does not need to come up with his/her
own way of managing such variables. Additionally, a support is provided
for generating constraints. In short, generating CSPs in such languages
and systems is considerably simpler than in the traditional programming
languages.
Next, the general methods used to solve CSPs and COPs rely on a small
number of reasonably well understood sets of techniques that we discussed in
the previous section. These techniques, including backtracking and branch
and bound search, and various variable and value choice heuristics and other
relevant parameters of the search, are often directly available to the program-
mer.
108 Constraint programming: a primer

Finally, several types of CSPs can be solved using specialised, domain spe-
cific methods. We already mentioned the examples of linear equations over
reals and linear inequalities over reals. In typical constraint programming
languages and systems these methods are directly available to the user in the
form of constraint solvers that can be readily invoked from a program.
Additionally, these languages and systems integrate such constraint solvers
with the general methods, such as built-in forms of constraint propagation,
so that the programmer can rely automatically on both types of methods.
As a result generating and solving of CSPs and COPs in constraint pro-
gramming languages and systems is considerably simpler than in conven-
tional programming languages. The resulting approach to programming,
i.e. constraint programming , is radically different from the usual one.
In particular, modelling plays an important role, while the programming
process is based on a judicious use of the available built-ins.
In the subsequent chapters we shall explain the ECLi PSe approach to
constraint programming. ECLi PSe , as already mentioned, extends Prolog
and consequently uses variables as unknowns, as in algebra or mathematical
logic. This makes it substantially easier to model CPSs and COPs since they
use the variables in the same sense. Further, ECLi PSe relies on a number
of libraries that facilitate the task of the programmer and simplify the task
of constraint programming in a dramatic way.

6.8 Summary
The aim of this chapter was to explain the basic principles of constraint
programming. To this end we introduced first the concepts of a constraint,
constraint satisfaction problem (CSP) and constrained optimisation prob-
lem (COP) and illustrated them by means of various examples. We also
mentioned the most common classes of constraints considered in practice.
Then we discussed the methods used to solve CSPs and COPs, notably
constraint propagation, backtracking search and branch and bound search.
We explained how these methods are supported in constraint programming
languages. In these languages specialised, domain specific methods are di-
rectly available in the form of constraint solvers.

6.9 Exercises

Exercise 6.1 Consider the following puzzle taken from Gardner [1979].
Ten cells numbered 0, . . ., 9 inscribe a 10-digit number such that each cell, say i,
 6.9 Exercises 109

indicates the total number of occurrences of the digit i in this number. Find this
number.
Formulate this problem as a CSP problem.
For your information: the answer is 6210001000.

Exercise 6.2 Consider the following problem:

Find a permutation of numbers 1, . . ., 10 such that
- 6 is put on 7th place,
- each number starting from the second one is either bigger by 3 or smaller by 2
than its predecessor.
Formulate this problem as a CSP.
For your information: the only solution is: 3 1 4 2 5 8 6 9 7 10.

Exercise 6.3 A magic square of order n is defined to be an n n array

made out of the integers 1, 2, . . ., n2 arranged so that the sum of every row,
column, and the two main diagonals is the same. For example
1 15 24 8 17
23 7 16 5 14
20 4 13 22 6
12 21 10 19 3
9 18 2 11 25
is a magic square of order 5, because each row, column and main diagonal
sums up to 65. Formulate the problem of finding a magic square of order n
as a CSP.
 7

Intermezzo: Iteration in ECLiPSe

7.1 Introduction 110

7.2 Iterating over lists and integer ranges 111
7.3 Iteration specifications in general 113
7.4 Arrays and iterations 121
7.5 fromto/4: the most general iterator 124
7.6 The n-queens problem 129
7.7 Summary 130
7.8 Exercises 132

7.1 Introduction

D ECLARATIVE LANGUAGES USE recursion a great deal. Often it

is used simply to iterate over a data structure, or an integer range,
in which case it makes the program hard to read. An example is
the LENGTH program given in Figure 3.1 on page 45:
% len(List, N) :- N is the length of the list List.

len([], 0).
len([_ | Ts], N) :- len(Ts, M), N is M+1.
It does seem a rather obscure and complicated way of expressing some-
thing quite simple. The difficulty is, of course, that declarative programs do
not allow the same variable to be reused for different things, so within the
pure declarative paradigm, we cannot just walk down the list incrementing
a counter.
ECLi PSe offers a syntax for encoding such iterations in a reasonably nat-
ural way. In this chapter we discuss these constructs in a systematic way. In
particular we show how they can be used to obtain a non-recursive solution
to the classic n-queens problem.

110
 7.2 Iterating over lists and integer ranges 111

7.2 Iterating over lists and integer ranges

The simplest form of iteration is over the members of a list. For the first
example we shall iterate over the list [a,b,c], writing each member of the
list on a new line using the writeln/1 built-in:
[eclipse 1]: ( foreach(El,[a,b,c]) do writeln(El) ).

a
b
c

El = El
Yes (0.00s cpu)
The user sees only one variable El, which appears to take at each iteration
the value of a different member of the list. However, internally ECLi PSe
turns this iteration into a recursion, so a different variable (El 1, El 2, etc.)
is introduced for each member of the list.
A similar iterator, count/3 is used for running a counter over a range of
integers:
[eclipse 2]: ( count(I,1,4) do writeln(I) ).

1
2
3
4

I = I
Yes (0.00s cpu)
The ECLi PSe iteration construct allows more than one iterator , so one
can iterate over a list and an integer range simultaneously. Thus we can write
out the elements of a list together with a number showing their position in
the list as follows:
[eclipse 3]: ( foreach(El,[a,b,c]), count(I,1,_)
do
writeln(I-El)
).

1 - a
112 Intermezzo: Iteration in ECLi PSe

2 - b
3 - c

El = El
I = I
Yes (0.00s cpu)
When two or more iterators appear on the same level, as in this example,
then they all iterate in step. In effect there is just one iteration, but at
each iteration all the iterators step forward together. In the above example,
therefore, at iteration 1, El takes the value a, whilst I takes the value 1.
At iteration 2, El takes the value b, whilst I takes the value 2. At the last
iteration 3, El takes the value c, whilst I takes the value 3. Because the
iteration stops when the end of the list is reached, we do not need to specify
a maximum for the integer range in the iterator count(I,1, ).
Later we shall encounter iterators at different levels, which is achieved by
writing an iteration construct within the query formed by another iteration
construct.
Coming back to the current example, with iterators at the same level,
we can even use the count iterator to compute the maximum, because it
automatically returns the value of I at the last iteration:
[eclipse 4]: ( foreach(El,[a,b,c]), count(I,1,Max)
do
writeln(I-El)
).

El = El
I = I
Max = 3
Yes (0.00s cpu)
1 - a
2 - b
3 - c
In the above query Max returns the length of the list. Thus we can find
the length of a list by this very construction we just do not bother to
write out the elements:
[eclipse 5]: ( foreach(_,[a,b,c]), count(_,1,Length)
do
true
 7.3 Iteration specifications in general 113

).

Length = 3
Yes (0.00s cpu)

The foreach/2 iterator walks down the list, and the count iterator in-
crements its index at each step. When the end of the list is reached, the
iteration stops, and the final value of the counter is returned as the value of
the variable Length. Curiously, the action taken at each iteration is noth-
ing! The atomic query true simply succeeds without doing anything. In
particular, we do not use the elements of the list or the counters that are
now represented by anonymous variables.

7.3 Iteration specifications in general

An iterator construction has the form

( IterationSpecs do Query )

where Query is any ECLi PSe query. The iteration specifiers IterationSpecs
is a comma-separated sequence of iterators, such as

foreach(El,List), count(I,1,Length)

in the example queries above.

In this section we shall go through the different iterators, exploring how
they can be used to process lists, ranges, compound terms (structures), and
arrays. We shall also see how to pass arguments into an iteration.

7.3.1 Constructing lists

We have seen how the iterator foreach allows us to walk down a list. It
can also be used to construct a list. To see how let us construct a list of
increasing integers, starting at 1. We shall give the length of the list as, say,
4:

[eclipse 6]: ( foreach(El,List), count(I,1,4) do El = I ).

El = El
List = [1, 2, 3, 4]
I = I
Yes (0.00s cpu)
114 Intermezzo: Iteration in ECLi PSe

Thus to construct a list of given length we can use the very same code we
used for measuring the length of a list, except that this time we input the
length, and output the list.
Notice that the variables El and I are not instantiated on completion of
the iteration. Their scope is local to the iteration construct, and even local to
each iteration. This means that we can construct a list of different variables
by simply replacing El = I in the above by true. To make clear these
are different variables, it is necessary to set the ECLi PSe output mode
to output the internal variable names.1 With the new output mode the
behaviour is as follows:
[eclipse 7]: ( foreach(El,List), count(I,1,4) do true ).

** Warning: Singleton local variable El in do-loop

(not used in loop body)
** Warning: Singleton local variable I in do-loop
(not used in loop body)
El = El_80
List = [El_386, El_390, El_394, El_398]
I = I_100
Yes (0.00s cpu)
As usual we can eliminate the warnings by using anonymous variables, but
this has no effect on the output itself:
[eclipse 8]: ( foreach(_,List), count(_,1,4) do true ).

List = [_339, _341, _343, _345]

Yes (0.00s cpu)
Now we are finally prepared to return to the LENGTH program discussed
in the introduction to this chapter. Because we can use the iterators both
for input and output, we could use exactly these iterators to implement the
len/2 predicate which works in both directions to compute the length
of a list, and to build a list of different variables of a specified length. The
resulting program is given in Figure 7.1.
We now have:

[eclipse 9]: len([a,b,c], N).

N = 3
1 To set the output mode one can use set flag(output mode,"VmQP").
 7.3 Iteration specifications in general 115

% len(List, N) :-
% N is the length of the list List
% or List is a list of variables of length N.

len(List, N) :-
( foreach(_,List), count(_,1,N) do true ).

Fig. 7.1 A non-recursive LENGTH program

Yes (0.00s cpu)

[eclipse 10]: len(Ls, 4).

Ls = [_159, _161, _163, _165]

Yes (0.00s cpu)

Finally, in case the maximum is an input (integer or expression), the

iterator for/3 can be used instead of count/3, with the same meaning:

[eclipse 11]: ( for(I,1,5) do write(I), write( ) ).

I = I
Yes (0.00s cpu)
1 2 3 4 5

7.3.2 Passing arguments into and out of iterations

Given an input list of integers, we can output a list of their successors as
follows:

[eclipse 12]: ( foreach(In,[1,2,3,4]),

foreach(Out,OutList)
do
Out is In+1
).

In = In
Out = Out
OutList = [2, 3, 4, 5]
Yes (0.00s cpu)
116 Intermezzo: Iteration in ECLi PSe

The reader should compare it with the MAP program from Figure 5.3 and
query No. 31 from page 84. Just as In and Out in the above query have
scope local to a single iteration, so we can add other local variables in the
body of the query. For example, we can input a list of characters and output
a list of their ASCII successors as follows:
[eclipse 13]: ( foreach(In,[a,b,c]),
foreach(Out,OutList)
do
char_code(In,InCode),
OutCode is InCode+1,
char_code(Out,OutCode)
).

In = In
Out = Out
OutList = [b, c, d]
InCode = InCode
OutCode = OutCode
Yes (0.00s cpu)
(The char_code/2 built-in provides a bi-directional conversion between
the characters and the ASCII codes.) Each variable occurring in a query
inside the iteration has local scope.
It is straightforward to pass a value, such as 5, into an iteration, as follows:
[eclipse 14]: ( foreach(In,[1,2,3]),
foreach(Out,OutList)
do
Out is In+5
).

In = In
Out = Out
OutList = [6,7,8]
Yes (0.00s cpu)
However, we cannot simply use a variable in place of the 5 in the above
query, because ECLi PSe treats any variable in the body of an iteration
construct as local, even if a variable with the same name also occurs outside
the iteration. For example the following behaviour is not really what we
want:
 7.3 Iteration specifications in general 117

[eclipse 15]: Var = 5,

( foreach(In,[1,2,3]),
foreach(Out,OutList)
do
Out is In+Var
).

*** Warning: Singleton local variable Var in do-loop,

maybe param(Var) missing?
instantiation fault in +(1, Var_487, Out_486) in module eclipse
Abort
To pass a (variable) argument into an iteration an extra iterator is used.
It is simply called param. We make the scope of Var global by passing it in
as a parameter, thus:
[eclipse 16]: Var = 5,
( foreach(In,[1,2,3]),
foreach(Out,OutList),
param(Var)
do
Out is In+Var
).

Var = 5
In = In
Out = Out
OutList = [6, 7, 8]
Yes (0.00s cpu)
We can now build a single clause program where the number to be added
to each element is supplied as an argument:
add_const(Increment, InList, OutList) :-
( foreach(In,InList),
foreach(Out,OutList),
param(Increment)
do
Out is In+Increment
).
Any number of parameters can be passed into an iteration. For example
we can increment positive elements by one amount and decrement negative
118 Intermezzo: Iteration in ECLi PSe

elements by a different amount. The resulting program is given in Figure

7.2. Note that here param takes two arguments.

add_const(Increment, Decrement, InList, OutList) :-

( foreach(In,InList),
foreach(Out,OutList),
param(Increment,Decrement)
do
( In >= 0 ->
Out is In+Increment
;
Out is In-Decrement
)
).

Fig. 7.2 An INCREMENT/DECREMENT program

We have then for example:

[eclipse 17]: add_const(3, 5, [2, -3, 5, -2, 4], OutList).

OutList = [5, -8, 8, -7, 7]

Yes (0.00s cpu)
The param iterator can also be used to thread the initial value of a variable
into the loop. Here is a simple example. Consider the clause
inc(N, Input, Output) :-
( foreach(In,Input),
foreach(N:Out,Output),
param(N)
do
Out is In+1
).
Then we have:
[eclipse 18]: inc(3, [4,5,6,7], Output).

Output = [3 : 5, 3 : 6, 3 : 7, 3 : 8]
Yes (0.00s cpu)
 7.3 Iteration specifications in general 119

We shall see this use of param in Section 8.4. Without param(N) we would
get the following behaviour (and a Singleton variables warning during
the compilation time):
[eclipse 19]: inc(3, [4,5,6,7], Output).

Output = [N : 5, N : 6, N : 7, N : 8]
Yes (0.00s cpu)

7.3.3 Iterations within iterations

Let us multiply now all the numbers in one range, say 1..3, by all the numbers
in another range, say 5..9. We can do this using an iteration in the body of
an outer iteration, thus:
[eclipse 20]: ( for(I,1,3)
do
( for(J,5,9),
param(I)
do
K is I*J,
write(K), write( )
)
).

I = I
J = J
K = K
Yes (0.00s cpu)
5 6 7 8 9 10 12 14 16 18 15 18 21 24 27
Notice the use of the param(I) to pass the argument I into the inner itera-
tion.
Because of the need to pass in parameters embedded iterations can be
complicated to write. Therefore a facility is provided to iterate over multiple
ranges using a single iterator multifor/3. The same example as above can
be written:
[eclipse 21]: ( multifor([I,J],[1,5],[3,9])
do
K is I*J,
write(K), write( )
120 Intermezzo: Iteration in ECLi PSe

).

I = I
J = J
K = K
Yes (0.00s cpu)
5 6 7 8 9 10 12 14 16 18 15 18 21 24 27

7.3.4 Iterating over structures

For applications where the number of components is fixed, it is usually more
efficient to use compound terms rather than lists to hold the data. Thus in-
stead of a list of variables [X,Y,Z], we have a term such as p(X,Y,Z). To iter-
ate over the arguments of a term, the appropriate iterator is foreacharg/2:

[eclipse 22]: ( foreacharg(Arg,p(a,b,c)) do writeln(Arg) ).

Arg = Arg
Yes (0.00s cpu)
a
b
c

The efficiency comes from being able to access components directly from
their location in the structure. In the following example we fill a structure
with values, and then access the third component:

[eclipse 23]: functor(Struct,p,3),

( foreach(El, [val1,val2,val3]),
foreacharg(El, Struct)
do
true
),
arg(3, Struct, X).

Struct = p(val1, val2, val3)

El = El
X = val3
Yes (0.00s cpu)
 7.4 Arrays and iterations 121

7.4 Arrays and iterations

InECLi PSe arrays are structures with the specific functor []. To create
them and specify their, possibly multi-dimensional, size, the built-in dim/2
is used. For example, to create a one-dimensional array Array of size 3 we
write:
[eclipse 24]: dim(Array,[3]).

Array = [](_162, _163, _164)

Yes (0.00s cpu)
In turn, the query dim(Array,[3,2]) creates a two-dimensional array Array
of size 3 by 2:
[eclipse 25]: dim(Array,[3,2]).

Array = []([](_174, _175), [](_171, _172), [](_168, _169))

Yes (0.00s cpu)
To assign a value to an array component or to access the value of an array
component one can use the ECLi PSe subscript/3 built-in:
[eclipse 26]: dim(Array,[3,2]), subscript(Array,[1,2],5).

Array = []([](_263, 5), [](_260, _261), [](_257, _258))

Yes (0.00s cpu)

[eclipse 27]: dim(Array,[3]), subscript(Array,[2],a),

subscript(Array,[2], X).

Array = [](_287, a, _289)

X = a
Yes (0.00s cpu)
The components of an array can also be accessed using the customary
subscript notation. For example:
[eclipse 28]: dim(Array,[3,2]), subscript(Array,[1,2],5),
X is Array[1,2] - 2, Y = f(Array[1,2]).

Array = []([](_444, 5), [](_441, _442), [](_438, _439))

X = 3
Y = f([]([](_444, 5), [](_441, _442), [](_438, _439))[1, 2])
122 Intermezzo: Iteration in ECLi PSe

Yes (0.00s cpu)

Note that the value of Y is not f(5) since the argument of f(Array[1,2])
is not evaluated. To assign f(5) to Y we would need to explicitly evalu-
ate f(Array[1,2]) using Z is f(Array[1,2]), Y = f(Z) instead of Y =
f(Array[1,2]).
In the context of arrays the is/2 evaluator is extended to allow us to
access array elements or to transform (a part of) a row or a column into a
list:
[eclipse 29]: A = []([](1, 2),
[](3, 4),
[](5, X)),
El is A[3,2],
Row is A[1, 1..2],
Col is A[2..3, 2].

A = []([](1, 2), [](3, 4), [](5, X))

X = X
El = X
Row = [1, 2]
Col = [4, X]
Yes (0.00s cpu)
We can also use it to transform a subarray into a list of lists:
[eclipse 30]: A = []([](1, 2),
[](3, 4),
[](5, X)),
Sub is A[2..3, 1..2].

A = []([](1, 2), [](3, 4), [](5, X))

X = X
Sub = [[3, 4], [5, X]]
Yes (0.00s cpu)
With arrays we can use the more general iterator foreachelem/2 which
iterates over all the elements of a, possibly multi-dimensional, array. The
query No. 23 from the previous subsection can be rewritten using arrays as:
[eclipse 31]: dim(Array,[3]),
( foreach(El, [val1,val2,val3]),
foreachelem(El, Array)
 7.4 Arrays and iterations 123

do
true
),
X is Array[3].

Array = [](val1, val2, val3)

El = El
X = val3
Yes (0.00s cpu)

However, we can now also build and access embedded components:

[eclipse 32]: dim(Array,[3,2]),

( foreach(El, [el11,el12,el21,el22,el31,el32]),
foreachelem(El, Array)
do
true
),
X is Array[2,2].

Array = []([](el11, el12), [](el21, el22), [](el31, el32))

El = El
X = el22
Yes (0.00s cpu)

A useful variant of foreachelem/2, with one more argument, addition-

ally returns the index of each component. This allows us, for example, to
transpose a matrix quite easily. The program is given in Figure 7.3. It
makes use of the fact that the dim/2 built-in also allows us to retrieve the
sizes of a given array. Using this program matrices of arbitrary sizes can be
transposed.
We can use this program on any 2 dimensional matrix:

[eclipse 33]: Matrix = []([](el11, el12),

[](el21, el22),
[](el31, el32)),
transpose(Matrix,Transpose).

Matrix = []([](el11, el12), [](el21, el22), [](el31, el32))

Transpose = []([](el11, el21, el31), [](el12, el22, el32))
Yes (0.00s cpu)
124 Intermezzo: Iteration in ECLi PSe

% transpose(Matrix, Transpose) :-
% Transpose is a transpose of the matrix Matrix.

transpose(Matrix, Transpose) :-
dim(Matrix,[R,C]),
dim(Transpose,[C,R]),
( foreachelem(El,Matrix,[I,J]),
param(Transpose)
do
subscript(Transpose,[J,I],El)
).

Fig. 7.3 The TRANSPOSE program

7.5 fromto/4: the most general iterator

The iterators introduced so far iterated over single components of the data
structure dealing with one element of the list/range/structure at a time.
In ECLi PSe a more general form of iterator is provided that allows us to
organise more complex forms of iteration and even replace all recursion by
iteration. This iterator, fromto/4, has four arguments:
fromto(First,In,Out,Last).
This iterator works by starting at the first iteration with In = First. The
first iteration computes a value for Out. Now this value is threaded at the
second iteration into In. So at the second iteration the value of In is the
value taken by Out at the first iteration. Subsequently at each iteration the
value of Out is threaded into the value of In at the next iteration. The
iteration stops when Out = Last. The control flow is informally depicted
in Figure 7.4.

fromto(First, In, Out, Last)

Fig. 7.4 The control flow in fromto/4

A simple example is iteration down a list:

[eclipse 34]: ( fromto([a,b,c],[Head | Tail],Tail,[])
do
 7.5 fromto/4: the most general iterator 125

writeln(Head)
).
a
b
c

Head = Head
Tail = Tail
Yes (0.00s cpu)
The reader may note that this is a possible implementation of the foreach/2
iterator. Indeed, all the other iterators can be implemented using fromto/4.
Our informal description of the behaviour of fromto is highly procedural.
However, the behaviour of the fromto/4 construct is more powerful than a
procedural description might suggest. In fact, this same construct can be
used to construct a value of First backwards from the value of Last. In
the following two examples we show fromto/4 using
its first argument as the full list,
its last argument as the full list.
In both of them fromto/4 and foreach/2 iterate in step.
[eclipse 35]: ( fromto([a,b,c],[Head | Tail],Tail,[]),
foreach(El,List)
do
El = Head
).

Head = Head
Tail = Tail
El = El
List = [a, b, c]
Yes (0.00s cpu)

[eclipse 36]: ( fromto([],Tail,[Head | Tail],[a,b,c]),

foreach(El,List)
do
El = Head
).

Tail = Tail
126 Intermezzo: Iteration in ECLi PSe

Head = Head
El = El
List = [c, b, a]
Yes (0.00s cpu)
The second query constructs the list [c,b,a] by putting one at a time
an element, being a new intermediate variable, on the front of an initially
empty list. The actual values of the list elements are unknown until the
iteration terminates, at which point all the intermediate variables, as well
as the output list List, become fully instantiated.
So using the fromto/4 iterator we can reverse a list in a simple and elegant
way. The program is given in Figure 7.5.

% rev(List, Reverse) :-
% Reverse is a reverse of the list List or
% List is a reverse of the list Reverse.

rev(List, Reverse) :-
( fromto([],Tail,[Head | Tail],Reverse),
foreach(El,List) do El = Head ).

Fig. 7.5 A non-recursive REVERSE program

It can be used in both directions:

[eclipse 37]: rev([a,b,c], Reverse), rev(List, Reverse).

Reverse = [c, b, a]
List = [a, b, c]
Yes (0.00s cpu)
The fromto/4 iterator can also be used to iterate down the lists in cases
where more than one element needs to be visible at each iteration. Let
us use this, first, to write a non-recursive version of the program ORDERED
from Figure 3.3 on page 48 that checks whether a list is an -ordered list
of numbers. The program is given in Figure 7.6.
Here the iterator
fromto(List,[This,Next | Rest],[Next | Rest],[_])
iterates down a non-empty list and at each iteration returns the tail of the
remaining list.
 7.5 fromto/4: the most general iterator 127

% ordered(Xs) :- Xs is an =<-ordered list of numbers.

ordered(List) :-
( List = [] ->
true
;
( fromto(List,[This,Next | Rest],[Next | Rest],[_])
do
This =< Next
)
).

Fig. 7.6 A non-recursive ORDERED program

The previous iterators have another limitation that all the different itera-
tors in a single iteration must have the same number of components. Thus
the output list/range/structure must have the same number of components
as the input. This makes it impossible to use the previous iterators to filter
a list, for example, constructing a list of only those elements that pass a
test. To take an input list of integers and output a list of only the positive
integers it is therefore necessary to use the fromto/4 iterator, thus:

[eclipse 38]: ( foreach(El,[2,-3,5,-2,4]),

fromto([],This,Next,PosList)
do
( El >= 0 -> Next = [El | This] ; Next = This )
).

El = El
This = This
Next = Next
PosList = [4, 5, 2]
Yes (0.00s cpu)

The output list is in the reverse order from the input list, because each new
positive integer is added in front of the previous ones. To have the output
come in the same order as the input, the following iteration is required:

[eclipse 39]: ( foreach(El,[2,-3,5,-2,4]),

fromto(PosList,This,Next,[])
128 Intermezzo: Iteration in ECLi PSe

do
( El >= 0 -> This = [El | Next] ; This = Next )
).

El = El
PosList = [2, 5, 4]
This = This
Next = Next
Yes (0.00s cpu)

The fromto/4 iterator is sometimes just used to hold a particular value,

such as the chosen one. Say, we have a list of items with names and values,
and we want the name of the cheapest item:

[eclipse 40]: ( foreach(El,[a:3,b:6,c:2,d:4]),

fromto(none:100,Old,New,Name:_)
do
El = _:ElVal,
Old = _:OldVal,
( ElVal < OldVal -> New = El ; New = Old )
).

Old = Old
New = New
Name = c
ElVal = ElVal
OldVal = OldVal
Yes (0.00s cpu)

Finally, the fromto/4 iterator is often used to check a stopping condition.

Let us copy one list into another until we reach a perfect square (i.e., a
square of an integer):2

[eclipse 41]: ( fromto([3,5,3,4,6,10],[In | Rest],Rest,_),

foreach(In,OutList),
fromto(go,_,Next,stop)
do
2 Here is one way to write perfect square/1 in ECLi PSe :
perfect square(In) :- Sqrt is fix(round(sqrt(In))), In =:= Sqrt*Sqrt.
The sqrt/1 built-in computes the positive square root of its argument (e.g., X is sqrt(9.61)
assigns 3.1 to X), round/1 rounds a real or integer number to the nearest integral value of the
same type (e.g., 2.3 to 2.0, 3 to 3 and 2.6 to 3.0) while fix/1 converts any number to an
integer, rounding towards 0 (e.g., 6.8 to 6, 6.0 to 6 and -7.9 to -7).
 7.6 The n-queens problem 129

( perfect_square(In) ->
Next = stop
;
Next = go
)
).

In = In
Rest = Rest
OutList = [3, 5, 3, 4]
Next = Next
Yes (0.00s cpu)
In this example the first fromto/4 iterator goes through the input list, while
the second fromto/4 iterator checks the stopping condition at each iteration.
This enables the iterations to stop before the input list has been exhausted.

7.6 The n-queens problem

As the final example of an ECLi PSe program that shows the use of iterators
we provide a non-recursive solution to the n-queens problem introduced in
Subsection 6.4.3. Recall that the problem calls for placing n queens on the
n n chess board, where n 3, so that they do not attack each other.

A natural way of encoding the program in ECLi PSe using iterators is

given in Figure 7.7.
In this program the select val/3 predicate is used to generate upon
demand successive integer values in the range [1..Number]. For each col-
umn J these values (held in the variable Qj) are tentatively assigned to the
subscripted variable QueenStruct[J], using the call
subscript(QueenStruct,[J],Qj),
upon which a check is made that the so placed queen does not come under
attack by the queens already placed in the previous columns.
The solution is generated as an array, for example:
[eclipse 43]: queens(QueenStruct, 8).

QueenStruct = [](1, 5, 8, 6, 3, 7, 2, 4)
Yes (0.02s cpu, solution 1, maybe more) ?
This solution corresponds to Figure 7.8.
130 Intermezzo: Iteration in ECLi PSe

% queens(QueenStruct, Number) :- The array QueenStruct

% is a solution to the Number-queens problem.

queens(QueenStruct, Number) :-
dim(QueenStruct,[Number]),
( for(J,1,Number),
param(QueenStruct,Number)
do
select_val(1, Number, Qj),
subscript(QueenStruct,[J],Qj),
( for(I,1,J-1),
param(J,Qj,QueenStruct)
do
QueenStruct[I] =\= Qj,
QueenStruct[I]-Qj =\= I-J,
QueenStruct[I]-Qj =\= J-I
)
).
augmented by the SELECT program of Figure 3.2.

Fig. 7.7 The QUEENS program

7.7 Summary
Iterators are typically used to iterate over data structures. They can also
be used as an alternative for recursion. In many situations the resulting
programs are simpler.
We introduced the iteration construct
( IterationSpecs do Query )
and the different iterators which can be used within the sequence of the itera-
tor specifiers IterationSpecs. In the process we also introduced the arrays
that are declared in ECLi PSe using the dim/2 built-in. The subscript/3
built-in allows us to assign a value to an array component or to test its value.
We discussed in turn the following iterators:3
For iterating over lists:
foreach(El,List) do Query(El).
Iterate Query(El) over each element El of the list List.
3 We also list here the most general versions of the discussed constructs.
 7.7 Summary 131

8
7

6
5
4
3
2
1
a b c d e f g h

Fig. 7.8 One of 92 solutions to the 8-queens problem

For iterating over an integer range:

count(I,MinExpr,Max) do Query(I).
Iterate Query(I) with I increasing at each iteration starting from
MinExpr. Max can be a number, in which case iteration stops when
I = Max, or it can be a variable, in which case it returns the value of
I when iteration is ended by another iterator.
for(I,MinExpr,MaxExpr) do Query(I).
Similar to count above, except MaxExpr must be an integer expression.
for(I,MinExpr,MaxExpr,Increment) do Query(I).
Iterate Query(I) with I ranging from MinExpr to MaxExpr
in increments of Increment.
For passing arguments or variables into an iteration:

( Iterator_1,
...,
Iterator_n,
param(Argument1,...,Variable,...)
do
Query(..., Argument1, ..., Variable, ...).
).

For iterating over one or more integer ranges:

multifor(List,MinList,MaxList) do Query(List).
Iterate Query(List) with List ranging over all combinations of
integers from MinList to MaxList.
132 Intermezzo: Iteration in ECLi PSe

multifor(List,MinList,MaxList,IncrementList) do Query(List).
As above, but with specified increments.
For iterating over compound terms, especially arrays:
foreacharg(X,Term) do Query(X).
Iterate Query(X) with X ranging over all the arguments of the structure
Term.
foreacharg(X,Term,Index) do Query(X,Index).
As before, with Index recording the argument position of X in the
term.
foreachelem(X,Array) do Query(X).
Iterate over all the elements of a multi-dimensional array.
foreachelem(X,Array,Index) do Query(X,Index).
As before, but Index is set to the multi-dimensional index of the
element.
foreachelem(Array,Index) do Query(Index).
As before, but only use the index.
The most general iterator:
fromto(First,In,Out,Last) do Query(In,Out).
Iterate Query(In,Out) starting with In = First, until Out = Last.

7.8 Exercises

Exercise 7.1 Count the number of non-variable elements in a list without

using recursion.

Exercise 7.2 Count the number of non-variable elements in an array with-

out using recursion.

Exercise 7.3 Write a non-recursive version of the the select val/3 pred-
icate from Figure 3.2 on page 46.

Exercise 7.4 Suggest an implementation of the

( count(I,K,L) do Body )
iterator using the fromto/4 iterator.
 8

Top-down search with passive constraints

8.1 Introduction 133

8.2 Solving CSPs using Prolog 134
8.3 Backtracking search in Prolog 135
8.4 Incomplete search 139
8.5 Non-logical variables 146
8.6 Counting the number of backtracks 149
8.7 Summary 153
8.8 Exercises 153

8.1 Introduction

I N THIS CHAPTER we discuss how to solve finite CSPs in Prolog by

means of top-down search. In top-down search one repeatedly tries
to assign a value to each decision variable in turn. During the search
process the decision variables and the values used can be reordered statically
or dynamically. Also, during the search process an assignment of a value
to a variable can be undone through backtracking. Further, the search can
be complete or incomplete depending on whether all possible assignments
are explored during the search. In what follows we discuss how these options
can be implemented in Prolog. Then we shall discuss the facilities offered by
ECLi PSe that allow us to count the number of backtracks, the most widely
used measure of efficiency concerned with top-down search.
Prolog supports only passive constraints, in the sense of Section 3.4. In
other words, the constraints can be used only as tests. This limitation
significantly affects the search process. First, the constraints can be invoked
only when all their variables are fully instantiated. As a result a natural
encoding of a problem has to be sometimes rewritten to avoid run-time
errors. Secondly, some obvious inferences that could reduce the search space
have to be implemented in an ad hoc way, even though in many situations,

133
134 Top-down search with passive constraints

for example in the case of Boolean constraints, each time the same principle
could be invoked. In the subsequent chapters we shall gradually relax these
limitations, first by allowing, in the next chapter, passive constraints without
the risk of run-time errors, and second, starting with Chapter 10, by allowing
active constraints.

8.2 Solving CSPs using Prolog

The most natural way to solve a finite CSP in Prolog is to use a program
that has the following shape:
solve(List):-
declareDomains(List),
search(List),
testConstraints(List).
Here List is a list of the CSP variables. The first atomic goal,
declareDomains(List) sets up the information about the domains of the
variables from List, while the atomic goal search(List) launches the ap-
propriately chosen search process. Often the information about the domains
is put directly into the atomic goal search(List). Finally, the atomic goal
testConstraints(List) tests whether the generated variable instantiations
satisfy the constraints of the considered CSP.
This approach is usually called generate and test and is obviously ex-
tremely inefficient. The reason is that during the search process all possible
variable instantiations are generated and their number is typically huge.
For example, for the SEND + MORE = MONEY problem formalised us-
ing eight decision variables S,E,N,D,M,O,R,Y the corresponding search tree
has 108 leaves and such an approach does not generate any solution in a
reasonable amount of time.
A better approach is to try to test the constraints as soon as possible.
For example, in the case of the SEND + MORE = MONEY problem it is
obviously better to test the constraint S =\= E as soon as the values for S
and E are generated, instead of only after the values for all decision variables
have been generated. Indeed, the combinations of values with S and E equal
are not considered then.
In general such an approach requires an appropriate interleaving of the
execution of the atomic goals search(List) and testConstraints(List)
and is not straightforward to achieve in a generic way. The suspend library
that will be discussed in the next chapter automatically ensures such an
interleaved execution.
 8.3 Backtracking search in Prolog 135

But we are running ahead of ourselves. In this chapter we focus solely on

the search process encoded in the search/1 procedure. Later, in Chapter
11, we shall see how this search process can be naturally combined with
constraint propagation.

8.3 Backtracking search in Prolog

We already discussed backtracking search in Section 6.6. Now we explain
how to program it in Prolog. In what follows we focus on labelling as
the branching method. Recall that labelling splits a finite domain of a
variable into singletons. This does not yet uniquely define the resulting
search procedure since we are still faced with the choice of the order in
which the variables are labelled and the choice in which the values in the
variable domains are selected. Let us discuss now the consequences of both
choices.

8.3.1 The variable ordering

It is perhaps surprising, but the order of the variables can have quite a
dramatic effect on the size of the search tree, even though it cannot change
the number of leaves. Let us take a very simple tree as an example for this
discussion. We have just two variables, X and Y. Suppose X has two possible
values, and Y has four. Assuming there are no constraints on the values of X
and Y, we can draw a complete search tree that finds all the possible pairs of
values for X and Y. In Figure 8.1, we order the variable X before Y. However,
if we label Y before X, we get the tree in Figure 8.2.

X=1 X=2

Y=1 Y=2 Y=3 Y=4 Y=1 Y=2 Y=3 Y=4

Fig. 8.1 A full search tree with two variables

Note that both trees have eight leaves corresponding to the eight possible
pairs of values for X and Y but have different number of internal nodes. The
136 Top-down search with passive constraints

Y=1 Y=2 Y=3 Y=4

X=1 X=1 X=1 X=1

X=2 X=2 X=2 X=2

Fig. 8.2 The search tree with a reversed variable ordering

second tree has more internal nodes, so it is harder to search. A way to

keep the number of internal nodes as small as possible is to label earlier
the variables with fewer choices. If we had four variables with respectively
one, two, three and four alternative values, then by ordering the variables
with the smallest domains first the number of internal nodes in the tree is
1 + (1 2) + (1 2 3) = 9. However, the worst ordering gives 4 + (4
3) + (4 3 2) = 40 internal nodes.
The real payoff from a better variable ordering comes when there are
constraints to be satisfied. To keep the example very simple, we will just
have one constraint, X + Y =:= 6. The program invoking this search is
shown in Figure 8.3 and the search tree explored by this program is depicted
in Figure 8.4.

search(X,Y) :-
member(X, [1,2]),
member(Y, [1,2,3,4]),
X+Y =:= 6.

Fig. 8.3 A two variable problem

The corresponding search behaviour, depth-first search with backtracking,

is automatically provided by the underlying Prolog computation model and
its implementation. The search tree has 11 nodes and all of them have to be
visited to reach the solution at the last leaf of the tree. Now, if the variables
are reordered, 13 nodes have to be visited to reach the solution. This is
shown in Figure 8.5.
 8.3 Backtracking search in Prolog 137

X=1 X=2

Y=1 Y=2 Y=3 Y=4 Y=1 Y=2 Y=3 Y=4

Fig. 8.4 Searching a two variable tree

Y=1 Y=2 Y=3 Y=4

X=1 X=1 X=1 X=1

X=2 X=2 X=2 X=2

Fig. 8.5 Searching another two variable tree

We shall see in Chapter 11 that with active constraints the variable order-
ing can have a tremendous influence on the number of search steps needed
to find a solution, or to prove that none exists.

8.3.2 The value ordering

If we could always choose the right value for each variable, then search would
be unnecessary: we would find a solution first time whenever there was one.
Would that life were so simple!
A surprising fact about different value orderings is that they do not affect
the size of the search tree. In fact, if there is no solution, then the same
number of search nodes needs to be explored whatever value ordering is
chosen. This remains true whatever the constraints are.
Where value orderings become important is for incomplete search. Most
practical problems have search trees so large that they cannot be explored
138 Top-down search with passive constraints

completely, and so value orderings are very important in practice. If the

program is only going to explore part of the search tree, it must explore the
part where there is most likely to be a good solution. We shall return to
this matter in the next section.

8.3.3 Variable and value ordering in Prolog

Controlling the variable and value ordering is very easy in Prolog. Assume
the variables are kept in a list, and each variable is paired with its domain,
thus [X-[1,2],Y-[1,2,3,4]]. A naive labelling procedure simply tries the
values and variables in the order they appear:

naive_search(List) :-
( foreach(Var-Domain,List) do member(Var, Domain) ).

The variable order and the value order are built into this search routine.
The following generic labelling procedure enables the programmer to choose
the next variable and the next value freely. As will be shown in Chapter 11,
in ECLi PSe these choices can be supported by information which is gleaned
during search. This generic search procedure can be written as in Figure
8.6.

% search(List) :-
% Assign values from the variable domains
% to all the Var-Domain pairs in List.

search(List) :-
( fromto(List, Vars, Rest, [])
do
choose_var(Vars, Var-Domain, Rest),
choose_val(Domain, Val),
Var = Val
).

Fig. 8.6 Generic search in ECLi PSe

To complete this labelling procedure we then write specific procedures

for choose var/3 and choose val/2. For example, the behaviour of the
naive search/1 procedure defined above is reproduced by encoding the
choose var/3 and choose val/2 predicates as in Figure 8.7.
 8.4 Incomplete search 139

% choose_var(List, Var, Rest) :-

% Var is a member of List and
% Rest contains all the other members.

choose_var(List, Var, Rest) :- List = [Var | Rest].

% choose_val(Domain, Val) :- Val is a member of Domain.

choose_val(Domain, Val) :- member(Val, Domain).

Fig. 8.7 Naive variable and value choice

An important point to note is that choose var/3 does not introduce any
choice points, because only one variable ordering is needed to carry out a
complete search. However choose val/2 does introduce choice points: in
fact the search tree is completely defined by the set of value choices at each
node of the search tree.

8.4 Incomplete search

In this section we shall assume that using some heuristics the domains of
the variables have already been ordered so that the better values appear
earlier. Then it is natural to ensure that an incomplete search will try values
appearing earlier in the variable domains.
The simplest form of incomplete search is to try only a limited number of
alternatives for each variable. This can be achieved by selecting a limited
number of domain values, and choosing only from those, as shown in the
program in Figure 8.8. For this purpose we specify a maximum number N of
values to be considered in each domain, and ensure that on backtracking the
search only tries at most the first N values from each domain. This is done
by using the select best/3 procedure which selects the appropriate values
in each domain. The atomic query select best(N, Domain, BestList)
succeeds only once and leaves no choice points. Consequently, if N = 1,
then no backtracking arises. The resulting search method is called then
greedy search.
The extra argument N in choose val/3 must be passed into the generic
search procedure from Figure 8.6 as a parameter.
Recall from Section 6.6 that the most general method of limiting the
search is obtained by allocating credit to each value choice, giving more
140 Top-down search with passive constraints

% choose_val(N, Domain, Val) :-

% Val is one of the N "best" values in Domain.

choose_val(N, Domain, Val) :-

select_best(N, Domain, BestList),
member(Val, BestList).

select_best(N, Domain, BestList) :-

( N >= length(Domain) ->
BestList = Domain
;
length(BestList, N),
append(BestList, _, Domain)
).

Fig. 8.8 Selecting only the N best values

credit to values which seem more likely to be a good choice. This credit
is then available for the search subtree which remains after choosing that
value for that variable. To program this procedure we modify accordingly
the generic search procedure from Figure 8.6. The resulting generic credit
search procedure is shown in Figure 8.9.
The difference between this generic search procedure and the one from
Figure 8.6 is that during the iteration down the variable list the credit is
being maintained. The initial credit equals Credit and in each loop iteration
is reduced from the current credit CurCredit to the new credit NewCredit.
This reduction takes place during the value selection to the variable Var,
in the call choose_val(Domain, Val, CurCredit, NewCredit). The new
credit NewCredit is then available for the remainder of the search concerned
with the list Rest.
This search/2 procedure needs to be completed by choosing the
choose var/3 procedure and the share credit/3 procedure that manages
the credit allocation. As an example of the latter let us encode the pro-
gram in Figure 8.8 as a credit allocation procedure. The corresponding
share credit/3 procedure is shown in Figure 8.10.
In this program the iteration is controlled by the amount of credit left.
The initial credit equals N and in each loop iteration is reduced from the
current credit CurCredit to the new credit NewCredit by 1, or to 0 if no
 8.4 Incomplete search 141

% search(List, Credit) :-
% Search for solutions with a given Credit.

search(List, Credit) :-
( fromto(List, Vars, Rest, []),
fromto(Credit, CurCredit, NewCredit, _)
do
choose_var(Vars, Var-Domain, Rest),
choose_val(Domain, Val, CurCredit, NewCredit),
Var = Val
).

choose_val(Domain, Val, CurCredit, NewCredit) :-

share_credit(Domain, CurCredit, DomCredList),
member(Val-NewCredit, DomCredList).

Fig. 8.9 Generic credit search

elements are left. The initial credit N is used to select the first N elements
of the domain Domain (or all the elements if the domain has fewer than N
elements). During the iteration the list DomCredList is built by assigning
to each selected element in turn the same, initial, credit N. The latter is
achieved by using the param(N) iterator.
We have then for example:
[eclipse 1]: share_credit([1,2,3,4,5,6,7,8,9], 5, DomCredList).

DomCredList = [1 - 5, 2 - 5, 3 - 5, 4 - 5, 5 - 5]
Yes (0.00s cpu)
[eclipse 2]: share_credit([1,2,3,4], 5, DomCredList).

DomCredList = [1 - 5, 2 - 5, 3 - 5, 4 - 5]
Yes (0.00s cpu)
Because we keep allocating to the selected elements the same, initial,
credit, in the original search/2 procedure from Figure 8.9 the current credit
CurCredit remains equal to the initial credit Credit, which ensures the
correct operation of each call
share_credit(Domain, CurCredit, DomCredList).
In particular the query
142 Top-down search with passive constraints

% share_credit(Domain, N, DomCredList) :-
% Admit only the first N values.

share_credit(Domain, N, DomCredList) :-
( fromto(N, CurCredit, NewCredit, 0),
fromto(Domain, [Val|Tail], Tail, _),
foreach(Val-N, DomCredList),
param(N)
do
( Tail = [] ->
NewCredit is 0
;
NewCredit is CurCredit - 1
)
).

Fig. 8.10 Admit the best N values

search([X-[1,2,3,4,5,6,7,8,9], Y-[1,2,3,4], Z-[1,2,3,4]], 5)

generates 80 solutions for (X,Y,Z), since with the above share_credit/3

procedure the admissible values are 1,2,3,4,5 for X and 1,2,3,4 for Y and
for Z.
We used this example of credit allocation because it is familiar to the
reader, but it is not a very natural use of credit-based search. A more
natural example of credit allocation is to allocate half the credit to the first
value in the domain, half of the remaining credit to the second value and so
on. When only one unit of credit remains, it is allocated to the next element
in the domain and no more alternatives are explored for this variable. This
focusses the search effort on the most promising part of the search tree. The
code is shown in Figure 8.11.
The expression fix(ceiling(CurCredit/2)), which appears in the code,
returns the smallest integer greater than or equal to CurCredit/2. In this
program the iteration is also controlled by the amount of credit left. How-
ever, in contrast to the share_credit/3 procedure of Figure 8.10, in each
loop iteration the current credit can now be reduced by a larger amount than
1. Also, the allocated credit is now diminishing. As a result the iteration
can terminate without retaining in DomCredList all domain values even if
the domain Domain has fewer elements than the initial credit N.
 8.4 Incomplete search 143

% share_credit(Domain, N, DomCredList) :-
% Allocate credit N by binary chop.

share_credit(Domain, N, DomCredList) :-
( fromto(N, CurCredit, NewCredit, 0),
fromto(Domain, [Val|Tail], Tail, _),
foreach(Val-Credit, DomCredList)
do
( Tail = [] ->
Credit is CurCredit
;
Credit is fix(ceiling(CurCredit/2))
),
NewCredit is CurCredit - Credit
).

Fig. 8.11 Allocating half remaining credit to next best value

For example, we have:

[eclipse 3]: share_credit([1,2,3,4,5,6,7,8,9],

5, DomCredList).

DomCredList = [1 - 3, 2 - 1, 3 - 1]
Yes (0.00s cpu)

[eclipse 4]: share_credit([1,2,3,4,5,6,7,8,9],

1000, DomCredList).

DomCredList = [1 - 500, 2 - 250, 3 - 125, 4 - 63,

5 - 31, 6 - 16, 7 - 8, 8 - 4, 9 - 3]
Yes (0.00s cpu)

As a result, with the above share_credit/3 procedure the query

search([X-[1,2,3,4,5,6,7,8,9], Y-[1,2,3,4], Z-[1,2,3,4]], 5)

now generates only five solutions for (X,Y,Z).
To further clarify this form of credit allocation consider the query

search([X-[1,2,3], Y-[1,2,3], Z-[1,2,3]], 8)

144 Top-down search with passive constraints

that produces eight solutions for (X,Y,Z):

(1,1,1), (1,1,2), (1,2,1), (1,3,1),

(2,1,1), (2,2,1), (3,1,1), (3,2,1),

These solutions correspond to the explored leaves in the search tree de-
picted in Figure 8.12. The explored part of the tree is put in bold.

4 2 2

2 1 1 1 1 1 1

Fig. 8.12 Credit based search with initial credit 8

The total credit allocated to all the values at a search node is equal to the
available credit N. Hence it follows that the total credit allocated throughout
the search tree is equal to the initial input credit at the top of the search
tree. Consequently, this form of credit sharing gives the programmer precise
control over the maximum number of nodes explored.
Another way to use credit is as a measure of distance from the preferred
left-hand branch of the search tree. This approach has been called limited
discrepancy search. The same amount of credit is allocated to the first
value as was input to the search node. An amount of credit one less than the
input is allocated to the second value; one less again to the third value; and
so on. The total credit allocated to alternative values, in this case, exceeds
the input credit. This share credit/3 procedure is shown in Figure 8.13.
So, as in the case of the share_credit/3 procedure from Figure 8.10 on
page 142, the same elements are selected. However, instead of the same
credit, N, now the diminishing credit, CurCredit, is allocated to these ele-
ments.
For example we now have:

[eclipse 5]: share_credit([1,2,3,4,5,6,7,8,9],

5, DomCredList).
 8.4 Incomplete search 145

% share_credit(Domain, N, DomCredList) :-
% Allocate credit N by discrepancy.

share_credit(Domain, N, DomCredList) :-
( fromto(N, CurrCredit, NewCredit, -1),
fromto(Domain, [Val|Tail], Tail, _),
foreach(Val-CurrCredit, DomCredList)
do
( Tail = [] ->
NewCredit is -1
;
NewCredit is CurrCredit-1
)
).

Fig. 8.13 Limited discrepancy credit allocation

DomCredList = [1 - 5, 2 - 4, 3 - 3, 4 - 2, 5 - 1, 6 - 0]
Yes (0.00s cpu)

[eclipse 6]: share_credit([1,2,3,4], 5, DomCredList).

DomCredList = [1 - 5, 2 - 4, 3 - 3, 4 - 2]
Yes (0.00s cpu)
Because of this difference the query
search([X-[1,2,3,4,5,6,7,8,9], Y-[1,2,3,4], Z-[1,2,3,4]], 4).
generates now 33 solutions for (X,Y,Z).
To further clarify the limited discrepancy credit allocation assume the
initial credit of 1 and consider the query
search([X-[1,2], Y-[1,2], Z-[1,2], U-[1,2], V-[1,2]], 1).
It produces six solutions for (X,Y,Z,U,V):
(1,1,1,1,1), (1,1,1,1,2), (1,1,1,2,1),
(1,1,2,1,1), (1,2,1,1,1), (2,1,1,1,1).
These solutions correspond to the explored leaves in the search tree depicted
in Figure 8.14. The explored part of the tree is put in bold.
146 Top-down search with passive constraints

Fig. 8.14 Limited discrepancy search tree

To complete the introduction to incomplete search we should mention

bounded backtrack search, and time limited search. These approaches do
not attempt to spread the search around the search tree. They explore
the tree strictly depth first and left-to-right, but simply stop search when
a certain limit is reached. Both are supported by the built-in search/6
predicate in ECLi PSe , which we will discuss in Chapter 11.

8.5 Non-logical variables

Occasionally one needs to resort to counting on the system level, for example
to compute the number of successful computations, nodes in a computation
tree, the number of times a query succeeds, etc. To deal with such matters
ECLi PSe offers non-logical variables. In general, non-logical variables
are atoms (so start with a lower case letter), which distinguishes them from
the customary ECLi PSe variables.
ECLi PSe provides four built-ins that allow us to manipulate non-logical
variables:

setval/2,
incval/1,
decval/1,
getval/2.

setval/2 assigns a value to a non-logical variable. In turn, incval/1

increments the value of the non-logical variable by 1, while decval/1 decre-
ments the value of the non-logical variable by 1. In both cases the current
value of the variable has to be an integer. Otherwise a run-time error arises.
Finally, getval/2 either assigns the value of a non-logical variable to a
 8.5 Non-logical variables 147

logical variable (this is the only way the values of non-logical variables can
be accessed) or tests whether a non-logical variable has a given value.
Thus we have for example:

[eclipse 7]: N is 3, setval(count,N),

incval(count), getval(count, M).

N = 3
M = 4
Yes (0.00s cpu)

[eclipse 8]: setval(step,a), getval(step,a).

Yes (0.00s cpu)

[eclipse 9]: setval(step,a), getval(step,b).

No (0.00s cpu)

The usefulness of non-logical variables stems from the fact that their val-
ues persist through backtracking. For example we have:

[eclipse 10]: setval(count,0),

( incval(count), fail
;
true
),
getval(count, N).

N = 1
Yes (0.00s cpu)

As a slightly less trivial example let us count now the number of times a
query succeeds. The appropriate program is given in Figure 8.15.
It is useful to see that the execution of the query succeed(Q, N) does
not instantiate any of the variables of the query Q. We have for example:

[eclipse 11]: succeed(X = a; X = b; X = c, N).

X = X
N = 3
Yes (0.00s cpu)
148 Top-down search with passive constraints

% succeed(Q, N) :-
% N is the number of times the query Q succeeds.

succeed(Q, N) :-
( setval(count,0),
Q,
incval(count),
fail
;
true
),
getval(count,N).

Fig. 8.15 Counting the number of successes

Of course, if the query Q diverges or succeeds infinitely often, the query

succeed(Q, N) diverges.
Finally, non-logical variables persist across the clauses. To clarify this
point and to show usefulness of this feature consider the problem of checking
whether a query succeeds exactly once. To this end we could run the query
succeed(Q, 1). This way, however, we let the computation continue after
it is clear that more than one solution exists. Moreover, this query can
diverge.
A better solution is to use the program given in Figure 8.16 and run
the query one(Q). During the execution of this query the non-logical vari-
able x is introduced and modified during the execution of the atomic query
not(twice(Q)) but is tested, using getval(x,1), only upon termination of
this query. The execution of the query one(Q) does not instantiate any of
the variables of the query Q.
The non-logical variables are ideal for simple programs but for large pro-
grams there is a risk of the same non-logical variable name being used for
different purposes, and therefore being updated by different procedures. So
the fact that these variables persist across the clauses can lead to erroneous
outcomes. For example, if count is modified during the execution of the
query Q, then the count reported by the call succeed(Q, N) can be wrong,
for example:

[eclipse 12]: succeed(setval(count,7), N).

 8.6 Counting the number of backtracks 149

% one(Q) :- The query Q succeeds exactly once.

one(Q) :-
not(twice(Q)), % The query Q succeeds at most once.
getval(x,1). % The query Q succeeds at least once.

% twice(Q) :- The query Q succeeds at least twice.

twice(Q) :-
setval(x,0),
Q,
incval(x),
getval(x,2), !.

Fig. 8.16 Checking that a query succeeds exactly once

N = 8
Yes (0.00s cpu)
To avoid variable clashes of this kind, ECLi PSe offers a special global vari-
able called a shelf whose name is automatically generated by the system.
Each shelf has a different name and so clashes are avoided. As our intention
is to focus on the basic concepts of ECLi PSe we do not pursue this topic
further.

8.6 Counting the number of backtracks

We now use the non-logical variables to count the number of backtracks,
which is the most widely used measure of search effort required to find a
solution, or to find all solutions. However, this can be done in alternative
ways which we illustrate with the following simple example:
[eclipse 13]: member(X, [1,2]), member(Y, [1,2]), X+Y =:= 3.

X = 1
Y = 2
Yes (0.00s cpu, solution 1, maybe more) ? ;

X = 2
Y = 1
150 Top-down search with passive constraints

Yes (0.00s cpu, solution 2, maybe more) ? ;

No (0.00s cpu)

The search behaviour associated with this query is illustrated in Figure

8.17.

X=1 X=2

1 2

Y=1 Y=2 Y=2

Y=1

3 4 5 6

Fig. 8.17 Search and backtracking

The number of upward arrows in this figure is six, and this is one measure
of the number of backtracks. The backtracking measure that is most often
used, however, only counts four backtracks in this search behaviour.
The reason is that when a failure lower in the search tree causes back-
tracking to a point several levels higher in the tree, this is usually counted as
a single backtrack. Consequently the upward arrows from node 4 to node 0
only count as a single backtrack. Similarly, the final failure at node 6 which
completes the search is only counted as a single backtrack.
To count backtracks according to these two measures, we augment our
generic search procedure from Figure 8.6 with three atomic goals: to ini-
tialise the backtrack count, to count backtracks, and to retrieve the final
count. The augmented generic search procedure is shown in Figure 8.18.
The backtrack information is recorded in non-logical variables and is
maintained through backtracking. The implementation of the backtracking
predicates that includes each backtrack step (yielding a total of six back-
tracks in the previous toy search example) is shown in Figure 8.19. In this
case we simply increment the backtracks variable on backtracking. The
predicate count_backtracks is implemented using the auxiliary predicate
on backtracking that takes a query as an argument and calls it on back-
tracking. Subsequently it fails so that backtracking continues.
 8.6 Counting the number of backtracks 151

% search(List, Backtracks) :-
% Find a solution and count the number of backtracks.

search(List, Backtracks) :-
init_backtracks,
( fromto(List, Vars, Rest,[])
do
choose_var(Vars, Var-Domain, Rest),
choose_val(Domain, Val),
Var = Val,
count_backtracks
),
get_backtracks(Backtracks).

Fig. 8.18 Generic search with backtrack count

init_backtracks :-
setval(backtracks,0).

get_backtracks(B) :-
getval(backtracks,B).

count_backtracks :-
on_backtracking(incval(backtracks)).

on_backtracking(_). % Until a failure happens do nothing.

% The second clause is entered
on_backtracking(Q) :- % on backtracking.
once(Q) % Query Q is called, but only once.
fail. % Backtracking continues afterwards.

Fig. 8.19 Predicates implementing backtrack counting

The implementation of the count backtracks predicate that counts an

uninterrupted sequence of backtracking steps as one backtrack (the clever
count backtracks) is more complicated. We provide in Figure 8.20 two
implementations. The first one uses the on backtracking predicate and is
152 Top-down search with passive constraints

more clear, though less efficient. The second one implements the counting
process directly.

% First implementation

count_backtracks :-
setval(single_step,true).
count_backtracks :-
on_backtracking(clever_count).

clever_count :-
( getval(single_step,true) ->
incval(backtracks)
;
true
),
setval(single_step, false).

% Second implementation

count_backtracks :-
setval(single_step,true).
count_backtracks :-
getval(single_step,true),
incval(backtracks),
setval(single_step,false).

Fig. 8.20 Two implementations of clever count backtracks

In Chapter 11 we shall use the implementation of the backtrack counting

based on the clever count_backtracks predicate to assess various ways
of implementing search. In the presence of ECLi PSe libraries that support
constraint programming the number of backtracks differs from the case of
Prolog programs. The reason is that the size of the search trees is then
different due to the constraint propagation process that we shall discuss
extensively in Chapter 10.
 8.7 Summary 153

8.7 Summary
In this chapter we discussed how finite CSPs can be solved in Prolog by
means of top-down search. This brought us to a discussion of complete and
incomplete forms of top-down search. For the complete top-down search we
discussed the impact of the variable ordering and the value ordering. For the
incomplete top-down search we focussed on the implementations of various
forms of credit-based search, including the limited discrepancy search.
We also discussed ECLi PSe facilities that allow us to count the number
of backtracks. They consist of non-logical variables and the following four
built-ins using which we can manipulate them:

setval/2,
incval/1,
decval/1,
getval/2.

The distinguishing feature of non-logical variables is that their values persist

through backtracking.

8.8 Exercises

Exercise 8.1 Count the number of variable elements in an array without

using recursion or the fromto/4 built-in, but using the non-logical variables
instead.

Exercise 8.2 Write a predicate all solutions(Query, List) that re-

turns all the solutions to a query as a list. For example if p/2 is defined
by:

p(a,Y) :- q(Y).
p(b,Y) :- r(Y).
q(1).
q(2).
r(1).

then the query all solutions(p(X,Y), List) should return

List = [p(a, 1), p(a, 2), p(b, 1)].

Hint. Use the non-logical variables.

154 Top-down search with passive constraints

Exercise 8.3 The naive search/1 predicate from page 138 can be alter-
natively defined by
naive_search1(Vars, Vals) :-
( foreach(V,Vars), param(Vals) do member(V,Vals) ).
It searches for all possible assignments of the values to the variables. For
example the query naive search1([X,Y,Z], [1,2]) returns
X=1, Y=1, Z=1
and then the following on backtracking
X=1, Y=1, Z=2,
X=1, Y=2, Z=1,
X=1, Y=2, Z=2,
etc.
Modify naive search1/2 to a predicate naive search2/2 that displays
its search tree using the names v1, v2, v3 and so on for the variables. For
example the query naive search2([X,Y,Z], [1,2]), fail should pro-
duce the following output:
v0 = 1 v1 = 1 v2 = 1
v2 = 2
v1 = 2 v2 = 1
v2 = 2
v0 = 2 v1 = 1 v2 = 1
v2 = 2
v1 = 2 v2 = 1
v2 = 2
Hint. Use the non-logical variables and the on backtracking/1 predicate
defined in Figure 8.19 on page 151.
 9

The suspend library

9.1 Introduction 155

9.2 Solving CSPs and COPs using ECLi PSe 156
9.3 Introducing the suspend library 157
9.4 Core constraints 160
9.5 User defined suspensions 167
9.6 Generating CSPs 170
9.7 Using the suspend library 175
9.8 Summary 180
9.9 Exercises 181

9.1 Introduction

I N THIS CHAPTER we introduce the first ECLi PSe library, called

suspend. It allows us to implement constraints. This library lacks
the important facility of constraint propagation (briefly mentioned in
Chapter 6) and consequently is of little use for solving computationally
intensive problems. On the other hand, it is used to implement all the
other, more sophisticated, constraint solvers in ECLi PSe and is an excellent
pedagogical tool for introducing constraints in the context of ECLi PSe .
The suspend library still supports only passive constraints and as a result
is only the first step towards realising constraint programming. It improves
upon the computation model of Prolog by automatically preventing the oc-
currence of run-time errors for arithmetic constraints, by delaying them until
all their variables are fully instantiated. This modified computation mecha-
nism is provided for a larger class of constraints that also includes Boolean
constraints. As we shall see, this leads to a more flexible programming style
in the presence of constraints.
In particular, using the suspend library we can easily generate CSPs and

155
156 The suspend library

solve them in a natural way. How this is done is the topic of the current
chapter.

9.2 Solving CSPs and COPs using ECLi PSe

Let us begin by explaining how CSPs are solved in presence of ECLi PSe
libraries. Each such ECLi PSe program has the following shape:
:- lib(my_library).

solve(List):-
declareDomains(List),
generateConstraints(List),
search(List).
The first line loads the appropriate library, here my_library. Next, List
is a list of the CSP variables. The atomic goal declareDomains(List)
generates the domains of the variables from List and the atomic goal
generateConstraints(List) generates the desired constraints. Finally,
the atomic goal search(List) launches the appropriately chosen search
process that results in solving the generated CSP (or establishing its incon-
sistency).
So we use here a different approach than in Chapter 8 where we discussed
how to solve CSPs using Prolog. Indeed, the constraints are now generated
first and only then the search is launched. This difference is fundamental
and in the presence of appropriate libraries supporting constraint processing
it can lead to substantial gains in computing time.
To solve a COP we also need to generate the cost function. This is often
done together with the generation of constraints. So conceptually we use
then an ECLi PSe program that looks as follows:
:- lib(my_library).

solve(List):-
declareDomains(List),
generateConstraints_and_Cost(List, Cost),
search(List, Cost).
The ECLi PSe system provides a collection of libraries that can be used to
generate and solve CSPs and COPs. Each such library offers the user a set of
built-ins that deal with specific constraints (for example linear constraints)
and specific methods used to solve them (for example branch and bound).
 9.3 Introducing the suspend library 157

In ECLi PSe each library, say my_library, can be accessed in three dif-
ferent ways.

By prefixing an atomic query with its name, for example

suspend:(2 < Y+1).

This is in particular needed when a reference is made to a predicate, like
<, that is already present in Prolog or in another library;
By loading it by means of the query lib(my_library);
By loading it using the declaration
:- lib(my_library).
at the beginning of the program.

9.3 Introducing the suspend library

Let us focus now on the suspend library. To properly understand its useful-
ness let us return to the Prolog approach to arithmetic discussed in Chapter
3. As already mentioned, its drawback is that when certain atomic goals are
not sufficiently instantiated a run-time error arises. One example of this is
[eclipse 1]: 2 < Y+1, Y = 3.
instantiation fault in +(Y, 1, _173)
Abort
This makes it difficult to generate CSPs in Prolog. The suspend library
allows us to overcome this difficulty by enabling certain atomic goals to
be delayed until the variables have become sufficiently instantiated for its
evaluation to go ahead. So using the suspend library we can suspend the
relevant query in the above example until it is ready for evaluation:
[eclipse 2]: suspend:(2 < Y+1), Y = 3.

Y = 3
Yes (0.00s cpu)
So in the presence of the suspend library a more complex computation
mechanism is used than in Prolog. In Prolog in each query the leftmost
atomic goal is selected. Now in each query the leftmost atomic goal that
is either non-arithmetic (that is, does not involve an arithmetic comparison
predicate) or ground is selected. We can make explicit how the Prolog selec-
tion rule is changed to handle delayed queries by writing a meta-interpreter
that does it.
158 The suspend library

We start by revising the META INTERPRETER program from Figure 4.4 on

page 72 as follows:

solve(true) :- !.
solve((A,B)) :- !, solve(A), solve(B).
solve(H) :- rule(H, Body), solve(Body).

rule(A, B) :-
functor(A,F,N),
is_dynamic(F/N),
clause(A, B).
rule(A, true) :- A.

The definition has been changed to use rule/2 instead of clause/2. The
is dynamic(F/N) built-in tests whether the predicate F/N has been declared
as dynamic (see page 70). We obtain this way a meta-interpreter for pro-
grams that use built-ins, for example the comparison predicates. As in the
programs SOLVE and SOLVE2 from Figures 4.4 and 4.5 the handling of built-
ins is just shifted to the system level.
The above meta-interpreter always chooses the leftmost atomic goal to
solve next. The modified meta-interpreter is given in Figure 9.1. Naturally,
the suspend library is not really implemented in this way.
To enable an atomic goal to be delayed (in case it is arithmetic and not
ground) we introduce two extra arguments to pass the delayed goals through
each call. Further, the new meta-interpreter has an extra clause that delays,
through a call of the postpone/1 predicate, an atomic goal if it is not ready
to be solved immediately. This is a non-ground atomic goal of the form
suspend:A. Such a goal is added to the currently delayed goal.
In the last clause for solve/3, after unifying the atomic goal H with the
head of a program clause, the meta-interpreter tries to solve the current
delayed goal, SuspIn, before solving the body Body. This is done in the
same way as solving the input query, using the call solve(SuspIn, true,
Susp2).
Finally, to handle suspend-qualified atomic goals a new rule is added to
the definition of rule/2. This clause is needed only when the suspend
library is not available or is not loaded.
This meta-interpreter now delays non-ground suspend-qualified goals as
required:

[eclipse 3]: solve((2 < Y + 1, Y = 3), true, SuspOut).

instantiation fault in +(Y, 1, _254)
 9.3 Introducing the suspend library 159

solve(true, Susp, Susp) :- !.

solve((A,B), SuspIn, SuspOut) :- !,
solve(A, SuspIn, Susp2),
solve(B, Susp2, SuspOut).
solve(A, Susp, (A,Susp)) :-
postpone(A), !.
solve(H, SuspIn, SuspOut) :-
rule(H, Body),
solve(SuspIn, true, Susp2),
solve(Body, Susp2, SuspOut).

rule(A, B) :-
functor(A,F,N),
is_dynamic(F/N),
clause(A, B).
rule(suspend:A, true) :- !, A.
rule(A, true) :- A.

postpone(suspend:A) :- not ground(A).

Fig. 9.1 The META INTERPRETER program for the suspend library

Abort

[eclipse 4]: solve((suspend:(2 < Y + 1),Y = 3), true, SuspOut).

Y = 3
SuspOut = true
Yes (0.00s cpu)

In general, a query solve(A, SuspIn, SuspOut) succeeds if the query A

succeeds using the new selection rule of the atomic goals. The current de-
layed goal is SuspIn and upon termination the new delayed goal is SuspOut.
The dynamic reordering of the atomic goals made provided through the
suspend library is more powerful than one might expect. A striking example
will be provided at the end of Subsection 9.4.2.
160 The suspend library

9.4 Core constraints

ECLi PSe supports a number of core constraints that are available in all
the constraint solvers where they make sense. These constraints comprise:

Boolean constraints which are the Boolean connectives neg/1, and/2,

or/2 and =>/2 (implies), the last three of which can also be written in
the infix form,
arithmetic constraints which are built using the arithmetic comparison
predicates and the operations considered in Section 3.2,
variable declarations (numeric ranges and numeric type declarations):
they are viewed as unary constraints, and
so-called reified constraints.

The programmer can use these constraints freely to model the problem,
and then select which constraint solver(s) to send them to, when experiment-
ing with algorithms for solving the problem. In particular, these constraints
are available in the suspend library.

9.4.1 Boolean constraints

In ECLi PSe the Boolean constants are represented as 0 for false and 1 for
true. Thus negation can be implemented as subtraction from 1, conjunction
can be implemented as minimum, and so on.
Let us illustrate the suspend behaviour on disjunction or:

[eclipse 5]: suspend:(X or Y), X = 0.

X = 0
Y = Y

Delayed goals:
suspend : (0 or Y)
Yes (0.00s cpu)

So the atomic goal X or Y got only partially instantiated X got instanti-

ated to 1 and the resulting atomic goal, 0 or Y, got delayed, so suspended.
This clarifies the displayed information.
This behaviour is not very powerful. The implementation of or in the
suspend library has failed to infer that Y must take the value 1. In general,
the delayed query is selected only once all its variables are instantiated. This
can lead to a success:
 9.4 Core constraints 161

[eclipse 6]: suspend:(X or Y), X = 0, Y = 1.

X = 0
Y = 1
Yes (0.00s cpu)
or to a failure:
[eclipse 7]: suspend:(X or Y), X = 0, Y = 0.

No (0.00s cpu)
The same holds for all the core constraints: nothing is done until the query
is completely instantiated (i.e. becomes ground).
To implement Boolean constraints in Prolog we need to follow a different
approach and list the true facts. For example the or constraint would be
defined by three facts:
1 or 1.
1 or 0.
0 or 1.
Then the query X or Y, X = 0, Y = 1 succeeds as well, but, in contrast to
the approach based on the suspend library, the computation now involves
backtracking.
In general, if a core constraint becomes fully instantiated it is either
deleted (when it succeeds) or it causes a failure (when it fails) which can
trigger backtracking, as in the case of customary Prolog computations.

9.4.2 Arithmetic constraints

The six arithmetic comparison predicates introduced in Section 3.3 are avail-
able as core constraints. Consequently, in the presence of the suspend li-
brary they can be used to model CSPs. So as before these predicates can
be used for testing:
[eclipse 8]: suspend:(1+2 =:= 3).

Yes (0.00s cpu)

[eclipse 9]: suspend:(1+2 > 2+3).

No (0.00s cpu)
but also for generating constraints in the form of delayed (atomic) goals:
162 The suspend library

[eclipse 10]: suspend:(2+4 > Y), suspend:(Y > 2).

Y = Y

Delayed goals:
suspend : (2 + 4 > Y)
suspend : (Y > 2)
Yes (0.00s cpu)
and for solving them through an instantiation:
[eclipse 11]: suspend:(2+4 > Y), suspend:(Y > 2), Y = 5.

Y = 5
Yes (0.00s cpu)
In particular, thanks to this delay mechanism we can first generate con-
straints and then solve them by a systematic generation of the candidates
values for the relevant variables.
Once the suspend library is loaded, these arithmetic constraints can be
written in a more compact way, using the $ prefix (though the equality and
disequality are written differently). So we have six arithmetic comparison
constraints:
less than, written as $<,
less than or equal, written as $=<,
equality, written as $=,
disequality, written as $\=,
greater than or equal, written as $>=,
greater than, written as $>.
In contrast to the corresponding original arithmetic comparison predi-
cates,
<, =<, =:=, =\=, =\=,>=, >,
they can be used in queries with variables, like in
[eclipse 12]: 1 + 2 $= Y.

Y = Y

Delayed goals:
suspend : (1 + 2 =:= Y)
 9.4 Core constraints 163

Yes (0.00s cpu)

This shows that once the suspend library is loaded 1 + 2 $= Y is indeed
a shorthand for suspend:(1 + 2 =:= Y). So, as long as only the suspend
library is loaded in ECLi PSe , there is no need to explicitly module qualify
each constraint that refers to an arithmetic comparison predicate.
The following example should clarify the difference between =/2, $=/2,
and =:=/2:
[eclipse 13]: X = 1 + 2.

X = 1 + 2
Yes (0.00s cpu)
[eclipse 14]: X $= 1 + 2.

X = X

Delayed goals:
suspend : (X =:= 1 + 2)
Yes (0.00s cpu)
[eclipse 15]: X =:= 1 + 2.
instantiation fault in X =:= 3 in module eclipse
Abort
In particular, the query X $= 1 + 2 is delayed, while its Prolog counterpart
X =:= 1 + 2, as already explained in Section 3.3, yields a run-time error.
A non-trivial example showing that the suspend library allows us to solve
more queries than Prolog is the following one. Let us modify the QUICKSORT
program of Figure 3.4 in Section 3.3 by treating the arithmetic comparison
operators as constraints, so by using the following modified definition of the
part/4 predicate:
part(_, [], [], []).
part(X, [Y|Xs], [Y|Ls], Bs) :- X $> Y, part(X, Xs, Ls, Bs).
part(X, [Y|Xs], Ls, [Y|Bs]) :- X $=< Y, part(X, Xs, Ls, Bs).
Then we can successfully run queries such as:
[eclipse 16]: qs([3.14, Y, 1, 5.5], [T, 2, U, Z]).

Y = 2
T = 1
U = 3.14
164 The suspend library

Z = 5.5
Yes (0.00s cpu, solution 1, maybe more) ? ;

No (0.00s cpu)
The same behaviour holds for all the considered modifications of QUICKSORT.

9.4.3 Variable declarations

As already mentioned, specific variable declarations are available in the
suspend library. Variable declarations will be important in the next chapter
in which we study two ECLi PSe libraries that allow us to automatically re-
duce variable domains. In contrast, variable declarations are hardly relevant
in the suspend library and the atomic goal declareDomains(List) could
equally well be deleted from the definition of the solve/1 predicate on page
156.
Indeed, suspending a variable declaration until the variable is instantiated
does not seem to be common sense. However, it does illustrate the fact
that in ECLi PSe a variable declaration is simply a unary constraint on
the variable, in principle no different from any other constraints. Also,
the presence of variable declarations allows us to write programs that we
can execute using both the suspend library and subsequent, more powerful
libraries.
In the suspend library two types of variable declaration are provided. The
first one is a range declaration. An example is
[eclipse 17]: suspend:(X :: 1..9).

X = X

Delayed goals:
suspend : (X :: 1 .. 9)
Yes (0.00s cpu)
This query generates a declaration of the variable X by stipulating that it
ranges over the integer interval [1..9]. The ECLi PSe prompt confirms
that this variable declaration is interpreted as a unary constraint. Once
the suspend library is loaded, we can simply write X :: 1..9 instead of
suspend:(X :: 1..9) and similarly with all other declarations here dis-
cussed. From now on we assume that the suspend library is loaded.
To abbreviate multiple declarations with the same domain a list notation
can be used:
 9.4 Core constraints 165

[S,E,N,D,M,O,R,Y] :: 0..9.
The domain bounds can be parameters that will be instantiated at the
run-time, as in the following example:
declare(M, N, List):- List :: M..N.
Then the query declare(0, 9, [S,E,N,D,M,O,R,Y]) achieves the effect of
the above declaration.
When handled by the suspend library the range is only used as a test
whether the variable becomes correctly instantiated:
[eclipse 18]: X :: 1..9, X = 5.

X = 5
Yes (0.00s cpu)
[eclipse 19]: X :: 1..9, X = 0.

No (0.00s cpu)
Finally, by using $:: instead of :: we can declare variables ranging
over a real interval:
[eclipse 20]: X :: 1..9, X = 2.5.

No (0.00s cpu)
[eclipse 21]: X $:: 1..9, X = 2.5.

X = 2.5
Yes (0.00s cpu)
The same can be achieved by using reals as bounds.
The second form of a variable declaration in the suspend library is pro-
vided by the integers/1 and reals/1 constraints. An example is
integers(X)
which declares the variable X ranging over the set of integers. Such decla-
rations can be viewed as a form of type declarations. As before, multiple
declarations can be written using the list notation. The following sample
queries illustrate the use of these two forms of variable declarations:
[eclipse 22]: integers([X,Y]), X = 3, Y = 4.

X = 3
166 The suspend library

Y = 4
Yes (0.00s cpu)
[eclipse 23]: integers([X,Y]), X = 3, Y = 4.5.

No (0.00s cpu)
[eclipse 24]: reals(X), X = [1,2.1,4].

X = [1, 2.1, 4]
Yes (0.00s cpu)
[eclipse 25]: reals(X), X = [1,2.1,a].

No (0.00s cpu)

Integrality can also be imposed implicitly on all the expressions in a con-

straint by using # instead of $:

[eclipse 26]: X $> 2, X = 2.5.

X = 2.5
Yes (0.00s cpu)
[eclipse 27]: X #> 2, X = 2.5.

No (0.00s cpu)

9.4.4 Reified constraints

Reified constraints are constraints that can be switched to true or false
by setting an extra Boolean variable. All the core constraints mentioned
above can be reified in this way, by adding an extra Boolean variable (so
one taking a value 0 or 1) as an argument of the constraint. Here are a few
examples, where we assume that the suspend library is loaded:

[eclipse 28]: $>(5, 4, 1).

Yes (0.00s cpu)

[eclipse 29]: $>(4, 5, 1).

No (0.00s cpu)
[eclipse 30]: $>(4, 5, Bool), Bool = 0.

Bool = 0
 9.5 User defined suspensions 167

Yes (0.00s cpu)

[eclipse 31]: $>(3, Y, 1), Y = 2.

Y = 2
Yes (0.00s cpu)
[eclipse 32]: $::(X, 1..9, 0), X = 10.

X = 10
Yes (0.00s cpu)
[eclipse 33]: or(1,0,B).

B = 1
Yes (0.00s cpu)

In the last example the binary disjunction or/2 is reified to the ternary
one, or/3.

9.5 User defined suspensions

The suspend library supports all the core constraints, but occasionally we
also need to suspend user-defined goals. This can be done in ECLi PSe using
the built-in suspend/3 that allows us to suspend arbitrary atomic goals.
To explain its use consider first an atomic goal referring to a built-in.
Suppose, for example, we wish to suspend the atomic goal X =:= 10 until
X is instantiated. To achieve this behaviour we pass the atomic goal as an
argument to the suspend/3 predicate:

[eclipse 34]: suspend(X =:= 10, 3, X -> inst).

X = X

Delayed goals:
X =:= 10
Yes (0.00s cpu)

The suspended atomic goal is executed as soon as X becomes instantiated:

[eclipse 35]: suspend(X =:= 10, 3, X -> inst), X is 2+8.

X = 10
Yes (0.00s cpu)
168 The suspend library

The second argument to suspend/3 is an integer (3 in the above exam-

ples), indicating the priority of the goal when it wakes up. When several
goals wake up as the result of the same event (such as X becoming instanti-
ated) the higher priority goals are executed earlier.
The third argument specifies the waking condition, written as
Term -> Cond. (Alternatively a list of such conditions can be given.) The
term Term can be any term that includes variables. The waking condition
applies to all the variables appearing in the term. The condition Cond is
normally inst.1
To illustrate the suspension of user-defined goals let us define a predicate
xor/2 that defines exclusive or. Of course it can be encoded simply as:
xor(1, 0).
xor(0, 1).
But this definition leads to an unnecessary backtracking search that we
now wish to avoid. We do this by suspending the atomic goal xor(X, Y)
and have it wake only after both variables are instantiated, by building a
chain of suspensions. We achieve this by means of a predicate susp xor/2,
defined in Figure 9.2.

susp_xor(X, Y) :-
( nonvar(X) ->
susp_y_xor(X, Y)
;
suspend(susp_y_xor(X,Y), 3, X -> inst)
).

susp_y_xor(X, Y) :-
( nonvar(Y) ->
xor(X, Y)
;
suspend(xor(X,Y), 3, Y -> inst)
).

xor(1, 0).
xor(0, 1).

Fig. 9.2 Exclusive or with a suspension

1 The full set of conditions can be found in the ECLi PSe documentation.
 9.5 User defined suspensions 169

This predicate waits using the call

suspend(susp_y_xor(X,Y), 3, X -> inst)

until X is instantiated and then it calls the second predicate, susp y xor/2.
(If X is already instantiated, then it calls susp y xor/2 immediately.) This
predicate in turn waits, using the call

suspend(xor(X,Y), 3, Y -> inst),

until Y is instantiated before finally calling xor/2. No matter in which order

the variables X and Y become instantiated the susp xor/2 predicate behaves
correctly:

[eclipse 36]: susp_xor(X, Y).

X = X
Y = Y

Delayed goals:
susp_y_xor(X, Y)

Yes (0.00s cpu)

[eclipse 37]: susp_xor(X, Y), X = 0.

X = 0
Y = Y

Delayed goals:
xor(0, Y)
Yes (0.00s cpu)
[eclipse 38]: susp_xor(X, Y), Y = 1.

X = X
Y = 1

Delayed goals:
susp_y_xor(X, 1)
Yes (0.00s cpu)
[eclipse 39]: susp_xor(X, Y), Y = 1, X = 0.

X = 0
170 The suspend library

Y = 1
Yes (0.00s cpu)
This should be contrasted with the behaviour of xor/2: the corresponding
queries (so queries 3739 with susp_xor replaced by xor) all succeed, but
each time a backtracking takes place.
Finally, of course there is no need to delay an atomic goal by means of
suspend/3 until all its arguments are instantiated. In fact, a more general
use of suspend/3 allows us to suspend an atomic goal until some of its
arguments are instantiated.
To explain the syntax and its use let us suspend the xor(X, Y) goal until
either X or Y have been instantiated:
[eclipse 40]: suspend(xor(X, Y), 2, [X, Y] -> inst).

X = X
Y = Y

Delayed goals:
xor(X, Y)
Yes (0.00s cpu)
[eclipse 41]: suspend(xor(X, Y), 2, [X, Y] -> inst), X = 1.

X = 1
Y = 0
Yes (0.00s cpu)
[eclipse 42]: suspend(xor(X, Y), 2, [X, Y] -> inst), Y = 1.

X = 0
Y = 1
Yes (0.00s cpu)
We shall discuss the suspend/3 built-in again in Subsection 10.3.1, where
we shall review its behaviour in the presence of constraint propagation.

9.6 Generating CSPs

Now that we know how to create constraints using the suspend library let
us see how we can generate specific CSPs. As an example, to generate the
CSP
hx 6= y, y 6= z, x 6= z ; x {0, 1}, y {0, 1}, z {0, 1}i
 9.6 Generating CSPs 171

we use the following query:

[eclipse 43]: [X,Y,Z] :: 0..1, X #\= Y, Y #\= Z, X #\= Z.

X = X
Y = Y
Z = Z

Delayed goals:
suspend : ([X, Y, Z] :: 0 .. 1)
suspend : (X #\= Y)
suspend : (Y #\= Z)
suspend : (X #\= Z)
Yes (0.00s cpu)

So the delayed goals correspond to the variable declarations and the

unsolved constraints on these variables. In other words, the execution of
the above query results in an answer that is a representation of the original
CSP.
Note that the corresponding ECLi PSe query with #\= replaced by =\=
yields a run-time error:

[eclipse 44]: [X,Y,Z] :: 0..1, X =\= Y, Y =\= Z, X =\= Z.

instantiation fault in X =\= Y
Abort

Indeed, during its execution the arithmetic comparison predicate =\= is ap-
plied to arguments that are not numbers.
So far, all the constraints were explicitly written. A powerful feature of
ECLi PSe is that the constraints and whole CSPs can also be generated. Let
us clarify it by the following example. Suppose we wish to generate the CSP

hx1 < x2 , . . ., xn1 < xn ; x1 {1..1000}, . . ., xn {1..1000}i,

where n > 1, independently of the value of n. This can be done by means of

the program given in Figure 9.3 that should be compared with the ORDERED
program from Figure 7.6 on page 127.
We can now run the query

List :: 1..1000, ordered(List).

with an arbitrary list of variables List. Each such query generates an ap-
propriate CSP. For example, to generate the CSP
172 The suspend library

% ordered(Xs) :- Xs is an <-ordered list of numbers.

ordered(List) :-
( List = [] ->
true
;
( fromto(List,[This,Next | Rest],[Next | Rest],[_])
do
This $< Next
)
).

Fig. 9.3 Another ORDERED program

hx < y, y < z, z < u, u < v ; x {1..1000}, y {1..1000},

z {1..1000}, u {1..1000}, v {1..1000}i

we use the following query:

[eclipse 45]: [X,Y,Z,U,V] :: 1..1000, ordered([X,Y,Z,U,V]).

X = X
Y = Y
Z = Z
U = U
V = V

Delayed goals:
suspend : ([X, Y, Z, U, V] :: 1 .. 1000)
suspend : (X < Y)
suspend : (Y < Z)
suspend : (Z < U)
suspend : (U < V)
Yes (0.00s cpu)

Earlier we used a query (query No. 43 on page 171) that given a list
of three variables posted a disequality constraint between each pair of ele-
ments of the list. Now we show how to define a generic version that we call
diff list/1 that does the same with a list of any number of variables. This
predicate corresponds to the all different constraint introduced in Sub-
 9.6 Generating CSPs 173

section 6.4.2. It is defined in Figure 9.4 using the fromto/4 and foreach/2
iterators.

% diff_list(List) :-
% List is a list of different variables.

diff_list(List) :-
( fromto(List,[X|Tail],Tail,[])
do
( foreach(Y, Tail),
param(X)
do
X $\= Y
)
).

Fig. 9.4 The DIFF LIST program

For example, the query

length(List, 3), List:: 1..100, diff_list(List).

produces a list List of three anonymous variables, each with the domain
1..100, and subsequently generates the disequality constraints between
these three variables. Recall from Section 7.3 that length/2 is a Prolog
built-in that given a list as the first argument computes its length, or given
a natural number n as the second argument generates a list of anonymous
(and thus distinct) variables of length n.
Here is the outcome of the interaction with ECLi PSe :

[eclipse 46]: length(List, 3), List:: 1..100, diff_list(List).

List = [_268, _330, _242]

Delayed goals:
suspend : ([_268, _330, _242] :: 1 .. 100)
suspend : (_268 =\= _330)
suspend : (_268 =\= _242)
suspend : (_330 =\= _242)
Yes (0.00s cpu)
174 The suspend library

As a final example consider the problem of generating a constraint of the

form
Xn
ai xi v,
i=1

where n 1, all ai s are given parameters and all xi s and v are variables
ranging over an integer interval. Recall that we used this constraint when
discussing the knapsack problem in Subsection 6.5.1. This constraint can
be generated in ECLi PSe as follows:
sigma(As, Xs, V) :-
sum(As, Xs, Out),
eval(Out) #=< V.

sum([], [], 0).

sum([A|As], [X|Xs], A*X + Y) :-
sum(As, Xs, Y).
Note the use of the eval/1 built-in in the constraint eval(Out) #=< V.
It is a wrapper used to indicate that a variable that occurs inside an arith-
metic constraint, here Out, will be bound to an expression at the run-time.
So it is only used in programs which generate constraints at run-time, like
the above clause defining sigma/3.
Now, during the execution of the query
[X,Y,Z,U] :: 1..1000, sigma([1,2,3,4], [X,Y,Z,U], V).
the local variable Out becomes instantiated to the sum 1*X + 2*Y + 3*Z +
4*U that is subsequently passed to the constraint eval(Out) $<= V and we
get the following outcome:
[eclipse 47]: [X,Y,Z,U] :: 1..1000,
sigma([1,2,3,4], [X,Y,Z,U], V).

X = X
Y = Y
Z = Z
U = U
V = V

Delayed goals:
suspend : ([X, Y, Z, U] :: 1 .. 1000)
suspend : (eval(1 * X + (2 * Y + (3 * Z + (4 * U + 0))))
 9.7 Using the suspend library 175

#=< V)
Yes (0.00s cpu)

9.7 Using the suspend library

Any program behaviour that can be achieved using the suspend library can
be equally achieved without it. However, reordering of the atomic goals so
that they are evaluated as soon as their arguments have become instantiated
requires careful programming. The key benefit of the suspend library is that
such a reordering can be achieved in a straightforward way. So using this
library flexible programs can be written in a simple and readable way. To
illustrate this point we now discuss three programs that compute solutions
to the CSPs studied in Section 6.4.

9.7.1 Map colouring problems

Recall that the problem was specified by
a set of regions,
a (smaller) set of colours,
a neighbour relation between pairs of regions.
The problem was to associate a colour with each region so that no two
neighbours have the same colour.
To model this problem in ECLi PSe we simply generate the appropriate
CSP discussed in Subsection 6.4.1 by means of an ECLi PSe program. So
each region is represented by a decision variable. The value of the variable,
when it is finally instantiated, is the colour chosen for that region.
Because the core constraints are numeric, the program has to use numbers
for the colours. So we represent the data as follows:
no_of_regions(4).

neighbour(1, 2).
neighbour(1, 3).
neighbour(1, 4).
neighbour(2, 3).
neighbour(2, 4).

colour(1). % red
colour(2). % yellow
colour(3). % blue
176 The suspend library

If the neighbours are recorded in a predicate, this problem is difficult to

solve in ECLi PSe without using the suspend library. The difficulty is that as
soon as a region is coloured the disequality should be checked between that
regions variable and the already instantiated variables of its neighbours.
This form of control is awkward to express and difficult to modify in case
the programmer decides to instantiate the variables in a different order.
In contrast, using the suspend library the map colouring program can be
written in a simple and generic way. It is presented in Figure 9.5. The vari-
able declarations are omitted as the variable domains are used only within
the search, through the calls of colour(R). In the program the counter vari-
ables I and J have the same start and stop values so we use the intuitively
clear shorthand multifor([I,J],1,Count) available in ECLi PSe .

% colour_map(Regions) :- Regions is a colouring of

% the map represented by the neighbour/2 predicate.

colour_map(Regions) :-
constraints(Regions),
search(Regions).

constraints(Regions) :-
no_of_regions(Count),
dim(Regions,[Count]),
( multifor([I,J],1,Count),
param(Regions)
do
( neighbour(I, J) ->
Regions[I] $\= Regions[J]
;
true
)
).

search(Regions) :-
( foreacharg(R,Regions) do colour(R) ).

Fig. 9.5 The MAP COLOURING program

 9.7 Using the suspend library 177

The suspend implementation of $\= ensures that the disequalities be-

tween the neighbouring countries are automatically tested as soon as both
neighbours are instantiated. The separation of the constraint generation
phase from the search allows us to modify the search routine, for example
to take into account a specific structure of the map, without any need to
change the constraint handling.
For the above data six solutions listed in Subsection 6.4.1 are generated,
starting with:

[eclipse 48]: colour_map(Regions).

Regions = [](1, 2, 3, 3)
Yes (0.00s cpu, solution 1, maybe more) ? ;

9.7.2 SEND + MORE = MONEY puzzle

Recall from Subsection 6.4.2 that the problem is to replace each letter by a
different digit so that the above sum is correct. The appropriate ECLi PSe
program consists simply of the formulation of the constraints followed by
the systematic selection of possible values for all variables. We choose here,
as an illustration, the third representation given in Subsection 6.4.2 with the
all_different constraint modelled by the earlier introduced diff_list/1
constraint. The program is given in Figure 9.6.
For compatibility with the next chapter we included in the program the
variable declarations even though they are not needed here. To generate
the possible values we use in the definition of search/1 the select val/3
predicate from Figure 3.2 on page 46.
The following interaction with ECLi PSe shows that this a very inefficient
solution:

[eclipse 49]: send(List).

List = [9, 5, 6, 7, 1, 0, 8, 2]
Yes (176.60s cpu, solution 1, maybe more) ? ;

No (183.94s cpu)

The reason is that each constraint is activated only when all its variables
become ground. So the fact that these constraints were generated first is
only of a limited use. For example, the combination of values S = 1, E =
1, N = 1 is not generated, since a failure arises after trying S = 1, E = 1.
178 The suspend library

% send(List):- List is a solution to

% the SEND + MORE = MONEY puzzle.

send(List):-
List = [S,E,N,D,M,O,R,Y],
List :: 0..9,
diff_list(List),
1000*S + 100*E + 10*N + D
+ 1000*M + 100*O + 10*R + E
$= 10000*M + 1000*O + 100*N + 10*E + Y,
S $\= 0,
M $\= 0,
search(List).

search(List) :-
( foreach(Var,List) do select_val(0, 9, Var) ).
augmented by the DIFF LIST program of Figure 9.4 and the SELECT
program of Figure 3.2.

Fig. 9.6 The SEND MORE MONEY program

However, all possible value combinations for the variables that are pairwise
different are still generated since the crucial equality constraint is used only
for testing. In the next chapter we shall see that in the presence of constraint
propagation this approach becomes considerably more realistic and efficient.

9.7.3 The n-queens problem

To solve this problem using the suspend library we use the second for-
malisation of it given in Subsection 6.4.3. So we assume that the solution
is generated in the form of a one-dimensional array QueenStruct of size
[1..Number] and just need to generate the following disequality constraints
for I [1..Number] and J [1..I-1]:

QueenStruct[I] $\= QueenStruct[J],

QueenStruct[I]-QueenStruct[J] $\= I-J,
QueenStruct[I]-QueenStruct[J] $\= J-I.
 9.7 Using the suspend library 179

The corresponding program is given in Figure 9.7. Again, for compatibil-

ity with the next chapter we included in the program the variable declara-
tions, in the line

QueenStruct[I] :: 1..Number,

even though they are not needed.

% queens(QueenStruct, Number) :- The array QueenStruct

% is a solution to the Number-queens problem.

queens(QueenStruct, Number) :-
dim(QueenStruct,[Number]),
constraints(QueenStruct, Number),
search(QueenStruct).

constraints(QueenStruct, Number) :-
( for(I,1,Number),
param(QueenStruct,Number)
do
QueenStruct[I] :: 1..Number,
( for(J,1,I-1),
param(I,QueenStruct)
do
QueenStruct[I] $\= QueenStruct[J],
QueenStruct[I]-QueenStruct[J] $\= I-J,
QueenStruct[I]-QueenStruct[J] $\= J-I
)
).

search(QueenStruct) :-
dim(QueenStruct,[N]),
( foreacharg(Col,QueenStruct),
param(N)
do
select_val(1, N, Col)
).
augmented by the SELECT program of Figure 3.2.

Fig. 9.7 Another QUEENS program

180 The suspend library

The structure of this program is different from that of the QUEENS program
given in Figure 7.7 on page 130. Namely, first constraints are generated and
then systematically the values for the subscripted variables are generated.
Unfortunately, because each constraint is activated only when the variable
QueenStruct becomes ground (i.e., when every queen has been placed) this
program is very inefficient:

[eclipse 50]: queens(QueenStruct, 8).

QueenStruct = [](1, 5, 8, 6, 3, 7, 2, 4)
Yes (341.84s cpu, solution 1, maybe more) ?

In the next chapter we shall reevaluate this approach in the presence of

constraint propagation. The reader should also consult Exercise 9.3.

9.8 Summary
In this chapter we discussed the suspend library of ECLi PSe . It provides a
support for passive constraints through a dynamic reordering of the atomic
goals. In the case of arithmetic constraints this prevents the occurrence of
run-time errors.
We introduced the following core constraints that are available in the
suspend library:

Boolean constraints:
and/2, or/2, neg/1, =>/2,
arithmetic comparison constraints on reals:
$<, $=<, $=, $\=, $>=, $>,
arithmetic comparison constraints on integers:
#<, #=<, #=, #\=, #>=, #>,
variable declarations:
range (::, $::, #::, for example M..N), integers/1 and reals/1.

We also discussed the suspend/3 built-in that allows us to delay arbitrary

atomic goals. Further, we introduced the eval/1 built-in used to indicate
that a variable within an arithmetic constraint will be bound to an expres-
sion at the run-time.
We also explained how CSPs can be generated and illustrated the intro-
duced facilities by presenting three programs that provide solutions to the
CSPs studied in Section 6.4.
 9.9 Exercises 181

9.9 Exercises

Exercise 9.1 The META INTERPRETER program for the suspend library
given in Figure 9.1 on page 159 outputs the suspended atomic goals in the
reverse order:

[eclipse 1]: solve((suspend:(2 < Y + 1), suspend:(X > 3)),

true, SuspOut).

Y = Y
X = X
SuspOut = suspend : (X > 3), suspend : (2 < Y + 1), true
Yes (0.00s cpu)

while we have

[eclipse 2]: suspend:(2 < Y + 1), suspend:(X > 3).

Y = Y
X = X

Delayed goals:
suspend : (2 < Y + 1)
suspend : (X > 3)
Yes (0.00s cpu)

Propose a modification that deals with this problem and that reproduces
the output of the suspend library, through a call of a solve/1 predicate.

Exercise 9.2 Write a version of the diff list/2 predicate from Figure 9.4
on page 173 that uses recursion instead of iteration.

Exercise 9.3 As already mentioned, the QUEENS program from Figure 9.7
on page 179 is very inefficient because each constraint is activated only when
the variable QueenStruct becomes ground. Propose a modification of this
program in which each binary constraint is activated as soon as both of its
variables are ground.

Exercise 9.4 By a clause we mean a disjunction of Boolean variables. We

represent it is a list. The clause is then true if at least one variable in it is
identical to 1.
182 The suspend library

Using the suspend/3 built-in write a procedure test cl/1 that imple-
ments the test whether a clause is true, by suspending on just one Boolean
variable.
Hint. Use the following filter/3 predicate in your implementation:
filter([], [], unsat).
filter([H | T], [H | Rest], Sat) :-
var(H), !, filter(T, Rest, Sat).
filter([0 | T], Rest, Sat) :- !,
filter(T, Rest, Sat).
filter([1 |_], [], sat).
The call filter(List, Rest, Sat) filters the instantiated variables (1
or 0) out of the clause List, leaving the remaining variables in Rest, and
recording in Sat if the clause is already satisfied.

Exercise 9.5 Modify the ORDERED program from Figure 9.3 on page 172
using the suspend/3 built-in so that it fails as early as possible. For example
the query ordered([W,X,Y,Z]), W = 2, Z = 2 should fail.
 Part IV

Programming with active constraints

 10

Constraint propagation in ECLiPSe

10.1 Introduction 185

10.2 The sd library 186
10.3 The ic library 189
10.4 Disjunctive constraints and reification 200
10.5 Summary 203
10.6 Exercises 204

10.1 Introduction

A SIGNIFICANT LIMITATION of the support for constraint pro-

cessing provided by the suspend library is that only passive con-
straints are admitted. As a result some obvious conclusions cannot
be inferred without search. As an example, recall that the query No. 5:
suspend:(X or Y), X = 0.
from page 160 did not allow us to infer that Y = 1. Such a conclusion could
only be established by explicitly instantiating Y and testing whether the
instantiation is correct.
Starting with this chapter we discuss libraries of ECLi PSe that provide
a more powerful support for constraint processing by admitting active con-
straints. So an evaluation of a constraint can affect its variables. In par-
ticular inferences such as Y = 1 above can be made automatically. This
additional support is called constraint propagation and, as already men-
tioned in Section 6.6, involves removal from the domains of variables some
values that do not participate in any solution. In the above example con-
straint propagation results in removing the value 0 from the domain of Y
which leaves 1 as the only possible alternative. As a result Y gets instantiated
to 1.

185
186 Constraint propagation in ECLi PSe

The constraint propagation is activated in an eager way: as soon as a new

constraint is encountered, a new round of the constraint propagation process
is triggered. This may reduce the domains of the previously considered
variables and lead in turn to a renewed activation of the already considered
constraints. The formal details of this process are omitted but are illustrated
by several examples.
In this chapter we introduce two libraries that support active constraints.
They are called sd and ic. The first one deals with constraints defined over
symbolic values, while the second deals with the core constraints, so con-
straints defined over Booleans, integers and reals, and reified constraints.
We focus here on the built-in facilities concerned with the constraint propa-
gation. The facilities supporting search will be discussed in the next chapter.

10.2 The sd library

We introduce here the ECLi PSe library sd (standing for symbolic domains)
which supports symbolic domains such as {a,b,c,d,e}. In sd the variables
are declared as follows:
X &:: [a,b,d],
[Y,Z] &:: [a,b,c,d]
In all examples below we assume that the sd library is loaded. Two binary
constraints are available on symbolic finite domains. These are:
equality, written as &=
disequality, written as &\=.
To generate the CSP
hx = y, y 6= z ; x {a, c, d}, y {a, b, c, d}, z {c}i
we thus use the following query:
[eclipse 1]: X &:: [a,c,d], Y &:: [a,b,c,d], Z &:: [c],
X &= Y, Y &\= Z.

X = X{[a, d]}
Y = Y{[a, d]}
Z = c

Delayed goals:
X{[a, d]} &= Y{[a, d]}
Yes (0.00s cpu)
 10.2 The sd library 187

We see here an important difference in ECLi PSe behaviour with respect

to the suspend library. Indeed, different variable domains from the declared
ones are produced. In fact, the produced answer corresponds to the following
CSP:
hx = y ; x {a, d}, y {a, d}, z {c}i.
So in the presence of the sd library an inference was made that led to a
removal of the value c from the domain of X and the values b,c from the
domain of Y. This inference can be easily justified by noting that because
of the equality constraint X &= Y the domains of X and Y have to be equal.
Further, because of the disequality constraint Y &\= Z the value c can be
removed from the domain of Y and hence the domain of X. At this moment
the disequality constraint Y &\= Z becomes solved (i.e., all possible variable
combinations satisfy it) so it is deleted. In the sd library such inferences are
done automatically.
In some situations, thanks to constraint propagation, all constraints be-
come solved and a solution is obtained. As an example reconsider the CSP
hx = y, y 6= z, z 6= u ; x {a, b, c}, y {a, b, d}, z {a, b}, u {b}i
from Section 6.3. To generate it in ECLi PSe we use the following query:
[eclipse 2]: X &:: [a,b,c], Y &:: [a,b,d], Z &:: [a,b],
U &:: [b], X &= Y, Y &\= Z, Z &\= U.

X = b
Y = b
Z = a
U = b
Yes (0.00s cpu)
So this query actually produces a unique solution to the CSP in question.
Finally, it is important to note that the constraint propagation for the
equality and disequality constraints considers each constraint separately
(and iterates the inferences until no new ones can be made). For exam-
ple, even though it is possible to infer that the CSP
hx 6= y, y 6= z, x 6= z ; x {a, b}, y {a, b}, z {a, b}i
has no solution, the sd library cannot make this inference:
[eclipse 3]: [X,Y,Z] &:: [a,b],
X &\= Y, Y &\= Z, X &\= Z.
188 Constraint propagation in ECLi PSe

X = X{[a, b]}
Y = Y{[a, b]}
Z = Z{[a, b]}

Delayed goals:
X{[a, b]} &\= Y{[a, b]}
Y{[a, b]} &\= Z{[a, b]}
X{[a, b]} &\= Z{[a, b]}
Yes (0.00s cpu)
As in the case of the suspend library we can also generate sequences of
constraints by means of ECLi PSe programs. For example, to generate the
CSP
hx1 = x2 , . . ., xn1 = xn ; x1 {l,r}, x2 {l,r,+}, . . ., xn {l,r,+}i
where n > 1, independently of the value of n, we can use the following
program:
equal([]).
equal([H | Ts]) :- equal(H, Ts).

equal(_, []).
equal(H, [Y | Ts]) :- H &= Y, equal(Y, Ts).
Then for each variable H and a list of variables List the query
H :: [l,r], List :: [l,r,+], equal([H | List])
generates an appropriate CSP on the variables in [H | List]. For example,
we have
[eclipse 4]: H &:: [l,r], [X,Y,Z] &:: [l,r,+],
equal([H | [X,Y,Z]]).

H = H{[l, r]}
X = X{[l, r]}
Y = Y{[l, r]}
Z = Z{[l, r]}

Delayed goals:
H{[l, r]} &= X{[l, r]}
X{[l, r]} &= Y{[l, r]}
Y{[l, r]} &= Z{[l, r]}
 10.3 The ic library 189

Yes (0.00s cpu)

So here, like in the previous examples of this section, an inference was
made that automatically led to reduced variable domains.

10.3 The ic library

We now move on to discuss the library ic (standing for interval constraints).
It greatly enhances the capabilities provided by the suspend library. The
ic library, just as the suspend library, supports all the core constraints,
originally introduced in Section 9.4. Recall that these constraints comprise
the variable declarations (viewed as unary constraints), Boolean constraints,
arithmetic constraints and reified constraints.
ic also supports a more powerful syntax for variable declarations. In
addition to the suspend library variable declarations we can also declare
variables ranging over more complex domains than ranges, for instance:
X :: [3,4,5,6,8,9],
[Y,Z] :: [1..3,5,7..9],
and use an alternative syntax for ranges:
Z :: [3..9].
In the case of other core constraints, the same syntax as in the suspend
library is used. However, in contrast to the suspend library, the constraint
propagation is now triggered. We now illustrate the resulting effect for the
Boolean constraints and arithmetic constraints.

10.3.1 Constraint propagation for Boolean constraints

To see the difference between the suspend and ic libraries let us return to
the query
X or Y, X = 0
already discussed in Section 9.4. We noted there that the suspend library
does not allow us to deduce that Y = 1. In contrast, once the ic library is
loaded this inference is made:
[eclipse 5]: ic:(X or Y), X = 0.

X = 0
Y = 1
Yes (0.00s cpu)
190 Constraint propagation in ECLi PSe

Analogous inferences are automatically made for other Boolean constraints.

In what follows we assume that all Boolean constraints are taken from the
ic library, so the explicit module qualification can be dropped.
To illustrate the cumulated effect of the underlying constraint propagation
consider digital circuits built out of the AND, OR and XOR gates. Each gate
generates an output value given two input values. In particular, recall that
XOR stands for the exclusive or defined by the following table:

XOR 0 1
0 0 1
1 1 0

We can view these gates as ternary (so reified) Boolean constraints. There-
fore the circuits built out of these gates can be naturally represented by a
sequence of such ternary constraints. In particular, consider the circuit de-
picted in Figure 10.1.

and

i1 xor or o2
i2 and

i3
xor o1

Fig. 10.1 Full adder circuit

This circuit is called full adder as it computes the binary sum i1 + i2 + i3

in the binary word o2 o1 . For example 1 + 1 + 0 yields 10. To reason about
this circuit we can use the following program:1

fullAdder(I1, I2, I3, O1, O2) :-

xor3(I1, I2, X1),
and(I1, I2, Y1),
xor3(X1, I3, O1),
and(I3, X1, Y2),
or(Y1, Y2, O2).

xor3(X, Y, Z) :- neg(X, X1), neg(Y, Y1), and(X, Y1, Z1),

and(X1, Y, Z2), or(Z1, Z2, Z).
1 In ECLi PSe xor/3 is a built-in with a different meaning, so we need to use another name to
define XOR.
 10.3 The ic library 191

The last rule defines xor3/3 by means of the and/3, or/3 and neg/2 con-
straints using the equivalence

X xor3 Y iff (X and not Y) or (Y and not X).

We would like now to show that i3 = 0 and o2 = 1 implies that i1 =

1, i2 = 1 and o1 = 0. To this end we simply execute the appropriate query:

[eclipse 6]: fullAdder(I1, I2, 0, O1, 1).

I1 = 1
I2 = 1
O1 = 0
Yes (0.00s cpu)

So we reached the desired conclusion using constraint propagation only.

Note that no backtracking takes place here. To run this query in Prolog
instead of using the constraints xor3/3, and/3 and or/3 we would have to
define them as predicates, listing the facts that represent their truth tables.
This would yield the same outcome at the cost of a considerable number of
backtracks that would start as soon as the first atomic goal, xor3(I1, I2,
X1), is encountered.
Let us discuss now an alternative definition of xor/3 using the suspend/3
built-in defined in Section 9.5. The following code generates the appropriate
constraints as soon as one argument of xor/3 is ground:

xor3(X, Y, Z) :-
[X,Y,Z]::0..1,
( ground(X) -> xor3g(X,Y,Z) ;
ground(Y) -> xor3g(Y,X,Z) ;
ground(Z) -> xor3g(Z,X,Y) ;
true -> suspend(xor3(X,Y,Z),2, [X,Y,Z] -> inst)
).

xor3g(0, Y, Z) :- Y = Z.
xor3g(1, Y, Z) :- neg(Y, Z).

Note that the xor3g/3 predicate is called each time with the arguments
appropriately swapped. For example, if X and Y are not ground and Z is,
the call xor3g(Z,X,Y) takes place. Then, if Z=0, the constraint X = Y is
generated and if Z=1, the constraint neg(X, Y) is generated.
192 Constraint propagation in ECLi PSe

10.3.2 Constraint propagation for linear constraints

In the ic library a built-in constraint propagation mechanism is also avail-
able for arithmetic constraints. Recall that the suspend library provided the
possibility to use # instead of $ when writing these constraints. Both ver-
sions are allowed in ic, as well. If the # syntax is used, all subexpressions
on both sides of the constraint are also constrained to be integral. Thus we
have
[eclipse 10]: [X,Y]::1..4, X/2 - Y/2 #= 1.

X = 4
Y = 2
Yes (0.00s cpu)
That is, this query has only one solution, X = 4 and Y = 2. The solution X
= 3 and Y = 1 is excluded since then X/2 and Y/2 are not integers.
Below we use the $ syntax rather than the # syntax so as to allow us to
freely switch between integer, finite domains and real, continuous domains.
If a variable without a domain is encountered in an ic constraint, it is
automatically initialised with an infinite domain -1.0Inf .. 1.0Inf.
Because of the built-in constraint propagation the generated CSP can
differ from the one actually written down in ECLi PSe . Since the constraint
propagation maintains equivalence, the resulting CSP is always equivalent
to the original one (i.e., it has the same set of solutions). As an example
involving linear constraints reconsider the predicate ordered/2 from Figure
9.3 on page 172. To generate the CSP
hx < y, y < z, z < u, u < v ; x {1..1000}, y {1..1000},
z {1..1000}, u {1..1000}, v {1..1000}i
we used there the query
[X,Y,Z,U,V] :: 1..1000, ordered([X,Y,Z,U,V]).
Now, however, a different, though equivalent, CSP is generated:
[eclipse 11]: [X,Y,Z,U,V] :: 1..1000, ordered([X,Y,Z,U,V]).

X = X{1 .. 996}
Y = Y{2 .. 997}
Z = Z{3 .. 998}
U = U{4 .. 999}
V = V{5 .. 1000}
 10.3 The ic library 193

Delayed goals:
ic : (Y{2 .. 997} - X{1 .. 996} > 0)
ic : (Z{3 .. 998} - Y{2 .. 997} > 0)
ic : (U{4 .. 999} - Z{3 .. 998} > 0)
ic : (V{5 .. 1000} - U{4 .. 999} > 0)
Yes (0.00s cpu)
The first five lines report the new domains of the considered variables.
Sometimes this built-in constraint propagation can result in solving the
generated constraints, that is, in producing a solved CSP. In this case the
final domains of its variables are generated. Here is an example:
[eclipse 12]: X :: [5..10], Y :: [3..7], X $< Y, X $\= 6.

X = 5
Y = Y{[6, 7]}
Yes (0.00s cpu)
Here X = 5 is a shorthand for X = X{[5]}, i.e., it stands for the degen-
erated variable declaration X :: [5..5]. So the answer corresponds to the
CSP h ; x {5}, y {6, 7}i with no constraints. If in the solved CSP all
the variable domains are singletons, a customary Prolog answer is produced.
For instance, we have
[eclipse 13]: X :: [5..10], Y :: [3..6], X $< Y.

X = 5
Y = 6
Yes (0.00s cpu)
Note that the corresponding ECLi PSe query
X :: [5..10], Y :: [3..6], X < Y
yields a run-time error:
[eclipse 14]: X :: [5..10], Y :: [3..6], X < Y.
instantiation fault in X{5 .. 10} < Y{3 .. 6}
Abort
So the use of the constraint $</2 instead of the arithmetic comparison
predicate </2 is essential here. Further, in some circumstances the constraint
propagation can detect that the generated CSP is inconsistent, like in this
example:
194 Constraint propagation in ECLi PSe

[eclipse 15]: [X,Y,Z] :: [1..1000], X $> Y, Y $> Z, Z $> X.

No (0.01s cpu)
Also here the replacement of $> by > leads to a run-time error.
So using the constraints and constraint propagation only we can solve
more queries than using Prolog and also more queries than using the suspend
library.
Finally, let us return to the problem of generating a constraint of the form
ni=1 ai xi v,
where n 1, all ai s are given parameters and all xi s and v are variables
ranging over an integer interval. We explained in Section 9.6 how to generate
it using the suspend library. Using the ic library and the foreach/2 iterator
this can be also done by means of the following non-recursive program:
sigma(As, Xs, V) :-
( foreach(A,As),
foreach(X,Xs),
foreach(Expr,Exprs)
do
Expr = A*X
),
sum(Exprs) $=< V.
We use here the built-in sum/1. In the context of the ic library the call
sum(List) returns the sum of the members of the list List.
Now, during the execution of the query
[X,Y,Z,U] :: 1..1000, sigma([1,2,3,4], [X,Y,Z,U], V)
the local variable Exprs becomes instantiated to the list [1*X, 2*Y, 3*Z,
4*U] that is subsequently passed to the constraint sum(Exprs) $=< V. Ad-
ditionally, some constraint propagation is automatically carried out and we
get the following outcome:
[eclipse 16]: [X,Y,Z,U] :: 1..1000,
sigma([1,2,3,4], [X,Y,Z,U], V).

X = X{1..1000}
Y = Y{1..1000}
Z = Z{1..1000}
U = U{1..1000}
 10.3 The ic library 195

V = V{10.0..1.0Inf}

Delayed goals:
ic : (4*U{1..1000} + 3*Z{1..1000} + 2*Y{1..1000} +
X{1..1000} - V{10.0..1.0Inf} =< 0)

So the original, implicit, domain of V, -1.0Inf..1.0Inf, got reduced to

10.0..1.0Inf. Recall that when the suspend library was used no domain
reduction was achieved.

10.3.3 SEND + MORE = MONEY puzzle revisited

A dramatic effect of the constraint propagation for linear constraints in
combination with the disequality constraints can be demonstrated using the
SEND + MORE = MONEY puzzle introduced in Subsection 6.4.2 and en-
coded in ECLi PSe in Figure 9.6 on page 178. Thanks to the constraint
propagation process all the variable domains become reduced before the
search process is launched. Indeed, given the variable declarations and con-
straints as in Figure 10.2 (the diff list/1 predicate is defined in Figure
9.4 on page 173) we get the following interaction with ECLi PSe :

[eclipse 17]: send(List).

List = [9, E{4..7}, N{5..8}, D{2..8}, 1, 0, R{2..8}, Y{2..8}]

There are 11 delayed goals.

send(List):-
List = [S,E,N,D,M,O,R,Y],
List :: 0..9,
diff_list(List),
1000*S + 100*E + 10*N + D
+ 1000*M + 100*O + 10*R + E
$= 10000*M + 1000*O + 100*N + 10*E + Y,
S $\= 0,
M $\= 0.

Fig. 10.2 Domains and constraints for the SEND MORE MONEY puzzle
196 Constraint propagation in ECLi PSe

In fact, now, because of the constraint propagation the solution is found

almost instantaneously. Indeed for the SEND MORE MONEY program from Fig-
ure 9.6 on page 178 (the one in which the definition of send/1 ends with
the call to search/1) we get:
[eclipse 18]: send(L).

L = [9, 5, 6, 7, 1, 0, 8, 2]
Yes (0.00s cpu, solution 1, maybe more) ? ;

No (0.00s cpu)
Recall that with the suspend library loaded instead of ic we got the final
answer after more than three minutes of CPU time.
It is important to realise that this significant domain reduction is achieved
thanks to the integrality constraints on the variables used. If we drop them
by changing the declaration
List :: 0..9
to
List :: 0.0..9.0,
very little can be inferred since there are many non-integer solutions to the
problem. Indeed, the only domain that gets reduced then is that of M which
becomes 0.0..1.102.
Finally, let us mention that the diff list/1 predicate is available in the
ic library as the alldifferent/1 built-in.

10.3.4 The n-queens problem revisited

Let us return now to the QUEENS program presented in Figure 9.7 on page
179. We noted there that it took more than 341 seconds of CPU time
to generate the first solution to the 8-queens problem. Now, thanks to the
constraint propagation the first solution is generated almost instantaneously:
[eclipse 19]: queens(QueenStruct, 8).

QueenStruct = [](1, 5, 8, 6, 3, 7, 2, 4)
Yes (0.01s cpu, solution 1, maybe more) ?
Here, in contrast to the program that solves the SEND + MORE =
MONEY puzzle, the constraint propagation has no effect before the search
 10.3 The ic library 197

process is launched. In fact, in the QUEENS program the constraint propa-

gation is activated only when a variable gets instantiated. This happens for
the first time during the search process. Each time a subscripted variable
QueenStruct[I] gets instantiated to a value some, possibly different, values
are removed from the domains of other subscripted variables by virtue of
the constraint propagation that is activated by the disequality constraints
in which QueenStruct[I] occurs.

10.3.5 Constraint propagation for non-linear constraints

The constraint propagation is also activated by non-linear arithmetic con-
straints, as the following three examples show.

[eclipse 20]: [X,Y] :: [1.0..1000.0], X*X + Y*Y $= 1000.

X = X{1.0 .. 31.606961258558218}
Y = Y{1.0 .. 31.606961258558218}

Delayed goals:
ic : (Y{1.0 .. 31.606961258558218} =:=
rsqr(_740{1.0 .. 999.0}))
ic : (X{1.0 .. 31.606961258558218} =:=
rsqr(_837{1.0 .. 999.0}))
ic : (_740{1.0 .. 999.0} + _837{1.0 .. 999.0} =:=
1000)
Yes (0.00s cpu)

The delayed goals show how each constraint was internally rewritten by
means of the auxiliary variables.
Integrality and non-linearity can be combined freely. Adding integrality
to the previous example tightens the bounds not just to the next integer
but, in this case, even further because additional constraint propagation
can take place:

[eclipse 21]: [X,Y] :: [1..1000], X*X + Y*Y $= 1000.

X = X{10 .. 30}
Y = Y{10 .. 30}

Delayed goals:
ic : (Y{10 .. 30} =:= rsqr(_657{100.0 .. 900.0}))
ic : (X{10 .. 30} =:= rsqr(_754{100.0 .. 900.0}))
198 Constraint propagation in ECLi PSe

ic : (_657{100.0 .. 900.0} + _754{100.0 .. 900.0} =:= 1000)

Yes (0.00s cpu)

In some examples constraint propagation alone is sufficient to solve non-

linear constraints:

[eclipse 22]: X :: [1..20], Y :: [9..11], Z :: [155..161],

X*Y $= Z.

X = 16
Y = 10
Z = 160
Yes (0.00s cpu)

In general, however, constraint propagation can only narrow the domain

bounds without determining a solution. For example, consider the following
integer constraint:

[eclipse 23]: X :: 1..1000, X*(X+1)*(X+2) $=< 1000.

X = X{1 .. 22}

There are 6 delayed goals.

Additional constraint propagation can be achieved by trying values near

the boundary of a variables domain, and narrowing the bounds if no solution
is found. The ic built-in to perform this kind of constraint propagation is
squash/3.

[eclipse 24]: X :: 1..1000, X*(X+1)*(X+2) $=< 1000,

squash([X],0.1,lin).

X = X{1 .. 9}

There are 6 delayed goals.

Yes (0.00s cpu)

The second parameter of squash/3 specifies how near the bounds to test
for infeasibility, and the third argument specifies whether to divide up the
domain linearly (lin), or logarithmically (log), which can work better for
large domains. We shall discuss this predicate in more detail in Section 13.5.
 10.3 The ic library 199

10.3.6 Direct variable domain modification

Finally, let us discuss the effect of a direct modification of a variable domain.
We say that an unsolved arithmetic constraint is attached to a variable,
say V, if V occurs in this constraint. Normally if the domain of V is reduced,
all the constraints attached to V are woken and may cause further domain
reductions through constraint propagation.
In ECLi PSe one can reduce the domain of a variable using the exclude/2
built-in. For a variable X the query exclude(X,El) removes the element El
from the domain of X. If the resulting domain is empty, a failure arises. How-
ever, if the domain is reduced by the call exclude(X,El), the constraints
attached to X are not automatically woken. This enables the programmer
to control the program execution more precisely. To wake the constraints
attached to X the programmer needs to call explicitly the built-in wake/0.
To see this consider the following two queries (till the end of this subsection
we simplify displayed information about the delayed goals):

[eclipse 25]: X :: 4..10, Y :: 3..7, X $< Y.

X = X{4 .. 6}
Y = Y{5 .. 7}

There is 1 delayed goal.

Yes (0.00s cpu)
[eclipse 26]: X :: 4..10, Y :: 3..7, X $< Y, X $\= 4.

X = X{[5, 6]}
Y = Y{[6, 7]}

There is 1 delayed goal.

Yes (0.00s cpu)

So the constraint X $\= 4, through constraint propagation, led to a reduc-

tion of the domain of Y from 5..7 to 6..7.
Now suppose that instead of using the constraint X $\= 4 we explicitly
remove 4 from the domain of X using the call exclude(X,4):2

[eclipse 27]: X :: 4..10, Y :: 3..7,

X $< Y, ic_kernel:exclude(X,4).

2 The exclude/2 command is hidden in the ECLi PSe module ic kernel because its behaviour
is non-standard. This explains the use of the ic kernel: prefix.
200 Constraint propagation in ECLi PSe

X = X{[5, 6]}
Y = Y{5 .. 7}

There is 1 delayed goal.

Yes (0.00s cpu)
So, since no wake/0 has been invoked, the final round of the constraint
propagation is not performed and the domain of Y remains here 5..7. The
constraint propagation takes place only after calling wake/0:
[eclipse 28]: X :: 4..10, Y :: 3..7,
X $< Y, ic_kernel:exclude(X,4), wake.

X = X{[5, 6]}
Y = Y{[6, 7]}

There is 1 delayed goal.

Yes (0.00s cpu)
In summary, the constraint propagation is triggered as soon as the CSP
represented by the set of unsolved constraints, i.e., delayed goals, is modified.
This modification can result from an addition of a constraint or from a
removal of an element from the domain of a variable. The set of delayed
constraints is usually called the constraint store. It is helpful to view the
corresponding variable declarations as part of the constraint store, that is
to view the constraint store as a CSP.
The constraint store dynamically changes during the computation. If,
due to constraint propagation, a delayed constraint becomes solved, it is re-
moved from the constraint store. If, due to the constraint propagation, the
constraint store becomes failed, a failure arises (which can trigger backtrack-
ing). In the presence of constraints a computation is successful if it yields
an answer (i.e., does not fail) and the constraint store is empty. These re-
marks about the constraint store apply to arbitrary type of constraints used
in ECLi PSe .3

10.4 Disjunctive constraints and reification

So far, the ECLi PSe programs that generated constraints were deterministic,
in the sense that part of the program used for generating constraints left no
choice points. This is good constraint programming practice.
3 In practice, for efficiency reasons, ECLi PSe does not always detect immediately whether a
constraint is solved.
 10.4 Disjunctive constraints and reification 201

In this section we shall discuss how to set up disjunctive constraints.

The easiest way is to use multiple clauses to define the constraint. Consider
for example the dist(X, Y, Z) constraint that states that |X-Y| $= Z. It
can be naturally defined by the following two clauses:

% dist(X, Y, Z) :- |X-Y| $= Z.

dist(X, Y, Z) :- X-Y $= Z.
dist(X, Y, Z) :- Y-X $= Z.

But this is not good constraint programming practice as it introduces a

choice point during the constraint setup. We now have

[eclipse 29]: X :: [1..4], Y :: [3..6], dist(X, Y, 1).

X = 4
Y = 3
Yes (0.00s cpu, solution 1, maybe more) ? ;

X = X{2 .. 4}
Y = Y{3 .. 5}

Delayed goals:
ic : (X{2 .. 4} - Y{3 .. 5} =:= -1)
Yes (0.00s cpu, solution 2)

This shows that first a solution to the alternative X-Y $= 1, is generated

and subsequently, upon hitting ; (that occurs in the

Yes (0.00s cpu, solution 1, maybe more) ? ;

line), a solution to the second alternative, X-Y $= -1, is generated.

The reason it is not good practice is because choice points belong to the
search routine, after all the constraints have been set up. A key advantage
of constraint programming is that each choice is made after the constraint
propagation takes place. Making choices before all the constraints have been
set up reduces the amount of constraint propagation that can occur and
therefore reduces the chances of making the right choice. More importantly,
it does not allow dynamic reordering of choices at execution time, which we
shall see later is crucial to the design of an efficient search routine.
For the above example it is possible to set up this same disjunctive con-
straint deterministically using the built-in mathematical operator abs/1:
202 Constraint propagation in ECLi PSe

% dist(X, Y, Z) :- |X-Y| $= Z.

dist(X, Y, Z) :- abs(X-Y) $= Z.

This implementation of dist/3 leads to domain reductions through con-

straint propagation before any choices are made:

[eclipse 30]: X :: [1..4], Y :: [3..6], abs(X-Y) $= 1.

X = X{2 .. 4}
Y = Y{3 .. 5}

Delayed goals:
ic : (_706{-1 .. 1} - X{2 .. 4} + Y{3 .. 5} =:= 0)
ic : (1 =:= abs(_706{-1 .. 1}))
ic : (_706{-1 .. 1} =:= +- 1)
Yes (0.00s cpu)

To get all solutions, it is necessary at search time to split a domain, for

example by choosing either X $> 3 or X $=< 3.
A more general way to handle disjunctive constraints is by using reifica-
tion, originally introduced in Subsection 9.4.4. Recall that for each of the
arithmetic constraints $=, $>, $=< etc. there is a ternary, reified, version with
an extra argument. This argument indicates whether the constraint holds or
not. Setting this extra argument to 1 enforces the constraint, while setting
it to 0 enforces its negation. Now, our example constraint dist(X, Y, Z)
can be imposed without introducing any choice points by exploiting these
additional variables. Indeed, if B1 is the extra argument indicating the truth
or falsity of one constraint, and B2 is the extra argument for the second con-
straint, then adding the constraint B1+B2 $= 1 enforces their disjunction:

dist(X, Y, Z) :-
$=(X-Y, Z, B1),
$=(Y-X, Z, B2),
B1+B2 $= 1.

To save introducing these extra arguments we can use the Boolean con-
straint or/2. Thus we can simply write:

dist(X, Y, Z) :- X-Y $= Z or Y-X $= Z.

To properly understand this definition recall that or/2 can be written

in the infix form and that on the account of the ambivalent syntax the
 10.5 Summary 203

arguments of or/2 can be queries. The interaction with ECLi PSe reveals
that this version is actually implemented by translating it into the previous
one using reified constraints:

[eclipse 31]: X :: [1..4], Y :: [3..6], X-Y $= 1 or Y-X $= 1.

X = X{1 .. 4}
Y = Y{3 .. 6}

Delayed goals:
=:=(-(Y{3 .. 6}) + X{1 .. 4}, 1, _914{[0, 1]})
=:=(Y{3 .. 6} - X{1 .. 4}, 1, _1003{[0, 1]})
-(_1003{[0, 1]}) - _914{[0, 1]} #=< -1
Yes (0.00s cpu)

We stress that or/2 is different from the Prolog disjunction ;/2. In fact,
the latter creates a choice point and if we used it here instead of or/2, we
would have reintroduced the initial dist/3 constraint.
Just as or/2 above, other Boolean constraints, i.e. neg/1, and/2, and
=>/2, can be used to form logical combinations of arbitrary constraints with-
out introducing any choice points.
The additional Boolean variables used in a reified constraint can be used
after all the constraints have been set up, during search, to make the choices
that were postponed during constraint setup.
Finally, let us mention that reified constraints cannot be further reified.
For example, the following query yields a run-time error:

[eclipse 32]: and(X, Y, 0) or neg(X, Y).

calling an undefined procedure and(X, Y, 0, _394)
in module eclipse
Abort

because of a failed attempt to reify the arguments of or/2.

10.5 Summary
In this chapter we discussed various facilities present in the libraries sd and
ic.
In the case of the sd library we explained how variables ranging over fi-
nite symbolic domains can be introduced and how constraints on these vari-
ables can be generated. In the case of the ic library we explained Boolean
204 Constraint propagation in ECLi PSe

constraints and arithmetic constraints on integers and reals. Finally, we

discussed reification as a means to handle disjunctive constraints.
From the point of view of programming two features were of crucial impor-
tance: the built-in facilities allowing us to generate CSPs, and the built-in
constraint propagation for the considered classes of constraints.
In the process we also introduced the following built-ins of ECLi PSe rel-
evant for constraint programming:
alldifferent/1,
sum/1,
abs/1,
squash/3,
exclude/2,
wake/0.

10.6 Exercises

Exercise 10.1 Suppose that the library ic is loaded. What is the ECLi PSe
output to the following queries? (Note that these are the queries from
Exercise 3.1 with is replaced by #=.)
1. X #= 7, X #= 6.

2. X #= 7, X #= X+1.

3. X #= X+1, X #= 7.

4. X #= 3, Y #= X+1.

5. Y #= X+1, X #= 3.

Exercise 10.2 Write a program that formalizes representation No. 4 of

the SEND + MORE = MONEY puzzle from Subsection 6.4.2 using the ic
library.

Exercise 10.3 Write a program that formalizes the first representation of

the n-queens problem from Subsection 6.4.3 using the ic library.
 11

Top-down search with active constraints

11.1 Introduction 205

11.2 Backtrack-free search 206
11.3 Shallow backtracking search 208
11.4 Backtracking search 211
11.5 Variable ordering heuristics 215
11.6 Value ordering heuristics 218
11.7 Constructing specific search behaviour 220
11.8 The search/6 generic search predicate 223
11.9 Summary 227
11.10 Exercises 229

11.1 Introduction

T HE PROBLEM OF solving finite CSPs is central to this book. Ini-

tially, in Chapter 8, we explained how to solve them in Prolog. In
Chapter 10 we noted that it is considerably more efficient to rely on
the constraint propagation automatically provided in the ic library. How-
ever, in the presence of constraint propagation the domains of the variables
shrink and it is important to organise the top-down search in such a manner
that during the search only the reduced domains and not the original do-
mains are considered. Additionally, as already mentioned in Chapter 6, the
top-down search can be complete or incomplete and can be set up in several
ways by specifying the variable ordering and the ordering of the domain val-
ues. The ic library provides a support for these aspects of top-down search.
The purpose of this chapter is to explain these facilities and illustrate them
by discussing various forms of top-down search.
Throughout the chapter we assume that the ic library is loaded. We
illustrate the impact of various heuristics by discussing corresponding so-
lutions to the n-queens problem and considering in each case the number

205
206 Top-down search with active constraints

of backtracks. To this end we use the clever way of counting discussed in

Section 8.6.

11.2 Backtrack-free search

We begin our exposition by programming in ECLi PSe the backtrack-free
search. It attempts to construct a solution to a finite CSP as follows.
We assume some fixed variable ordering. We start with the empty assign-
ment (of values to decision variables). Then successively we try to extend
the current partial assignment defined on some variables to the next vari-
able in the assumed ordering. This is done by assigning to each variable
the minimum (or maximum) value in its current domain. If by proceeding
this way the final variable gets instantiated, a solution is found. Otherwise
a failure is reported.
This search procedure relies crucially on the constraint propagation in that
each time the current and not the original variable domain is used. This
makes it a meaningful search procedure, as opposed to the corresponding
backtrack-free search in Prolog. However, this search procedure is clearly
incomplete in that it may not find a solution if one exists. Indeed, it can take
a wrong turn already at the beginning, by instantiating the first variable
to a value that does not participate in a solution.
This procedure can be implemented using the get min/2 built-in. Given a
variable X, the query get min(X, Min) returns in Min the minimum value in
the current domain of X. So the query get min(X, X) instantiates X to the
minimum value in its current domain. The dual built-in get max/2 returns
the maximum value in the current variable domain.
To illustrate the use of the backtrack-free search let us return to the
definition of the variable domains and constraints for the n-queens problem
given in the program from Figure 9.7 on page 179. By combining it with an
appropriate use of the get min/2 built-in we obtain a program that realises
for the n-queens problem the backtrack-free search in the presence of the
constraint propagation for arithmetic constraints. It is given in Figure 11.1.
Interestingly, this program generates a solution for only two values of
N < 500, namely 5 and 7. For N equal to 7 we get the following solution:
[eclipse 1]: queens(QueenStruct, 7).

QueenStruct = [](1, 3, 5, 7, 2, 4, 6)
Yes (0.00s cpu, solution 1, maybe more) ?
In fact, for N equal to 7, after placing the first three queens the constraint
 11.2 Backtrack-free search 207

queens(QueenStruct, Number) :-
dim(QueenStruct,[Number]),
constraints(QueenStruct, Number),
backtrack_free(QueenStruct).

constraints(QueenStruct, Number) :-
( for(I,1,Number),
param(QueenStruct,Number)
do
QueenStruct[I] :: 1..Number,
(for(J,1,I-1),
param(I,QueenStruct)
do
QueenStruct[I] $\= QueenStruct[J],
QueenStruct[I]-QueenStruct[J] $\= I-J,
QueenStruct[I]-QueenStruct[J] $\= J-I
)
).

backtrack_free(QueenStruct):-
( foreacharg(Col,QueenStruct) do get_min(Col,Col) ).

Fig. 11.1 A backtrack-free solution to the n-queens problem

propagation already determines the unique positions for the remaining four
queens. This can be seen by omitting the final call to the backtrack_free/1
and by executing the following query:

[eclipse 2]: queens([](1,3,5,X4,X5,X6,X7), 7).

X4 = 7
X5 = 2
X6 = 4
X7 = 6
Yes (0.00s cpu)

In case the decision variables are organised in a list, the backtrack-free

search procedure is defined as follows:
208 Top-down search with active constraints

backtrack_free(List) :-
( foreach(Var,List) do get_min(Var,Var) ).

The following query illustrates the effect of this search procedure:

[eclipse 3]: List = [X,Y,Z], List :: 1..3,

alldifferent(List), backtrack_free(List).

List = [1, 2, 3]
X = 1
Y = 2
Z = 3
Yes (0.00s cpu)

11.3 Shallow backtracking search

By a shallow backtracking search we mean a search that differs from
the backtrack-free search in that only a limited form of backtracking is
allowed, inside the value selection process. More precisely, in the shallow
backtracking search, when we try to assign a value to the current variable,
the constraint propagation process is automatically triggered. If it yields a
failure, the next value is tried. This process continues until the first value
is found for which the constraint propagation does not yield a failure. Once
such a value is found its choice is final and we proceed to the next variable.
As before, if by proceeding this way the final variable gets instantiated, a
solution is found. Otherwise a failure is reported.
Again, this search procedure is incomplete so a failure does not necessarily
mean that no solution exists. To encode this search we need built-ins that
allow us to access the current variable domain.
One of such built-ins is get domain as list/2 that takes as the first
argument a finite domain variable X and generates in the second argument
the ordered list of the values of the current domain of X. For example:

[eclipse 4]: X :: [1..4], get_domain_as_list(X, Domain),

X #\= 2, get_domain_as_list(X, NewDomain).

X = X{[1, 3, 4]}
Domain = [1, 2, 3, 4]
NewDomain = [1, 3, 4]
Yes (0.00s cpu)
 11.3 Shallow backtracking search 209

Once the list Domain is produced, we can enumerate all its members by
means of the member/2 predicate, as in the following query:

[eclipse 5]: X :: [1..4], X #\= 2,

get_domain_as_list(X, Domain), member(X, Domain).

X = 1
Domain = [1, 3, 4]
Yes (0.00s cpu, solution 1, maybe more) ? ;

X = 3
Domain = [1, 3, 4]
Yes (0.00s cpu, solution 2, maybe more) ? ;

X = 4
Domain = [1, 3, 4]
Yes (0.00s cpu, solution 3)

This combination of get_domain_as_list/2 and member/2 is so often

used that ECLi PSe provides a built-in indomain/1, defined logically as
indomain(X) :-
get_domain_as_list(X, Domain),
member(X, Domain).
but implemented more efficiently, without constructing the Domain list.
We can now define the shallow backtracking search for a list of variables
by means of the following procedure:

shallow_backtrack(List) :-
( foreach(Var,List) do once(indomain(Var)) ).

Recall from Section 4.2 that once(Q) generates only the first solution to the
query Q.
The following two queries illustrate the difference between the backtrack-
free search and shallow backtracking:

[eclipse 6]: List = [X,Y,Z], X :: 1..3, [Y,Z] :: 1..2,

alldifferent(List), backtrack_free(List).

No (0.00s cpu)
[eclipse 7]: List = [X,Y,Z], X :: 1..3, [Y,Z] :: 1..2,
alldifferent(List), shallow_backtrack(List).
210 Top-down search with active constraints

List = [3, 1, 2]
X = 3
Y = 1
Z = 2
Yes (0.00s cpu)

As an example of the use of shallow backtracking, let us return to the

SEND + MORE = MONEY puzzle and reconsider the send/1 procedure
from Figure 10.2 on page 195, the one in which only the variable domains
and constraints were defined. Then the unique solution to this puzzle can be
generated by adding to it the shallow backtracking search, that is by means
of the program in Figure 11.2.

send(List):-
List = [S,E,N,D,M,O,R,Y],
List :: 0..9,
diff_list(List),
1000*S + 100*E + 10*N + D
+ 1000*M + 100*O + 10*R + E
$= 10000*M + 1000*O + 100*N + 10*E + Y,
S $\= 0,
M $\= 0,
shallow_backtrack(List).

shallow_backtrack(List) :-
( foreach(Var,List) do once(indomain(Var)) ).
augmented by the diff list/2 procedure of Figure 9.4.

Fig. 11.2 The SEND MORE MONEY program with shallow backtracking

Indeed, we have then:

[eclipse 8]: send(L).

L = [9, 5, 6, 7, 1, 0, 8, 2]
Yes (0.00s cpu)
 11.4 Backtracking search 211

11.4 Backtracking search

We considered already the backtracking search for CSPs on finite domains
in the example programs discussed in Section 9.7. In these programs the
backtracking search was realised by a systematic enumeration of the original
variable domains. This ensured completeness of this search procedure.
However, in the presence of the ic library this approach does not take
into account the fact that the domains of the variables shrink during the
constraint propagation phase. As a result, during the so defined search
procedure values are generated that lie outside the current variable domains,
which is completely unneeded. A more appropriate approach consists of
enumerating the current variable domains. This brings us to a revision of
the generic search procedure introduced in Figure 8.6 on page 138.
Recall that this procedure required the initial domain to be explicitly asso-
ciated with each variable and the choose val/2 predicate used the variable
domain as an input argument. With the ic library loaded we can access
instead the current variable domain using the get domain as list/2 built-
in. However, for the purpose of the backtracking process it is more efficient
to use instead the indomain/1 built-in which both retrieves a value from
the current domain and instantiates the variable to that value.
As a result, we can simplify the generic search procedure from Figure
8.6 to the one shown in Figure 11.3. In the presence of finite domains this
procedure emerged as a standard in constraint programming.

% search(List) :- Assign values from their domains

% to all the variables in List.

search_with_dom(List) :-
( fromto(List, Vars, Rest, [])
do
choose_var(Vars, Var, Rest),
indomain(Var).
).

choose_var(List, Var, Rest) :- List = [Var | Rest].

Fig. 11.3 Generic search using associated variable domains

Because this generic search procedure is so common, it is available as a

built-in predicate in the ic library that for historical standardisation rea-
212 Top-down search with active constraints

sons is named labeling/1. (indomain/1 and labeling/1 built-ins are also

available in the sd library.) As in the choose_var/2 procedure the variable
ordering of the original input list is adopted, the definition of labeling/1
can be simplified to

labeling(List):- ( foreach(Var,List) do indomain(Var) ).

The following query illustrates this form of search:

[eclipse 9]: List = [W,X,Y,Z], W::1..4, [X,Y,Z] :: 1..3,

alldifferent(List), labeling(List).

List = [4, 1, 2, 3]
W = 4
X = 1
Y = 2
Z = 3
Yes (0.00s cpu, solution 1, maybe more) ? ;

In total 6 solutions are produced. In contrast, backtrack-free/1 and

shallow_backtrack/1 procedures both fail here.
Conceptually, for a list of variables List, the query labeling(List) re-
alises a backtracking search for the generated CSP on the variables in List
that takes into account the underlying constraint propagation process. To il-
lustrate the gain achieved by the presence of constraint propagation consider
the following query:

[eclipse 10]: List = [X,Y,Z], List :: 1..100,

X $> Y, Y*Y $>= Z, Y + Z $>= 2*X + 50.

List = [X{5 .. 49}, Y{4 .. 48}, Z{12 .. 100}]

X = X{5 .. 49}
Y = Y{4 .. 48}
Z = Z{12 .. 100}

There are 4 delayed goals.

Yes (0.00s cpu)

So the constraint propagation led to a substantial domain reduction but was

not sufficient to find a solution and we need to perform a search. This can
be done by just invoking labeling/1:
 11.4 Backtracking search 213

[eclipse 11]: List = [X,Y,Z], List :: 1..100,

X $> Y, Y*Y $>= Z, Y + Z $>= 2*X + 50,
labeling(List).

List = [9, 8, 60]

X = 9
Y = 8
Z = 60
Yes (0.00s cpu, solution 1, maybe more) ?
Several more solutions are returned upon backtracking.
During the execution of this query the call labeling(List) unfolds into
indomain(X), indomain(Y), indomain(Z),
so the backtracking involves the enumeration of the elements of the cur-
rent domain of X, which is [5..49]. Additionally, each time the calls
indomain(Y) and indomain(Z) are reached, the current domains of Y and
Z are reduced to respective subsets of [4..48] and [12..100] because of
the triggered constraint propagation process. If we used here instead the
backtracking search as in Section 9.7, the search would involve the original,
larger, domains of X, Y and Z which are all [1..100].
As another illustration of the difference between the backtracking search
here considered and the one used in Chapter 9 let us return again to the
definition of the domains and constraints for the n-queens problem used in
the program from Figure 9.7 on page 179. By combining it with an appro-
priate use of indomain/1 we obtain a program that takes full advantage of
the constraint propagation during the search process. The program is given
in Figure 11.4.
With the ic library loaded this program is considerably faster than the one
given in Figure 9.7 on page 179 in which the search involved the enumeration
of the original domains. For example, it finds the first solution to the 20-
queens problem 21 times faster.
In general it is important to realise which variables should be passed as
arguments to the labeling procedure. To see this recall from Subsection
10.3.2 the sigma/3 predicate that was defined by
sigma(As, Xs, V) :-
( foreach(A,As),
foreach(X,Xs),
foreach(Expr,Exprs)
do
214 Top-down search with active constraints

queens(QueenStruct, Number) :-
dim(QueenStruct,[Number]),
constraints(QueenStruct, Number),
backtrack_search(QueenStruct).

constraints(QueenStruct, Number) :-
( for(I,1,Number),
param(QueenStruct,Number)
do
QueenStruct[I] :: 1..Number,
(for(J,1,I-1),
param(I,QueenStruct)
do
QueenStruct[I] $\= QueenStruct[J],
QueenStruct[I]-QueenStruct[J] $\= I-J,
QueenStruct[I]-QueenStruct[J] $\= J-I
)
).

backtrack_search(QueenStruct):-
( foreacharg(Col,QueenStruct) do indomain(Col) ).

Fig. 11.4 A backtracking solution to the n-queens problem

Expr = A*X
),
sum(Exprs) $=< V.

and that generated the constraint

ni=1 ai xi v.

Here n 1, all ai s are given parameters and all xi s and v are variables
ranging over an integer interval.
The appropriate labelling involves here only the variables in the list Xs
and V. Indeed, first of all we wish to call sigma/3 with the first argument in-
stantiated, so the variables in As do not need to be enumerated. Second, the
local variables A,X,Expr and Exprs are only used to construct the relevant
constraint and should not be enumerated. Finally, if we always call sigma/3
with the last argument V instantiated, its enumeration is not needed either.
 11.5 Variable ordering heuristics 215

11.5 Variable ordering heuristics

In the implementations of search provided so far, i.e. in backtrack_free/1,
shallow_backtrack/1 and the labeling/1 procedures, the ordering of the
variables as in the input variable list is adopted and in each domain the
values are enumerated proceeding from the smallest one to the largest one.
To solve a computationally intensive problem it is often useful to use different
heuristics that allow us to alter these orderings. They naturally fall into two
categories: the variable ordering heuristics and the value ordering heuristics.
To illustrate the possible choices for the first category we construct a
generic n-queens program with a parameter Heur that we will use to distin-
guish between alternative heuristics. We use the definition of the domains
and constraints for the n-queens problem as in the program from Figure 11.4
but add a call to the struct to list/2 predicate that converts a structure
(here an array) to a list by simply stepping through the arguments and
copying them into a list. The generic search/2 procedure has now the ex-
tra parameter Heur. The program is given in Figure 11.5. To complete it we
now systematically add to it the clauses defining the search/2 procedure.

naive heuristic
This heuristic simply reproduces the behaviour of the built-in labeling/1.
So it labels the variables in the order they appear in the list and tries values
starting with the smallest element in the current domain and trying the rest
in increasing order.
Let us call this heuristic naive and define the first clause of search/2
accordingly:

search(naive, List) :- labeling(List).

The naive heuristic reproduces the behaviour of the program from Figure
11.4 on page 214 and finds a solution to the 8-queens problem after only 10
backtracks, and can solve the 16-queens problems with 542 backtracks, but
the number of backtracks grows very quickly with an increasing number of
queens. With a limit of 50 seconds, even the 32-queens problem times out.

middle out heuristic

For the n-queens problem it makes sense to start labelling queens near the
centre of the board earlier. Indeed, a queen placed near the centre of the
board tends to constrain the others more than a queen near the edge and
216 Top-down search with active constraints

% queens(Heur, Queens, Number) :-

% Solve the Number-queens problem using
% heuristic Heur, generating answer Queens.

queens(Heur, Queens, Number) :-

dim(QueenStruct,[Number]),
constraints(QueenStruct, Number),
struct_to_list(QueenStruct, Queens),
search(Heur, Queens).

constraints(QueenStruct, Number) :-
( for(I,1,Number),
param(QueenStruct,Number)
do
QueenStruct[I] :: 1..Number,
(for(J,1,I-1),
param(I,QueenStruct)
do
QueenStruct[I] $\= QueenStruct[J],
QueenStruct[I]-QueenStruct[J] $\= I-J,
QueenStruct[I]-QueenStruct[J] $\= J-I
)
).

struct_to_list(Struct, List) :-
( foreacharg(Arg,Struct),
foreach(Var,List)
do
Var = Arg
).

Fig. 11.5 Generic n-queens search program

hence yields a stronger constraint propagation. This can be achieved by re-

ordering the list of queens so as to start with the middle queen and hop back
and forth to the next queen before and the next queen after. The reordering
is encoded using three predefined predicates, defined in an ECLi PSe library
called lists:
 11.5 Variable ordering heuristics 217

:- lib(lists).

middle_out(List, MOutList) :-
halve(List, FirstHalf, LastHalf),
reverse(FirstHalf, RevFirstHalf),
splice(LastHalf, RevFirstHalf, MOutList).
The call halve(List, Front, Back) splits its first argument in the middle
while the call splice(Odds, Evens, List) merges the first two lists by
interleaving their elements.
The middle out heuristic simply reorders the list of queens and labels
them as before:
search(middle_out, List) :-
middle_out(List, MOutList),
labeling(MOutList).
Interestingly, the query queens(middle out, Queens, 8) succeeds with-
out any backtracking, and a first solution to the 16-queens problem now
needs only 17 backtracks. However, for 32 queens the search still times out
after 50 seconds.

first fail heuristic

first fail is a widely applicable heuristic that selects as the next variable
to label the one with the fewest values remaining in its domain. This is a
dynamic heuristic, as the sizes of the domains of the unlabelled variables are
reduced during search. Indeed, for the n-queens problem all the variables
initially have domains of the same size, so any improvement in efficiency
brought by the first fail heuristic depends upon changes in the domain
sizes which have occurred earlier during the search.
This heuristic is supported by the delete/5 built-in of the library ic.
The call
delete(Var, Vars, Rest, Arg, Heur)
removes a variable Var from a list of variables Vars, leaving the remaining
list Rest. The parameter Arg enables this predicate to handle not only lists
of variables, but more generally lists of compound terms in which the variable
occurs as an argument. The final parameter Heur is the criterion used to
select the variable to remove. For our current heuristic this parameter is
input as first fail.
This heuristic is encoded as the following clause for search/2:
218 Top-down search with active constraints

search(first_fail, List) :-
( fromto(List, Vars, Rest, [])
do
delete(Var, Vars, Rest, 0, first_fail),
indomain(Var)
).
On the small problem instances of the n-queens problem such as 8-queens
the impact of first fail is unspectacular. Indeed, the search requires 10
backtracks to find the first solution to 8-queens, which is no better than the
naive heuristic. However, on larger instances its impact is dramatic. Indeed,
this heuristic enables a first solution to the 32-queens problem to be found
with only four backtracks. Even the 75-queens problem succumbs after 818
backtracks.

middle out combined with first fail (moff) heuristic

Naturally, the static variable ordering heuristic middle out can be combined
with the dynamic first fail heuristic, to create a new heuristic, which
we shall call here moff. The following clause for search/2 encodes this
combined heuristic:
search(moff,List) :-
middle_out(List, MOutList),
( fromto(MOutList, Vars, Rest, [])
do
delete(Var, Vars, Rest, 0, first_fail),
indomain(Var)
).
The moff heuristic solves the 8-queen and 16-queen problem without any
backtracking, and also solves the 75-queens problem, but it still needs 719
backtracks before finding the first solution.

11.6 Value ordering heuristics

The middle out heuristic was motivated by our observation that the middle
queens had more impact on the remaining queens. This also holds for value
choice: choosing a location near the centre of the board will also lead to
more conflicts. Since the middle positions must be filled sooner or later, it
makes sense to try and fill them earlier and propagate the consequences of
these choices.
 11.6 Value ordering heuristics 219

ECLi PSe provides an indomain/2 built-in whose second argument gives

the programmer control over the order in which values are tried. In particu-
lar indomain(Var,middle) tries first values near the middle of the current
domain of Var, and only explores the minimum and maximum values after
all the others. In turn, indomain(Var,min) reproduces the behaviour of
indomain(Var) while indomain(Var,max) starts the enumeration from the
largest value downwards.
Our final heuristic for the n-queens problem combines the moff variable
ordering heuristic (i.e., the middle-out static variable ordering combined
with the first-fail dynamic variable ordering) and the middle-out value or-
dering heuristic. We call it the moffmo heuristic and encode it as follows:

search(moffmo, List) :-
middle_out(List, MOutList),
( fromto(MOutList, Vars, Rest, [])
do
delete(Var, Vars, Rest, 0, first_fail),
indomain(Var,middle)
).

The number of backtracks for each introduced heuristic on various sizes

of the n-queens problem is summarised in Table 11.6.

naive middle out first fail moff moffmo

8-queens 10 0 10 0 3
16-queens 542 17 3 0 3
32-queens 4 1 7
75-queens 818 719 0
120-queens 0

Table 11.1 Number of backtracks for solving n-queens with different heuristics

Of course for different combinatorial problems different value ordering

heuristics may be advantageous. For example, for graph colouring problems
the value ordering heuristic which works best is to try to avoid re-using
colours. Since a standard value ordering method such as the one used
in indomain/1 will always try the colours in the same order, it leads to
colour reuse by default. A better value ordering heuristic for graph-colouring
is therefore indomain(Var,random).
220 Top-down search with active constraints

For scheduling and packing problems it is usually best to label each vari-
able by its smallest possible value, so using indomain(Var,min). (In fact
indomain(Var) uses this same value ordering.)

11.7 Constructing specific search behaviour

The n-queens problem is symmetrical in that you can either place a queen
in every column, or equivalently place a queen in every row. Indeed, if the
problem is solved by placing a queen at the ni th square in each column i,
it is also solved by placing a queen in the ni th square in every row. The
solutions can be shown to be the same by just turning the board 90 degrees.
Consequently, when the middle out variable ordering heuristic finds a
solution without backtracking, then so should the indomain(Var,middle)
value ordering heuristic. We can now take a step further and create a mod-
ification of the indomain(Var,middle) value ordering heuristic that back-
tracks in exactly the same way as the middle out variable ordering heuristic.
By rotating the output the resulting program produces the same solutions
in the same order.
To achieve this search behaviour we use the middle out/2 predicate de-
fined on page 216 but this time to appropriately reorder the domain values,
through a call

middle_out(OrigDom, MOutDom).

Next, instead of iterating over the list of decision variables (using

foreach(Var,MOutList) for the middle out variable ordering heuristics)
we now iterate over the domain values (using foreach(Val,MOutDom)). Fur-
ther, instead of selecting a value for the current decision variable through
a call to indomain/1, we now select a variable to which a given value is
assigned, through an appropriate call to member/2, as in:

[eclipse 12]: member(2, [X,Y,Z]).

X = 2
Y = Y
Z = Z
Yes (0.00s cpu, solution 1, maybe more) ? ;

X = X
Y = 2
Z = Z
 11.7 Constructing specific search behaviour 221

Yes (0.00s cpu, solution 2, maybe more) ? ;

X = X
Y = Y
Z = 2
Yes (0.00s cpu, solution 3)
Finally, to make the output the same, we rotate the board at the end,
so that if in the solution the ith queen is in square j, then in the output the
jth queen is in square i.
Accordingly, the n-queens program in Figure 11.6 produces exactly the
same results in the same order as the middle out variable ordering heuristic,
that is the queries
queens(middle_out, Queens, Number).
and
queens(rotate, Queens, Number).
produce the same solutions in the same order. Surprisingly, when we count
the number of backtracks required to find the first solution, we get the results
in Table 11.2.

middle out rotate

8-queens 0 0
16-queens 17 0
24-queens 194 12
27-queens 1161 18
29-queens 25120 136

Table 11.2 Number of backtracks for solving n-queens with middle out and rotate
heuristics

This difference occurs because the constraint propagation behaviour is

not the same for columns and for rows. In our model we have a single
(queen) variable for each column whose value is the square that the queen
occupies in that column. If all the values for a column (i.e. for a variable)
are excluded except one, then this variable is immediately instantiated to
the remaining value. This is the behaviour exhibited by the rotate variable
ordering heuristics.
However, if all the values in a row are excluded except one, then in our
222 Top-down search with active constraints

queens(rotate, Queens, Number) :-

dim(QueenStruct,[Number]),
constraints(QueenStruct, Number),
struct_to_list(QueenStruct, QList),
search(rotate, QList),
rotate(QueenStruct, Queens).

search(rotate, QList) :-
middle_out_dom(QList, MOutDom),
( foreach(Val,MOutDom),
param(QList)
do
member(Val, QList)
).

middle_out_dom([Q | _], MOutDom) :-

get_domain_as_list(Q,OrigDom),
middle_out(OrigDom, MOutDom).

rotate(QueenStruct, Queens) :-
dim(QueenStruct,[N]),
dim(RQueens,[N]),
( foreachelem(Q,QueenStruct,[I]),
param(RQueens)
do
subscript(RQueens,[Q],I)
),
struct_to_list(RQueens, Queens).
augmented by the constraints/2 and struct to list/2 procedures of
Figure 11.5 and middle out/2 procedure from page 216.

Fig. 11.6 Rotated middle out heuristic for n-queens

model it means that only one (queen) variable remains which still has a
square in that row in its domain. In this case, which corresponds to the
middle out variable ordering heuristics, constraint propagation does not
instantiate the variable. This phenomenon is illustrated in Figure 11.7. In
row 5 only the field e5 is available. Since in every row exactly one queen
 11.8 The search/6 generic search predicate 223

needs to be placed, it follows that in column 5 the queen has to be placed

in the field e5. However, the variables represent columns, so the constraint
propagation only reduces the domain of the variable representing column e
to the set {2,3,5}.

8
7

6
5
4
3
2
1
a b c d e f g h

Fig. 11.7 Limitation of constraint propagation

The way to achieve the same amount of constraint propagation and ulti-
mately the same number of backtracks for the two symmetrical heuristics
is to add another board and link the two, so that if queen i is placed on
square j on one board, then queen j is automatically placed on square i
on the other. This redundant modelling of the problem improves all the n-
queens problem benchmark results, and is a typical way to improve problem
solving efficiency.

11.8 The search/6 generic search predicate

The different variable and value choice heuristics, together with a number
of incomplete search options are provided through a single generic ECLi PSe
predicate called search/6. The six arguments of
search(List, Arg, VarChoice, ValChoice, Method, Options)
are best illustrated by some examples of its use.
The first argument, List, is a list of decision variables when the second
argument, Arg, is 0 or a list of compound terms when the second argu-
ment, Arg, is > 0. Assume for a moment that Arg = 0. By instantiating
appropriately the third argument, VarChoice, to the (predicate) name of a
224 Top-down search with active constraints

predefined or user-defined variable ordering heuristic and the fourth argu-

ment, ValChoice, to the (predicate) name of a predefined or user-defined
value ordering heuristic, we can encode different combinations of heuristics.
As an example in Figure 11.8 we define using the search/6 predicate the
already discussed heuristics for the n-queens problem.

% search(Heur, List) :-
% Find a labelling of List using
% the combination Heur of heuristics.

search(naive, List) :-
search(List,0,input_order,indomain,complete,[]).

search(middle_out, List) :-
middle_out(List, MOutList),
search(MOutList,0,input_order,indomain,complete,[]).

search(first_fail, List) :-
search(List,0,first_fail,indomain,complete,[]).

search(moff, List) :-
middle_out(List, MOutList),
search(MOutList,0,first_fail,indomain,complete,[]).

search(moffmo, List) :-
middle_out(List, MOutList),
search(MOutList,0,first_fail,
indomain_middle,complete,[]).
augmented by the middle out/2 procedure from page 216.

Fig. 11.8 Generic search for solving n-queens

So input order and first fail are predefined variable ordering heuris-
tics corresponding to naive and first fail, while indomain and
indomain middle are predefined value ordering heuristics corresponding to
indomain(Var,min) (i.e., indomain(Var)) and indomain(Var,middle).
In all the clauses the fifth argument, Method, is instantiated to complete
which refers to a complete search. Other, incomplete, forms of search are
 11.8 The search/6 generic search predicate 225

also allowed. For example, by instantiating Method to lds(2) we specify a

limited discrepancy search with up to two discrepancies (see Section 8.4).
The sixth argument, Options, allows the programmer to specify options,
such as counting backtracks, or drawing the search tree as a graph. An
example use might be as in the following query:

search(Queens,0,first_fail,
indomain_middle,lds(2),[backtrack(B)]),
writeln(backtracks - B).

Finally, consider the second argument, Arg, of search/6 when Arg > 0.
As already mentioned in that case the first argument, List, should be a list
of compound terms. The idea is that from an object-oriented viewpoint it
makes sense to admit any list of compound terms (representing objects) as
input, as long as during search we can extract the decision variable from the
current compound term. This is done by instantiating the second argument,
Arg, to the argument of each term that contains the decision variable.
Suppose for example that we keep with each queen variable an integer
recording which column this queen belongs to. Each queen will then be
represented by the term q(Col,Q), where Col is an integer recording the
column and Q a domain variable which will represent the queens position.
The first argument, List, of search/6 will then be a list of such terms. In
this case the decision variable is the second argument of each term, so the
second argument of search/6 is set to 2.
We mentioned already that the VarChoice and ValChoice parameters of
search/6 can be instantiated to user-defined heuristics. Let us illustrate
now the usefulness of this facility. Suppose that for performance debug-
ging purposes of a variable ordering heuristic for a given search method
(such as limited discrepancy search) we want to output the column and
the position whenever a queen is labelled, so that we know the order in
which the queens are being selected for labelling. To do this we introduce a
user-defined show choice value ordering heuristic that displays the choice
currently being made, and set the fourth argument, ValChoice, of search/6
to show choice. The resulting program is shown in Figure 11.9.
The following example interaction with this program illustrates the be-
haviour of the first fail variable ordering heuristic for lds(2), the limited
discrepancy search with up to two discrepancies.

[eclipse 13]: queens(debug, Queens, first_fail, lds(2), 6).

col(1) - square(1)
col(2) - square(3)
226 Top-down search with active constraints

queens(debug, Queens, VarChoice, Method, Number) :-

dim(QueenStruct,[Number]),
constraints(QueenStruct, Number),
struct_to_list(QueenStruct, Queens),
struct_to_queen_list(QueenStruct, QList),
search(QList, 2, VarChoice, show_choice, Method, []).

struct_to_queen_list(Struct, QList) :-
( foreacharg(Var,Struct),
count(Col,1,_),
foreach(Term,QList)
do
Term = q(Col, Var)
).

show_choice(q(I, Var)) :-
indomain(Var),
writeln(col(I):square(Var)).
augmented by the constraints/2 and struct to list/2 procedures of
Figure 11.5.

Fig. 11.9 Code for performance debugging of n-queens

col(1) - square(1)
col(2) - square(3)
col(2) - square(4)
col(1) - square(2)
col(2) - square(4)
col(3) - square(6)
col(4) - square(1)
col(5) - square(3)
col(6) - square(5)

Queens = [2, 4, 6, 1, 3, 5]
Yes (0.00s cpu, solution 1, maybe more) ?

We can now illustrate the flexibility of the search/6 predicate, by using

it to run our own credit based value ordering heuristic procedure. The pro-
cedure share_credit/2 defined in Figure 8.11 on page 143 assigned more
 11.9 Summary 227

credit to our earlier choices and less to later ones. To interface to such proce-
dures the fourth argument, ValChoice, of search/6 allows us to thread an
extra argument, for example the initial credit Var, through the procedure.
An example interface is shown in Figure 11.10 where the ValChoice argu-
ment of search/6 is instantiated to our_credit_search(Var, _). The re-
sulting search/2 procedure also uses the first fail variable choice heuris-
tic, to show the orthogonality of variable and value ordering heuristics, how-
ever complex these heuristics may be.
To complete the chapter we present experimental results in Table 11.3 that
show that each time we change the amount of credit, we also get different
backtracking behaviour. This suggests that insufficient credit is eliminating
some correct choices, and consequently suggests that our credit based search
is not a good heuristic for solving the n-queens problem.

Credit(N) N=10 N=20 N=40 N=100

8-queens failed 6 10 16
16-queens 7 13 25 45
32-queens 9 9 14 21
75-queens failed failed 7 6
120-queens failed failed 21 3

Table 11.3 Number of backtracks for solving n-queens with different credit limits

11.9 Summary
In this chapter we studied the support provided in the ic library for top-
down search for finite CSPs in the presence of constraint propagation. We
discussed in turn the backtrack-free search, shallow backtracking and back-
tracking search. We also defined various variable and value ordering heuris-
tics and showed how by using them specific forms of search behaviour can
be defined.
Further, we discussed a generic search procedure search/6 that allows
us to customise the top-down search process by using predefined or user-
defined variable ordering and value ordering heuristics, and various, possibly
incomplete, forms of search.
In our presentation we introduced the following ic library built-ins:

get min/2,
228 Top-down search with active constraints

% search(our_credit_search(Var,_), List) :-
% Find a labelling of List using our own credit based
% search procedure with an initial credit of Var.

search(our_credit_search(Var,_), List) :-
search(List, 0, first_fail,
our_credit_search(Var,_), complete, []).

% our_credit_search(Var, TCredit, NCredit) :-

% Select a value for Var reducing the remaining
% credit from TCredit to NCredit.

our_credit_search(Var, TCredit, NCredit) :-

get_domain_as_list(Var,Domain),
share_credit(Domain, TCredit, DomCredList),
member(Val-Credit, DomCredList),
Var = Val,
NCredit = Credit.

% share_credit(DomList, InCredit, DomCredList) :-

% Share the credit in InCredit amongst the
% values in DomList producing DomCredList.

share_credit(DomList, InCredit, DomCredList) :-

( fromto(InCredit, TCredit, NCredit, 0),
fromto(DomList, [Val | Tail], Tail, _),
foreach(Val-Credit, DomCredList)
do
Credit is fix(ceiling(TCredit/2)),
NCredit is TCredit - Credit
).

Fig. 11.10 Credit based search for solving n-queens

get max/2,
get_domain_as_list/2,
indomain/1,
labeling/1,
 11.10 Exercises 229

delete/5,
indomain/2,
search/6.
Throughout the chapter we assumed that the domain splitting consists of
an enumeration of all the domain elements. But other natural possibilities
exist. For example, instead of such a complete domain decomposition we
could instead split the domain only into two parts, one consisting of a single
element and the other of the remaining elements. Alternatively, we could
split the domain into two, roughly equal parts.
These alternatives can be programmed in ECLi PSe in a straightforward
way. We shall return to the last one in Chapter 13 on solving continuous
CSPs, as for finite domains this approach can be easily obtained by a minor
modification of the corresponding domain splitting method there defined.

11.10 Exercises

Exercise 11.1 Write a version of the queens/2 predicate from Figure 11.4
on page 214 that uses lists instead of arrays and recursion instead of itera-
tion.

Exercise 11.2 Write an ECLi PSe program that solves the problem from
Exercise 6.1 on page 108.

Exercise 11.3 Write an ECLi PSe program that solves the problem from
Exercise 6.2 on page 109.

Exercise 11.4 Write an ECLi PSe program that generates magic squares of
order 5. (For their definition see Exercise 6.3 on page 109.)
 12

Optimisation with active constraints

12.1 The minimize/2 built-in 230

12.2 The knapsack problem 232
12.3 The coins problem 234
12.4 The currency design problem 235
12.5 Generating Sudoku puzzles 240
12.6 The bb min/3 built-in 246
12.7 When the number of variables is unknown 253
12.8 Summary 254
12.9 Exercises 255

L ET US TURN our attention now to the problem of solving finite

constrained optimisation problems (COPs) as defined in Section 6.5.
ECLi PSe allows us to solve them by combining constraint propa-
gation with various forms of the branch and bound search. This form of
search does not need to be programmed explicitly as it is supported by the
branch and bound library. In this chapter we discuss two main built-ins
available in this library.

12.1 The minimize/2 built-in

The most common built-in in this library is minimize/2. Its appropriate use
is best explained by recalling from Chapter 9 the general form of an ECLi PSe
program that solves a constrained optimisation problem with finite domains.
In the current context it looks as follows:
:- lib(ic).
:- lib(branch_and_bound).

solveOpt(List):-

230
 12.1 The minimize/2 built-in 231

declareDomains(List),
generateConstraints_and_Cost(List, Cost),
minimize(search(List), Cost).
The call generateCons and Cost(List, Cost) generates the constraints
and defines the cost function that is subsequently used to find the optimal
solution. The call
minimize(search(List), Cost)
realises in ECLi PSe the basic branch and bound algorithm discussed in
Section 6.6, augmented with the constraint propagation for constraints on
finite domains. It computes, using the call search(List), a solution to the
generated CSP for which the value of the cost function defined in Cost is
minimal. At the time of the call Cost has to be a variable, i.e., it should
not be instantiated.
To illustrate the internal working of minimize/2 consider the following
query:
[eclipse 1]: minimize(member(X,[5,6,5,3,4,2]), X).
Found a solution with cost 5
Found a solution with cost 3
Found a solution with cost 2
Found no solution with cost -1.0Inf .. 1

X = 2
Yes (0.00s cpu)
So ECLi PSe reports information about each newly found improved value
of the cost function. The member/2 built-in generates upon backtracking
the successive elements of the input list, so only the solutions with the cost
5, 3 and 2 are reported. Upon termination of the search the values of the
variables for the first optimal solution found are printed.
If we need to use a non-variable cost function, we introduce it by means
of a constraint, as in the following example:
[eclipse 2]: Cost #= X+1,
minimize(member(X,[5,6,5,3,4,2]), Cost).
Found a solution with cost 6
Found a solution with cost 4
Found a solution with cost 3
Found no solution with cost -1.0Inf .. 2
232 Optimisation with active constraints

Cost = 3
X = 2
Yes (0.00s cpu)

In general, to solve a constrained optimisation problem we need to launch

a complete top-down search, for example using the labeling/1 built-in, in
the presence of the cost function. As an example consider the following
artificial problem.

Problem Find a solution to the equation x3 + y 2 = z 3 such that x, y, z

[100..500], with the minimal value of z x y.

Here is a solution in ECLi PSe :

find(X, Y, Z) :-
[X,Y,Z] :: [100..500],
X*X*X + Y*Y #= Z*Z*Z,
Cost #= Z-X-Y,
minimize(labeling([X,Y,Z]), Cost).

The interaction with ECLi PSe yields the following outcome:

[eclipse 3]: find(X, Y, Z).

Found a solution with cost -180
Found a solution with cost -384
Found no solution with cost -1.0Inf .. -385

X = 110
Y = 388
Z = 114
Yes (12.34s cpu)

Of course, in general the efficiency of the program and the order in which
the values of the cost function are generated depend on the adopted search
strategy.

12.2 The knapsack problem

We now explain how to solve in ECLi PSe the knapsack problem introduced
in Subsection 6.5.1. Let us recall first its formulation as a COP.
We have n objects with volumes a1 , ..., an and values b1 , ..., bn and the
knapsack volume v. Further, we have n Boolean variables x1 , ..., xn . The
 12.2 The knapsack problem 233

inclusion of the object i in a collection is modelled by setting the value of

xi to 1.
So we have just one constraint:
n
X
ai x i v (12.1)
i=1

that models the requirement that the collection fits in the knapsack. The
cost function which we wish to minimise is:
n
X
bi xi .
i=1

It models the requirement that we seek a solution to (12.1) for which the
sum is maximal.
The appropriate program is given in Figure 12.1. It uses twice the pred-
icate sigma/3. Its variant was originally discussed in Subsection 10.3.2,
where we explained how the constraint (12.1) can be generated in the ic
library. The discussion at the end of Section 11.4 explains why the labeling
procedure is applied only to the list Xs.

knapsack(Volumes, Values, Capacity, Xs) :-

Xs :: [0..1],
sigma(Volumes, Xs, Volume),
Volume $=< Capacity,
sigma(Values, Xs, Value),
Cost $= -Value,
minimize(labeling(Xs), Cost).

sigma(List1, List2, Value) :-

( foreach(V1,List1),
foreach(V2,List2),
foreach(Prod,ProdList)
do
Prod = V1*V2
),
Value $= sum(ProdList).

Fig. 12.1 The KNAPSACK program

234 Optimisation with active constraints

As an example assume we have a knapsack with capacity 60 and that we

have five objects with the volumes and values as given in Table 12.1.

volume value

52 100
23 60
35 70
15 15
7 15

Table 12.1 Volumes and values of knapsack objects

We get then:

[eclipse 4]: knapsack([52,23,35,15,7], [100,60,70,15,15],

60, [X1,X2,X3,X4,X5]).

Found a solution with cost 0

Found a solution with cost -15
Found a solution with cost -30
Found a solution with cost -70
Found a solution with cost -85
Found a solution with cost -100
Found a solution with cost -130
Found no solution with cost -260.0 .. -131.0

X1 = 0
X2 = 1
X3 = 1
X4 = 0
X5 = 0
Yes (0.01s cpu)

12.3 The coins problem

To see another example of the use of the minimize/2 built-in let us return
now to the coins problem introduced in Subsection 6.5.2. Recall that the
task was to find the minimum number of euro cent coins that allows us to
pay exactly any amount smaller than one euro.
 12.4 The currency design problem 235

We formalized this problem using six variables x1 , x2 , x5 , x10 , x20 , x50 rang-
ing over the domain [0..99] that denote the appropriate amounts of the coins,
and for each i [1..99] six variables xi1 , xi2 , xi5 , xi10 , xi20 , xi50 that are used to
state that the amount of i cents can be exactly paid. The appropriate con-
straints are:
xi1 + 2xi2 + 5xi5 + 10xi10 + 20xi20 + 50xi50 = i,

0 xij ,

xij xj
for all i [1..99] and j {1, 2, 5, 10, 20, 50} and the cost function is
x1 + x2 + x5 + x10 + x20 + x50 .
The program is a direct translation of this representation into ECLi PSe .
So we use a list Coins of six variables corresponding to the variables x1 , x2 ,
x5 , x10 , x20 , x50 , each ranging over the domain 0..99.
Then for each price between 1 and 99 we impose the above three con-
straints using the predicate price cons/4. The variables xi1 , xi2 , xi5 , xi10 , xi20 ,
xi50 are kept in the list CoinsforPrice and the complete list of these lists,
with i [1..99], is kept in the list Pockets. Though we are not interested
in the values of the variables assembled in the list Pockets, we keep them
so as to be able to ensure there really is a feasible labelling for them after
setting the appropriate amounts of the coins to variables in the list Coins.
Finally, we use sum(Coins) as the cost function. The program is given in
Figure 12.2.
This optimisation problem is easy to solve. The optimal solution of 8 is
found very quickly:
[eclipse 5]: solve(Coins, Min).
Found a solution with cost 8
Found no solution with cost 1.0 .. 7.0

Coins = [1, 2, 1, 1, 2, 1]
Min = 8
Yes (0.08s cpu)

12.4 The currency design problem

In the last problem the values of the six coins were fixed: 1, 2, 5, 10, 20
and 50. But suppose now that we have the freedom of choosing the values
236 Optimisation with active constraints

solve(Coins, Min) :-
init_vars(Values, Coins),
coin_cons(Values, Coins, Pockets),
Min #= sum(Coins),
minimize((labeling(Coins), check(Pockets)), Min).

init_vars(Values, Coins) :-
Values = [1,2,5,10,20,50],
length(Coins, 6),
Coins :: 0..99.

coin_cons(Values, Coins, Pockets) :-

( for(Price,1,99),
foreach(CoinsforPrice,Pockets),
param(Coins,Values)
do
price_cons(Price, Coins, Values, CoinsforPrice)
).

price_cons(Price, Coins, Values, CoinsforPrice) :-

( foreach(V,Values),
foreach(C,CoinsforPrice),
foreach(Coin,Coins),
foreach(Prod,ProdList)
do
Prod = V*C,
0 #=< C,
C #=< Coin
),
Price #= sum(ProdList).

check(Pockets) :-
( foreach(CoinsforPrice,Pockets)
do
once(labeling(CoinsforPrice))
).

Fig. 12.2 The COINS program

 12.4 The currency design problem 237

of these six coins. Can we then design a set of six coins for which we can
solve the above problem with fewer than eight coins? This is the currency
design problem introduced in Subsection 6.5.2. Computationally this is a
much more challenging problem. The reason is that the constraints become
now non-linear. Indeed, recall that each previous constraint
xi1 + 2xi2 + 5xi5 + 10xi10 + 20xi20 + 50xi50 = i
becomes now
v1 xi1 + v2 xi2 + v3 xi3 + v4 xi4 + v5 xi5 + v6 xi6 = i,
where v1 , . . ., v6 are the values of the coins and for each i [1..99], xi1 , . . ., xi6
are the numbers of these coins that allow one to pay the amount i.
On the other hand, from the programming point of view the solution is a
minor modification of the previous program. Indeed, to solve this problem
instead of the list of coin values
Values = [1,2,5,10,20,50]
we simply use now the list
Values = [V1, V2, V3, V4, V5, V6]
of six constraint variables such that 0 #< V1 #< ... #< V6 #< 100, that
represent the values of the coins for which an optimal solution is sought.
Unfortunately, even though the resulting program finds a solution with
eight coins relatively quickly, the computation concerned with the proof of
optimality takes an excessive amount of time.
A remedy consists of using implied constraints that make a big difference
to performance. The idea is that if we have coins with value V 1 and V 2,
where V 1 < V 2, then it is never needed to have enough coins with value V 1
to equal or exceed the value V 2. Indeed, if we had so many V 1-valued coins,
we could pay the amount that uses them using a V 2-valued coin instead.
The reason is that in any solution there are always enough coins to make
up any amount up to V 1 1.
Additionally, for the coin with the largest value we can impose the con-
straint that we only need to pay amounts smaller than 100. The resulting
program is presented in Figure 12.3. The additional constraints are gener-
ated using the predicate clever_cons/2. The increasing/1 predicate is a
simplified version of the ordered/1 predicate introduced in Figure 9.3 on
page 172, in which we do not deal with the case of the empty lists.
The following interaction with ECLi PSe shows that no solution with seven
coins exists.
238 Optimisation with active constraints

design_currency(Values, Coins) :-
init_vars(Values, Coins),
coin_cons(Values, Coins, Pockets),
clever_cons(Values, Coins),
Min #= sum(Coins),
minimize((labeling(Values), labeling(Coins),
check(Pockets)), Min).

init_vars(Values, Coins) :-
length(Values, 6),
Values :: 1..99,
increasing(Values),
length(Coins, 6),
Coins :: 0..99.

increasing(List) :-
( fromto(List,[This,Next | Rest],[Next | Rest],[_])
do
This #< Next
).

clever_cons(Values, Coins) :-
( fromto(Values,[V1 | NV],NV,[]),
fromto(Coins,[N1 | NN],NN,[])
do
( NV = [V2 | _] ->
N1*V1 #< V2
;
N1*V1 #< 100
)
).
augmented by the procedures coin cons/3 and check/1 of Figure 12.2.

Fig. 12.3 The CURRENCY program

[eclipse 6]: design_currency(Values, Coins).

Found a solution with cost 19
[...]
 12.4 The currency design problem 239

Found a solution with cost 8

Found no solution with cost 1.0 .. 7.0

Values = [1, 2, 3, 4, 11, 33]

Coins = [1, 1, 0, 2, 2, 2]
Yes (297.98s cpu)

In other words, it is impossible to design a set of values for 6 coins such

that a set of 7 coins would allow us to pay exactly any amount smaller than
one euro.
Call now a collection of n (possibly identical) coins of some values an n
coins system if it allows us to pay exactly any amount smaller than one
euro. Using the COINS program we showed that an eight coins system exists.
Further, using the additional constraints that we formalized using the
predicate clever_cons/2 we can prove that if an n coins system exists,
then also an n coins system with all coins different exists. Indeed, if a V-
valued coin appears more than once, say N times, then its N th occurrence
can be replaced by a different coin with face value N V . By the addi-
tional constraints it is a new coin. Repeating this replacement process we
end up with an n coins system with all coins different. For example, this
transformation transforms the above solution with

Values = [1, 2, 3, 4, 11, 33]

Coins = [1, 1, 0, 2, 2, 2]

into a solution with

Values = [1, 2, 4, 8, 11, 22, 33, 66]

Coins = [1, 1, 1, 1, 1, 1, 1, 1].

So using the CURRENCY program we showed that no six coins system exists.
The final question is of course whether a seven coins system exists. To
answer it, in view of the above transformation, it suffices to change in the
CURRENCY program the parameter 6 to 7, add the constraint Min #= 7, and
run the query design currency(Values, [1,1,1,1,1,1,1]).
Note also that in this case the call of clever_cons/2 is unneeded. This
yields after 5 seconds a solution

Values = [1, 2, 3, 6, 12, 25, 50].

So an optimal coin system consists of seven coins and 1, 2, 3, 6, 12, 25,

50 is one possible choice.
240 Optimisation with active constraints

12.5 Generating Sudoku puzzles

Sudoku puzzles gained popularity in Japan in the eighties. They became
recently extremely popular throughout the world and nowadays many news-
papers publish them (often referring to some copyrighted websites). A Su-
doku puzzle consists of a 9 by 9 grid divided into 9 squares of size 3 by 3,
partially filled with the digits 1 through 9. The problem consists of filling
the empty fields so that that every row, every column, and every 3 by 3
square contains the digits 1 through 9. It is required that the Sudoku puzzle
has a unique solution. An example is given in Figure 12.4.

2 1 7
8
9 3
8 9 6 1
6 2
5 3
9 4 5 6
7 3 8
2

Fig. 12.4 A Sudoku puzzle

In this section we consider the problem of generating Sudoku puzzles. We

begin with a simpler problem of solving Sudoku puzzles. We represent such
puzzles by means of an 9 by 9 array partially filled with the digits 1 through
9, with the remaining fields filled by the anonymous variables. For example,
here is the array representing the Sudoku puzzle from Figure 12.4:

[](
[](_, 2, 1, _, _, 7, _, _, _),
[](_, _, _, _, _, 8, _, _, _),
[](_, _, _, _, _, _, 9, 3, _),
[](_, 8, _, 9, _, _, 6, _, 1),
[](_, _, _, _, 6, _, _, 2, _),
[](_, _, _, _, _, 5, _, _, 3),
[](_, 9, 4, _, _, _, _, 5, 6),
[](7, _, _, 3, _, _, 8, _, _),
[](_, _, _, _, _, 2, _, _, _))
 12.5 Generating Sudoku puzzles 241

The program solving Sudoku puzzles is very simple: it suffices to state

the alldifferent constraints for each row, column and 3 by 3 square and
launch the search. In the program we use the alldifferent/1 built-in from
the library ic global that leads to a stronger constraint propagation than
the alldifferent/1 built-in from the ic library. For example, we have

[eclipse 7]: [X,Y,Z] :: [0..1], ic:alldifferent([X,Y,Z]).

X = X{[0, 1]}
Y = Y{[0, 1]}
Z = Z{[0, 1]}

There are 3 delayed goals.

Yes (0.00s cpu)
[eclipse 8]: [X,Y,Z] :: [0..1], ic_global:alldifferent([X,Y,Z]).

No (0.00s cpu)

The program is given in Figure 12.5. It uses two built-ins that were not
discussed so far: flatten/2 and term variables/2. The former transforms
a list of lists (in general, a list of lists of . . . lists) into a list. So the query

S is Board[I..I+2,J..J+2], flatten(S, SubSquare).

assigns to the variable SubSquare the list of elements filling the subsquare
Board[I..I+2,J..J+2] of Board. In turn, the term variables/2 built-in
computes the list of variables that appear in a term. It is similar to the
vars/2 predicate defined in Figure 5.2 on page 83 (see also Exercise 5.2 on
page 85). Finally, the multifor([I,J],1,9,3) shorthand stands for the
iteration of I and of J from 1 to 9 in the increments of 3, i.e., through the
values 1,4,7. The solution to the puzzle from Figure 12.4 is easily found:

[eclipse 9]: solve([](

[](_, 2, 1, _, _, 7, _, _, _),
[](_, _, _, _, _, 8, _, _, _),
[](_, _, _, _, _, _, 9, 3, _),
[](_, 8, _, 9, _, _, 6, _, 1),
[](_, _, _, _, 6, _, _, 2, _),
[](_, _, _, _, _, 5, _, _, 3),
[](_, 9, 4, _, _, _, _, 5, 6),
[](7, _, _, 3, _, _, 8, _, _),
[](_, _, _, _, _, 2, _, _, _))).
242 Optimisation with active constraints

solve(Board) :-
sudoku(Board),
print_board(Board).

sudoku(Board) :-
constraints(Board),
search(Board).

constraints(Board) :-
dim(Board,[9,9]),
Board[1..9,1..9] :: [1..9],
( for(I,1,9), param(Board)
do
Row is Board[I,1..9],
alldifferent(Row),
Col is Board[1..9,I],
alldifferent(Col)
),
( multifor([I,J],1,9,3),
param(Board)
do
S is Board[I..I+2,J..J+2],
flatten(S, SubSquare),
alldifferent(SubSquare)
).

search(Board) :-
term_variables(Board, Vars),
labeling(Vars).

print_board(Board) :-
( foreachelem(El,Board,[_,J])
do
( J =:= 1 -> nl ; true ),
write( ),
( var(El) -> write(_) ; write(El) )
).

Fig. 12.5 The SUDOKU program

 12.5 Generating Sudoku puzzles 243

9 2 1 5 3 7 4 6 8
4 3 6 2 9 8 1 7 5
8 5 7 4 1 6 9 3 2
2 8 5 9 7 3 6 4 1
1 7 3 8 6 4 5 2 9
6 4 9 1 2 5 7 8 3
3 9 4 7 8 1 2 5 6
7 6 2 3 5 9 8 1 4
5 1 8 6 4 2 3 9 7
Yes (0.20s cpu)

To generate the Sudoku puzzles we consider two different approaches.

The first program starts by randomly generating a solution to a Sudoku
puzzle. Then repeatedly fields are randomly removed (i.e., replaced by an
anonymous variable) until a problem with more than one solution is created.
The problem just before the last one is then a Sudoku puzzle and is printed.
Using the branch and bound search a Sudoku puzzle with the minimum
number of filled fields is sought and each successively better solution is
printed.
The resulting program is given in Figure 12.6. It relies on three support-
ing predicates, remove_random/3, board_to_list/2 and list_to_board/2
with the expected meaning. They are defined in Figure 12.8 on page 247.
The program performs the branch and bound search using the minimize/2
built-in. The crucial procedure is drop until/3. Its first argument, List,
represents the contents of the initial solution while the second argument,
Final, represents the contents of the final solution, both in a list form. The
third argument, Min, represents the number of (non-dropped) values in the
final solution.
To test that a puzzle has a unique solution we use within the drop until/3
procedure the predicate one/1 defined in Figure 8.16 on page 149. Recall
that the call one(Q) does not instantiate the variables of the query Q. If
there is only one solution, the iteration continues by dropping further values.
Otherwise, if there are multiple solutions, the iteration stops and returns the
previous solution.
The search is launched by the query generate. It fairly quickly generates
puzzles with 2325 filled fields.
In the above approach each printed Sudoku puzzle is a subset of the
original randomly generated solution. The second program takes a dual
approach. It generates a Sudoku puzzle by repeatedly randomly filling fields
244 Optimisation with active constraints
generate :-
sudoku_random(Board),
minimize(search(Board,Min), Min).

sudoku_random(Board) :-
constraints(Board),
term_variables(Board, Vars),
( foreach(X,Vars) do indomain(X,random) ).

search(Board, Min) :-
board_to_list(Board, List),
drop_until(List, Final, Min),
list_to_board(Final, FinalBoard),
print_board(FinalBoard), nl.

drop_until(List, Final, Min) :-

( fromto(List, This, Next, Final),
fromto(ok, _, Cont, stop)
do
remove_random(This, _, Succ),
list_to_board(Succ, Test),
( one(sudoku(Test)) ->
Next = Succ, Cont = ok
;
Next = This, Cont = stop
)
),
length(Final, Min).
augmented by the procedures constraints/1, sudoku/1 and
print board/1 of Figure 12.5; board to list/2, list to board/2
and remove random/3 given below in Figure 12.8;
and one/1 of Figure 8.16.

Fig. 12.6 The GEN SUDOKU program

until a Sudoku puzzle (i.e., a puzzle with a unique solution) is created. The
generated puzzle is printed. Subsequently, using the branch and bound
search a Sudoku puzzle with the minimum number of filled fields is sought.
The program is given in Figure 12.7.
 12.5 Generating Sudoku puzzles 245
generate :-
constraints(Board),
minimize(search2(Board,Min), Min).

search2(Board, Min) :-
board_to_list(Board, List),
add_until(Board, List, Final, Min),
list_to_board(Final, FinalBoard),
print_board(FinalBoard), nl.

add_until(Board, List, Final, Min) :-

( fromto([], Used, [[I,J]-Val | Used], Final),
fromto(List, TFree, NFree, _),
fromto(ok, _, Cont, stop),
count(Ct,1,Min),
param(Board, Min)
do
Min #>= Ct,
remove_random(TFree, [I,J]-Val, NFree),
var(Val),
indomain(Val,random),
ct(search(Board), N),
( N = 1 -> Cont = stop ;
N = multiple -> Cont = ok )
).

ct(Q, N) :-
( not twice(Q) -> getval(x,N) ; N = multiple ).
augmented by the procedures constraints/1, search/1 and
print board/1 of Figure 12.5; board to list/2, list to board/2
and remove random/3 given below in Figure 12.8;
and twice/1 of Figure 8.16.

Fig. 12.7 Another GEN SUDOKU program

The crucial procedure, add until/4, is now more involved. Its first ar-
gument, Board, represents the constrained Sudoku board. The second ar-
gument, List, is a list of the remaining unfilled squares in the board. The
third argument, Final, as before represents the contents of the final solution
246 Optimisation with active constraints

in a list form, and the fourth argument, Min, is the number of entries in this
final solution.
In contrast to the previous program the search for the minimum uses a
constraint, Min #>= Ct, to prune the search tree, which is needed to curtail
fruitless searches for solutions that have too many entries.
The call remove_random(TFree, [I,J]-Val, NFree) removes an(other)
entry from the board. The test if Val is a variable avoids explicitly selecting
entries whose values are already entailed by constraint propagation from the
previously chosen entries. In turn, the call ct(search(Board), N) is used
to complete the board after instantiating the next value. Since the con-
straints were posted at the start, this program uses the search/1 procedure
instead of the more costly sudoku/1.
Also ct/2 is used instead of one/1, because this enables the program to
distinguish the case where the current board has multiple solutions from the
case where there are none. (Recall that x is the non-logical variable used in
the procedure one/2 defined in Figure 8.16.) Finally, the statement
( N = 1 -> Cont = stop ;
N = multiple -> Cont = ok )
is used to avoid extending a partial assignment that is not a partial solution.
(Note that if N = 0 a failure arises.) If there are no solutions, the program
immediately backtracks and finds a different value for Val.
The program can be easily modified to one in which one first randomly
generates a solution to a Sudoku puzzle and then randomly copies its fields
into the Board array until a Sudoku puzzle is created.
We conclude by listing in Figure 12.8 the supporting predicates needed
for the two previous programs.

12.6 The bb min/3 built-in

In general, the computation of a solution to an optimisation problem consists
of:
the process of finding successively better solutions, before an optimum
solution is found,
the proof of optimality, obtained by searching for an even better solution,
and ultimately failing.
For problems such as the currency design problem from the last section,
where the optimum is almost exactly known in advance, the first process can
be avoided by constraining the cost function to be at or near the optimum.
 12.6 The bb min/3 built-in 247

remove_random(List, Val, Next) :-

length(List, Len),
Index:: 1..Len,
indomain(Index,random),
remove_nth(Index, List, Val, Next).

remove_nth(1, [H | T], H, T) :- !.
remove_nth(N, [H | T], Val, [H | Tail]) :-
N1 is N-1,
remove_nth(N1, T, Val, Tail).

board_to_list(Board, List) :-
( foreachelem(Val,Board,[I,J]),
foreach([I,J]-Val,List)
do
true
).

list_to_board(List, Board) :-
dim(Board,[9,9]),
( foreach([I,J]-Val,List),
param(Board)
do
subscript(Board,[I,J],Val)
).

Fig. 12.8 Supporting predicates for GEN SUDOKU

Benchmark optimisation problems are typically of this kind. For most opti-
misation problems, however, it is necessary to trawl through a sequence of
better and better solutions en route to the optimum. Here is a trace of a
scheduling program:
Cost: 863 in time: 0.110000000000582
... [20 intermediate solutions found in quick succession]
Cost: 784 in time: 2.04300000000148
... [10 more solutions - the time for each is growing]
Cost: 754 in time: 7.65100000000166
Cost: 752 in time: 8.75300000000061
248 Optimisation with active constraints

Cost: 750 in time: 21.8120000000017

Cost: 743 in time: 30.8549999999996
Cost: 742 in time: 32.3169999999991
... [5 solutions found quickly]
Cost: 723 in time: 34.2400000000016
Cost: 716 in time: 34.3600000000006
Cost: 710 in time: 34.5900000000001
Cost: 709 in time: 42.1610000000001
... [5 more solutions found quickly]
Cost: 685 in time: 44.5740000000005
Cost: 684 in time: 44.8050000000003
Cost: 683 in time: 51.7350000000006
Cost: 677 in time: 58.1140000000014
Found no solution with cost 349.0 .. 676.0
The uneven progress is typical: it takes some 13 seconds to improve the
cost from 752 to 750, but the improvement from 723 to 710 takes under
half a second. It took 7 seconds to improve the second best solution to an
optimal one and the time to prove optimality (not shown above) was also
around 7 seconds.
To better manage the computation process concerned with an optimisa-
tion problem ECLi PSe offers the bb min/3 built-in. The minimize/2 built-
in is a special case of bb min/3. However, bb min/3 has one extra argument
which allows the programmer to specify a number of options, some of which
we discussed in Subsection 6.6.3, and which can be used to reduce the com-
putation time needed for finding an optimal solution or a solution the cost
of which is close to the optimum.
We begin by exploring the following two options:
delta,
factor,
that are concerned with the improvement of the cost of each new solution.
Such options for bb min are specified using a special syntax which we explain
by returning to the trivial matter of finding the minimal element in the list
[5,6,5,3,4,2].
Let us specify that in the branch and bound search the cost of each new
solution has to be better by the value of 3 w.r.t. the previous one. This is
done by means of the following query:
[eclipse 10]: bb_min(member(X,[5,6,5,3,4,2]),
X,
 12.6 The bb min/3 built-in 249

bb_options{delta:3}).
Found a solution with cost 5
Found a solution with cost 2
Found no solution with cost -1.0Inf .. -1

X = 2
Yes (0.00s cpu)

Note that in the search the solution with cost 3 was skipped. Of course, the
branch and bound search with such an option can cease to be complete:

[eclipse 11]: bb_min(member(X,[5,6,5,3,4,2]),

X,
bb_options{delta:4}).
Found a solution with cost 5
Found no solution with cost -1.0Inf .. 1

X = 5
Yes (0.00s cpu)

Still we can conclude that the cost of the generated solution is at most delta
from the optimum.
A more important option is factor that governs how much better the
next solution should be than the last. The improvement factor is a number
between 0 and 1, which relates the improvement to the current cost upper
bound and lower bound.
Setting the factor to 1 puts the new upper bound at the last found
best cost this is the default used by the standard minimize/2 predicate.
Setting the factor to 0.01 sets the new upper bound almost to the cost lower
bound (factor 0.0 is not accepted by bb_min/3): typically it is easy to prove
there is no solution with a cost near this lower bound, and the optimisation
terminates quickly. More interesting is a setting around 0.9, which looks for
a 10% improvement at each step. In the case of our scheduling example this
factor gives the following trace:

Cost: 863 in time: 0.120000000002619

Cost: 811 in time: 0.200000000004366
Cost: 764 in time: 0.30000000000291
Cost: 716 in time: 0.430000000000291
Cost: 677 in time: 2.95400000000518
Found no solution with cost 349.0 .. 644.2
250 Optimisation with active constraints

This represents a tremendous saving in the time taken to improve, from the
same initial solution as before 863, to the optimal solution of 677. Notice
that the cost lower bound is 349. So after finding an initial solution of 863,
the new cost bound is set to 349 + 0.9 (863 349) = 811.6. The next
bounds are successively 764.8, 722.5, 679.3, 644.2, the last one being below
the optimum and therefore yielding no solution.
In this example the optimum solution is found, but in fact this search only
proves there is no solution better than 644: had there been further solutions
between 677 and 644 they would not have been found by this procedure.
What we get is a guarantee that the best solution 677 is within 10% or
more precisely within 0.1 (677 349) of the optimum.
The benefit is that not only is the time taken for the improvement process
much shorter (it is reduced from 58 seconds to around 3 seconds), but also
the time taken for the proof of optimality. It is reduced from 7 seconds to
around 0.2 seconds. The reason is that it is much easier to prove there is
no solution better than 644.2 than to prove there is no solution better than
677.
As another example let us modify the CURRENCY program from Figure 12.3
on page 238 so that we call for an improvement by a factor of 2/3. To this
end we replace the call

minimize((labeling(Values), labeling(Coins), check(Pockets)),

Min)

by

bb_min((labeling(Values), labeling(Coins), check(Pockets)),

Min,
bb_options{factor:0.666}).

An optimal solution is found then considerably faster:

[eclipse 12]: design_currency(Values, Coins).

Found a solution with cost 19
Found a solution with cost 12
Found a solution with cost 8
Found no solution with cost 1.0 .. 5.662

Values = [1, 2, 3, 4, 11, 33]

Coins = [1, 1, 0, 2, 2, 2]
Yes (29.26s cpu)
 12.6 The bb min/3 built-in 251

On the other hand, with the option factor:0.2 instead of factor:0.666

the above query exhibits the following behaviour:
[eclipse 13]: design_currency(Values, Coins).
Found a solution with cost 19
Found no solution with cost 1.0 .. 4.6

Values = [1, 2, 3, 4, 5, 6]
Coins = [1, 1, 0, 1, 0, 16]
Yes (8.78s cpu)
This confirms that, just as the delta option, the factor option can lead
to incompleteness.
Other important options for bb min/2 control the strategy used in the
branch and bound search and are in this sense more fundamental than delta
and factor. These strategy options are:
strategy:continue,
strategy:restart,
strategy:dichotomic.
The continue strategy is the default used by minimize/2: when a new
optimum is found search continues from the same point as before. The
restart strategy restarts the search whenever a new optimum is found.
The consequence can be that parts of the search tree that have already been
explored previously are explored again. However, because of the dynamic
variable choice heuristics, restarting can be more efficient than continuing,
if the tightened cost focusses the heuristic on the right variables.
As an example consider the following quotation from Hardy [1992]:
I had ridden in taxi cab number 1729 and remarked that the number seemed to
me rather a dull one, and that I hoped it was not an unfavorable omen. No, he
[Srinivasa Ramanujan] replied, it is a very interesting number; it is the smallest
number expressible as the sum of two cubes in two different ways.

To prove this property of 1729 it suffices to use the predicate hardy/2

defined in Figure 12.9 in which the restart strategy is employed.
The answer is produced instantaneously:
[eclipse 14]: hardy([X1,X2,Y1,Y2], Z).
Found a solution with cost 1729
Found no solution with cost 2.0 .. 1728.0

X1 = 1
252 Optimisation with active constraints

hardy([X1,X2,Y1,Y2], Z) :-
X1 #> 0, X2 #> 0,
Y1 #> 0, Y2 #> 0,
X1 #\= Y1, X1 #\= Y2,
X1^3 + X2^3 #= Z,
Y1^3 + Y2^3 #= Z,
bb_min(labeling([X1,X2,Y1,Y2]),
Z,
bb_options{strategy:restart}).

Fig. 12.9 The hardy/2 predicate

X2 = 12
Y1 = 9
Y2 = 10
Z = 1729

Yes (0.01s cpu)

In contrast, if we try to solve this problem using the continue strategy

instead, i.e., using the minimize/2 built-in, the optimality of the solution is
not computed in a meaningful amount of time.
Finally, the dichotomic strategy, like the factor option, seeks a faster
improvement of the cost but without sacrificing optimality. With this strat-
egy the search can seek improvements of the cost by, say, 10% or more, but
if no solution is found, then the failed upper bound is set as a new lower
bound, and the search resumes within the gap. On our scheduling example,
dichotomic search with factor 0.9 gives the following trace:

Cost: 863 in time: 0.120999999999185

Cost: 811 in time: 0.201000000000931
Cost: 764 in time: 0.290999999997439
Cost: 716 in time: 0.421000000002095
Cost: 677 in time: 2.92500000000291
Found no solution with cost 349.0 .. 644.2
Found no solution with cost 644.2 .. 673.72
Found no solution with cost 673.72 .. 676
 12.7 When the number of variables is unknown 253

Other options are available and so are other versions of bb min with more
arguments that give the programmer even more control over the computa-
tions that deal with the optimisation.

12.7 When the number of variables is unknown

It is useful to realise that some constrained optimisation problems with finite
domains cannot be solved by means of the branch and bound search. This
is the case for COPs with an unknown number of variables. Such COPs can
be easily generated in ECLi PSe . They still can be solved by means of the
customary backtracking combined with the constraint propagation, without
any use of the minimize/2 or bb min/3 built-ins. As an example consider
the following problem concerning natural numbers:

Problem Given m check whether it can be written as a sum of at least two

different cubes. If yes, produce the smallest solution in the number of cubes.

Even if m is fixed, say 100, it is not clear what the number of variables of
the corresponding CSP is. In other words, it not clear how to set up a CSP
for which we are to find an optimal solution. In this case the most natural
solution consists of systematically picking a candidate number n [2..m]
and of trying to find an appropriate combination of n different cubes. This
boils down to solving a CSP with n variables. If no such combination exists,
the candidate number is incremented and the procedure is repeated.
We can limit the variable domains by noting that

x31 + + x3n = M and 0 < x1 < x2 < < xn

implies

xn b 3 M c and n b 3 M c,

where the latter holds since n xn . To keep things simple in the pro-
gram presented in Figure 12.10 we actually use b M c (as theexpression
fix(round(sqrt(M)))) as the bound, since the computation of b 3 M c would
have to be programmed explicitly.
Then the following interaction with ECLi PSe shows that a solution exists
for N equal to 1000 000, in contrast to N equal to 1000:

[eclipse 15]: cubes(1000000, Qs).

Qs = [16, 68, 88]

254 Optimisation with active constraints

cubes(M, Qs) :-
K is fix(round(sqrt(M))),
N :: [2..K],
indomain(N),
length(Qs, N),
Qs :: [1..K],
increasing(Qs),
( foreach(Q,Qs),
foreach(Expr,Exprs)
do
Expr = Q*Q*Q
),
sum(Exprs) #= M,
labeling(Qs), !.
augmented by the procedure increasing/1 of Figure 12.3.

Fig. 12.10 The CUBES program

Yes (0.28s cpu)

[eclipse 16]: cubes(1000, Qs).

No (0.07s cpu)

12.8 Summary
In this chapter we explained how finite COPs can be solved using ECLi PSe .
To this end we studied the facilities provided by the branch and bound li-
brary, notably the minimize/2 built-in. It realises the basic form of the
branch and bound search. We illustrated the use of minimize/2 by con-
sidering the knapsack problem, the coins problem and the currency design
problem originally introduced in Section 6.5. Also we showed how Sudoku
puzzles can be generated in a simple way by combining constraint propaga-
tion with the branch and bound search.
Further, we introduced the bb min/3 built-in, of which minimize/2 is
a special case, that allows us to program various versions of the branch
and bound search by selecting appropriate options. We also clarified that
some finite domain COPs cannot be solved by means of the branch and
bound search when the number of variables can be unknown. In that case
 12.9 Exercises 255

the search for an optimal solution can be programmed using the customary
backtracking combined with the constraint propagation.

12.9 Exercises

Exercise 12.1 Solving an optimisation problem sometimes requires a bit

of experimentation. Consider the following program:
create_list_exprlist(N, List, ExprList) :-
length(List, N),
Max is 2*N,
List :: 0 .. Max,
alldifferent(List),
( for(C,1,N),
for(D,N,1,-1),
foreach(Var,List),
foreach(Expr,ExprList)
do
Expr #= C*D*Var
).
Find minimal solutions to the following two problems:
(i) prob1(List, Min) :-
create_list_exprlist(9, List, ExprList),
Min #= sum(ExprList).
(ii) prob2(List, Min) :-
create_list_exprlist(16, List, ExprList),
maxlist(ExprList,Min).
maxlist/2 is a built-in in the ic library for which maxlist(List,X) holds
if X the maximum value of the list List.
Hint. Each of these problems calls for different search/6 and bb min/3
strategies.

Exercise 12.2 Given an n n chess board, find the domination number,

which is the minimum number of queens needed to attack or occupy every
square. For the 8 8 chess board, the queens domination number is 5.

Exercise 12.3 Given an integer-sized square and a list of integer-sized

square tiles pack the square with the tiles so that the number of uncovered
unit squares is minimal.
 13

Constraints on reals

13.1 Introduction 256

13.2 Three classes of problems 257
13.3 Constraint propagation 261
13.4 Splitting one domain 263
13.5 Search 264
13.6 The built-in search predicate locate 265
13.7 Shaving 267
13.8 Optimisation on continuous variables 272
13.9 Summary 275
13.10 Exercises 275

13.1 Introduction

I N THE PREVIOUS three chapters we have assumed that the variables

have finite domains. Accordingly all the problems have a finite search
space which can be explored by instantiating the variables in all pos-
sible ways. In the next two chapters we will consider problems involving
variables with infinite domains. Because the problem search space is no
longer always finite, we may need to resort to different ways of solving these
problems. In this chapter we consider constraints on reals, sometimes called
constraints over continuous variables.
Constraints on reals are sometimes handled by artificially discretising the
domain, for example by only allowing 60 time points per hour, or distances
in multiples of a metre. The resulting problem has finite domain variables
and can be solved by labelling and propagation.
However, such a discretisation is artificial. More importantly, if an ac-
curate solution is needed the discretisation needs to be fine, and then the
variable domain sizes become huge, making the search space for labelling too
large. The most important drawback of discretisation is that it may elimi-

256
 13.2 Three classes of problems 257

nate solutions completely! For example, suppose we discretise a continuous

domain by admitting points which are exactly representable with N decimal
places. Then any problem whose solutions are not exactly representable in
this way becomes insoluble. A simple example would be to find a right angle
triangle with two equal sides, and return the length of the hypotenuse: this
has no solutions representable exactly to N decimal places.
Consequently, instead of the discretisation we rather modify the ingre-
dients of the approach used so far, i.e. top-down search and constraint
propagation.

13.2 Three classes of problems

In general, it is useful to distinguish three classes of problems involving
continuous variables:
those whose continuous variables are all dependent on other, finite domain,
variables. These can be solved by searching for the values of the finite
domain variables and then deriving the values of the continuous variables,
those which have a finite search space, but involve mathematical con-
straints which yield real number values for the variables,
those which have an infinite search space, and possibly infinitely many so-
lutions, which can only be represented intensionally by means of formulas,
for instance 0.0 X 10.0.
Let us illustrate these problems by means of examples.

13.2.1 A problem with dependent continuous variables

We wish to build a cylindrical compost heap using a length of chicken wire.
The volume of compost should be at least 2 cubic metres. The wire comes
in lengths of 2, 3, 4 and 5 metres, and widths of 50, 100 and 200 centimetres.
Which length and width of wire do we need?
The variables in this problem are the length of wire, L, the width W and
the volume surrounded by the wire, V , see Figure 13.1.
The constraints are L {2, 3, 4, 5}, W {50, 100, 200}, V 2 and V =
(W/100) (L2 /(4 )).
This problem can be solved in the usual way using search and propagation,
as follows:
:- lib(ic).
:- lib(branch_and_bound).
258 Constraints on reals

Fig. 13.1 A cylinder

compost_1(W, L, V) :-
W :: [50, 100, 200],
L :: 2..5,
V $>= 2.0,
V $= (W/100)*(L^2/(4*pi)),
minimize(labeling([W,L]),V).

Notice that in ECLi PSe pi/0 is a built-in constant which returns the value
of .
We get then:

[eclipse 1]: compost_1(W, L, V).

Found a solution with cost
2.5464790894703251__2.546479089470326
Found no solution with cost
2.5464790894703251 .. 1.546479089470326

W = 200
L = 4
V = 2.5464790894703251__2.546479089470326

There are 6 delayed goals.

Yes (0.00s cpu)

So the generated solution reports a number of delayed goals. The reasons

for this will be explained in Section 13.5. Also, the value of V is returned in
the form of an interval.
 13.2 Three classes of problems 259

13.2.2 A problem with a finite search space involving continuous

variables
To illustrate this class of problems we use a simple geometrical problem.
It involves two circles of radius 2, the first one centered on (0,0) and the
second one centered on (1,1). We seek a pont that lies on both circles, and
additionally lies to the right of the line X = Y :
circles(X, Y) :-
4 $= X^2 + Y^2,
4 $= (X - 1)^2 + (Y - 1)^2,
X $>= Y.
The conjunction of these three constraints has exactly one solution, see
Figure 13.2.

(1,1)
(0,0)
(X,Y)

Fig. 13.2 Two intersecting circles

To solve this problem we must specify a form of search that focusses on

the finite set of (two) potential solutions where the circles intersect. One
form of search which achieves this is described in Section 13.4 below.

13.2.3 A problem with infinitely many solutions

The most general class of problems with continuous variables are problems
that have an infinite space of potential or actual solutions.
By changing an input parameter, or a finite domain variable, into a con-
tinuous variable, many problems can be made infinite in this sense. For
example, suppose in the compost heap problem the length and width of
chicken wire are both continuous variables. We then seek pairs of values
that yield a volume of 2:
260 Constraints on reals

compost_2(W, L) :-
W :: 50.0..200.0,
L :: 2.0..5.0,
V $= 2.0,
V $= (W/100)*(L^2/(4*pi)).

In this case ECLi PSe returns an interval of possible values for W and L:

[eclipse 2]: compost_2(W, L).

W = W{100.53096491487334 .. 200.0}
L = L{3.5449077018110313 .. 5.0}

There are 11 delayed goals.

Yes (0.00s cpu)

So, as in the discrete case, ECLi PSe returns a number of delayed goals.
This answer means that there is no width below 100.53096491487334 and no
length below 3.5449077018110313 from which we can build a 2 cubic metre
compost heap. However, it is not possible to list all the pairs of values for
W and L within the above intervals that yield a volume of 2 cubic metres.
Instead, the delayed goals record additional constraints on the variables W
and L that must be satisfied in any solution.
In this example every value for W in the final interval has a compatible
value for L which together make up a solution to the problem. In general, it
is not the case that all values in the final intervals form part of a solution.
A trivial example is:

[eclipse 3]: [X, Y] :: -10.0 .. 10.0, X $= 1/Y.

X = X{-10.0 .. 10.0}
Y = Y{-10.0 .. 10.0}

There are 3 delayed goals.

Yes (0.00s cpu)

Clearly the delayed goals on X and Y are unsatisfiable for any value of Y in
the interval -0.1 .. 0.1. We shall see in Section 13.4, to elicit compatible
solutions for X and Y, this example requires some additional search and
constraint propagation.
 13.3 Constraint propagation 261

13.3 Constraint propagation

In principle to solve the constraints on continuous variables we can use
the same constraint propagation techniques as the ones used before for fi-
nite domain variables. However, because the variable domains are infinite,
additional complications can arise. First of all constraint propagation for
the constraints on reals can be computationally costly because a single con-
straint propagation step may yield a very small reduction in the domain size
of the variables. We can illustrate this very simply using an inconsistent pair
of constraints. Consider the following definition:

incons(X, Y, Diff) :-
[X,Y] :: 0.0..10.0,
X $>= Y + Diff,
Y $>= X + Diff.

If Diff > 0, then the query incons(X, Y, Diff) should fail even without
any information about the domains of X and Y since X $>= Y + Diff and Y
$>=X + Diff implies X $>= X + 2*Diff. However, in the ic library each
constraint is considered separately and the inconsistency is achieved only by
the repeated shaving of the domain bounds of X and Y by the amount Diff.
We can try smaller and smaller values of Diff and see how the time needed
to detect the inconsistency grows. (The term 1e-n represents in ECLi PSe
10n .)

[eclipse 4]: incons(X, Y, 1e-4).

No (0.09s cpu)

[eclipse 5]: incons(X, Y, 1e-5).

No (0.82s cpu)

[eclipse 6]: incons(X, Y, 1e-6).

No (8.40s cpu)

[eclipse 7]: incons(X, Y, 1e-7).

No (159.81s cpu)

[eclipse 8]: incons(X, Y, 1e-8).

X = X{0.0 .. 10.0}
Y = Y{0.0 .. 10.0}
262 Constraints on reals

Delayed goals:
ic : (Y{0.0 .. 10.0} - X{0.0 .. 10.0} =< -1e-8)
ic : (-(Y{0.0 .. 10.0}) + X{0.0 .. 10.0} =< -1e-8)
Yes (0.00s cpu)

The increasing amount of time required to recognise the inconsistency for

each query is due to the increasing number of constraint propagation steps
required. The last query, shockingly, does not fail because instead of eval-
uating the constraints ECLi PSe delays them. This is, arguably, the most
convenient behaviour because it would have taken an excessive amount of
time to detect the inconsistency by means of the built-in constraint propa-
gation.
The mechanism governing this process is a setting termed the propaga-
tion threshold which specifies a maximum amount by which the bounds
of a variable may change without ECLi PSe actually updating the bounds
and waking any delayed goals. The default propagation threshold is 1e-8.
The last query does not perform any constraint propagation because the
change at the first step is too small and consequently no bound updates are
performed.
For some applications the default threshold of 1e-8 is already too small
(or too costly), and so the programmer has a facility to change this setting
using the set threshold/1 built-in, for example:

[eclipse 9]: set_threshold(1e-3), incons(X, Y, 1e-4).

X = X{0.0 .. 10.0}
Y = Y{0.0 .. 10.0}

Delayed goals:
ic : (Y{0.0 .. 10.0} - X{0.0 .. 10.0} =< -0.0001)
ic : (-(Y{0.0 .. 10.0}) + X{0.0 .. 10.0} =< -0.0001)
Yes (0.00s cpu)

Controlling constraint propagation using the threshold is a question of

experimentation, to be used in case the default threshold yields poor per-
formance. One important point, however, is that changing the threshold
cannot generate wrong answers. Any goals that are delayed are reported in
the solution, and should remind the programmer that the solution is only
true subject to these delayed constraints being satisfied.
 13.4 Splitting one domain 263

13.4 Splitting one domain

Constraint propagation alone is not enough to solve constraint satisfaction
problems on reals. Take for example the predicate circles/2 discussed in
Section 13.2. Constraint propagation produces an interval for each variable
which includes both solutions and all the points in between:

[eclipse 10]: circles(X, Y).

X = X{-1.0000000000000004 .. 2.0000000000000004}
Y = Y{-1.0000000000000004 .. 2.0000000000000004}

There are 13 delayed goals.

Yes (0.00s cpu)

To isolate the unique solution properly it is necessary to split the domain

of X into two halves and try each domain separately:

[eclipse 11]: circles(X, Y), (X $>= 1.5 ; X $=< 1.5).

X = X{1.8228756488546369 .. 1.8228756603552694}
Y = Y{-0.82287567032498 .. -0.82287564484820042}

There are 12 delayed goals.

Yes (0.00s cpu, solution 1, maybe more) ? ;

No (0.00s cpu)

So the first subdomain, with X >= 1.5, tightens the lower bound on X and
enables the constraint propagation to reduce the domain of X to the interval

1.8228756488546369 .. 1.8228756603552694

and the domain of Y to the interval

-0.82287567032498 .. -0.82287564484820042.

The second subdomain, with X =< 1.5, tightens the upper bound on X
and enables the constraint propagation to detect that there is no solution in
this domain for X. Supplementing this computation by a reasoning showing
that a unique solution exists we can conclude that this solution lies in the
rectangle determined by the above two intervals. So even though we did not
compute the precise solution we could determine its location with a very
high degree of precision.
264 Constraints on reals

Domain splitting can also be effective for problems with infinitely many
solutions, as illustrated by the example X = 1/Y introduced in Section 13.2.
In this case we need to split the domain into three subdomains, X ,
 X  and  X for some small . With this splitting we can achieve
a precise cover of the solution set:1

[eclipse 12]: [X, Y] :: -10.0..10.0, X $= 1/Y,

(X :: -inf .. -0.001
;
X :: -0.001 .. 0.001
;
X :: 0.001 .. inf
).

X = X{-10.0 .. -0.099999999999999992}
Y = Y{-10.0 .. -0.099999999999999992}

There are 4 delayed goals.

Yes (0.00s cpu, solution 1, maybe more) ? ;

X = X{0.099999999999999992 .. 10.0}
Y = Y{0.099999999999999992 .. 10.0}

There are 4 delayed goals.

Yes (0.00s cpu, solution 2)

13.5 Search
For continuous variables a general search method is to associate with each
variable a domain in the form of an interval of possible values, and at each
search node to narrow one interval. Complete search can be maintained by
exploring a different subinterval at each subbranch, but ensuring that the
union of the subintervals covers the input interval at the node.
Because the domains are continuous, this search tree is infinitely deep:
you can go on dividing an interval forever without reducing it to a point.
Consequently a precision has to be specified which is the maximum allowed
width of any interval on completion of the search. During the search, when
the interval associated with a variable is smaller than this precision, this
1 inf is the built-in representation of the infinity.
 13.6 The built-in search predicate locate 265

variable is never selected again. Search stops when all the variables have
intervals smaller than the precision.
If the search fails after exploring all the alternative subintervals at each
node, then (assuming the subintervals cover the input interval at the node)
this represents a proof that the problem has no solution. In other words,
the search failure for continuous variables is sound.
On the other hand if the search stops without failure, this is no guaran-
tee that there really is a solution within the final intervals of the variables
(unless they are all reduced by constraint propagation to a single point). A
similar phenomenon has already been illustrated in Section 13.3 using the
incons/3 predicate, where constraint propagation stops even though there
is no solution satisfying the constraints.
Accordingly each answer to a problem on reals is in fact only a condi-
tional solution. It has two components:

A real interval for each variable. Each interval is smaller than the given
precision.
A set of constraints in the form of delayed goals. These constraints are
neither solved nor unsatisfiable when considered on the final real intervals.

An example is the answer to the query No. 9 in the previous section.

13.6 The built-in search predicate locate

Naturally ECLi PSe provides a built-in predicate locate/2 that makes it
easy to search over continuous variables. This built-in recursively splits
the domains of specified variables until their sizes fall within the specified
precision.
The simplest version of locate/2 has two arguments, specifying a list
of variables to search on, and a final precision. Let us illustrate its use by
considering the problem of finding zeroes of the polynomial 2x5 5x4 +5,
depicted in Figure 13.3.
The difficulty lies in the fact that in general, by the celebrated Galois
theorem, the zeroes of polynomials of degree higher than four cannot be
expressed as radicals, that is by means of the four arithmetic operations and
the root extraction. In fact, 2x5 5x4 +5 is one of such polynomials, found
in the nineteenth century by the Norwegian mathematician Niels Henrik
Abel. To cope with this problem we shall ask for the (conditional) solutions
in the form of intervals. To ensure an acceptable computation time we search
for solutions within the interval -1000000.0 .. 1000000.0. In ECLi PSe
266 Constraints on reals

10

-5

-10

-10 -5 0 5 10

Fig. 13.3 The diagram of the polynomial 2 x5 5 x4 + 5

we can represent this interval using the syntax 1e-6 .. 1e6, and so the
problem is set up as follows:

abel(X) :-
X :: -1e6 .. 1e6,
2*X^5 - 5*X^4 + 5 $= 0.

Firstly we set the search precision to 1.0. This is sufficient to distinguish

one of the solutions from the others:

[eclipse 13]: abel(X), locate([X], 1.0).

X = X{-0.92435801492603964 .. -0.92435801492602809}

There are 5 delayed goals.

Yes (0.00s cpu, solution 1, maybe more) ;

X = X{1.1711620684831605 .. 2.4280727927074546}

There are 5 delayed goals.

Yes (0.00s cpu, solution 2)

The delayed goals specify the additional conditions on X under which the
 13.7 Shaving 267

query is satisfied. To separate out all three solutions we use a smaller pre-
cision, such as 0.001:
[eclipse 14]: abel(X), locate([X], 0.001).

X = X{-0.92435801492603964 .. -0.92435801492602809}

There are 5 delayed goals.

Yes (0.00s cpu, solution 1, maybe more) ;

X = X{1.1711620684831605 .. 1.1711620684832094}
There are 5 delayed goals.
Yes (0.00s cpu, solution 2, maybe more) ;

X = X{2.4280727927071966 .. 2.4280727927074546}
There are 5 delayed goals.
Yes (0.01s cpu, solution 3)

13.7 Shaving
While constraint propagation and domain splitting work fine on many prob-
lems, there are problems where even for very narrow intervals constraint
propagation is unable to recognise infeasibility. The consequence is that too
many alternative conditional solutions are returned.
The following predicate encodes the so-called Reimers system:
reimer(X, Y, Z) :-
[X,Y,Z] :: -1.0 .. 1.0,
X^2 - Y^2 + Z^2 $= 0.5,
X^3 - Y^3 + Z^3 $= 0.5,
X^4 - Y^4 + Z^4 $= 0.5.
We will search for the solutions using the ECLi PSe query
reimer(X, Y, Z), locate([X,Y,Z], Prec).
for different levels of precision Prec. The number of conditional solutions
returned for different levels of precision i.e. different final interval widths
are presented in Table 13.1.
For the wide intervals it is clear that several solutions are being captured
within a single interval, and decreasing the interval width increases the num-
ber of distinct conditional solutions. At a width of around 1e-10 the intervals
are narrow enough to detect that certain conditional solutions are infeasible,
268 Constraints on reals

search precision solution count CPU seconds

1.0 6 0.09
0.01 8 2.08
1e-5 13 5.41
1e-8 11 7.23
1e-10 10 7.80
1e-12 11 8.25
1e-14 24 8.61

Table 13.1 Reimers problem without shaving

so the number of distinct conditional solutions decreases. However, as the

intervals become narrower many narrow intervals are returned in the vicin-
ity of each actual solution. These figures would lead one to believe that the
problem has 10 distinct real solutions, but this is surprisingly not the case.
To clarify these matters we return to the predicate squash/3 which we
encountered previously in Subsection 10.3.2. This predicate implements a
special form of constraint propagation, called shaving , where a variable is
constrained to take a value or to lie within a narrow interval near its
bound (upper or lower), and the results of this constraint are propagated
to the other variables to determine whether the problem is still feasible. If
not, the domain of the original variable is reduced to exclude the infeasible
region and constraint propagation continues.
For constraints on continuous variables shaving is a very important tech-
nique. Recall that the arguments of the squash/3 predicate are:

a list of variables to which this constraint propagation technique is applied,

an interval width the size of the region within which the variable is
constrained to lie during propagation,
a parameter which can take two values, lin or log. This parameter guides
the way domains are split. If it is set to lin, the split is linear (i.e. the
arithmetic mean of the bounds is used). If it is set to log, the split is
logarithmic (i.e., the geometric mean of the bounds is used).2

To demonstrate its effect we will tackle Reimers problem again, but this
time applying shaving to the intervals returned by the search routine. We
will again try different precisions for locate/2 and for squash/3 we will use
2 Recall that the arithmetic mean of a and b is (a + b)/2 while the geometric mean of a and b is

a b.
 13.7 Shaving 269

an interval width of 1e-15. For the third parameter we will always choose
the value lin. The query is thus:

reimer(X, Y, Z), locate([X,Y,Z], Prec),

squash([X,Y,Z], 1e-15, lin).

search precision solution count CPU seconds

1.0 4 6.34
0.01 4 7.56
1e-5 4 8.58
1e-10 4 8.66
1e-14 16 8.44

Table 13.2 Reimers problem with shaving

Clearly there are at most four different solutions. The shaving procedure
removes infeasible intervals. Note, however, that very narrow intervals very
close to a real solution still cannot be removed by squash/3. In particular,
if we set the precision to 1e-14, then we get 16 conditional solutions.
Shaving can be used as a (polynomial) alternative to the (exponential)
interval splitting method of narrowing intervals. Because it does not in-
troduce extra choice points it applies well to problems of the second class
introduced in Section 13.2, that is problems which have a finite search space
but involve mathematical constraints.
As another example of the use of shaving we shall consider a program that
has been often used to illustrate solving constraints on continuous variables.
The program specifies a financial application, concerning how to pay off a
mortgage, and is given in Figure 13.4.
It expresses the relationship between the four variables:
Loan (the loan),
Payment (the fixed monthly amount paid off),
Interest (fixed, but compounded, monthly interest rate),
Time (the time in months taken to pay off a loan).
Three of the variables are continuous, Loan, Payment and Interest, and
the other, Time, is integer valued.
This program can be used in many ways, for example to compute the
duration of a given loan. To answer this question we need to ask for the
270 Constraints on reals

mortgage(Loan, _Payment, _Interest, 0) :-

Loan $= 0.

mortgage(Loan, Payment, Interest, Time) :-

Loan $> 0,
Time $>= 1,
NewLoan $= Loan*(1+Interest) - Payment,
NewTime $= Time - 1,
mortgage(NewLoan, Payment, Interest, NewTime).

Fig. 13.4 The MORTGAGE program

amount of time needed to pay off at least the borrowed amount, as in the
following query:

[eclipse 15]: minimize((X $>= 50000, X$=< 50700,

mortgage(X, 700, 0.01, T)),
T).
Found a solution with cost 126
Found no solution with cost -1.0Inf .. 125

X = 50019.564804291353__50019.56480429232
T = 126
Yes (0.06s cpu)

Note that the query mortgage(50000, 700, 0.01, T) simply fails, since
the payments of 700 do not add up to 50000 precisely.
It is more difficult to compute the required regular payment from a fixed
loan, interest rate and payoff time. At least we know the payment is greater
than 0 and less than the total loan, so we can provide an initial interval for
it:

[eclipse 16]: Payment :: 0.0 .. 200000.0,

mortgage(200000, Payment, 0.01, 360).

Payment = Payment{0.0 .. 200000.0}

There are 718 delayed goals.

Yes (0.02s cpu, solution 1, maybe more) ? ;
 13.7 Shaving 271

No (0.02s cpu)

This is completely uninformative. Similarly, we may want to compute the

actual interest rate corresponding to a given loan, regular payment and time.
Also here we can provide an initial interval for the interest rate, between 0.0
and 1.0:

[eclipse 17]: Interest :: 0.0 .. 1.0,

mortgage(60000, 1500, Interest, 60).

Interest = Interest{0.0 .. 0.5}

There are 257 delayed goals.

Yes (0.01s cpu, solution 1, maybe more) ? ;

No (0.01s cpu)

Again the answer is completely uninformative. However, using shaving

we can extract solutions to both queries with a high degree of precision:

[eclipse 18]: Payment :: 0.0 .. 200000.0,

mortgage(200000, Payment, 0.01, 360),
squash([Payment], 1e-5, log).

Payment = Payment{2057.1348846043343 .. 2057.2332029240038}

There are 360 delayed goals.

Yes (3.72s cpu, solution 1, maybe more) ? ;

No (3.72s cpu)

[eclipse 19]: Interest :: 0.0 .. 1.0,

mortgage(60000, 1500, Interest, 60),
squash([Interest], 1e-5, log).

Interest = Interest{0.0143890380859375 .. 0.01439666748046875}

There are 237 delayed goals.

Yes (14.48s cpu, solution 1, maybe more) ? ;

No (14.48s cpu)
272 Constraints on reals

13.8 Optimisation on continuous variables

The branch and bound procedure embodied in the minimize/2 and bb min/3
built-ins introduced in Chapter 12 relies on the fact that the search proce-
dure instantiates the variable constrained to the cost function. (Otherwise
a run-time error results.) In the case of constraints on continuous variables
this cannot be guaranteed. Indeed, in general the search stops when all
the variables have intervals smaller than the chosen precision. As a result
these two built-ins cannot be used directly for optimisation on continuous
variables. In this section we will outline one way of adapting them to cope
with a continuous optimisation function.
In the case of constraints on continuous variables the backtracking pro-
cess systematically returns conditional solutions in the form of small boxes
(of size smaller than the precision). When searching for an optimum of a
continuous cost function, with each box an interval for the cost function is
returned. To ensure that the the variable constrained to the cost function is
instantiated, we set it each time to the lower bound of this returned interval.
After this modification we can reuse the previous routine.
What can we conclude about the results produced by a so-modified branch
and bound procedure? If the sets of constraints associated with each re-
turned box are all satisfiable, we know that upon termination of this mod-
ified branch and bound the minimum lies in the interval of the size of the
set precision, with the lower bound Min.
However, if some (possibly all) of these sets of constraints are unsatisfiable,
we can only use the final value Min of the cost function as a conditional lower
bound: we do not know whether an actual solution exists, but if it does then
the value of the cost function is at least Min.
To clarify this approach suppose, for example, we wish to minimize some
expression involving the variables X and Y under the following constraints,
expressed in ECLi PSe :
cons(X, Y) :-
[X,Y] :: -100.0 .. 100.0,
X*2 + 2*Y^2 $< 10,
X + 3*Y $=< 5,
2*X - Y $=< 10.
Suppose, furthermore, we search for a solution by using locate/2 to recur-
sively split the intervals associated with the variables X and Y, using
locate([X,Y], 0.01).
During branch and bound search, whenever a conditional solution is re-
 13.8 Optimisation on continuous variables 273

turned, it will associate an interval with each of X and Y. If we introduce a

new variable OptExpr constrained to be equal to the expression being op-
timised, then constraint propagation will produce an interval for OptExpr.
Since we are minimising, the lower bound of this interval is the best possible
solution found so far if it is feasible! We use this lower bound to constrain
the value of the optimised expression from now on. If no better solution
is found, then the conditional solution returned is the last returned best
conditional solution.
The appropriate optimisation predicate opt 1 can be encoded using the
minimize/2 built-in introduced in Section 12.1:
opt_1(Query, Expr, Min) :-
OptExpr $= eval(Expr),
minimize(( Query, get_min(OptExpr,Min) ), Min).
Recall from Section 9.6 that the eval/1 built-in indicates that its argu-
ment is a variable that will be bound to an expression at the run-time. We
can now solve our constrained optimisation problem. To be specific suppose
the optimisation function is 6-X:
[eclipse 20]: cons(X, Y),
opt_1( locate([X,Y], 0.01), (6-X), Min ).

Found a solution with cost 5.9977375094284113

Found a solution with cost 4.9949650990605292
Found a solution with cost 3.9640446836405561
Found a solution with cost 2.9593568281590485
Found a solution with cost 1.9412953187340669
Found no solution with cost -1.0Inf .. 0.94129531873406691

X = X{3.9941057914745426 .. 4.0587046812659331}
Y = Y{-0.0012435228262296953 .. 0.0044940668232825009}
Min = 1.9412953187340669

There are 4 delayed goals.

Yes (6.08s cpu)
This output shows that optimality (even in the conditional sense) has not
exactly been proven: there is no solution with a value less than
0.941 295 318 734 066 91, but there is a gap between this and the best con-
ditional solution found, which is 1.941 295 318 734 0669. The reason is that
the branch and bound default is tailored for integer problems: it seeks a
274 Constraints on reals

solution which is a whole unit (i.e., 1.0) better than the previous one. So we
can only conclude that no minimum below 0.941 295 318 734 066 91 exists.
For continuous problems it therefore makes sense to override the default
and specify a minimum improvement sought at each step. The smaller
the improvement the nearer the procedure can guarantee to come to the
actual optimum (if it exists). However, smaller improvements can result in
increased solving times.
We conclude this chapter by encoding a modified optimisation procedure
that gives the programmer control over the required improvement at each
step. The appropriate optimisation predicate uses the bb min/3 built-in:
opt_2(Query, Expr, Improvement, Min) :-
OptExpr $= eval(Expr),
bb_min(( Query, get_min(OptExpr,Min) ),
Min,
bb_options{delta:Improvement}).
By choosing an improvement of first 0.5 and then 0.1 we can illustrate the
advantage of smaller improvements:3
[eclipse 21]: cons(X, Y),
opt_2(locate([X, Y], 0.01), (6-X), 0.5, Min).

Found a solution with cost 5.9977375094284113

... 9 intermediate solutions

Found no solution with cost -1.0Inf .. 0.7349623277551629

X = X{4.6891966915344225 .. 4.7650376722448371}
Y = Y{-0.0012435228262296953 .. 0.0044940668232825009}
Min = 1.2349623277551629

There are 4 delayed goals.

Yes (6.16s cpu)

[eclipse 22]: cons(X, Y),

opt_2(locate([X, Y], 0.01), (6-X), 0.1, Min).

3 In this problem both queries happen to take a similar time: this is because the branch and
bound is not causing any pruning of the search space. Sometimes a branch and bound search
with more intermediate solutions takes an exponentially longer time.
 13.9 Summary 275

Found a solution with cost 5.9977375094284113

... 42 intermediate solutions

Found no solution with cost -1.0Inf .. 0.97958067273231519

X = X{4.8421052712298724 .. 4.9204193272676848}
Y = Y{-0.0012435228262296953 .. 0.0044940668232825009}
Min = 1.0795806727323152

There are 4 delayed goals.

Yes (6.02s cpu)
Because the computed minima are only conditional (subject to the delayed
constraints) we can only conclude that no minimum below
0.979 580 672 732 315 19 exists.

13.9 Summary
In this chapter we discussed the problem of solving constraints on reals. To
this end we considered in turn:
three natural classes of problems involving continuous variables,
constraint propagation on continuous variables,
search,
optimisation over continuous optimisation functions.
We also indicated the importance of a specialised constraint propagation
technique, called shaving.
In the process we introduced the following built-ins:
the constant pi, which represents in ECLi PSe ,
the number and exponent syntax for real numbers, e.g. 1e-8,
set threshold/1, used to set the propagation threshold,
locate/2, to perform search over continuous variables,
squash/3, to perform shaving.

13.10 Exercises

Exercise 13.1 When optimising using the predicate opt 2/3 defined in
Section 13.8, it is necessary to compromise. A tighter precision means more
276 Constraints on reals

optimisation steps, but a looser precision means there may be no feasible

solution in the conditional solution returned at each branch and bound step.
Extend the definition of opt 2(Query, Expr, Imp) to include a call to
squash(Vars, 1e-8, log) each time a new conditional solution is returned
by branch and bound. Apply it to the list of the variables appearing in
Query.

Exercise 13.2 Apply the same modification to minimise the expression

X 2 + (1/Y )2 under the constraints X, Y 10.0 . . . 10.0, X Y = 1.

Exercise 13.3 A piano with length L and width W has to be moved around
a right-angle corner in a corridor, see Figure 13.5. Coming into the corner
the corridor has width Corr1 and going out it has a possibly different width
Corr2. Given L = 2, W = 1 and Corr1 = 1.5, what is the minimum value
for Corr2?
Corr1

L Corr2

Fig. 13.5 A corridor

To make this a little easier here are the underlying mathematical con-
straints.
The envelope of the piano as it moves around the corner can be
expressed in terms of a variable T . The envelope describes two coor-
dinates X and Y as follows:
X = W T 3 + L (1 T 2 )1/2 ,
Y = L T + W (1 T 2 )3/2 .
At the critical point X = Corr1 and Y = Corr2.
T can take values in the interval MinT .. MaxT, where
 13.10 Exercises 277
2 4L2 )1/2 1/2
 
MinT = 1/2 (9W 6W ,
 2 4L2 )1/2
 1/2
MaxT = 1/2 + (9W 6W .
The problem is then to write a predicate which finds the minimum possible
value for Corr2, taking parameters W, L and Corr1.
 14

Linear constraints over continuous and

integer variables

14.1 Introduction 278

14.2 The eplex library 279
14.3 Solving satisfiability problems using eplex 284
14.4 Repeated solver waking 285
14.5 The transportation problem 289
14.6 The linear facility location problem 294
14.7 The non-linear facility location problem 296
14.8 Summary 302
14.9 Exercises 303

14.1 Introduction

W E ALREADY DISCUSSED linear constraints on reals in Section

6.3 and used them in Subsection 6.5.4 to formalise the facility
location problem. These constraints form an important class be-
cause of their wide range of applicability. As a result over many years in
the field of operation research several efficient methods have been developed
for handling them and ways of modelling problems using only constraints
of this form have been devised. In contrast, the problem of solving linear
constraints over integer variables is in general NP-complete and as a result
can be used to express any constraint satisfaction problem.
There exist both scalable polynomial algorithms for solving linear con-
straints on reals and, in principle exponential, Simplex-based algorithms
which are both scalable and can be embedded in branch and bound algo-
rithms for handling linear constraints over integer variables, see, e.g., Pa-
padimitriou and Steiglitz [1982].
Linear constraints together with integrality constraints1 are at the core of
1 An integrality constraint on a variable enforces the variable to take an integer value.

278
 14.2 The eplex library 279

Operational Research. There are several software packages available which

perform constraint solving and search on problems expressed in terms of
linear and integrality constraints. ECLi PSe provides an interface to the
Xpress-MP package from Dash Optimisation, which was used to design and
run the examples in this book. It also provides an interface to the CPLEX
package from ILOG.
These interfaces are available through the ECLi PSe library eplex that
consists of the eplex constraint mapping module augmented with a number
of auxiliary built-ins that facilitate interaction with the external solver. In
this chapter we discuss the eplex library and illustrate its use to solve
problems that deal both with linear and non-linear constraints on integers
and reals.

14.2 The eplex library

14.2.1 Motivation
To properly motivate the eplex library note first the following limitation of
the ic library. Consider the predicate

incons(W, X, Y, Z) :-
W+X+Y+Z $>= 10,
W+X+Y $= 5,
Z $=< 4.

Clearly the query incons(W, X, Y, Z) should fail. However, when we in-

voke it using the ic library we get the following outcome:

[eclipse 1]: [W,X,Y,Z]:: 0..10, incons(W, X, Y, Z).

W = W{0 .. 5}
X = X{0 .. 5}
Y = Y{0 .. 5}
Z = Z{0 .. 4}

Delayed goals:
ic : (-(Z{0 .. 4}) - Y{0 .. 5} - X{0 .. 5} - W{0 .. 5}
=< -10)
ic : (Y{0 .. 5} + X{0 .. 5} + W{0 .. 5} =:= 5)
Yes (0.00s cpu)
280 Linear constraints over continuous and integer variables

So the inconsistency cannot be detected in ic without search. This is in

contrast to the eplex solver. To execute the query incons(W, X, Y, Z)
using eplex the appropriate ECLi PSe code is :
[eclipse 2]: eplex_solver_setup(min(0)),
incons(W, X, Y, Z),
eplex_solve(_).

No (0.00s cpu)
So the eplex solver correctly detects the inconsistency even without any
initial domains for the variables.
A key difference is that, in contrast to ic, the eplex solver only handles
constraints after it has been explicitly initialised by the program. (The
rationale for this facility is that the same program can initialise more than
one linear solver instance through the eplex library, and send different
constraints to different instances.) The first line initialises the eplex solver
by stating whether one searches for a minimum of the cost function or a
maximum.
In general, the argument of the eplex solver setup/1 built-in is either
min(Cost) or max(Cost), where the Cost variable is constrained to the
expression defining the cost function. In our example we are not interested
in the finding an optimal solution, so just use 0 as the cost function. The
last line launches the eplex solver using the eplex solve/1 built-in. In
general its argument returns the optimal value of the expression defining
the cost function and specified at setup time. Here we are not interested in
the optimal value, so use an anonymous variable as the argument.

14.2.2 Accessing solver information

The eplex library admits variable declarations and a limited form of linear
constraints, using the same syntax as the ic library. The following classes
of constraints are not supported:

the integer versions of linear constraints, such as X #=< Y,

strict inequalities and disequality, so X $< Y, X $\= Y, X $> Y,
Boolean constraints,
reified constraints.
Let us give now an example of a successful evaluation using eplex:
[eclipse 3]: eplex_solver_setup(min(0)),
 14.2 The eplex library 281

W+X+Y+Z $>= 10,

W+X+Y $= 5,
Z $=< 6,
eplex_solve(_).

W = W{-1e+20 .. 1e+20 @ 5.0}

X = X{-1e+20 .. 1e+20 @ 0.0}
Y = Y{-1e+20 .. 1e+20 @ 0.0}
Z = Z{-1e+20 .. 6.0 @ 5.0}
Yes (0.00s cpu)
The eplex solver does not tighten the domains of the variables, except for
Z whose upper bound is explicitly set by the constraint Z $=< 6. However, it
does compute a solution which satisfies all the constraints, and this solution
is shown in the output, after the @ sign: W = 5.0, X = 0.0, Y = 0.0,
Z = 5.0. In fact this is just one solution and there are many others, for
example W = 0.0, X = 5.0, Y = 0.0, Z = 5.0.
In other words, the solution computed by eplex satisfies the constraints,
but is not entailed by them. Many problems tackled by eplex have infinitely
many solutions and it is impossible to generate all of them. Consequently
eplex does not provide any means of computing alternative solutions, for
example using backtracking. On the other hand, eplex does provide a
mechanism for extracting the current solution, using the eplex var get/3
built in:
[eclipse 4]: eplex_solver_setup(min(0)),
W+X+Y+Z $>= 10,
W+X+Y $= 5,
Z $=< 6,
eplex_solve(_),
eplex_var_get(W, typed_solution, Val).

...
Val = 5.0
Yes (0.00s cpu)
Alternatively, it is possible to extract information about the complete
solver state using the eplex get/2 built-in:2
[eclipse 5]: eplex_solver_setup(min(0)),
2 In this output, for brevity and clarity, we have set the variable printing options not to display
any of the variable attributes.
282 Linear constraints over continuous and integer variables

W+X+Y+Z $>= 10,

W+X+Y $= 5,
Z $=< 6,
eplex_solve(_),
eplex_get(constraints, Cons),
eplex_get(vars, Vars),
eplex_get(typed_solution, Vals).

...
Cons = [Y + X + W =:= 5.0, Z + Y + X + W >= 10.0]
Vars = (W, X, Y, Z)
Vals = (5.0, 0.0, 0.0, 5.0)
Yes (0.00s cpu)

Often it is useful to instantiate the problem variables to the eplex solution,

which can be done using the following predicate:

return_solution :-
eplex_get(vars, Vars),
eplex_get(typed_solution, Vals),
Vars = Vals.

14.2.3 Integrality constraints

Handling integer variables dramatically increases the range of problems that
can be solved using eplex, but they also require very different solving meth-
ods. Integrality constraints in the resulting class of optimisation problems,
called mixed integer programming problems, cannot in general be solved
without resorting to search. They are dealt with within the eplex solver
using a search method transparent to the user.
A domain declaration with integer bounds does not automatically impose
integrality within eplex, in order to leave the option open for the integrality
to be imposed by another solver or search routine. To impose an integrality
constraint on a list List of variables in eplex it is necessary to explicitly
call integers(List).
Consider a simple example:

lptest(W, X) :-
eplex_solver_setup(min(X)),
[W,X] :: 0..10,
2*W+X $= 5,
 14.2 The eplex library 283

eplex_solve(_),
return_solution.

Neither W nor X are constrained to be integers. The optimal solution

returned by eplex is W = 2.5, X = 0.0. For this query ECLi PSe issues a
warning:

[eclipse 6]: lptest(W, X).

WARNING: Imposing integer bounds on variable(s) [W, X]

for eplex instance eplex does not impose integer type.

W = 2.5
X = 0.0
Yes (0.02s cpu)

In the definition of miptest/2 below, the integrality of W is enforced in

eplex:

miptest(W, X) :-
eplex_solver_setup(min(X)),
[W,X] :: 0.0..10.0,
integers([W]),
2*W+X $= 5,
eplex_solve(_),
return_solution.

Now the solver performs a mixed integer programming algorithm and finds
an optimal solution with W integral:

[eclipse 7]: miptest(W, X).

W = 2
X = 1.0
Yes (0.00s cpu)

The internal (branch and bound) search method involving integer vari-
ables is as follows.

(i) Solve the linear constraints in eplex, as if all the variables were con-
tinuous.
(ii) From the solution select an integer variable, W , whose eplex value
val is non-integral. If there is no such variable the problem is solved.
284 Linear constraints over continuous and integer variables

(iii) Let valint be the smallest integer larger than val and valint the largest
integer smaller than val. Then val int = valint + 1.
Make a search choice: add either the constraint W val int or
the constraint W valint . Return to (i). On backtracking try the
alternative choice.
(iv) To find an optimal solution, use the branch and bound search de-
scribed in Subsection 6.6.3.

14.3 Solving satisfiability problems using eplex

Thanks to the integrality constraints the eplex solver can also be used to
solve sets of Boolean constraints. The underlying computation relies on
the branch and bound search. Usually, a different representation is used,
called a propositional satisfiability problem. It involves Boolean vari-
ables and one deals with sets of clauses which in this case denote arbitrary
disjunctions of literals. In turn, a literal is either a variable or a negated
variable.
The encoding is simple: Boolean variables are modelled as 0-1 variables;
a negated variable not X is modelled as 1-X and a clause, say X or Y or
Z, is encoded as an inequality X+Y+Z $>= 1. For example, the satisfiability
problem
X or Y,
Y or Z,
Z or X,
not X or not Y,
not Y or not Z,
is modelled in eplex as follows:

sat(X, Y, Z) :-
eplex_solver_setup(min(0)),
integers([X,Y,Z]),
[X,Y,Z] :: 0..1,
X+Y $>= 1, Y+Z $>= 1, Z+X $>= 1,
(1-X) + (1-Y) $>= 1, (1-Y) + (1-Z) $>= 1,
eplex_solve(_),
return_solution.

Running the query sat(X, Y, Z) yields a unique solution:

[eclipse 8]: sat(X, Y, Z).

 14.4 Repeated solver waking 285

X = 1
Y = 0
Z = 1
Yes (0.01s cpu)

14.4 Repeated solver waking

The most interesting uses of the linear constraint solver are ones where
a set of linear constraints appears as part of a larger problem with other
complicating constraints. One way of exploiting the linear constraints is to
solve the linear subproblem at each node within a branch and bound search.
New linear constraints may be added on the different branches below the
node and an extended linear problem can be solved at each subnode.

14.4.1 A simple non-linear problem

To illustrate this on a toy example, we modify the previous problem and
amend
the minimisation expression, so as to minimise not X but W +
X instead. Because the square root function is non-linear it cannot be
represented within eplex. The simplest approach seems to be to minimise
W + X, and instead of returning the value of W + X as the optimum, to
return W + X. Unfortunately the solution that minimises W + X is
not, in general, the solution that minimises W + X. Indeed, reconsider the
constraints in our previous example:
cons(W, X) :-
[W,X] :: 0..10,
integers([W]),
2*W+X $= 5.
The minimum
value of W + X is 3, when X = 1, but the minimum value
of W + X is approximately 2.24, when X = 5. To find the true minimum
we include a variable SqrtX in our model, and use a linear approximation
of the constraint SqrtX = X. A linear approximation is, of course, a set
of linear constraints that approximate the non-linear constraint.
For this problem we generate new linear constraints during search which
are added to the initial linear approximation to make it better and better.
Search stops when the optimal values of X and SqrtX which satisfy the
current linear approximation are near enough in the sense that abs(SqrtX

X) . In what follows we choose for  the value 1e-5.

286 Linear constraints over continuous and integer variables

Our non-linear search nlsearch/2 is thus defined by two clauses. The

first one states
that the search can halt when the absolute difference between
SqrtX and X is no more than :
nlsearch(X, SqrtX) :-
eplex_var_get(X, solution, Val),
eplex_var_get(SqrtX, solution, SqrtVal),
abs(sqrt(Val)-SqrtVal) =< 1e-5, !,
return_solution.
If the difference is not small enough, then search continues by making
a choice, adding some additional linear constraints, and searching further.
This is the second clause defining nlsearch/2 which, because of the cut in
the previous clause, is only taken if the gap is still not small enough:
nlsearch(X, SqrtX) :-
make_choice(X),
add_linear_cons(X, SqrtX),
nlsearch(X, SqrtX).
The predicates used in this second clause are defined as follows.
make choice/1.
The idea is to reduce the interval associated with X at each search
step until the linear constraints on the square root sufficiently closely
approximate the actual square root. This is achieved by simply splitting
the interval associated with X at its current eplex value:
make_choice(X) :-
eplex_var_get(X, solution, Val),
( X $>= Val ; X $=< Val ).
add linear cons/2.
This constraint adds a better
and better approximation to the non-
linear constraint SqrtX = X. Because the square root of X increases
with
X, we can constrain the square root of X to fall within the range
[ M in.. M ax], where
[M
in..M ax] is the
current
domain of X. More-
over, we also have X M in X and X M ax X.
We can express these constraints using the eplex var get bounds/3
built-in that extracts the bounds of an eplex variable:
add_linear_cons(X, SqrtX) :-
eplex_var_get_bounds(X, Min, Max),
SqrtMin is sqrt(Min),
 14.4 Repeated solver waking 287

SqrtMax is sqrt(Max),
SqrtX $>= SqrtMin,
SqrtX $=< SqrtMax,
SqrtX*SqrtMin $=< X,
SqrtX*SqrtMax $>= X.

14.4.2 Setting up the eplex solver with triggers

The nlsearch/2 predicate can be embedded in a branch and bound search
in which the linear solver is invoked at each search node. To set up the
linear solver so that it can be called again, automatically, whenever a new
constraint is added, we use the eplex solver setup/4 built-in. It does
the job of both eplex solver setup/1 and eplex solve/1. It has four
arguments:

eplex_solver_setup(OptFun, Cost, Options, Triggers),

where

The optimisation function OptFun is as for eplex solver setup/1 (for

example min(X)).
The Cost argument is the optimal value of this function, which is the same
as the argument of eplex solve/1. Because the solver runs immediately,
as well as waking up and running again when an event triggers it, it is
necessary to impose initial bounds on the optimisation variables, in this
case X and Cost. They would normally be initialised to the same range.
The third argument is a list of options. We will not choose any options
for the considered example, so we will use the the empty list [].
The final argument specifies trigger conditions which will automatically
wake the linear solver. We will specify two trigger conditions, bounds
and new constraint, which wake the solver whenever a variable bound
is tightened, or a new linear constraint is imposed.

In the following we assume that we are dealing with a minimisation prob-

lem. The most important feature of eplex solver setup/4, which we will
exploit immediately, is that each time the solver runs, and finds a new mini-
mum, this is imposed as a new lower bound on the variable Cost. This lower
bound is indeed a valid logical consequence of the optimisation procedure,
and it remains a lower bound no matter how many more linear constraints
are added to the problem (they can only make the problem harder to solve,
thus making the minimum value worse, i.e., higher).
288 Linear constraints over continuous and integer variables

14.4.3 Non-linear optimisation by branch and bound

Whilst the eplex solver imposes tighter lower bounds on the Cost variable,
the branch and bound built-in minimize(Query, Cost) imposes a tighter
upper bound on Cost each time a better solution has been found. If, at
any point in the search tree, the lower bound on Cost imposed by the linear
solver becomes greater than the upper bound imposed by the minimize/2
predicate, the search fails at this point, and the whole subtree below the
current node is pruned away.
The ECLi PSe predicate that employs this approach to solve our non-linear
minimisation problem is defined as follows:

nonlin(W, X, Cost) :-
Cost :: 0.0..inf, %1
eplex_solver_setup(min(Cost),
Cost,
[],
[bounds, new_constraint]), %2
cons(W, X), %3
add_linear_cons(X, SqrtX), %4
Cost $= W+SqrtX, %5
minimize(nlsearch(X, SqrtX), Cost), %6
return_solution, %7
eplex_cleanup. %8

At the first line the optimisation variable is given an initial domain. At

line 3 the problem constraints are imposed. Line 4 imposes the initial lin-
ear approximation to the square root. Line 5 constrains the optimisation
variable W , and line 6 specifies the search.
At line 7 the predicate return solution is invoked to instantiate to its
eplex value any variables in the initial query nonlin(W, X, Cost) that
were not already instantiated during search (in this case W ).
Finally, at line 8, the final predicate eplex cleanup/0 is needed to stop
the linear solver because otherwise it remains suspended even after the
branch and bound search has ended, still waiting for a new event to ac-
tivate it.
The initial non-linear optimisation problem can now be solved:

[eclipse 9]: nonlin(W, X, Cost).

W = 0
X = 5.0
 14.5 The transportation problem 289

Cost = 2.23606797749979
Yes (0.03s cpu)

14.5 The transportation problem

We now explain how to use the eplex to solve the facility location problem
introduced in Subsection 6.5.4. Recall that the problem is concerned with a
choice of a set of facilities, for instance warehouses, to best serve a given set of
customers with a given set of demands. Each customer is characterised by its
demand for the provided product. For each potential warehouse location and
each client there are transportation costs of the product from the warehouse
to the client, per unit of supplied quantity. The objective is to satisfy the
demands of all customers minimising the total cost, including transportation
costs.
We will tackle this problem in three stages, illustrating increasingly large
and challenging classes of optimisation problems. The first stage is the
transportation problem: how can a set of customers be served most effi-
ciently from a set of warehouses. The second stage is the facility location
problem: at which locations, from a finite set of alternatives, should we open
warehouses so as to most efficiently serve a set of customers. Notice that
the transportation problem is a subproblem of the facility location problem.
The third stage is the non-linear facility location problem, a non-linear ex-
tension of the facility location problem where the capacity of each warehouse
is a variable, and the cost of opening it has a non-linear relationship with its
capacity. These three problems belong respectively to the following problem
classes:

linear problems that consists of linear constraints on continuous variables,

integer linear problems that consists of linear constraints on both contin-
uous and discrete variables,
non-linear problems that consists of non-linear constraints on both con-
tinuous and discrete variables.
To represent the data of the transportation problem we use the following
five predicates:

cust count/1; its argument is an integer, the number of customers,

warehouse count/1; its argument is an integer, the number of ware-
houses,
capacities/1; its argument is a list of numbers, one for each warehouse,
demands/1; its argument is a list of numbers, one for each customer,
290 Linear constraints over continuous and integer variables

transport costs/1; its argument is a two-dimensional array of numbers,

one for each customer and warehouse.

Here is an example data file, illustrated in Figure 14.1:

cust_count(4).
warehouse_count(3).

capacities([600,500,600]).

demands([400,200,200,300]).

transport_costs([]([](5,4,1),
[](3,3,2),
[](4,2,6),
[](2,4,4))).

customer 1
demand: 400
st: 5
t co
warehouse 1
t cost: 3
capacity: 600
customer 2
demand: 200

warehouse 2
tc
capacity: 500 ost
:4
customer 3
demand: 200
warehouse 3
t
co

capacity: 600
st
:2

customer 4
demand: 300

Fig. 14.1 A transportation problem

 14.5 The transportation problem 291

This problem has a decision variable for each customer and warehouse
reflecting the amount of goods transported from the warehouse to the cus-
tomer. To ensure the model is generic we use a matrix to hold all the decision
variables. We solve it using the predicate transport/2 defined by:

transport(Supplies, Cost) :-
eplex_solver_setup(min(Cost)),
init_vars(Supplies),
supply_cons(Supplies),
cost_expr(Supplies, Cost),
eplex_solve(Cost),
return_solution.

This has the usual structure of a constraint program, except the first and
last lines which are specific to eplex. The arguments contain the decision
variables, as discussed above. The objective is to minimise the variable Cost.
The data structures containing the decision variables are initialised as
follows:

init_vars(Supplies) :-
init_supplies(Supplies).

init_supplies(Supplies) :-
cust_count(CustCt),
warehouse_count(WCt),
dim(Supplies,[CustCt,WCt]),
( foreachelem(S, Supplies)
do
0 $=< S
).

So to perform the initialisation we retrieve the number of customers from

cust count/1 and the number of warehouses from warehouse count/1. No-
tice that we have to explicitly constrain the supplied quantities, i.e., the
Supply entries, to be non-negative.
The constraints are expressed more or less as they were written in the
problem specification in Subsection 6.5.4. We formulate them using the
sum/1 built-in introduced in Subsection 10.3.2:

supply_cons(Supplies) :-
capacity_cons(Supplies),
demand_cons(Supplies).
292 Linear constraints over continuous and integer variables

capacity_cons(Supplies) :-
capacities(Capacities),
( count(WHouse, 1, _),
foreach(Cap, Capacities),
param(Supplies)
do
cust_count(CustCt),
Cap $>= sum(Supplies[1..CustCt, WHouse])
).

demand_cons(Supplies) :-
demands(Demands),
( count(Cust, 1, _),
foreach(Demand, Demands),
param(Supplies)
do
warehouse_count(WCt),
sum(Supplies[Cust, 1..WCt]) $>= Demand
).

So we stipulate that the sum of the supplies from each warehouse does
not exceed its capacity and the sum of supplies to each customer meets or
exceeds its demand.
Finally, we define the cost expression which is the sum of the resulting
transportation costs:

cost_expr(Supplies, CostExpr ) :-
transport_expr(Supplies, TransportExpr),
CostExpr $= TransportExpr.

transport_expr(Supplies, sum(TExprList)) :-
transport_costs(TransportCosts),
( foreachelem(TCost, TransportCosts),
foreachelem(Qty, Supplies),
foreach(TExpr, TExprList)
do
TExpr = TCost*Qty
).
 14.5 The transportation problem 293

Notice the use of = to construct the expression TExpr.3 If we denote by

SC,W the decision variable representing the amount of goods supplied to
customer C from warehouse W , the constraints generated by ECLi PSe are
as follows:
S1,1 + S1,2 + S1,3 400,
S2,1 + S2,2 + S2,3 200,
S3,1 + S3,2 + S3,3 200,
S4,1 + S4,2 + S4,3 300,
600 S1,1 + S2,1 + S3,1 + S4,1 ,
500 S1,2 + S2,2 + S3,2 + S4,2 ,
600 S1,3 + S2,3 + S3,3 + S4,3 .
Additionally all the variables are constrained to be positive. The expres-
sion minimised by ECLi PSe is:
5 S1,1 + 4 S1,2 + 1 S1,3 +
3 S2,1 + 3 S2,2 + 2 S2,3 +
4 S3,1 + 2 S3,2 + 6 S3,3 +
2 S4,1 + 4 S4,2 + 4 S4,3 .
The optimal solution returned by ECLi PSe for this problem instance is
1800.0:

[eclipse 10]: transport(Supplies, Cost).

Supplies = []([](0.0, 0.0, 400.0),

[](0.0, 0.0, 200.0),
[](0.0, 200.0, 0.0),
[](300.0, 0.0, 0.0))
Cost = 1800.0
Yes (0.00s cpu)

3 This is not a numerical equation, but simply a clearer way of writing:

( foreachelem(TCost, TransportCosts),
foreachelem(Qty, Supplies),
foreach(TCost*Qty, TExprList)
do
true
)
294 Linear constraints over continuous and integer variables

14.6 The linear facility location problem

Recall that the facility location problem has a decision variable associated
with each potential warehouse location. This is a binary variable, with 0
representing that the location is not chosen, and 1 that it is chosen, for an
open warehouse. These binary decision variables make the problem harder
to solve. (In fact, the class of facility location problems is NP-hard.)
To model this problem in ECLi PSe we introduce, in addition to the trans-
portation problem data and variables:
a new data item generated by setup costs/1; its argument is a list of
numbers, one for each potential warehouse location, for example

setup_costs([100,800,400]).
a list OpenWhs of Boolean variables, one for each potential warehouse
location.
The data structures are now initialised by the following modified init vars/2
predicate:
init_vars(OpenWhs, Supplies) :-
init_openwhs(OpenWhs),
init_supplies(Supplies).

init_openwhs(OpenWhs) :-
warehouse_count(WCt),
length(OpenWhs, WCt),
OpenWhs :: 0.0..1.0,
integers(OpenWhs).
The key extensions to the transportation problem are in the constraints
on how much goods can be supplied to a customer from a warehouse, and an
extension to the cost function. Consider first the modified supply constraints
with the amended constraints on the supply from each warehouse:
supply_cons(OpenWhs, Supplies) :-
capacity_cons(Supplies, OpenWhs),
demand_cons(Supplies).

capacity_cons(Supplies, OpenWhs) :-
capacities(Capacities),
( count(WHouse, 1, _),
foreach(OpenWh, OpenWhs),
 14.6 The linear facility location problem 295

foreach(Cap, Capacities),
param(Supplies)
do
cust_count(CustCt),
Cap*OpenWh $>= sum(Supplies[1..CustCt, WHouse])
).
So if the warehouse WHouse is not chosen, then the corresponding OpenWh
variable is 0. The constraint ensures that in this case the sum of all the
supplies from this warehouse is 0. However, if the warehouse WHouse is
chosen, then OpenWh = 1 and the constraint reduces to the one used in the
previous section.
This way of handling Boolean variables, such as OpenWhs, is standard in
mixed integer programming. The constraint M*Bool $>= Expr is called a
big M constraint. Our big M constraint is:
Cap*OpenWh $>= sum(Supplies[1..CustCt, WHouse]).
This example is especially interesting as, by choosing M equal to the ca-
pacity Cap of the warehouse, we have created a constraint that does two
things at the same time. If the warehouse is not chosen, it precludes any
supply from that warehouse, and if the warehouse is chosen, it restricts its
total supply to be no more than the capacity of the warehouse.
The second extension of the transportation problem model is an additional
argument in the cost function, which is now defined as follows:
cost_expr(OpenWhs, Supplies, Cost) :-
setup_expr(OpenWhs, SetupExpr),
transport_expr(Supplies, TransportExpr),
Cost $= SetupExpr+TransportExpr.

setup_expr(OpenWhs, sum(SExprList)) :-
setup_costs(SetupCosts),
( foreach(OpenWh, OpenWhs),
foreach(SetupCost, SetupCosts),
foreach(SExpr, SExprList)
do
SExpr = OpenWh*SetupCost
).
(The transport expr/2 predicate is defined in the previous section.) So
the cost variable Cost is now constrained to be equal to the sum of the setup
costs of the chosen warehouses plus the resulting transportation costs:
296 Linear constraints over continuous and integer variables

Cost $= SetupExpr+TransportExpr.
We encode now the facility location problem similarly to the transporta-
tion problem:

warehouse(OpenWhs, Supplies, Cost) :-

eplex_solver_setup(min(Cost)),
init_vars(OpenWhs, Supplies),
supply_cons(OpenWhs, Supplies),
cost_expr(OpenWhs, Supplies, Cost),
eplex_solve(Cost),
return_solution.

The problem instance defined by the example data items above yields the
following result:

[eclipse 11]: warehouse(OpenWhs, Supplies, Cost).

OpenWhs = [1, 0, 1]
Supplies = []([](0.0, 0.0, 400.0),
[](0.0, 0.0, 200.0),
[](200.0, 0.0, 0.0),
[](300.0, 0.0, 0.0))
Cost = 2700.0
Yes (0.00s cpu)

If the integrality constraints on the Boolean variables are dropped, the

optimal cost Cost improves to 2570.0.

14.7 The non-linear facility location problem

The non-linear facility location problem has the same data and decision
variables as the previous problem, except that now the capacity of each
warehouse is a decision variable instead of being given as input data.
The cost of setting up a warehouse, in this non-linear version of the ware-
house problem, includes not only a standard (initial) setup cost, but also
an extra cost proportional to the square root of the capacity of the ware-
house. This extra cost reflects the fact the larger warehouses cost more, but
the cost does not go up linearly with the capacity of the warehouse: larger
warehouses are effectively cheaper per unit capacity.
Just as the non-linear facility location problem is an extension of the
original facility location problem, the solution is an extension of the original
 14.7 The non-linear facility location problem 297

facility location solution. This is a nice feature of constraint programming,

and contrasts with traditional modelling and optimisation formalisms.
As for the non-linear problem discussed in Section 14.4, the program for
solving the non-linear warehouse problem uses not only the eplex library
but also the branch and bound library.
The definition of the main predicate for the non-linear facility location
problem combines the structure of the predicate nonlin/2 of Section 14.4
with that of warehouse/3 of Section 14.6.
nlwarehouse(OpenWhs, Capacities, Cost) :-
Cost :: 0.0..inf,
eplex_solver_setup(min(Cost),
Cost,
[],
[new_constraint, deviating_bounds]),
init_vars(OpenWhs, Supplies, Capacities, SqrtCaps),
supply_cons(OpenWhs, Supplies, Capacities),
cost_expr(OpenWhs, Supplies, SqrtCaps, Cost),
minimize(nlsearch(Capacities, SqrtCaps), Cost),
return_solution,
eplex_cleanup.
The waking conditions, or triggers, on the linear solver ensure it is acti-
vated whenever
a new linear constraint is added to the problem, or
a bound on a variable becomes tightened so as to exclude its linear optimal
value from the previous activation of the linear solver.
The second trigger deviating bounds is newly introduced here. It can avoid
a lot of unnecessary calls to the solver which the simpler trigger bounds
would have made.
The decision variables are initialised as before, but this time we introduce
two new lists of variables, Capacities and SqrtCaps:
init_vars(OpenWhs, Supplies, Capacities, SqrtCaps) :-
init_openwhs(OpenWhs),
init_supplies(Supplies),
init_capacities(Capacities, SqrtCaps).
Capacities and SqrtCaps are, of course, the warehouse capacity vari-
ables, and a corresponding list of variables which will, eventually, be instan-
tiated to the square root of the warehouse capacities. The square roots are
298 Linear constraints over continuous and integer variables

needed only for the cost expression. The newly introduced predicate
init capacities/2 is defined as follows:

init_capacities(Capacities, SqrtCaps) :-
warehouse_count(WCt),
length(Capacities, WCt),
length(SqrtCaps, WCt),
total_demand(CapUPB),
SqrtCapUPB is sqrt(CapUPB),
Capacities :: 0.0..CapUPB,
SqrtCaps :: 0.0..SqrtCapUPB,
( foreach(Cap, Capacities),
foreach(SqrtCap, SqrtCaps)
do
add_linear_cons(Cap, SqrtCap)
).

total_demand(CapUPB) :-
demands(Demands),
CapUPB is sum(Demands).

So we stipulate that the upper bound on the capacity variables is the total
demand from all the customers this is the largest warehouse capacity that
could ever be needed.
The add linear cons/2 predicate, originally described in Section 14.4,
posts a set of linear constraints that approximate the square root relation.
The next program section posts the problem constraints:

supply_cons(OpenWhs, Supplies, Capacities) :-

capacity_cons(Supplies, OpenWhs, Capacities),
demand_cons(Supplies).

capacity_cons(Supplies, OpenWhs, Capacities) :-

total_demand(BigM),
( count(WHouse, 1, _),
foreach(OpenWh, OpenWhs),
foreach(Cap, Capacities),
param(Supplies, BigM)
do
cust_count(CustCt),
Cap $>= sum(Supplies[1..CustCt, WHouse]),
 14.7 The non-linear facility location problem 299

BigM*OpenWh $>= sum(Supplies[1..CustCt, WHouse])

).

(The demand cons/2 predicate is defined in Section 14.5.) Since the capac-
ity Cap is now a variable, the expression Cap*OpenWh is no longer linear, and
therefore it cannot be used in a linear constraint, as in the capacity cons/2
predicate in the previous section. Instead, the constraint
Cap $>= sum(Supplies[1..CustCt, WHouse])
stating that there is sufficient capacity in the warehouse to meet the supply
commitment is added separately from the constraint

BigM*OpenWh $>= sum(Supplies[1..CustCt, WHouse])

stating that a non-zero supply is only possible from an open warehouse.
The cost expression is defined largely as before, except that the auxiliary
setup exp/3 predicate now includes a subexpression involving the square
root of the warehouse capacity:

cost_expr(OpenWhs, Supplies, SqrtCaps, Cost) :-

setup_expr(OpenWhs, SqrtCaps, SetupExpr),
transport_expr(Supplies, TransportExpr),
Cost $= SetupExpr+TransportExpr.

setup_expr(OpenWhs, SqrtCaps, sum(SExprList)) :-

setup_costs(SetupCosts),
( foreach(OpenWh, OpenWhs),
foreach(SqrtCap, SqrtCaps),
foreach(SetupCost, SetupCosts),
foreach(SExpr, SExprList)
do
SExpr = (OpenWh+SqrtCap*0.1)*SetupCost
).

The value 0.1 in the expression (OpenWh+SqrtCap*0.1)*SetupCost se-

lects the balance between the fixed setup cost and the non-linear cost related
to capacity. Experimentally this particular choice of coefficient leads to an
interesting optimisation behaviour. Any other coefficient could be used here
instead.
Up to now no constraint has been imposed to ensure that SqrtCap is
indeed the square root of the chosen warehouse capacity. This relationship
is enforced by the following nlsearch/2 predicate:
300 Linear constraints over continuous and integer variables

nlsearch(Capacities, SqrtCaps) :-
( too_different(Capacities, SqrtCaps, Cap, SqrtCap) ->
make_choice(Cap),
add_linear_cons(Cap, SqrtCap),
nlsearch(Capacities, SqrtCaps)
;
return_solution
).

too_different(Capacities, SqrtCaps, Cap, SqrtCap) :-

get_pair(Capacities, SqrtCaps, Cap, SqrtCap),
eplex_var_get(Cap, solution, CapVal),
eplex_var_get(SqrtCap, solution, SqrtCapVal),
abs(sqrt(CapVal)-SqrtCapVal) $>= 1e-5.

get_pair([X | _], [Y | _], X, Y).

get_pair([_ | TX], [_ | TY], X, Y) :-
get_pair(TX, TY, X, Y).

The predicate too different/4 checks through the warehouse capacities

Cap and the associated square root variables SqrtCap until it finds a pair
for which the value of SqrtCap, held internally in the linear solver, is more
than  (here 1e-5) apart from the square root of the linear solvers value for
Cap.
If it finds such a pair, it returns it and makes a search choice by calling the
make choice/1 predicate defined in Section 14.4, subsequently imposes the
new (better) linear approximation of the square root function, and continues
the search by invoking nlsearch/2 again.
Whilst the naive linear approximation described earlier works, the rate
at which the algorithm converges to the optimal solution is poor. Faster
convergence is achieved by sharpening the linear approximation. A more
precise approximation is

(SqrtM ax SqrtM in)/(M ax M in) (SqrtX SqrtM in)/(X M in).

It holds because the function X is concave, so in the interval between
the current lower bound M in of X and its upper bound M ax
the square
root of X cannot fall below the straight line joining M in and M ax. This
can be expressed as a linear constraint in the form:

(SqrtM ax SqrtM in) (X M in) (SqrtX SqrtM in) (M ax M in).

 14.7 The non-linear facility location problem 301

We therefore add this linear constraint to the definition of

add linear cons/2:
add_linear_cons(X, SqrtX) :-
eplex_var_get_bounds(X, Min, Max),
SqrtMin is sqrt(Min), SqrtMax is sqrt(Max),
SqrtX $>= SqrtMin, SqrtX $=< SqrtMax,
SqrtX*SqrtMin $=< X,
SqrtX*SqrtMax $>= X,
(SqrtMax-SqrtMin)*(X-Min) $=< (SqrtX-SqrtMin)*(Max-Min).
If each SqrtCap variable is sufficiently (< ) close to its Cap variable, then
the search terminates and we have a solution. The variables are instantiated
to their values in the linear solver, and nlsearch/2 succeeds. The minimi-
sation, however, continues because nlsearch/2 was originally invoked from
the call
minimize(nlsearch(Capacities, SqrtCaps), Cost).
Only when nlsearch/2 fails does the branch and bound minimisation rou-
tine finally terminate successfully.
The non-linear facility location problem is solved thus in ECLi PSe :
[eclipse 12]: nlwarehouse(OpenWhs, Capacities, Cost).
Found a solution with cost 3944.3710282197303
Found a solution with cost 3903.40269486325
Found no solution with cost
3574.3828982755144 .. 3902.40269486325

OpenWhs = [1, 0, 1]
Capacities = [500.0, 0.0, 600.0]
Cost = 3903.40269486325
Yes (0.06s cpu)
This program scales up in a reasonable way. For a set of benchmarks
with 10 warehouses and 50 customers the program typically finds a solution
quickly and proves optimality in about 20 seconds.

14.7.1 Multiple Solvers

Finally, let us mention that if both ic and eplex are loaded, then some
constraints become ambiguous. For example X $>= 0 is both an ic and
an eplex constraint. In this case it is necessary to explicitly qualify each
302 Linear constraints over continuous and integer variables

constraint goal that could be ambiguous, specifying which solver should

handle it, thus:

ic: (X $>= 2)

or

eplex: (X $>= 2).

Unambiguous constraints, such as X #>= 2 do not need to be qualified.

Here is a simple program that can run with both solvers loaded:

test(X, Y) :-
ic: (X $>= 2),
Y #>= X.

14.8 Summary
In this chapter we introduced the eplex solver that allows us to solve linear
and non-linear constraints on integers and reals. We discussed in turn:

the limitations of the ic library,

the eplex library: constraint syntax and facilities,
handling of integrality and its use to solve satisfiability problems,
repeated eplex solver waking (within a branch and bound search).

Then we illustrated these techniques by considering three, gradually more

complex, problems: a transportation problem, a (linear) facility location
problem and a non-linear version of the facility location problem.
In the process we introduced the following built-ins:

eplex solver setup/1, to set up the eplex solver,

eplex solve/1, to invoke the eplex solver,
eplex var get/3, to extract information about solver variables,
eplex get/2, to extract information about the solver state,
integers/1, to impose integrality constraints in eplex, so that the inter-
nal eplex branch and bound search will be used,
eplex var get bounds/3, to extract the bounds of an eplex variable,
eplex solver setup/4, to set up the eplex solver with trigger conditions
for repeated waking,
eplex cleanup/0, to close the eplex solver, and kill its associated delayed
goals.
 14.9 Exercises 303

14.9 Exercises

Exercise 14.1 A company produces two types of products T1 and T2,

which requires the following resources to produce each unit of the product:

Resource T1 T2
Labour (hours) 9 6
Pumps (units) 1 1
Tubing (m) 12 16

The amounts of profit per unit of products are: T1 : 350, T2 : 300.

They have the following resources available: 1566 hours of labour, 200
pumps, and 2880 metres of tubing.
(i) Write a program to maximise the profit for the company, using eplex
as a black box solver. Write a predicate that returns the profit and
the values for T1 and T2.
(ii) What program change is required to answer this question: what profit
can be achieved if exactly 150 units of T1 are required?
(iii) What would the profit be if fractional numbers of products could be
produced?

Exercise 14.2 Explore the behaviour of the program solving the non-linear
facility location problem and presented in Section 14.7 with different:
(i) linear approximations
Add the following linear approximations one by one to determine
their effect on the search performance:
SqrtX $>= SqrtMin and SqrtX $=< SqrtMax,
SqrtX*SqrtMin $=< X and SqrtX*SqrtMax $>= X,
(SqrtMax-SqrtMin)*(X-Min) $=< (SqrtX-SqrtMin)*(Max-Min).
(ii) setup costs
Try the following setup costs:
SExpr = (OpenWh+SqrtCap*0.3)*SetupCost,
SExpr = (OpenWh+SqrtCap*0.1)*SetupCost,
SExpr = (OpenWh+SqrtCap*0.01)*SetupCost.
(iii) trigger conditions
Add the following trigger conditions one at a time:
304 Linear constraints over continuous and integer variables

new constraint,
bounds,
deviating bounds.
(iv) precision
Try varying the allowed difference between SqrtCapVal and
sqrt(CapVal) using:
10,
1,
1e-5.
 Solutions to selected exercises

Exercise 1.1
[eclipse 1]: X = 7, X = 6.

No (0.00s cpu)
[eclipse 2]: X = 7, X = X+1.

No (0.00s cpu)
[eclipse 3]: X = X+1, X = 7.

No (0.00s cpu)
[eclipse 4]: X = 3, Y = X+1.

Y = 3 + 1
X = 3
Yes (0.00s cpu)
[eclipse 5]: Y = X+1, X = 3.

Y = 3 + 1
X = 3
Yes (0.00s cpu)

Exercise 1.2
[eclipse 1]: book(X, apt), book(X, Y), Y \= apt.

X = constraint_logic_programming
Y = wallace
Yes (0.00s cpu)
Exercise 1.3
[eclipse 1]: Z = 2, q(X).

Z = 2
X = X
Yes (0.00s cpu)

305
306 Solutions to selected exercises

[eclipse 2]: Z = 2, q(Z).

Z = 2
Yes (0.00s cpu)
[eclipse 3]: Z = 2, q(X), Y = Z+1.

Z = 2
X = X
Y = 2 + 1

Exercise 1.4
[eclipse 1]: p(Behold), q(Behold).

Behold = [prolog, works, for, solving(problems), among, others]

Yes (0.00s cpu, solution 1, maybe more) ? ;

No (0.00s cpu)

Exercise 1.5
The first solution is obtained twice, by splitting Xs into [] and Xs or into Xs and
[]. In each case, by appending Xs to [] or [] to Xs we get back Xs. The following
modified definition of rotate/2 removes the possibility of such redundancies:
rotate(Xs, Ys) :- app(As, Bs, Xs), Bs \= [], app(Bs, As, Ys).

Exercise 2.1
The result is X=0, Y=0.
Exercise 2.2
The first procedure call fails, while the other two succeed.
Exercise 2.3
app(X1, X2, X3):
begin
begin new Ys;
X1 = []; X2 = Ys; X3 = Ys
end
orelse
begin new X,Xs,Ys,Zs;
X1 = [X | Xs]; X2 = Ys; X3 = [X | Zs];
app(Xs, Ys, Zs)
end
end

Exercise 3.1
[eclipse 1]: X is 7, X is 6.

No (0.00s cpu)
 Solutions to selected exercises 307

[eclipse 2]: X is 7, X is X+1.

No (0.00s cpu)
[eclipse 3]: X is X+1, X is 7.
instantiation fault in +(X, 1, X)
Abort
[eclipse 4]: X is 3, Y is X+1.

Y = 4
X = 3
Yes (0.00s cpu)
[eclipse 5]: Y is X+1, X is 3.
instantiation fault in +(X, 1, Y)
Abort

Exercise 3.2

[eclipse 1]: Z is 2, q(X).

Z = 2
X = X
Yes (0.00s cpu)
[eclipse 2]: Z is 2, q(Z).

Z = 2
Yes (0.00s cpu)
[eclipse 3]: Z is 2, q(X), Y is Z+1.

Z = 2
X = X
Y = 3
Yes (0.00s cpu)
[eclipse 4]: Z = 2, inc(X), X = 2.
instantiation fault in +(X, 1, Z) in module eclipse
Abort
[eclipse 5]: X = 2, inc(X), Z = 2.

X = 2
Z = 2
Yes (0.00s cpu)

Exercise 3.3
An example is the query min(1+1, 3, Z). For the first program it yields the
answer Z = 1+1 and for the second program the answer Z = 2.
Exercise 3.4
Assume the declarations
:- op(500, yfx, have).
:- op(400, yfx, and).
308 Solutions to selected exercises

The following interaction with ECLi PSe provides the answers.

[eclipse 1]: have(and(she,he), luck) = she and he have luck.

Yes (0.00s cpu)

[eclipse 2]: have(they, and(and(time,money),luck)) =
they have time and money and luck.

Yes (0.00s cpu)

Exercise 3.5
Here is a possible declaration that ensures that has length/2 is properly defined
and queries such as [a,b,c] has length X+1 are correctly parsed:
:- op(800, xfy, has_length).

Exercise 4.1
Here is one possible solution.

len(X, N) :-
( integer(N) ->
len1(X, N)
;
len2(X, N)
).

len1([], 0) :- !.
len1([_ | Ts], N) :- M is N-1, len1(Ts, M).

len2([], 0).
len2([_ | Ts], N) :- len2(Ts, M), N is M+1.

Exercise 4.2
This exercise illustrates a subtlety in using cut. Consider the query add(a, [a],
[a,a]). For the first program it yields the answer No and for the second program
the answer Yes.
Exercise 4.3
while(B, S) :-
( B ->
S,
while(B, S)
;
true
).

Exercise 5.1
Here is one possible solution.
 Solutions to selected exercises 309

vars(Term, []) :- atomic(Term), !.

vars(Term, [Term]) :- var(Term), !.
vars([H|T], VarList) :- !,
vars(H, L1),
vars(T, L2),
app(L1, L2, VarList).
vars(Term, VarList) :-
Term =.. [_|Args],
vars(Args, VarList).

Exercise 5.3
Here is one possible solution.
subs(Input, _X, _Term, Input) :- atomic(Input), !.
subs(Input, X, Term, Term) :- var(Input), X == Input, !.
subs(Input, _X, _Term, Input) :- var(Input), !.

subs(Input, X, Term, Output) :-

Input =.. [F|Args],
subslist(Args, X, Term, ArgsOut),
Output =.. [F|ArgsOut].

subslist([], _X, _Term, []).

subslist([Head|Tail], X, Term, [Output|OutputList]) :-
subs(Head, X, Term, Output),
subslist(Tail, X, Term, OutputList).

Exercise 7.1
( foreach(V,List),
fromto(0,This,Next,Count)
do
(nonvar(V) -> Next is This+1 ; Next = This)
).

Exercise 7.2
( foreachelem(El,Array),
fromto(0,This,Next,Count)
do
(nonvar(El) -> Next is This+1 ; Next = This)
).

Exercise 7.3
Here is one possible solution.
select_val(Min, Max, Val) :-
( count(I,Min,Val),
fromto(go,_,Continue,stop),
param(Max)
310 Solutions to selected exercises

do
( Continue = stop
;
I < Max, Continue = go
)
).
This is an example where the recursive encoding (provided in the SELECT program
from Figure 3.2) is, perhaps, easier to follow.
Exercise 7.4
( count(I,K,L) do Body ) is equivalent to
( L1 is L+1, fromto(K,I,M,L1) do M is I+1, Body ).
Exercise 8.1
setval(c,0),
( foreachelem(El,Array)
do
( var(El) -> true ; incval(c) )
),
getval(c,Count).

Exercise 8.2
Here is one possible solution.
all_solutions(Query, _List) :-
setval(sols,[]),
Query,
getval(sols,Old),
append(Old,[Query],New),
setval(sols,New),
fail.
all_solutions(_, List) :-
getval(sols,List).

Exercise 8.3
Here is one possible solution.
naive_search2(Vars,Vals) :-
( foreach(V,Vars),
count(Indent,0,_),
param(Vals)
do
setval(first,true),
member(V,Vals),
( getval(first,false) -> nl_indent(Indent) ; true ),
write(v), write(Indent), write( = ),
write(V), write(\t),
on_backtracking(setval(first,false))
).
 Solutions to selected exercises 311

nl_indent(N) :-
nl,
( for(_I,1,N) do write(\t) ).

on_backtracking(_).
on_backtracking(Q) :-
once(Q),
fail.
(\t stands for the TAB symbol.)
Exercise 9.1
The following solve/1 predicate does the job:
solve(A) :-
solve(A, true, Susp),
write_susp(Susp).

write_susp(true) :- !.
write_susp((A,B)) :-
nl, writeln(Delayed goals:),
write_susp1((A,B)).
write_susp1((A,B)) :-
( B = true -> true ; write_susp1(B) ),
write(\t), writeln(A).
We have then for example:
[eclipse 1]: solve((suspend:(2 < Y + 1), suspend:(X > 3))).

Delayed goals:
suspend : (2 < Y + 1)
suspend : (X > 3)

Y = Y
X = X
Yes (0.00s cpu)

Exercise 9.2
In the solution below the predicate diff_list/2 is defined by a double recursion
using the auxiliary predicate out of/2.
diff_list([]).
diff_list([E1|Tail]):-
out_of(E1, Tail),
diff_list(Tail).

out_of(_, []).
out_of(E1, [E2|Tail]):-
E1 $\= E2,
out_of(E1, Tail).
312 Solutions to selected exercises

Exercise 9.3
The following modification of the constraints/2 predicate does the job:
constraints(QueenStruct, Number) :-
( for(I,1,Number),
param(QueenStruct,Number)
do
QueenStruct[I] :: 1..Number,
( for(J,1,I-1),
param(I,QueenStruct)
do
subscript(QueenStruct,[I],Qi),
subscript(QueenStruct,[J],Qj),
Qi $\= Qj,
Qi-Qj $\= I-J,
Qi-Qj $\= J-I
)
).
We get then a substantial speed up w.r.t. the original solution, for example:
[eclipse 1]: queens(X, 8).

X = [](1, 5, 8, 6, 3, 7, 2, 4)
Yes (0.02s cpu, solution 1, maybe more) ?

Exercise 9.4
Here is one possible solution.
filter([], [], unsat).
filter([H | T], [H | Rest], Sat) :-
var(H), !, filter(T, Rest, Sat).
filter([0 | T], Rest, Sat) :- !,
filter(T, Rest, Sat).
filter([1 |_], [], sat).

test_cl(List) :-
filter(List, Rest, Sat),
( Sat = sat -> true ;
Rest = [Var|_] -> suspend(test_cl(Rest), 2, Var->inst)
).

Exercise 9.5
Here is one possible solution.
ordered(List) :-
( fromto(List,[X|Successors],Successors,[]),
fromto([],Predecessors,[X|Predecessors],_RList)
do
( ground(X) ->
cg(X, Predecessors, Successors)
;
 Solutions to selected exercises 313

suspend(cg(X,Predecessors,Successors),2,X->inst)
)
).

cg(X, Predecessors, Successors) :-

check_gap(X, <, Successors),
check_gap(X, >, Predecessors).

check_gap(_, _, []) :- !.
check_gap(X, Comp, [Y|Rest]) :-
var(Y), !,
check_gap(X, Comp, Rest).
check_gap(X, Comp, [Y|_]) :-
comp(X, Comp, Y).

comp(X, <, Y) :- X < Y.

comp(X, >, Y) :- X > Y.
In this program as soon as an element X of the input list becomes instantiated,
it is checked using the call cg(X, Predecessors, Successors) against its nearest
instantiated predecessor and its nearest instantiated successor. Thanks to these
checks the query ordered(List) detects a failure if the input list contains two
integers in the wrong order.
Exercise 10.1
[eclipse 1]: X #= 7, X #= 6.

No (0.00s cpu)
[eclipse 2]: X #= 7, X #= X+1.

No (0.00s cpu)
[eclipse 3]: X #= X+1, X #= 7.

No (0.00s cpu)
[eclipse 4]: X #= 3, Y #= X+1.

X = 3
Y = 4
Yes (0.00s cpu)
[eclipse 5]: Y #= X+1, X #= 3.

Y = 4
X = 3
Yes (0.00s cpu)

Exercise 11.1
Here is one possible solution.
queens(Queens, Number):-
length(Queens, Number),
314 Solutions to selected exercises

Queens :: [1..Number],
constraints(Queens),
labeling(Queens).

constraints(Queens):-
noHorizontalCheck(Queens),
noDiagonalCheck(Queens).

noHorizontalCheck(Queens):-
diff_list(Queens).

noDiagonalCheck([]).
noDiagonalCheck([Q | Queens]):-
noCheck(Q, Queens, 1),
noDiagonalCheck(Queens).

noCheck(_, [], _).

noCheck(Q, [Q1 | Queens], I):-
Q $\= Q1-I,
Q $\= Q1+I,
I1 is I+1,
noCheck(Q, Queens, I1).
The constraints dealing with the diagonals are generated using an auxiliary pred-
icate noCheck/3. Given a variable (a queen) Q, and a list of variables Queens rep-
resenting the list of the queens to the right of Q, the query noCheck(Q, Queens,
1) generates the constraints stating that the queen Q does not attack along a di-
agonal any queen from Queens. Assuming Queens = [X1, ..., Xk], this is done
by generating the constraints
Q $\= X1 - 1, Q $\= X1 + 1, ..., Q $\= Xk - k, Q $\= Xk + k.
First the constraints Q $\= X1 - 1, Q $\= X1 + 1 are generated. The recursive
call of noCheck/3 deals with the remaining constraints. The second argument of
noCheck/3 becomes the tail of the list, here [X2, ..., Xk], and the third one gets
incremented by 1. The recursion terminates when the list becomes empty.
The use of the length/2 built-in allows one to call queens/2 with the first
argument uninstantiated.
Exercise 11.2
The solution below depends crucially on the constraint Psum(List) #= N. This
9
constraint holds since for any N -digit number we have i=0 occ(i) = N , where
occ(i) is the number of occurrences of the digit i in the number.
solve(List) :-
N is 10,
length(List,N),
List :: 0..9,
sum(List) #= N,
labeling(List),
correct(List).
 Solutions to selected exercises 315

correct(List) :-
( foreach(El,List),
count(I,0,_),
param(List)
do
occ(I, List, El)
).

occ(I, List, Count) :-

( foreach(El,List),
fromto(0,This,Next,Count),
param(I)
do
(El = I -> Next is This+1 ; Next = This)
).

The predicate occ/3 is actually present in ECLi PSe as the occurrences/3 built-
in in the ic global library. For a solution using this built-in see www.eclipse-clp.
org/examples/magicseq.ecl.txt.
Exercise 11.3
Here is one possible solution.

solve(X) :-
N is 10,
dim(X,[N]),
dim(Y,[N]),
subscript(X,[7],6),
( for(I,1,N),
param(X,Y,N)
do
X[I] :: 1..N,
( for(J,1,I-1),
param(I,X)
do
X[I] $\= X[J]
),
( I >= 2 ->
Y[I] $= X[I]-X[I-1],
Y[I] :: [3,-2]
;
true
)
),
( foreacharg(I,X) do indomain(I) ).

Exercise 12.1
For the first problem the continue strategy is much better than restart strat-
egy, while for the second problem the converse holds. Further, labeling(List)
performs well for the first problem, while for the second problem it is much better
316 Solutions to selected exercises

to use search(List,0,first_fail,indomain,complete,[]). We have for exam-

ple:
[eclipse 1]: prob1(List, Max),
bb_min(labeling(List),
Max,
bb_options{strategy:continue}).

[...]
List = [7, 5, 3, 1, 0, 2, 4, 6, 8]
Max = 530
Yes (6.48s cpu)

[eclipse 2]: prob1(List, Max),

bb_min(labeling(List),
Max,
bb_options{strategy:restart}).

[...]
List = [7, 5, 3, 1, 0, 2, 4, 6, 8]
Max = 530
Yes (87.19s cpu)

[eclipse 3]: prob2(List, Max),

bb_min(search(List,0,first_fail,indomain,complete,[]),
Max,
bb_options{strategy:restart}).

[...]
List = [14, 12, 10, 8, 6, 0, 1, 2, 3, 4, 5, 7, 9, 11, 13, 15]
Max = 468
Yes (1.57s cpu)

[eclipse 4]: prob2(List, Max),

bb_min(search(List,0,first_fail,indomain,complete,[]),
Max,
bb_options{strategy:continue}).

[...]
List = [12, 13, 10, 8, 6, 2, 3, 0, 1, 4, 5, 7, 9, 11, 14, 15]
Max = 468
Yes (104.27s cpu)

[eclipse 5]: prob2(List, Max),

bb_min(labeling(List),
Max,
bb_options{strategy:restart}).

[...]
List = [12, 13, 10, 8, 0, 1, 2, 3, 4, 5, 6, 7, 9, 11, 14, 15]
Max = 468
 Solutions to selected exercises 317

Yes (3326.68s cpu)

Exercise 12.2
Here is one possible solution.
domination(N,Min) :-
dim(Board,[N,N]),
matrix_to_list(Board, Vars),
Vars :: 0..1,
sum(Vars) $= Min,
( multifor([Row,Col],[1,1],[N,N]),
param(Board)
do
covered(Row, Col, Board)
),
minimize(labeling(Vars),Min),
print_board(Board).

covered(Row, Col, Board) :-

( foreachelem(X,Board,[I,J]),
fromto([],This,Next,AttackList),
param(Row,Col)
do
( attack(Row,Col,I,J) -> Next = [X|This] ; Next = This )
),
sum(AttackList) $>= 1.

attack(R,C,I,J) :-
R =:= I ; C =:= J ; R-C =:= I-J ; R+C =:= I+J.

matrix_to_list(Board, List) :-
( foreachelem(El,Board),
foreach(Var,List)
do
Var = El
).

print_board(Board) :-
( foreachelem(El,Board,[_,J])
do
( J =:= 1 -> nl ; true ),
write( ), write(El)
).
The call covered(Row, Col, Board) checks that the list of fields that attack
the square [Row,Col] (constructed in the variable Attacklist) contains at least
one queen (using the constraint sum(AttackList) $>= 1).
Exercise 12.3
Here is one possible solution with a sample input problem.
% problem(Nr, SquareSize, TileSizes)
318 Solutions to selected exercises

problem(1, 10, [7,6,5,4,3,2,1]).

squares(ProblemNr, Bs, Xs, Ys, Min) :-

problem(ProblemNr, Size, Sizes),
squares(Size, Sizes, Bs, Xs, Ys, Min).

squares(Size, Sizes, Bs, Xs, Ys, Waste) :-

( foreach(X,Xs),
foreach(Y,Ys),
foreach(B,Bs),
foreach(S,Sizes),
param(Size)
do
B :: 0..1,
Sd is Size-S+1,
[X,Y] :: 1..Sd
),
waste_expr(Size, Sizes, Bs, Waste),
no_overlap(Bs, Xs, Ys, Sizes),
capacity(Bs, Xs, Sizes, Size,Waste),
capacity(Bs, Ys, Sizes, Size,Waste),
append(Xs,Ys,XYs),
append(Bs,XYs,Vars),
minimize(search(Vars, 0, smallest, indomain, complete, []),
Waste).

waste_expr(Size, Sizes, Bs, Waste) :-

( foreach(S,Sizes),
foreach(B,Bs),
foreach(Used,UsedList)
do
Used #= B*S*S
),
Waste #= Size*Size - sum(UsedList),
Waste #>= 0.

no_overlap(Bs, Xs, Ys, Sizes) :-

( foreach(B,Bs),
fromto(Xs, [X|XXs], XXs, []),
fromto(Ys, [Y|YYs], YYs, []),
fromto(Sizes, [S|SSs], SSs, [])
do
( foreach(X1,XXs),
foreach(Y1,YYs),
foreach(S1,SSs),
param(B,X,Y,S)
do
no_overlap(B, X, Y, S, X1, Y1, S1)
)
 Solutions to selected exercises 319

).

no_overlap(B, X1, Y1, S1, X2, Y2, S2) :-

B #=<
(X1+S1 #=< X2 or X2+S2 #=< X1 or Y1+S1 #=< Y2 or Y2+S2 #=< Y1).

capacity(Bs, Cs, Sizes, Size,Waste) :-

( for(Pos,1,Size),
foreach(W,Ws),
param(Bs,Cs,Size,Sizes)
do
( foreach(C,Cs),
foreach(B,Bs),
foreach(S,Sizes),
param(Pos),
fromto(Sum,S*Bool*B+Sum1,Sum1,0)
do
::(C, Pos-S+1..Pos, Bool)
),
W #= eval(Size)-Sum,
W #>= 0
),
Waste #= sum(Ws).
Here ::/3 is the reified form of the variable declaration built-in ::/2 (see Sub-
section 9.4.4). We have then for example:
[eclipse 3]: squares(1, Bs, Xs, Ys, Min).
Found a solution with cost 100

[...]

Found a solution with cost 45

Bs = [0, 0, 1, 1, 1, 1, 1]
Xs = [1, 1, 1, 1, 6, 9, 1]
Ys = [1, 1, 1, 6, 1, 1, 10]
Min = 45
Yes (6.58s cpu)
 Bibliographic remarks

Chapter 1
The unification problem was introduced and solved in Robinson [1965]. The
MartelliMontanari unification algorithm was introduced in Martelli
and Montanari [1982]. The use of unification for computing is due to Kowal-
ski [1974]. The concept of a pure Prolog was popularised by the first edition
of Sterling and Shapiro [1994].

Chapter 2
The language L0 defined in Section 2.2 and the translation of pure Prolog
into it is based on Apt and Smaus [2001], where it is shown that this lan-
guage supports declarative programming. The declarative interpretation of
L0 programs is closely related to the so-called Clark completion of logic
programs introduced in Clark [1978], further discussed for example in Apt
and van Emden [1982].
The equivalence between the declarative interpretations of pure Prolog
programs and L0 programs w.r.t. successful queries follows from the results
of Apt and van Emden [1982] that compare the least Herbrand model of a
logic program with the models of Clark completion.

Chapter 4
OKeefe [1990] discusses in detail the proper uses of the cut. Sterling and
Shapiro [1994] provides an in-depth discussion of various meta-interpreters
written in Prolog. Apt and Turini [1995] is a collection of articles discussing
the meta-programming in Prolog.

Chapter 6
This chapter draws on material from Apt [2003]. The SEND + MORE
= MONEY puzzle is due to H. E. Dudeney and appeared in England, in

320
 Bibliographic remarks 321

the July 1924 issue of Strand Magazine. The representation of the SEND
+ MORE = MONEY puzzle and of the n-queens problem as a CSP is
discussed in Van Hentenryck [1989].
According to Russell and Norvig [2003] the eight queens problem was
originally published anonymously in the German chess magazine Schach in
1848 and its generalisation to the n-queens problem in Netto [1901]. The
coins problem is discussed in Wallace, Novello and Schimpf [1997].

Chapter 7
The iterators in ECLi PSe are designed and implemented by Joachim Schimpf.
In Schimpf [2002] an early version of the iterators and their implementation
in ECLi PSe is described. It is shown in Voronkov [1992] that in logic pro-
gramming appropriately powerful iterators have the same power as recur-
sion. The fromto/4 is an example of such an iterator. The QUEENS program
given in Figure 7.8 of Section 7.6 is modelled after an analogous solution in
the Alma-0 language of Apt et al. [1998]. The summary of the iterators is
based on the ECLi PSe user manual, see www.eclipse-clp.org/doc/bips/
kernel/control/do-2.html.

Chapter 8
The credit-based search was introduced in Beldiceanu et al. [1997]. The
limited discrepancy search is due to Harvey and Ginsberg [1995].

Chapter 10
The illustration of the constraint propagation for Boolean constraints by
means of the full adder example is from Fruhwirth [1995]. The use of the
SEND + MORE = MONEY puzzle to discuss the effect of constraint prop-
agation is from Van Hentenryck [1989, page 143]. The details of the con-
straint propagation mechanisms implemented in the sd and ic libraries are
discussed in Chapters 6 and 7 of Apt [2003].

Chapter 11
Exercise 11.2 is originally from Van Hentenryck [1989, pages 155157].

Chapter 12
The use of constraint logic programming to solve Sudoku puzzles can be
traced back to Carlson [1995]. Various approaches to solving them based
on different forms of constraint propagation are analysed in detail in Simo-
nis [2005]. Exercise 12.2 is from Wikipedia [2006]. Exercise 12.3 is a modi-
fied version of an exercise from www.eclipse-clp.org/examples/square_
tiling.pl.txt.
322 Bibliographic remarks

Chapter 13
The first comprehensive account of solving constraints on reals based on
constraint propagation is Van Hentenryck [1997] where a modelling language
Numerica is introduced. The Reimers system is taken from Granvilliers
[2003]. The MORTGAGE program was originally discussed in Jaffar and Lassez
[1987]. Exercise 13.3 is taken from Monfroy, Rusinowitch and Schott [1996].

Chapter 14
The facility location problem is a standard problem in operations research,
see, e.g., Vazirani [2001].
 Bibliography

K. R. Apt
[2003] Principles of Constraint Programming, Cambridge University Press,
Cambridge, England. Cited on page(s) 320, 321.
K. R. Apt, J. Brunekreef, V. Partington and A. Schaerf
[1998] Alma-0: An imperative language that supports declarative programming,
ACM Toplas, 20, pp. 10141066. Cited on page(s) 321.
K. R. Apt and J. Smaus
[2001] Rule-based versus procedure-based view of logic programming, Joint
Bulletin of the Novosibirsk Computing Center and Institute of Infor-
matics Systems, Series: Computer Science, 16, pp. 7597. Available via
http://www.cwi.nl/~apt. Cited on page(s) 320.
K. R. Apt and F. Turini
[1995] eds., Meta-logics and Logic Programming, The MIT Press, Cambridge,
Massachusetts. Cited on page(s) 320.
K. R. Apt and M. van Emden
[1982] Contributions to the theory of logic programming, Journal of the ACM,
29, pp. 841862. Cited on page(s) 320.
N. Beldiceanu, E. Bourreau, P. Chan and D. Rivreau
[1997] Partial search strategy in CHIP, in: Proceedings of the Second Interna-
tional Conference on Meta-Heuristics. Sophia-Antipolis, France. Cited
on page(s) 321.
I. Bratko
[2001] PROLOG Programming for Artificial Intelligence, International Com-
puter Science Series, Addison-Wesley, Harlow, England, third edn.
Cited on page(s) xiii.
B. Carlson
[1995] Compiling and Executing Finite Domain Constraints, PhD thesis, Upp-
sala University and SICS. Cited on page(s) 321.
K. L. Clark
[1978] Negation as failure, in: Logic and Databases, H. Gallaire and J. Minker,
eds., Plenum Press, New York, pp. 293322. Cited on page(s) 320.

323
324 Bibliography

A. Colmerauer
[1990] An introduction to Prolog III, Communications of the ACM, 33, pp. 69
90. Cited on page(s) x.
A. Colmerauer and P. Roussel
[1996] The birth of Prolog, in: History of Programming Languages, T. J. Bergin
and R. G. Gibson, eds., ACM Press/Addison-Wesley, Reading, Mas-
sachusetts, pp. 331367. Cited on page(s) x.
M. Dincbas, P. V. Hentenryck, H. Simonis, A. Aggoun, T. Graf and
F. Berthier
[1988] The constraint logic programming language CHIP, in: Proceedings of
the International Conference on Fifth Generation Computer Systems,
Tokyo, pp. 693702. Cited on page(s) xi.
T. Fruhwirth
[1995] Constraint Handling Rules, in: Constraint Programming: Basics and
Trends, A. Podelski, ed., LNCS 910, Springer-Verlag, Berlin, pp. 90
107. (Chatillon-sur-Seine Spring School, France, May 1994). Cited on
page(s) 321.
M. Gardner
[1979] Mathematical Circus, Random House, New York, NY. Cited on page(s)
108.
L. Granvilliers
[2003] RealPaver Users Manual, Version 0.3. http://www.sciences.
univ-nantes.fr/info/perso/permanents/granvil/realpaver/src/
realpaver-0.3.pdf. Cited on page(s) 322.
G. H. Hardy
[1992] A Mathematicians Apology, Cambridge University Press, Cambridge,
England, reprint edn. Cited on page(s) 251.
W. D. Harvey and M. L. Ginsberg
[1995] Limited discrepancy search, in: Proceedings of the Fourteenth Inter-
national Joint Conference on Artificial Intelligence (IJCAI-95); Vol.
1, C. S. Mellish, ed., Morgan Kaufmann, Montreal, Quebec, Canada,
pp. 607615. Cited on page(s) 321.
C. A. R. Hoare
[1962] Quicksort, BCS Computer Journal, 5, pp. 1015. Cited on page(s) 49.
J. Jaffar and J.-L. Lassez
[1987] Constraint logic programming, in: POPL87: Proceedings 14th ACM
Symposium on Principles of Programming Languages, ACM, pp. 111
119. Cited on page(s) xi, 322.
J. Jaffar, S. Michayov, P. J. Stuckey and R. H. C. Yap
[1992] The CLP(R) language and system, ACM Transactions on Programming
Languages and Systems, 14, pp. 339395. Cited on page(s) xi.
R. Kowalski
[1974] Predicate logic as a programming language, in: Proceedings IFIP74,
North-Holland, New York, pp. 569574. Cited on page(s) ix, 320.
 Bibliography 325

A. Mackworth
[1977] Consistency in networks of relations, Artificial Intelligence, 8, pp. 99
118. Cited on page(s) x.
A. Martelli and U. Montanari
[1982] An efficient unification algorithm, ACM Transactions on Programming
Languages and Systems, 4, pp. 258282. Cited on page(s) 320.
E. Monfroy, M. Rusinowitch and R. Schott
[1996] Implementing non-linear constraints with cooperative solvers, in: Pro-
ceedings of the ACM Symposium on Applied Computing (SAC 96), ACM
Press, New York, NY, USA, pp. 6372. Cited on page(s) 322.
U. Montanari
[1974] Networks of constraints: fundamental properties and applications to pic-
ture processing, Information Science, 7, pp. 95132. Also Technical Re-
port, Carnegie Mellon University, 1971. Cited on page(s) x.
E. Netto
[1901] Lehrbuch der Combinatorik, Teubner, Stuttgart. Cited on page(s) 321.
R. OKeefe
[1990] The Craft of Prolog, The MIT Press, Cambridge, Massachusetts. Cited
on page(s) 320.
C. H. Papadimitriou and K. Steiglitz
[1982] Combinatorial Optimization: Algorithms and Complexity, Prentice-Hall,
Englewood Cliffs, NJ. Cited on page(s) 278.
J. Robinson
[1965] A machine-oriented logic based on the resolution principle, J. ACM, 12,
pp. 2341. Cited on page(s) ix, 320.
S. Russell and P. Norvig
[2003] Artifical Intelligence: A Modern Approach, Prentice-Hall, Englewood
Cliffs, NJ, second edn. Cited on page(s) 321.
J. Schimpf
[2002] Logical loops, in: Proceedings of the 2002 International Conference on
Logic Programming, P. J. Stuckey, ed., vol. 2401 of Lecture Notes in
Computer Science, Springer-Verlag, pp. 224238. Cited on page(s) 321.
H. Simonis
[2005] Sudoku as a constraint problem, in: Proceedings of the Fourth Interna-
tional Workshop on Modelling and Reformulating Constraint Satisfac-
tion Problems, B. Hnich, P. Prosser and B. Smith, eds. Sitges, Spain.
Cited on page(s) 321.
L. Sterling and E. Shapiro
[1994] The Art of Prolog, The MIT Press, Cambridge, Massachusetts, sec-
ond edn. Cited on page(s) xiii, 320.
I. Sutherland
[1963] Sketchpad, A Man-Machine Graphical Communication System, Garland
Publishing, New York. Cited on page(s) x.
326 Bibliography

P. Van Hentenryck
[1989] Constraint Satisfaction in Logic Programming, The MIT Press, Cam-
bridge, Massachusetts. Cited on page(s) xi, 321.
[1997] Numerica: A Modeling Language for Global Optimization, The MIT
Press, Cambridge, Massachusetts. Cited on page(s) 322.
V. V. Vazirani
[2001] Approximation Algorithms, Springer-Verlag, Berlin. Cited on page(s)
322.
A. Voronkov
[1992] Logic programming with bounded quantifiers, in: Logic Programming
and Automated ReasoningProc. 2nd Russian Conference on Logic
Programming, A. Voronkov, ed., LNCS 592, Springer-Verlag, Berlin,
pp. 486514. Cited on page(s) 321.
M. Wallace, S. Novello and J. Schimpf
[1997] ECLi PSe : A platform for constraint logic programming, ICL Systems
Journal, 12, pp. 159200. Cited on page(s) 321.
Wikipedia
[2006] Eight queens puzzle. Entry http://en.wikipedia.org/wiki/Queens_
problem in Wikipedia. Cited on page(s) 321.
 Index

./2, 81 strategy:continue option, 251

//, 42 strategy:dichotomic option, 251
1e-n, 261 strategy:restart option, 251
::/2, 164, 186, 189 char code/2, 116
;/2, 62 clause/2, 69
</2, 47 compound/1, 77
=../2, 82 count/3, 111
=/2, 18 decval/1, 146
=:=/2, 47 delete/5, 217
=</2, 47 dim/2, 121
==/2, 78 div/2, 42
=>/2, 160 dynamic, 70
>/2, 47 eplex cleanup/0, 288
>=/2, 47 eplex get/2, 281
,/2, 60 eplex solve/1, 280
->/2, 62 eplex solver setup/1, 280
=\=/2, 47 eplex solver setup/4, 287
\==/2, 78 eplex var get/3, 281
#</2, 166 eplex var get bounds/3, 286
#=/2, 166 eval/1, 174
#=</2, 166 exclude/2, 199
#>/2, 166 fail/0, 65
#>=/2, 166 fix/1, 128
#\=/2, 166 flatten/2, 241
$</2, 162 for/3, 115
$=/2, 162 foreach/2, 111
$=</2, 162 foreacharg/2, 120
$>/2, 162 foreachelem/2, 122, 123
$>=/2, 162 fromto/4, 124
$\=/2, 162 functor/3, 80
&=/2, 186 fx, 55
&\=, 186 fxx, 54
!/0, 65 fxy, 54
abs/1, 42 fy, 55
alldifferent/1, 196, 241 get domain as list/2, 208
and/2, 160 get max/2, 206
append/3, 24 get min/2, 206
arg/3, 81 getval/2, 146
atom/1, 77 ground/1, 77
atomic/1, 77 halt/0, 17
bb min/3, 248 incval/1, 146
delta option, 248 indomain/1, 209
factor option, 248 indomain/2, 219

327
328 Index

inf/0, 264 Colmerauer, A., ix

integer/1, 75 comparison predicate, see arithmetic
integers/1, 165 comparison predicate
is/2, 43, 122 conditional solution, 265
is dynamic/1, 158 conjunction, 92
labeling/1, 212 constant, 5
length/1, 46 numeric, 92
length/2, 44 constrained optimisation problem, see COP
locate/2, 265 constraint, 90
maxlist/2, 255 active, 51
member/2, 67 all different, 96
minimize/2, 230 arithmetic, 93, 161
mod/2, 42 Boolean, 92, 160
multifor/3, 119 core, 160
neg/1, 160 disjunctive, 201
nonvar/1, 77 implied, 105
not/1, 64 integrality, 282
number/1, 49, 77 linear, 93
once/1, 65 passive, 51
or/2, 160 reified, 166
pi/0, 258 satisfiable, 90
reals/1, 165 solution to, 90
rem/2, 42 solved, 90
round/1, 128 unsatisfiable, 90
search/6, 223 constraint propagation, 103, 185
set threshold/1, 262 constraint satisfaction problem, see CSP
setval/2, 146 constraint store, 200
sqrt/1, 128 continuous variable, see variable
squash/3, 198, 268 COP, 98
subscript/3, 121 cryptarithmetic problem, 95
sum/1, 194 CSP, 90
suspend/3, 167 consistent, 91
term variables/2, 241 failed, 91
true/0, 70 feasible, 91
var/1, 76 inconsistent, 91
wake/0, 199 infeasible, 91
write/1, 62 solution to, 91
writeln/1, 111 solved, 91
xf, 55 cut, 65
xfx, 54
xfy, 54 decision variable, see variable
yf, 55 declarative programming, 35
yfx, 54 disjunction, 62, 92

equivalent CSPs, 91
ambivalent syntax, 59
expression
answer, 12
arithmetic, 43, 93
arithmetic comparison predicate, 47
Boolean, 92
arithmetic constraint, see constraint
linear, 92
arithmetic evaluator, 43
array, 121 fact, 6
atom, 5 functor, 5, 59
atomic goal, 5
gae (ground arithmetic expression), 43
backtracking, 15, 31 goal, 6
binding order, 43
Boolean constraint, see constraint heuristics
bracketless prefix form, 42 value choice, 105
variable choice, 104
choice point, 15
Clark completion, 320 implication, 8
clause, 5, 284 infix form, 42
 Index 329

interpretation, 90 greedy, 139

declarative, 9 incomplete, 105, 133
procedural, 9 limited discrepancy, 144
iterator, 111 shallow backtracking, 208
top-down, 103
Kowalski, R. A. K., ix search tree, 104
shaving, 268
labelling, 104 shelf, 149
linear constraint, see constraint statement
list, 21 alternative choice, 30
empty, 21 atomic, 30
head of, 21 block, 30
tail of, 21 structure, 59
logic program, 4 substitution, 11
Sutherland, I., x
Mackworth, A., x
Martelli, M., 12 term, 5
meta-variable, 61 base, 4
mgu, 12 compound, 5
mixed integer programming, 282 ground, 5
modelling, 94
Montanari, U., x, 12 unification algorithm, 12
most general unifier, see mgu unification problem, 12
unifier, 11
negation, 64, 92
number, 42 variable, 4
anonymous, 7
occur check, 19 Boolean, 92
operator, 52 continuous, 101, 256
arithmetic, 42 decision, 91
declaration of, 52 local, 33
logical, 20
predicate, 5 non-logical, 146
arity of, 5 variable declaration, 164
dynamic, 70
name of, 5
prefix form, 42
procedure, 6
definition of, 6
program, 6
Prolog, x
propagation threshold, 262
propositional satisfiability problem, 284
pure Prolog, 4
pure Prolog program, see program

query, 6
atomic, 6

range declaration, 164

recursion, 25
Robinson, J. A., ix
rule, 6
body of, 6
head of, 6

Schimpf, J., 321

search
backtrack-free, 206
backtracking, 104
branch and bound, 106
complete, 105, 133